GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-11 11:19:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix alignment issue in ploaddup for AVX512

Browse Source
This commit is contained in:
Gael Guennebaud 2018-09-28 16:57:32 +02:00
parent e95696acb3
commit 626942d9dd
2 changed files with 5 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@ -466,7 +466,9 @@ EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) {
// {a0, a0 a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7} // {a0, a0 a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7}
template <> template <>
EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) { EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) {
__m256i low_half = _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); // an unaligned load is required here as there is no requirement
// on the alignment of input pointer 'from'
__m256i low_half = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
__m512 even_elements = _mm512_castsi512_ps(_mm512_cvtepu32_epi64(low_half)); __m512 even_elements = _mm512_castsi512_ps(_mm512_cvtepu32_epi64(low_half));
__m512 pairs = _mm512_permute_ps(even_elements, _MM_SHUFFLE(2, 2, 0, 0)); __m512 pairs = _mm512_permute_ps(even_elements, _MM_SHUFFLE(2, 2, 0, 0));
return pairs; return pairs;

2
test/packetmath.cpp
View File

@ -227,6 +227,7 @@ template<typename Scalar> void packetmath()
if(PacketSize>1) if(PacketSize>1)
{ {
// apply different offsets to check that ploaddup is robust to unaligned inputs
for(int offset=0;offset<4;++offset) for(int offset=0;offset<4;++offset)
{ {
for(int i=0;i<PacketSize/2;++i) for(int i=0;i<PacketSize/2;++i)
@ -238,6 +239,7 @@ template<typename Scalar> void packetmath()
if(PacketSize>2) if(PacketSize>2)
{ {
// apply different offsets to check that ploadquad is robust to unaligned inputs
for(int offset=0;offset<4;++offset) for(int offset=0;offset<4;++offset)
{ {
for(int i=0;i<PacketSize/4;++i) for(int i=0;i<PacketSize/4;++i)