portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-12 15:10:46 +01:00
Code Releases Activity

hip-friendly

Browse Source
This commit is contained in:
Christoph Lehner 2021-03-19 11:33:23 +01:00
parent 49ecbc81d4
commit 2bb374daea
2 changed files with 17 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
Grid/algorithms/CoarsenedMatrix.h
View File

@ -442,6 +442,8 @@ public:
for(int p=0; p<geom.npoint; p++) for(int p=0; p<geom.npoint; p++)
points[p] = geom.points_dagger[p]; points[p] = geom.points_dagger[p];
auto points_p = &points[0];
RealD* dag_factor_p = &dag_factor[0]; RealD* dag_factor_p = &dag_factor[0];
accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, { accelerator_for(sss, Grid()->oSites()*nbasis, Nsimd, {
@ -453,7 +455,7 @@ public:
StencilEntry *SE; StencilEntry *SE;
for(int p=0;p<geom_v.npoint;p++){ for(int p=0;p<geom_v.npoint;p++){
int point = points[p]; int point = points_p[p];
SE=Stencil_v.GetEntry(ptype,point,ss); SE=Stencil_v.GetEntry(ptype,point,ss);
@ -708,6 +710,8 @@ public:
for(int p=0; p<npoint; p++) for(int p=0; p<npoint; p++)
points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p; points[p] = (dag && !hermitian) ? geom.points_dagger[p] : p;
auto points_p = &points[0];
Vector<Aview> AcceleratorViewContainer; Vector<Aview> AcceleratorViewContainer;
for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead)); for(int p=0;p<npoint;p++) AcceleratorViewContainer.push_back(a[p].View(AcceleratorRead));
Aview *Aview_p = & AcceleratorViewContainer[0]; Aview *Aview_p = & AcceleratorViewContainer[0];
@ -728,7 +732,7 @@ public:
StencilEntry *SE; StencilEntry *SE;
for(int p=0;p<npoint;p++){ for(int p=0;p<npoint;p++){
int point = points[p]; int point = points_p[p];
SE=st_v.GetEntry(ptype,point,ss); SE=st_v.GetEntry(ptype,point,ss);
if(SE->_is_local) { if(SE->_is_local) {
@ -754,7 +758,7 @@ public:
StencilEntry *SE; StencilEntry *SE;
for(int p=0;p<npoint;p++){ for(int p=0;p<npoint;p++){
int point = points[p]; int point = points_p[p];
SE=st_v.GetEntry(ptype,point,ss); SE=st_v.GetEntry(ptype,point,ss);
if(SE->_is_local) { if(SE->_is_local) {

12
Grid/lattice/Lattice_transfer.h
View File

@ -360,16 +360,22 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
autoView( coarseData_ , coarseData, AcceleratorWrite); autoView( coarseData_ , coarseData, AcceleratorWrite);
autoView( fineData_ , fineData, AcceleratorRead); autoView( fineData_ , fineData, AcceleratorRead);
auto coarseData_p = &coarseData_[0];
auto fineData_p = &fineData_[0];
Coordinate fine_rdimensions = fine->_rdimensions; Coordinate fine_rdimensions = fine->_rdimensions;
Coordinate coarse_rdimensions = coarse->_rdimensions; Coordinate coarse_rdimensions = coarse->_rdimensions;
vobj zz = Zero();
accelerator_for(sc,coarse->oSites(),1,{ accelerator_for(sc,coarse->oSites(),1,{
// One thread per sub block // One thread per sub block
Coordinate coor_c(_ndimension); Coordinate coor_c(_ndimension);
Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions); // Block coordinate Lexicographic::CoorFromIndex(coor_c,sc,coarse_rdimensions); // Block coordinate
coarseData_[sc]=Zero();
vobj cd = zz;
for(int sb=0;sb<blockVol;sb++){ for(int sb=0;sb<blockVol;sb++){
int sf; int sf;
@ -379,9 +385,11 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d]; for(int d=0;d<_ndimension;d++) coor_f[d]=coor_c[d]*block_r[d] + coor_b[d];
Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions); Lexicographic::IndexFromCoor(coor_f,sf,fine_rdimensions);
coarseData_[sc]=coarseData_[sc]+fineData_[sf]; cd=cd+fineData_p[sf];
} }
coarseData_p[sc] = cd;
}); });
return; return;
} }