portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-16 14:57:05 +01:00
Code Releases Activity

Merge branch 'develop' into feature/hmc_generalise

Browse Source
This commit is contained in:
Guido Cossu
2017-01-25 11:33:53 +00:00
parent 0baa20d292 d65e81518f
commit 17629b8d9e
88 changed files with 7904 additions and 430 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
lib/simd/Grid_avx.h
View File

@ -204,6 +204,29 @@ namespace Optimization {
}
};
struct MultRealPart{
inline __m256 operator()(__m256 a, __m256 b){
__m256 ymm0;
ymm0 = _mm256_shuffle_ps(a,a,_MM_SELECT_FOUR_FOUR(2,2,0,0)); // ymm0 <- ar ar,
return _mm256_mul_ps(ymm0,b); // ymm0 <- ar bi, ar br
}
inline __m256d operator()(__m256d a, __m256d b){
__m256d ymm0;
ymm0 = _mm256_shuffle_pd(a,a,0x0); // ymm0 <- ar ar, ar,ar b'00,00
return _mm256_mul_pd(ymm0,b); // ymm0 <- ar bi, ar br
}
};
struct MaddRealPart{
inline __m256 operator()(__m256 a, __m256 b, __m256 c){
__m256 ymm0 = _mm256_moveldup_ps(a); // ymm0 <- ar ar,
return _mm256_add_ps(_mm256_mul_ps( ymm0, b),c);
}
inline __m256d operator()(__m256d a, __m256d b, __m256d c){
__m256d ymm0 = _mm256_shuffle_pd( a, a, 0x0 );
return _mm256_add_pd(_mm256_mul_pd( ymm0, b),c);
}
};
struct MultComplex{
// Complex float
inline __m256 operator()(__m256 a, __m256 b){
@ -618,7 +641,9 @@ namespace Optimization {
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::MultRealPart MultRealPartSIMD;
typedef Optimization::MaddRealPart MaddRealPartSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;