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

Initial commit

Browse Source
This commit is contained in:
Antonin Portelli 2023-01-13 19:00:20 +00:00
commit 2a64c41a8c
10 changed files with 2404 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
.gitignore vendored Normal file
View File

@ -0,0 +1,15 @@
/.vscode
build*
.buildutils
autom4te.cache
config.*
compile
ar-lib
aclocal.m4
configure
depcomp
install-sh
missing
Makefile.in
.DS_Store
*~

224
Grid/Benchmark_IO.cpp Normal file
View File

@ -0,0 +1,224 @@
/*
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
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
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include "Benchmark_IO.hpp"
#ifndef BENCH_IO_LMIN
#define BENCH_IO_LMIN 8
#endif
#ifndef BENCH_IO_LMAX
#define BENCH_IO_LMAX 32
#endif
#ifndef BENCH_IO_NPASS
#define BENCH_IO_NPASS 10
#endif
#ifdef HAVE_LIME
using namespace Grid;
std::string filestem(const int l)
{
return "iobench_l" + std::to_string(l);
}
int vol(const int i)
{
return BENCH_IO_LMIN + 2 * i;
}
int volInd(const int l)
{
return (l - BENCH_IO_LMIN) / 2;
}
template <typename Mat>
void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
{
auto nr = data[0].rows(), nc = data[0].cols();
Eigen::MatrixXd sqSum(nr, nc);
double n = static_cast<double>(data.size());
assert(n > 1.);
mean = Mat::Zero(nr, nc);
sqSum = Mat::Zero(nr, nc);
for (auto &d : data)
{
mean += d;
sqSum += d.cwiseProduct(d);
}
stdDev = ((sqSum - mean.cwiseProduct(mean) / n) / (n - 1.)).cwiseSqrt();
mean /= n;
}
#define grid_printf(...) \
{ \
char _buf[1024]; \
sprintf(_buf, __VA_ARGS__); \
MSG << _buf; \
}
enum
{
sRead = 0,
sWrite = 1,
gRead = 2,
gWrite = 3
};
int main(int argc, char **argv)
{
Grid_init(&argc, &argv);
int64_t threads = GridThread::GetThreads();
auto mpi = GridDefaultMpi();
unsigned int nVol = (BENCH_IO_LMAX - BENCH_IO_LMIN) / 2 + 1;
unsigned int nRelVol = (BENCH_IO_LMAX - 24) / 2 + 1;
std::vector<Eigen::MatrixXd> perf(BENCH_IO_NPASS, Eigen::MatrixXd::Zero(nVol, 4));
std::vector<Eigen::VectorXd> avPerf(BENCH_IO_NPASS, Eigen::VectorXd::Zero(4));
std::vector<int> latt;
MSG << "Grid is setup to use " << threads << " threads" << std::endl;
MSG << "MPI partition " << mpi << std::endl;
for (unsigned int i = 0; i < BENCH_IO_NPASS; ++i)
{
MSG << BIGSEP << std::endl;
MSG << "Pass " << i + 1 << "/" << BENCH_IO_NPASS << std::endl;
MSG << BIGSEP << std::endl;
MSG << SEP << std::endl;
MSG << "Benchmark std write" << std::endl;
MSG << SEP << std::endl;
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl;
writeBenchmark<LatticeFermion>(latt, filestem(l), stdWrite<LatticeFermion>);
perf[i](volInd(l), sWrite) = BinaryIO::lastPerf.mbytesPerSecond;
}
MSG << SEP << std::endl;
MSG << "Benchmark std read" << std::endl;
MSG << SEP << std::endl;
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl;
readBenchmark<LatticeFermion>(latt, filestem(l), stdRead<LatticeFermion>);
perf[i](volInd(l), sRead) = BinaryIO::lastPerf.mbytesPerSecond;
}
#ifdef HAVE_LIME
MSG << SEP << std::endl;
MSG << "Benchmark Grid C-Lime write" << std::endl;
MSG << SEP << std::endl;
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl;
writeBenchmark<LatticeFermion>(latt, filestem(l), limeWrite<LatticeFermion>);
perf[i](volInd(l), gWrite) = BinaryIO::lastPerf.mbytesPerSecond;
}
MSG << SEP << std::endl;
MSG << "Benchmark Grid C-Lime read" << std::endl;
MSG << SEP << std::endl;
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
latt = {l * mpi[0], l * mpi[1], l * mpi[2], l * mpi[3]};
MSG << "-- Local volume " << l << "^4" << std::endl;
readBenchmark<LatticeFermion>(latt, filestem(l), limeRead<LatticeFermion>);
perf[i](volInd(l), gRead) = BinaryIO::lastPerf.mbytesPerSecond;
}
#endif
avPerf[i].fill(0.);
for (int f = 0; f < 4; ++f)
for (int l = 24; l <= BENCH_IO_LMAX; l += 2)
{
avPerf[i](f) += perf[i](volInd(l), f);
}
avPerf[i] /= nRelVol;
}
Eigen::MatrixXd mean(nVol, 4), stdDev(nVol, 4), rob(nVol, 4);
Eigen::VectorXd avMean(4), avStdDev(4), avRob(4);
// double n = BENCH_IO_NPASS;
stats(mean, stdDev, perf);
stats(avMean, avStdDev, avPerf);
rob.fill(100.);
rob -= 100. * stdDev.cwiseQuotient(mean.cwiseAbs());
avRob.fill(100.);
avRob -= 100. * avStdDev.cwiseQuotient(avMean.cwiseAbs());
MSG << BIGSEP << std::endl;
MSG << "SUMMARY" << std::endl;
MSG << BIGSEP << std::endl;
MSG << "Summary of individual results (all results in MB/s)." << std::endl;
MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n",
"L", "std read", "std dev", "std write", "std dev",
"Grid read", "std dev", "Grid write", "std dev");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
grid_printf("%4d %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
l, mean(volInd(l), sRead), stdDev(volInd(l), sRead),
mean(volInd(l), sWrite), stdDev(volInd(l), sWrite),
mean(volInd(l), gRead), stdDev(volInd(l), gRead),
mean(volInd(l), gWrite), stdDev(volInd(l), gWrite));
}
MSG << std::endl;
MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s\n",
"L", "std read", "std write", "Grid read", "Grid write");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{
grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n",
l, rob(volInd(l), sRead), rob(volInd(l), sWrite),
rob(volInd(l), gRead), rob(volInd(l), gWrite));
}
MSG << std::endl;
MSG << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX << "^4 (all results in MB/s)." << std::endl;
MSG << "Every second colum gives the standard deviation of the previous column." << std::endl;
MSG << std::endl;
grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n",
"std read", "std dev", "std write", "std dev",
"Grid read", "std dev", "Grid write", "std dev");
grid_printf("%12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n",
avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite),
avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
MSG << std::endl;
MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)" << std::endl;
MSG << std::endl;
grid_printf("%12s %12s %12s %12s\n",
"std read", "std write", "Grid read", "Grid write");
grid_printf("%12.1f %12.1f %12.1f %12.1f\n",
avRob(sRead), avRob(sWrite), avRob(gRead), avRob(gWrite));
Grid_finalize();
return EXIT_SUCCESS;
}
#else
int main(int argc, char **argv) {}
#endif

