portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-04 05:54:32 +00:00
Code Releases Activity

Merge branch 'develop' into feature/hadrons

Browse Source
This commit is contained in:
Antonin Portelli
2017-08-29 11:03:53 -06:00
parent 5846566728 4b4c2a715b
commit a56e3b40c4
20 changed files with 1174 additions and 184 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
lib/qcd/action/fermion/CayleyFermion5D.cc
View File

@@ -414,7 +414,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0; i < Ls; i++){
as[i] = 1.0;
omega[i] = gamma[i]*zolo_hi; //NB reciprocal relative to Chroma NEF code
// assert(fabs(omega[i])>0.0);
assert(omega[i]!=Coeff_t(0.0));
bs[i] = 0.5*(bpc/omega[i] + bmc);
cs[i] = 0.5*(bpc/omega[i] - bmc);
}
@@ -429,7 +429,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
for(int i=0;i<Ls;i++){
bee[i]=as[i]*(bs[i]*(4.0-this->M5) +1.0);
// assert(fabs(bee[i])>0.0);
assert(bee[i]!=Coeff_t(0.0));
cee[i]=as[i]*(1.0-cs[i]*(4.0-this->M5));
beo[i]=as[i]*bs[i];
ceo[i]=-as[i]*cs[i];
@@ -455,11 +455,17 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
dee[i] = bee[i];
if ( i < Ls-1 ) {
assert(bee[i]!=Coeff_t(0.0));
assert(bee[0]!=Coeff_t(0.0));
lee[i] =-cee[i+1]/bee[i]; // sub-diag entry on the ith column
leem[i]=mass*cee[Ls-1]/bee[0];
for(int j=0;j<i;j++) leem[i]*= aee[j]/bee[j+1];
for(int j=0;j<i;j++) {
assert(bee[j+1]!=Coeff_t(0.0));
leem[i]*= aee[j]/bee[j+1];
}
uee[i] =-aee[i]/bee[i]; // up-diag entry on the ith row
@@ -478,7 +484,7 @@ void CayleyFermion5D<Impl>::SetCoefficientsInternal(RealD zolo_hi,std::vector<Co
{
Coeff_t delta_d=mass*cee[Ls-1];
for(int j=0;j<Ls-1;j++) {
// assert(fabs(bee[j])>0.0);
assert(bee[j] != Coeff_t(0.0));
delta_d *= cee[j]/bee[j];
}
dee[Ls-1] += delta_d;

48
lib/qcd/action/fermion/WilsonCompressor.h
View File

@@ -238,7 +238,33 @@ template<typename HCS,typename HS,typename S> using WilsonCompressor = WilsonCom
template<class vobj,class cobj>
class WilsonStencil : public CartesianStencil<vobj,cobj> {
public:
double timer0;
double timer1;
double timer2;
double timer3;
double timer4;
double timer5;
double timer6;
uint64_t callsi;
void ZeroCountersi(void)
{
timer0=0;
timer1=0;
timer2=0;
timer3=0;
timer4=0;
timer5=0;
timer6=0;
callsi=0;
}
void Reporti(int calls)
{
if ( timer0 ) std::cout << GridLogMessage << " timer0 (HaloGatherOpt) " <<timer0/calls <<std::endl;
if ( timer1 ) std::cout << GridLogMessage << " timer1 (Communicate) " <<timer1/calls <<std::endl;
if ( timer2 ) std::cout << GridLogMessage << " timer2 (CommsMerge ) " <<timer2/calls <<std::endl;
if ( timer3 ) std::cout << GridLogMessage << " timer3 (commsMergeShm) " <<timer3/calls <<std::endl;
if ( timer4 ) std::cout << GridLogMessage << " timer4 " <<timer4 <<std::endl;
}
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
std::vector<int> same_node;
@@ -252,6 +278,7 @@ public:
: CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances) ,
same_node(npoints)
{
ZeroCountersi();
surface_list.resize(0);
};
@@ -261,7 +288,6 @@ public:
// 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++){
@@ -282,17 +308,28 @@ public:
{
std::vector<std::vector<CommsRequest_t> > reqs;
this->HaloExchangeOptGather(source,compress);
this->CommunicateBegin(reqs);
this->CommunicateComplete(reqs);
double t1=usecond();
// Asynchronous MPI calls multidirectional, Isend etc...
// this->CommunicateBegin(reqs);
// this->CommunicateComplete(reqs);
// Non-overlapped directions within a thread. Asynchronous calls except MPI3, threaded up to comm threads ways.
this->Communicate();
double t2=usecond(); timer1 += t2-t1;
this->CommsMerge(compress);
double t3=usecond(); timer2 += t3-t2;
this->CommsMergeSHM(compress);
double t4=usecond(); timer3 += t4-t3;
}
template <class compressor>
void HaloExchangeOptGather(const Lattice<vobj> &source,compressor &compress)
{
this->Prepare();
double t0=usecond();
this->HaloGatherOpt(source,compress);
double t1=usecond();
timer0 += t1-t0;
callsi++;
}
template <class compressor>
@@ -304,7 +341,9 @@ public:
typedef typename compressor::SiteHalfSpinor SiteHalfSpinor;
typedef typename compressor::SiteHalfCommSpinor SiteHalfCommSpinor;
this->mpi3synctime_g-=usecond();
this->_grid->StencilBarrier();
this->mpi3synctime_g+=usecond();
assert(source._grid==this->_grid);
this->halogtime-=usecond();
@@ -323,7 +362,6 @@ public:
int dag = compress.dag;
int face_idx=0;
if ( dag ) {
// std::cout << " Optimised Dagger compress " <<std::endl;
assert(same_node[Xp]==this->HaloGatherDir(source,XpCompress,Xp,face_idx));
assert(same_node[Yp]==this->HaloGatherDir(source,YpCompress,Yp,face_idx));
assert(same_node[Zp]==this->HaloGatherDir(source,ZpCompress,Zp,face_idx));

