From ddf775363147fc7ee778b42c21b642f085193f55 Mon Sep 17 00:00:00 2001
From: Gael Guennebaud <g.gael@free.fr>
Date: Thu, 1 Aug 2013 16:26:57 +0200
Subject: [PATCH] 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<Derived,OtherDerived,false,false> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.derived()); }
   template<typename ActualDerived, typename ActualOtherDerived>
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { other.evalTo(dst); return dst; }
 };
 template<typename Derived, typename OtherDerived>
@@ -544,6 +545,7 @@ struct assign_selector<Derived,OtherDerived,false,true> {
   EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& run(Derived& dst, const OtherDerived& other) { return dst.lazyAssign(other.transpose()); }
   template<typename ActualDerived, typename ActualOtherDerived>
+  EIGEN_DEVICE_FUNC
   static EIGEN_STRONG_INLINE Derived& evalTo(ActualDerived& dst, const ActualOtherDerived& other) { Transpose<ActualDerived> dstTrans(dst); other.evalTo(dstTrans); return dst; }
 };
 template<typename Derived, typename OtherDerived>
@@ -556,18 +558,21 @@ struct assign_selector<Derived,OtherDerived,true,true> {
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
   return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other)
 {
   return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
 }
 
 template<typename Derived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other)
 {
   return internal::assign_selector<Derived,Derived>::run(derived(), other.derived());
@@ -575,6 +580,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& ot
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
 {
   return internal::assign_selector<Derived,OtherDerived>::run(derived(), other.derived());
@@ -582,6 +588,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<Othe
 
 template<typename Derived>
 template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other)
 {
   return internal::assign_selector<Derived,OtherDerived,false>::evalTo(derived(), other.derived());
@@ -589,6 +596,7 @@ EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<Othe
 
 template<typename Derived>
 template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
 EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
 {
   return internal::assign_selector<Derived,OtherDerived,false>::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<typename Derived, bool Big = (Derived::SizeAtCompileTime>=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<typename Derived>
 struct setIdentity_impl<Derived, true>
 {
   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<typename Derived> class DenseBase
     /** swaps *this with the expression \a other.
       *
       */
-    template<typename OtherDerived> EIGEN_DEVICE_FUNC
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(const DenseBase<OtherDerived>& other,
               int = OtherDerived::ThisConstantIsPrivateInPlainObjectBase)
     {
@@ -401,7 +402,8 @@ template<typename Derived> class DenseBase
     /** swaps *this with the matrix or array \a other.
       *
       */
-    template<typename OtherDerived> EIGEN_DEVICE_FUNC
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(PlainObjectBase<OtherDerived>& other)
     {
       SwapWrapper<Derived>(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<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(MatrixBase<OtherDerived> 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<typename Derived> class MatrixBase
     template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; };
     template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; };
 
-    template<unsigned int Mode> typename TriangularViewReturnType<Mode>::Type triangularView();
-    template<unsigned int Mode> typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename TriangularViewReturnType<Mode>::Type triangularView();
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
 
     template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SelfAdjointView<Derived, UpLo> Type; };
     template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView<const Derived, UpLo> Type; };
 
-    template<unsigned int UpLo> typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
-    template<unsigned int UpLo> typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
+    template<unsigned int UpLo> 
+    EIGEN_DEVICE_FUNC
+    typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
 
     const SparseView<Derived> sparseView(const Scalar& m_reference = Scalar(0),
                                          const typename NumTraits<Scalar>::Real& m_epsilon = NumTraits<Scalar>::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<const Derived> asDiagonal() const;
     const PermutationWrapper<const Derived> asPermutation() const;
 
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity();
+    EIGEN_DEVICE_FUNC
     Derived& setIdentity(Index rows, Index cols);
 
     bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
@@ -430,11 +440,17 @@ template<typename Derived> class MatrixBase
     };
     #endif // EIGEN_PARSED_BY_DOXYGEN
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     typename cross_product_return_type<OtherDerived>::type
     cross(const MatrixBase<OtherDerived>& other) const;
+    
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
+    
+    EIGEN_DEVICE_FUNC
     PlainObject unitOrthogonal(void) const;
+    
     Matrix<Scalar,3,1> 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<typename T> struct GenericNumTraits
                    >::type NonInteger;
   typedef T Nested;
 
