fjosw/lattice-benchmarks
1
0
Fork 0
forked from portelli/lattice-benchmarks
Code Issues Pull Requests Packages Projects Releases Wiki Activity

main #2

Merged
fjosw merged 16 commits from portelli/lattice-benchmarks:main into main 2023-01-27 16:35:49 +00:00
Conversation 0 Commits 16 Files Changed 28 +190 -163
3 changed files with 190 additions and 163 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit c6349bb93f - Show all commits

14
Grid/.clang-format Normal file
View File

@ -0,0 +1,14 @@
{
BasedOnStyle: LLVM,
UseTab: Never,
IndentWidth: 2,
TabWidth: 2,
BreakBeforeBraces: Allman,
AllowShortIfStatementsOnASingleLine: false,
IndentCaseLabels: false,
ColumnLimit: 0,
AccessModifierOffset: -4,
NamespaceIndentation: All,
FixNamespaceComments: false,
SortIncludes: true,
}

5
Grid/Benchmark_ITT.cpp → Grid/Benchmark_Grid.cpp
View File

@ -2,6 +2,8 @@
Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
This is a refactoring of Benchmark_ITT.cpp from Grid
This program is free software; you can redistribute it and/or
modify it under the terms of the GNU General Public License
as published by the Free Software Foundation; either version 2
@ -67,6 +69,7 @@ Gamma::Algebra Gmu[] = {
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT};
struct controls
{
int Opt;
@ -76,7 +79,7 @@ struct controls
class Benchmark
{
public:
public:
static void Decomposition(void)
{

