portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-04 19:25:56 +01:00
Code Releases Activity

Verified

Browse Source
This commit is contained in:
azusayamaguchi 2017-02-21 23:01:25 +00:00
parent bf7e3f20d4
commit 1c30e9a961
8 changed files with 163 additions and 153 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
benchmarks/Benchmark_staggered.cc
View File

@ -115,7 +115,7 @@ int main (int argc, char ** argv)
ImprovedStaggeredFermionR Ds(Umu,Umu,Grid,RBGrid,mass,c1,c2,u0,params);
std::cout<<GridLogMessage << "Calling Ds"<<std::endl;
int ncall=100000;
int ncall=1000;
double t0=usecond();
for(int i=0;i<ncall;i++){
Ds.Dhop(src,result,0);

4
lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
View File

@ -338,12 +338,12 @@ void ImprovedStaggeredFermion<Impl>::DhopInternal(StencilImpl &st, LebesgueOrder
if (dag == DaggerYes) {
PARALLEL_FOR_LOOP
for (int sss = 0; sss < in._grid->oSites(); sss++) {
Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), sss, sss, in, out);
Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
}
} else {
PARALLEL_FOR_LOOP
for (int sss = 0; sss < in._grid->oSites(); sss++) {
Kernels::DhopSite(st, lo, U, UUU, st.CommBuf(), sss, sss, in, out);
Kernels::DhopSite(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
}
}
};

19
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -228,9 +228,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
const FermionField &in, FermionField &out,int dag)
{
Compressor compressor;
int LLs = in._grid->_rdimensions[0];
st.HaloExchange(in,compressor);
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
@ -241,28 +239,11 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), LLs, sU,in, out);
}
} else {
#if 1
PARALLEL_FOR_LOOP
for (int ss = 0; ss < U._grid->oSites(); ss++) {
int sU=ss;
Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out);
}
#else
#pragma omp parallel
{
for(int i=0;i<10;i++){
int len = U._grid->oSites();
int me,mywork,myoff;
GridThread::GetWorkBarrier(len,me, mywork,myoff);
for (int ss = myoff; ss < myoff+mywork; ss++) {
int sU=ss;
int sF=LLs*sU;
Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),LLs,sU,in,out);
}
GridThread::ThreadBarrier();
}
}
#endif
}
}

56
lib/qcd/action/fermion/StaggeredKernels.cc
View File

@ -186,32 +186,31 @@ template <class Impl>
void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs, int sU,
const FermionField &in, FermionField &out) {
int dag(1);
SiteSpinor naik;
SiteSpinor naive;
int oneLink =0;
int threeLink=1;
Real scale;
if(dag) scale = -1.0;
else scale = 1.0;
int dag=1;
switch(Opt) {
#ifdef AVX512
//FIXME; move the sign into the Asm routine
case OptInlineAsm:
DhopSiteAsm(st,lo,U,UUU,buf,LLs,sU,in,out);
for(int s=0;s<LLs;s++) {
int sF=s+LLs*sU;
out._odata[sF]=-out._odata[sF];
}
break;
#endif
case OptHandUnroll:
DhopSiteDepthHand(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
DhopSiteHand(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
break;
case OptGeneric:
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
out._odata[sF] =scale*(naive+naik);
out._odata[sF] =-naive-naik;
}
break;
default:
@ -223,17 +222,13 @@ void StaggeredKernels<Impl>::DhopSiteDag(StencilImpl &st, LebesgueOrder &lo, Dou
template <class Impl>
void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionField &in, FermionField &out) {
int dag(0);
int oneLink =0;
int threeLink=1;
SiteSpinor naik;
SiteSpinor naive;
static int once;
int sF=LLs*sU;
int sU, const FermionField &in, FermionField &out)
{
int oneLink =0;
int threeLink=1;
SiteSpinor naik;
SiteSpinor naive;
int dag=0;
switch(Opt) {
#ifdef AVX512
case OptInlineAsm:
@ -241,22 +236,23 @@ void StaggeredKernels<Impl>::DhopSite(StencilImpl &st, LebesgueOrder &lo, Double
break;
#endif
case OptHandUnroll:
DhopSiteDepthHand(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
break;
DhopSiteHand(st,lo,U,UUU,buf,LLs,sU,in,out,dag);
break;
case OptGeneric:
for(int s=0;s<LLs;s++){
DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
out._odata[sF] =naive+naik;
}
for(int s=0;s<LLs;s++){
int sF=LLs*sU+s;
// assert(sF<in._odata.size());
// assert(sU< U._odata.size());
// assert(sF>=0); assert(sU>=0);
DhopSiteDepth(st,lo,U,buf,sF,sU,in,naive,oneLink);
DhopSiteDepth(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
out._odata[sF] =naive+naik;
}
break;
default:
assert(0);
break;
}
};
template <class Impl>