-  static inline Real epsilon() { return std::numeric_limits<T>::epsilon(); }
+  EIGEN_DEVICE_FUNC
+  static inline Real epsilon()
+  {
+    #if defined(__CUDA_ARCH__)
+    return internal::device::numeric_limits<T>::epsilon();
+    #else
+    return std::numeric_limits<T>::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<T>::max)(); }
-  EIGEN_DEVICE_FUNC
   static inline T lowest()  { return IsInteger ? (std::numeric_limits<T>::min)() : (-(std::numeric_limits<T>::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<Derived>::type
     enum { NeedsToAlign = SizeAtCompileTime != Dynamic && (internal::traits<Derived>::Flags & AlignedBit) != 0 };
     EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
 
+    EIGEN_DEVICE_FUNC
     Base& base() { return *static_cast<Base*>(this); }
+    EIGEN_DEVICE_FUNC
     const Base& base() const { return *static_cast<const Base*>(this); }
 
     EIGEN_DEVICE_FUNC
@@ -351,6 +353,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::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<Derived>::run(*this, nbRows, nbCols);
@@ -363,6 +366,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::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<Derived>::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<Derived>::type
       *
       * When values are appended, they will be uninitialized.
       */
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResize(Index size)
     {
       internal::conservative_resize_like_impl<Derived>::run(*this, size);
@@ -405,6 +411,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * appended to the matrix they will copied from \c other.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase<OtherDerived>& other)
     {
       internal::conservative_resize_like_impl<Derived,OtherDerived>::run(*this, other);
@@ -647,6 +654,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 
     template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE void _init2(Index nbRows, Index nbCols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
     {
       EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
@@ -670,6 +678,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
       * data pointers.
       */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void _swap(DenseBase<OtherDerived> const & other)
     {
       enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic };
@@ -790,6 +799,7 @@ struct conservative_resize_like_impl<Derived,OtherDerived,true>
 template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
 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<typename MatrixTypeA, typename MatrixTypeB>
 struct matrix_swap_impl<MatrixTypeA, MatrixTypeB, true>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(MatrixTypeA& a, MatrixTypeB& b)
   {
     static_cast<typename MatrixTypeA::Base&>(a).m_storage.swap(static_cast<typename MatrixTypeB::Base&>(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<typename MatrixType, unsigned int UpLo> 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<typename MatrixType, unsigned int UpLo> 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<typename MatrixType, unsigned int UpLo> 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<MatrixTypeNestedCleaned*>(&m_matrix); }
 
     /** Efficient self-adjoint matrix times vector/matrix product */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     SelfadjointProductMatrix<MatrixType,Mode,false,OtherDerived,0,OtherDerived::IsVectorAtCompileTime>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
@@ -113,6 +124,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
 
     /** Efficient vector/matrix times self-adjoint matrix product */
     template<typename OtherDerived> friend
+    EIGEN_DEVICE_FUNC
     SelfadjointProductMatrix<OtherDerived,0,OtherDerived::IsVectorAtCompileTime,MatrixType,Mode,false>
     operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
     {
@@ -132,6 +144,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar)
       */
     template<typename DerivedU, typename DerivedV>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha = Scalar(1));
 
     /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
@@ -145,6 +158,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar)
       */
     template<typename DerivedU>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1));
 
 /////////// Cholesky module ///////////
@@ -159,11 +173,14 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
     /** Return type of eigenvalues() */
     typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType;
 
+    EIGEN_DEVICE_FUNC
     EigenvaluesReturnType eigenvalues() const;
+    EIGEN_DEVICE_FUNC
     RealScalar operatorNorm() const;
     
     #ifdef EIGEN2_SUPPORT
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& operator=(const MatrixBase<OtherDerived>& other)
     {
       enum {
@@ -174,6 +191,7 @@ template<typename MatrixType, unsigned int UpLo> class SelfAdjointView
       return *this;
     }
     template<typename OtherMatrixType, unsigned int OtherMode>
+    EIGEN_DEVICE_FUNC
     SelfAdjointView& operator=(const TriangularView<OtherMatrixType, OtherMode>& other)
     {
       enum {
@@ -209,6 +227,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), U
     row = (UnrollCount-1) % Derived1::RowsAtCompileTime
   };
 
+  EIGEN_DEVICE_FUNC
   static inline void run(Derived1 &dst, const Derived2 &src)
   {
     triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), UnrollCount-1, ClearOpposite>::run(dst, src);
@@ -223,6 +242,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Upper), U
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, 0, ClearOpposite>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(Derived1 &, const Derived2 &) {}
 };
 
