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

Merged

Browse Source 
Merge branch 'master' of https://github.com/coppolachan/Grid into coppolachan-master

Conflicts:
	lib/simd/Grid_vector_types.h
This commit is contained in:
Peter Boyle
2015-05-19 15:05:07 +01:00
parent ee8cf77071 7fb3221d47
commit 3fe7275332
16 changed files with 526 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
lib/simd/.dirstamp Normal file
View File

194
lib/simd/Grid_sse4.h Normal file
View File

@@ -0,0 +1,194 @@
//----------------------------------------------------------------------
/*! @file Grid_sse4.h
@brief Optimization libraries
*/
// Time-stamp: <2015-05-19 17:06:51 neo>
//----------------------------------------------------------------------
#include <pmmintrin.h>
namespace Optimization {
struct Vsplat{
//Complex float
inline __m128 operator()(float a, float b){
return _mm_set_ps(b,a,b,a);
}
// Real float
inline __m128 operator()(float a){
return _mm_set_ps(a,a,a,a);
}
//Complex double
inline __m128d operator()(double a, double b){
return _mm_set_pd(b,a);
}
//Real double
inline __m128d operator()(double a){
return _mm_set_pd(a,a);
}
//Integer
inline __m128i operator()(Integer a){
return _mm_set1_epi32(a);
}
};
struct Vstore{
//Float
inline void operator()(__m128 a, float* F){
_mm_store_ps(F,a);
}
//Double
inline void operator()(__m128d a, double* D){
_mm_store_pd(D,a);
}
//Integer
inline void operator()(__m128i a, Integer* I){
_mm_store_si128((__m128i *)I,a);
}
};
struct Vset{
// Complex float
inline __m128 operator()(Grid::ComplexF *a){
return _mm_set_ps(a[1].imag(), a[1].real(),a[0].imag(),a[0].real());
}
// Complex double
inline __m128d operator()(Grid::ComplexD *a){
return _mm_set_pd(a[0].imag(),a[0].real());
}
// Real float
inline __m128 operator()(float *a){
return _mm_set_ps(a[3],a[2],a[1],a[0]);
}
// Real double
inline __m128d operator()(double *a){
return _mm_set_pd(a[1],a[0]);
}
// Integer
inline __m128i operator()(Integer *a){
return _mm_set_epi32(a[0],a[1],a[2],a[3]);
}
};
struct Reduce{
//Complex float
inline Grid::ComplexF 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]);
}
//Complex double
inline Grid::ComplexD operator()(__m128d in){
printf("Missing complex double implementation -> FIX\n");
