ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #2823 from Ruyk/master

Browse Source 
use SYCL team_reduce to mimic cuda __syncthreads_count
This commit is contained in:
Axel Kohlmeyer
2021-07-06 16:38:18 -04:00
committed by GitHub
parent 3b480397d1 e2a78cbed6
commit 5f7766a71c
1 changed files with 6 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/KOKKOS/npair_kokkos.cpp
View File

@ -610,7 +610,9 @@ void NeighborKokkosExecute<DeviceType>::build_ItemGPU(typename Kokkos::TeamPolic
if (test) return;
#else
dev.team_barrier();
int not_done = (i >= 0 && i <= nlocal);
dev.team_reduce(Kokkos::Max<int>(not_done));
if(not_done == 0) return;
#endif
if (i >= 0 && i < nlocal) {
@ -1053,13 +1055,14 @@ void NeighborKokkosExecute<DeviceType>::build_ItemSizeGPU(typename Kokkos::TeamP
other_x[MY_II + 4 * atoms_per_bin] = radi;
}
other_id[MY_II] = i;
// FIXME_SYCL
#ifndef KOKKOS_ENABLE_SYCL
int test = (__syncthreads_count(i >= 0 && i <= nlocal) == 0);
if (test) return;
#else
dev.team_barrier();
int not_done = (i >= 0 && i <= nlocal);
dev.team_reduce(Kokkos::Max<int>(not_done));
if(not_done == 0) return;
#endif
if (i >= 0 && i < nlocal) {