portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-04 14:04:32 +00:00
Code Releases Activity

Cleaning up simd files

Browse Source
This commit is contained in:
neo
2015-05-26 13:31:10 +09:00
parent 48cc816136
commit 4dbaa389c8
3 changed files with 64 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
TODO
View File

@@ -1,8 +1,8 @@
================================================================ ================================================================
*** Hacks and bug fixes to clean up and Audits *** Hacks and bug fixes to clean up and Audits
================================================================ ================================================================
* Base class to share common code between vRealF, VComplexF etc... * Base class to share common code between vRealF, VComplexF etc... done
- Performance check on Guido's reimplementation strategy - Performance check on Guido's reimplementation strategy - (GUIDO) tested and no difference was found, merged
* FIXME audit * FIXME audit

90
lib/Grid_simd.h
View File

@@ -95,100 +95,10 @@ namespace Grid {
template<> inline void zeroit(RealF &arg){ arg=0; }; template<> inline void zeroit(RealF &arg){ arg=0; };
template<> inline void zeroit(RealD &arg){ arg=0; }; template<> inline void zeroit(RealD &arg){ arg=0; };
// Eventually delete this part
#if defined (SSE4)
typedef __m128 fvec;
typedef __m128d dvec;
typedef __m128 cvec;
typedef __m128d zvec;
typedef __m128i ivec;
#endif
#if defined (AVX1) || defined (AVX2)
typedef __m256 fvec;
typedef __m256d dvec;
typedef __m256 cvec;
typedef __m256d zvec;
typedef __m256i ivec;
#endif
#if defined (AVX512)
typedef __m512 fvec;
typedef __m512d dvec;
typedef __m512 cvec;
typedef __m512d zvec;
typedef __m512i ivec;
#endif
#if defined (QPX)
typedef float fvec __attribute__ ((vector_size (16))); // QPX has same SIMD width irrespective of precision
typedef float cvec __attribute__ ((vector_size (16)));
typedef vector4double dvec;
typedef vector4double zvec;
#endif
#if defined (AVX1) || defined (AVX2) || defined (AVX512)
inline void v_prefetch0(int size, const char *ptr){
for(int i=0;i<size;i+=64){ // Define L1 linesize above// What about SSE?
_mm_prefetch(ptr+i+4096,_MM_HINT_T1);
_mm_prefetch(ptr+i+512,_MM_HINT_T0);
}
}
#else
inline void v_prefetch0(int size, const char *ptr){};
#endif
//////////////////////////////////////////////////////////
// Permute
// Permute 0 every ABCDEFGH -> BA DC FE HG
// Permute 1 every ABCDEFGH -> CD AB GH EF
// Permute 2 every ABCDEFGH -> EFGH ABCD
// Permute 3 possible on longer iVector lengths (512bit = 8 double = 16 single)
// Permute 4 possible on half precision @512bit vectors.
//////////////////////////////////////////////////////////
template<class vsimd>
inline void Gpermute(vsimd &y,const vsimd &b,int perm){
union {
fvec f;
decltype(vsimd::v) v;
} conv;
conv.v = b.v;
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 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
#ifdef SSE4
case 1: conv.f = _mm_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(2,3,0,1)); break;
case 0: conv.f = _mm_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(1,0,3,2));break;
#endif
#ifdef AVX512
// 16 floats=> permutes
// Permute 0 every abcd efgh ijkl mnop -> badc fehg jilk nmpo
// Permute 1 every abcd efgh ijkl mnop -> cdab ghef jkij opmn
// Permute 2 every abcd efgh ijkl mnop -> efgh abcd mnop ijkl
// Permute 3 every abcd efgh ijkl mnop -> ijkl mnop abcd efgh
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;
#endif
#ifdef QPX
#error not implemented
#endif
default: assert(0); break;
}
y.v=conv.v;
};
}; };
#include <simd/Grid_vector_types.h> #include <simd/Grid_vector_types.h>
namespace Grid { namespace Grid {
// NB: Template the following on "type Complex" and then implement *,+,- for // NB: Template the following on "type Complex" and then implement *,+,- for

