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

Benchmark prep

Browse Source
This commit is contained in:
Peter Boyle
2017-08-25 09:25:54 +01:00
parent b49bec0cec
commit c3b1263e75
10 changed files with 494 additions and 71 deletions
Download Patch File Download Diff File Expand all files Collapse all files

322
benchmarks/Benchmark_ITT.cc
View File

@@ -32,6 +32,19 @@ using namespace std;
using namespace Grid;
using namespace Grid::QCD;
typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
std::vector<int> L_list;
std::vector<int> Ls_list;
std::vector<double> mflop_list;
double mflop_ref;
double mflop_ref_err;
int NN_global;
struct time_statistics{
double mean;
@@ -95,13 +108,15 @@ public:
static void Comms(void)
{
int Nloop=100;
int Nloop=1000;
int nmu=0;
int maxlat=32;
std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
std::vector<int> mpi_layout = GridDefaultMpi();
for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
std::vector<double> t_time(Nloop);
time_statistics timestat;
@@ -133,13 +148,14 @@ public:
bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
}
int ncomm;
int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
int ncomm;
double dbytes;
std::vector<double> times(Nloop);
for(int i=0;i<Nloop;i++){
double start=usecond();
std::vector<CartesianCommunicator::CommsRequest_t> requests;
dbytes=0;
ncomm=0;
@@ -150,7 +166,6 @@ public:
if (mpi_layout[mu]>1 ) {
ncomm++;
int xmit_to_rank;
int recv_from_rank;
if ( dir == mu ) {
@@ -160,18 +175,18 @@ public:
int comm_proc = mpi_layout[mu]-1;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
}
#if 1
tbytes= Grid.StencilSendToRecvFromBegin(requests,
(void *)&xbuf[dir][0],
xmit_to_rank,
(void *)&rbuf[dir][0],
recv_from_rank,
bytes,dir);
Grid.StencilSendToRecvFromComplete(requests,dir);
#endif
requests.resize(0);
tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
(void *)&rbuf[dir][0], recv_from_rank,
bytes,dir);
#ifdef GRID_OMP
#pragma omp atomic
#endif
ncomm++;
#ifdef GRID_OMP
#pragma omp atomic
#endif
dbytes+=tbytes;
}
}
@@ -181,13 +196,15 @@ public:
}
timestat.statistics(t_time);
// for(int i=0;i<t_time.size();i++){
// std::cout << i<<" "<<t_time[i]<<std::endl;
// }
dbytes=dbytes*ppn;
double xbytes = dbytes*0.5;
double rbytes = dbytes*0.5;
double bidibytes = dbytes;
std::cout<<GridLogMessage << std::setw(4) << lat<<"\t"<<Ls<<"\t"
<<std::setw(11) << bytes<< std::fixed << std::setprecision(1) << std::setw(7)
<<std::right<< xbytes/timestat.mean<<" "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " "
@@ -196,7 +213,8 @@ public:
<< bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
}
}
}
return;
@@ -218,7 +236,7 @@ public:
std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
uint64_t lmax=48;
#define NLOOP (50*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
#define NLOOP (100*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
GridSerialRNG sRNG; sRNG.SeedFixedIntegers(std::vector<int>({45,12,81,9}));
for(int lat=8;lat<=lmax;lat+=4){
@@ -253,8 +271,7 @@ public:
}
};
static void DWF(int Ls,int L)
static double DWF5(int Ls,int L)
{
RealD mass=0.1;
RealD M5 =1.8;
@@ -262,6 +279,7 @@ public:
double mflops;
double mflops_best = 0;
double mflops_worst= 0;
std::vector<double> mflops_all;
///////////////////////////////////////////////////////
// Set/Get the layout & grid size
@@ -274,6 +292,189 @@ public:
GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global=NN;
uint64_t SHM=NP/NN;
std::vector<int> internal;
if ( SHM == 1 ) internal = std::vector<int>({1,1,1,1});
else if ( SHM == 2 ) internal = std::vector<int>({2,1,1,1});
else if ( SHM == 4 ) internal = std::vector<int>({2,2,1,1});
else if ( SHM == 8 ) internal = std::vector<int>({2,2,2,1});
else assert(0);
std::vector<int> nodes({mpi[0]/internal[0],mpi[1]/internal[1],mpi[2]/internal[2],mpi[3]/internal[3]});
std::vector<int> latt4({local[0]*nodes[0],local[1]*nodes[1],local[2]*nodes[2],local[3]*nodes[3]});
