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 'develop' into feature/hmc_generalise

Browse Source
This commit is contained in:
Guido Cossu
2016-10-20 17:04:41 +01:00
parent 977d844394 7af9b87318
commit 977b0a6dd9
28 changed files with 1119 additions and 348 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@ -186,10 +186,10 @@ namespace Grid {
inline void InsertForce5D(GaugeField &mat, FermionField &Btilde, FermionField &Atilde,int mu){
int Ls=Btilde._grid->_fdimensions[0];
GaugeLinkField tmp(mat._grid);
tmp = zero;
PARALLEL_FOR_LOOP
PARALLEL_FOR_LOOP
for(int sss=0;sss<tmp._grid->oSites();sss++){
int sU=sss;
for(int s=0;s<Ls;s++){
@ -198,7 +198,7 @@ namespace Grid {
}
}
PokeIndex<LorentzIndex>(mat,tmp,mu);
}
};

209
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -42,11 +42,11 @@ const std::vector<int> WilsonFermion5DStatic::displacements({1,1,1,1,-1,-1,-1,-1
// 5d lattice for DWF.
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _M5,const ImplParams &p) :
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
GridRedBlackCartesian &FourDimRedBlackGrid,
RealD _M5,const ImplParams &p) :
Kernels(p),
_FiveDimGrid (&FiveDimGrid),
_FiveDimRedBlackGrid(&FiveDimRedBlackGrid),
@ -135,10 +135,10 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
/*
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
RealD _M5,const ImplParams &p) :
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
RealD _M5,const ImplParams &p) :
{
int nsimd = Simd::Nsimd();
@ -175,6 +175,73 @@ WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
}
*/
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Number of Dhop Calls : " << DhopCalls << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Total Communication time : " << DhopCommTime
<< " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Total Compute time : "
<< DhopComputeTime << " us" << std::endl;
std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
RealD mflops = 1344*volume*DhopCalls/DhopComputeTime;
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NP << std::endl;
}
if ( DerivCalls > 0 ) {
std::cout << GridLogMessage << "#### Deriv calls report "<< std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Number of Deriv Calls : " <<DerivCalls <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Total Communication time : " <<DerivCommTime <<" us"<<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D CommTime/Calls : " <<DerivCommTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Total Compute time : " <<DerivComputeTime <<" us"<<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D ComputeTime/Calls : " <<DerivComputeTime/DerivCalls<<" us" <<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Total Dhop Compute time : " <<DerivDhopComputeTime <<" us"<<std::endl;
std::cout << GridLogMessage << "WilsonFermion5D Dhop ComputeTime/Calls : " <<DerivDhopComputeTime/DerivCalls<<" us" <<std::endl;
RealD mflops = 144*volume*DerivCalls/DerivDhopComputeTime;
std::cout << GridLogMessage << "Average mflops/s per call : " << mflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NP << std::endl;
}
if (DerivCalls > 0 || DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil"<<std::endl; Stencil.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd"<<std::endl; StencilOdd.Report();
}
}
template<class Impl>
void WilsonFermion5D<Impl>::ZeroCounters(void) {
DhopCalls = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DerivCalls = 0;
DerivCommTime = 0;
DerivComputeTime = 0;
DerivDhopComputeTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
template<class Impl>
void WilsonFermion5D<Impl>::ImportGauge(const GaugeField &_Umu)
{
@ -215,12 +282,13 @@ PARALLEL_FOR_LOOP
template<class Impl>
void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
DoubledGaugeField & U,
GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
{
DerivCalls++;
assert((dag==DaggerNo) ||(dag==DaggerYes));
conformable(st._grid,A._grid);
@ -231,51 +299,53 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
