portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-16 23:07:05 +01:00
Code Releases Activity

Merge branch 'develop' into feature/CG_repro

Browse Source
This commit is contained in:
Guido Cossu
2016-12-05 05:07:01 +00:00
parent afbbcd2194 e27c6b217c
commit b2dc17e160
16 changed files with 481 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
lib/simd/Grid_avx.h
View File

@ -170,7 +170,7 @@ namespace Optimization {
}
//Integer
inline __m256i operator()(__m256i a, __m256i b){
#if defined (AVX1) || defined (AVXFMA4)
#if defined (AVX1) || defined (AVXFMA) || defined (AVXFMA4)
__m128i a0,a1;
__m128i b0,b1;
a0 = _mm256_extractf128_si256(a,0);
@ -198,7 +198,7 @@ namespace Optimization {
}
//Integer
inline __m256i operator()(__m256i a, __m256i b){
#if defined (AVX1) || defined (AVXFMA4)
#if defined (AVX1) || defined (AVXFMA) || defined (AVXFMA4)
__m128i a0,a1;
__m128i b0,b1;
a0 = _mm256_extractf128_si256(a,0);
@ -219,7 +219,7 @@ namespace Optimization {
struct MultComplex{
// Complex float
inline __m256 operator()(__m256 a, __m256 b){
#if defined (AVX1)
#if defined (AVX1)
__m256 ymm0,ymm1,ymm2;
ymm0 = _mm256_shuffle_ps(a,a,_MM_SELECT_FOUR_FOUR(2,2,0,0)); // ymm0 <- ar ar,
ymm0 = _mm256_mul_ps(ymm0,b); // ymm0 <- ar bi, ar br
@ -236,7 +236,7 @@ namespace Optimization {
a_imag = _mm256_mul_ps( a_imag,tmp ); // (Ai, Ai) * (Bi, Br) = Ai Bi, Ai Br
return _mm256_maddsub_ps( a_real, b, a_imag ); // Ar Br , Ar Bi +- Ai Bi = ArBr-AiBi , ArBi+AiBr
#endif
#if defined (AVX2)
#if defined (AVX2) || defined (AVXFMA)
__m256 a_real = _mm256_moveldup_ps( a ); // Ar Ar
__m256 a_imag = _mm256_movehdup_ps( a ); // Ai Ai
a_imag = _mm256_mul_ps( a_imag, _mm256_shuffle_ps( b,b, _MM_SELECT_FOUR_FOUR(2,3,0,1) )); // (Ai, Ai) * (Bi, Br) = Ai Bi, Ai Br
@ -267,7 +267,7 @@ namespace Optimization {
IF IMM0[3] = 0
THEN DEST[255:192]=SRC2[191:128] ELSE DEST[255:192]=SRC2[255:192] FI; // Ox5 r<->i ; 0xC unchanged
*/
#if defined (AVX1)
#if defined (AVX1)
__m256d ymm0,ymm1,ymm2;
ymm0 = _mm256_shuffle_pd(a,a,0x0); // ymm0 <- ar ar, ar,ar b'00,00
ymm0 = _mm256_mul_pd(ymm0,b); // ymm0 <- ar bi, ar br
@ -282,7 +282,7 @@ namespace Optimization {
a_imag = _mm256_mul_pd( a_imag, _mm256_permute_pd( b, 0x5 ) ); // (Ai, Ai) * (Bi, Br) = Ai Bi, Ai Br
return _mm256_maddsub_pd( a_real, b, a_imag ); // Ar Br , Ar Bi +- Ai Bi = ArBr-AiBi , ArBi+AiBr
#endif
#if defined (AVX2)
#if defined (AVX2) || defined (AVXFMA)
__m256d a_real = _mm256_movedup_pd( a ); // Ar Ar
__m256d a_imag = _mm256_shuffle_pd(a,a,0xF);//aiai
a_imag = _mm256_mul_pd( a_imag, _mm256_permute_pd( b, 0x5 ) ); // (Ai, Ai) * (Bi, Br) = Ai Bi, Ai Br
@ -323,7 +323,7 @@ namespace Optimization {
#if defined (AVXFMA4)
a= _mm256_macc_ps(b,c,a);
#endif
#if defined (AVX2)
#if defined (AVX2) || defined (AVXFMA)
a= _mm256_fmadd_ps( b, c, a);
#endif
}
@ -335,7 +335,7 @@ namespace Optimization {
#if defined (AVXFMA4)
a= _mm256_macc_pd(b,c,a);
#endif
#if defined (AVX2)
#if defined (AVX2) || defined (AVXFMA)
a= _mm256_fmadd_pd( b, c, a);
#endif
}
@ -350,7 +350,7 @@ namespace Optimization {
}
// Integer
inline __m256i operator()(__m256i a, __m256i b){
#if defined (AVX1)
#if defined (AVX1) || defined (AVXFMA)
__m128i a0,a1;
__m128i b0,b1;
a0 = _mm256_extractf128_si256(a,0);

8
lib/simd/Grid_avx512.h
View File

@ -86,13 +86,13 @@ namespace Optimization {
struct Vstream{
//Float
inline void operator()(float * a, __m512 b){
//_mm512_stream_ps(a,b);
_mm512_store_ps(a,b);
_mm512_stream_ps(a,b);
// _mm512_store_ps(a,b);
}
//Double
inline void operator()(double * a, __m512d b){
//_mm512_stream_pd(a,b);
_mm512_store_pd(a,b);
_mm512_stream_pd(a,b);
// _mm512_store_pd(a,b);
}
};

18
lib/simd/Grid_qpx.h
View File

@ -244,7 +244,22 @@ namespace Optimization {
return a*b;
}
};
struct Div{
// Real double
inline vector4double operator()(vector4double a, vector4double b){
return vec_swdiv(a, b);
}
// Real float
FLOAT_WRAP_2(operator(), inline)
// Integer
inline int operator()(int a, int b){
return a/b;
}
};
struct Conj{
// Complex double
inline vector4double operator()(vector4double v){
@ -413,6 +428,7 @@ template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;

4
lib/simd/Grid_vector_types.h
View File

@ -44,7 +44,7 @@ directory
#ifdef SSE4
#include "Grid_sse4.h"
#endif
#if defined(AVX1) || defined(AVX2) || defined(AVXFMA4)
#if defined(AVX1) || defined (AVXFMA) || defined(AVX2) || defined(AVXFMA4)
#include "Grid_avx.h"
#endif
#if defined AVX512
@ -130,7 +130,7 @@ class Grid_simd {
Vector_type v;
static inline int Nsimd(void) {
static inline constexpr int Nsimd(void) {
return sizeof(Vector_type) / sizeof(Scalar_type);
}