portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Merge branch 'feature/gen-simd' into feature/doxygen

Browse Source 
# Conflicts:
#	Makefile.am
#	configure.ac
This commit is contained in:
Antonin Portelli
2016-11-19 13:11:13 +01:00
parent cb02b7088f 433afd36f5
commit 97cddda49e
109 changed files with 6830 additions and 5158 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
lib/simd/Grid_avx.h
View File

@ -365,6 +365,18 @@ namespace Optimization {
}
};
struct Div{
// Real float
inline __m256 operator()(__m256 a, __m256 b){
return _mm256_div_ps(a,b);
}
// Real double
inline __m256d operator()(__m256d a, __m256d b){
return _mm256_div_pd(a,b);
}
};
struct Conj{
// Complex single
inline __m256 operator()(__m256 in){
@ -437,14 +449,13 @@ namespace Optimization {
};
#if defined (AVX2) || defined (AVXFMA4)
#define _mm256_alignr_epi32(ret,a,b,n) ret=(__m256) _mm256_alignr_epi8((__m256i)a,(__m256i)b,(n*4)%16)
#define _mm256_alignr_epi64(ret,a,b,n) ret=(__m256d) _mm256_alignr_epi8((__m256i)a,(__m256i)b,(n*8)%16)
#if defined (AVX2)
#define _mm256_alignr_epi32_grid(ret,a,b,n) ret=(__m256) _mm256_alignr_epi8((__m256i)a,(__m256i)b,(n*4)%16)
#define _mm256_alignr_epi64_grid(ret,a,b,n) ret=(__m256d) _mm256_alignr_epi8((__m256i)a,(__m256i)b,(n*8)%16)
#endif
#if defined (AVX1)
#define _mm256_alignr_epi32(ret,a,b,n) { \
#if defined (AVX1) || defined (AVXFMA)
#define _mm256_alignr_epi32_grid(ret,a,b,n) { \
__m128 aa, bb; \
\
aa = _mm256_extractf128_ps(a,1); \
@ -458,7 +469,7 @@ namespace Optimization {
ret = _mm256_insertf128_ps(ret,aa,0); \
}
#define _mm256_alignr_epi64(ret,a,b,n) { \
#define _mm256_alignr_epi64_grid(ret,a,b,n) { \
__m128d aa, bb; \
\
aa = _mm256_extractf128_pd(a,1); \
@ -474,19 +485,6 @@ namespace Optimization {
#endif
inline std::ostream & operator << (std::ostream& stream, const __m256 a)
{
const float *p=(const float *)&a;
stream<< "{"<<p[0]<<","<<p[1]<<","<<p[2]<<","<<p[3]<<","<<p[4]<<","<<p[5]<<","<<p[6]<<","<<p[7]<<"}";
return stream;
};
inline std::ostream & operator<< (std::ostream& stream, const __m256d a)
{
const double *p=(const double *)&a;
stream<< "{"<<p[0]<<","<<p[1]<<","<<p[2]<<","<<p[3]<<"}";
return stream;
};
struct Rotate{
static inline __m256 rotate(__m256 in,int n){
@ -518,11 +516,10 @@ namespace Optimization {
__m256 tmp = Permute::Permute0(in);
__m256 ret;
if ( n > 3 ) {
_mm256_alignr_epi32(ret,in,tmp,n);
_mm256_alignr_epi32_grid(ret,in,tmp,n);
} else {
_mm256_alignr_epi32(ret,tmp,in,n);
_mm256_alignr_epi32_grid(ret,tmp,in,n);
}
// std::cout << " align epi32 n=" <<n<<" in "<<tmp<<in<<" -> "<< ret <<std::endl;
return ret;
};
@ -531,18 +528,15 @@ namespace Optimization {
__m256d tmp = Permute::Permute0(in);
__m256d ret;
if ( n > 1 ) {
_mm256_alignr_epi64(ret,in,tmp,n);
_mm256_alignr_epi64_grid(ret,in,tmp,n);
} else {
_mm256_alignr_epi64(ret,tmp,in,n);
_mm256_alignr_epi64_grid(ret,tmp,in,n);
}
// std::cout << " align epi64 n=" <<n<<" in "<<tmp<<in<<" -> "<< ret <<std::endl;
return ret;
};
};
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m256>::operator()(__m256 in){
@ -631,6 +625,7 @@ namespace Optimization {
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;