return Grid::ComplexD(0,0); // FIXME wrong
}
};
/////////////////////////////////////////////////////
// Arithmetic operations
/////////////////////////////////////////////////////
struct Sum{
//Complex/Real float
inline __m128 operator()(__m128 a, __m128 b){
return _mm_add_ps(a,b);
}
//Complex/Real double
inline __m128d operator()(__m128d a, __m128d b){
return _mm_add_pd(a,b);
}
//Integer
inline __m128i operator()(__m128i a, __m128i b){
return _mm_add_epi32(a,b);
}
};
struct Sub{
//Complex/Real float
inline __m128 operator()(__m128 a, __m128 b){
return _mm_sub_ps(a,b);
}
//Complex/Real double
inline __m128d operator()(__m128d a, __m128d b){
return _mm_sub_pd(a,b);
}
//Integer
inline __m128i operator()(__m128i a, __m128i b){
return _mm_sub_epi32(a,b);
}
};
struct MultComplex{
// Complex float
inline __m128 operator()(__m128 a, __m128 b){
__m128 ymm0,ymm1,ymm2;
ymm0 = _mm_shuffle_ps(a,a,_MM_SHUFFLE(2,2,0,0)); // ymm0 <- ar ar,
ymm0 = _mm_mul_ps(ymm0,b); // ymm0 <- ar bi, ar br
ymm1 = _mm_shuffle_ps(b,b,_MM_SHUFFLE(2,3,0,1)); // ymm1 <- br,bi
ymm2 = _mm_shuffle_ps(a,a,_MM_SHUFFLE(3,3,1,1)); // ymm2 <- ai,ai
ymm1 = _mm_mul_ps(ymm1,ymm2); // ymm1 <- br ai, ai bi
return _mm_addsub_ps(ymm0,ymm1);
}
// Complex double
inline __m128d operator()(__m128d a, __m128d b){
__m128d ymm0,ymm1,ymm2;
ymm0 = _mm_shuffle_pd(a,a,0x0); // ymm0 <- ar ar,
ymm0 = _mm_mul_pd(ymm0,b); // ymm0 <- ar bi, ar br
ymm1 = _mm_shuffle_pd(b,b,0x1); // ymm1 <- br,bi b01
ymm2 = _mm_shuffle_pd(a,a,0x3); // ymm2 <- ai,ai b11
ymm1 = _mm_mul_pd(ymm1,ymm2); // ymm1 <- br ai, ai bi
return _mm_addsub_pd(ymm0,ymm1);
}
};
struct Mult{
// Real float
inline __m128 operator()(__m128 a, __m128 b){
return _mm_mul_ps(a,b);
}
// Real double
inline __m128d operator()(__m128d a, __m128d b){
return _mm_mul_pd(a,b);
}
// Integer
inline __m128i operator()(__m128i a, __m128i b){
return _mm_mul_epi32(a,b);
}
};
}
// Here assign types
namespace Grid {
typedef __m128 SIMD_Ftype; // Single precision type
typedef __m128d SIMD_Dtype; // Double precision type
typedef __m128i SIMD_Itype; // Integer type
// Function names
typedef Optimization::Vsplat VsplatSIMD;
typedef Optimization::Vstore VstoreSIMD;
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Vset VsetSIMD;
}

