portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-08-02 20:57:06 +01:00
Code Releases Activity

Updating to modify non-inlining permute routines and hopefully get better reg use and

Browse Source 
enhance performance.
This commit is contained in:
Peter Boyle
2015-09-25 08:55:04 -07:00
parent 5ef42add2d
commit 64d64d1ab6
11 changed files with 212 additions and 141 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
lib/simd/Grid_imci.h
View File

@@ -255,7 +255,36 @@ namespace Optimization {
};
struct Permute{
static inline __m512 Permute0(__m512 in){
return _mm512_permute4f128_ps(in,(_MM_PERM_ENUM)_MM_SELECT_FOUR_FOUR(1,0,3,2));
};
static inline __m512 Permute1(__m512 in){
return _mm512_permute4f128_ps(in,(_MM_PERM_ENUM)_MM_SELECT_FOUR_FOUR(2,3,0,1));
};
static inline __m512 Permute2(__m512 in){
return _mm512_swizzle_ps(in,_MM_SWIZ_REG_BADC);
};
static inline __m512 Permute3(__m512 in){
return _mm512_swizzle_ps(in,_MM_SWIZ_REG_CDAB);
};
static inline __m512d Permute0(__m512d in){// Hack no intrinsic for 256 swaps of __m512d
return (__m512d)_mm512_permute4f128_ps((__m512)in,(_MM_PERM_ENUM)_MM_SELECT_FOUR_FOUR(1,0,3,2));
};
static inline __m512d Permute1(__m512d in){
return _mm512_swizzle_pd(in,_MM_SWIZ_REG_BADC);
};
static inline __m512d Permute2(__m512d in){
return _mm512_swizzle_pd(in,_MM_SWIZ_REG_CDAB);
};
static inline __m512d Permute3(__m512d in){
return in;
};
};
//////////////////////////////////////////////
@@ -315,25 +344,6 @@ namespace Grid {
}
// Gpermute utilities consider coalescing into 1 Gpermute
template < typename VectorSIMD >
inline void Gpermute(VectorSIMD &y,const VectorSIMD &b, int perm ) {
union {
__m512 f;
decltype(VectorSIMD::v) v;
} conv;
conv.v = b.v;
switch(perm){
case 3: conv.f = _mm512_swizzle_ps(conv.f,_MM_SWIZ_REG_CDAB); break;
case 2: conv.f = _mm512_swizzle_ps(conv.f,_MM_SWIZ_REG_BADC); break;
case 1 : conv.f = _mm512_permute4f128_ps(conv.f,(_MM_PERM_ENUM)_MM_SHUFFLE(2,3,0,1)); break;
case 0 : conv.f = _mm512_permute4f128_ps(conv.f,(_MM_PERM_ENUM)_MM_SHUFFLE(1,0,3,2)); break;
default: assert(0); break;
}
y.v=conv.v;
};
// Function name aliases
typedef Optimization::Vsplat VsplatSIMD;