273
Grid/Benchmark_IO.hpp Normal file
View File

@ -0,0 +1,273 @@
/*
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
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
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#ifndef Benchmark_IO_hpp_
#define Benchmark_IO_hpp_
#include <Grid/Grid.h>
#define MSG std::cout << GridLogMessage
#define SEP \
"-----------------------------------------------------------------------------"
#define BIGSEP \
"============================================================================="
#ifdef HAVE_LIME
namespace Grid
{
template <typename Field>
using WriterFn = std::function<void(const std::string, Field &)>;
template <typename Field>
using ReaderFn = std::function<void(Field &, const std::string)>;
// AP 06/10/2020: Standard C version in case one is suspicious of the C++ API
//
// template <typename Field>
// void stdWrite(const std::string filestem, Field &vec)
// {
// std::string rankStr = std::to_string(vec.Grid()->ThisRank());
// std::FILE *file = std::fopen((filestem + "." + rankStr + ".bin").c_str(), "wb");
// size_t size;
// uint32_t crc;
// GridStopWatch ioWatch, crcWatch;
// size = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
// autoView(vec_v, vec, CpuRead);
// crcWatch.Start();
// crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
// std::fwrite(&crc, sizeof(uint32_t), 1, file);
// crcWatch.Stop();
// MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
// ioWatch.Start();
// std::fwrite(vec_v.cpu_ptr, sizeof(typename Field::scalar_object), vec.Grid()->lSites(), file);
// ioWatch.Stop();
// std::fclose(file);
// size *= vec.Grid()->ProcessorCount();
// auto &p = BinaryIO::lastPerf;
// p.size = size;
// p.time = ioWatch.useconds();
// p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
// MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
// << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
// MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
// }
//
// template <typename Field>
// void stdRead(Field &vec, const std::string filestem)
// {
// std::string rankStr = std::to_string(vec.Grid()->ThisRank());
// std::FILE *file = std::fopen((filestem + "." + rankStr + ".bin").c_str(), "rb");
// size_t size;
// uint32_t crcRead, crcData;
// GridStopWatch ioWatch, crcWatch;
// size = vec.Grid()->lSites()*sizeof(typename Field::scalar_object);
// crcWatch.Start();
// std::fread(&crcRead, sizeof(uint32_t), 1, file);
// crcWatch.Stop();
// {
// autoView(vec_v, vec, CpuWrite);
// ioWatch.Start();
// std::fread(vec_v.cpu_ptr, sizeof(typename Field::scalar_object), vec.Grid()->lSites(), file);
// ioWatch.Stop();
// std::fclose(file);
// }
// {
// autoView(vec_v, vec, CpuRead);
// crcWatch.Start();
// crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
// crcWatch.Stop();
// }
// MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
// assert(crcData == crcRead);
// size *= vec.Grid()->ProcessorCount();
// auto &p = BinaryIO::lastPerf;
// p.size = size;
// p.time = ioWatch.useconds();
// p.mbytesPerSecond = size/1024./1024./(ioWatch.useconds()/1.e6);
// MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
// << ", " << p.mbytesPerSecond << " MB/s" << std::endl;
// MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
// }
template <typename Field>
void stdWrite(const std::string filestem, Field &vec)
{
std::string rankStr = std::to_string(vec.Grid()->ThisRank());
std::ofstream file(filestem + "." + rankStr + ".bin", std::ios::out | std::ios::binary);
size_t size, sizec;
uint32_t crc;
GridStopWatch ioWatch, crcWatch;
size = vec.Grid()->lSites() * sizeof(typename Field::scalar_object);
sizec = size / sizeof(char); // just in case of...
autoView(vec_v, vec, CpuRead);
crcWatch.Start();
crc = GridChecksum::crc32(vec_v.cpu_ptr, size);
file.write(reinterpret_cast<char *>(&crc), sizeof(uint32_t) / sizeof(char));
crcWatch.Stop();
MSG << "Std I/O write: Data CRC32 " << std::hex << crc << std::dec << std::endl;
ioWatch.Start();
file.write(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
file.flush();
ioWatch.Stop();
size *= vec.Grid()->ProcessorCount();
auto &p = BinaryIO::lastPerf;
p.size = size;
p.time = ioWatch.useconds();
p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
MSG << "Std I/O write: Wrote " << p.size << " bytes in " << ioWatch.Elapsed()
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
MSG << "Std I/O write: checksum overhead " << crcWatch.Elapsed() << std::endl;
}
template <typename Field>
void stdRead(Field &vec, const std::string filestem)
{
std::string rankStr = std::to_string(vec.Grid()->ThisRank());
std::ifstream file(filestem + "." + rankStr + ".bin", std::ios::in | std::ios::binary);
size_t size, sizec;
uint32_t crcRead, crcData;
GridStopWatch ioWatch, crcWatch;
size = vec.Grid()->lSites() * sizeof(typename Field::scalar_object);
sizec = size / sizeof(char); // just in case of...
crcWatch.Start();
file.read(reinterpret_cast<char *>(&crcRead), sizeof(uint32_t) / sizeof(char));
crcWatch.Stop();
{
autoView(vec_v, vec, CpuWrite);
ioWatch.Start();
file.read(reinterpret_cast<char *>(vec_v.cpu_ptr), sizec);
ioWatch.Stop();
}
{
autoView(vec_v, vec, CpuRead);
crcWatch.Start();
crcData = GridChecksum::crc32(vec_v.cpu_ptr, size);
crcWatch.Stop();
}
MSG << "Std I/O read: Data CRC32 " << std::hex << crcData << std::dec << std::endl;
assert(crcData == crcRead);
size *= vec.Grid()->ProcessorCount();
auto &p = BinaryIO::lastPerf;
p.size = size;
p.time = ioWatch.useconds();
p.mbytesPerSecond = size / 1024. / 1024. / (ioWatch.useconds() / 1.e6);
MSG << "Std I/O read: Read " << p.size << " bytes in " << ioWatch.Elapsed()
<< ", " << p.mbytesPerSecond << " MB/s" << std::endl;
MSG << "Std I/O read: checksum overhead " << crcWatch.Elapsed() << std::endl;
}
template <typename Field>
void limeWrite(const std::string filestem, Field &vec)
{
emptyUserRecord record;
ScidacWriter binWriter(vec.Grid()->IsBoss());
binWriter.open(filestem + ".lime.bin");
binWriter.writeScidacFieldRecord(vec, record);
binWriter.close();
}
template <typename Field>
void limeRead(Field &vec, const std::string filestem)
{
emptyUserRecord record;
ScidacReader binReader;
binReader.open(filestem + ".lime.bin");
binReader.readScidacFieldRecord(vec, record);
binReader.close();
}
inline void makeGrid(std::shared_ptr<GridBase> &gPt,
const std::shared_ptr<GridCartesian> &gBasePt,
const unsigned int Ls = 1, const bool rb = false)
{
if (rb)
{
if (Ls > 1)
{
gPt.reset(SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, gBasePt.get()));
}
else
{
gPt.reset(SpaceTimeGrid::makeFourDimRedBlackGrid(gBasePt.get()));
}
}
else
{
if (Ls > 1)
{
gPt.reset(SpaceTimeGrid::makeFiveDimGrid(Ls, gBasePt.get()));
}
else
{
gPt = gBasePt;
}
}
}
template <typename Field>
void writeBenchmark(const Coordinate &latt, const std::string filename,
const WriterFn<Field> &write,
const unsigned int Ls = 1, const bool rb = false)
{
auto mpi = GridDefaultMpi();
auto simd = GridDefaultSimd(latt.size(), Field::vector_type::Nsimd());
std::shared_ptr<GridCartesian> gBasePt(SpaceTimeGrid::makeFourDimGrid(latt, simd, mpi));
std::shared_ptr<GridBase> gPt;
std::random_device rd;
makeGrid(gPt, gBasePt, Ls, rb);
GridBase *g = gPt.get();
GridParallelRNG rng(g);
Field vec(g);
rng.SeedFixedIntegers({static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd()),
static_cast<int>(rd()), static_cast<int>(rd())});
random(rng, vec);
write(filename, vec);
}
template <typename Field>
void readBenchmark(const Coordinate &latt, const std::string filename,
const ReaderFn<Field> &read,
const unsigned int Ls = 1, const bool rb = false)
{
auto mpi = GridDefaultMpi();
auto simd = GridDefaultSimd(latt.size(), Field::vector_type::Nsimd());
std::shared_ptr<GridCartesian> gBasePt(SpaceTimeGrid::makeFourDimGrid(latt, simd, mpi));
std::shared_ptr<GridBase> gPt;
makeGrid(gPt, gBasePt, Ls, rb);
GridBase *g = gPt.get();
Field vec(g);
read(vec, filename);
}
}
#endif // LIME
#endif // Benchmark_IO_hpp_

