GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-05-01 00:04:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

Enable vectorization with half-packets

Browse Source
This commit is contained in:
Gael Guennebaud 2015-08-07 20:05:31 +02:00
parent 6245591349
commit febcce34f1
5 changed files with 151 additions and 11 deletions

6
Eigen/src/Core/AssignEvaluator.h

@ -29,8 +29,8 @@ struct copy_using_evaluator_traits
{
typedef typename DstEvaluator::XprType Dst;
typedef typename Dst::Scalar DstScalar;
// TODO recursively find best packet size
typedef typename packet_traits<DstScalar>::type PacketType;
// TODO distinguish between linear traversal and inner-traversals
typedef typename find_best_packet<DstScalar,Dst::SizeAtCompileTime>::type PacketType;
enum {
DstFlags = DstEvaluator::Flags,
@ -55,7 +55,7 @@ private:
: int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
: int(Dst::MaxRowsAtCompileTime),
MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
PacketSize = packet_traits<DstScalar>::size
PacketSize = unpacket_traits<PacketType>::size
};
enum {

3
Eigen/src/Core/DenseBase.h

@ -76,7 +76,6 @@ template<typename Derived> class DenseBase
* It is an alias for the Scalar type */
typedef Scalar value_type;
typedef typename internal::packet_traits<Scalar>::type PacketScalar;
typedef typename NumTraits<Scalar>::Real RealScalar;
typedef internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
@ -178,6 +177,8 @@ template<typename Derived> class DenseBase
InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret,
OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret
};
typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar;
enum { IsPlainObjectBase = 0 };

3
Eigen/src/Core/Matrix.h

@ -141,7 +141,8 @@ struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
{
private:
// TODO find ideal packet size
typedef typename packet_traits<_Scalar>::type PacketScalar;
enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret };
typedef typename find_best_packet<_Scalar,size>::type PacketScalar;
enum {
row_major_bit = _Options&RowMajor ? RowMajorBit : 0,
is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic,

22
Eigen/src/Core/util/XprHelper.h

@ -126,6 +126,28 @@ template<typename T> struct unpacket_traits
};
};
template<int Size, typename PacketType,
bool Stop = (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
struct find_best_packet_helper;
template< int Size, typename PacketType>
struct find_best_packet_helper<Size,PacketType,true>
{
typedef PacketType type;
};
template<int Size, typename PacketType>
struct find_best_packet_helper<Size,PacketType,false>
{
typedef typename find_best_packet_helper<Size,typename unpacket_traits<PacketType>::half>::type type;
};
template<typename T, int Size>
struct find_best_packet
{
typedef typename find_best_packet_helper<Size,typename packet_traits<T>::type>::type type;
};
#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
template<int ArrayBytes, int AlignmentBytes,
bool Match = bool((ArrayBytes%AlignmentBytes)==0),