84
lib/simd/Grid_avx512.h
View File

@ -32,6 +32,16 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
namespace Grid{
namespace Optimization {
union u512f {
__m512 v;
float f[16];
};
union u512d {
__m512d v;
double f[8];
};
struct Vsplat{
//Complex float
@ -221,6 +231,17 @@ namespace Optimization {
}
};
struct Div{
// Real float
inline __m512 operator()(__m512 a, __m512 b){
return _mm512_div_ps(a,b);
}
// Real double
inline __m512d operator()(__m512d a, __m512d b){
return _mm512_div_pd(a,b);
}
};
struct Conj{
// Complex single
@ -350,7 +371,67 @@ namespace Optimization {
//////////////////////////////////////////////
// Some Template specialization
// Hack for CLANG until mm512_reduce_add_ps etc... are implemented in GCC and Clang releases
#ifndef __INTEL_COMPILER
#warning "Slow reduction due to incomplete reduce intrinsics"
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m512>::operator()(__m512 in){
__m512 v1,v2;
v1=Optimization::Permute::Permute0(in); // avx 512; quad complex single
v1= _mm512_add_ps(v1,in);
v2=Optimization::Permute::Permute1(v1);
v1 = _mm512_add_ps(v1,v2);
v2=Optimization::Permute::Permute2(v1);
v1 = _mm512_add_ps(v1,v2);
u512f conv; conv.v = v1;
return Grid::ComplexF(conv.f[0],conv.f[1]);
}
//Real float Reduce
template<>
inline Grid::RealF Reduce<Grid::RealF, __m512>::operator()(__m512 in){
__m512 v1,v2;
v1 = Optimization::Permute::Permute0(in); // avx 512; octo-double
v1 = _mm512_add_ps(v1,in);
v2 = Optimization::Permute::Permute1(v1);
v1 = _mm512_add_ps(v1,v2);
v2 = Optimization::Permute::Permute2(v1);
v1 = _mm512_add_ps(v1,v2);
v2 = Optimization::Permute::Permute3(v1);
v1 = _mm512_add_ps(v1,v2);
u512f conv; conv.v=v1;
return conv.f[0];
}
//Complex double Reduce
template<>
inline Grid::ComplexD Reduce<Grid::ComplexD, __m512d>::operator()(__m512d in){
__m512d v1;
v1 = Optimization::Permute::Permute0(in); // sse 128; paired complex single
v1 = _mm512_add_pd(v1,in);
v1 = Optimization::Permute::Permute1(in); // sse 128; paired complex single
v1 = _mm512_add_pd(v1,in);
u512d conv; conv.v = v1;
return Grid::ComplexD(conv.f[0],conv.f[1]);
}
//Real double Reduce
template<>
inline Grid::RealD Reduce<Grid::RealD, __m512d>::operator()(__m512d in){
__m512d v1,v2;
v1 = Optimization::Permute::Permute0(in); // avx 512; quad double
v1 = _mm512_add_pd(v1,in);
v2 = Optimization::Permute::Permute1(v1);
v1 = _mm512_add_pd(v1,v2);
v2 = Optimization::Permute::Permute2(v1);
v1 = _mm512_add_pd(v1,v2);
u512d conv; conv.v = v1;
return conv.f[0];
}
#else
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m512>::operator()(__m512 in){
@ -362,7 +443,6 @@ namespace Optimization {
return _mm512_reduce_add_ps(in);
}
//Complex double Reduce
template<>
inline Grid::ComplexD Reduce<Grid::ComplexD, __m512d>::operator()(__m512d in){
@ -382,6 +462,7 @@ namespace Optimization {
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
}
#endif
}
@ -418,6 +499,7 @@ namespace Optimization {
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;

