portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Benchmarks modified for new GPU constructs

Browse Source
This commit is contained in:
Peter Boyle 2019-06-15 12:52:56 +01:00
parent 0184719216
commit 0561c2edeb
6 changed files with 61 additions and 84 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
benchmarks/Benchmark_IO_vs_dir.cc
View File

@ -5,7 +5,6 @@
"============================================================================="
using namespace Grid;
using namespace QCD;
int main (int argc, char ** argv)
{

4
benchmarks/Benchmark_ITT.cc
View File

@ -153,8 +153,8 @@ public:
dbytes=0;
ncomm=0;
thread_loop( (int dir=0;dir<8;dir++),{
thread_for(dir,8,{
double tbytes;
int mu =dir % 4;

2
benchmarks/Benchmark_comms.cc
View File

@ -485,7 +485,7 @@ int main (int argc, char ** argv)
dbytes=0;
ncomm=0;
thread_loop( (int dir=0;dir<8;dir++),{
thread_for(dir,8,{
double tbytes;
int mu =dir % 4;

2
benchmarks/Benchmark_memory_asynch.cc
View File

@ -78,7 +78,7 @@ int main (int argc, char ** argv)
}
double start=usecond();
thread_loop( (int t=0;t<threads;t++),{
thread_for(t,threads,{
auto x_t = x[t].View();
sum[t] = x_t[0];
for(int i=0;i<Nloop;i++){

36
benchmarks/Benchmark_meson_field.cc
View File

@ -30,8 +30,6 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
using namespace std;
using namespace Grid;
using namespace Grid::QCD;
#include "Grid/util/Profiling.h"
@ -67,7 +65,7 @@ void sliceInnerProductMesonField(std::vector< std::vector<ComplexD> > &mat,
// sum across these down to scalars
// splitting the SIMD
std::vector<vector_type,alignedAllocator<vector_type> > lvSum(rd*Lblock*Rblock);
thread_loop( (int r = 0; r < rd * Lblock * Rblock; r++),{
thread_for(r, rd * Lblock * Rblock,{
lvSum[r] = Zero();
});
@ -79,7 +77,7 @@ void sliceInnerProductMesonField(std::vector< std::vector<ComplexD> > &mat,
std::cout << GridLogMessage << " Entering first parallel loop "<<std::endl;
// Parallelise over t-direction doesn't expose as much parallelism as needed for KNL
thread_loop((int r=0;r<rd;r++),{
thread_for(r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@ -114,7 +112,7 @@ void sliceInnerProductMesonField(std::vector< std::vector<ComplexD> > &mat,
std::cout << GridLogMessage << " Entering second parallel loop "<<std::endl;
// Sum across simd lanes in the plane, breaking out orthog dir.
thread_loop((int rt=0;rt<rd;rt++),{
thread_for(rt,rd,{
Coordinate icoor(Nd);
@ -199,12 +197,12 @@ void sliceInnerProductMesonFieldGamma(std::vector< std::vector<ComplexD> > &mat,
int MFlvol = ld*Lblock*Rblock*Ngamma;
std::vector<vector_type,alignedAllocator<vector_type> > lvSum(MFrvol);
thread_loop( (int r = 0; r < MFrvol; r++),{
thread_for(r,MFrvol,{
lvSum[r] = Zero();
});
std::vector<scalar_type > lsSum(MFlvol);
thread_loop( (int r = 0; r < MFlvol; r++),{
thread_for( r,MFlvol,{
lsSum[r]=scalar_type(0.0);
});
@ -215,7 +213,7 @@ void sliceInnerProductMesonFieldGamma(std::vector< std::vector<ComplexD> > &mat,
std::cout << GridLogMessage << " Entering first parallel loop "<<std::endl;
// Parallelise over t-direction doesn't expose as much parallelism as needed for KNL
thread_loop((int r=0;r<rd;r++),{
thread_for(r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@ -256,7 +254,7 @@ void sliceInnerProductMesonFieldGamma(std::vector< std::vector<ComplexD> > &mat,
std::cout << GridLogMessage << " Entering second parallel loop "<<std::endl;
// Sum across simd lanes in the plane, breaking out orthog dir.
thread_loop((int rt=0;rt<rd;rt++),{
thread_for(rt,rd,{
iScalar<vector_type> temp;
Coordinate icoor(Nd);
@ -347,12 +345,12 @@ void sliceInnerProductMesonFieldGamma1(std::vector< std::vector<ComplexD> > &mat
int MFlvol = ld*Lblock*Rblock;
Vector<SpinMatrix_v > lvSum(MFrvol);
thread_loop( (int r = 0; r < MFrvol; r++),{
thread_for(r,MFrvol,{
lvSum[r] = Zero();
});
Vector<SpinMatrix_s > lsSum(MFlvol);
thread_loop( (int r = 0; r < MFlvol; r++),{
thread_for(r,MFlvol,{
lsSum[r]=scalar_type(0.0);
});
@ -363,7 +361,7 @@ void sliceInnerProductMesonFieldGamma1(std::vector< std::vector<ComplexD> > &mat
std::cout << GridLogMessage << " Entering first parallel loop "<<std::endl;
// Parallelise over t-direction doesn't expose as much parallelism as needed for KNL
thread_loop((int r=0;r<rd;r++),{
thread_for(r,rd,{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@ -398,7 +396,7 @@ void sliceInnerProductMesonFieldGamma1(std::vector< std::vector<ComplexD> > &mat
std::cout << GridLogMessage << " Entering second parallel loop "<<std::endl;
// Sum across simd lanes in the plane, breaking out orthog dir.
thread_loop((int rt=0;rt<rd;rt++),{
thread_for(rt,rd,{
Coordinate icoor(Nd);
ExtractBuffer<SpinMatrix_s> extracted(Nsimd);
@ -425,7 +423,7 @@ void sliceInnerProductMesonFieldGamma1(std::vector< std::vector<ComplexD> > &mat
});
std::cout << GridLogMessage << " Entering third parallel loop "<<std::endl;
thread_loop((int t=0;t<fd;t++)
thread_for(t,fd,
{
int pt = t / ld; // processor plane
int lt = t % ld;
@ -490,13 +488,13 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
int MFlvol = ld*Lblock*Rblock*Nmom;
Vector<SpinMatrix_v > lvSum(MFrvol);
thread_loop( (int r = 0; r < MFrvol; r++),
thread_for(r,MFrvol,
{
lvSum[r] = Zero();
});
Vector<SpinMatrix_s > lsSum(MFlvol);
thread_loop( (int r = 0; r < MFlvol; r++),
thread_for(r,MFlvol,
{
lsSum[r]=scalar_type(0.0);
});
@ -508,7 +506,7 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
std::cout << GridLogMessage << " Entering first parallel loop "<<std::endl;
// Parallelise over t-direction doesn't expose as much parallelism as needed for KNL
thread_loop((int r=0;r<rd;r++),
thread_for(r,rd,
{
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
@ -552,7 +550,7 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
std::cout << GridLogMessage << " Entering second parallel loop "<<std::endl;
// Sum across simd lanes in the plane, breaking out orthog dir.
thread_loop((int rt=0;rt<rd;rt++),
thread_for(rt,rd,
{
Coordinate icoor(Nd);
@ -582,7 +580,7 @@ void sliceInnerProductMesonFieldGammaMom(std::vector< std::vector<ComplexD> > &m
});
std::cout << GridLogMessage << " Entering third parallel loop "<<std::endl;
thread_loop((int t=0;t<fd;t++),
thread_for(t,fd,
{
int pt = t / ld; // processor plane
int lt = t % ld;

