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

Using a single slice rather than the typical two provides about 5% speedup (sometimes more or less) on some GFX9 systems for single node. [ROCm/rccl commit: 2f6b20c00a]
2025-05-29 16:17:35 -07:00
parent 19ed482121
commit 4277b5aa88
9 changed files with 66 additions and 10 deletions
@@ -288,7 +288,7 @@ 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;
+  uint32_t redOpArgIsPtr:1, regUsed:1, netRegUsed:1, oneNode:1, direct:2, isOneRPN:1, rcclUseOneSlice:1;
  uint32_t root:30, connIndex:2;
  uint16_t pivotA2ANumBiRings;
  void* recvbuff;