mirror of
https://github.com/paboyle/Grid.git
synced 2025-06-17 15:27:06 +01:00
Merge branch 'feature/fft-opt' into feature/feynman-rules
# Conflicts: # lib/FFT.h # lib/qcd/action/fermion/WilsonFermion5D.h # tests/core/Test_fft.cc
This commit is contained in:
@ -184,44 +184,37 @@ void WilsonFermion5D<Impl>::Report(void)
|
||||
|
||||
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;
|
||||
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;
|
||||
RealD mflops = 1344*volume*DhopCalls/DhopComputeTime/2; // 2 for red black counting
|
||||
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;
|
||||
std::cout << GridLogMessage << "Average mflops/s per call per rank : " << 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;
|
||||
|
||||
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();
|
||||
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();
|
||||
}
|
||||
}
|
||||
|
||||
@ -275,7 +268,7 @@ PARALLEL_FOR_LOOP
|
||||
for(int s=0;s<Ls;s++){
|
||||
int sU=ss;
|
||||
int sF = s+Ls*sU;
|
||||
Kernels::DiracOptDhopDir(Stencil,Umu,Stencil.comm_buf,sF,sU,in,out,dirdisp,gamma);
|
||||
Kernels::DiracOptDhopDir(Stencil,Umu,Stencil.CommBuf(),sF,sU,in,out,dirdisp,gamma);
|
||||
}
|
||||
}
|
||||
};
|
||||
@ -327,8 +320,7 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
|
||||
assert(sF < B._grid->oSites());
|
||||
assert(sU < U._grid->oSites());
|
||||
|
||||
Kernels::DiracOptDhopDir(st, U, st.comm_buf, sF, sU, B, Btilde, mu,
|
||||
gamma);
|
||||
Kernels::DiracOptDhopDir(st, U, st.CommBuf(), sF, sU, B, Btilde, mu, gamma);
|
||||
|
||||
////////////////////////////
|
||||
// spin trace outer product
|
||||
@ -342,10 +334,10 @@ void WilsonFermion5D<Impl>::DerivInternal(StencilImpl & st,
|
||||
}
|
||||
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::DhopDeriv( GaugeField &mat,
|
||||
const FermionField &A,
|
||||
const FermionField &B,
|
||||
int dag)
|
||||
void WilsonFermion5D<Impl>::DhopDeriv(GaugeField &mat,
|
||||
const FermionField &A,
|
||||
const FermionField &B,
|
||||
int dag)
|
||||
{
|
||||
conformable(A._grid,FermionGrid());
|
||||
conformable(A._grid,B._grid);
|
||||
@ -358,9 +350,9 @@ 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);
|
||||
@ -376,9 +368,9 @@ 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);
|
||||
@ -393,10 +385,9 @@ 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);
|
||||
|
||||
@ -413,27 +404,25 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
|
||||
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);
|
||||
Kernels::DiracOptDhopSiteDag(st, lo, U, st.CommBuf(), sF, sU, LLs, 1, in, out);
|
||||
}
|
||||
#ifdef AVX512
|
||||
} else if (stat.is_init() ) {
|
||||
|
||||
int nthreads;
|
||||
stat.start();
|
||||
#pragma omp parallel
|
||||
#pragma omp parallel
|
||||
{
|
||||
#pragma omp master
|
||||
#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);
|
||||
}
|
||||
#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.CommBuf(),sF,sU,LLs,1,in,out);
|
||||
}
|
||||
stat.exit(mythread);
|
||||
}
|
||||
stat.accum(nthreads);
|
||||
@ -443,8 +432,7 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
|
||||
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);
|
||||
Kernels::DiracOptDhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out);
|
||||
}
|
||||
}
|
||||
DhopComputeTime+=usecond();
|
||||
@ -454,6 +442,7 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int dag)
|
||||
{
|
||||
DhopCalls++;
|
||||
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
|
||||
conformable(in._grid,out._grid); // drops the cb check
|
||||
|
||||
@ -465,6 +454,7 @@ void WilsonFermion5D<Impl>::DhopOE(const FermionField &in, FermionField &out,int
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int dag)
|
||||
{
|
||||
DhopCalls++;
|
||||
conformable(in._grid,FermionRedBlackGrid()); // verifies half grid
|
||||
conformable(in._grid,out._grid); // drops the cb check
|
||||
|
||||
@ -476,6 +466,7 @@ void WilsonFermion5D<Impl>::DhopEO(const FermionField &in, FermionField &out,int
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::Dhop(const FermionField &in, FermionField &out,int dag)
|
||||
{
|
||||
DhopCalls+=2;
|
||||
conformable(in._grid,FermionGrid()); // verifies full grid
|
||||
conformable(in._grid,out._grid);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user