801
Grid/Benchmark_ITT.cpp Normal file
View File

@ -0,0 +1,801 @@
/*
Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
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
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <Grid/Grid.h>
using namespace Grid;
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;
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();
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;
}
};
void comms_header()
{
std::cout << GridLogMessage << " L "
<< "\t"
<< " Ls "
<< "\t"
<< "bytes\t MB/s uni (err/min/max) \t\t MB/s bidi (err/min/max)" << std::endl;
};
Gamma::Algebra Gmu[] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT};
struct controls
{
int Opt;
int CommsOverlap;
Grid::CartesianCommunicator::CommunicatorPolicy_t CommsAsynch;
};
class Benchmark
{
public:
static void Decomposition(void)
{
int threads = GridThread::GetThreads();
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "= Grid is setup to use " << threads << " threads" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "Grid Default Decomposition patterns\n";
std::cout << GridLogMessage << "\tOpenMP threads : " << GridThread::GetThreads() << std::endl;
std::cout << GridLogMessage << "\tMPI tasks : " << GridCmdVectorIntToString(GridDefaultMpi()) << std::endl;
std::cout << GridLogMessage << "\tvReal : " << sizeof(vReal) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vReal::Nsimd())) << std::endl;
std::cout << GridLogMessage << "\tvRealF : " << sizeof(vRealF) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vRealF::Nsimd())) << std::endl;
std::cout << GridLogMessage << "\tvRealD : " << sizeof(vRealD) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vRealD::Nsimd())) << std::endl;
std::cout << GridLogMessage << "\tvComplex : " << sizeof(vComplex) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vComplex::Nsimd())) << std::endl;
std::cout << GridLogMessage << "\tvComplexF : " << sizeof(vComplexF) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vComplexF::Nsimd())) << std::endl;
std::cout << GridLogMessage << "\tvComplexD : " << sizeof(vComplexD) * 8 << "bits ; " << GridCmdVectorIntToString(GridDefaultSimd(4, vComplexD::Nsimd())) << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
}
static void Comms(void)
{
int Nloop = 200;
int nmu = 0;
int maxlat = 32;
Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
Coordinate 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;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
std::cout << GridLogMessage << "= Benchmarking threaded STENCIL halo exchange in " << nmu << " dimensions" << std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl;
comms_header();
for (int lat = 16; lat <= maxlat; lat += 8)
{
// for(int Ls=8;Ls<=8;Ls*=2){
{
int Ls = 12;
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);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank / Nnode;
std::vector<HalfSpinColourVectorD *> xbuf(8);
std::vector<HalfSpinColourVectorD *> rbuf(8);
// Grid.ShmBufferFreeAll();
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);
// bzero((void *)xbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
// bzero((void *)rbuf[d],lat*lat*lat*Ls*sizeof(HalfSpinColourVectorD));
}
// int ncomm;
double dbytes;
for (int dir = 0; dir < 8; dir++)
{
int mu = dir % 4;
if (mpi_layout[mu] > 1)
{
std::vector<double> times(Nloop);
for (int i = 0; i < Nloop; i++)
{
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);
}
else
{
int comm_proc = mpi_layout[mu] - 1;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
}
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
}
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;
}
}
for (int d = 0; d < 8; d++)
{
acceleratorFreeDevice(xbuf[d]);
acceleratorFreeDevice(rbuf[d]);
}
}
}
return;
}
static void Memory(void)
{
const int Nvec = 8;
typedef Lattice<iVector<vReal, Nvec>> LatticeVec;
typedef iVector<vReal, Nvec> Vec;
Coordinate simd_layout = GridDefaultSimd(Nd, vReal::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "= Benchmarking a*x + y bandwidth" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " L "
<< "\t\t"
<< "bytes"
<< "\t\t\t"
<< "GB/s"
<< "\t\t"
<< "Gflop/s"
<< "\t\t seconds"
<< "\t\tGB/s / node" << std::endl;
std::cout << GridLogMessage << "----------------------------------------------------------" << std::endl;
// uint64_t NP;
uint64_t NN;
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)
{
Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2], lat * mpi_layout[3]});
int64_t vol = latt_size[0] * latt_size[1] * latt_size[2] * latt_size[3];
GridCartesian Grid(latt_size, simd_layout, mpi_layout);
// NP= Grid.RankCount();
NN = Grid.NodeCount();
Vec rn;
random(sRNG, rn);
LatticeVec z(&Grid);
z = Zero();
LatticeVec x(&Grid);
x = Zero();
LatticeVec y(&Grid);
y = Zero();
double a = 2.0;
uint64_t Nloop = NLOOP;
double start = usecond();
for (int i = 0; i < Nloop; i++)
{
z = a * x - y;
}
double stop = usecond();
double time = (stop - start) / Nloop * 1000;
double flops = vol * Nvec * 2; // mul,add
double bytes = 3.0 * vol * Nvec * sizeof(Real);
std::cout << GridLogMessage << std::setprecision(3)
<< lat << "\t\t" << bytes << " \t\t" << bytes / time << "\t\t" << flops / time << "\t\t" << (stop - start) / 1000. / 1000.
<< "\t\t" << bytes / time / NN << std::endl;
}
};
static void SU4(void)
{
const int Nc4 = 4;
typedef Lattice<iMatrix<vComplexF, Nc4>> LatticeSU4;
Coordinate simd_layout = GridDefaultSimd(Nd, vComplexF::Nsimd());
Coordinate mpi_layout = GridDefaultMpi();
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "= Benchmarking z = y*x SU(4) bandwidth" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " L "
<< "\t\t"
<< "bytes"
<< "\t\t\t"
<< "GB/s"
<< "\t\t"
<< "Gflop/s"
<< "\t\t seconds"
<< "\t\tGB/s / node" << std::endl;
std::cout << GridLogMessage << "----------------------------------------------------------" << std::endl;
uint64_t NN;
uint64_t lmax = 32;
GridSerialRNG sRNG;
sRNG.SeedFixedIntegers(std::vector<int>({45, 12, 81, 9}));
for (int lat = 8; lat <= lmax; lat += 8)
{
Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2], lat * mpi_layout[3]});
int64_t vol = latt_size[0] * latt_size[1] * latt_size[2] * latt_size[3];
GridCartesian Grid(latt_size, simd_layout, mpi_layout);
NN = Grid.NodeCount();
LatticeSU4 z(&Grid);
z = Zero();
LatticeSU4 x(&Grid);
x = Zero();
LatticeSU4 y(&Grid);
y = Zero();
// double a=2.0;
uint64_t Nloop = NLOOP;
double start = usecond();
for (int i = 0; i < Nloop; i++)
{
z = x * y;
}
double stop = usecond();
double time = (stop - start) / Nloop * 1000;
double flops = vol * Nc4 * Nc4 * (6 + (Nc4 - 1) * 8); // mul,add
double bytes = 3.0 * vol * Nc4 * Nc4 * 2 * sizeof(RealF);
std::cout << GridLogMessage << std::setprecision(3)
<< lat << "\t\t" << bytes << " \t\t" << bytes / time << "\t\t" << flops / time << "\t\t" << (stop - start) / 1000. / 1000.
<< "\t\t" << bytes / time / NN << std::endl;
}
};
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();
Coordinate mpi = GridDefaultMpi();
assert(mpi.size() == 4);
Coordinate local({L, L, L, L});
Coordinate latt4({local[0] * mpi[0], local[1] * mpi[1], local[2] * mpi[2], local[3] * mpi[3]});
GridCartesian *TmpGrid = SpaceTimeGrid::makeFourDimGrid(latt4,
GridDefaultSimd(Nd, vComplex::Nsimd()),
GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global = NN;
uint64_t SHM = NP / NN;
///////// Welcome message ////////////
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "Benchmark DWF on " << L << "^4 local volume " << std::endl;
std::cout << GridLogMessage << "* Nc : " << Nc << std::endl;
std::cout << GridLogMessage << "* Global volume : " << GridCmdVectorIntToString(latt4) << std::endl;
std::cout << GridLogMessage << "* Ls : " << Ls << std::endl;
std::cout << GridLogMessage << "* ranks : " << NP << std::endl;
std::cout << GridLogMessage << "* nodes : " << NN << std::endl;
std::cout << GridLogMessage << "* ranks/node : " << SHM << std::endl;
std::cout << GridLogMessage << "* ranks geom : " << GridCmdVectorIntToString(mpi) << std::endl;
std::cout << GridLogMessage << "* Using " << threads << " threads" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
///////// Lattice Init ////////////
GridCartesian *UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
GridRedBlackCartesian *UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian *FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid);
GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(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(FGrid);
RNG5.SeedFixedIntegers(seeds5);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
typedef DomainWallFermionF Action;
typedef typename Action::FermionField Fermion;
typedef LatticeGaugeFieldF Gauge;
///////// Source preparation ////////////
Gauge Umu(UGrid);
SU<Nc>::HotConfiguration(RNG4, Umu);
Fermion src(FGrid);
random(RNG5, src);
Fermion src_e(FrbGrid);
Fermion src_o(FrbGrid);
Fermion r_e(FrbGrid);
Fermion r_o(FrbGrid);
Fermion r_eo(FGrid);
Action Dw(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
{
pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd, src_o, src);
const int num_cases = 4;
std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
controls Cases[] = {
{WilsonKernelsStatic::OptGeneric, WilsonKernelsStatic::CommsThenCompute, CartesianCommunicator::CommunicatorPolicyConcurrent},
{WilsonKernelsStatic::OptGeneric, WilsonKernelsStatic::CommsAndCompute, CartesianCommunicator::CommunicatorPolicyConcurrent},
{WilsonKernelsStatic::OptGeneric, WilsonKernelsStatic::CommsThenCompute, CartesianCommunicator::CommunicatorPolicySequential},
{WilsonKernelsStatic::OptGeneric, WilsonKernelsStatic::CommsAndCompute, CartesianCommunicator::CommunicatorPolicySequential}};