130
lib/simd/Grid_vector_types.h
View File

@@ -1,6 +1,14 @@
//---------------------------------------------------------------------------
/*! @file Grid_vector_types.h
@brief Defines templated class Grid_simd to deal with inner vector types
*/
// Time-stamp: <2015-05-19 17:20:36 neo>
//---------------------------------------------------------------------------
#ifndef GRID_VECTOR_TYPES
#define GRID_VECTOR_TYPES
#include "Grid_sse4.h"
namespace Grid {
@@ -13,31 +21,32 @@ namespace Grid {
struct RealPart< std::complex<T> >{
typedef T type;
};
////////////////////////////////////////////////////////
////////////////////////////////////////////////////////
// Check for complexity with type traits
template <typename T>
struct is_complex : std::false_type {};
template < typename T >
struct is_complex< std::complex<T> >: std::true_type {};
////////////////////////////////////////////////////////
// Define the operation templates functors
template < class SIMD, class Operation >
SIMD binary(SIMD src_1, SIMD src_2, Operation op){
// general forms to allow for vsplat syntax
// need explicit declaration of types when used since
// clang cannot automatically determine the output type sometimes
template < class Out, class Input1, class Input2, class Operation >
Out binary(Input1 src_1, Input2 src_2, Operation op){
return op(src_1, src_2);
}
}
template < class SIMD, class Operation >
SIMD unary(SIMD src, Operation op){
template < class SIMDout, class Input, class Operation >
SIMDout unary(Input src, Operation op){
return op(src);
}
}
///////////////////////////////////////////////
/*
@brief Grid_simd class for the SIMD vector type operations
*/
template < class Scalar_type, class Vector_type >
class Grid_simd {
@@ -69,76 +78,91 @@ namespace Grid {
};
///////////////////////////////////////////////
// mac, mult, sub, add, adj
// Should do an AVX2 version with mac.
///////////////////////////////////////////////
friend inline void mac (Grid_simd * __restrict__ y,const Grid_simd * __restrict__ a,const Grid_simd *__restrict__ x){ *y = (*a)*(*x)+(*y); };
friend inline void mult(Grid_simd * __restrict__ y,const Grid_simd * __restrict__ l,const Grid_simd *__restrict__ r){ *y = (*l) * (*r); }
friend inline void sub (Grid_simd * __restrict__ y,const Grid_simd * __restrict__ l,const Grid_simd *__restrict__ r){ *y = (*l) - (*r); }
friend inline void add (Grid_simd * __restrict__ y,const Grid_simd * __restrict__ l,const Grid_simd *__restrict__ r){ *y = (*l) + (*r); }
//not for integer types... FIXME
friend inline Grid_simd adj(const Grid_simd &in){ return conjugate(in); }
//////////////////////////////////
// Initialise to 1,0,i
//////////////////////////////////
///////////////////////////////////////////////
// Initialise to 1,0,i for the correct types
///////////////////////////////////////////////
// if not complex overload here
friend inline void vone(Grid_simd &ret) { vsplat(ret,1.0); }
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0.0); }
template < class S = Scalar_type,typename std::enable_if < !is_complex < S >::value, int >::type = 0 >
friend inline void vone(Grid_simd &ret) { vsplat(ret,1.0); }
template < class S = Scalar_type,typename std::enable_if < !is_complex < S >::value, int >::type = 0 >
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0.0); }
// overload for complex type
template < class S = Scalar_type,typename std::enable_if < is_complex < S >::value, int >::type = 0 >
friend inline void vone(Grid_simd &ret) { vsplat(ret,1.0,0.0); }
friend inline void vone(Grid_simd &ret) { vsplat(ret,1.0,0.0); }
template < class S = Scalar_type,typename std::enable_if < is_complex < S >::value, int >::type = 0 >
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0.0,0.0); }
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0.0,0.0); }// use xor?
// For integral type
template < class S = Scalar_type,typename std::enable_if < std::is_integral < S >::value, int >::type = 0 >
friend inline void vone(Grid_simd &ret) { vsplat(ret,1); }
friend inline void vone(Grid_simd &ret) { vsplat(ret,1); }
template < class S = Scalar_type,typename std::enable_if < std::is_integral < S >::value, int >::type = 0 >
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0); }
friend inline void vzero(Grid_simd &ret) { vsplat(ret,0); }
template < class S = Scalar_type,typename std::enable_if < std::is_integral < S >::value, int >::type = 0 >
