portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Corrected AVX regression error. Tested.

Browse Source
This commit is contained in:
Guido Cossu
2015-05-27 10:49:33 +09:00
parent 13707e0808
commit 26ec41288d
14 changed files with 3634 additions and 4381 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
lib/simd/Grid_avx.h
View File

@ -299,7 +299,7 @@ namespace Optimization {
//////////////////////////////////////////////
// Some Template specialization
template < typename vtype >
void permute(vtype a, vtype b, int perm) {
void permute(vtype &a, vtype &b, int perm) {
union {
__m256 f;
vtype v;
@ -320,11 +320,16 @@ namespace Optimization {
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m256>::operator()(__m256 in){
__m256 v1,v2;
union {
__m256 v;
float f[8];
} conv;
Optimization::permute(v1,in,0); // sse 128; paired complex single
v1 = _mm256_add_ps(v1,in);
Optimization::permute(v2,v1,1); // avx 256; quad complex single
v1 = _mm256_add_ps(v1,v2);
return Grid::ComplexF(v1[0],v1[1]);
conv.v = v1;
return Grid::ComplexF(conv.f[0],conv.f[1]);
}
//Real float Reduce
template<>

4
lib/simd/Grid_vector_types.h
View File

@ -77,9 +77,7 @@ inline void Gpermute(vsimd &y,const vsimd &b,int perm){
switch (perm){
#if defined(AVX1)||defined(AVX2)
// 8x32 bits=>3 permutes
case 2:
conv.f = _mm256_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(2,3,0,1));
break;
case 2: conv.f = _mm256_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(2,3,0,1)); break;
case 1: conv.f = _mm256_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(1,0,3,2)); break;
case 0: conv.f = _mm256_permute2f128_ps(conv.f,conv.f,0x01); break;
#endif