portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-08-02 20:57:06 +01:00
Code Releases Activity

Remove Ls Vec cases from benchmarks

Browse Source
This commit is contained in:
Peter Boyle
2019-06-04 20:45:35 +01:00
parent 7e27a5213a
commit 47c063f984
4 changed files with 7 additions and 392 deletions
Download Patch File Download Diff File Expand all files Collapse all files

168
benchmarks/Benchmark_dwf.cc
View File

@@ -38,9 +38,6 @@ struct scal {
Gamma::Algebra::GammaT
};
typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
typedef WilsonFermion5D<DomainWallVec5dImplF> WilsonFermion5DF;
typedef WilsonFermion5D<DomainWallVec5dImplD> WilsonFermion5DD;
int main (int argc, char ** argv)
{
@@ -247,171 +244,6 @@ int main (int argc, char ** argv)
DwH.Report();
}
if (1)
{
std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::Dhop "<<std::endl;
std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
if ( sizeof(Real)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(Real)==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::OptGpu ) std::cout << GridLogMessage<< "* Using Gpu 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;
typedef WilsonFermion5D<DomainWallVec5dImplR> WilsonFermion5DR;
LatticeFermion ssrc(sFGrid);
LatticeFermion sref(sFGrid);
LatticeFermion sresult(sFGrid);
WilsonFermion5DR sDw(Umu,*sFGrid,*sFrbGrid,*sUGrid,*sUrbGrid,M5);
localConvert(src,ssrc);
std::cout<<GridLogMessage<< "src norms "<< norm2(src)<<" " <<norm2(ssrc)<<std::endl;
FGrid->Barrier();
sDw.Dhop(ssrc,sresult,0);
sDw.ZeroCounters();
double t0=usecond();
for(int i=0;i<ncall;i++){
__SSC_START;
sDw.Dhop(ssrc,sresult,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;
std::cout<<GridLogMessage << "Called Dw s_inner "<<ncall<<" times in "<<t1-t0<<" us"<<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<< "res norms "<< norm2(result)<<" " <<norm2(sresult)<<std::endl;
sDw.Report();
RealD sum=0;
err=Zero();
localConvert(sresult,err);
err = err - ref;
sum = norm2(err);
std::cout<<GridLogMessage<<" difference between normal ref and simd is "<<sum<<std::endl;
if(sum > 1.0e-4 ){
std::cout<< "sD REF\n " <<ref << std::endl;
std::cout<< "sD ERR \n " <<err <<std::endl;
}
// assert(sum < 1.0e-4);
err=Zero();
localConvert(sresult,err);
err = err - result;
sum = norm2(err);
std::cout<<GridLogMessage<<" difference between normal result and simd is "<<sum<<std::endl;
if(sum > 1.0e-4 ){
std::cout<< "sD REF\n " <<result << std::endl;
std::cout<< "sD ERR \n " << err <<std::endl;
}
assert(sum < 1.0e-4);
if(1){
std::cout << GridLogMessage<< "*********************************************************" <<std::endl;
std::cout << GridLogMessage<< "* Benchmarking WilsonFermion5D<DomainWallVec5dImplR>::DhopEO "<<std::endl;
std::cout << GridLogMessage<< "* Vectorising fifth dimension by "<<vComplex::Nsimd()<<std::endl;
if ( sizeof(Real)==4 ) std::cout << GridLogMessage<< "* SINGLE precision "<<std::endl;
if ( sizeof(Real)==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::OptGpu ) std::cout << GridLogMessage<< "* Using Gpu 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;
LatticeFermion sr_eo(sFGrid);
LatticeFermion ssrc_e (sFrbGrid);
LatticeFermion ssrc_o (sFrbGrid);
LatticeFermion sr_e (sFrbGrid);
LatticeFermion sr_o (sFrbGrid);
pickCheckerboard(Even,ssrc_e,ssrc);
pickCheckerboard(Odd,ssrc_o,ssrc);
// setCheckerboard(sr_eo,ssrc_o);
// setCheckerboard(sr_eo,ssrc_e);
sr_e = Zero();
sr_o = Zero();
FGrid->Barrier();
sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
sDw.ZeroCounters();
// sDw.stat.init("DhopEO");
double t0=usecond();
for (int i = 0; i < ncall; i++) {
sDw.DhopEO(ssrc_o, sr_e, DaggerNo);
}
double t1=usecond();
FGrid->Barrier();
// sDw.stat.print();
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 << "sDeo mflop/s = "<< flops/(t1-t0)<<std::endl;
std::cout<<GridLogMessage << "sDeo mflop/s per rank "<< flops/(t1-t0)/NP<<std::endl;
std::cout<<GridLogMessage << "sDeo mflop/s per node "<< flops/(t1-t0)/NN<<std::endl;
sDw.Report();
sDw.DhopEO(ssrc_o,sr_e,DaggerNo);
sDw.DhopOE(ssrc_e,sr_o,DaggerNo);
sDw.Dhop (ssrc ,sresult,DaggerNo);
pickCheckerboard(Even,ssrc_e,sresult);
pickCheckerboard(Odd ,ssrc_o,sresult);
ssrc_e = ssrc_e - sr_e;
RealD error = norm2(ssrc_e);
std::cout<<GridLogMessage << "sE norm diff "<< norm2(ssrc_e)<< " vec nrm"<<norm2(sr_e) <<std::endl;
ssrc_o = ssrc_o - sr_o;
error+= norm2(ssrc_o);
std::cout<<GridLogMessage << "sO norm diff "<< norm2(ssrc_o)<< " vec nrm"<<norm2(sr_o) <<std::endl;
if(( error>1.0e-4) ) {
setCheckerboard(ssrc,ssrc_o);
setCheckerboard(ssrc,ssrc_e);
std::cout<< "DIFF\n " <<ssrc << std::endl;
setCheckerboard(ssrc,sr_o);
setCheckerboard(ssrc,sr_e);
std::cout<< "CBRESULT\n " <<ssrc << std::endl;
std::cout<< "RESULT\n " <<sresult<< std::endl;
}
assert(error<1.0e-4);
}
if(0){
std::cout << "Single cache warm call to sDw.Dhop " <<std::endl;
for(int i=0;i< PerformanceCounter::NumTypes(); i++ ){
sDw.Dhop(ssrc,sresult,0);
PerformanceCounter Counter(i);
Counter.Start();
sDw.Dhop(ssrc,sresult,0);
Counter.Stop();
Counter.Report();
}
}
}
if (1)
{ // Naive wilson dag implementation
ref = Zero();