718
lib/simd/Grid_generic.h
View File

@ -6,8 +6,7 @@
Copyright (C) 2015
Author: Peter Boyle <paboyle@ph.ed.ac.uk>
Author: neo <cossu@post.kek.jp>
Author: Antonin Portelli <antonin.portelli@me.com>
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -27,133 +26,352 @@ Author: neo <cossu@post.kek.jp>
*************************************************************************************/
/* END LEGAL */
static_assert(GEN_SIMD_WIDTH % 16u == 0, "SIMD vector size is not an integer multiple of 16 bytes");
//#define VECTOR_LOOPS
// playing with compiler pragmas
#ifdef VECTOR_LOOPS
#ifdef __clang__
#define VECTOR_FOR(i, w, inc)\
_Pragma("clang loop unroll(full) vectorize(enable) interleave(enable) vectorize_width(w)")\
for (unsigned int i = 0; i < w; i += inc)
#elif defined __INTEL_COMPILER
#define VECTOR_FOR(i, w, inc)\
_Pragma("simd vectorlength(w*8)")\
for (unsigned int i = 0; i < w; i += inc)
#else
#define VECTOR_FOR(i, w, inc)\
for (unsigned int i = 0; i < w; i += inc)
#endif
#else
#define VECTOR_FOR(i, w, inc)\
for (unsigned int i = 0; i < w; i += inc)
#endif
namespace Grid {
namespace Optimization {
template<class vtype>
union uconv {
float f;
vtype v;
// type traits giving the number of elements for each vector type
template <typename T> struct W;
template <> struct W<double> {
constexpr static unsigned int c = GEN_SIMD_WIDTH/16u;
constexpr static unsigned int r = GEN_SIMD_WIDTH/8u;
};
union u128f {
float v;
float f[4];
};
union u128d {
double v;
double f[2];
template <> struct W<float> {
constexpr static unsigned int c = GEN_SIMD_WIDTH/8u;
constexpr static unsigned int r = GEN_SIMD_WIDTH/4u;
};
// SIMD vector types
template <typename T>
struct vec {
alignas(GEN_SIMD_WIDTH) T v[W<T>::r];
};
typedef vec<float> vecf;
typedef vec<double> vecd;
struct Vsplat{
//Complex float
inline u128f operator()(float a, float b){
u128f out;
out.f[0] = a;
out.f[1] = b;
out.f[2] = a;
out.f[3] = b;
// Complex
template <typename T>
inline vec<T> operator()(T a, T b){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 2)
{
out.v[i] = a;
out.v[i+1] = b;
}
return out;
}
// Real float
inline u128f operator()(float a){
u128f out;
out.f[0] = a;
out.f[1] = a;
out.f[2] = a;
out.f[3] = a;
// Real
template <typename T>
inline vec<T> operator()(T a){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 1)
{
out.v[i] = a;
}
return out;
}
//Complex double
inline u128d operator()(double a, double b){
u128d out;
out.f[0] = a;
out.f[1] = b;
return out;
}
//Real double
inline u128d operator()(double a){
u128d out;
out.f[0] = a;
out.f[1] = a;
return out;
}
//Integer
// Integer
inline int operator()(Integer a){
return a;
}
};
struct Vstore{
//Float
inline void operator()(u128f a, float* F){
memcpy(F,a.f,4*sizeof(float));
}
//Double
inline void operator()(u128d a, double* D){
memcpy(D,a.f,2*sizeof(double));
// Real
template <typename T>
inline void operator()(vec<T> a, T *D){
*((vec<T> *)D) = a;
}
//Integer
inline void operator()(int a, Integer* I){
I[0] = a;
inline void operator()(int a, Integer *I){
*I = a;
}
};
struct Vstream{
//Float
inline void operator()(float * a, u128f b){
memcpy(a,b.f,4*sizeof(float));
// Real
template <typename T>
inline void operator()(T * a, vec<T> b){
*((vec<T> *)a) = b;
}
//Double
inline void operator()(double * a, u128d b){
memcpy(a,b.f,2*sizeof(double));
}
};
struct Vset{
// Complex float
inline u128f operator()(Grid::ComplexF *a){
u128f out;
out.f[0] = a[0].real();
out.f[1] = a[0].imag();
out.f[2] = a[1].real();
out.f[3] = a[1].imag();
// Complex
template <typename T>