FermionField Btilde(B._grid);
FermionField Atilde(B._grid);
DerivCommTime-=usecond();
st.HaloExchange(B,compressor);
DerivCommTime+=usecond();
Atilde=A;
for(int mu=0;mu<Nd;mu++){
DerivComputeTime-=usecond();
for (int mu = 0; mu < Nd; mu++) {
////////////////////////////////////////////////////////////////////////
// Flip gamma if dag
////////////////////////////////////////////////////////////////////////
int gamma = mu;
if ( !dag ) gamma+= Nd;
if (!dag) gamma += Nd;
////////////////////////
// Call the single hop
////////////////////////
PARALLEL_FOR_LOOP
for(int sss=0;sss<U._grid->oSites();sss++){
for(int s=0;s<Ls;s++){
int sU=sss;
int sF = s+Ls*sU;
DerivDhopComputeTime -= usecond();
PARALLEL_FOR_LOOP
for (int sss = 0; sss < U._grid->oSites(); sss++) {
for (int s = 0; s < Ls; s++) {
int sU = sss;
int sF = s + Ls * sU;
assert ( sF< B._grid->oSites());
assert ( sU< U._grid->oSites());
assert(sF < B._grid->oSites());
assert(sU < U._grid->oSites());
Kernels::DiracOptDhopDir(st,U,st.comm_buf,sF,sU,B,Btilde,mu,gamma);
////////////////////////////
// spin trace outer product
////////////////////////////
Kernels::DiracOptDhopDir(st, U, st.comm_buf, sF, sU, B, Btilde, mu,
gamma);
////////////////////////////
// spin trace outer product
////////////////////////////
}
}
Impl::InsertForce5D(mat,Btilde,Atilde,mu);
DerivDhopComputeTime += usecond();
Impl::InsertForce5D(mat, Btilde, Atilde, mu);
}
DerivComputeTime += usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopDeriv( GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
const FermionField &A,
const FermionField &B,
int dag)
{
conformable(A._grid,FermionGrid());
conformable(A._grid,B._grid);
@ -288,9 +358,9 @@ void WilsonFermion5D<Impl>::DhopDeriv( GaugeField &mat,
template<class Impl>
void WilsonFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
const FermionField &A,
const FermionField &B,
int dag)
{
conformable(A._grid,FermionRedBlackGrid());
conformable(GaugeRedBlackGrid(),mat._grid);
@ -306,9 +376,9 @@ void WilsonFermion5D<Impl>::DhopDerivEO(GaugeField &mat,
template<class Impl>
void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
const FermionField &A,
const FermionField &B,
int dag)
const FermionField &A,
const FermionField &B,
int dag)
{
conformable(A._grid,FermionRedBlackGrid());
conformable(GaugeRedBlackGrid(),mat._grid);
@ -323,32 +393,61 @@ void WilsonFermion5D<Impl>::DhopDerivOE(GaugeField &mat,
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
DoubledGaugeField & U,
const FermionField &in, FermionField &out,int dag)
{
DhopCalls++;
// assert((dag==DaggerNo) ||(dag==DaggerYes));
Compressor compressor(dag);
int LLs = in._grid->_rdimensions[0];
DhopCommTime-=usecond();
st.HaloExchange(in,compressor);
DhopCommTime+=usecond();
DhopComputeTime-=usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
if ( dag == DaggerYes ) {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSiteDag(st,lo,U,st.comm_buf,sF,sU,LLs,1,in,out);
if (dag == DaggerYes) {
PARALLEL_FOR_LOOP
for (int ss = 0; ss < U._grid->oSites(); ss++) {
int sU = ss;
int sF = LLs * sU;
Kernels::DiracOptDhopSiteDag(st, lo, U, st.comm_buf, sF, sU, LLs, 1, in,
out);
}
#ifdef AVX512
} else if (stat.is_init() ) {
int nthreads;
stat.start();
#pragma omp parallel
{
#pragma omp master
nthreads = omp_get_num_threads();
int mythread = omp_get_thread_num();
stat.enter(mythread);
#pragma omp for nowait
for(int ss=0;ss<U._grid->oSites();ss++)
{
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSite(st,lo,U,st.comm_buf,sF,sU,LLs,1,in,out);
}
stat.exit(mythread);
}
stat.accum(nthreads);
#endif
} else {
PARALLEL_FOR_LOOP
for(int ss=0;ss<U._grid->oSites();ss++){
int sU=ss;
int sF=LLs*sU;
Kernels::DiracOptDhopSite(st,lo,U,st.comm_buf,sF,sU,LLs,1,in,out);
PARALLEL_FOR_LOOP
for (int ss = 0; ss < U._grid->oSites(); ss++) {
int sU = ss;
int sF = LLs * sU;
Kernels::DiracOptDhopSite(st, lo, U, st.comm_buf, sF, sU, LLs, 1, in,
out);
}
}
DhopComputeTime+=usecond();
}