for (int c = 0; c < num_cases; c++)
{
WilsonKernelsStatic::Comms = Cases[c].CommsOverlap;
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::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;
std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
int nwarm = 10;
double t0 = usecond();
FGrid->Barrier();
for (int i = 0; i < nwarm; i++)
{
Dw.DhopEO(src_o, r_e, DaggerNo);
}
FGrid->Barrier();
double t1 = usecond();
uint64_t ncall = 500;
FGrid->Broadcast(0, &ncall, sizeof(ncall));
// std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
Dw.ZeroCounters();
time_statistics timestat;
std::vector<double> t_time(ncall);
for (uint64_t i = 0; i < ncall; i++)
{
t0 = usecond();
Dw.DhopEO(src_o, r_e, DaggerNo);
t1 = usecond();
t_time[i] = t1 - t0;
}
FGrid->Barrier();
double volume = Ls;
for (int mu = 0; mu < Nd; mu++)
volume = volume * latt4[mu];
// Nc=3 gives
// 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 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;
#endif
double flops = (fps * 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 << "Deo FlopsPerSite is " << fps << std::endl;
std::cout << GridLogMessage << std::fixed << std::setprecision(1) << "Deo mflop/s = " << mflops << " (" << mf_err << ") " << mf_lo << "-" << mf_hi << std::endl;
std::cout << GridLogMessage << std::fixed << std::setprecision(1) << "Deo mflop/s per rank " << mflops / NP << std::endl;
std::cout << GridLogMessage << std::fixed << std::setprecision(1) << "Deo mflop/s per node " << mflops / NN << std::endl;
}
std::cout << GridLogMessage << "==================================================================================" << 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 << 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 Staggered(int L)
{
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();
Coordinate mpi = GridDefaultMpi();
assert(mpi.size() == 4);
Coordinate local({L, L, L, L});
Coordinate latt4({local[0] * mpi[0], local[1] * mpi[1], local[2] * mpi[2], local[3] * mpi[3]});
GridCartesian *TmpGrid = SpaceTimeGrid::makeFourDimGrid(latt4,
GridDefaultSimd(Nd, vComplex::Nsimd()),
GridDefaultMpi());
uint64_t NP = TmpGrid->RankCount();
uint64_t NN = TmpGrid->NodeCount();
NN_global = NN;
uint64_t SHM = NP / NN;
///////// Welcome message ////////////
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "Benchmark ImprovedStaggered on " << L << "^4 local volume " << std::endl;
std::cout << GridLogMessage << "* Global volume : " << GridCmdVectorIntToString(latt4) << std::endl;
std::cout << GridLogMessage << "* ranks : " << NP << std::endl;
std::cout << GridLogMessage << "* nodes : " << NN << std::endl;
std::cout << GridLogMessage << "* ranks/node : " << SHM << std::endl;
std::cout << GridLogMessage << "* ranks geom : " << GridCmdVectorIntToString(mpi) << std::endl;
std::cout << GridLogMessage << "* Using " << threads << " threads" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
///////// Lattice Init ////////////
GridCartesian *FGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(FGrid);
///////// RNG Init ////////////
std::vector<int> seeds4({1, 2, 3, 4});
GridParallelRNG RNG4(FGrid);
RNG4.SeedFixedIntegers(seeds4);
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
RealD mass = 0.1;
RealD c1 = 9.0 / 8.0;
RealD c2 = -1.0 / 24.0;
RealD u0 = 1.0;
typedef ImprovedStaggeredFermionF Action;
typedef typename Action::FermionField Fermion;
typedef LatticeGaugeFieldF Gauge;
Gauge Umu(FGrid);
SU<Nc>::HotConfiguration(RNG4, Umu);
typename Action::ImplParams params;
Action Ds(Umu, Umu, *FGrid, *FrbGrid, mass, c1, c2, u0, params);
///////// Source preparation ////////////
Fermion src(FGrid);
random(RNG4, src);
Fermion src_e(FrbGrid);
Fermion src_o(FrbGrid);
Fermion r_e(FrbGrid);
Fermion r_o(FrbGrid);
Fermion r_eo(FGrid);
{
pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd, src_o, src);
const int num_cases = 4;
std::string fmt("G/S/C ; G/O/C ; G/S/S ; G/O/S ");
controls Cases[] = {
{StaggeredKernelsStatic::OptGeneric, StaggeredKernelsStatic::CommsThenCompute, CartesianCommunicator::CommunicatorPolicyConcurrent},
{StaggeredKernelsStatic::OptGeneric, StaggeredKernelsStatic::CommsAndCompute, CartesianCommunicator::CommunicatorPolicyConcurrent},
{StaggeredKernelsStatic::OptGeneric, StaggeredKernelsStatic::CommsThenCompute, CartesianCommunicator::CommunicatorPolicySequential},
{StaggeredKernelsStatic::OptGeneric, StaggeredKernelsStatic::CommsAndCompute, CartesianCommunicator::CommunicatorPolicySequential}};
for (int c = 0; c < num_cases; c++)
{
StaggeredKernelsStatic::Comms = Cases[c].CommsOverlap;
StaggeredKernelsStatic::Opt = Cases[c].Opt;
CartesianCommunicator::SetCommunicatorPolicy(Cases[c].CommsAsynch);
std::cout << GridLogMessage << "==================================================================================" << std::endl;
if (StaggeredKernelsStatic::Opt == StaggeredKernelsStatic::OptGeneric)
std::cout << GridLogMessage << "* Using GENERIC Nc StaggeredKernels" << std::endl;
if (StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsAndCompute)
std::cout << GridLogMessage << "* Using Overlapped Comms/Compute" << std::endl;
if (StaggeredKernelsStatic::Comms == StaggeredKernelsStatic::CommsThenCompute)
std::cout << GridLogMessage << "* Using sequential Comms/Compute" << std::endl;
std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
int nwarm = 10;
double t0 = usecond();
FGrid->Barrier();
for (int i = 0; i < nwarm; i++)
{
Ds.DhopEO(src_o, r_e, DaggerNo);
}
FGrid->Barrier();
double t1 = usecond();
uint64_t ncall = 500;
FGrid->Broadcast(0, &ncall, sizeof(ncall));
// std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
Ds.ZeroCounters();
time_statistics timestat;
std::vector<double> t_time(ncall);
for (uint64_t i = 0; i < ncall; i++)
{
t0 = usecond();
Ds.DhopEO(src_o, r_e, DaggerNo);
t1 = usecond();
t_time[i] = t1 - t0;
}
FGrid->Barrier();
double volume = 1;
for (int mu = 0; mu < Nd; mu++)
volume = volume * latt4[mu];
double flops = (1146.0 * 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) << "Deo mflop/s = " << mflops << " (" << mf_err << ") " << mf_lo << "-" << mf_hi << std::endl;
std::cout << GridLogMessage << std::fixed << std::setprecision(1) << "Deo mflop/s per rank " << mflops / NP << std::endl;
std::cout << GridLogMessage << std::fixed << std::setprecision(1) << "Deo mflop/s per node " << mflops / NN << std::endl;
}
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << L << "^4 Deo Best mflop/s = " << mflops_best << " ; " << mflops_best / NN << " per node " << std::endl;
std::cout << GridLogMessage << L << "^4 Deo Worst mflop/s = " << mflops_worst << " ; " << mflops_worst / NN << " per node " << 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;
}
};
int main(int argc, char **argv)
{
Grid_init(&argc, &argv);
CartesianCommunicator::SetCommunicatorPolicy(CartesianCommunicator::CommunicatorPolicySequential);
#ifdef KNL
LebesgueOrder::Block = std::vector<int>({8, 2, 2, 2});
#else
LebesgueOrder::Block = std::vector<int>({2, 2, 2, 2});
#endif
Benchmark::Decomposition();
int do_su4 = 1;
int do_memory = 1;
int do_comms = 1;
int sel = 4;
std::vector<int> L_list({8, 12, 16, 24, 32});
int selm1 = sel - 1;
std::vector<double> wilson;
std::vector<double> dwf4;
std::vector<double> staggered;
int Ls = 1;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Wilson dslash 4D vectorised" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
for (int l = 0; l < L_list.size(); l++)
{
wilson.push_back(Benchmark::DWF(Ls, L_list[l]));
}
Ls = 12;
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++)
{
double result = Benchmark::DWF(Ls, L_list[l]);
dwf4.push_back(result);
}
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Improved Staggered dslash 4D vectorised" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
for (int l = 0; l < L_list.size(); l++)
{
double result = Benchmark::Staggered(L_list[l]);
staggered.push_back(result);
}
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\t DWF4 \t\t Staggered" << std::endl;
for (int l = 0; l < L_list.size(); l++)
{
std::cout << GridLogMessage << L_list[l] << " \t\t " << wilson[l] << " \t\t " << dwf4[l] << " \t\t " << staggered[l] << std::endl;
}
std::cout << GridLogMessage << "==================================================================================" << std::endl;
int NN = NN_global;
if (do_memory)
{
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Memory benchmark " << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
Benchmark::Memory();
}
if (do_su4)
{
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " SU(4) benchmark " << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
Benchmark::SU4();
}
if (do_comms)
{
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Communications benchmark " << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
Benchmark::Comms();
}
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Per Node Summary table Ls=" << Ls << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " L \t\t Wilson\t\t DWF4\t\t Staggered " << 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 " << staggered[l] / NN << std::endl;
}
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << "==================================================================================" << std::endl;
std::cout << GridLogMessage << " Comparison point result: " << 0.5 * (dwf4[sel] + dwf4[selm1]) / NN << " Mflop/s per node" << std::endl;
std::cout << GridLogMessage << " Comparison point is 0.5*(" << dwf4[sel] / NN << "+" << dwf4[selm1] / NN << ") " << std::endl;
std::cout << std::setprecision(3);
std::cout << GridLogMessage << "==================================================================================" << std::endl;
Grid_finalize();
}

