portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-03 21:44:33 +00:00
Code Releases Activity

First version of the doube prec for testing purposes

Browse Source 
It does not compile single and double version at the same time
This commit is contained in:
Guido Cossu
2016-10-28 17:20:04 +01:00
parent ac99a56237
commit e1042aef77
3 changed files with 80 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

77
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@@ -53,12 +53,13 @@ WilsonKernels<Impl >::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,
}
#if defined(AVX512)
#include <simd/Intel512wilson.h>
#if defined(GRID_DEFAULT_PRECISION_SINGLE)
///////////////////////////////////////////////////////////
// If we are AVX512 specialise the single precision routine
///////////////////////////////////////////////////////////
#include <simd/Intel512wilson.h>
#include <simd/Intel512single.h>
static Vector<vComplexF> signs;
@@ -78,6 +79,7 @@ static Vector<vComplexF> signs;
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
#define FX(A) WILSONASM_ ##A
#define COMPLEX_TYPE vComplexF
#undef KERNEL_DAG
template<> void
@@ -113,8 +115,79 @@ template<> void
WilsonKernels<DomainWallVec5dImplF>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef COMPLEX_TYPE
#endif
#endif //Single precision
#if defined(GRID_DEFAULT_PRECISION_DOUBLE)
//temporary separating the two sections
//for debug in isolation
//can be unified
///////////////////////////////////////////////////////////
// If we are AVX512 specialise the double precision routine
///////////////////////////////////////////////////////////
#include <simd/Intel512double.h>
static Vector<vComplexD> signs;
int setupSigns(void ){
Vector<vComplexD> bother(2);
signs = bother;
vrsign(signs[0]);
visign(signs[1]);
return 1;
}
static int signInit = setupSigns();
#define label(A) ilabel(A)
#define ilabel(A) ".globl\n" #A ":\n"
#define MAYBEPERM(A,perm) if (perm) { A ; }
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN(ptr,pf)
#define FX(A) WILSONASM_ ##A
#define COMPLEX_TYPE vComplexD
#undef KERNEL_DAG
template<> void
WilsonKernels<WilsonImplD>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#define KERNEL_DAG
template<> void
WilsonKernels<WilsonImplD>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef VMOVIDUP
#undef VMOVRDUP
#undef MAYBEPERM
#undef MULT_2SPIN
#undef FX
#define FX(A) DWFASM_ ## A
#define MAYBEPERM(A,B)
#define VMOVIDUP(A,B,C) VBCASTIDUPf(A,B,C)
#define VMOVRDUP(A,B,C) VBCASTRDUPf(A,B,C)
#define MULT_2SPIN(ptr,pf) MULT_ADDSUB_2SPIN_LS(ptr,pf)
#undef KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplD>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#define KERNEL_DAG
template<> void
WilsonKernels<DomainWallVec5dImplD>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,SiteHalfSpinor *buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#undef COMPLEX_TYPE
#endif //Double precision
#endif //AVX512
#define INSTANTIATE_ASM(A)\
template void WilsonKernels<A>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U, SiteHalfSpinor *buf,\

4
lib/qcd/action/fermion/WilsonKernelsAsmBody.h
View File

@@ -5,7 +5,9 @@
const uint64_t plocal =(uint64_t) & in._odata[0];
// vComplexF isigns[2] = { signs[0], signs[1] };
vComplexF *isigns = &signs[0];
//COMPLEX_TYPE is vComplexF of vComplexD depending
//on the chosen precision
COMPLEX_TYPE *isigns = &signs[0];
MASK_REGS;
int nmax=U._grid->oSites();

6
lib/simd/Grid_avx512.h
View File

@@ -371,14 +371,8 @@ namespace Optimization {
// Some Template specialization
// Hack for CLANG until mm512_reduce_add_ps etc... are implemented in GCC and Clang releases
<<<<<<< HEAD
#define GNU_CLANG_COMPILER
#ifdef GNU_CLANG_COMPILER
=======
#ifndef __INTEL_COMPILER
#warning "Slow reduction due to incomplete reduce intrinsics"
>>>>>>> develop
//Complex float Reduce
template<>
inline Grid::ComplexF Reduce<Grid::ComplexF, __m512>::operator()(__m512 in){