inline vec<T> operator()(std::complex<T> *a){
vec<T> out;
VECTOR_FOR(i, W<T>::c, 1)
{
out.v[2*i] = a[i].real();
out.v[2*i+1] = a[i].imag();
}
return out;
}
// Complex double
inline u128d operator()(Grid::ComplexD *a){
u128d out;
out.f[0] = a[0].real();
out.f[1] = a[0].imag();
return out;
}
// Real float
inline u128f operator()(float *a){
u128f out;
out.f[0] = a[0];
out.f[1] = a[1];
out.f[2] = a[2];
out.f[3] = a[3];
return out;
}
// Real double
inline u128d operator()(double *a){
u128d out;
out.f[0] = a[0];
out.f[1] = a[1];
// Real
template <typename T>
inline vec<T> operator()(T *a){
vec<T> out;
out = *((vec<T> *)a);
return out;
}
// Integer
inline int operator()(Integer *a){
return a[0];
return *a;
}
};
/////////////////////////////////////////////////////
// Arithmetic operations
/////////////////////////////////////////////////////
struct Sum{
// Complex/Real
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 1)
{
out.v[i] = a.v[i] + b.v[i];
}
return out;
}
//I nteger
inline int operator()(int a, int b){
return a + b;
}
};
struct Sub{
// Complex/Real
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 1)
{
out.v[i] = a.v[i] - b.v[i];
}
return out;
}
//Integer
inline int operator()(int a, int b){
return a-b;
}
};
struct Mult{
// Real
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 1)
{
out.v[i] = a.v[i]*b.v[i];
}
return out;
}
// Integer
inline int operator()(int a, int b){
return a*b;
}
};
#define cmul(a, b, c, i)\
c[i] = a[i]*b[i] - a[i+1]*b[i+1];\
c[i+1] = a[i]*b[i+1] + a[i+1]*b[i];
struct MultComplex{
// Complex
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::c, 1)
{
cmul(a.v, b.v, out.v, 2*i);
}
return out;
}
};
#undef cmul
struct Div{
// Real
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::r, 1)
{
out.v[i] = a.v[i]/b.v[i];
}
return out;
}
};
#define conj(a, b, i)\
b[i] = a[i];\
b[i+1] = -a[i+1];
struct Conj{
// Complex
template <typename T>
inline vec<T> operator()(vec<T> a){
vec<T> out;
VECTOR_FOR(i, W<T>::c, 1)
{
conj(a.v, out.v, 2*i);
}
return out;
}
};
#undef conj
#define timesmi(a, b, i)\
b[i] = a[i+1];\
b[i+1] = -a[i];
struct TimesMinusI{
// Complex
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::c, 1)
{
timesmi(a.v, out.v, 2*i);
}
return out;
}
};
#undef timesmi
#define timesi(a, b, i)\
b[i] = -a[i+1];\
b[i+1] = a[i];
struct TimesI{
// Complex
template <typename T>
inline vec<T> operator()(vec<T> a, vec<T> b){
vec<T> out;
VECTOR_FOR(i, W<T>::c, 1)
{
timesi(a.v, out.v, 2*i);
}
return out;
}
};
#undef timesi
//////////////////////////////////////////////
// Some Template specialization
#define perm(a, b, n, w)\
unsigned int _mask = w >> (n + 1);\
VECTOR_FOR(i, w, 1)\
{\
b[i] = a[i^_mask];\
}
#define DECL_PERMUTE_N(n)\
template <typename T>\
static inline vec<T> Permute##n(vec<T> in) {\
vec<T> out;\
perm(in.v, out.v, n, W<T>::r);\
return out;\
}
struct Permute{
DECL_PERMUTE_N(0);
DECL_PERMUTE_N(1);
DECL_PERMUTE_N(2);
DECL_PERMUTE_N(3);
};
#undef perm
#undef DECL_PERMUTE_N
#define rot(a, b, n, w)\
VECTOR_FOR(i, w, 1)\
{\
b[i] = a[(i + n)%w];\
}
struct Rotate{
template <typename T>
static inline vec<T> rotate(vec<T> in, int n){
vec<T> out;
rot(in.v, out.v, n, W<T>::r);
return out;
}
};
#undef rot
#define acc(v, a, off, step, n)\
for (unsigned int i = off; i < n; i += step)\
{\
a += v[i];\
}
template <typename Out_type, typename In_type>
struct Reduce{
//Need templated class to overload output type
@ -164,316 +382,67 @@ namespace Optimization {
return 0;
}
};
/////////////////////////////////////////////////////
// Arithmetic operations