///////// Welcome message ////////////
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "Benchmark DWF Ls vec on "<<L<<"^4 local volume "<<std::endl;
std::cout<<GridLogMessage << "* Global volume : "<<GridCmdVectorIntToString(latt4)<<std::endl;
std::cout<<GridLogMessage << "* Ls : "<<Ls<<std::endl;
std::cout<<GridLogMessage << "* MPI ranks : "<<GridCmdVectorIntToString(mpi)<<std::endl;
std::cout<<GridLogMessage << "* Intranode : "<<GridCmdVectorIntToString(internal)<<std::endl;
std::cout<<GridLogMessage << "* nodes : "<<GridCmdVectorIntToString(nodes)<<std::endl;
std::cout<<GridLogMessage << "* Using "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
///////// Lattice Init ////////////
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * sUGrid = SpaceTimeGrid::makeFourDimDWFGrid(latt4,GridDefaultMpi());
GridRedBlackCartesian * sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
GridCartesian * sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls,UGrid);
GridRedBlackCartesian * sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls,UGrid);
///////// RNG Init ////////////
std::vector<int> seeds4({1,2,3,4});
std::vector<int> seeds5({5,6,7,8});
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
GridParallelRNG RNG5(sFGrid); RNG5.SeedFixedIntegers(seeds5);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
///////// Source preparation ////////////
LatticeFermion src (sFGrid); random(RNG5,src);
LatticeFermion tmp (sFGrid);
RealD N2 = 1.0/::sqrt(norm2(src));
src = src*N2;
LatticeGaugeField Umu(UGrid); SU3::HotConfiguration(RNG4,Umu);
WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
LatticeFermion src_e (sFrbGrid);
LatticeFermion src_o (sFrbGrid);
LatticeFermion r_e (sFrbGrid);
LatticeFermion r_o (sFrbGrid);
LatticeFermion r_eo (sFGrid);
LatticeFermion err (sFGrid);
{
pickCheckerboard(Even,src_e,src);
pickCheckerboard(Odd,src_o,src);
#if defined(AVX512)
const int num_cases = 6;
std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
#else
const int num_cases = 4;
std::string fmt("U/S ; U/O ; G/S ; G/O ");
#endif
controls Cases [] = {
#ifdef AVX512
{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
#endif
{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptGeneric , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptGeneric , QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential }
};
for(int c=0;c<num_cases;c++) {
QCD::WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
QCD::WilsonKernelsStatic::Opt = Cases[c].Opt;
CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptGeneric ) std::cout << GridLogMessage<< "* Using GENERIC Nc WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptHandUnroll) std::cout << GridLogMessage<< "* Using Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm ) std::cout << GridLogMessage<< "* Using Asm Nc=3 WilsonKernels" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsAndCompute ) std::cout << GridLogMessage<< "* Using Overlapped Comms/Compute" <<std::endl;
if ( WilsonKernelsStatic::Comms == WilsonKernelsStatic::CommsThenCompute) std::cout << GridLogMessage<< "* Using sequential comms compute" <<std::endl;
if ( sizeof(Real)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(Real)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int nwarm = 200;
double t0=usecond();
sFGrid->Barrier();
for(int i=0;i<nwarm;i++){
sDw.DhopEO(src_o,r_e,DaggerNo);
}
sFGrid->Barrier();
double t1=usecond();
// uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
// if (ncall < 500) ncall = 500;
uint64_t ncall = 1000;
sFGrid->Broadcast(0,&ncall,sizeof(ncall));
// std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
sDw.ZeroCounters();
time_statistics timestat;
std::vector<double> t_time(ncall);
for(uint64_t i=0;i<ncall;i++){
t0=usecond();
sDw.DhopEO(src_o,r_e,DaggerNo);
t1=usecond();
t_time[i] = t1-t0;
}
sFGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=(1344.0*volume)/2;