107
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@@ -124,22 +124,24 @@ WilsonFermion5D<Impl>::WilsonFermion5D(GaugeField &_Umu,
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;
// std::cout << GridLogMessage << " SurfaceLists "<< Stencil.surface_list.size()
// <<" " << StencilEven.surface_list.size()<<std::endl;
}
template<class Impl>
void WilsonFermion5D<Impl>::Report(void)
{
std::vector<int> latt = GridDefaultLatt();
RealD volume = Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD NP = _FourDimGrid->_Nprocessors;
RealD NN = _FourDimGrid->NodeCount();
RealD volume = Ls;
std::vector<int> latt = _FourDimGrid->GlobalDimensions();
for(int mu=0;mu<Nd;mu++) volume=volume*latt[mu];
if ( DhopCalls > 0 ) {
std::cout << GridLogMessage << "#### Dhop calls report " << std::endl;
@@ -185,6 +187,11 @@ void WilsonFermion5D<Impl>::Report(void)
std::cout << GridLogMessage << "WilsonFermion5D StencilEven"<<std::endl; StencilEven.Report();
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd" <<std::endl; StencilOdd.Report();
}
if ( DhopCalls > 0){
std::cout << GridLogMessage << "WilsonFermion5D Stencil Reporti()" <<std::endl; Stencil.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilEven Reporti()"<<std::endl; StencilEven.Reporti(DhopCalls);
std::cout << GridLogMessage << "WilsonFermion5D StencilOdd Reporti()" <<std::endl; StencilOdd.Reporti(DhopCalls);
}
}
template<class Impl>
@@ -204,6 +211,9 @@ void WilsonFermion5D<Impl>::ZeroCounters(void) {
Stencil.ZeroCounters();
StencilEven.ZeroCounters();
StencilOdd.ZeroCounters();
Stencil.ZeroCountersi();
StencilEven.ZeroCountersi();
StencilOdd.ZeroCountersi();
}
@@ -380,7 +390,6 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
{
#ifdef GRID_OMP
// assert((dag==DaggerNo) ||(dag==DaggerYes));
typedef CartesianCommunicator::CommsRequest_t CommsRequest_t;
Compressor compressor(dag);
@@ -389,46 +398,70 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
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);
st.CommsMergeSHM(compressor);// Could do this inside parallel region overlapped with comms
DhopFaceTime+=usecond();
// Perhaps use omp task and region
#pragma omp parallel
double ctime=0;
double ptime=0;
//////////////////////////////////////////////////////////////////////////////////////////////////////
// Ugly explicit thread mapping introduced for OPA reasons.
//////////////////////////////////////////////////////////////////////////////////////////////////////
#pragma omp parallel reduction(max:ctime) reduction(max:ptime)
{
int tid = omp_get_thread_num();
int nthreads = omp_get_num_threads();
int me = omp_get_thread_num();
int myoff, mywork;
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();
int ncomms = CartesianCommunicator::nCommThreads;
if (ncomms == -1) ncomms = 1;
assert(nthreads > ncomms);
if (tid >= ncomms) {
double start = usecond();
nthreads -= ncomms;
int ttid = tid - ncomms;
int n = U._grid->oSites();
int chunk = n / nthreads;
int rem = n % nthreads;
int myblock, myn;
if (ttid < rem) {
myblock = ttid * chunk + ttid;
myn = chunk+1;
} else {
myblock = ttid*chunk + rem;
myn = chunk;
}
// do the compute
if (dag == DaggerYes) {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSiteDag(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
} else {
for (int ss = myblock; ss < myblock+myn; ++ss) {
int sU = ss;
int sF = LLs * sU;
Kernels::DhopSite(st,lo,U,st.CommBuf(),sF,sU,LLs,1,in,out,1,0);
}
}
ptime = usecond() - start;
}
{
double start = usecond();
st.CommunicateThreaded();
ctime = usecond() - start;
}
}
DhopCommTime += ctime;
DhopComputeTime+=ptime;
// First to enter, last to leave timing
st.CollateThreads();
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();
@@ -449,11 +482,9 @@ void WilsonFermion5D<Impl>::DhopInternalOverlappedComms(StencilImpl & st, Lebesg
#else
assert(0);
#endif
}
template<class Impl>
void WilsonFermion5D<Impl>::DhopInternalSerialComms(StencilImpl & st, LebesgueOrder &lo,
DoubledGaugeField & U,