6
lib/qcd/action/fermion/StaggeredKernels.h
View File

@ -57,11 +57,11 @@ public:
int sF, int sU, const FermionField &in, SiteSpinor &out,int threeLink);
void DhopSiteDepthHandLocal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
void DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, SiteSpinor * buf,
int sF, int sU, const FermionField &in, SiteSpinor&out,int threeLink);
void DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,SiteSpinor * buf,
int Lls, int sU, const FermionField &in, FermionField &out, int dag);
void DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U, DoubledGaugeField &UUU,SiteSpinor * buf,
int LLs, int sU, const FermionField &in, FermionField &out, int dag);
void DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU, SiteSpinor * buf,
int LLs, int sU, const FermionField &in, FermionField &out);

159
lib/qcd/action/fermion/StaggeredKernelsAsm.cc
View File

@ -517,7 +517,7 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
: : "r" (a0) : "%r8" ); \
#define PF_GAUGE_XYZT(a0)
#define PF_GAUGE_XYZTa(a0) \
#define PF_GAUGE_XYZTa(a0) \
asm ( \
"movq %0, %%r8 \n\t" \
VPREFETCH1(0,%%r8) \
@ -578,10 +578,10 @@ namespace QCD {
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionField &in, FermionField &out)
DoubledGaugeField &U,
DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionField &in, FermionField &out)
{
assert(0);
@ -611,35 +611,35 @@ void StaggeredKernels<Impl>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
o0 = SE0->_offset; \
l0 = SE0->_is_local; \
p0 = SE0->_permute; \
CONDITIONAL_MOVE(l0,o0,addr0); \
CONDITIONAL_MOVE(l0,o0,addr0); \
PF_CHI(addr0); \
\
SE1=st.GetEntry(ptype,Y+skew,sF); \
o1 = SE1->_offset; \
l1 = SE1->_is_local; \
p1 = SE1->_permute; \
CONDITIONAL_MOVE(l1,o1,addr1); \
\
SE1=st.GetEntry(ptype,Y+skew,sF); \
o1 = SE1->_offset; \
l1 = SE1->_is_local; \
p1 = SE1->_permute; \
CONDITIONAL_MOVE(l1,o1,addr1); \
PF_CHI(addr1); \
\
SE2=st.GetEntry(ptype,Z+skew,sF); \
o2 = SE2->_offset; \
l2 = SE2->_is_local; \
p2 = SE2->_permute; \
CONDITIONAL_MOVE(l2,o2,addr2); \
\
SE2=st.GetEntry(ptype,Z+skew,sF); \
o2 = SE2->_offset; \
l2 = SE2->_is_local; \
p2 = SE2->_permute; \
CONDITIONAL_MOVE(l2,o2,addr2); \
PF_CHI(addr2); \
\
SE3=st.GetEntry(ptype,T+skew,sF); \
o3 = SE3->_offset; \
l3 = SE3->_is_local; \
