mirror of
https://github.com/paboyle/Grid.git
synced 2024-11-10 07:55:35 +00:00
Adding extra benchmark
This commit is contained in:
paboyle
parent
c7ba47bdc7
commit
e8dddb1596
174
benchmarks/Benchmark_zmm.cc
Normal file
174
benchmarks/Benchmark_zmm.cc
Normal file
@ -0,0 +1,174 @@
|
||||
/*************************************************************************************
|
||||
|
||||
Grid physics library, www.github.com/paboyle/Grid
|
||||
|
||||
Source file: ./tests/Test_zmm.cc
|
||||
|
||||
Copyright (C) 2015
|
||||
|
||||
Author: paboyle <paboyle@ph.ed.ac.uk>
|
||||
|
||||
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.
|
||||
|
||||
See the full license in the file "LICENSE" in the top level distribution directory
|
||||
*************************************************************************************/
|
||||
/* END LEGAL */
|
||||
#include <Grid.h>
|
||||
#include <PerfCount.h>
|
||||
#include <simd/Intel512wilson.h>
|
||||
|
||||
|
||||
using namespace Grid;
|
||||
using namespace Grid::QCD;
|
||||
|
||||
void ZmulF(void *ptr1,void *ptr2,void *ptr3);
|
||||
void Zmul(void *ptr1,void *ptr2,void *ptr3);
|
||||
void WilsonDslashAvx512(void *ptr1,void *ptr2,void *ptr3);
|
||||
void WilsonDslashAvx512F(void *ptr1,void *ptr2,void *ptr3);
|
||||
void TimesIAvx512F(void *ptr1,void *ptr3);
|
||||
void TimesIAvx512(void *ptr1,void *ptr3);
|
||||
void TimesMinusIAvx512F(void *ptr1,void *ptr3);
|
||||
void TimesMinusIAvx512(void *ptr1,void *ptr3);
|
||||
|
||||
|
||||
int bench(std::ofstream &os, std::vector<int> &latt4,int Ls);
|
||||
|
||||
int main(int argc,char **argv)
|
||||
{
|
||||
Grid_init(&argc,&argv);
|
||||
std::ofstream os("zmm.dat");
|
||||
|
||||
os << "#V Ls Lxy Lzt C++ Asm OMP L1 " <<std::endl;
|
||||
for(int L=4;L<32;L+=2){
|
||||
for(int m=1;m<=2;m++){
|
||||
for(int Ls=8;Ls<=16;Ls+=8){
|
||||
std::vector<int> grid({L,L,m*L,m*L});
|
||||
bench(os,latt4,Ls);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
int bench(std::ofstream &os, std::vector<int> &latt4,int Ls)
|
||||
{
|
||||
|
||||
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(latt4, GridDefaultSimd(Nd,vComplex::Nsimd()),GridDefaultMpi());
|
||||
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
|
||||
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
|
||||
GridRedBlackCartesian * FrbGrid = SpaceTimeGrid::makeFiveDimRedBlackGrid(Ls,UGrid);
|
||||
|
||||
std::vector<int> simd_layout = GridDefaultSimd(Nd,vComplex::Nsimd());
|
||||
std::vector<int> mpi_layout = GridDefaultMpi();
|
||||
int threads = GridThread::GetThreads();
|
||||
|
||||
std::vector<int> seeds4({1,2,3,4});
|
||||
std::vector<int> seeds5({5,6,7,8});
|
||||
|
||||
GridSerialRNG sRNG; sRNG.SeedFixedIntegers(seeds4);
|
||||
|
||||
LatticeFermion src (FGrid);
|
||||
LatticeFermion tmp (FGrid);
|
||||
LatticeFermion srce(FrbGrid);
|
||||
|
||||
LatticeFermion resulto(FrbGrid); resulto=zero;
|
||||
LatticeFermion resulta(FrbGrid); resulta=zero;
|
||||
LatticeFermion junk(FrbGrid); junk=zero;
|
||||
LatticeFermion diff(FrbGrid);
|
||||
LatticeGaugeField Umu(UGrid);
|
||||
|
||||
double mfc, mfa, mfo, mfl1;
|
||||
|
||||
GridParallelRNG RNG4(UGrid); RNG4.SeedFixedIntegers(seeds4);
|
||||
GridParallelRNG RNG5(FGrid); RNG5.SeedFixedIntegers(seeds5);
|
||||
random(RNG5,src);
|
||||
#if 1
|
||||
random(RNG4,Umu);
|
||||
#else
|
||||
int mmu=2;
|
||||
std::vector<LatticeColourMatrix> U(4,UGrid);
|
||||
for(int mu=0;mu<Nd;mu++){
|
||||
U[mu] = PeekIndex<LorentzIndex>(Umu,mu);
|
||||
if ( mu!=mmu ) U[mu] = zero;
|
||||
if ( mu==mmu ) U[mu] = 1.0;
|
||||
PokeIndex<LorentzIndex>(Umu,U[mu],mu);
