portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Update from the gods upstream

Browse Source
This commit is contained in:
Henrique B. R.
2021-09-24 16:39:56 +01:00
parent 428b8ba907 67e08aa952
commit 30e5311b43
135 changed files with 4674 additions and 1138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

93
benchmarks/Benchmark_ITT.cc
View File

@ -133,34 +133,30 @@ public:
std::vector<HalfSpinColourVectorD *> xbuf(8);
std::vector<HalfSpinColourVectorD *> rbuf(8);
Grid.ShmBufferFreeAll();
//Grid.ShmBufferFreeAll();
uint64_t bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
for(int d=0;d<8;d++){
xbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
rbuf[d] = (HalfSpinColourVectorD *)Grid.ShmBufferMalloc(lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
// bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
// bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
}
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();
for(int dir=0;dir<8;dir++) {
int mu =dir % 4;
if (mpi_layout[mu]>1 ) {
dbytes=0;
ncomm=0;
std::vector<double> times(Nloop);
for(int i=0;i<Nloop;i++){
thread_for(dir,8,{
double tbytes;
int mu =dir % 4;
if (mpi_layout[mu]>1 ) {
dbytes=0;
double start=usecond();
int xmit_to_rank;
int recv_from_rank;
if ( dir == mu ) {
int comm_proc=1;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
@ -168,40 +164,38 @@ public:
int comm_proc = mpi_layout[mu]-1;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
}
tbytes= Grid.StencilSendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
(void *)&rbuf[dir][0], recv_from_rank,
bytes,dir);
thread_critical {
ncomm++;
dbytes+=tbytes;
}
Grid.SendToRecvFrom((void *)&xbuf[dir][0], xmit_to_rank,
(void *)&rbuf[dir][0], recv_from_rank,
bytes);
dbytes+=bytes;
double stop=usecond();
t_time[i] = stop-start; // microseconds
}
});
Grid.Barrier();
double stop=usecond();
t_time[i] = stop-start; // microseconds
timestat.statistics(t_time);
dbytes=dbytes*ppn;
double xbytes = dbytes*0.5;
double bidibytes = dbytes;
std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
<< bytes << " \t "
<<xbytes/timestat.mean<<" \t "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " \t "
<<xbytes/timestat.max <<" "<< xbytes/timestat.min
<< "\t\t"<< bidibytes/timestat.mean<< " " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
<< bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
}
}
timestat.statistics(t_time);
dbytes=dbytes*ppn;
double xbytes = dbytes*0.5;
// double rbytes = dbytes*0.5;
double bidibytes = dbytes;
std::cout<<GridLogMessage << lat<<"\t"<<Ls<<"\t "
<< bytes << " \t "
<<xbytes/timestat.mean<<" \t "<< xbytes*timestat.err/(timestat.mean*timestat.mean)<< " \t "
<<xbytes/timestat.max <<" "<< xbytes/timestat.min
<< "\t\t"<< bidibytes/timestat.mean<< " " << bidibytes*timestat.err/(timestat.mean*timestat.mean) << " "
<< bidibytes/timestat.max << " " << bidibytes/timestat.min << std::endl;
}
}
for(int d=0;d<8;d++){
acceleratorFreeDevice(xbuf[d]);
acceleratorFreeDevice(rbuf[d]);
}
}
}
return;
}
static void Memory(void)
@ -281,7 +275,6 @@ public:
uint64_t lmax=32;
#define NLOOP (1000*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+=8){
@ -445,7 +438,7 @@ public:
// 1344= 3*(2*8+6)*2*8 + 8*3*2*2 + 3*4*2*8
// 1344 = Nc* (6+(Nc-1)*8)*2*Nd + Nd*Nc*2*2 + Nd*Nc*Ns*2
// double flops=(1344.0*volume)/2;
#if 1
#if 0
double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + Nd*Nc*Ns + Nd*Nc*Ns*2;
#else
double fps = Nc* (6+(Nc-1)*8)*Ns*Nd + 2*Nd*Nc*Ns + 2*Nd*Nc*Ns*2;
@ -716,12 +709,12 @@ int main (int argc, char ** argv)
if ( do_su4 ) {
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
std::cout<<GridLogMessage << " SU(4) benchmark " <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
Benchmark::SU4();
}
if ( do_comms && (NN>1) ) {
if ( do_comms ) {
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;

2
benchmarks/Benchmark_comms_host_device.cc
View File

@ -53,7 +53,7 @@ struct time_statistics{
void header(){
std::cout <<GridLogMessage << " L "<<"\t"<<" Ls "<<"\t"
<<std::setw(11)<<"bytes\t\t"<<"MB/s uni (err/min/max)"<<"\t\t"<<"MB/s bidi (err/min/max)"<<std::endl;
<<std::setw(11)<<"bytes\t\t"<<"MB/s uni"<<"\t"<<"MB/s bidi"<<std::endl;
};
int main (int argc, char ** argv)

28
benchmarks/Benchmark_dwf.cc
View File

@ -236,34 +236,6 @@ int main (int argc, char ** argv)
Dw.Report();
}
DomainWallFermionRL DwH(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
if (0) {
FGrid->Barrier();
DwH.ZeroCounters();
DwH.Dhop(src,result,0);
double t0=usecond();
for(int i=0;i<ncall;i++){
__SSC_START;
DwH.Dhop(src,result,0);
__SSC_STOP;
}
double t1=usecond();
FGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=single_site_flops*volume*ncall;
std::cout<<GridLogMessage << "Called half prec comms Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
std::cout<<GridLogMessage << "mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "mflop/s per rank = "<< flops/(t1-t0)/NP<<std::endl;
std::cout<<GridLogMessage << "mflop/s per node = "<< flops/(t1-t0)/NN<<std::endl;
err = ref-result;
std::cout<<GridLogMessage << "norm diff "<< norm2(err)<<std::endl;
assert (norm2(err)< 1.0e-3 );
DwH.Report();
}
if (1)
{ // Naive wilson dag implementation
ref = Zero();

2
benchmarks/Benchmark_dwf_fp32.cc
View File

@ -182,7 +182,7 @@ int main (int argc, char ** argv)
std::cout << GridLogMessage<< "*****************************************************************" <<std::endl;
DomainWallFermionF Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
int ncall =1000;
int ncall =3000;
if (1) {
FGrid->Barrier();

25
benchmarks/Benchmark_gparity.cc
View File

@ -2,7 +2,6 @@
#include <sstream>
using namespace std;
using namespace Grid;
;
template<class d>
struct scal {
@ -118,30 +117,6 @@ int main (int argc, char ** argv)
Dw.Report();
}
std::cout << GridLogMessage<< "* SINGLE/HALF"<<std::endl;
GparityDomainWallFermionFH DwH(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
if (1) {
FGrid->Barrier();
DwH.ZeroCounters();
DwH.Dhop(src,result,0);
double t0=usecond();
for(int i=0;i<ncall;i++){
__SSC_START;
DwH.Dhop(src,result,0);
__SSC_STOP;
}
double t1=usecond();
FGrid->Barrier();
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
double flops=2*1320*volume*ncall;
std::cout<<GridLogMessage << "Called half prec comms Dw "<<ncall<<" times in "<<t1-t0<<" us"<<std::endl;
std::cout<<GridLogMessage << "mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "mflop/s per rank = "<< flops/(t1-t0)/NP<<std::endl;
std::cout<<GridLogMessage << "mflop/s per node = "<< flops/(t1-t0)/NN<<std::endl;
DwH.Report();
}
GridCartesian * UGrid_d = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid_d = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid_d);