portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-13 01:05:36 +00:00
Code Releases Activity

offload more of mgrid to GPU

Browse Source
This commit is contained in:
Peter Boyle 2019-12-13 22:08:11 -05:00
parent 736b19485e
commit d18994eddc
2 changed files with 12 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
Grid/algorithms/CoarsenedMatrix.h
View File

@ -253,27 +253,6 @@ public:
Chebyshev<FineField> ChebFilt (0.03,64.0,500); Chebyshev<FineField> ChebFilt (0.03,64.0,500);
Chebyshev<FineField> ChebDependent(0.01,64.0,200); Chebyshev<FineField> ChebDependent(0.01,64.0,200);
#if 0
auto latt_size = FineGrid->GlobalDimensions();
Coordinate Fourier[dependent] = {
Coordinate({0, 0,0,0,0}),
Coordinate({0, 1,0,0,0}),
Coordinate({0,-1,0,0,0}),
Coordinate({0,0, 1,0,0}),
Coordinate({0,0,-1,0,0}),
Coordinate({0,0,0, 1,0}),
Coordinate({0,0,0,-1,0}),
Coordinate({0,0,0,0, 1}),
Coordinate({0,0,0,0,-1})
};
ComplexD ci(0.0,1.0);
Lattice<CComplex> C(FineGrid);
Lattice<CComplex> coor(FineGrid);
FineField save(FineGrid);
FineField tmp (FineGrid);
#endif
FineField noise(FineGrid); FineField noise(FineGrid);
FineField Mn(FineGrid); FineField Mn(FineGrid);
@ -295,17 +274,6 @@ public:
if(b==bb) { if(b==bb) {
ChebFilt(hermop,noise,Mn); ChebFilt(hermop,noise,Mn);
} else { } else {
#if 0
C=Zero();
for(int mu=0;mu<5;mu++){
RealD TwoPiL = M_PI * 2.0/ latt_size[mu];
LatticeCoordinate(coor,mu);
C = C + (TwoPiL * Fourier[dep][mu]) * coor;
}
C = exp(C*ci); // Fourier phase
noise=C*save;
hermop.Op(noise,Mn); std::cout<<GridLogMessage << "noise ["<<b<<"] <n|MdagM|n> "<<norm2(Mn)<<std::endl;
#endif
ChebDependent(hermop,noise,Mn); ChebDependent(hermop,noise,Mn);
} }

20
Grid/lattice/Lattice_transfer.h
View File

@ -111,7 +111,10 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
auto fineData_ = fineData.View(); auto fineData_ = fineData.View();
auto coarseData_ = coarseData.View(); auto coarseData_ = coarseData.View();
// Loop over coars parallel, and then loop over fine associated with coarse. ////////////////////////////////////////////////////////////////////////////////////////////////////////
// To make this lock free, loop over coars parallel, and then loop over fine associated with coarse.
// Otherwise do finee inner product per site, and make the update atomic
////////////////////////////////////////////////////////////////////////////////////////////////////////
thread_for( sf, fine->oSites(), { thread_for( sf, fine->oSites(), {
int sc; int sc;
Coordinate coor_c(_ndimension); Coordinate coor_c(_ndimension);
@ -120,10 +123,11 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d]; for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions); Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
thread_critical { for(int i=0;i<nbasis;i++) {
for(int i=0;i<nbasis;i++) { auto Basis_ = Basis[i].View();
auto Basis_ = Basis[i].View(); auto ip = innerProduct(Basis_[sf],fineData_[sf]);
coarseData_[sc](i)=coarseData_[sc](i) + innerProduct(Basis_[sf],fineData_[sf]); thread_critical {
coarseData_[sc](i)=coarseData_[sc](i) + ip;
} }
} }
}); });
@ -160,7 +164,7 @@ inline void blockZAXPY(Lattice<vobj> &fineZ,
auto fineY_ = fineY.View(); auto fineY_ = fineY.View();
auto coarseA_= coarseA.View(); auto coarseA_= coarseA.View();
thread_for(sf, fine->oSites(), { accelerator_for(sf, fine->oSites(), 1, {
int sc; int sc;
Coordinate coor_c(_ndimension); Coordinate coor_c(_ndimension);
@ -196,7 +200,7 @@ inline void blockInnerProduct(Lattice<CComplex> &CoarseInner,
fine_inner = localInnerProduct(fineX,fineY); fine_inner = localInnerProduct(fineX,fineY);
blockSum(coarse_inner,fine_inner); blockSum(coarse_inner,fine_inner);
thread_for(ss, coarse->oSites(),{ accelerator_for(ss, coarse->oSites(), 1, {
CoarseInner_[ss] = coarse_inner_[ss]; CoarseInner_[ss] = coarse_inner_[ss];
}); });
} }
@ -321,7 +325,7 @@ inline void blockPromote(const Lattice<iVector<CComplex,nbasis > > &coarseData,
auto coarseData_ = coarseData.View(); auto coarseData_ = coarseData.View();
// Loop with a cache friendly loop ordering // Loop with a cache friendly loop ordering
thread_for(sf,fine->oSites(),{ acceelerator_for(sf,fine->oSites(),1,{
int sc; int sc;
Coordinate coor_c(_ndimension); Coordinate coor_c(_ndimension);
Coordinate coor_f(_ndimension); Coordinate coor_f(_ndimension);