p3 = SE3->_permute; \
CONDITIONAL_MOVE(l3,o3,addr3); \
\
SE3=st.GetEntry(ptype,T+skew,sF); \
o3 = SE3->_offset; \
l3 = SE3->_is_local; \
p3 = SE3->_permute; \
CONDITIONAL_MOVE(l3,o3,addr3); \
PF_CHI(addr3); \
\
gauge0 =(uint64_t)&UU._odata[sU]( X ); \
gauge1 =(uint64_t)&UU._odata[sU]( Y ); \
gauge2 =(uint64_t)&UU._odata[sU]( Z ); \
gauge0 =(uint64_t)&UU._odata[sU]( X ); \
gauge1 =(uint64_t)&UU._odata[sU]( Y ); \
gauge2 =(uint64_t)&UU._odata[sU]( Z ); \
gauge3 =(uint64_t)&UU._odata[sU]( T );
// This is the single precision 5th direction vectorised kernel
#include <simd/Intel512single.h>
template <> void StaggeredKernels<StaggeredVec5dImplF>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
@ -762,6 +762,14 @@ template <> void StaggeredKernels<StaggeredVec5dImplD>::DhopSiteAsm(StencilImpl
VPERM0(Chi_11,Chi_11) \
VPERM0(Chi_12,Chi_12) );
#define PERMUTE01 \
if ( p0 ) { PERMUTE_DIR3; }\
if ( p1 ) { PERMUTE_DIR2; }
#define PERMUTE23 \
if ( p2 ) { PERMUTE_DIR1; }\
if ( p3 ) { PERMUTE_DIR0; }
// This is the single precision 5th direction vectorised kernel
#include <simd/Intel512single.h>
@ -785,35 +793,50 @@ template <> void StaggeredKernels<StaggeredImplF>::DhopSiteAsm(StencilImpl &st,
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
addr0 = (uint64_t) &out._odata[sF];
REDUCE(addr0);
}
REDUCEa(addr0);
}
#else
assert(0);
assert(0);
#endif
}
#include <simd/Intel512double.h>
template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st, LebesgueOrder &lo,
DoubledGaugeField &U,
@ -835,31 +858,47 @@ template <> void StaggeredKernels<StaggeredImplD>::DhopSiteAsm(StencilImpl &st,
StencilEntry *SE2;
StencilEntry *SE3;
for(int s=0;s<LLs;s++){
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
// Xp, Yp, Zp, Tp
PREPARE(Xp,Yp,Zp,Tp,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,0,U);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xp,Yp,Zp,Tp,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
PREPARE(Xm,Ym,Zm,Tm,8,UUU);
LOAD_CHI(addr0,addr1,addr2,addr3);
MULT_ADD_LS(gauge0,gauge1,gauge2,gauge3);
LOAD_CHIa(addr0,addr1);
PERMUTE01;
MULT_ADD_XYZT(gauge0,gauge1);
LOAD_CHIa(addr2,addr3);
PERMUTE23;
MULT_ADD_XYZT(gauge2,gauge3);
addr0 = (uint64_t) &out._odata[sF];
REDUCE(addr0);
}
REDUCEa(addr0);
}
#else
assert(0);
assert(0);
#endif
}

