portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-19 16:55:37 +01:00
Code Releases Activity

CUDA cub refuses to reduce vSpinColourMatrix, breaking up into smaller parts like already done for HIP case.

Browse Source
This commit is contained in:
dbollweg 2024-02-27 12:41:45 -05:00
parent b507fe209c
commit 3c9012676a
1 changed files with 3 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
Grid/lattice/Lattice_slicesum_core.h
View File

@ -119,18 +119,13 @@ template<class vobj> inline void sliceSumReduction_cub_large(const vobj *Data, V
template<class vobj> inline void sliceSumReduction_cub(const Lattice<vobj> &Data, Vector<vobj> &lvSum, const int rd, const int e1, const int e2, const int stride, const int ostride, const int Nsimd)
{
autoView(Data_v, Data, AcceleratorRead);
#if defined(GRID_CUDA)
sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
#elif defined (GRID_HIP) //hipcub cannot deal with large vobjs that don't fit in shared memory, therefore separate into _small/_large.
if constexpr (sizeof(vobj) <= 256) {
autoView(Data_v, Data, AcceleratorRead); //hipcub/cub cannot deal with large vobjs so we split into small/large case.
if constexpr (sizeof(vobj) <= 256) {
sliceSumReduction_cub_small(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
}
else {
sliceSumReduction_cub_large(&Data_v[0], lvSum, rd, e1, e2, stride, ostride, Nsimd);
}
#endif
}
#endif
@ -215,4 +210,4 @@ template<class vobj> inline void sliceSumReduction(const Lattice<vobj> &Data, Ve
}
NAMESPACE_END(Grid);
NAMESPACE_END(Grid);