/////////////////////////////////////////////////////
struct Sum{
//Complex/Real float
inline u128f operator()(u128f a, u128f b){
u128f out;
out.f[0] = a.f[0] + b.f[0];
out.f[1] = a.f[1] + b.f[1];
out.f[2] = a.f[2] + b.f[2];
out.f[3] = a.f[3] + b.f[3];
return out;
}
//Complex/Real double
inline u128d operator()(u128d a, u128d b){
u128d out;
out.f[0] = a.f[0] + b.f[0];
out.f[1] = a.f[1] + b.f[1];
return out;
}
//Integer
inline int operator()(int a, int b){
return a + b;
}
};
struct Sub{
//Complex/Real float
inline u128f operator()(u128f a, u128f b){
u128f out;
out.f[0] = a.f[0] - b.f[0];
out.f[1] = a.f[1] - b.f[1];
out.f[2] = a.f[2] - b.f[2];
out.f[3] = a.f[3] - b.f[3];
return out;
}
//Complex/Real double
inline u128d operator()(u128d a, u128d b){
u128d out;
out.f[0] = a.f[0] - b.f[0];
out.f[1] = a.f[1] - b.f[1];
return out;
}
//Integer
inline int operator()(int a, int b){
return a-b;
}
};
struct MultComplex{
// Complex float
inline u128f operator()(u128f a, u128f b){
u128f out;
out.f[0] = a.f[0]*b.f[0] - a.f[1]*b.f[1];
out.f[1] = a.f[0]*b.f[1] + a.f[1]*b.f[0];
out.f[2] = a.f[2]*b.f[2] - a.f[3]*b.f[3];
out.f[3] = a.f[2]*b.f[3] + a.f[3]*b.f[2];
return out;
}
// Complex double
inline u128d operator()(u128d a, u128d b){
u128d out;
out.f[0] = a.f[0]*b.f[0] - a.f[1]*b.f[1];
out.f[1] = a.f[0]*b.f[1] + a.f[1]*b.f[0];
return out;
}
};
struct Mult{
//CK: Appear unneeded
// inline float mac(float a, float b,double c){
// return 0;
// }
// inline double mac(double a, double b,double c){
// return 0;
// }
// Real float
inline u128f operator()(u128f a, u128f b){
u128f out;
out.f[0] = a.f[0]*b.f[0];
out.f[1] = a.f[1]*b.f[1];
out.f[2] = a.f[2]*b.f[2];
out.f[3] = a.f[3]*b.f[3];
return out;
}
// Real double
inline u128d operator()(u128d a, u128d b){
u128d out;
out.f[0] = a.f[0]*b.f[0];
out.f[1] = a.f[1]*b.f[1];
return out;
}
// Integer
inline int operator()(int a, int b){
return a*b;
}
};
struct Conj{
// Complex single
inline u128f operator()(u128f in){
u128f out;
out.f[0] = in.f[0];
out.f[1] = -in.f[1];
out.f[2] = in.f[2];
out.f[3] = -in.f[3];
return out;
}
// Complex double
inline u128d operator()(u128d in){
u128d out;
out.f[0] = in.f[0];
out.f[1] = -in.f[1];
return out;
}
// do not define for integer input
};
struct TimesMinusI{
//Complex single
inline u128f operator()(u128f in, u128f ret){ //note ret is ignored
u128f out;
out.f[0] = in.f[1];
out.f[1] = -in.f[0];
out.f[2] = in.f[3];
out.f[3] = -in.f[2];
return out;
}
//Complex double
inline u128d operator()(u128d in, u128d ret){
u128d out;
out.f[0] = in.f[1];
out.f[1] = -in.f[0];
return out;
}
};
struct TimesI{
//Complex single
inline u128f operator()(u128f in, u128f ret){ //note ret is ignored
u128f out;
out.f[0] = -in.f[1];
out.f[1] = in.f[0];
out.f[2] = -in.f[3];
out.f[3] = in.f[2];
return out;
}
//Complex double
inline u128d operator()(u128d in, u128d ret){
u128d out;
out.f[0] = -in.f[1];
out.f[1] = in.f[0];
return out;
}
};
//////////////////////////////////////////////
// Some Template specialization
struct Permute{
//We just have to mirror the permutes of Grid_sse4.h
static inline u128f Permute0(u128f in){ //AB CD -> CD AB
u128f out;
out.f[0] = in.f[2];
out.f[1] = in.f[3];
out.f[2] = in.f[0];
out.f[3] = in.f[1];
return out;
};
static inline u128f Permute1(u128f in){ //AB CD -> BA DC
u128f out;
out.f[0] = in.f[1];
out.f[1] = in.f[0];
out.f[2] = in.f[3];
out.f[3] = in.f[2];
return out;
};
static inline u128f Permute2(u128f in){
return in;
};
static inline u128f Permute3(u128f in){
return in;
};
static inline u128d Permute0(u128d in){ //AB -> BA
u128d out;
