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

Cheaper threadfence for gfx942 in postPeer [1/N]: enable for single node allreduce (#1766)

Browse Source 
Boosts single node bfloat16 allreduce performance by up to 20% for some data sizes and provides gating with the RCCL_GFX942_CHEAP_FENCE_OFF environment variable

[ROCm/rccl commit: 11fabf1de1]
This commit is contained in:
alex-breslow-amd
2025-07-22 07:15:15 -07:00
committed by GitHub
parent bd63518944
commit cbb648505a
10 changed files with 116 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/device.h
+1 -1
View File
@@ -305,7 +305,7 @@ struct alignas(16) ncclDevWorkColl {
// nChannels == (channelHi - channelLo) + 1
uint32_t channelLo:8, channelHi:8;
uint32_t nWarps:8;
uint32_t redOpArgIsPtr:1, regUsed:1, netRegUsed:1, oneNode:1, direct:2, isOneRPN:1, rcclUseOneSlice:1;
uint32_t redOpArgIsPtr:1, regUsed:1, netRegUsed:1, oneNode:1, direct:2, isOneRPN:1, rcclUseOneSlice:1, gfx942CheapFenceOff:1;
uint32_t root:30, connIndex:2;
uint16_t pivotA2ANumBiRings:15, profilerEnabled:1;
void* recvbuff;