66
lib/simd/Grid_vector_types.h
View File

@@ -2,7 +2,7 @@
/*! @file Grid_vector_types.h /*! @file Grid_vector_types.h
@brief Defines templated class Grid_simd to deal with inner vector types @brief Defines templated class Grid_simd to deal with inner vector types
*/ */
// Time-stamp: <2015-05-26 12:05:39 neo> // Time-stamp: <2015-05-26 13:22:36 neo>
//--------------------------------------------------------------------------- //---------------------------------------------------------------------------
#ifndef GRID_VECTOR_TYPES #ifndef GRID_VECTOR_TYPES
#define GRID_VECTOR_TYPES #define GRID_VECTOR_TYPES
@@ -16,7 +16,9 @@
#if defined AVX512 #if defined AVX512
#include "Grid_knc.h" #include "Grid_knc.h"
#endif #endif
#if defined QPX
#include "Grid_qpx.h"
#endif
namespace Grid { namespace Grid {
@@ -33,8 +35,6 @@ namespace Grid {
template <typename Condition, typename ReturnType> using EnableIf = Invoke<std::enable_if<Condition::value, ReturnType>>; template <typename Condition, typename ReturnType> using EnableIf = Invoke<std::enable_if<Condition::value, ReturnType>>;
template <typename Condition, typename ReturnType> using NotEnableIf= Invoke<std::enable_if<!Condition::value, ReturnType>>; template <typename Condition, typename ReturnType> using NotEnableIf= Invoke<std::enable_if<!Condition::value, ReturnType>>;
//////////////////////////////////////////////////////// ////////////////////////////////////////////////////////
// Check for complexity with type traits // Check for complexity with type traits
template <typename T> struct is_complex : std::false_type {}; template <typename T> struct is_complex : std::false_type {};
@@ -57,6 +57,58 @@ namespace Grid {
/////////////////////////////////////////////// ///////////////////////////////////////////////
//////////////////////////////////////////////////////////
// Permute
// Permute 0 every ABCDEFGH -> BA DC FE HG
// Permute 1 every ABCDEFGH -> CD AB GH EF
// Permute 2 every ABCDEFGH -> EFGH ABCD
// Permute 3 possible on longer iVector lengths (512bit = 8 double = 16 single)
// Permute 4 possible on half precision @512bit vectors.
//////////////////////////////////////////////////////////
template<class vsimd>
inline void Gpermute(vsimd &y,const vsimd &b,int perm){
union {
SIMD_Ftype f;
decltype(vsimd::v) v;
} conv;
conv.v = b.v;
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 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
#ifdef SSE4
case 1: conv.f = _mm_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(2,3,0,1)); break;
case 0: conv.f = _mm_shuffle_ps(conv.f,conv.f,_MM_SHUFFLE(1,0,3,2));break;
#endif
#ifdef AVX512
// 16 floats=> permutes
// Permute 0 every abcd efgh ijkl mnop -> badc fehg jilk nmpo
// Permute 1 every abcd efgh ijkl mnop -> cdab ghef jkij opmn
// Permute 2 every abcd efgh ijkl mnop -> efgh abcd mnop ijkl
// Permute 3 every abcd efgh ijkl mnop -> ijkl mnop abcd efgh
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;
#endif
#ifdef QPX
#error not implemented
#endif
default: assert(0); break;
}
y.v=conv.v;
};
///////////////////////////////////////
/* /*
@brief Grid_simd class for the SIMD vector type operations @brief Grid_simd class for the SIMD vector type operations
*/ */
@@ -380,6 +432,12 @@ namespace Grid {
typedef Grid_simd< std::complex< double >, SIMD_Dtype > vComplexD; typedef Grid_simd< std::complex< double >, SIMD_Dtype > vComplexD;
typedef Grid_simd< Integer , SIMD_Itype > vInteger; typedef Grid_simd< Integer , SIMD_Itype > vInteger;
} }
#endif #endif