@@ -234,6 +254,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), U
     row = (UnrollCount-1) % Derived1::RowsAtCompileTime
   };
 
+  EIGEN_DEVICE_FUNC
   static inline void run(Derived1 &dst, const Derived2 &src)
   {
     triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), UnrollCount-1, ClearOpposite>::run(dst, src);
@@ -248,6 +269,7 @@ struct triangular_assignment_selector<Derived1, Derived2, (SelfAdjoint|Lower), U
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, 0, ClearOpposite>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(Derived1 &, const Derived2 &) {}
 };
 
@@ -255,6 +277,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Upper, Dynamic, ClearOpposite>
 {
   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<Derived1, Derived2, SelfAdjoint|Upper, Dyn
 template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, SelfAdjoint|Lower, Dynamic, ClearOpposite>
 {
+  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<typename ExpressionType> class SwapWrapper
     EIGEN_DENSE_PUBLIC_INTERFACE(SwapWrapper)
     typedef typename internal::packet_traits<Scalar>::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<typename ExpressionType> 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<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void copyCoeff(Index rowId, Index colId, const DenseBase<OtherDerived>& other)
     {
       OtherDerived& _other = other.const_cast_derived();
@@ -81,6 +93,7 @@ template<typename ExpressionType> class SwapWrapper
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void copyCoeff(Index index, const DenseBase<OtherDerived>& other)
     {
       OtherDerived& _other = other.const_cast_derived();
@@ -115,6 +128,7 @@ template<typename ExpressionType> class SwapWrapper
       _other.template writePacket<LoadMode>(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<typename Derived> class TriangularBase : public EigenBase<Derived>
     typedef typename internal::traits<Derived>::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<typename Other>
+    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<typename Derived> class TriangularBase : public EigenBase<Derived>
     }
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC
     inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
     inline Derived& derived() { return *static_cast<Derived*>(this); }
     #endif // not EIGEN_PARSED_BY_DOXYGEN
 
     template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
     void evalTo(MatrixBase<DenseDerived> &other) const;
     template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
     void evalToLazy(MatrixBase<DenseDerived> &other) const;
 
+    EIGEN_DEVICE_FUNC
     DenseMatrixType toDenseMatrix() const
     {
       DenseMatrixType res(rows(), cols());
@@ -189,36 +204,52 @@ template<typename _MatrixType, unsigned int _Mode> 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<typename Other> TriangularView&  operator+=(const DenseBase<Other>& other) { return *this = m_matrix + other.derived(); }
+    /** \sa MatrixBase::operator+=() */    
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularView&  operator+=(const DenseBase<Other>& other) { return *this = m_matrix + other.derived(); }
     /** \sa MatrixBase::operator-=() */
-    template<typename Other> TriangularView&  operator-=(const DenseBase<Other>& other) { return *this = m_matrix - other.derived(); }
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularView&  operator-=(const DenseBase<Other>& other) { return *this = m_matrix - other.derived(); }
     /** \sa MatrixBase::operator*=() */
+    EIGEN_DEVICE_FUNC
     TriangularView&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = m_matrix * other; }
     /** \sa MatrixBase::operator/=() */
+    EIGEN_DEVICE_FUNC
     TriangularView&  operator/=(const typename internal::traits<MatrixType>::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<typename _MatrixType, unsigned int _Mode> 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<MatrixTypeNestedCleaned*>(&m_matrix); }
 
     /** Assigns a triangular matrix to a triangular part of a dense matrix */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     TriangularView& operator=(const TriangularBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     TriangularView& operator=(const MatrixBase<OtherDerived>& other);
 
+    EIGEN_DEVICE_FUNC
     TriangularView& operator=(const TriangularView& other)
     { return *this = other.nestedExpression(); }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void lazyAssign(const TriangularBase<OtherDerived>& other);
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void lazyAssign(const MatrixBase<OtherDerived>& other);
 
     /** \sa MatrixBase::conjugate() */
+    EIGEN_DEVICE_FUNC
     inline TriangularView<MatrixConjugateReturnType,Mode> conjugate()
     { return m_matrix.conjugate(); }
     /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
     inline const TriangularView<MatrixConjugateReturnType,Mode> conjugate() const
     { return m_matrix.conjugate(); }
 
     /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
     inline const TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> adjoint() const
     { return m_matrix.adjoint(); }
 
