portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-26 05:35:55 +01:00
Code Releases Activity

Provide info if EE term is trivial (m^2 factor)

Browse Source 
Better timing in staggered 4d case
This commit is contained in:
Azusa Yamaguchi 2018-04-26 10:10:07 +01:00
parent 31d83ee046
commit 2cbb72a81c
5 changed files with 101 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

75
lib/qcd/action/fermion/ImprovedStaggeredFermion.cc
View File

@ -330,6 +330,7 @@ void ImprovedStaggeredFermion<Impl>::DhopDerivEO(GaugeField &mat, const FermionF
template <class Impl>
void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &out, int dag) {
DhopCalls+=2;
conformable(in._grid, _grid); // verifies full grid
conformable(in._grid, out._grid);
@ -340,6 +341,7 @@ void ImprovedStaggeredFermion<Impl>::Dhop(const FermionField &in, FermionField &
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField &out, int dag) {
DhopCalls+=1;
conformable(in._grid, _cbgrid); // verifies half grid
conformable(in._grid, out._grid); // drops the cb check
@ -351,6 +353,7 @@ void ImprovedStaggeredFermion<Impl>::DhopOE(const FermionField &in, FermionField
template <class Impl>
void ImprovedStaggeredFermion<Impl>::DhopEO(const FermionField &in, FermionField &out, int dag) {
DhopCalls+=1;
conformable(in._grid, _cbgrid); // verifies half grid
conformable(in._grid, out._grid); // drops the cb check
@ -403,13 +406,18 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
int len = U._grid->oSites();
const int LLs = 1;
DhopTotalTime -= usecond();
DhopFaceTime -= usecond();
st.Prepare();
st.HaloGather(in,compressor);
st.CommsMergeSHM(compressor);
DhopFaceTime += usecond();
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
DhopComputeTime -= usecond();
#pragma omp parallel
{
int tid = omp_get_thread_num();
@ -451,10 +459,14 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
st.CommunicateThreaded();
}
}
DhopComputeTime += usecond();
// First to enter, last to leave timing
DhopFaceTime -= usecond();
st.CommsMerge(compressor);
DhopFaceTime -= usecond();
DhopComputeTime2 -= usecond();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
parallel_for (int ss = 0; ss < sz; ss++) {
@ -468,6 +480,7 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalOverlappedComms(StencilImpl &st
Kernels::DhopSite(st,lo,U,UUU,st.CommBuf(),1,sU,in,out,0,1);
}
}
DhopComputeTime2 += usecond();
#else
assert(0);
#endif
@ -483,9 +496,14 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
{
assert((dag == DaggerNo) || (dag == DaggerYes));
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
Compressor compressor;
st.HaloExchange(in, compressor);
DhopCommTime += usecond();
DhopComputeTime -= usecond();
if (dag == DaggerYes) {
parallel_for (int sss = 0; sss < in._grid->oSites(); sss++) {
Kernels::DhopSiteDag(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
@ -495,8 +513,65 @@ void ImprovedStaggeredFermion<Impl>::DhopInternalSerialComms(StencilImpl &st, Le
Kernels::DhopSite(st, lo, U, UUU, st.CommBuf(), 1, sss, in, out);
}
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
};
////////////////////////////////////////////////////////////////
// Reporting
////////////////////////////////////////////////////////////////
template<class Impl>
void ImprovedStaggeredFermion<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = 1; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _grid->_Nprocessors;
RealD NN = _grid->NodeCount();
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Number of DhopEO Calls : "
<< DhopCalls << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion TotalTime /Calls : "
<< DhopTotalTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion CommTime /Calls : "
<< DhopCommTime / DhopCalls << " us" << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion ComputeTime/Calls : "
<< DhopComputeTime / DhopCalls << " us" << std::endl;
// Average the compute time
_grid->GlobalSum(DhopComputeTime);
DhopComputeTime/=NP;
RealD mflops = 1154*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 rank : " << mflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node : " << mflops/NN << std::endl;
RealD Fullmflops = 1154*volume*DhopCalls/(DhopTotalTime)/2; // 2 for red black counting
std::cout << GridLogMessage << "Average mflops/s per call (full) : " << Fullmflops << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per rank (full): " << Fullmflops/NP << std::endl;
std::cout << GridLogMessage << "Average mflops/s per call per node (full): " << Fullmflops/NN << std::endl;
std::cout << GridLogMessage << "ImprovedStaggeredFermion Stencil" <<std::endl; Stencil.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "ImprovedStaggeredFermion StencilOdd" <<std::endl; StencilOdd.Report();
}
template<class Impl>
void ImprovedStaggeredFermion<Impl>::ZeroCounters(void)
{
DhopCalls = 0;
DhopTotalTime = 0;
DhopCommTime = 0;
DhopComputeTime = 0;
DhopFaceTime = 0;
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
}
FermOpStaggeredTemplateInstantiate(ImprovedStaggeredFermion);
//AdjointFermOpTemplateInstantiate(ImprovedStaggeredFermion);

