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

Use relaxed atomics for LL on GFX11 (#859)

Browse Source 
[ROCm/rccl commit: 6a0a6a37d9]
This commit is contained in:
Wenkai Du
2023-08-21 16:28:39 -07:00
committed by GitHub
parent 75e3927f50
commit 5983f0e371
3 changed files with 30 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/enqueue.cc
+1 -1
View File
@@ -319,7 +319,7 @@ static ncclResult_t addP2pToPlan(
struct ncclConnInfo* conn = isSendNotRecv ?
&comm->channels[channelId].peers[peer]->send[1].conn : &comm->channels[channelId].peers[peer]->recv[1].conn;
// do not use LL on gfx11
info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold() && comm->topo->nodes[GPU].nodes[0].gpu.gcn/100 != 11) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold()) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
struct ncclProxyOp proxyOp = {};
NCCLCHECK(ncclProxyComputeP2p(&info, &proxyOp));