|
||||
}
|
||||
#endif
|
||||
pickCheckerboard(Even,srce,src);
|
||||
|
||||
RealD mass=0.1;
|
||||
RealD M5 =1.8;
|
||||
DomainWallFermionR Dw(Umu,*FGrid,*FrbGrid,*UGrid,*UrbGrid,mass,M5);
|
||||
|
||||
std::cout<<GridLogMessage << "Calling Dw"<<std::endl;
|
||||
int ncall=50;
|
||||
double t0=usecond();
|
||||
for(int i=0;i<ncall;i++){
|
||||
Dw.DhopOE(srce,resulto,0);
|
||||
}
|
||||
double t1=usecond();
|
||||
|
||||
double volume=Ls; for(int mu=0;mu<Nd;mu++) volume=volume*latt4[mu];
|
||||
double flops=1344*volume/2;
|
||||
|
||||
mfc = flops*ncall/(t1-t0);
|
||||
std::cout<<GridLogMessage << "Called C++ Dw"<< " mflop/s = "<< mfc<<std::endl;
|
||||
|
||||
QCD::WilsonFermion5DStatic::AsmOptDslash=1;
|
||||
t0=usecond();
|
||||
for(int i=0;i<ncall;i++){
|
||||
Dw.DhopOE(srce,resulta,0);
|
||||
}
|
||||
t1=usecond();
|
||||
mfa = flops*ncall/(t1-t0);
|
||||
std::cout<<GridLogMessage << "Called ASM Dw"<< " mflop/s = "<< mfa<<std::endl;
|
||||
|
||||
t0=usecond();
|
||||
for(int i=0;i<1;i++){
|
||||
Dw.DhopInternalOMPbench(Dw.StencilOdd,Dw.LebesgueEvenOdd,Dw.UmuEven,srce,junk,dag);
|
||||
}
|
||||
t1=usecond();
|
||||
mfo = flops*100/(t1-t0);
|
||||
std::cout<<GridLogMessage << "Called ASM-OMP Dw"<< " mflop/s = "<< mfo<<std::endl;
|
||||
|
||||
t0=usecond();
|
||||
for(int i=0;i<1;i++){
|
||||
Dw.DhopInternalL1bench(Dw.StencilOdd,Dw.LebesgueEvenOdd,Dw.UmuEven,srce,junk,dag);
|
||||
}
|
||||
t1=usecond();
|
||||
mfl1= flops*100/(t1-t0);
|
||||
std::cout<<GridLogMessage << "Called ASM-L1 Dw"<< " mflop/s = "<< mfl1<<std::endl;
|
||||
|
||||
os << latt4[0]*latt4[1]*latt4[2]*latt4[3]<< " "<<Ls<<" "<< latt4[0] <<" " <<latt4[2]<< " "
|
||||
<< mfc<<" "
|
||||
<< mfa<<" "
|
||||
<< mfo<<" "
|
||||
<< mfl1<<std::endl;
|
||||
|
||||
for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
|
||||
Dw.DhopOE(srce,resulta,0);
|
||||
PerformanceCounter Counter(i);
|
||||
Counter.Start();
|
||||
Dw.DhopOE(srce,resulta,0);
|
||||
Counter.Stop();
|
||||
Counter.Report();
|
||||
}
|
||||
//resulta = (-0.5) * resulta;
|
||||
|
||||
diff = resulto-resulta;
|
||||
std::cout<<GridLogMessage << "diff "<< norm2(diff)<<std::endl;
|
||||
std::cout<<std::endl;
|
||||
}
|
||||
|
||||
|
||||
@ -450,7 +450,7 @@ void WilsonFermion5D<Impl>::DhopInternalOMPbench(StencilImpl & st, LebesgueOrder
|
||||
|
||||
#pragma omp parallel
|
||||
{
|
||||
for(int jjj=0;jjj<1000;jjj++){
|
||||
for(int jjj=0;jjj<100;jjj++){
|
||||
#pragma omp barrier
|
||||
dslashtime -=usecond();
|
||||
if ( dag == DaggerYes ) {
|
||||
@ -538,6 +538,124 @@ void WilsonFermion5D<Impl>::DhopInternalOMPbench(StencilImpl & st, LebesgueOrder
|
||||
alltime+=usecond();
|
||||
}
|
||||
|
||||
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::DhopInternalL1bench(StencilImpl & st, LebesgueOrder &lo,
|
||||
DoubledGaugeField & U,
|
||||
const FermionField &in, FermionField &out,int dag)
|
||||
{
|
||||
// assert((dag==DaggerNo) ||(dag==DaggerYes));
|
||||
alltime-=usecond();
|
||||
Compressor compressor(dag);
|
||||
|
||||
// Assume balanced KMP_AFFINITY; this is forced in GridThread.h
|
||||
|
||||
int threads = GridThread::GetThreads();
|
||||
int HT = GridThread::GetHyperThreads();
|
||||
int cores = GridThread::GetCores();
|
||||
int nwork = U._grid->oSites();
|
||||
|
||||
commtime -=usecond();
|
||||
auto handle = st.HaloExchangeBegin(in,compressor);
|
||||
st.HaloExchangeComplete(handle);
|
||||
commtime +=usecond();
|
||||
|
||||
jointime -=usecond();
|
||||
jointime +=usecond();
|
||||
|
||||
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
|
||||
// Not loop ordering and data layout.
