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

Use One Slice per Basic Primitive for AllReduce, ReduceScatter, AllGather (#1681) for Single Node on Some GFX9 Systems

Browse Source 
Using a single slice rather than the typical two provides about 5% speedup (sometimes more or less) on some GFX9 systems for single node.
This commit is contained in:
alex-breslow-amd
2025-05-29 16:17:35 -07:00
committed by GitHub
parent 12517a957e
commit 2f6b20c00a
9 changed files with 66 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/enqueue.cc
+1
View File
@@ -332,6 +332,7 @@ ncclResult_t ncclTasksRegAndEnqueue(struct ncclComm* comm) {
devWork.redOpArg = task->opDev.scalarArg;
devWork.redOpArgIsPtr = task->opDev.scalarArgIsPtr;
devWork.oneNode = (comm->nNodes == 1);
devWork.rcclUseOneSlice = comm->rcclUseOneSlice;
devWork.isOneRPN = comm->isOneRPN;
devWork.netRegUsed = devWork.regUsed = 0;
if (task->regBufType & NCCL_NET_REG_BUFFER)