100
benchmarks/Benchmark_su3_gpu.cc
View File

@ -71,26 +71,21 @@ int main (int argc, char ** argv)
auto z_v = z.View();
const uint64_t Nsite = x_v.size();
const uint64_t nsimd = vComplex::Nsimd();
const uint64_t NN = Nsite*nsimd;
for(int64_t i=0;i<Nwarm;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = xx*yy;
insertLane(lane,z_v[ss],zz);
coalescedWrite(z_v[ss],zz);
});
}
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = xx*yy;
insertLane(lane,z_v[ss],zz);
coalescedWrite(z_v[ss],zz);
});
}
double stop=usecond();
@ -126,26 +121,21 @@ int main (int argc, char ** argv)
auto z_v = z.View();
const uint64_t Nsite = x_v.size();
const uint64_t nsimd = vComplex::Nsimd();
const uint64_t NN = Nsite*nsimd;
for(int64_t i=0;i<Nwarm;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = xx*yy;
insertLane(lane,x_v[ss],zz);
coalescedWrite(x_v[ss],zz);
});
}
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = xx*yy;
insertLane(lane,x_v[ss],zz);
coalescedWrite(x_v[ss],zz);
});
}
double stop=usecond();
@ -182,28 +172,23 @@ int main (int argc, char ** argv)
auto z_v = z.View();
const uint64_t Nsite = x_v.size();
const uint64_t nsimd = vComplex::Nsimd();
const uint64_t NN = Nsite*nsimd;
for(int64_t i=0;i<Nwarm;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
auto zz = extractLane(lane,z_v[ss]);
zz = zz + xx * yy;
insertLane(lane,z_v[ss],zz);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = coalescedRead(z_v[ss]);
zz = zz+xx*yy;
coalescedWrite(z_v[ss],zz);
});
}
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
auto zz = extractLane(lane,z_v[ss]);
zz = zz + xx * yy;
insertLane(lane,x_v[ss],zz);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = coalescedRead(z_v[ss]);
zz = zz+xx*yy;
coalescedWrite(z_v[ss],zz);
});
}
double stop=usecond();
@ -241,28 +226,23 @@ int main (int argc, char ** argv)
auto w_v = z.View();
const uint64_t Nsite = x_v.size();
const uint64_t nsimd = vComplex::Nsimd();
const uint64_t NN = Nsite*nsimd;
for(int64_t i=0;i<Nwarm;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
auto zz = extractLane(lane,z_v[ss]);
auto ww = zz + xx * yy;
insertLane(lane,w_v[ss],ww);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = coalescedRead(z_v[ss]);
auto ww = zz+xx*yy;
coalescedWrite(w_v[ss],ww);
});
}
double start=usecond();
for(int64_t i=0;i<Nloop;i++){
accelerator_loopN( sss, NN, {
uint64_t lane = sss % nsimd;
uint64_t ss = sss / nsimd;
auto xx = extractLane(lane,x_v[ss]);
auto yy = extractLane(lane,y_v[ss]);
auto zz = extractLane(lane,z_v[ss]);
auto ww = zz + xx * yy;
insertLane(lane,w_v[ss],ww);
accelerator_for( ss, Nsite, nsimd ,{
auto xx = coalescedRead(x_v[ss]);
auto yy = coalescedRead(y_v[ss]);
auto zz = coalescedRead(z_v[ss]);
auto ww = zz+xx*yy;
coalescedWrite(w_v[ss],ww);
});
}
double stop=usecond();