251
Grid/Benchmark_comms_host_device.cpp Normal file
View File

@ -0,0 +1,251 @@
/*
Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
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
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <Grid/Grid.h>
using namespace std;
using namespace Grid;
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();
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;
}
};
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)
{
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;
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::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;
header();
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]});
GridCartesian Grid(latt_size,simd_layout,mpi_layout);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
RealD ppn = Nrank/Nnode;
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);
}
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++){
ncomm=0;
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);
}
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::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){
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);
RealD Nrank = Grid._Nprocessors;
RealD Nnode = Grid.NodeCount();
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);
}
int ncomm;
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++;
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);
}
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;
}
}
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;
Grid_finalize();
}

425
Grid/Benchmark_dwf_fp32.cpp Normal file
View File

@ -0,0 +1,425 @@
/*
Copyright © 2015 Peter Boyle <paboyle@ph.ed.ac.uk>
Copyright © 2022 Antonin Portelli <antonin.portelli@me.com>
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
of the License, or (at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include <Grid/Grid.h>
#ifdef GRID_CUDA
#define CUDA_PROFILE
#endif
#ifdef CUDA_PROFILE
#include <cuda_profiler_api.h>
#endif
using namespace std;
using namespace Grid;
template <class d>
struct scal
{
d internal;
};
Gamma::Algebra Gmu[] = {
Gamma::Algebra::GammaX,
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT};
int main(int argc, char **argv)
{
Grid_init(&argc, &argv);
int threads = GridThread::GetThreads();
Coordinate latt4 = GridDefaultLatt();
int Ls = 16;
for (int i = 0; i < argc; i++)
if (std::string(argv[i]) == "-Ls")
{
std::stringstream ss(argv[i + 1]);
ss >> Ls;
}
GridLogLayout();
long unsigned int single_site_flops = 8 * Nc * (7 + 16 * Nc);
GridCartesian *UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
GridRedBlackCartesian *UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian *FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid);
GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGrid);
std::cout << GridLogMessage << "Making s innermost grids" << std::endl;
GridCartesian *sUGrid = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(), GridDefaultMpi());
GridRedBlackCartesian *sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
GridCartesian *sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls, UGrid);
GridRedBlackCartesian *sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls, UGrid);
std::vector<int> seeds4({1, 2, 3, 4});
std::vector<int> seeds5({5, 6, 7, 8});
std::cout << GridLogMessage << "Initialising 4d RNG" << std::endl;
GridParallelRNG RNG4(UGrid);
RNG4.SeedUniqueString(std::string("The 4D RNG"));
std::cout << GridLogMessage << "Initialising 5d RNG" << std::endl;
GridParallelRNG RNG5(FGrid);
RNG5.SeedUniqueString(std::string("The 5D RNG"));
std::cout << GridLogMessage << "Initialised RNGs" << std::endl;
LatticeFermionF src(FGrid);
random(RNG5, src);
#if 0
src = Zero();
{
Coordinate origin({0,0,0,latt4[2]-1,0});
SpinColourVectorF tmp;
tmp=Zero();
tmp()(0)(0)=Complex(-2.0,0.0);
std::cout << " source site 0 " << tmp<<std::endl;
pokeSite(tmp,src,origin);
}
#else
RealD N2 = 1.0 / ::sqrt(norm2(src));
src = src * N2;
#endif
LatticeFermionF result(FGrid);
result = Zero();
LatticeFermionF ref(FGrid);
ref = Zero();
LatticeFermionF tmp(FGrid);
LatticeFermionF err(FGrid);
std::cout << GridLogMessage << "Drawing gauge field" << std::endl;
LatticeGaugeFieldF Umu(UGrid);
SU<Nc>::HotConfiguration(RNG4, Umu);
std::cout << GridLogMessage << "Random gauge initialised " << std::endl;
#if 0
Umu=1.0;
for(int mu=0;mu<Nd;mu++){
LatticeColourMatrixF ttmp(UGrid);
ttmp = PeekIndex<LorentzIndex>(Umu,mu);
// if (mu !=2 ) ttmp = 0;
// ttmp = ttmp* pow(10.0,mu);
PokeIndex<LorentzIndex>(Umu,ttmp,mu);
}
std::cout << GridLogMessage << "Forced to diagonal " << std::endl;
#endif
////////////////////////////////////
// Naive wilson implementation
////////////////////////////////////
// replicate across fifth dimension
// LatticeGaugeFieldF Umu5d(FGrid);
std::vector<LatticeColourMatrixF> U(4, UGrid);
for (int mu = 0; mu < Nd; mu++)
{
U[mu] = PeekIndex<LorentzIndex>(Umu, mu);
}
std::cout << GridLogMessage << "Setting up Cshift based reference " << std::endl;
if (1)
{
ref = Zero();
for (int mu = 0; mu < Nd; mu++)
{
tmp = Cshift(src, mu + 1, 1);
{
autoView(tmp_v, tmp, CpuWrite);
autoView(U_v, U[mu], CpuRead);
for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
{
for (int s = 0; s < Ls; s++)
{
tmp_v[Ls * ss + s] = U_v[ss] * tmp_v[Ls * ss + s];
}
}
}
ref = ref + tmp - Gamma(Gmu[mu]) * tmp;
{
autoView(tmp_v, tmp, CpuWrite);
autoView(U_v, U[mu], CpuRead);
autoView(src_v, src, CpuRead);
for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
{
for (int s = 0; s < Ls; s++)
{
tmp_v[Ls * ss + s] = adj(U_v[ss]) * src_v[Ls * ss + s];
}
}
}
tmp = Cshift(tmp, mu + 1, -1);
ref = ref + tmp + Gamma(Gmu[mu]) * tmp;
}
ref = -0.5 * ref;
}
RealD mass = 0.1;
RealD M5 = 1.8;
RealD NP = UGrid->_Nprocessors;
RealD NN = UGrid->NodeCount();
std::cout << GridLogMessage << "*****************************************************************" << std::endl;
std::cout << GridLogMessage << "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" << std::endl;
std::cout << GridLogMessage << "*****************************************************************" << std::endl;
std::cout << GridLogMessage << "*****************************************************************" << std::endl;
std::cout << GridLogMessage << "* Benchmarking DomainWallFermionR::Dhop " << std::endl;
std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd() << std::endl;
std::cout << GridLogMessage << "* VComplexF size is " << sizeof(vComplexF) << " B" << std::endl;
