portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-11 03:46:55 +01:00
Code Releases Activity

Build a list of whats on the surface

Browse Source
This commit is contained in:
paboyle
2017-04-24 17:06:15 +01:00
parent 916e9e1d3e
commit 56277a11c8
5 changed files with 101 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -242,6 +242,7 @@ public:
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
std::vector<int> surface_list;
WilsonStencil(GridBase *grid,
int npoints,
@ -249,11 +250,33 @@ public:
const std::vector<int> &directions,
const std::vector<int> &distances)
: CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
same_node(npoints)
same_node(npoints)
{
assert(npoints==8);// or 10 if do naive DWF 5d red black ?
surface_list.resize(0);
};
void BuildSurfaceList(int Ls,int vol4){
// find same node for SHM
// Here we know the distance is 1 for WilsonStencil
for(int point=0;point<this->_npoints;point++){
same_node[point] = this->SameNode(point);
std::cout << " dir " <<point<<" same_node " <<same_node[point]<<std::endl;
}
for(int site = 0 ;site< vol4;site++){
int local = 1;
for(int point=0;point<this->_npoints;point++){
if( (!this->GetNodeLocal(site*Ls,point)) && (!same_node[point]) ){
local = 0;
}
}
if(local == 0) {
surface_list.push_back(site);
}
}
}
template < class compressor>
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
{

36
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -117,6 +117,19 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
// Allocate the required comms buffer
ImportGauge(_Umu);
// Build lists of exterior only nodes
int LLs = FourDimGrid._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>
@ -406,6 +419,8 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
// 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.
#if 1
#if 0
#pragma omp parallel
{
@ -422,6 +437,27 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
else Kernels::DhopSite (st,lo,U,st.CommBuf(),sF,myoff,LLs,mywork,in,out,0,1);
if ( me==0 ) DhopComputeTime2+=usecond();
}// end parallel region
#else
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {
#pragma omp parallel for schedule(static,1)
for (int ss = 0; ss < st.surface_list.size(); 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 {
#pragma omp parallel for schedule(static,1)
for (int ss = 0; ss < st.surface_list.size(); 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();
#endif
#else
DhopComputeTime2-=usecond();
if (dag == DaggerYes) {