out.f[0] = in.f[1];
out.f[1] = in.f[0];
return out;
};
static inline u128d Permute1(u128d in){
return in;
};
static inline u128d Permute2(u128d in){
return in;
};
static inline u128d Permute3(u128d in){
return in;
};
};
template < typename vtype >
void permute(vtype &a, vtype b, int perm) {
};
struct Rotate{
static inline u128f rotate(u128f in,int n){
u128f out;
switch(n){
case 0:
out.f[0] = in.f[0];
out.f[1] = in.f[1];
out.f[2] = in.f[2];
out.f[3] = in.f[3];
break;
case 1:
out.f[0] = in.f[1];
out.f[1] = in.f[2];
out.f[2] = in.f[3];
out.f[3] = in.f[0];
break;
case 2:
out.f[0] = in.f[2];
out.f[1] = in.f[3];
out.f[2] = in.f[0];
out.f[3] = in.f[1];
break;
case 3:
out.f[0] = in.f[3];
out.f[1] = in.f[0];
out.f[2] = in.f[1];
out.f[3] = in.f[2];
break;
default: assert(0);
}
return out;
}
static inline u128d rotate(u128d in,int n){
u128d out;
switch(n){
case 0:
out.f[0] = in.f[0];
out.f[1] = in.f[1];
break;
case 1:
out.f[0] = in.f[1];
out.f[1] = in.f[0];
break;
default: assert(0);
}
return out;
}
};
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, u128f>::operator()(u128f in){ //2 complex
return Grid::ComplexF(in.f[0] + in.f[2], in.f[1] + in.f[3]);
template <>
inline Grid::ComplexF Reduce<Grid::ComplexF, vecf>::operator()(vecf in){
float a = 0.f, b = 0.f;
acc(in.v, a, 0, 2, W<float>::r);
acc(in.v, b, 1, 2, W<float>::r);
return Grid::ComplexF(a, b);
}
//Real float Reduce
template<>
inline Grid::RealF Reduce<Grid::RealF, u128f>::operator()(u128f in){ //4 floats
return in.f[0] + in.f[1] + in.f[2] + in.f[3];
inline Grid::RealF Reduce<Grid::RealF, vecf>::operator()(vecf in){
float a = 0.;
acc(in.v, a, 0, 1, W<float>::r);
return a;
}
//Complex double Reduce
template<>
inline Grid::ComplexD Reduce<Grid::ComplexD, u128d>::operator()(u128d in){ //1 complex
return Grid::ComplexD(in.f[0],in.f[1]);
inline Grid::ComplexD Reduce<Grid::ComplexD, vecd>::operator()(vecd in){
double a = 0., b = 0.;
acc(in.v, a, 0, 2, W<double>::r);
acc(in.v, b, 1, 2, W<double>::r);
return Grid::ComplexD(a, b);
}
//Real double Reduce
template<>
inline Grid::RealD Reduce<Grid::RealD, u128d>::operator()(u128d in){ //2 doubles
return in.f[0] + in.f[1];
inline Grid::RealD Reduce<Grid::RealD, vecd>::operator()(vecd in){
double a = 0.f;
acc(in.v, a, 0, 1, W<double>::r);
return a;
}
//Integer Reduce
template<>
inline Integer Reduce<Integer, int>::operator()(int in){
// FIXME unimplemented
printf("Reduce : Missing integer implementation -> FIX\n");
assert(0);
return in;
}
}
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
typedef Optimization::u128f SIMD_Ftype; // Single precision type
typedef Optimization::u128d SIMD_Dtype; // Double precision type
typedef Optimization::vecf SIMD_Ftype; // Single precision type
typedef Optimization::vecd SIMD_Dtype; // Double precision type
typedef int SIMD_Itype; // Integer type
// prefetch utilities
inline void v_prefetch0(int size, const char *ptr){};
inline void prefetch_HINT_T0(const char *ptr){};
// Gpermute function
template < typename VectorSIMD >
inline void Gpermute(VectorSIMD &y,const VectorSIMD &b, int perm ) {
Optimization::permute(y.v,b.v,perm);
}
// Function name aliases
typedef Optimization::Vsplat VsplatSIMD;
typedef Optimization::Vstore VstoreSIMD;
@ -481,16 +450,13 @@ namespace Optimization {
typedef Optimization::Vstream VstreamSIMD;
template <typename S, typename T> using ReduceSIMD = Optimization::Reduce<S,T>;
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;
typedef Optimization::TimesMinusI TimesMinusISIMD;
typedef Optimization::TimesI TimesISIMD;
}