if (sizeof(RealF) == 4)
std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
if (sizeof(RealF) == 8)
std::cout << GridLogMessage << "* DOUBLE precision " << std::endl;
#ifdef GRID_OMP
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;
#endif
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;
std::cout << GridLogMessage << "*****************************************************************" << std::endl;
DomainWallFermionF Dw(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
int ncall = 300;
if (1)
{
FGrid->Barrier();
Dw.ZeroCounters();
Dw.Dhop(src, result, 0);
std::cout << GridLogMessage << "Called warmup" << std::endl;
double t0 = usecond();
for (int i = 0; i < ncall; i++)
{
__SSC_START;
Dw.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;
auto nsimd = vComplex::Nsimd();
auto simdwidth = sizeof(vComplex);
// RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors
double data_rf = volume * ((2 * Nd + 1) * Nd * Nc + 2 * Nd * Nc * Nc) * simdwidth / nsimd * ncall / (1024. * 1024. * 1024.);
// mem: Nd Wilson * Ls, Nd gauge, Nc colors
double data_mem = (volume * (2 * Nd + 1) * Nd * Nc + (volume / Ls) * 2 * Nd * Nc * Nc) * simdwidth / nsimd * ncall / (1024. * 1024. * 1024.);
std::cout << GridLogMessage << "Called Dw " << ncall << " times in " << t1 - t0 << " us" << std::endl;
// std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
// std::cout<<GridLogMessage << "norm ref "<< norm2(ref)<<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;
std::cout << GridLogMessage << "RF GiB/s (base 2) = " << 1000000. * data_rf / ((t1 - t0)) << std::endl;
std::cout << GridLogMessage << "mem GiB/s (base 2) = " << 1000000. * data_mem / ((t1 - t0)) << std::endl;
err = ref - result;
std::cout << GridLogMessage << "norm diff " << norm2(err) << std::endl;
// exit(0);
if ((norm2(err) > 1.0e-4))
{
/*
std::cout << "RESULT\n " << result<<std::endl;
std::cout << "REF \n " << ref <<std::endl;
std::cout << "ERR \n " << err <<std::endl;
*/
std::cout << GridLogMessage << "WRONG RESULT" << std::endl;
FGrid->Barrier();
exit(-1);
}
assert(norm2(err) < 1.0e-4);
Dw.Report();
}
if (1)
{ // Naive wilson dag implementation
ref = Zero();
for (int mu = 0; mu < Nd; mu++)
{
// ref = src - Gamma(Gamma::Algebra::GammaX)* src ; // 1+gamma_x
tmp = Cshift(src, mu + 1, 1);
{
autoView(ref_v, ref, CpuWrite);
autoView(tmp_v, tmp, CpuRead);
autoView(U_v, U[mu], CpuRead);
for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
{
for (int s = 0; s < Ls; s++)
{
int i = s + Ls * ss;
ref_v[i] += U_v[ss] * (tmp_v[i] + Gamma(Gmu[mu]) * tmp_v[i]);
;
}
}
}
{
autoView(tmp_v, tmp, CpuWrite);
autoView(U_v, U[mu], CpuRead);
autoView(src_v, src, CpuRead);
for (int ss = 0; ss < U[mu].Grid()->oSites(); ss++)
{
for (int s = 0; s < Ls; s++)
{
tmp_v[Ls * ss + s] = adj(U_v[ss]) * src_v[Ls * ss + s];
}
}
}
// tmp =adj(U[mu])*src;
tmp = Cshift(tmp, mu + 1, -1);
{
autoView(ref_v, ref, CpuWrite);
autoView(tmp_v, tmp, CpuRead);
for (int i = 0; i < ref_v.size(); i++)
{
ref_v[i] += tmp_v[i] - Gamma(Gmu[mu]) * tmp_v[i];
;
}
}
}
ref = -0.5 * ref;
}
// dump=1;
Dw.Dhop(src, result, 1);
std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl;
std::cout << GridLogMessage << "Called DwDag" << std::endl;
std::cout << GridLogMessage << "norm dag result " << norm2(result) << std::endl;
std::cout << GridLogMessage << "norm dag ref " << norm2(ref) << std::endl;
err = ref - result;
std::cout << GridLogMessage << "norm dag diff " << norm2(err) << std::endl;
if ((norm2(err) > 1.0e-4))
{
/*
std::cout<< "DAG RESULT\n " <<ref << std::endl;
std::cout<< "DAG sRESULT\n " <<result << std::endl;
std::cout<< "DAG ERR \n " << err <<std::endl;
*/
}
LatticeFermionF src_e(FrbGrid);
LatticeFermionF src_o(FrbGrid);
LatticeFermionF r_e(FrbGrid);
LatticeFermionF r_o(FrbGrid);
LatticeFermionF r_eo(FGrid);
std::cout << GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec" << std::endl;
pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd, src_o, src);
std::cout << GridLogMessage << "src_e" << norm2(src_e) << std::endl;
std::cout << GridLogMessage << "src_o" << norm2(src_o) << std::endl;
// S-direction is INNERMOST and takes no part in the parity.
std::cout << GridLogMessage << "*********************************************************" << std::endl;
std::cout << GridLogMessage << "* Benchmarking DomainWallFermionF::DhopEO " << std::endl;
std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd() << std::endl;
if (sizeof(RealF) == 4)
std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
if (sizeof(RealF) == 8)
std::cout << GridLogMessage << "* DOUBLE precision " << std::endl;
#ifdef GRID_OMP
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;
#endif
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;
std::cout << GridLogMessage << "*********************************************************" << std::endl;
{
Dw.ZeroCounters();
FGrid->Barrier();
Dw.DhopEO(src_o, r_e, DaggerNo);
double t0 = usecond();
for (int i = 0; i < ncall; i++)
{
#ifdef CUDA_PROFILE
if (i == 10)
cudaProfilerStart();
#endif
Dw.DhopEO(src_o, r_e, DaggerNo);
#ifdef CUDA_PROFILE
if (i == 20)
cudaProfilerStop();
#endif
}
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) / 2.0;
std::cout << GridLogMessage << "Deo mflop/s = " << flops / (t1 - t0) << std::endl;
std::cout << GridLogMessage << "Deo mflop/s per rank " << flops / (t1 - t0) / NP << std::endl;
std::cout << GridLogMessage << "Deo mflop/s per node " << flops / (t1 - t0) / NN << std::endl;
Dw.Report();
}
Dw.DhopEO(src_o, r_e, DaggerNo);
Dw.DhopOE(src_e, r_o, DaggerNo);
Dw.Dhop(src, result, DaggerNo);
std::cout << GridLogMessage << "r_e" << norm2(r_e) << std::endl;
std::cout << GridLogMessage << "r_o" << norm2(r_o) << std::endl;
std::cout << GridLogMessage << "res" << norm2(result) << std::endl;
setCheckerboard(r_eo, r_o);
setCheckerboard(r_eo, r_e);
err = r_eo - result;
std::cout << GridLogMessage << "norm diff " << norm2(err) << std::endl;
if ((norm2(err) > 1.0e-4))
{
/*
std::cout<< "Deo RESULT\n " <<r_eo << std::endl;
std::cout<< "Deo REF\n " <<result << std::endl;
std::cout<< "Deo ERR \n " << err <<std::endl;
*/
}
pickCheckerboard(Even, src_e, err);
pickCheckerboard(Odd, src_o, err);
std::cout << GridLogMessage << "norm diff even " << norm2(src_e) << std::endl;
std::cout << GridLogMessage << "norm diff odd " << norm2(src_o) << std::endl;
assert(norm2(src_e) < 1.0e-4);
assert(norm2(src_o) < 1.0e-4);
Grid_finalize();
exit(0);
}

