portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-09 21:50:45 +01:00
Code Releases Activity

Accelerator loop attempt at speed up

Browse Source
This commit is contained in:
Peter Boyle 2019-12-14 05:28:16 -05:00
parent 152b525a4d
commit 9e15474999
1 changed files with 35 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
Grid/lattice/Lattice_transfer.h
View File

@ -106,6 +106,7 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d]; block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]); assert(block_r[d]*coarse->_rdimensions[d] == fine->_rdimensions[d]);
} }
int blockVol = fine->oSites()/coarse->oSites();
coarseData=Zero(); coarseData=Zero();
@ -113,20 +114,26 @@ inline void blockProject(Lattice<iVector<CComplex,nbasis > > &coarseData,
auto coarseData_ = coarseData.View(); auto coarseData_ = coarseData.View();
//////////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////////
// To make this lock free, 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 // Otherwise do fine inner product per site, and make the update atomic
//////////////////////////////////////////////////////////////////////////////////////////////////////// ////////////////////////////////////////////////////////////////////////////////////////////////////////
thread_for( sf, fine->oSites(), { accelerator_for( sc, coarse->oSites(), {
int sc;
Coordinate coor_c(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d];
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
for(int i=0;i<nbasis;i++) { Coordinate coor_c(_ndimension);
auto Basis_ = Basis[i].View(); Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions); // Block coordinate
auto ip = innerProduct(Basis_[sf],fineData_[sf]); coarseData_[sc]=Zero();
thread_critical {
for(int sb=0;sb<blockVol;sb++){
Coordinate coor_b(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_b,sb,block_r);
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);
for(int i=0;i<nbasis;i++) {
auto Basis_ = Basis[i].View();
auto ip = innerProduct(Basis_[sf],fineData_[sf]);
coarseData_[sc](i)=coarseData_[sc](i) + ip; coarseData_[sc](i)=coarseData_[sc](i) + ip;
} }
} }
@ -230,23 +237,29 @@ inline void blockSum(Lattice<vobj> &coarseData,const Lattice<vobj> &fineData)
for(int d=0 ; d<_ndimension;d++){ for(int d=0 ; d<_ndimension;d++){
block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d]; block_r[d] = fine->_rdimensions[d] / coarse->_rdimensions[d];
} }
int blockVol = fine->oSites()/coarse->oSites();
// Turn this around to loop threaded over sc and interior loop // Turn this around to loop threaded over sc and interior loop
// over sf would thread better // over sf would thread better
coarseData=Zero();
auto coarseData_ = coarseData.View(); auto coarseData_ = coarseData.View();
auto fineData_ = fineData.View(); auto fineData_ = fineData.View();
thread_for(sf,fine->oSites(),{ accelerator_for(sc,coarse->oSites(),1,{
int sc;
// One thread per sub block
Coordinate coor_c(_ndimension); Coordinate coor_c(_ndimension);
Coordinate coor_f(_ndimension); Lexicographic::CoorFromIndex(coor_c,sc,coarse->_rdimensions); // Block coordinate
coarseData_[sc]=Zero();
Lexicographic::CoorFromIndex(coor_f,sf,fine->_rdimensions);
for(int d=0;d<_ndimension;d++) coor_c[d]=coor_f[d]/block_r[d]; for(int sb=0;sb<blockVol;sb++){
Lexicographic::IndexFromCoor(coor_c,sc,coarse->_rdimensions);
int sf;
thread_critical { Coordinate coor_b(_ndimension);
Coordinate coor_f(_ndimension);
Lexicographic::CoorFromIndex(coor_b,sb,block_r); // Block sub coordinate
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);
coarseData_[sc]=coarseData_[sc]+fineData_[sf]; coarseData_[sc]=coarseData_[sc]+fineData_[sf];
} }