aboutsummaryrefslogtreecommitdiff
path: root/test/vectorization_logic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'test/vectorization_logic.cpp')
-rw-r--r--test/vectorization_logic.cpp114
1 files changed, 62 insertions, 52 deletions
diff --git a/test/vectorization_logic.cpp b/test/vectorization_logic.cpp
index 83c1439ad..97c0bdad9 100644
--- a/test/vectorization_logic.cpp
+++ b/test/vectorization_logic.cpp
@@ -22,6 +22,14 @@
#include "main.h"
#include <typeinfo>
+// Disable "ignoring attributes on template argument"
+// for packet_traits<Packet*>
+// => The only workaround would be to wrap _m128 and the likes
+// within wrappers.
+#if EIGEN_GNUC_AT_LEAST(6,0)
+ #pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
using internal::demangle_flags;
using internal::demangle_traversal;
using internal::demangle_unrolling;
@@ -29,6 +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)
@@ -53,6 +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)
@@ -109,26 +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
@@ -141,14 +151,6 @@ struct vectorization_logic
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(Matrix44(),Matrix44()+Matrix44(),
InnerVectorizedTraversal,InnerUnrolling));
@@ -157,11 +159,11 @@ struct vectorization_logic
EIGEN_UNALIGNED_VECTORIZE ? InnerUnrolling : NoUnrolling));
VERIFY(test_assign(Matrix1(),Matrix1()+Matrix1(),
- (Matrix1::InnerSizeAtCompileTime % PacketSize)==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal,
+ (int(Matrix1::InnerSizeAtCompileTime) % int(PacketSize))==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal,
CompleteUnrolling));
VERIFY(test_assign(Matrix1u(),Matrix1()+Matrix1(),
- EIGEN_UNALIGNED_VECTORIZE ? ((Matrix1::InnerSizeAtCompileTime % PacketSize)==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal)
+ EIGEN_UNALIGNED_VECTORIZE ? ((int(Matrix1::InnerSizeAtCompileTime) % int(PacketSize))==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal)
: LinearTraversal, CompleteUnrolling));
VERIFY(test_assign(Matrix44c().col(1),Matrix44c().col(2)+Matrix44c().col(3),
@@ -177,24 +179,25 @@ struct vectorization_logic
VERIFY(test_assign(Matrix33c().row(2),Matrix33c().row(1)+Matrix33c().row(1),
LinearTraversal,CompleteUnrolling));
VERIFY(test_assign(Vector3(),Vector3()+Vector3(),
- EIGEN_UNALIGNED_VECTORIZE ? (HalfPacketSize==1 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : (HalfPacketSize==1 ? InnerVectorizedTraversal : LinearTraversal), CompleteUnrolling));
+ sizeof(Scalar)==16 ? InnerVectorizedTraversal : (EIGEN_UNALIGNED_VECTORIZE ? LinearVectorizedTraversal : LinearTraversal), CompleteUnrolling));
VERIFY(test_assign(Matrix33c().col(0),Matrix33c().col(1)+Matrix33c().col(1),
- EIGEN_UNALIGNED_VECTORIZE ? (HalfPacketSize==1 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : (HalfPacketSize==1 ? SliceVectorizedTraversal : LinearTraversal),
- ((!EIGEN_UNALIGNED_VECTORIZE) && HalfPacketSize==1) ? NoUnrolling : CompleteUnrolling));
+ EIGEN_UNALIGNED_VECTORIZE ? (sizeof(Scalar)==16 ? InnerVectorizedTraversal : LinearVectorizedTraversal)
+ : (sizeof(Scalar)==16 ? SliceVectorizedTraversal : LinearTraversal),
+ ((!EIGEN_UNALIGNED_VECTORIZE) && (sizeof(Scalar)==16)) ? NoUnrolling : CompleteUnrolling));
VERIFY(test_assign(Matrix3(),Matrix3().cwiseProduct(Matrix3()),
LinearVectorizedTraversal,CompleteUnrolling));
