portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 04:37:05 +01:00
Code Releases Activity

Experimental: Multiple communicators to see if we can avoid thread locks in --enable-comms=mpit

Browse Source
This commit is contained in:
paboyle
2017-06-24 23:10:24 +01:00
parent 869b99ec1e
commit 54e94360ad
13 changed files with 139 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -379,7 +379,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
Compressor compressor(dag);
@ -388,46 +387,46 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
std::vector<std::vector<CommsRequest_t> > reqs;
// Rely on async comms; start comms before merge of local data
DhopComputeTime-=usecond();
DhopCommTime-=usecond();
st.CommunicateBegin(reqs);
DhopFaceTime-=usecond();
st.CommsMergeSHM(compressor);
DhopFaceTime+=usecond();
// Perhaps use omp task and region
#pragma omp parallel
{
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int myoff, mywork;
// Should time this somehow; hard as the threads fork nowait
st.CommunicateThreaded();
GridThread::GetWork(len,me-1,mywork,myoff,nthreads-1);
int sF = LLs * myoff;
if ( me == 0 ) {
st.CommunicateComplete(reqs);
DhopCommTime+=usecond();
} else {
// Interior links in stencil
if ( me==1 ) DhopComputeTime-=usecond();
if (dag == DaggerYes) Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
else Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,1,0);
if ( me==1 ) DhopComputeTime+=usecond();
if (dag == DaggerYes) {
#pragma omp for
for (int ss = 0; ss < U._grid->oSites(); ss++) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
} else {
#pragma omp for
for (int ss = 0; ss < U._grid->oSites(); ss++) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
}
#pragma omp single
DhopComputeTime+=usecond();
#pragma omp taskwait
#pragma omp single
DhopCommTime+=usecond();
} // Closes parallel region and waits the comms (I hope)
DhopFaceTime-=usecond();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
// Load imbalance alert. Should use dynamic schedule OMP for loop
// Perhaps create a list of only those sites with face work, and
// load balance process the list.
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
int sz=st.surface_list.size();
@ -448,11 +447,9 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,