portelli/lattice-benchmarks
3
1
Fork 2
Code Issues Pull Requests 2 Packages Projects Releases Wiki Activity

Grid column formatting

Browse Source
This commit is contained in:
Antonin Portelli 2023-01-25 17:45:30 +00:00
parent 514d81d351
commit 03badefec7
5 changed files with 214 additions and 193 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/.clang-format
View File

@ -6,7 +6,7 @@
BreakBeforeBraces: Allman, BreakBeforeBraces: Allman,
AllowShortIfStatementsOnASingleLine: false, AllowShortIfStatementsOnASingleLine: false,
IndentCaseLabels: false, IndentCaseLabels: false,
ColumnLimit: 0, ColumnLimit: 90,
AccessModifierOffset: -4, AccessModifierOffset: -4,
NamespaceIndentation: All, NamespaceIndentation: All,
FixNamespaceComments: false, FixNamespaceComments: false,

133
Grid/Benchmark_Grid.cpp
View File

@ -122,87 +122,82 @@ class Benchmark
for (int lat = 16; lat <= maxlat; lat += 8) 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);
uint64_t bytes = lat * lat * lat * Ls * sizeof(HalfSpinColourVectorD);
for (int d = 0; d < 8; d++)
{ {
int Ls = 12; 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));
}
Coordinate latt_size({lat * mpi_layout[0], double dbytes;
lat * mpi_layout[1],
lat * mpi_layout[2],
lat * mpi_layout[3]});
GridCartesian Grid(latt_size, simd_layout, mpi_layout); for (int dir = 0; dir < 8; dir++)
RealD Nrank = Grid._Nprocessors; {
RealD Nnode = Grid.NodeCount(); int mu = dir % 4;
RealD ppn = Nrank / Nnode; if (mpi_layout[mu] > 1)
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; std::vector<double> times(Nloop);
double dbytes; for (int i = 0; i < Nloop; i++)
for (int dir = 0; dir < 8; dir++)
{
int mu = dir % 4;
if (mpi_layout[mu] > 1)
{ {
std::vector<double> times(Nloop); dbytes = 0;
for (int i = 0; i < Nloop; i++) double start = usecond();
int xmit_to_rank;
int recv_from_rank;
if (dir == mu)
{ {
int comm_proc = 1;
dbytes = 0; Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
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); 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;
dbytes = dbytes * ppn; double stop = usecond();
double xbytes = dbytes * 0.5; t_time[i] = stop - start; // microseconds
double bidibytes = dbytes;
std::cout << GridLogMessage << lat << "\t" << Ls << "\t "
<< bytes << " \t "
<< xbytes / timestat.mean << " \t " << xbytes * timestat.err / (timestat.mean * timestat.mean) << " \t "
<< xbytes / timestat.max << " " << xbytes / timestat.min
<< "\t\t" << bidibytes / timestat.mean << " " << bidibytes * timestat.err / (timestat.mean * timestat.mean) << " "
<< bidibytes / timestat.max << " " << bidibytes / timestat.min << std::endl;
} }
timestat.statistics(t_time);
dbytes = dbytes * ppn;
double xbytes = dbytes * 0.5;
double 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++) }
{ for (int d = 0; d < 8; d++)
acceleratorFreeDevice(xbuf[d]); {
acceleratorFreeDevice(rbuf[d]); acceleratorFreeDevice(xbuf[d]);
} acceleratorFreeDevice(rbuf[d]);
} }
} }
return; return;