VERIFY(test_assign(Matrix<Scalar,17,17>(),Matrix<Scalar,17,17>()+Matrix<Scalar,17,17>(),
- HalfPacketSize==1 ? InnerVectorizedTraversal :
+ sizeof(Scalar)==16 ? InnerVectorizedTraversal :
EIGEN_UNALIGNED_VECTORIZE ? LinearVectorizedTraversal :
- LinearTraversal,
+ LinearTraversal,
NoUnrolling));
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(),
@@ -207,6 +210,12 @@ struct vectorization_logic
VERIFY(test_redux(Vector1(),
LinearVectorizedTraversal,CompleteUnrolling));
+ VERIFY(test_redux(Vector1().array()*Vector1().array(),
+ LinearVectorizedTraversal,CompleteUnrolling));
+
+ VERIFY(test_redux((Vector1().array()*Vector1().array()).col(0),
+ LinearVectorizedTraversal,CompleteUnrolling));
+
VERIFY(test_redux(Matrix<Scalar,PacketSize,3>(),
LinearVectorizedTraversal,CompleteUnrolling));
@@ -216,8 +225,13 @@ 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));
@@ -269,25 +283,21 @@ struct vectorization_logic_half
typedef Matrix<Scalar,5*PacketSize,7,ColMajor> Matrix57;
typedef Matrix<Scalar,3*PacketSize,5,ColMajor> Matrix35;
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)
+ (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
@@ -306,14 +316,6 @@ struct vectorization_logic_half
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));
@@ -322,7 +324,7 @@ struct vectorization_logic_half
EIGEN_UNALIGNED_VECTORIZE ? InnerUnrolling : NoUnrolling));
VERIFY(test_assign(Matrix1u(),Matrix1()+Matrix1(),
- EIGEN_UNALIGNED_VECTORIZE ? ((Matrix1::InnerSizeAtCompileTime % PacketSize)==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : LinearTraversal,CompleteUnrolling));
+ EIGEN_UNALIGNED_VECTORIZE ? ((int(Matrix1::InnerSizeAtCompileTime) % int(PacketSize))==0 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : LinearTraversal,CompleteUnrolling));
if(PacketSize>1)
{
@@ -330,17 +332,20 @@ struct vectorization_logic_half
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),
- EIGEN_UNALIGNED_VECTORIZE ? (PacketSize==1 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : LinearTraversal,CompleteUnrolling));
+ EIGEN_UNALIGNED_VECTORIZE ? (sizeof(Scalar)==16 ? InnerVectorizedTraversal : LinearVectorizedTraversal)
+ : (sizeof(Scalar)==16 ? SliceVectorizedTraversal : LinearTraversal),
+ ((!EIGEN_UNALIGNED_VECTORIZE) && (sizeof(Scalar)==16)) ? NoUnrolling : 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>(),
- EIGEN_UNALIGNED_VECTORIZE ? (PacketSize==1 ? InnerVectorizedTraversal : LinearVectorizedTraversal) : LinearTraversal,
+ sizeof(Scalar)==16 ? InnerVectorizedTraversal : (EIGEN_UNALIGNED_VECTORIZE ? LinearVectorizedTraversal : 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),
- EIGEN_UNALIGNED_VECTORIZE ? InnerVectorizedTraversal : DefaultTraversal,PacketSize>4?InnerUnrolling:CompleteUnrolling));
+ EIGEN_UNALIGNED_VECTORIZE ? InnerVectorizedTraversal : DefaultTraversal,InnerUnrolling+CompleteUnrolling));
+
VERIFY(test_assign(Vector1(),Matrix11()*Vector1(),
InnerVectorizedTraversal,CompleteUnrolling));
@@ -361,16 +366,21 @@ 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
}
};
@@ -380,7 +390,7 @@ template<typename Scalar> struct vectorization_logic_half<Scalar,false>
static void run() {}
};
-void test_vectorization_logic()
+EIGEN_DECLARE_TEST(vectorization_logic)
{
#ifdef EIGEN_VECTORIZE
generated by cgit v1.2.3 (git 2.25.1) at 2024-05-12 23:57:10 +0000