48
lib/qcd/action/fermion/StaggeredKernelsHand.cc
View File

@ -91,34 +91,31 @@ namespace QCD {
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionField &in, FermionField &out, int dag) {
SiteSpinor naik;
SiteSpinor naive;
int oneLink =0;
int threeLink=1;
int skew(0);
Real scale(1.0);
if(dag) scale = -1.0;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
DhopSiteDepthHandLocal(st,lo,U,buf,sF,sU,in,naive,oneLink);
DhopSiteDepthHandLocal(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
out._odata[sF] =scale*(naive+naik);
}
void StaggeredKernels<Impl>::DhopSiteHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,DoubledGaugeField &UUU,
SiteSpinor *buf, int LLs,
int sU, const FermionField &in, FermionField &out, int dag)
{
SiteSpinor naik;
SiteSpinor naive;
int oneLink =0;
int threeLink=1;
int skew(0);
Real scale(1.0);
if(dag) scale = -1.0;
for(int s=0;s<LLs;s++){
int sF=s+LLs*sU;
DhopSiteDepthHand(st,lo,U,buf,sF,sU,in,naive,oneLink);
DhopSiteDepthHand(st,lo,UUU,buf,sF,sU,in,naik,threeLink);
out._odata[sF] =scale*(naive+naik);
}
}
template <class Impl>
void StaggeredKernels<Impl>::DhopSiteDepthHandLocal(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
SiteSpinor *buf, int sF,
int sU, const FermionField &in, SiteSpinor &out,int threeLink) {
void StaggeredKernels<Impl>::DhopSiteDepthHand(StencilImpl &st, LebesgueOrder &lo, DoubledGaugeField &U,
SiteSpinor *buf, int sF,
int sU, const FermionField &in, SiteSpinor &out,int threeLink)
{
typedef typename Simd::scalar_type S;
typedef typename Simd::vector_type V;
@ -300,7 +297,6 @@ void StaggeredKernels<Impl>::DhopSiteDepthHandLocal(StencilImpl &st, LebesgueOrd
vstream(out()()(1),even_1+odd_1);
vstream(out()()(2),even_2+odd_2);
}
}
FermOpStaggeredTemplateInstantiate(StaggeredKernels);

22
tests/core/Test_staggered5Dvec.cc
View File

@ -57,34 +57,33 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
std::vector<int> seeds({1,2,3,4});
/*
GridParallelRNG pRNG4(UGrid);
GridParallelRNG pRNG5(FGrid);
pRNG4.SeedFixedIntegers(seeds);
pRNG5.SeedFixedIntegers(seeds);
*/
typedef typename ImprovedStaggeredFermion5DR::FermionField FermionField;
typedef typename ImprovedStaggeredFermion5DR::ComplexField ComplexField;
typename ImprovedStaggeredFermion5DR::ImplParams params;
FermionField src (FGrid); src=zero;
// random(pRNG5,src);
FermionField src (FGrid);
random(pRNG5,src);
/*
std::vector<int> site({0,0,0,0,0});
std::vector<int> site({0,1,2,0,0});
ColourVector cv = zero;
cv()()(0)=1.0;
src = zero;
pokeSite(cv,src,site);
*/
FermionField result(FGrid); result=zero;
FermionField tmp(FGrid); tmp=zero;
FermionField err(FGrid); tmp=zero;
FermionField phi (FGrid); phi=1.0;//random(pRNG5,phi);
FermionField chi (FGrid); chi=1.0;//random(pRNG5,chi);
FermionField phi (FGrid); random(pRNG5,phi);
FermionField chi (FGrid); random(pRNG5,chi);
LatticeGaugeField Umu(UGrid); Umu=1.0; //SU3::HotConfiguration(pRNG4,Umu);
LatticeGaugeField Umu(UGrid);
SU3::HotConfiguration(pRNG4,Umu);
/*
for(int mu=1;mu<4;mu++){
@ -103,7 +102,7 @@ int main (int argc, char ** argv)
RealD c2=-1.0/24.0;
RealD u0=1.0;
ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,c1,c2,u0,params);
ImprovedStaggeredFermion5DR Ds(Umu,Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,c1,c2,u0,params);
ImprovedStaggeredFermionVec5dR sDs(Umu,Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,mass,c1,c2,u0,params);
std::cout<<GridLogMessage<<"=========================================================="<<std::endl;
@ -127,7 +126,6 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
std::cout<<GridLogMessage << "mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "Calling vectorised staggered operator"<<std::endl;
QCD::StaggeredKernelsStatic::Opt=QCD::StaggeredKernelsStatic::OptInlineAsm;