86
Grid/Benchmark_IO.cpp
View File

@ -32,23 +32,13 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
#ifdef HAVE_LIME #ifdef HAVE_LIME
using namespace Grid; using namespace Grid;
std::string filestem(const int l) std::string filestem(const int l) { return "iobench_l" + std::to_string(l); }
{
return "iobench_l" + std::to_string(l);
}
int vol(const int i) int vol(const int i) { return BENCH_IO_LMIN + 2 * i; }
{
return BENCH_IO_LMIN + 2 * i;
}
int volInd(const int l) int volInd(const int l) { return (l - BENCH_IO_LMIN) / 2; }
{
return (l - BENCH_IO_LMIN) / 2;
}
template <typename Mat> template <typename Mat> void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
{ {
auto nr = data[0].rows(), nc = data[0].cols(); auto nr = data[0].rows(), nc = data[0].cols();
Eigen::MatrixXd sqSum(nr, nc); Eigen::MatrixXd sqSum(nr, nc);
@ -66,11 +56,11 @@ void stats(Mat &mean, Mat &stdDev, const std::vector<Mat> &data)
mean /= n; mean /= n;
} }
#define grid_printf(...) \ #define grid_printf(...) \
{ \ { \
char _buf[1024]; \ char _buf[1024]; \
sprintf(_buf, __VA_ARGS__); \ sprintf(_buf, __VA_ARGS__); \
MSG << _buf; \ MSG << _buf; \
} }
enum enum
@ -173,47 +163,49 @@ int main(int argc, char **argv)
MSG << "SUMMARY" << std::endl; MSG << "SUMMARY" << std::endl;
MSG << BIGSEP << std::endl; MSG << BIGSEP << std::endl;
MSG << "Summary of individual results (all results in MB/s)." << 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 << "Every second colum gives the standard deviation of the previous column."
<< std::endl;
MSG << std::endl; MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n", grid_printf("%4s %12s %12s %12s %12s %12s %12s %12s %12s\n", "L", "std read", "std dev",
"L", "std read", "std dev", "std write", "std dev", "std write", "std dev", "Grid read", "std dev", "Grid write", "std dev");
"Grid read", "std dev", "Grid write", "std dev");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2) 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", grid_printf("%4d %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n", l,
l, mean(volInd(l), sRead), stdDev(volInd(l), sRead), mean(volInd(l), sRead), stdDev(volInd(l), sRead), mean(volInd(l), sWrite),
mean(volInd(l), sWrite), stdDev(volInd(l), sWrite), stdDev(volInd(l), sWrite), mean(volInd(l), gRead),
mean(volInd(l), gRead), stdDev(volInd(l), gRead), stdDev(volInd(l), gRead), mean(volInd(l), gWrite),
mean(volInd(l), gWrite), stdDev(volInd(l), gWrite)); stdDev(volInd(l), gWrite));
} }
MSG << std::endl; MSG << std::endl;
MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)" << std::endl; MSG << "Robustness of individual results, in %. (rob = 100% - std dev / mean)"
<< std::endl;
MSG << std::endl; MSG << std::endl;
grid_printf("%4s %12s %12s %12s %12s\n", grid_printf("%4s %12s %12s %12s %12s\n", "L", "std read", "std write", "Grid read",
"L", "std read", "std write", "Grid read", "Grid write"); "Grid write");
for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2) for (int l = BENCH_IO_LMIN; l <= BENCH_IO_LMAX; l += 2)
{ {
grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n", grid_printf("%4d %12.1f %12.1f %12.1f %12.1f\n", l, rob(volInd(l), sRead),
l, rob(volInd(l), sRead), rob(volInd(l), sWrite), rob(volInd(l), sWrite), rob(volInd(l), gRead), rob(volInd(l), gWrite));
rob(volInd(l), gRead), rob(volInd(l), gWrite));
} }
MSG << std::endl; 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 << "Summary of results averaged over local volumes 24^4-" << BENCH_IO_LMAX
MSG << "Every second colum gives the standard deviation of the previous column." << std::endl; << "^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; MSG << std::endl;
grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n", grid_printf("%12s %12s %12s %12s %12s %12s %12s %12s\n", "std read", "std dev",
"std read", "std dev", "std write", "std dev", "std write", "std dev", "Grid read", "std dev", "Grid 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),
grid_printf("%12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f %12.1f\n", avStdDev(sRead), avMean(sWrite), avStdDev(sWrite), avMean(gRead),
avMean(sRead), avStdDev(sRead), avMean(sWrite), avStdDev(sWrite), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
avMean(gRead), avStdDev(gRead), avMean(gWrite), avStdDev(gWrite));
MSG << std::endl; MSG << std::endl;
MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)" << std::endl; MSG << "Robustness of volume-averaged results, in %. (rob = 100% - std dev / mean)"
<< std::endl;
MSG << std::endl; MSG << std::endl;
grid_printf("%12s %12s %12s %12s\n", grid_printf("%12s %12s %12s %12s\n", "std read", "std write", "Grid read",
"std read", "std write", "Grid read", "Grid write"); "Grid write");
grid_printf("%12.1f %12.1f %12.1f %12.1f\n", grid_printf("%12.1f %12.1f %12.1f %12.1f\n", avRob(sRead), avRob(sWrite), avRob(gRead),
avRob(sRead), avRob(sWrite), avRob(gRead), avRob(gWrite)); avRob(gWrite));
Grid_finalize(); Grid_finalize();