339
Grid/LICENSE Normal file
View File

@ -0,0 +1,339 @@
GNU GENERAL PUBLIC LICENSE
Version 2, June 1991
Copyright (C) 1989, 1991 Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The licenses for most software are designed to take away your
freedom to share and change it. By contrast, the GNU General Public
License is intended to guarantee your freedom to share and change free
software--to make sure the software is free for all its users. This
General Public License applies to most of the Free Software
Foundation's software and to any other program whose authors commit to
using it. (Some other Free Software Foundation software is covered by
the GNU Lesser General Public License instead.) You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
this service if you wish), that you receive source code or can get it
if you want it, that you can change the software or use pieces of it
in new free programs; and that you know you can do these things.
To protect your rights, we need to make restrictions that forbid
anyone to deny you these rights or to ask you to surrender the rights.
These restrictions translate to certain responsibilities for you if you
distribute copies of the software, or if you modify it.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must give the recipients all the rights that
you have. You must make sure that they, too, receive or can get the
source code. And you must show them these terms so they know their
rights.
We protect your rights with two steps: (1) copyright the software, and
(2) offer you this license which gives you legal permission to copy,
distribute and/or modify the software.
Also, for each author's protection and ours, we want to make certain
that everyone understands that there is no warranty for this free
software. If the software is modified by someone else and passed on, we
want its recipients to know that what they have is not the original, so
that any problems introduced by others will not reflect on the original
authors' reputations.
Finally, any free program is threatened constantly by software
patents. We wish to avoid the danger that redistributors of a free
program will individually obtain patent licenses, in effect making the
program proprietary. To prevent this, we have made it clear that any
patent must be licensed for everyone's free use or not licensed at all.
The precise terms and conditions for copying, distribution and
modification follow.
GNU GENERAL PUBLIC LICENSE
TERMS AND CONDITIONS FOR COPYING, DISTRIBUTION AND MODIFICATION
0. This License applies to any program or other work which contains
a notice placed by the copyright holder saying it may be distributed
under the terms of this General Public License. The "Program", below,
refers to any such program or work, and a "work based on the Program"
means either the Program or any derivative work under copyright law:
that is to say, a work containing the Program or a portion of it,
either verbatim or with modifications and/or translated into another
language. (Hereinafter, translation is included without limitation in
the term "modification".) Each licensee is addressed as "you".
Activities other than copying, distribution and modification are not
covered by this License; they are outside its scope. The act of
running the Program is not restricted, and the output from the Program
is covered only if its contents constitute a work based on the
Program (independent of having been made by running the Program).
Whether that is true depends on what the Program does.
1. You may copy and distribute verbatim copies of the Program's
source code as you receive it, in any medium, provided that you
conspicuously and appropriately publish on each copy an appropriate
copyright notice and disclaimer of warranty; keep intact all the
notices that refer to this License and to the absence of any warranty;
and give any other recipients of the Program a copy of this License
along with the Program.
You may charge a fee for the physical act of transferring a copy, and
you may at your option offer warranty protection in exchange for a fee.
2. You may modify your copy or copies of the Program or any portion
of it, thus forming a work based on the Program, and copy and
distribute such modifications or work under the terms of Section 1
above, provided that you also meet all of these conditions:
a) You must cause the modified files to carry prominent notices
stating that you changed the files and the date of any change.
b) You must cause any work that you distribute or publish, that in
whole or in part contains or is derived from the Program or any
part thereof, to be licensed as a whole at no charge to all third
parties under the terms of this License.
c) If the modified program normally reads commands interactively
when run, you must cause it, when started running for such
interactive use in the most ordinary way, to print or display an
announcement including an appropriate copyright notice and a
notice that there is no warranty (or else, saying that you provide
a warranty) and that users may redistribute the program under
these conditions, and telling the user how to view a copy of this
License. (Exception: if the Program itself is interactive but
does not normally print such an announcement, your work based on
the Program is not required to print an announcement.)
These requirements apply to the modified work as a whole. If
identifiable sections of that work are not derived from the Program,
and can be reasonably considered independent and separate works in
themselves, then this License, and its terms, do not apply to those
sections when you distribute them as separate works. But when you
distribute the same sections as part of a whole which is a work based
on the Program, the distribution of the whole must be on the terms of
this License, whose permissions for other licensees extend to the
entire whole, and thus to each and every part regardless of who wrote it.
Thus, it is not the intent of this section to claim rights or contest
your rights to work written entirely by you; rather, the intent is to
exercise the right to control the distribution of derivative or
collective works based on the Program.
In addition, mere aggregation of another work not based on the Program
with the Program (or with a work based on the Program) on a volume of
a storage or distribution medium does not bring the other work under
the scope of this License.
3. You may copy and distribute the Program (or a work based on it,
under Section 2) in object code or executable form under the terms of
Sections 1 and 2 above provided that you also do one of the following:
a) Accompany it with the complete corresponding machine-readable
source code, which must be distributed under the terms of Sections
1 and 2 above on a medium customarily used for software interchange; or,
b) Accompany it with a written offer, valid for at least three
years, to give any third party, for a charge no more than your
cost of physically performing source distribution, a complete
machine-readable copy of the corresponding source code, to be
distributed under the terms of Sections 1 and 2 above on a medium
customarily used for software interchange; or,
c) Accompany it with the information you received as to the offer
to distribute corresponding source code. (This alternative is
allowed only for noncommercial distribution and only if you
received the program in object code or executable form with such
an offer, in accord with Subsection b above.)
The source code for a work means the preferred form of the work for
making modifications to it. For an executable work, complete source
code means all the source code for all modules it contains, plus any
associated interface definition files, plus the scripts used to
control compilation and installation of the executable. However, as a
special exception, the source code distributed need not include
anything that is normally distributed (in either source or binary
form) with the major components (compiler, kernel, and so on) of the
operating system on which the executable runs, unless that component
itself accompanies the executable.
If distribution of executable or object code is made by offering
access to copy from a designated place, then offering equivalent
access to copy the source code from the same place counts as
distribution of the source code, even though third parties are not
compelled to copy the source along with the object code.
4. You may not copy, modify, sublicense, or distribute the Program
except as expressly provided under this License. Any attempt
otherwise to copy, modify, sublicense or distribute the Program is
void, and will automatically terminate your rights under this License.
However, parties who have received copies, or rights, from you under
this License will not have their licenses terminated so long as such
parties remain in full compliance.
5. You are not required to accept this License, since you have not
signed it. However, nothing else grants you permission to modify or
distribute the Program or its derivative works. These actions are
prohibited by law if you do not accept this License. Therefore, by
modifying or distributing the Program (or any work based on the
Program), you indicate your acceptance of this License to do so, and
all its terms and conditions for copying, distributing or modifying
the Program or works based on it.
6. Each time you redistribute the Program (or any work based on the
Program), the recipient automatically receives a license from the
original licensor to copy, distribute or modify the Program subject to
these terms and conditions. You may not impose any further
restrictions on the recipients' exercise of the rights granted herein.
You are not responsible for enforcing compliance by third parties to
this License.
7. If, as a consequence of a court judgment or allegation of patent
infringement or for any other reason (not limited to patent issues),
conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot
distribute so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you
may not distribute the Program at all. For example, if a patent
license would not permit royalty-free redistribution of the Program by
all those who receive copies directly or indirectly through you, then
the only way you could satisfy both it and this License would be to
refrain entirely from distribution of the Program.
If any portion of this section is held invalid or unenforceable under
any particular circumstance, the balance of the section is intended to
apply and the section as a whole is intended to apply in other
circumstances.
It is not the purpose of this section to induce you to infringe any
patents or other property right claims or to contest validity of any
such claims; this section has the sole purpose of protecting the
integrity of the free software distribution system, which is
implemented by public license practices. Many people have made
generous contributions to the wide range of software distributed
through that system in reliance on consistent application of that
system; it is up to the author/donor to decide if he or she is willing
to distribute software through any other system and a licensee cannot
impose that choice.
This section is intended to make thoroughly clear what is believed to
be a consequence of the rest of this License.
8. If the distribution and/or use of the Program is restricted in
certain countries either by patents or by copyrighted interfaces, the
original copyright holder who places the Program under this License
may add an explicit geographical distribution limitation excluding
those countries, so that distribution is permitted only in or among
countries not thus excluded. In such case, this License incorporates
the limitation as if written in the body of this License.
9. The Free Software Foundation may publish revised and/or new versions
of the General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the Program
specifies a version number of this License which applies to it and "any
later version", you have the option of following the terms and conditions
either of that version or of any later version published by the Free
Software Foundation. If the Program does not specify a version number of
this License, you may choose any version ever published by the Free Software
Foundation.
10. If you wish to incorporate parts of the Program into other free
programs whose distribution conditions are different, write to the author
to ask for permission. For software which is copyrighted by the Free
Software Foundation, write to the Free Software Foundation; we sometimes
make exceptions for this. Our decision will be guided by the two goals
of preserving the free status of all derivatives of our free software and
of promoting the sharing and reuse of software generally.
NO WARRANTY
11. BECAUSE THE PROGRAM IS LICENSED FREE OF CHARGE, THERE IS NO WARRANTY
FOR THE PROGRAM, TO THE EXTENT PERMITTED BY APPLICABLE LAW. EXCEPT WHEN
OTHERWISE STATED IN WRITING THE COPYRIGHT HOLDERS AND/OR OTHER PARTIES
PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESSED
OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE ENTIRE RISK AS
TO THE QUALITY AND PERFORMANCE OF THE PROGRAM IS WITH YOU. SHOULD THE
PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF ALL NECESSARY SERVICING,
REPAIR OR CORRECTION.
12. IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MAY MODIFY AND/OR
REDISTRIBUTE THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES,
INCLUDING ANY GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
OUT OF THE USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED
TO LOSS OF DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY
YOU OR THIRD PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER
PROGRAMS), EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE
POSSIBILITY OF SUCH DAMAGES.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
convey the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
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 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License along
with this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
Also add information on how to contact you by electronic and paper mail.
If the program is interactive, make it output a short notice like this
when it starts in an interactive mode:
Gnomovision version 69, Copyright (C) year name of author
Gnomovision comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, the commands you use may
be called something other than `show w' and `show c'; they could even be
mouse-clicks or menu items--whatever suits your program.
You should also get your employer (if you work as a programmer) or your
school, if any, to sign a "copyright disclaimer" for the program, if
necessary. Here is a sample; alter the names:
Yoyodyne, Inc., hereby disclaims all copyright interest in the program
`Gnomovision' (which makes passes at compilers) written by James Hacker.
<signature of Ty Coon>, 1 April 1989
Ty Coon, President of Vice
This General Public License does not permit incorporating your program into
proprietary programs. If your program is a subroutine library, you may
consider it more useful to permit linking proprietary applications with the
library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License.

