Merge with default.

This commit is contained in:
Gael Guennebaud
2018-09-23 21:52:58 +02:00
12 changed files with 143 additions and 93 deletions

View File

@@ -30,7 +30,14 @@ static int nb_storeu;
EIGEN_DECLARE_TEST(unalignedcount)
{
#if defined(EIGEN_VECTORIZE_AVX)
#if defined(EIGEN_VECTORIZE_AVX512)
VectorXf a(48), b(48);
VERIFY_ALIGNED_UNALIGNED_COUNT(a += b, 6, 0, 3, 0);
VERIFY_ALIGNED_UNALIGNED_COUNT(a.segment(0,48) += b.segment(0,48), 3, 3, 3, 0);
VERIFY_ALIGNED_UNALIGNED_COUNT(a.segment(0,48) -= b.segment(0,48), 3, 3, 3, 0);
VERIFY_ALIGNED_UNALIGNED_COUNT(a.segment(0,48) *= 3.5, 3, 0, 3, 0);
VERIFY_ALIGNED_UNALIGNED_COUNT(a.segment(0,48) /= 3.5, 3, 0, 3, 0);
#elif defined(EIGEN_VECTORIZE_AVX)
VectorXf a(40), b(40);
VERIFY_ALIGNED_UNALIGNED_COUNT(a += b, 10, 0, 5, 0);
VERIFY_ALIGNED_UNALIGNED_COUNT(a.segment(0,40) += b.segment(0,40), 5, 5, 5, 0);

View File

@@ -37,6 +37,7 @@ using internal::demangle_unrolling;
template<typename Dst, typename Src>
bool test_assign(const Dst&, const Src&, int traversal, int unrolling)
{
EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src);
typedef internal::copy_using_evaluator_traits<internal::evaluator<Dst>,internal::evaluator<Src>, internal::assign_op<typename Dst::Scalar,typename Src::Scalar> > traits;
bool res = traits::Traversal==traversal;
if(unrolling==InnerUnrolling+CompleteUnrolling)
@@ -61,6 +62,7 @@ bool test_assign(const Dst&, const Src&, int traversal, int unrolling)
template<typename Dst, typename Src>
bool test_assign(int traversal, int unrolling)
{
EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src);
typedef internal::copy_using_evaluator_traits<internal::evaluator<Dst>,internal::evaluator<Src>, internal::assign_op<typename Dst::Scalar,typename Src::Scalar> > traits;
bool res = traits::Traversal==traversal && traits::Unrolling==unrolling;
if(!res)
@@ -117,26 +119,26 @@ struct vectorization_logic
typedef Matrix<Scalar,Dynamic,1> VectorX;
typedef Matrix<Scalar,Dynamic,Dynamic> MatrixXX;
typedef Matrix<Scalar,PacketSize,PacketSize> Matrix11;
typedef Matrix<Scalar,2*PacketSize,2*PacketSize> Matrix22;
typedef Matrix<Scalar,(Matrix11::Flags&RowMajorBit)?8:2*PacketSize,(Matrix11::Flags&RowMajorBit)?2*PacketSize:8> Matrix22;
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)
(PacketSize==16 ? 8 : PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 2 : 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),
(PacketSize==16 ? 8 : PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 2 : 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)
(PacketSize==16 ? 4 : PacketSize==8 ? 4 : PacketSize==4 ? 6 : PacketSize==2 ? ((Matrix11::Flags&RowMajorBit)?2:3) : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 12 : 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
@@ -202,7 +204,7 @@ struct vectorization_logic
VERIFY(test_assign(Matrix11(), Matrix11()+Matrix11(),InnerVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix11(),Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(2,3)+Matrix<Scalar,17,17>().template block<PacketSize,PacketSize>(8,4),
VERIFY(test_assign(Matrix11(),Matrix<Scalar,21,21>().template block<PacketSize,PacketSize>(2,3)+Matrix<Scalar,21,21>().template block<PacketSize,PacketSize>(3,2),
(EIGEN_UNALIGNED_VECTORIZE) ? InnerVectorizedTraversal : DefaultTraversal, CompleteUnrolling|InnerUnrolling));
VERIFY(test_assign(Vector1(),Matrix11()*Vector1(),
@@ -230,8 +232,13 @@ struct vectorization_logic
VERIFY(test_redux(Matrix44(),
LinearVectorizedTraversal,NoUnrolling));
VERIFY(test_redux(Matrix44().template block<(Matrix1::Flags&RowMajorBit)?4:PacketSize,(Matrix1::Flags&RowMajorBit)?PacketSize:4>(1,2),
DefaultTraversal,CompleteUnrolling));
if(PacketSize>1) {
VERIFY(test_redux(Matrix44().template block<(Matrix1::Flags&RowMajorBit)?4:PacketSize,(Matrix1::Flags&RowMajorBit)?PacketSize:4>(1,2),
SliceVectorizedTraversal,CompleteUnrolling));
VERIFY(test_redux(Matrix44().template block<(Matrix1::Flags&RowMajorBit)?2:PacketSize,(Matrix1::Flags&RowMajorBit)?PacketSize:2>(1,2),
DefaultTraversal,CompleteUnrolling));
}
VERIFY(test_redux(Matrix44c().template block<2*PacketSize,1>(1,2),
LinearVectorizedTraversal,CompleteUnrolling));
@@ -289,19 +296,19 @@ struct vectorization_logic_half
// 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)
(PacketSize==16 ? 8 : PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 2 : 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),
(PacketSize==16 ? 8 : PacketSize==8 ? 4 : PacketSize==4 ? 2 : PacketSize==2 ? 1 : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 2 : 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)
(PacketSize==16 ? 4 : PacketSize==8 ? 4 : PacketSize==4 ? 6 : PacketSize==2 ? ((Matrix11::Flags&RowMajorBit)?2:3) : /*PacketSize==1 ?*/ 1),
(PacketSize==16 ? 12 : 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
@@ -354,7 +361,8 @@ struct vectorization_logic_half
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),
EIGEN_UNALIGNED_VECTORIZE ? InnerVectorizedTraversal : DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling));
EIGEN_UNALIGNED_VECTORIZE ? InnerVectorizedTraversal : DefaultTraversal,InnerUnrolling+CompleteUnrolling));
VERIFY(test_assign(Vector1(),Matrix11()*Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
@@ -375,16 +383,21 @@ struct vectorization_logic_half
VERIFY(test_redux(Matrix35(),
LinearVectorizedTraversal,CompleteUnrolling));
VERIFY(test_redux(Matrix57().template block<PacketSize,3>(1,0),
DefaultTraversal,CompleteUnrolling));
VERIFY(test_redux(Matrix57().template block<PacketSize==1?2:PacketSize,3>(1,0),
SliceVectorizedTraversal,CompleteUnrolling));
if(PacketSize>1) {
VERIFY(test_redux(Matrix57().template block<PacketSize,2>(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)));
>(DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling)));
VERIFY((test_assign(Matrix57(), Matrix<Scalar,5*PacketSize,3>()*Matrix<Scalar,3,7>(),
InnerVectorizedTraversal, InnerUnrolling|CompleteUnrolling)));
InnerVectorizedTraversal, InnerUnrolling+CompleteUnrolling)));
#endif
}
};