     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
     inline TriangularView<Transpose<MatrixType>,TransposeMode> transpose()
     {
       EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
       return m_matrix.const_cast_derived().transpose();
     }
     /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
     inline const TriangularView<Transpose<MatrixType>,TransposeMode> transpose() const
     {
       return m_matrix.transpose();
@@ -278,6 +322,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
 
     /** Efficient triangular matrix times vector/matrix product */
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     TriangularProduct<Mode,true,MatrixType,false,OtherDerived, OtherDerived::IsVectorAtCompileTime>
     operator*(const MatrixBase<OtherDerived>& rhs) const
     {
@@ -288,6 +333,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
 
     /** Efficient vector/matrix times triangular matrix product */
     template<typename OtherDerived> friend
+    EIGEN_DEVICE_FUNC
     TriangularProduct<Mode,false,OtherDerived,OtherDerived::IsVectorAtCompileTime,MatrixType,false>
     operator*(const MatrixBase<OtherDerived>& lhs, const TriangularView& rhs)
     {
@@ -326,26 +372,32 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     #endif // EIGEN2_SUPPORT
 
     template<int Side, typename Other>
+    EIGEN_DEVICE_FUNC
     inline const internal::triangular_solve_retval<Side,TriangularView, Other>
     solve(const MatrixBase<Other>& other) const;
 
     template<int Side, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void solveInPlace(const MatrixBase<OtherDerived>& other) const;
 
     template<typename Other>
+    EIGEN_DEVICE_FUNC
     inline const internal::triangular_solve_retval<OnTheLeft,TriangularView, Other> 
     solve(const MatrixBase<Other>& other) const
     { return solve<OnTheLeft>(other); }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void solveInPlace(const MatrixBase<OtherDerived>& other) const
     { return solveInPlace<OnTheLeft>(other); }
 
+    EIGEN_DEVICE_FUNC
     const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
     {
       EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
       return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
     }
+    EIGEN_DEVICE_FUNC
     SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
     {
       EIGEN_STATIC_ASSERT((Mode&UnitDiag)==0,PROGRAMMING_ERROR);
@@ -353,18 +405,21 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(TriangularBase<OtherDerived> const & other)
     {
       TriangularView<SwapWrapper<MatrixType>,Mode>(const_cast<MatrixType&>(m_matrix)).lazyAssign(other.derived());
     }
 
     template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
     void swap(MatrixBase<OtherDerived> const & other)
     {
       SwapWrapper<MatrixType> swaper(const_cast<MatrixType&>(m_matrix));
       TriangularView<SwapWrapper<MatrixType>,Mode>(swaper).lazyAssign(other.derived());
     }
 
+    EIGEN_DEVICE_FUNC
     Scalar determinant() const
     {
       if (Mode & UnitDiag)
@@ -377,6 +432,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     
     // TODO simplify the following:
     template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
     {
       setZero();
@@ -384,12 +440,14 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     }
     
     template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator+=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
     {
       return assignProduct(other,1);
     }
     
     template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator-=(const ProductBase<ProductDerived, Lhs,Rhs>& other)
     {
       return assignProduct(other,-1);
@@ -397,6 +455,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     
     
     template<typename ProductDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator=(const ScaledProduct<ProductDerived>& other)
     {
       setZero();
@@ -404,12 +463,14 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
     }
     
     template<typename ProductDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator+=(const ScaledProduct<ProductDerived>& other)
     {
       return assignProduct(other,other.alpha());
     }
     
     template<typename ProductDerived>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& operator-=(const ScaledProduct<ProductDerived>& other)
     {
       return assignProduct(other,-other.alpha());
@@ -418,6 +479,7 @@ template<typename _MatrixType, unsigned int _Mode> class TriangularView
   protected:
     
     template<typename ProductDerived, typename Lhs, typename Rhs>
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE TriangularView& assignProduct(const ProductBase<ProductDerived, Lhs,Rhs>& 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<Derived1, Derived2, Mode, UnrollCount-1, ClearOpposite>::run(dst, src);
@@ -467,6 +530,7 @@ struct triangular_assignment_selector
 template<typename Derived1, typename Derived2, unsigned int Mode, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, Mode, 0, ClearOpposite>
 {
+  EIGEN_DEVICE_FUNC
   static inline void run(Derived1 &, const Derived2 &) {}
 };
 