92
Grid/Benchmark_comms_host_device.cpp
View File

@ -34,8 +34,7 @@ struct time_statistics
mean = sum / v.size(); mean = sum / v.size();
std::vector<double> diff(v.size()); std::vector<double> diff(v.size());
std::transform(v.begin(), v.end(), diff.begin(), [=](double x) std::transform(v.begin(), v.end(), diff.begin(), [=](double x) { return x - mean; });
{ return x - mean; });
double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0); double sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), 0.0);
err = std::sqrt(sq_sum / (v.size() * (v.size() - 1))); err = std::sqrt(sq_sum / (v.size() * (v.size() - 1)));
@ -50,8 +49,7 @@ void header()
std::cout << GridLogMessage << " L " std::cout << GridLogMessage << " L "
<< "\t" << "\t"
<< " Ls " << " Ls "
<< "\t" << "\t" << std::setw(11) << "bytes\t\t"
<< std::setw(11) << "bytes\t\t"
<< "MB/s uni" << "MB/s uni"
<< "\t" << "\t"
<< "MB/s bidi" << std::endl; << "MB/s bidi" << std::endl;
@ -64,7 +62,8 @@ int main(int argc, char **argv)
Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd()); Coordinate simd_layout = GridDefaultSimd(Nd, vComplexD::Nsimd());
Coordinate mpi_layout = GridDefaultMpi(); Coordinate mpi_layout = GridDefaultMpi();
int threads = GridThread::GetThreads(); 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 Nloop = 250;
int nmu = 0; int nmu = 0;
@ -73,13 +72,21 @@ int main(int argc, char **argv)
if (mpi_layout[mu] > 1) if (mpi_layout[mu] > 1)
nmu++; 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); std::vector<double> t_time(Nloop);
// time_statistics timestat; // time_statistics timestat;
std::cout << GridLogMessage << "====================================================================================================" << std::endl; std::cout << GridLogMessage
std::cout << GridLogMessage << "= Benchmarking sequential halo exchange from host memory " << std::endl; << "========================================================================="
std::cout << GridLogMessage << "====================================================================================================" << std::endl; "==========================="
<< std::endl;
std::cout << GridLogMessage
<< "= Benchmarking sequential halo exchange from host memory " << std::endl;
std::cout << GridLogMessage
<< "========================================================================="
"==========================="
<< std::endl;
header(); header();
for (int lat = 8; lat <= maxlat; lat += 4) for (int lat = 8; lat <= maxlat; lat += 4)
@ -87,9 +94,7 @@ int main(int argc, char **argv)
for (int Ls = 8; Ls <= 8; Ls *= 2) for (int Ls = 8; Ls <= 8; Ls *= 2)
{ {
Coordinate latt_size({lat * mpi_layout[0], Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2],
lat * mpi_layout[1],
lat * mpi_layout[2],
lat * mpi_layout[3]}); lat * mpi_layout[3]});
GridCartesian Grid(latt_size, simd_layout, mpi_layout); GridCartesian Grid(latt_size, simd_layout, mpi_layout);
@ -127,22 +132,16 @@ int main(int argc, char **argv)
{ {
std::vector<CommsRequest_t> requests; std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank); Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0], Grid.SendToRecvFrom((void *)&xbuf[mu][0], xmit_to_rank,
xmit_to_rank, (void *)&rbuf[mu][0], recv_from_rank, bytes);
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
} }
comm_proc = mpi_layout[mu] - 1; comm_proc = mpi_layout[mu] - 1;
{ {
std::vector<CommsRequest_t> requests; std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank); Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], xmit_to_rank,
xmit_to_rank, (void *)&rbuf[mu + 4][0], recv_from_rank, bytes);
(void *)&rbuf[mu + 4][0],
recv_from_rank,
bytes);
} }
} }
Grid.Barrier(); Grid.Barrier();
@ -154,17 +153,24 @@ int main(int argc, char **argv)
double bidibytes = xbytes + rbytes; double bidibytes = xbytes + rbytes;
std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t" std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t"
<< std::setw(11) << bytes << std::fixed << std::setprecision(1) << std::setw(7) << " " << std::setw(11) << bytes << std::fixed << std::setprecision(1)
