portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 00:07:05 +01:00
Code Releases Activity

Merge branch 'develop' of https://github.com/paboyle/Grid into feature/rare_kaon

Browse Source 
# Conflicts:
#	lib/qcd/action/fermion/WilsonFermion5D.cc
#	tests/hadrons/Test_hadrons_rarekaon.cc
This commit is contained in:
Lanny91
2017-05-04 16:22:33 +01:00
parent 49331a3e72 679ae98b14
commit ca1077c560
109 changed files with 4537 additions and 3328 deletions
Download Patch File Download Diff File Expand all files Collapse all files

110
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -117,49 +117,20 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
// Allocate the required comms buffer
ImportGauge(_Umu);
// Build lists of exterior only nodes
int LLs = FiveDimGrid._rdimensions[0];
int vol4;
vol4=FourDimGrid.oSites();
Stencil.BuildSurfaceList(LLs,vol4);
vol4=FourDimRedBlackGrid.oSites();
StencilEven.BuildSurfaceList(LLs,vol4);
StencilOdd.BuildSurfaceList(LLs,vol4);
std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
<<" " << StencilEven.surface_list.size()<<std::endl;
}
/*
template<class Impl>
WilsonFermion5D<Impl>::WilsonFermion5D(int simd,GaugeField &_Umu,
GridCartesian &FiveDimGrid,
GridRedBlackCartesian &FiveDimRedBlackGrid,
GridCartesian &FourDimGrid,
RealD _M5,const ImplParams &p) :
{
int nsimd = Simd::Nsimd();
// some assertions
assert(FiveDimGrid._ndimension==5);
assert(FiveDimRedBlackGrid._ndimension==5);
assert(FiveDimRedBlackGrid._checker_dim==0); // Checkerboard the s-direction
assert(FourDimGrid._ndimension==4);
// Dimension zero of the five-d is the Ls direction
Ls=FiveDimGrid._fdimensions[0];
assert(FiveDimGrid._processors[0] ==1);
assert(FiveDimGrid._simd_layout[0] ==nsimd);
assert(FiveDimRedBlackGrid._fdimensions[0]==Ls);
assert(FiveDimRedBlackGrid._processors[0] ==1);
assert(FiveDimRedBlackGrid._simd_layout[0]==nsimd);
// Other dimensions must match the decomposition of the four-D fields
for(int d=0;d<4;d++){
assert(FiveDimRedBlackGrid._fdimensions[d+1]==FourDimGrid._fdimensions[d]);
assert(FiveDimRedBlackGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FourDimGrid._simd_layout[d]=1);
assert(FiveDimRedBlackGrid._simd_layout[d+1]==1);
assert(FiveDimGrid._fdimensions[d+1] ==FourDimGrid._fdimensions[d]);
assert(FiveDimGrid._processors[d+1] ==FourDimGrid._processors[d]);
assert(FiveDimGrid._simd_layout[d+1] ==FourDimGrid._simd_layout[d]);
}
{
}
}
*/
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
@ -396,6 +367,7 @@ void WilsonFermion5D<Impl>::DhopInternal(StencilImpl & st, LebesgueOrder &lo,
DhopTotalTime+=usecond();
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,
@ -409,12 +381,21 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
int LLs = in._grid->_rdimensions[0];
int len = U._grid->oSites();
DhopFaceTime-=usecond();
st.HaloExchangeOptGather(in,compressor);
DhopFaceTime+=usecond();
std::vector<std::vector<CommsRequest_t> > reqs;
// Rely on async comms; start comms before merge of local data
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();
@ -425,8 +406,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
int sF = LLs * myoff;
if ( me == 0 ) {
DhopCommTime-=usecond();
st.CommunicateBegin(reqs);
st.CommunicateComplete(reqs);
DhopCommTime+=usecond();
} else {
@ -439,28 +418,37 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
}
DhopFaceTime-=usecond();
st.CommsMerge();
st.CommsMerge(compressor);
DhopFaceTime+=usecond();
#pragma omp parallel
{
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int myoff, mywork;
GridThread::GetWork(len,me,mywork,myoff,nthreads);
int sF = LLs * myoff;
// Exterior links in stencil
if ( me==0 ) DhopComputeTime2-=usecond();
if (dag == DaggerYes) Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,0,1);
else Kernels::DhopSite (st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,0,1);
if ( me==0 ) DhopComputeTime2+=usecond();
}// end parallel region
// 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();
parallel_for (int ss = 0; ss < sz; ss++) {
int sU = st.surface_list[ss];
int sF = LLs * sU;
Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,0,1);
}
} else {
int sz=st.surface_list.size();
parallel_for (int ss = 0; ss < sz; ss++) {
int sU = st.surface_list[ss];
int sF = LLs * sU;
Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,0,1);
}
}
DhopComputeTime2+=usecond();
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,