|
||||
// Designed to create
|
||||
// - per thread reuse in L1 cache for U
|
||||
// - 8 linear access unit stride streams per thread for Fermion for hw prefetchable.
|
||||
|
||||
#pragma omp parallel
|
||||
{
|
||||
for(int jjj=0;jjj<100;jjj++){
|
||||
#pragma omp barrier
|
||||
dslashtime -=usecond();
|
||||
if ( dag == DaggerYes ) {
|
||||
if( this->HandOptDslash ) {
|
||||
#pragma omp for
|
||||
for(int ss=0;ss<U._grid->oSites();ss++){
|
||||
int sU=0;
|
||||
for(int s=0;s<Ls;s++){
|
||||
int sF = s+Ls*sU;
|
||||
Kernels::DiracOptHandDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
|
||||
#pragma omp for
|
||||
for(int ss=0;ss<U._grid->oSites();ss++){
|
||||
{
|
||||
int sd;
|
||||
for(sd=0;sd<Ls;sd++){
|
||||
int sU=0;
|
||||
int sF = sd+Ls*sU;
|
||||
Kernels::DiracOptDhopSiteDag(st,U,st.comm_buf,sF,sU,in,out);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
} else {
|
||||
if( this->AsmOptDslash ) {
|
||||
// for(int i=0;i<1;i++){
|
||||
// for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
|
||||
// PerformanceCounter Counter(i);
|
||||
// Counter.Start();
|
||||
|
||||
#pragma omp for
|
||||
for(int t=0;t<threads;t++){
|
||||
|
||||
int hyperthread = t%HT;
|
||||
int core = t/HT;
|
||||
|
||||
int sswork, swork,soff,ssoff, sU,sF;
|
||||
|
||||
GridThread::GetWork(nwork,core,sswork,ssoff,cores);
|
||||
GridThread::GetWork(Ls , hyperthread, swork, soff,HT);
|
||||
|
||||
for(int ss=0;ss<sswork;ss++){
|
||||
for(int s=soff;s<soff+swork;s++){
|
||||
|
||||
sU=0;
|
||||
sF = s+Ls*sU;
|
||||
Kernels::DiracOptAsmDhopSite(st,U,st.comm_buf,sF,sU,in,out);
|
||||
}
|
||||
}
|
||||
}
|
||||
// Counter.Stop();
|
||||
// Counter.Report();
|
||||
// }
|
||||
} else if( this->HandOptDslash ) {
|
||||
#pragma omp for
|
||||
|
||||
for(int ss=0;ss<U._grid->oSites();ss++){
|
||||
int sU=0;
|
||||
for(int s=0;s<Ls;s++){
|
||||
int sF = s+Ls*sU;
|
||||
Kernels::DiracOptHandDhopSite(st,U,st.comm_buf,sF,sU,in,out);
|
||||
}
|
||||
}
|
||||
} else {
|
||||
#pragma omp for
|
||||
for(int ss=0;ss<U._grid->oSites();ss++){
|
||||
int sU=0;
|
||||
for(int s=0;s<Ls;s++){
|
||||
int sF = s+Ls*sU;
|
||||
Kernels::DiracOptDhopSite(st,U,st.comm_buf,sF,sU,in,out);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
dslashtime +=usecond();
|
||||
alltime+=usecond();
|
||||
}
|
||||
|
||||
|
||||
template<class Impl>
|
||||
void WilsonFermion5D<Impl>::DhopInternalCommsOverlapCompute(StencilImpl & st, LebesgueOrder &lo,
|
||||
DoubledGaugeField & U,
|
||||
|
||||
Loading…
Reference in New Issue
Block a user