portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-19 16:55:37 +01:00
Code Releases Activity

Don't need the numerical rounding tolerance in multigrid

Browse Source
This commit is contained in:
Peter Boyle 2023-12-22 18:10:23 -05:00
parent 551f6c4edd
commit b7c7000d0d
1 changed files with 31 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
Grid/lattice/Lattice_transfer.h
View File

@ -301,6 +301,37 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
// std::cout << GridLogPerformance << " blockProject : blockZaxpy : "<<t_za<<" us"<<std::endl;
}
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockProjectFast(Lattice<iVector<CComplex,nbasis > > &coarseData,
const Lattice<vobj> &fineData,
const VLattice &Basis)
{
GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid();
Lattice<iScalar<CComplex>> ip(coarse);
Lattice<vobj> fineDataRed = fineData;
autoView( coarseData_ , coarseData, AcceleratorWrite);
autoView( ip_ , ip, AcceleratorWrite);
RealD t_IP=0;
RealD t_co=0;
for(int v=0;v<nbasis;v++) {
t_IP-=usecond();
blockInnerProductD(ip,Basis[v],fineData); // ip = <basis|fine>
t_IP+=usecond();
t_co-=usecond();
accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
convertType(coarseData_[sc](v),ip_[sc]);
});
t_co+=usecond();
}
// std::cout << GridLogPerformance << " blockProjectFast : blockInnerProduct : "<<t_IP<<" us"<<std::endl;
// std::cout << GridLogPerformance << " blockProjectFast : conv : "<<t_co<<" us"<<std::endl;
}
// This only minimises data motion from CPU to GPU
// there is chance of better implementation that does a vxk loop of inner products to data share
// at the GPU thread level