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

Imporved General coarsened matrix

Browse Source
This commit is contained in:
Peter Boyle 2023-10-18 22:41:53 -04:00
parent 541772313c
commit 7cc3435ba8
1 changed files with 49 additions and 32 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

81
Grid/algorithms/multigrid/GeneralCoarsenedMatrix.h
View File

@ -47,6 +47,7 @@ public:
typedef Lattice<siteVector> CoarseVector; typedef Lattice<siteVector> CoarseVector;
typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix; typedef Lattice<iMatrix<CComplex,nbasis > > CoarseMatrix;
typedef iMatrix<CComplex,nbasis > Cobj; typedef iMatrix<CComplex,nbasis > Cobj;
typedef iVector<CComplex,nbasis > Cvec;
typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field typedef Lattice< CComplex > CoarseScalar; // used for inner products on fine field
typedef Lattice<Fobj > FineField; typedef Lattice<Fobj > FineField;
typedef CoarseVector Field; typedef CoarseVector Field;
@ -136,61 +137,76 @@ public:
CoarseVector tin=in; CoarseVector tin=in;
texch-=usecond(); texch-=usecond();
CoarseVector pin = Cell.Exchange(tin); // CoarseVector pin = Cell.Exchange(tin);
CoarseVector pin = Cell.ExchangeTest(tin);
texch+=usecond(); texch+=usecond();
CoarseVector pout(pin.Grid()); pout=Zero(); CoarseVector pout(pin.Grid());
int npoint = geom.npoint; int npoint = geom.npoint;
typedef LatticeView<Cobj> Aview; typedef LatticeView<Cobj> Aview;
typedef LatticeView<Cvec> Vview;
const int Nsimd = CComplex::Nsimd(); const int Nsimd = CComplex::Nsimd();
int64_t osites=pin.Grid()->oSites(); int64_t osites=pin.Grid()->oSites();
// int gsites=pin.Grid()->gSites();
RealD flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd(); RealD flops = 1.0* npoint * nbasis * nbasis * 8.0 * osites * CComplex::Nsimd();
RealD bytes = 1.0*osites*sizeof(siteMatrix)*npoint RealD bytes = 1.0*osites*sizeof(siteMatrix)*npoint
+ 2.0*osites*sizeof(siteVector)*npoint; + 2.0*osites*sizeof(siteVector)*npoint;
// for(int point=0;point<npoint;point++){
// conformable(A[point],pin);
// }
{ {
tviews-=usecond();
autoView( in_v , pin, AcceleratorRead); autoView( in_v , pin, AcceleratorRead);
autoView( out_v , pout, AcceleratorWrite); autoView( out_v , pout, AcceleratorWriteDiscard);
autoView( Stencil_v , Stencil, AcceleratorRead); autoView( Stencil_v , Stencil, AcceleratorRead);
// Static and prereserve to keep UVM region live and not resized across multiple calls
Vector<Aview> AcceleratorViewContainer; AcceleratorViewContainer.reserve(npoint);
Vector<Vview> AcceleratorVecViewContainer; AcceleratorVecViewContainer.reserve(npoint);
std::vector<CoarseVector> outp(npoint,pin.Grid());
tviews-=usecond();
for(int p=0;p<npoint;p++) {
AcceleratorViewContainer.push_back( A[p].View(AcceleratorRead));
AcceleratorVecViewContainer.push_back(outp[p].View(AcceleratorWrite));
}
tviews+=usecond(); tviews+=usecond();
auto Aview_p = &AcceleratorViewContainer[0];
auto Vview_p = &AcceleratorVecViewContainer[0];
tmult-=usecond(); tmult-=usecond();
for(int point=0;point<npoint;point++){ accelerator_for(spb, osites*nbasis*npoint, Nsimd, {
std::cout << GridLogMessage<< "View "<<point<<"/"<<npoint<<std::endl; typedef decltype(coalescedRead(in_v[0](0))) calcComplex;
tviews-=usecond(); int32_t ss = spb/(nbasis*npoint);
autoView( A_v, A[point],AcceleratorRead); int32_t bp = spb%(nbasis*npoint);
tviews+=usecond(); int32_t b = bp/npoint;
std::cout << GridLogMessage<< "Mult "<<point<<"/"<<npoint<<std::endl; int32_t point= bp%npoint;
accelerator_for(sss, osites*nbasis, Nsimd, { auto SE = Stencil_v.GetEntry(point,ss);
auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd);
typedef decltype(coalescedRead(in_v[0])) calcVector; auto res = coalescedRead(Aview_p[point][ss](b,0))*nbr(0);
for(int bb=1;bb<nbasis;bb++) {
int ss = sss/nbasis; res = res + coalescedRead(Aview_p[point][ss](b,bb))*nbr(bb);
int b = sss%nbasis; }
coalescedWrite(Vview_p[point][ss](b),res);
auto SE = Stencil_v.GetEntry(point,ss); });
auto nbr = coalescedReadGeneralPermute(in_v[SE->_offset],SE->_permute,Nd); accelerator_for(sb, osites*nbasis, Nsimd, {
auto res = out_v(ss)(b); int ss = sb/nbasis;
for(int bb=0;bb<nbasis;bb++) { int b = sb%nbasis;
res = res + coalescedRead(A_v[ss](bb,b))*nbr(bb); auto res = coalescedRead(Vview_p[0][ss](b));
} for(int point=1;point<npoint;point++){
coalescedWrite(out_v[ss](b),res); res = res + coalescedRead(Vview_p[point][ss](b));
}); }
} coalescedWrite(out_v[ss](b),res);
});
tmult+=usecond(); tmult+=usecond();
for(int p=0;p<npoint;p++) {
AcceleratorViewContainer[p].ViewClose();
AcceleratorVecViewContainer[p].ViewClose();
}
} }
text-=usecond(); text-=usecond();
std::cout << GridLogMessage<< "Extract "<<std::endl;
out = Cell.Extract(pout); out = Cell.Extract(pout);
text+=usecond(); text+=usecond();
ttot+=usecond(); ttot+=usecond();
@ -200,6 +216,7 @@ public:
std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl; std::cout << GridLogPerformance<<"Coarse Mult mult "<<tmult<<" us"<<std::endl;
std::cout << GridLogPerformance<<"Coarse Mult ext "<<text<<" us"<<std::endl; std::cout << GridLogPerformance<<"Coarse Mult ext "<<text<<" us"<<std::endl;
std::cout << GridLogPerformance<<"Coarse Mult tot "<<ttot<<" us"<<std::endl; std::cout << GridLogPerformance<<"Coarse Mult tot "<<ttot<<" us"<<std::endl;
std::cout << GridLogPerformance<<std::endl;
std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl; std::cout << GridLogPerformance<<"Coarse Kernel flop/s "<< flops/tmult<<" mflop/s"<<std::endl;
std::cout << GridLogPerformance<<"Coarse Kernel bytes/s"<< bytes/tmult<<" MB/s"<<std::endl; std::cout << GridLogPerformance<<"Coarse Kernel bytes/s"<< bytes/tmult<<" MB/s"<<std::endl;
std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl; std::cout << GridLogPerformance<<"Coarse overall flops/s "<< flops/ttot<<" mflop/s"<<std::endl;