14
lib/qcd/action/fermion/WilsonFermion5D.h
View File

@ -31,6 +31,8 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#ifndef GRID_QCD_WILSON_FERMION_5D_H
#define GRID_QCD_WILSON_FERMION_5D_H
#include <Grid/Stat.h>
namespace Grid {
namespace QCD {
@ -60,6 +62,18 @@ namespace Grid {
public:
INHERIT_IMPL_TYPES(Impl);
typedef WilsonKernels<Impl> Kernels;
PmuStat stat;
void Report(void);
void ZeroCounters(void);
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DerivCalls;
double DerivCommTime;
double DerivComputeTime;
double DerivDhopComputeTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base

23
lib/qcd/action/fermion/WilsonKernels.h
View File

@ -196,24 +196,11 @@ namespace Grid {
WilsonKernels(const ImplParams &p = ImplParams());
};
///////////////////////////////////////////////////////////
// Default to no assembler implementation
///////////////////////////////////////////////////////////
template<class Impl>
void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
{
assert(0);
}
template<class Impl>
void WilsonKernels<Impl >::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
{
assert(0);
}
}
}
#endif

42
lib/qcd/action/fermion/WilsonKernelsAsm.cc
View File

@ -31,9 +31,30 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#include <Grid.h>
namespace Grid {
namespace QCD {
///////////////////////////////////////////////////////////
// Default to no assembler implementation
///////////////////////////////////////////////////////////
template<class Impl>
void WilsonKernels<Impl >::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
{
assert(0);
}
template<class Impl>
void WilsonKernels<Impl >::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out)
{
assert(0);
}
#if defined(AVX512)
@ -102,6 +123,27 @@ namespace Grid {
#include <qcd/action/fermion/WilsonKernelsAsmBody.h>
#endif
#define INSTANTIATE_ASM(A)\
template void WilsonKernels<A>::DiracOptAsmDhopSite(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,\
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,\
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out);\
template void WilsonKernels<A>::DiracOptAsmDhopSiteDag(StencilImpl &st,LebesgueOrder & lo,DoubledGaugeField &U,\
std::vector<SiteHalfSpinor,alignedAllocator<SiteHalfSpinor> > &buf,\
int ss,int ssU,int Ls,int Ns,const FermionField &in, FermionField &out);\
INSTANTIATE_ASM(WilsonImplF);
INSTANTIATE_ASM(WilsonImplD);
INSTANTIATE_ASM(ZWilsonImplF);
INSTANTIATE_ASM(ZWilsonImplD);
INSTANTIATE_ASM(GparityWilsonImplF);
INSTANTIATE_ASM(GparityWilsonImplD);
INSTANTIATE_ASM(DomainWallVec5dImplF);
INSTANTIATE_ASM(DomainWallVec5dImplD);
INSTANTIATE_ASM(ZDomainWallVec5dImplF);
INSTANTIATE_ASM(ZDomainWallVec5dImplD);
}
}

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

@ -134,7 +134,9 @@
////////////////////////////////
// Xm
////////////////////////////////
#ifndef STREAM_STORE
basep= (uint64_t) &out._odata[ss];
#endif
// basep= st.GetPFInfo(nent,plocal); nent++;
if ( local ) {
LOAD64(%r10,isigns); // times i => shuffle and xor the real part sign bit
@ -229,7 +231,9 @@
LOAD_CHI(base);
}
base= (uint64_t) &out._odata[ss];
#ifndef STREAM_STORE
PREFETCH_CHIMU(base);
#endif
{
MULT_2SPIN_DIR_PFTM(Tm,basep);
}

6
lib/qcd/action/pseudofermion/TwoFlavourEvenOddRatio.h
View File

@ -131,9 +131,11 @@ namespace Grid{
Vpc.MpcDag(PhiOdd,Y); // Y= Vdag phi
X=zero;
ActionSolver(Mpc,Y,X); // X= (MdagM)^-1 Vdag phi
Mpc.Mpc(X,Y); // Y= Mdag^-1 Vdag phi
//Mpc.Mpc(X,Y); // Y= Mdag^-1 Vdag phi
// Multiply by Ydag
RealD action = real(innerProduct(Y,X));
RealD action = norm2(Y);
//RealD action = norm2(Y);
// The EE factorised block; normally can replace with zero if det is constant (gauge field indept)
// Only really clover term that creates this. Leave the EE portion as a future to do to make most