12
lib/simd/Grid_imci.h
View File

@ -236,6 +236,17 @@ namespace Optimization {
}
};
struct Div{
// Real float
inline __m512 operator()(__m512 a, __m512 b){
return _mm512_div_ps(a,b);
}
// Real double
inline __m512d operator()(__m512d a, __m512d b){
return _mm512_div_pd(a,b);
}
};
struct Conj{
// Complex single
@ -429,6 +440,7 @@ namespace Optimization {
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;

15
lib/simd/Grid_sse4.h
View File

@ -224,6 +224,18 @@ namespace Optimization {
}
};
struct Div{
// Real float
inline __m128 operator()(__m128 a, __m128 b){
return _mm_div_ps(a,b);
}
// Real double
inline __m128d operator()(__m128d a, __m128d b){
return _mm_div_pd(a,b);
}
};
struct Conj{
// Complex single
inline __m128 operator()(__m128 in){
@ -372,6 +384,8 @@ namespace Optimization {
}
}
//////////////////////////////////////////////////////////////////////////////////////
// Here assign types
@ -398,6 +412,7 @@ namespace Optimization {
// Arithmetic operations
typedef Optimization::Sum SumSIMD;
typedef Optimization::Sub SubSIMD;
typedef Optimization::Div DivSIMD;
typedef Optimization::Mult MultSIMD;
typedef Optimization::MultComplex MultComplexSIMD;
typedef Optimization::Conj ConjSIMD;

90
lib/simd/Grid_vector_types.h
View File

@ -38,7 +38,7 @@ directory
#ifndef GRID_VECTOR_TYPES
#define GRID_VECTOR_TYPES
#ifdef GENERIC_VEC
#ifdef GEN
#include "Grid_generic.h"
#endif
#ifdef SSE4
@ -77,38 +77,24 @@ struct RealPart<std::complex<T> > {
//////////////////////////////////////
// demote a vector to real type
//////////////////////////////////////
// type alias used to simplify the syntax of std::enable_if
template <typename T>
using Invoke = typename T::type;
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 T> using Invoke = typename T::type;
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> >;
////////////////////////////////////////////////////////
// Check for complexity with type traits
template <typename T>
struct is_complex : public std::false_type {};
template <>
struct is_complex<std::complex<double> > : public std::true_type {};
template <>
struct is_complex<std::complex<float> > : public std::true_type {};
template <typename T> struct is_complex : public std::false_type {};
template <> struct is_complex<std::complex<double> > : public std::true_type {};
template <> struct is_complex<std::complex<float> > : public std::true_type {};
template <typename T>
using IfReal = Invoke<std::enable_if<std::is_floating_point<T>::value, int> >;
template <typename T>
using IfComplex = Invoke<std::enable_if<is_complex<T>::value, int> >;
template <typename T>
using IfInteger = Invoke<std::enable_if<std::is_integral<T>::value, int> >;
template <typename T> using IfReal = Invoke<std::enable_if<std::is_floating_point<T>::value, int> >;
template <typename T> using IfComplex = Invoke<std::enable_if<is_complex<T>::value, int> >;
template <typename T> using IfInteger = Invoke<std::enable_if<std::is_integral<T>::value, int> >;
template <typename T>
using IfNotReal =
Invoke<std::enable_if<!std::is_floating_point<T>::value, int> >;
template <typename T>
using IfNotComplex = Invoke<std::enable_if<!is_complex<T>::value, int> >;
template <typename T>
using IfNotInteger = Invoke<std::enable_if<!std::is_integral<T>::value, int> >;
template <typename T> using IfNotReal = Invoke<std::enable_if<!std::is_floating_point<T>::value, int> >;
template <typename T> using IfNotComplex = Invoke<std::enable_if<!is_complex<T>::value, int> >;
template <typename T> using IfNotInteger = Invoke<std::enable_if<!std::is_integral<T>::value, int> >;
////////////////////////////////////////////////////////
// Define the operation templates functors
@ -285,6 +271,20 @@ class Grid_simd {
return a * b;
}
//////////////////////////////////
// Divides
//////////////////////////////////
friend inline Grid_simd operator/(const Scalar_type &a, Grid_simd b) {
Grid_simd va;
vsplat(va, a);
return va / b;
}
friend inline Grid_simd operator/(Grid_simd b, const Scalar_type &a) {
Grid_simd va;
vsplat(va, a);
return b / a;
}
///////////////////////
// Unary negation
///////////////////////
@ -428,7 +428,6 @@ inline void rotate(Grid_simd<S,V> &ret,Grid_simd<S,V> b,int nrot)
ret.v = Optimization::Rotate::rotate(b.v,2*nrot);
}
template <class S, class V>
inline void vbroadcast(Grid_simd<S,V> &ret,const Grid_simd<S,V> &src,int lane){
S* typepun =(S*) &src;
@ -512,7 +511,6 @@ template <class S, class V, IfInteger<S> = 0>
inline void vfalse(Grid_simd<S, V> &ret) {
vsplat(ret, 0);
}
template <class S, class V>
inline void zeroit(Grid_simd<S, V> &z) {
vzero(z);
@ -530,7 +528,6 @@ inline void vstream(Grid_simd<S, V> &out, const Grid_simd<S, V> &in) {
typedef typename S::value_type T;
binary<void>((T *)&out.v, in.v, VstreamSIMD());
}
template <class S, class V, IfInteger<S> = 0>
inline void vstream(Grid_simd<S, V> &out, const Grid_simd<S, V> &in) {
out = in;
@ -569,6 +566,34 @@ inline Grid_simd<S, V> operator*(Grid_simd<S, V> a, Grid_simd<S, V> b) {
return ret;
};
// Distinguish between complex types and others
template <class S, class V, IfComplex<S> = 0>
inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
typedef Grid_simd<S, V> simd;
simd ret;
simd den;
typename simd::conv_t conv;
ret = a * conjugate(b) ;
den = b * conjugate(b) ;
auto real_den = toReal(den);
ret.v=binary<V>(ret.v, real_den.v, DivSIMD());
return ret;
};
// Real/Integer types
template <class S, class V, IfNotComplex<S> = 0>
inline Grid_simd<S, V> operator/(Grid_simd<S, V> a, Grid_simd<S, V> b) {
Grid_simd<S, V> ret;
ret.v = binary<V>(a.v, b.v, DivSIMD());
return ret;
};
///////////////////////
// Conjugate
///////////////////////
@ -582,7 +607,6 @@ template <class S, class V, IfNotComplex<S> = 0>
inline Grid_simd<S, V> conjugate(const Grid_simd<S, V> &in) {
return in; // for real objects
}
// Suppress adj for integer types... // odd; why conjugate above but not adj??
template <class S, class V, IfNotInteger<S> = 0>
inline Grid_simd<S, V> adj(const Grid_simd<S, V> &in) {
@ -596,14 +620,12 @@ template <class S, class V, IfComplex<S> = 0>
inline void timesMinusI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in) {
ret.v = binary<V>(in.v, ret.v, TimesMinusISIMD());
}
template <class S, class V, IfComplex<S> = 0>
inline Grid_simd<S, V> timesMinusI(const Grid_simd<S, V> &in) {
Grid_simd<S, V> ret;
timesMinusI(ret, in);
return ret;
}
template <class S, class V, IfNotComplex<S> = 0>
inline Grid_simd<S, V> timesMinusI(const Grid_simd<S, V> &in) {
return in;
@ -616,14 +638,12 @@ template <class S, class V, IfComplex<S> = 0>
inline void timesI(Grid_simd<S, V> &ret, const Grid_simd<S, V> &in) {
ret.v = binary<V>(in.v, ret.v, TimesISIMD());
}
template <class S, class V, IfComplex<S> = 0>
inline Grid_simd<S, V> timesI(const Grid_simd<S, V> &in) {
Grid_simd<S, V> ret;
timesI(ret, in);
return ret;
}
template <class S, class V, IfNotComplex<S> = 0>
inline Grid_simd<S, V> timesI(const Grid_simd<S, V> &in) {
return in;

