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

Merge sendrecv kernel from NCCL 2.7.3

Browse Source 
This commit was cherry-picked and modified from
https://github.com/NVIDIA/nccl/commit/5949d96f36d050e59d05872f8bbffd2549318e95
This commit is contained in:
Wenkai Du
2020-06-29 08:47:46 -07:00
parent b90735c935
commit 964c4c2061
3 changed files with 51 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/proxy.cc
+2 -2
View File
@@ -152,13 +152,13 @@ ncclResult_t ncclProxySaveP2p(struct ncclInfo* info, struct ncclChannel* channel
args.dtype = info->datatype;
if (info->delta > 0 && info->sendbytes >= 0) {
int peersend = (info->comm->rank+info->delta)%info->comm->nRanks;
args.nsteps = DIVUP(info->sendbytes, info->comm->buffSizes[NCCL_PROTO_SIMPLE]/NCCL_STEPS);
args.nsteps = DIVUP(info->sendbytes, info->comm->buffSizes[NCCL_PROTO_SIMPLE]/NCCL_STEPS/SENDRECV_SLICEFACTOR);
if (args.nsteps == 0) args.nsteps = 1;
NCCLCHECK(SaveProxy<proxySend>(peersend, &args));
}
if (info->delta > 0 && info->recvbytes >= 0) {
int peerrecv = (info->comm->nRanks+info->comm->rank-info->delta)%info->comm->nRanks;
args.nsteps = DIVUP(info->recvbytes, info->comm->buffSizes[NCCL_PROTO_SIMPLE]/NCCL_STEPS);
args.nsteps = DIVUP(info->recvbytes, info->comm->buffSizes[NCCL_PROTO_SIMPLE]/NCCL_STEPS/SENDRECV_SLICEFACTOR);
if (args.nsteps == 0) args.nsteps = 1;
NCCLCHECK(SaveProxy<proxyRecv>(peerrecv, &args));
}