<< std::right << xbytes / mean << " " << std::setw(7) << " " << std::right << xbytes / mean << " "
<< "\t\t" << std::setw(7) << bidibytes / mean << std::endl; << "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
} }
} }
} }
} }
std::cout << GridLogMessage << "====================================================================================================" << std::endl; std::cout << GridLogMessage
std::cout << GridLogMessage << "= Benchmarking sequential halo exchange from GPU memory " << std::endl; << "========================================================================="
std::cout << GridLogMessage << "====================================================================================================" << std::endl; "==========================="
<< std::endl;
std::cout << GridLogMessage
<< "= Benchmarking sequential halo exchange from GPU memory " << std::endl;
std::cout << GridLogMessage
<< "========================================================================="
"==========================="
<< std::endl;
header(); header();
for (int lat = 8; lat <= maxlat; lat += 4) for (int lat = 8; lat <= maxlat; lat += 4)
@ -172,9 +178,7 @@ int main(int argc, char **argv)
for (int Ls = 8; Ls <= 8; Ls *= 2) for (int Ls = 8; Ls <= 8; Ls *= 2)
{ {
Coordinate latt_size({lat * mpi_layout[0], Coordinate latt_size({lat * mpi_layout[0], lat * mpi_layout[1], lat * mpi_layout[2],
lat * mpi_layout[1],
lat * mpi_layout[2],
lat * mpi_layout[3]}); lat * mpi_layout[3]});
GridCartesian Grid(latt_size, simd_layout, mpi_layout); GridCartesian Grid(latt_size, simd_layout, mpi_layout);
@ -212,22 +216,16 @@ int main(int argc, char **argv)
{ {
std::vector<CommsRequest_t> requests; std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank); Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu][0], Grid.SendToRecvFrom((void *)&xbuf[mu][0], xmit_to_rank,
xmit_to_rank, (void *)&rbuf[mu][0], recv_from_rank, bytes);
(void *)&rbuf[mu][0],
recv_from_rank,
bytes);
} }
comm_proc = mpi_layout[mu] - 1; comm_proc = mpi_layout[mu] - 1;
{ {
std::vector<CommsRequest_t> requests; std::vector<CommsRequest_t> requests;
Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank); Grid.ShiftedRanks(mu, comm_proc, xmit_to_rank, recv_from_rank);
Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], Grid.SendToRecvFrom((void *)&xbuf[mu + 4][0], xmit_to_rank,
xmit_to_rank, (void *)&rbuf[mu + 4][0], recv_from_rank, bytes);
(void *)&rbuf[mu + 4][0],
recv_from_rank,
bytes);
} }
} }
Grid.Barrier(); Grid.Barrier();
@ -239,8 +237,8 @@ int main(int argc, char **argv)
double bidibytes = xbytes + rbytes; double bidibytes = xbytes + rbytes;
std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t" std::cout << GridLogMessage << std::setw(4) << lat << "\t" << Ls << "\t"
<< std::setw(11) << bytes << std::fixed << std::setprecision(1) << std::setw(7) << " " << std::setw(11) << bytes << std::fixed << std::setprecision(1)
<< std::right << xbytes / mean << " " << std::setw(7) << " " << std::right << xbytes / mean << " "
<< "\t\t" << std::setw(7) << bidibytes / mean << std::endl; << "\t\t" << std::setw(7) << bidibytes / mean << std::endl;
} }
} }
@ -253,9 +251,15 @@ int main(int argc, char **argv)
} }
} }
std::cout << GridLogMessage << "====================================================================================================" << std::endl; std::cout << GridLogMessage
<< "========================================================================="
"==========================="
<< std::endl;
std::cout << GridLogMessage << "= All done; Bye Bye" << std::endl; std::cout << GridLogMessage << "= All done; Bye Bye" << std::endl;
std::cout << GridLogMessage << "====================================================================================================" << std::endl; std::cout << GridLogMessage
<< "========================================================================="
"==========================="
<< std::endl;
Grid_finalize(); Grid_finalize();
} }