12
Grid/Makefile.am Normal file
View File

@ -0,0 +1,12 @@
ACLOCAL_AMFLAGS = -I .buildutils/m4
bin_PROGRAMS = \
Benchmark_comms_host_device \
Benchmark_dwf_fp32 \
Benchmark_ITT \
Benchmark_IO
Benchmark_comms_host_device_SOURCES = Benchmark_comms_host_device.cpp
Benchmark_dwf_fp32_SOURCES = Benchmark_dwf_fp32.cpp
Benchmark_ITT_SOURCES = Benchmark_ITT.cpp
Benchmark_IO_SOURCES = Benchmark_IO.cpp

6
Grid/bootstrap.sh Executable file
View File

@ -0,0 +1,6 @@
#!/usr/bin/env bash
set -euo pipefail
mkdir -p .buildutils/m4
autoreconf -fvi

58
Grid/configure.ac Normal file
View File

@ -0,0 +1,58 @@
AC_PREREQ([2.69])
AC_INIT([lattice-bench], [0.1], [antonin.portelli@me.com])
AC_CANONICAL_BUILD
AC_CANONICAL_HOST
AC_CANONICAL_TARGET
AC_CONFIG_SRCDIR([Benchmark_ITT.cpp])
AC_CONFIG_MACRO_DIR([.buildutils/m4])
AC_CONFIG_HEADERS([config.h])
AM_INIT_AUTOMAKE([-Wall -Werror foreign])
m4_ifdef([AM_SILENT_RULES],[AM_SILENT_RULES([yes])])
# Checks for programs.
AC_PROG_CXX
AC_PROG_CC
AC_PROG_RANLIB
AM_PROG_AR
AC_LANG([C++])
AC_ARG_WITH([grid],
[AS_HELP_STRING([--with-grid=<prefix>],
[try this for a non-standard install prefix of Grid])],
[PATH="$with_grid/bin$PATH_SEPARATOR$PATH"]
[CXXFLAGS="$CXXFLAGS -I$with_grid/include"]
[LDFLAGS="$LDFLAGS -L$with_grid/lib"])
AC_CHECK_PROG([GRIDCONF],[grid-config],[yes])
if test x"$GRIDCONF" != x"yes" ; then
AC_MSG_ERROR([grid-config not found])
fi
CXXFLAGS="$CXXFLAGS `grid-config --cxxflags`"
LDFLAGS="$LDFLAGS `grid-config --ldflags`"
CXXFLAGS="$AM_CXXFLAGS $CXXFLAGS"
LDFLAGS="$AM_LDFLAGS $LDFLAGS"
LIBS=" -lGrid $LIBS `grid-config --libs`"
AC_MSG_CHECKING([that a minimal Grid program compiles]);
AC_LINK_IFELSE(
[AC_LANG_SOURCE([[
#include <Grid/Grid.h>
using namespace Grid;
int main(int argc, char *argv[])
{
Grid_init(&argc, &argv);
Grid_finalize();
return 0;
}
]])],
[AC_MSG_RESULT([yes])],
[AC_MSG_RESULT([no])]
[AC_MSG_ERROR([Could not compile a minimal Grid program])])
AC_SUBST([AM_CXXFLAGS])
AC_SUBST([AM_LDFLAGS])
AC_CONFIG_FILES([Makefile])
AC_OUTPUT