portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-11-04 05:54:32 +00:00
Code Releases Activity

Reproducing CG can be more useful now

Browse Source
This commit is contained in:
Peter Boyle
2024-03-06 00:03:16 +00:00
parent 517822fdd2
commit 9b5f741e85
1 changed files with 58 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

111
tests/Test_dwf_mixedcg_prec.cc
View File

@@ -30,27 +30,16 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
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)
{
char hostname[HOST_NAME_MAX+1];
gethostname(hostname, HOST_NAME_MAX+1);
std::string host(hostname);
Grid_init(&argc,&argv);
const int Ls=12;
std::cout << GridLogMessage << "::::: NB: to enable a quick bit reproducibility check use the --checksums flag. " << std::endl;
{
GridCartesian * UGrid = SpaceTimeGrid::makeFourDimGrid(GridDefaultLatt(), GridDefaultSimd(Nd,vComplexD::Nsimd()),GridDefaultMpi());
GridRedBlackCartesian * UrbGrid = SpaceTimeGrid::makeFourDimRedBlackGrid(UGrid);
GridCartesian * FGrid = SpaceTimeGrid::makeFiveDimGrid(Ls,UGrid);
@@ -92,7 +81,14 @@ int main (int argc, char ** argv)
SchurDiagMooeeOperator<DomainWallFermionD,LatticeFermionD> HermOpEO(Ddwf);
SchurDiagMooeeOperator<DomainWallFermionF,LatticeFermionF> HermOpEO_f(Ddwf_f);
std::cout << GridLogMessage << "::::::::::::: Starting mixed CG" << std::endl;
int nsecs=600;
if( GridCmdOptionExists(argv,argv+argc,"--seconds") ){
std::string arg = GridCmdOptionPayload(argv,argv+argc,"--seconds");
GridCmdOptionInt(arg,nsecs);
}
std::cout << GridLogMessage << "::::::::::::: Starting mixed CG for "<<nsecs <<" seconds" << std::endl;
MixedPrecisionConjugateGradient<LatticeFermionD,LatticeFermionF> mCG(1.0e-8, 10000, 50, FrbGrid_f, HermOpEO_f, HermOpEO);
double t1,t2,flops;
double MdagMsiteflops = 1452; // Mobius (real coeffs)
@@ -101,7 +97,14 @@ int main (int argc, char ** argv)
std:: cout << " MdagM site flops = "<< 4*MdagMsiteflops<<std::endl;
std:: cout << " CG site flops = "<< CGsiteflops <<std::endl;
int iters;
for(int i=0;i<10;i++){
time_t start = time(NULL);
uint32_t csum, csumref;
csumref=0;
int iter=0;
do {
std::cerr << "******************* SINGLE PRECISION SOLVE "<<iter<<std::endl;
result_o = Zero();
t1=usecond();
mCG(src_o,result_o);
@@ -111,10 +114,28 @@ int main (int argc, char ** argv)
flops+= CGsiteflops*FrbGrid->gSites()*iters;
std::cout << " SinglePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
std::cout << " SinglePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
}
std::cout << GridLogMessage << "::::::::::::: Starting regular CG" << std::endl;
csum = crc(result_o);
if ( csumref == 0 ) {
csumref = csum;
} else {
if ( csum != csumref ) {
std::cerr << host<<" FAILURE " <<iter <<" csum "<<std::hex<<csum<< " != "<<csumref <<std::dec<<std::endl;
assert(0);
} else {
std::cout << host <<" OK " <<iter <<" csum "<<std::hex<<csum<<std::dec<<" -- OK! "<<std::endl;
}
}
iter ++;
} while (time(NULL) < (start + nsecs/2) );
std::cout << GridLogMessage << "::::::::::::: Starting double precision CG" << std::endl;
ConjugateGradient<LatticeFermionD> CG(1.0e-8,10000);
for(int i=0;i<1;i++){
csumref=0;
int i=0;
do {
std::cerr << "******************* DOUBLE PRECISION SOLVE "<<i<<std::endl;
result_o_2 = Zero();
t1=usecond();
CG(HermOpEO,src_o,result_o_2);
@@ -122,46 +143,30 @@ int main (int argc, char ** argv)
iters = CG.IterationsToComplete;
flops = MdagMsiteflops*4*FrbGrid->gSites()*iters;
flops+= CGsiteflops*FrbGrid->gSites()*iters;
std::cout << " DoublePrecision iterations/sec "<< iters/(t2-t1)*1000.*1000.<<std::endl;
std::cout << " DoublePrecision GF/s "<< flops/(t2-t1)/1000.<<std::endl;
}
// MemoryManager::Print();
csum = crc(result_o);
if ( csumref == 0 ) {
csumref = csum;
} else {
if ( csum != csumref ) {
std::cerr << i <<" csum "<<std::hex<<csum<< " != "<<csumref <<std::dec<<std::endl;
assert(0);
} else {
std::cout << i <<" csum "<<std::hex<<csum<<std::dec<<" -- OK! "<<std::endl;
}
}
i++;
} while (time(NULL) < (start + nsecs) );
LatticeFermionD diff_o(FrbGrid);
RealD diff = axpy_norm(diff_o, -1.0, result_o, result_o_2);
std::cout << GridLogMessage << "::::::::::::: Diff between mixed and regular CG: " << diff << std::endl;
#ifdef HAVE_LIME
if( GridCmdOptionExists(argv,argv+argc,"--checksums") ){
assert(diff < 1e-4);
std::string file1("./Propagator1");
emptyUserRecord record;
uint32_t nersc_csum;
uint32_t scidac_csuma;
uint32_t scidac_csumb;
typedef SpinColourVectorD FermionD;
typedef vSpinColourVectorD vFermionD;
BinarySimpleMunger<FermionD,FermionD> munge;
std::string format = getFormatString<vFermionD>();
BinaryIO::writeLatticeObject<vFermionD,FermionD>(result_o,file1,munge, 0, format,
nersc_csum,scidac_csuma,scidac_csumb);
std::cout << GridLogMessage << " Mixed checksums "<<std::hex << scidac_csuma << " "<<scidac_csumb<<std::endl;
BinaryIO::writeLatticeObject<vFermionD,FermionD>(result_o_2,file1,munge, 0, format,
nersc_csum,scidac_csuma,scidac_csumb);
std::cout << GridLogMessage << " CG checksums "<<std::hex << scidac_csuma << " "<<scidac_csumb<<std::endl;
}
#endif
}
MemoryManager::Print();
Grid_finalize();
}