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

Improving the reduction to go through our on permute.

Browse Source 
Must also do this for avx512
This commit is contained in:
Peter Boyle
2015-05-27 16:07:17 +01:00
parent 64753ea633
commit b72ca15bd2
4 changed files with 89 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
lib/simd/Grid_sse4.h
View File

@ -10,6 +10,21 @@
#include <pmmintrin.h>
namespace Optimization {
template<class vtype>
union uconv {
__m128 f;
vtype v;
};
union u128f {
__m128 v;
float f[4];
};
union u128d {
__m128d v;
double f[2];
};
struct Vsplat{
//Complex float
@ -50,7 +65,6 @@ namespace Optimization {
};
struct Vstream{
//Float
inline void operator()(float * a, __m128 b){
@ -64,8 +78,6 @@ namespace Optimization {
};
struct Vset{
// Complex float
inline __m128 operator()(Grid::ComplexF *a){
@ -102,9 +114,6 @@ namespace Optimization {
}
};
/////////////////////////////////////////////////////
// Arithmetic operations
/////////////////////////////////////////////////////
@ -138,7 +147,6 @@ namespace Optimization {
}
};
struct MultComplex{
// Complex float
inline __m128 operator()(__m128 a, __m128 b){
@ -177,7 +185,6 @@ namespace Optimization {
}
};
struct Conj{
// Complex single
inline __m128 operator()(__m128 in){
@ -216,57 +223,61 @@ namespace Optimization {
__m128d tmp = _mm_shuffle_pd(in,in,0x1);
return _mm_addsub_pd(_mm_setzero_pd(),tmp); // r,-i
}
};
//////////////////////////////////////////////
// Some Template specialization
template < typename vtype >
void permute(vtype &a, vtype b, int perm) {
uconv<vtype> conv;
conv.v = b;
switch(perm){
case 3: break; //empty for SSE4
case 2: break; //empty for 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;
default: assert(0); break;
}
a=conv.v;
};
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m128>::operator()(__m128 in){
union {
__m128 v1;
float f[4];
} u128;
u128.v1 = _mm_add_ps(in, _mm_shuffle_ps(in,in, 0b01001110)); // FIXME Prefer to use _MM_SHUFFLE macros
return Grid::ComplexF(u128.f[0], u128.f[1]);
__m128 v1; // two complex
Optimization::permute(v1,in,0);
v1= _mm_add_ps(v1,in);
u128f conv; conv.v=v1;
return Grid::ComplexF(conv.f[0],conv.f[1]);
}
//Real float Reduce
template<>
inline Grid::RealF Reduce<Grid::RealF, __m128>::operator()(__m128 in){
// FIXME Hack
const Grid::RealF * ptr = (const Grid::RealF *) &in;
Grid::RealF ret = 0;
for(int i=0;i< 4 ;i++){ // 4 number of simd lanes for float
ret = ret+ptr[i];
}
return ret;
__m128 v1,v2; // quad single
Optimization::permute(v1,in,0);
v1= _mm_add_ps(v1,in);
Optimization::permute(v2,v1,1);
v1 = _mm_add_ps(v1,v2);
u128f conv; conv.v=v1;
return conv.f[0];
}
//Complex double Reduce
template<>
inline Grid::ComplexD Reduce<Grid::ComplexD, __m128d>::operator()(__m128d in){
printf("Reduce : Missing good complex double implementation -> FIX\n");
return Grid::ComplexD(in[0], in[1]); // inefficient
u128d conv; conv.v = in;
return Grid::ComplexD(conv.f[0],conv.f[1]);
}
//Real double Reduce
template<>
inline Grid::RealD Reduce<Grid::RealD, __m128d>::operator()(__m128d in){
// FIXME Hack
const Grid::RealD * ptr =(const Grid::RealD *) &in;
Grid::RealD ret = 0;
for(int i=0;i< 2 ;i++){// 2 number of simd lanes for float
ret = ret+ptr[i];
}
return ret;
__m128d v1;
Optimization::permute(v1,in,0); // avx 256; quad double
v1 = _mm_add_pd(v1,in);
u128d conv; conv.v = v1;
return conv.f[0];
}
//Integer Reduce
@ -276,12 +287,6 @@ namespace Optimization {
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
}
}
//////////////////////////////////////////////////////////////////////////////////////
@ -292,27 +297,13 @@ namespace Grid {
typedef __m128d SIMD_Dtype; // Double precision type
typedef __m128i SIMD_Itype; // Integer type
inline void v_prefetch0(int size, const char *ptr){}; // prefetch utilities
// Gpermute function
template < typename VectorSIMD >
inline void Gpermute(VectorSIMD &y,const VectorSIMD &b, int perm ) {
union {
__m128 f;
decltype(VectorSIMD::v) v;
} conv;
conv.v = b.v;
switch(perm){
case 3: break; //empty for SSE4
case 2: break; //empty for 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;
default: assert(0); break;
}
y.v=conv.v;
};
Optimization::permute(y.v,b.v,perm);
}
// Function name aliases