double mf_hi, mf_lo, mf_err;
timestat.statistics(t_time);
mf_hi = flops/timestat.min;
mf_lo = flops/timestat.max;
mf_err= flops/timestat.min * timestat.err/timestat.mean;
mflops = flops/timestat.mean;
mflops_all.push_back(mflops);
if ( mflops_best == 0 ) mflops_best = mflops;
if ( mflops_worst== 0 ) mflops_worst= mflops;
if ( mflops>mflops_best ) mflops_best = mflops;
if ( mflops<mflops_worst) mflops_worst= mflops;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s = "<< mflops << " ("<<mf_err<<") " << mf_lo<<"-"<<mf_hi <<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per rank "<< mflops/NP<<std::endl;
std::cout<<GridLogMessage << std::fixed << std::setprecision(1)<<"sDeo mflop/s per node "<< mflops/NN<<std::endl;
sDw.Report();
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Best mflop/s = "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Worst mflop/s = "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Performance Robustness = "<< mflops_worst/mflops_best <<std::endl;
std::cout<<GridLogMessage <<fmt << std::endl;
std::cout<<GridLogMessage ;
for(int i=0;i<mflops_all.size();i++){
std::cout<<mflops_all[i]/NN<<" ; " ;
}
std::cout<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
return mflops_best;
}
static double DWF(int Ls,int L)
{
RealD mass=0.1;
RealD M5 =1.8;
double mflops;
double mflops_best = 0;
double mflops_worst= 0;
std::vector<double> mflops_all;
///////////////////////////////////////////////////////
// Set/Get the layout & grid size
///////////////////////////////////////////////////////
int threads = GridThread::GetThreads();
std::vector<int> mpi = GridDefaultMpi(); assert(mpi.size()==4);
std::vector<int> local({L,L,L,L});
GridCartesian * TmpGrid = SpaceTimeGrid::makeFourDimGrid(std::vector<int>({64,64,64,64}),
GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global=NN;
uint64_t SHM=NP/NN;
std::vector<int> internal;
@@ -364,13 +565,15 @@ public:
#if defined(AVX512)
const int num_cases = 6;
std::string fmt("A/S ; A/O ; U/S ; U/O ; G/S ; G/O ");
#else
const int num_cases = 4;
std::string fmt("U/S ; U/O ; G/S ; G/O ");
#endif
controls Cases [] = {
#ifdef AVX512
{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptInlineAsm , QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
#endif
{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsThenCompute ,CartesianCommunicator::CommunicatorPolicySequential },
{ QCD::WilsonKernelsStatic::OptHandUnroll, QCD::WilsonKernelsStatic::CommsAndCompute ,CartesianCommunicator::CommunicatorPolicySequential },
@@ -394,7 +597,7 @@ public:
if ( sizeof(Real)==8 ) std::cout << GridLogMessage<< "* DOUBLE precision "<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int nwarm = 10;
int nwarm = 200;
double t0=usecond();
FGrid->Barrier();
for(int i=0;i<nwarm;i++){
@@ -402,7 +605,10 @@ public:
}
FGrid->Barrier();
double t1=usecond();
uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
// uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
// if (ncall < 500) ncall = 500;
uint64_t ncall = 1000;
FGrid->Broadcast(0,&ncall,sizeof(ncall));
// std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
@@ -428,7 +634,7 @@ public:
mf_err= flops/timestat.min * timestat.err/timestat.mean;
mflops = flops/timestat.mean;
mflops_all.push_back(mflops);
if ( mflops_best == 0 ) mflops_best = mflops;
if ( mflops_worst== 0 ) mflops_worst= mflops;
if ( mflops>mflops_best ) mflops_best = mflops;
@@ -450,12 +656,20 @@ public:
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best mflop/s = "<< mflops_best <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s = "<< mflops_worst<<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best mflop/s = "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s = "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Performance Robustness = "<< mflops_worst/mflops_best <<std::endl;
std::cout<<GridLogMessage <<fmt << std::endl;
std::cout<<GridLogMessage ;