94
Grid/Benchmark_dwf_fp32.cpp
View File

@ -28,17 +28,13 @@ along with this program. If not, see <http://www.gnu.org/licenses/>.
using namespace std; using namespace std;
using namespace Grid; using namespace Grid;
template <class d> template <class d> struct scal
struct scal
{ {
d internal; d internal;
}; };
Gamma::Algebra Gmu[] = { Gamma::Algebra Gmu[] = {Gamma::Algebra::GammaX, Gamma::Algebra::GammaY,
Gamma::Algebra::GammaX, Gamma::Algebra::GammaZ, Gamma::Algebra::GammaT};
Gamma::Algebra::GammaY,
Gamma::Algebra::GammaZ,
Gamma::Algebra::GammaT};
int main(int argc, char **argv) int main(int argc, char **argv)
{ {
@ -59,13 +55,15 @@ int main(int argc, char **argv)
long unsigned int single_site_flops = 8 * Nc * (7 + 16 * Nc); long unsigned int single_site_flops = 8 * Nc * (7 + 16 * Nc);
GridCartesian *UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi()); GridCartesian *UGrid = SpaceTimeGrid::makeFourDimGrid(
GridDefaultLatt(), GridDefaultSimd(Nd, vComplexF::Nsimd()), GridDefaultMpi());
GridRedBlackCartesian *UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid); GridRedBlackCartesian *UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian *FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid); GridCartesian *FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls, UGrid);
GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGrid); GridRedBlackCartesian *FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls, UGrid);
std::cout << GridLogMessage << "Making s innermost grids" << std::endl; std::cout << GridLogMessage << "Making s innermost grids" << std::endl;
GridCartesian *sUGrid = SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(), GridDefaultMpi()); GridCartesian *sUGrid =
SpaceTimeGrid::makeFourDimDWFGrid(GridDefaultLatt(), GridDefaultMpi());
GridRedBlackCartesian *sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid); GridRedBlackCartesian *sUrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(sUGrid);
GridCartesian *sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls, UGrid); GridCartesian *sFGrid = SpaceTimeGrid::makeFiveDimDWFGrid(Ls, UGrid);
GridRedBlackCartesian *sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls, UGrid); GridRedBlackCartesian *sFrbGrid = SpaceTimeGrid::makeFiveDimDWFRedBlackGrid(Ls, UGrid);
@ -177,13 +175,24 @@ int main(int argc, char **argv)
RealD NP = UGrid->_Nprocessors; RealD NP = UGrid->_Nprocessors;
RealD NN = UGrid->NodeCount(); RealD NN = UGrid->NodeCount();
std::cout << GridLogMessage << "*****************************************************************" << std::endl; std::cout << GridLogMessage
std::cout << GridLogMessage << "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm" << std::endl; << "*****************************************************************"
std::cout << GridLogMessage << "*****************************************************************" << std::endl; << std::endl;
std::cout << GridLogMessage << "*****************************************************************" << std::endl; std::cout << GridLogMessage
std::cout << GridLogMessage << "* Benchmarking DomainWallFermionR::Dhop " << std::endl; << "* Kernel options --dslash-generic, --dslash-unroll, --dslash-asm"
std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd() << std::endl; << std::endl;
std::cout << GridLogMessage << "* VComplexF size is " << sizeof(vComplexF) << " B" << 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) if (sizeof(RealF) == 4)
std::cout << GridLogMessage << "* SINGLE precision " << std::endl; std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
if (sizeof(RealF) == 8) if (sizeof(RealF) == 8)
@ -200,7 +209,9 @@ int main(int argc, char **argv)
std::cout << GridLogMessage << "* Using Nc=3 WilsonKernels" << std::endl; std::cout << GridLogMessage << "* Using Nc=3 WilsonKernels" << std::endl;
if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm) if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm)
std::cout << GridLogMessage << "* Using Asm Nc=3 WilsonKernels" << std::endl; std::cout << GridLogMessage << "* Using Asm Nc=3 WilsonKernels" << std::endl;
std::cout << GridLogMessage << "*****************************************************************" << std::endl; std::cout << GridLogMessage
<< "*****************************************************************"
<< std::endl;
DomainWallFermionF Dw(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5); DomainWallFermionF Dw(Umu, *FGrid, *FrbGrid, *UGrid, *UrbGrid, mass, M5);
int ncall = 300; int ncall = 300;
@ -230,19 +241,29 @@ int main(int argc, char **argv)
auto simdwidth = sizeof(vComplex); auto simdwidth = sizeof(vComplex);
// RF: Nd Wilson * Ls, Nd gauge * Ls, Nc colors // 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.); 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 // 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.); 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 << "Called Dw " << ncall << " times in " << t1 - t0
