portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-10-31 12:04:33 +00:00
Code Releases Activity

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

Browse Source
This commit is contained in:
peterx.a.boyle
2021-08-10 05:41:18 -07:00
parent 50181f16e5 75030637cc
commit 1eda4d8e0b
2 changed files with 44 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/communicator/SharedMemoryMPI.cc
View File

@@ -890,7 +890,7 @@ void SharedMemory::SetCommunicator(Grid_MPI_Comm comm)
}
#endif
SharedMemoryTest();
//SharedMemoryTest();
}
//////////////////////////////////////////////////////////////////
// On node barrier

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;