15
lib/qcd/action/fermion/ImprovedStaggeredFermion.h
View File

@ -49,6 +49,18 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
FermionField _tmp;
FermionField &tmp(void) { return _tmp; }
////////////////////////////////////////
// Performance monitoring
////////////////////////////////////////
void Report(void);
void ZeroCounters(void);
double DhopTotalTime;
double DhopCalls;
double DhopCommTime;
double DhopComputeTime;
double DhopComputeTime2;
double DhopFaceTime;
///////////////////////////////////////////////////////////////
// Implement the abstract base
///////////////////////////////////////////////////////////////
@ -142,7 +154,8 @@ class ImprovedStaggeredFermion : public StaggeredKernels<Impl>, public ImprovedS
// protected:
public:
// any other parameters of action ???
virtual int isTrivialEE(void) { return 1; };
virtual RealD Mass(void) { return mass; }
RealD mass;
RealD u0;
RealD c1;

9
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.cc
View File

@ -291,14 +291,12 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOr
DoubledGaugeField & U,DoubledGaugeField & UUU,
const FermionField &in, FermionField &out,int dag)
{
DhopTotalTime-=usecond();
#ifdef GRID_OMP
if ( StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute )
DhopInternalOverlappedComms(st,lo,U,UUU,in,out,dag);
else
#endif
DhopInternalSerialComms(st,lo,U,UUU,in,out,dag);
DhopTotalTime+=usecond();
}
template<class Impl>
@ -412,6 +410,7 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
//double t1=usecond();
DhopTotalTime -= usecond();
DhopCommTime -= usecond();
st.HaloExchange(in,compressor);
@ -432,6 +431,12 @@ void ImprovedStaggeredFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st,
}
DhopComputeTime += usecond();
DhopTotalTime += usecond();
//double t2=usecond();
//std::cout << __FILE__ << " " << __func__ << " Total Time " << DhopTotalTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Total Time Org " << t2-t1 << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Comml Time " << DhopCommTime << std::endl;
//std::cout << __FILE__ << " " << __func__ << " Compute Time " << DhopComputeTime << std::endl;
}
/*CHANGE END*/

3
lib/qcd/action/fermion/ImprovedStaggeredFermion5D.h
View File

@ -178,6 +178,9 @@ namespace QCD {
// Data members require to support the functionality
///////////////////////////////////////////////////////////////
public:
virtual int isTrivialEE(void) { return 1; };
virtual RealD Mass(void) { return mass; }
GridBase *_FourDimGrid;
GridBase *_FourDimRedBlackGrid;

2
lib/qcd/action/fermion/WilsonFermion.h
View File

@ -135,6 +135,8 @@ class WilsonFermion : public WilsonKernels<Impl>, public WilsonFermionStatic {
// protected:
public:
virtual RealD Mass(void) { return mass; }
virtual int isTrivialEE(void) { return 1; };
RealD mass;
GridBase *_grid;