AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #810 from nusislam/tune-send-recv

Browse Source 
device: fine-tune RCCL send-recv on MI250/MI200
This commit is contained in:
Nusrat Islam
2023-07-25 10:18:12 -05:00
committed by GitHub
parent 8bab4f04b7 b2a0a0bd3a
commit 47f754e6f5
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/collectives/device/sendrecv.h
+8
View File
@@ -202,13 +202,21 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
if (args->proto == NCCL_PROTO_LL) {
runRecv<ProtoLL>(tid, nthreads, group, args);
} else {
#if defined(__gfx90a__)
runRecv<ProtoSimple<1,1,8>>(tid, nthreads, group, args);
#else
runRecv<ProtoSimple<1,1>>(tid, nthreads, group, args);
#endif
}
} else {
if (args->proto == NCCL_PROTO_LL) {
runSend<ProtoLL>(tid, nthreads, group, args);
} else {
#if defined(__gfx90a__)
runSend<ProtoSimple<1,1,8>>(tid, nthreads, group, args);
#else
runSend<ProtoSimple<1,1>>(tid, nthreads, group, args);
#endif
}
}
}