for(int i=0;i<mflops_all.size();i++){
std::cout<<mflops_all[i]/NN<<" ; " ;
}
std::cout<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
return mflops_best;
}
};
@@ -493,26 +707,66 @@ int main (int argc, char ** argv)
// empty for now
}
int sel=2;
std::vector<int> L_list({8,12,16,24});
std::vector<double> wilson;
std::vector<double> dwf4;
std::vector<double> dwf5;
if ( do_wilson ) {
int Ls=1;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Wilson dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
Benchmark::DWF(Ls,16);
Benchmark::DWF(Ls,24);
Benchmark::DWF(Ls,32);
for(int l=0;l<L_list.size();l++){
wilson.push_back(Benchmark::DWF(1,L_list[l]));
}
}
int Ls=16;
if ( do_dwf ) {
int Ls=16;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
Benchmark::DWF(Ls,8);
Benchmark::DWF(Ls,12);
Benchmark::DWF(Ls,16);
Benchmark::DWF(Ls,24);
for(int l=0;l<L_list.size();l++){
dwf4.push_back(Benchmark::DWF(Ls,L_list[l]));
}
}
if ( do_dwf ) {
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
for(int l=0;l<L_list.size();l++){
dwf5.push_back(Benchmark::DWF5(Ls,L_list[l]));
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "L \t\t Wilson \t DWF4 \t DWF5 " <<std::endl;
for(int l=0;l<L_list.size();l++){
std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t "<<dwf4[l]<<" \t "<<dwf5[l] <<std::endl;
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
int NN=NN_global;
std::cout<<GridLogMessage << " L \t\t Wilson\t\t DWF4 \t\t DWF5 " <<std::endl;
for(int l=0;l<L_list.size();l++){
std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]/NN<<" \t "<<dwf4[l]/NN<<" \t "<<dwf5[l] /NN<<std::endl;
}
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Comparison point result: " << dwf4[sel]/NN <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}
Grid_finalize();
}

30
benchmarks/Benchmark_comms.cc
View File

@@ -92,11 +92,16 @@ int main (int argc, char ** argv)
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
Vector<Vector<HalfSpinColourVectorD> > xbuf(8,Vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
Vector<Vector<HalfSpinColourVectorD> > rbuf(8,Vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);
std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
int ncomm;
int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
for(int mu=0;mu<8;mu++){
xbuf[mu].resize(lat*lat*lat*Ls);
rbuf[mu].resize(lat*lat*lat*Ls);
// std::cout << " buffers " << std::hex << (uint64_t)&xbuf[mu][0] <<" " << (uint64_t)&rbuf[mu][0] <<std::endl;
}
for(int i=0;i<Nloop;i++){
double start=usecond();
@@ -112,7 +117,6 @@ int main (int argc, char ** argv)
int comm_proc=1;
int xmit_to_rank;
int recv_from_rank;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFromBegin(requests,
(void *)&xbuf[mu][0],
@@ -172,9 +176,14 @@ int main (int argc, char ** argv)
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
Vector<Vector<HalfSpinColourVectorD> > xbuf(8,Vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
Vector<Vector<HalfSpinColourVectorD> > rbuf(8,Vector<HalfSpinColourVectorD>(lat*lat*lat*Ls));
std::vector<Vector<HalfSpinColourVectorD> > xbuf(8);
std::vector<Vector<HalfSpinColourVectorD> > rbuf(8);
for(int mu=0;mu<8;mu++){
xbuf[mu].resize(lat*lat*lat*Ls);
rbuf[mu].resize(lat*lat*lat*Ls);
// std::cout << " buffers " << std::hex << (uint64_t)&xbuf[mu][0] <<" " << (uint64_t)&rbuf[mu][0] <<std::endl;
}
int ncomm;
int bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
@@ -493,14 +502,9 @@ int main (int argc, char ** argv)
int comm_proc = mpi_layout[mu]-1;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
}
tbytes= Grid.StencilSendToRecvFromBegin(requests,
(void *)&xbuf[dir][0],
xmit_to_rank,
(void *)&rbuf[dir][0],
recv_from_rank,
bytes,dir);
Grid.StencilSendToRecvFromComplete(requests,dir);
requests.resize(0);
tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
(void *)&rbuf[dir][0], recv_from_rank, bytes,dir);
#pragma omp atomic
dbytes+=tbytes;