128
test/vectorization_logic.cpp

@ -101,11 +101,16 @@ bool test_redux(const Xpr&, int traversal, int unrolling)
return res;
}
template<typename Scalar, bool Enable = internal::packet_traits<Scalar>::Vectorizable> struct vectorization_logic
template<typename Scalar, bool Enable = internal::packet_traits<Scalar>::Vectorizable>
struct vectorization_logic
{
typedef internal::packet_traits<Scalar> PacketTraits;
typedef typename internal::packet_traits<Scalar>::type PacketType;
typedef typename internal::unpacket_traits<PacketType>::half HalfPacketType;
enum {
PacketSize = PacketTraits::size
PacketSize = internal::unpacket_traits<PacketType>::size,
HalfPacketSize = internal::unpacket_traits<HalfPacketType>::size
};
static void run()
{
@ -179,9 +184,9 @@ template<typename Scalar, bool Enable = internal::packet_traits<Scalar>::Vectori
VERIFY(test_assign(Matrix3(),Matrix3().cwiseQuotient(Matrix3()),
PacketTraits::HasDiv ? LinearVectorizedTraversal : LinearTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix<Scalar,17,17>(),Matrix<Scalar,17,17>()+Matrix<Scalar,17,17>(),
LinearTraversal,NoUnrolling));
HalfPacketSize==1 ? InnerVectorizedTraversal : LinearTraversal,NoUnrolling));
VERIFY(test_assign(Matrix11(),Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(2,3)+Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(8,4),
DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling));
@ -221,8 +226,6 @@ template<typename Scalar, bool Enable = internal::packet_traits<Scalar>::Vectori
VERIFY(test_redux(VectorX(10),
LinearVectorizedTraversal,NoUnrolling));
}
};
@ -231,6 +234,113 @@ template<typename Scalar> struct vectorization_logic<Scalar,false>
static void run() {}
};
template<typename Scalar, bool Enable = !internal::is_same<typename internal::unpacket_traits<typename internal::packet_traits<Scalar>::type>::half,
typename internal::packet_traits<Scalar>::type>::value >
struct vectorization_logic_half
{
typedef internal::packet_traits<Scalar> PacketTraits;
typedef typename internal::unpacket_traits<typename internal::packet_traits<Scalar>::type>::half PacketType;
enum {
PacketSize = internal::unpacket_traits<PacketType>::size
};
static void run()
{
typedef Matrix<Scalar,PacketSize,1> Vector1;
typedef Matrix<Scalar,PacketSize,PacketSize> Matrix11;
typedef Matrix<Scalar,5*PacketSize,7,ColMajor> Matrix57;
typedef Matrix<Scalar,5*PacketSize,7,DontAlign|ColMajor> Matrix57u;
// typedef Matrix<Scalar,(Matrix11::Flags&RowMajorBit)?16:4*PacketSize,(Matrix11::Flags&RowMajorBit)?4*PacketSize:16> Matrix44;
// typedef Matrix<Scalar,(Matrix11::Flags&RowMajorBit)?16:4*PacketSize,(Matrix11::Flags&RowMajorBit)?4*PacketSize:16,DontAlign|EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION> Matrix44u;
// typedef Matrix<Scalar,4*PacketSize,4*PacketSize,ColMajor> Matrix44c;
// typedef Matrix<Scalar,4*PacketSize,4*PacketSize,RowMajor> Matrix44r;
typedef Matrix<Scalar,
(PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==8 ? 2 : PacketSize==4 ? 2 : PacketSize==2 ? 2 : /*PacketSize==1 ?*/ 1)
> Matrix1;
typedef Matrix<Scalar,
(PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==8 ? 2 : PacketSize==4 ? 2 : PacketSize==2 ? 2 : /*PacketSize==1 ?*/ 1),
DontAlign|((Matrix1::Flags&RowMajorBit)?RowMajor:ColMajor)> Matrix1u;
// this type is made such that it can only be vectorized when viewed as a linear 1D vector
typedef Matrix<Scalar,
(PacketSize==8 ? 4 : PacketSize==4 ? 6 : PacketSize==2 ? ((Matrix11::Flags&RowMajorBit)?2:3) : /*PacketSize==1 ?*/ 1),
(PacketSize==8 ? 6 : PacketSize==4 ? 2 : PacketSize==2 ? ((Matrix11::Flags&RowMajorBit)?3:2) : /*PacketSize==1 ?*/ 3)
> Matrix3;
#if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT
VERIFY(test_assign(Vector1(),Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1()+Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1().cwiseProduct(Vector1()),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1().template cast<Scalar>(),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1()+Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector1(),Vector1().cwiseProduct(Vector1()),
InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix57(),Matrix57()+Matrix57(),
InnerVectorizedTraversal,InnerUnrolling));
VERIFY(test_assign(Matrix57u(),Matrix57()+Matrix57(),
LinearTraversal,NoUnrolling));
VERIFY(test_assign(Matrix1u(),Matrix1()+Matrix1(),
LinearTraversal,CompleteUnrolling));
if(PacketSize>1)
{
typedef Matrix<Scalar,3,3,ColMajor> Matrix33c;
VERIFY(test_assign(Matrix33c().row(2),Matrix33c().row(1)+Matrix33c().row(1),
LinearTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix33c().col(0),Matrix33c().col(1)+Matrix33c().col(1),
LinearTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix3(),Matrix3().cwiseQuotient(Matrix3()),
PacketTraits::HasDiv ? LinearVectorizedTraversal : LinearTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix<Scalar,17,17>(),Matrix<Scalar,17,17>()+Matrix<Scalar,17,17>(),
LinearTraversal,NoUnrolling));
VERIFY(test_assign(Matrix11(),Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(2,3)+Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(8,4),
DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling));
}
VERIFY(test_redux(Matrix3(),
LinearVectorizedTraversal,CompleteUnrolling));
VERIFY(test_redux(Matrix57(),
LinearVectorizedTraversal,CompleteUnrolling));
VERIFY(test_redux(Matrix57().template block<PacketSize,3>(1,0),
DefaultTraversal,CompleteUnrolling));
VERIFY((test_assign<
Map<Matrix<Scalar,EIGEN_PLAIN_ENUM_MAX(2,PacketSize),EIGEN_PLAIN_ENUM_MAX(2,PacketSize)>, AlignedMax, InnerStride<3*PacketSize> >,
Matrix<Scalar,EIGEN_PLAIN_ENUM_MAX(2,PacketSize),EIGEN_PLAIN_ENUM_MAX(2,PacketSize)>
>(DefaultTraversal,CompleteUnrolling)));
VERIFY((test_assign(Matrix57(), Matrix<Scalar,5*PacketSize,3>()*Matrix<Scalar,3,7>(),
InnerVectorizedTraversal, CompleteUnrolling)));
#endif
}
};
template<typename Scalar> struct vectorization_logic_half<Scalar,false>
{
static void run() {}
};
void test_vectorization_logic()
{
@ -242,6 +352,12 @@ void test_vectorization_logic()
CALL_SUBTEST( vectorization_logic<std::complex<float> >::run() );
CALL_SUBTEST( vectorization_logic<std::complex<double> >::run() );
CALL_SUBTEST( vectorization_logic_half<int>::run() );
CALL_SUBTEST( vectorization_logic_half<float>::run() );
CALL_SUBTEST( vectorization_logic_half<double>::run() );
CALL_SUBTEST( vectorization_logic_half<std::complex<float> >::run() );
CALL_SUBTEST( vectorization_logic_half<std::complex<double> >::run() );
if(internal::packet_traits<float>::Vectorizable)
{
VERIFY(test_assign(Matrix<float,3,3>(),Matrix<float,3,3>()+Matrix<float,3,3>(),