2
lib/simd/Intel512avx.h
View File

@ -53,7 +53,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define ZMULMEM2SPd(O,P,tmp,B,C,Briir,Biirr,Criir,Ciirr)\
VSHUFMEMd(O,P,tmp) \
VMULMEMd(O,P,B,Biirr) \
VMULMEMd(O,P,B,Biirr) \
VMULMEMd(O,P,C,Ciirr) \
VMULd(tmp,B,Briir) \
VMULd(tmp,C,Criir)

9
lib/simd/Intel512common.h
View File

@ -37,7 +37,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
// Opcodes common
////////////////////////////////////////////////////////////////////////////////////////////////////
#define MASK_REGS \
__asm__ ("mov $0xAAAA, %%eax \n"\
__asm__ ("mov $0xAAAA, %%eax \n"\
"kmovw %%eax, %%k6 \n"\
"mov $0x5555, %%eax \n"\
"kmovw %%eax, %%k7 \n" : : : "%eax");
@ -138,9 +138,14 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define ZLOADf(OFF,PTR,ri,ir) VLOADf(OFF,PTR,ir) VSHUFf(ir,ri)
#define ZLOADd(OFF,PTR,ri,ir) VLOADd(OFF,PTR,ir) VSHUFd(ir,ri)
#define STREAM_STORE
#ifdef STREAM_STORE
#define VSTOREf(OFF,PTR,SRC) "vmovntps " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#define VSTOREd(OFF,PTR,SRC) "vmovntpd " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#else
#define VSTOREf(OFF,PTR,SRC) "vmovaps " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#define VSTOREd(OFF,PTR,SRC) "vmovapd " #SRC "," #OFF "*64(" #PTR ")" ";\n"
#endif
// Swaps Re/Im ; could unify this with IMCI
#define VSHUFd(A,DEST) "vpshufd $0x4e," #A "," #DEST ";\n"