334
Grid/Benchmark_comms_host_device.cpp
View File

@ -21,231 +21,241 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
using namespace std;
using namespace Grid;
struct time_statistics{
struct time_statistics
{
double mean;
double err;
double min;
double max;
void statistics(std::vector<double> v){
double sum = std::accumulate(v.begin(), v.end(), 0.0);
mean = sum / v.size();
void statistics(std::vector<double> v)
{
double sum = std::accumulate(v.begin(), v.end(), 0.0);
mean = sum / v.size();
std::vector<double> diff(v.size());
std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
err = std::sqrt(sq_sum / (v.size()*(v.size() - 1)));
std::vector<double> diff(v.size());
std::transform(v.begin(), v.end(), diff.begin(), [=](double x)
{ return x - mean; });
double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
err = std::sqrt(sq_sum / (v.size() * (v.size() - 1)));
auto result = std::minmax_element(v.begin(), v.end());
min = *result.first;
max = *result.second;
}
auto result = std::minmax_element(v.begin(), v.end());
min = *result.first;
max = *result.second;
}
};
void header(){
std::cout <<GridLogMessage << " L "<<"\t"<<" Ls "<<"\t"
<<std::setw(11)<<"bytes\t\t"<<"MB/s uni"<<"\t"<<"MB/s bidi"<<std::endl;
};
int main (int argc, char ** argv)
void header()
{
Grid_init(&argc,&argv);
std::cout << GridLogMessage << " L "
<< "\t"
<< " Ls "
<< "\t"
<< std::setw(11) << "bytes\t\t"
<< "MB/s uni"
<< "\t"
<< "MB/s bidi" << std::endl;
};
Coordinate simd_layout = GridDefaultSimd(Nd,vComplexD::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
int main(int argc, char **argv)
{
Grid_init(&argc, &argv);
Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
int threads = GridThread::GetThreads();
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
std::cout << GridLogMessage << "Grid is setup to use " << threads << " threads" << std::endl;
int Nloop=250;
int nmu=0;
int maxlat=32;
for(int mu=0;mu<Nd;mu++) if (mpi_layout[mu]>1) nmu++;
int Nloop = 250;
int nmu = 0;
int maxlat = 32;
for (int mu = 0; mu < Nd; mu++)
if (mpi_layout[mu] > 1)
nmu++;
std::cout << GridLogMessage << "Number of iterations to average: "<< Nloop << std::endl;
std::cout << GridLogMessage << "Number of iterations to average: " << Nloop << std::endl;
std::vector<double> t_time(Nloop);
// time_statistics timestat;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from host memory "<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
std::cout << GridLogMessage << "= Benchmarking sequential halo exchange from host memory " << std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
header();
for(int lat=8;lat<=maxlat;lat+=4){
for(int Ls=8;Ls<=8;Ls*=2){
for (int lat = 8; lat <= maxlat; lat += 4)
{
for (int Ls = 8; Ls <= 8; Ls *= 2)
{
Coordinate latt_size ({lat*mpi_layout[0],
lat*mpi_layout[1],
lat*mpi_layout[2],
lat*mpi_layout[3]});
Coordinate latt_size({lat * mpi_layout[0],
lat * mpi_layout[1],
lat * mpi_layout[2],
lat * mpi_layout[3]});
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridCartesian Grid(latt_size, simd_layout, mpi_layout);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
RealD ppn = Nrank / Nnode;
std::vector<std::vector<HalfSpinColourVectorD> > xbuf(8);
std::vector<std::vector<HalfSpinColourVectorD> > rbuf(8);
std::vector<std::vector<HalfSpinColourVectorD>> xbuf(8);
std::vector<std::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);
for (int mu = 0; mu < 8; mu++)
{
xbuf[mu].resize(lat * lat * lat * Ls);
rbuf[mu].resize(lat * lat * lat * Ls);
}
uint64_t bytes=lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
uint64_t bytes = lat * lat * lat * Ls * sizeof(HalfSpinColourVectorD);
int ncomm;
for(int mu=0;mu<4;mu++){
if (mpi_layout[mu]>1 ) {
double start=usecond();
for(int i=0;i<Nloop;i++){
for (int mu = 0; mu < 4; mu++)
{
if (mpi_layout[mu] > 1)
{
double start = usecond();
for (int i = 0; i < Nloop; i++)
{
ncomm=0;
ncomm = 0;
ncomm++;
int comm_proc = 1;
int xmit_to_rank;
int recv_from_rank;
ncomm++;
int comm_proc=1;
int xmit_to_rank;
int recv_from_rank;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
}
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
}
comm_proc = mpi_layout[mu] - 1;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0],
xmit_to_rank,
(void *)&rbuf[mu + 4][0],
recv_from_rank,
bytes);
}
}
Grid.Barrier();
double stop = usecond();
double mean = (stop - start) / Nloop;
double dbytes = bytes * ppn;
double xbytes = dbytes * 2.0 * ncomm;
double rbytes = xbytes;
double bidibytes = xbytes + rbytes;
comm_proc = mpi_layout[mu]-1;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu+4][0],
xmit_to_rank,
(void *)&rbuf[mu+4][0],
recv_from_rank,
bytes);
}
}
Grid.Barrier();
double stop=usecond();
double mean=(stop-start)/Nloop;
double dbytes = bytes*ppn;
double xbytes = dbytes*2.0*ncomm;
double rbytes = xbytes;
double bidibytes = xbytes+rbytes;
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/mean<<" "
<< "\t\t"<<std::setw(7)<< bidibytes/mean<< std::endl;
}
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 / mean << " "
<< "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
}
}
}
}
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking sequential halo exchange from GPU memory "<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
std::cout << GridLogMessage << "= Benchmarking sequential halo exchange from GPU memory " << std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
header();
for(int lat=8;lat<=maxlat;lat+=4){
for(int Ls=8;Ls<=8;Ls*=2){
for (int lat = 8; lat <= maxlat; lat += 4)
{
for (int Ls = 8; Ls <= 8; Ls *= 2)
{
Coordinate latt_size ({lat*mpi_layout[0],
lat*mpi_layout[1],
lat*mpi_layout[2],
lat*mpi_layout[3]});
Coordinate latt_size({lat * mpi_layout[0],
lat * mpi_layout[1],
lat * mpi_layout[2],
lat * mpi_layout[3]});
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
GridCartesian Grid(latt_size, simd_layout, mpi_layout);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
RealD ppn = Nrank / Nnode;
std::vector<HalfSpinColourVectorD *> xbuf(8);
std::vector<HalfSpinColourVectorD *> rbuf(8);
uint64_t bytes = lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD);
for(int d=0;d<8;d++){
xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
uint64_t bytes = lat * lat * lat * Ls * sizeof(HalfSpinColourVectorD);
for (int d = 0; d < 8; d++)
{
xbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
rbuf[d] = (HalfSpinColourVectorD *)acceleratorAllocDevice(bytes);
}
int ncomm;
for(int mu=0;mu<4;mu++){
if (mpi_layout[mu]>1 ) {
double start=usecond();
for(int i=0;i<Nloop;i++){
for (int mu = 0; mu < 4; mu++)
{
if (mpi_layout[mu] > 1)
{
double start = usecond();
for (int i = 0; i < Nloop; i++)
{
ncomm=0;
ncomm = 0;
ncomm++;
int comm_proc = 1;
int xmit_to_rank;
int recv_from_rank;
ncomm++;
int comm_proc=1;
int xmit_to_rank;
int recv_from_rank;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
}
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0],
xmit_to_rank,
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
}
comm_proc = mpi_layout[mu] - 1;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0],
xmit_to_rank,
(void *)&rbuf[mu + 4][0],
recv_from_rank,
bytes);
}
}
Grid.Barrier();
double stop = usecond();
double mean = (stop - start) / Nloop;
double dbytes = bytes * ppn;
double xbytes = dbytes * 2.0 * ncomm;
double rbytes = xbytes;
double bidibytes = xbytes + rbytes;
comm_proc = mpi_layout[mu]-1;
{
std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu,comm_proc,xmit_to_rank,recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu+4][0],
xmit_to_rank,
(void *)&rbuf[mu+4][0],
recv_from_rank,
bytes);
}
}
Grid.Barrier();
double stop=usecond();
double mean=(stop-start)/Nloop;
double dbytes = bytes*ppn;
double xbytes = dbytes*2.0*ncomm;
double rbytes = xbytes;
double bidibytes = xbytes+rbytes;
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/mean<<" "
<< "\t\t"<<std::setw(7)<< bidibytes/mean<< std::endl;
}
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 / mean << " "
<< "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
}
}
for(int d=0;d<8;d++){
acceleratorFreeDevice(xbuf[d]);
acceleratorFreeDevice(rbuf[d]);
for (int d = 0; d < 8; d++)
{
acceleratorFreeDevice(xbuf[d]);
acceleratorFreeDevice(rbuf[d]);
}
}
}
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= All done; Bye Bye"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
std::cout << GridLogMessage << "= All done; Bye Bye" << std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
Grid_finalize();
}