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

AVX512 shaken out

This commit is contained in:
paboyle
2016-03-28 00:38:05 -06:00
parent 61469252fe
commit ad80f61fba
2 changed files with 141 additions and 281 deletions

View File

@ -246,26 +246,30 @@ namespace Optimization {
struct TimesMinusI{
//Complex single
inline __m512 operator()(__m512 in, __m512 ret){
__m512 tmp = _mm512_mask_sub_ps(in,0xaaaa,_mm512_setzero_ps(),in); // real -imag
return _mm512_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(1,0,3,2)); // 0x4E??
//__m512 tmp = _mm512_mask_sub_ps(in,0xaaaa,_mm512_setzero_ps(),in); // real -imag
//return _mm512_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(2,3,1,0)); // 0x4E??
__m512 tmp = _mm512_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1));
return _mm512_mask_sub_ps(tmp,0xaaaa,_mm512_setzero_ps(),tmp);
}
//Complex double
inline __m512d operator()(__m512d in, __m512d ret){
__m512d tmp = _mm512_mask_sub_pd(in,0xaa,_mm512_setzero_pd(),in); // real -imag
return _mm512_shuffle_pd(tmp,tmp,0x55);
//__m512d tmp = _mm512_mask_sub_pd(in,0xaa,_mm512_setzero_pd(),in); // real -imag
//return _mm512_shuffle_pd(tmp,tmp,0x55);
__m512d tmp = _mm512_shuffle_pd(in,in,0x55);
return _mm512_mask_sub_pd(tmp,0xaa,_mm512_setzero_pd(),tmp);
}
};
struct TimesI{
//Complex single
inline __m512 operator()(__m512 in, __m512 ret){
__m512 tmp = _mm512_shuffle_ps(tmp,tmp,_MM_SELECT_FOUR_FOUR(1,0,3,2));
return _mm512_mask_sub_ps(tmp,0xaaaa,_mm512_setzero_ps(),tmp);
__m512 tmp = _mm512_shuffle_ps(in,in,_MM_SELECT_FOUR_FOUR(2,3,0,1));
return _mm512_mask_sub_ps(tmp,0x5555,_mm512_setzero_ps(),tmp);
}
//Complex double
inline __m512d operator()(__m512d in, __m512d ret){
__m512d tmp = _mm512_shuffle_pd(tmp,tmp,0x55);
return _mm512_mask_sub_pd(tmp,0xaa,_mm512_setzero_pd(),tmp);
__m512d tmp = _mm512_shuffle_pd(in,in,0x55);
return _mm512_mask_sub_pd(tmp,0x55,_mm512_setzero_pd(),tmp);
}