<< " us" << std::endl;
// std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl; // std::cout<<GridLogMessage << "norm result "<< norm2(result)<<std::endl;
// std::cout<<GridLogMessage << "norm ref "<< norm2(ref)<<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 = " << flops / (t1 - t0) << std::endl;
std::cout << GridLogMessage << "mflop/s per rank = " << flops / (t1 - t0) / NP << std::endl; std::cout << GridLogMessage << "mflop/s per rank = " << flops / (t1 - t0) / NP
std::cout << GridLogMessage << "mflop/s per node = " << flops / (t1 - t0) / NN << std::endl; << std::endl;
std::cout << GridLogMessage << "RF GiB/s (base 2) = " << 1000000. * data_rf / ((t1 - t0)) << std::endl; std::cout << GridLogMessage << "mflop/s per node = " << flops / (t1 - t0) / NN
std::cout << GridLogMessage << "mem GiB/s (base 2) = " << 1000000. * data_mem / ((t1 - t0)) << std::endl; << 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; err = ref - result;
std::cout << GridLogMessage << "norm diff " << norm2(err) << std::endl; std::cout << GridLogMessage << "norm diff " << norm2(err) << std::endl;
// exit(0); // exit(0);
@ -313,7 +334,9 @@ int main(int argc, char **argv)
} }
// dump=1; // dump=1;
Dw.Dhop(src, result, 1); Dw.Dhop(src, result, 1);
std::cout << GridLogMessage << "Compare to naive wilson implementation Dag to verify correctness" << std::endl; std::cout << GridLogMessage
<< "Compare to naive wilson implementation Dag to verify correctness"
<< std::endl;
std::cout << GridLogMessage << "Called DwDag" << std::endl; std::cout << GridLogMessage << "Called DwDag" << std::endl;
std::cout << GridLogMessage << "norm dag result " << norm2(result) << std::endl; std::cout << GridLogMessage << "norm dag result " << norm2(result) << std::endl;
std::cout << GridLogMessage << "norm dag ref " << norm2(ref) << std::endl; std::cout << GridLogMessage << "norm dag ref " << norm2(ref) << std::endl;
@ -333,7 +356,8 @@ int main(int argc, char **argv)
LatticeFermionF r_o(FrbGrid); LatticeFermionF r_o(FrbGrid);
LatticeFermionF r_eo(FGrid); LatticeFermionF r_eo(FGrid);
std::cout << GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec" << std::endl; std::cout << GridLogMessage << "Calling Deo and Doe and //assert Deo+Doe == Dunprec"
<< std::endl;
pickCheckerboard(Even, src_e, src); pickCheckerboard(Even, src_e, src);
pickCheckerboard(Odd, src_o, src); pickCheckerboard(Odd, src_o, src);
@ -341,9 +365,12 @@ int main(int argc, char **argv)
std::cout << GridLogMessage << "src_o" << norm2(src_o) << std::endl; std::cout << GridLogMessage << "src_o" << norm2(src_o) << std::endl;
// S-direction is INNERMOST and takes no part in the parity. // S-direction is INNERMOST and takes no part in the parity.
std::cout << GridLogMessage << "*********************************************************" << std::endl; std::cout << GridLogMessage
std::cout << GridLogMessage << "* Benchmarking DomainWallFermionF::DhopEO " << std::endl; << "*********************************************************" << std::endl;
std::cout << GridLogMessage << "* Vectorising space-time by " << vComplexF::Nsimd() << 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) if (sizeof(RealF) == 4)
std::cout << GridLogMessage << "* SINGLE precision " << std::endl; std::cout << GridLogMessage << "* SINGLE precision " << std::endl;
if (sizeof(RealF) == 8) if (sizeof(RealF) == 8)
@ -360,7 +387,8 @@ int main(int argc, char **argv)
std::cout << GridLogMessage << "* Using Nc=3 WilsonKernels" << std::endl; std::cout << GridLogMessage << "* Using Nc=3 WilsonKernels" << std::endl;
if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm) if (WilsonKernelsStatic::Opt == WilsonKernelsStatic::OptInlineAsm)
std::cout << GridLogMessage << "* Using Asm Nc=3 WilsonKernels" << std::endl; std::cout << GridLogMessage << "* Using Asm Nc=3 WilsonKernels" << std::endl;
std::cout << GridLogMessage << "*********************************************************" << std::endl; std::cout << GridLogMessage
<< "*********************************************************" << std::endl;
{ {
Dw.ZeroCounters(); Dw.ZeroCounters();
FGrid->Barrier(); FGrid->Barrier();
@ -387,8 +415,10 @@ int main(int argc, char **argv)
double flops = (single_site_flops * volume * ncall) / 2.0; 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 = " << 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 rank " << flops / (t1 - t0) / NP
std::cout << GridLogMessage << "Deo mflop/s per node " << flops / (t1 - t0) / NN << std::endl; << std::endl;
std::cout << GridLogMessage << "Deo mflop/s per node " << flops / (t1 - t0) / NN
<< std::endl;
Dw.Report(); Dw.Report();
} }
Dw.DhopEO(src_o, r_e, DaggerNo); Dw.DhopEO(src_o, r_e, DaggerNo);