friend inline void vtrue (Grid_simd &ret){vsplat(ret,0xFFFFFFFF);}
template < class S = Scalar_type,typename std::enable_if < std::is_integral < S >::value, int >::type = 0 >
friend inline void vfalse(vInteger &ret){vsplat(ret,0);}
// do not compile if real or integer, send an error message from the compiler
template < class S = Scalar_type,typename std::enable_if < is_complex < S >::value, int >::type = 0 >
friend inline void vcomplex_i(Grid_simd &ret){ vsplat(ret,0.0,1.0);}
////////////////////////////////////
// Arithmetic operator overloads +,-,*
////////////////////////////////////
friend inline Grid_simd operator + (Grid_simd a, Grid_simd b)
{
vComplexF ret;
// FIXME call the binary op
Grid_simd ret;
ret.v = binary<Vector_type>(a.v, b.v, SumSIMD());
return ret;
};
friend inline Grid_simd operator - (Grid_simd a, Grid_simd b)
{
vComplexF ret;
// FIXME call the binary op
Grid_simd ret;
ret.v = binary<Vector_type>(a.v, b.v, SubSIMD());
return ret;
};
friend inline Grid_simd operator * (Grid_simd a, Grid_simd b)
// Distinguish between complex types and others
template < class S = Scalar_type, typename std::enable_if < is_complex < S >::value, int >::type = 0 >
friend inline Grid_simd operator * (Grid_simd a, Grid_simd b)
{
vComplexF ret;
// FIXME call the binary op
Grid_simd ret;
ret.v = binary<Vector_type>(a.v,b.v, MultComplexSIMD());
return ret;
};
// Real/Integer types
template < class S = Scalar_type,typename std::enable_if < !is_complex < S >::value, int >::type = 0 >
friend inline Grid_simd operator * (Grid_simd a, Grid_simd b)
{
Grid_simd ret;
ret.v = binary<Vector_type>(a.v,b.v, MultSIMD());
return ret;
};
////////////////////////////////////////////////////////////////////////
// FIXME: gonna remove these load/store, get, set, prefetch
////////////////////////////////////////////////////////////////////////
friend inline void vset(Grid_simd &ret, Scalar_type *a){
// FIXME set
ret.v = unary<Vector_type>(a, VsetSIMD());
}
///////////////////////
@@ -147,34 +171,33 @@ namespace Grid {
// overload if complex
template < class S = Scalar_type >
friend inline void vsplat(Grid_simd &ret, typename std::enable_if< is_complex < S >::value, S>::type c){
Real a= real(c);
Real b= imag(c);
Real a = real(c);
Real b = imag(c);
vsplat(ret,a,b);
}
// this only for the complex version
template < class S = Scalar_type, typename std::enable_if < is_complex < S >::value, int >::type = 0 >
friend inline void vsplat(Grid_simd &ret,Real a, Real b){
// FIXME add operator
ret.v = binary<Vector_type>(a, b, VsplatSIMD());
}
//if real fill with a, if complex fill with a in the real part
//if real fill with a, if complex fill with a in the real part (first function above)
friend inline void vsplat(Grid_simd &ret,Real a){
// FIXME add operator
ret.v = unary<Vector_type>(a, VsplatSIMD());
}
friend inline void vstore(const Grid_simd &ret, Scalar_type *a){
//FIXME
binary<void>(ret.v, (Real*)a, VstoreSIMD());
}
friend inline void vprefetch(const Grid_simd &v)
{
_mm_prefetch((const char*)&v.v,_MM_HINT_T0);
}
friend inline Scalar_type Reduce(const Grid_simd & in)
{
// FIXME add operator
@@ -221,6 +244,7 @@ namespace Grid {
inline Grid_simd &operator *=(const Grid_simd &r) {
*this = (*this)*r;
return *this;
// return (*this)*r; ?
}
inline Grid_simd &operator +=(const Grid_simd &r) {
*this = *this+r;
@@ -233,6 +257,12 @@ namespace Grid {
friend inline void permute(Grid_simd &y,Grid_simd b,int perm)
{
Gpermute<Grid_simd>(y,b,perm);
}
/*
friend inline void permute(Grid_simd &y,Grid_simd b,int perm)
{
Gpermute<Grid_simd>(y,b,perm);
@@ -253,7 +283,7 @@ namespace Grid {
{
Gextract<Grid_simd,Scalar_type>(y,extracted);
}
*/
};// end of Grid_simd class definition
@@ -286,11 +316,11 @@ namespace Grid {
// Define available types (now change names to avoid clashing)
typedef __m128 SIMD_type;// decided at compilation time
typedef Grid_simd< float , SIMD_type > MyRealF;
typedef Grid_simd< double , SIMD_type > MyRealD;
typedef Grid_simd< std::complex< float > , SIMD_type > MyComplexF;
typedef Grid_simd< std::complex< double >, SIMD_type > MyComplexD;
typedef Grid_simd< float , SIMD_Ftype > MyRealF;
typedef Grid_simd< double , SIMD_Dtype > MyRealD;
typedef Grid_simd< std::complex< float > , SIMD_Ftype > MyComplexF;
typedef Grid_simd< std::complex< double >, SIMD_Dtype > MyComplexD;