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

Add batched block project/promote functions

Browse Source
This commit is contained in:
Raoul Hodgson 2023-02-14 14:37:10 +00:00
parent 796abfad80
commit be528b6d27
1 changed files with 45 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
Grid/lattice/Lattice_transfer.h
View File

@ -288,7 +288,34 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
blockZAXPY(fineDataRed,ip,Basis[v],fineDataRed);
}
}
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void batchBlockProject(std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
const std::vector<Lattice<vobj>> &fineData,
const VLattice &Basis)
{
int NBatch = fineData.size();
GridBase * fine = fineData[0].Grid();
GridBase * coarse= coarseData[0].Grid();
Lattice<iScalar<CComplex>> ip(coarse);
std::vector<Lattice<vobj>> fineDataCopy = fineData;
autoView(ip_, ip, AcceleratorWrite);
for(int v=0;v<nbasis;v++) {
for (int k=0; k<NBatch; k++) {
autoView( coarseData_ , coarseData[k], AcceleratorWrite);
blockInnerProductD(ip,Basis[v],fineDataCopy[k]); // ip = <basis|fine>
accelerator_for( sc, coarse->oSites(), vobj::Nsimd(), {
convertType(coarseData_[sc](v),ip_[sc]);
});
// improve numerical stability of projection
// |fine> = |fine> - <basis|fine> |basis>
ip=-ip;
blockZAXPY(fineDataCopy[k],ip,Basis[v],fineDataCopy[k]);
}
}
}
template<class vobj,class vobj2,class CComplex>
inline void blockZAXPY(Lattice<vobj> &fineZ,
@ -590,6 +617,24 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
}
#endif
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void batchBlockPromote(const std::vector<Lattice<iVector<CComplex,nbasis>>> &coarseData,
std::vector<Lattice<vobj>> &fineData,
const VLattice &Basis)
{
int NBatch = fineData.size();
GridBase * fine = fineData[0].Grid();
GridBase * coarse = coarseData[0].Grid();
for (int k=0; k<NBatch; k++)
fineData[k]=Zero();
for (int i=0;i<nbasis;i++) {
for (int k=0; k<NBatch; k++) {
Lattice<iScalar<CComplex>> ip = PeekIndex<0>(coarseData[k],i);
blockZAXPY(fineData[k],ip,Basis[i],fineData[k]);
}
}
}
// Useful for precision conversion, or indeed anything where an operator= does a conversion on scalars.
// Simd layouts need not match since we use peek/poke Local
template<class vobj,class vvobj>