@@ -475,6 +539,7 @@ struct triangular_assignment_selector<Derived1, Derived2, Upper, Dynamic, ClearO
 {
   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)
@@ -493,6 +558,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, Lower, Dynamic, ClearOpposite>
 {
   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<Derived1, Derived2, StrictlyUpper, Dynamic
 {
   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)
@@ -530,6 +597,7 @@ template<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, StrictlyLower, Dynamic, ClearOpposite>
 {
   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<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, UnitUpper, Dynamic, ClearOpposite>
 {
   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<typename Derived1, typename Derived2, bool ClearOpposite>
 struct triangular_assignment_selector<Derived1, Derived2, UnitLower, Dynamic, ClearOpposite>
 {
   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<bool Condition, typename T> struct enable_if;
 template<typename T> struct enable_if<true,T>
 { typedef T type; };
 
+#if defined(__CUDA_ARCH__)
+template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
 
+namespace device {
+template<typename T> struct numeric_limits
+{
+  EIGEN_DEVICE_FUNC
+  static T epsilon() { return 0; }
+};
+template<> struct numeric_limits<float>
+{
+  EIGEN_DEVICE_FUNC
+  static float epsilon() { return __FLT_EPSILON__; }
+};
+template<> struct numeric_limits<double>
+{
+  EIGEN_DEVICE_FUNC
+  static double epsilon() { return __DBL_EPSILON__; }
+};
+
+}
+
+#else
+template<typename T> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> class SelfAdjointEigenSolver
       *
       * \sa eigenvalues()
       */
+    EIGEN_DEVICE_FUNC
     const MatrixType& eigenvectors() const
     {
       eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
@@ -247,6 +253,7 @@ template<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> 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<typename _MatrixType> class SelfAdjointEigenSolver
       static_cast<GeneralizedSelfAdjointEigenSolver<MatrixType>*>(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<typename _MatrixType> class SelfAdjointEigenSolver
   */
 namespace internal {
 template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC
 static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n);
 }
 
 template<typename MatrixType>
+EIGEN_DEVICE_FUNC
 SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 ::compute(const MatrixType& matrix, int options)
 {
@@ -481,6 +497,7 @@ namespace internal {
   
 template<typename SolverType,int Size,bool IsComplex> 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<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
   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;
-    using std::atan2;
-    using std::cos;
-    using std::sin;
+    EIGEN_USING_STD_MATH(sqrt)
+    EIGEN_USING_STD_MATH(atan2)
+    EIGEN_USING_STD_MATH(cos)
+    EIGEN_USING_STD_MATH(sin)
     const Scalar s_inv3 = Scalar(1.0)/Scalar(3.0);
     const Scalar s_sqrt3 = sqrt(Scalar(3.0));
 
@@ -531,15 +549,16 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
 
     // Sort in increasing order.
     if (roots(0) >= 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<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
 };
 
 // 2x2 direct eigenvalues decomposition, code from Hauke Heibel
-template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2,false>
+template<typename SolverType> 
+struct direct_selfadjoint_eigenvalues<SolverType,2,false>
 {
   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<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2
     roots(1) = t1 + t0;
   }
   
+  EIGEN_DEVICE_FUNC
   static inline void run(SolverType& solver, const MatrixType& mat, int options)
   {
     using std::sqrt;
@@ -728,6 +750,7 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2
 }
 
 template<typename MatrixType>
+EIGEN_DEVICE_FUNC
 SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 ::computeDirect(const MatrixType& matrix, int options)
 {
@@ -737,6 +760,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 
 namespace internal {
 template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+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<Scalar>::Real RealScalar;
   typedef typename Derived::Index Index;
   typedef Matrix<Scalar,2,1> Vector2;
+  EIGEN_DEVICE_FUNC
   static inline VectorType run(const Derived& src)
   {
     VectorType perp = VectorType::Zero(src.size());
@@ -154,6 +155,7 @@ struct unitOrthogonal_selector<Derived,3>
   typedef typename plain_matrix_type<Derived>::type VectorType;
   typedef typename traits<Derived>::Scalar Scalar;
   typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
   static inline VectorType run(const Derived& src)
   {
     VectorType perp;
@@ -192,6 +194,7 @@ template<typename Derived>
 struct unitOrthogonal_selector<Derived,2>
 {
   typedef typename plain_matrix_type<Derived>::type VectorType;
+  EIGEN_DEVICE_FUNC
   static inline VectorType run(const Derived& src)
   { return VectorType(-numext::conj(src.y()), numext::conj(src.x())).normalized(); }
 };