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

multRHS initial support -- needs optimisation for multi project/promote.

Browse Source 
Bug fix in freeing intermediate grids to stop double free
This commit is contained in:
Peter Boyle 2023-11-23 18:18:35 -05:00
parent 639cc6f73a
commit 031f85247c
2 changed files with 34 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
Grid/lattice/Lattice_transfer.h
View File

@ -265,8 +265,8 @@ inline auto localInnerProductD(const Lattice<vobj> &lhs,const Lattice<vobj> &rhs
//////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj,class CComplex,int nbasis,class VLattice> template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData, inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
const Lattice<vobj> &fineData, const Lattice<vobj> &fineData,
const VLattice &Basis) const VLattice &Basis)
{ {
GridBase * fine = fineData.Grid(); GridBase * fine = fineData.Grid();
GridBase * coarse= coarseData.Grid(); GridBase * coarse= coarseData.Grid();
@ -300,6 +300,7 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
// std::cout << GridLogPerformance << " blockProject : conv : "<<t_co<<" us"<<std::endl; // std::cout << GridLogPerformance << " blockProject : conv : "<<t_co<<" us"<<std::endl;
// std::cout << GridLogPerformance << " blockProject : blockZaxpy : "<<t_za<<" us"<<std::endl; // std::cout << GridLogPerformance << " blockProject : blockZaxpy : "<<t_za<<" us"<<std::endl;
} }
// This only minimises data motion from CPU to GPU // 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 // there is chance of better implementation that does a vxk loop of inner products to data share
// at the GPU thread level // at the GPU thread level
@ -1802,5 +1803,32 @@ void Grid_unsplit(std::vector<Lattice<Vobj> > & full,Lattice<Vobj> & split)
} }
} }
//////////////////////////////////////////////////////
// MultiRHS interface support for coarse space
// -- Simplest possible implementation to begin with
//////////////////////////////////////////////////////
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockProjectMany(Lattice<iVector<CComplex,nbasis > > &coarseIP,
Lattice<iVector<CComplex,nbasis > > &coarseTMP,
const VLattice &fineData, // Basis and fineData necessarily same type
const VLattice &Basis)
{
for(int r=0;r<fineData.size();r++){
blockProject(coarseTMP,fineData[r],Basis);
InsertSliceLocal(coarseTMP, coarseIP,r,r,0);
}
}
template<class vobj,class CComplex,int nbasis,class VLattice>
inline void blockPromoteMany(Lattice<iVector<CComplex,nbasis > > &coarseIP,
Lattice<iVector<CComplex,nbasis > > &coarseTMP,
const VLattice &fineData, // Basis and fineData necessarily same type
const VLattice &Basis)
{
for(int r=0;r<fineData.size();r++){
ExtractSliceLocal(coarseTMP, coarseIP,r,r,0);
blockPromote(coarseTMP,fineData[r],Basis);
}
}
NAMESPACE_END(Grid); NAMESPACE_END(Grid);

5
Grid/lattice/PaddedCell.h
View File

@ -234,8 +234,11 @@ public:
} }
void DeleteGrids(void) void DeleteGrids(void)
{ {
Coordinate processors=unpadded_grid->_processors;
for(int d=0;d<grids.size();d++){ for(int d=0;d<grids.size();d++){
delete grids[d]; if ( processors[d] > 1 ) {
delete grids[d];
}
} }
grids.resize(0); grids.resize(0);
}; };