Single-node AllGather and ReduceScatter Optimization (#2019)

* Single-node performance tuning

* Normalizing value to individual rank

src/rccl_wrap.cc

@@ -38,14 +38,20 @@ RCCL_PARAM(DirectAllGatherThreshold, "DIRECT_ALLGATHER_THRESHOLD", 75497472);
 void rcclUpdateCollectiveProtocol(struct ncclComm* comm, size_t const& nBytes, struct ncclTaskColl* info) {
   // Honor user input for protocol choice
   static int userProtocolInput = -2;
+  size_t sizePerRank = rcclGetSizePerRank(info->func, nBytes, comm->nRanks);
   if (userProtocolInput == -2) {
     const char *protoStr = getenv("NCCL_PROTO");
     userProtocolInput = !protoStr ? 0 : 1;
   }
-  if (!userProtocolInput && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950") && comm->nNodes == 1 && (info->func == ncclFuncAllGather) && nBytes <= 524288) {
+  if (!userProtocolInput && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950") && comm->nNodes == 1 && (info->func == ncclFuncAllGather) && sizePerRank <= 88448) {
+    // Change LL protocol threshold
+    info->protocol = NCCL_PROTO_LL;
+  } else if (!userProtocolInput && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx950") && comm->nNodes == 1 && (info->func == ncclFuncReduceScatter) && sizePerRank <= 175488) {
+    // Change LL protocol threshold
+    info->protocol = NCCL_PROTO_LL;
+  } else if (!userProtocolInput && IsArchMatch(comm->topo->nodes[GPU].nodes[0].gpu.gcn, "gfx942") && comm->nNodes == 1 && (info->func == ncclFuncReduceScatter) && sizePerRank <= 352128) {
     // Change LL protocol threshold
     info->protocol = NCCL_PROTO_LL;
-
   } else if(!userProtocolInput && comm->nNodes >= 2 && (info->func == ncclFuncReduceScatter || info->func == ncclFuncAllGather || info->func == ncclFuncAllReduce || info->func == ncclFuncBroadcast || info->func == ncclFuncReduce)) {
     auto tunableIndex = rcclGetTunableIndex(info->func);
     auto llMin = comm->minMaxLLRange[tunableIndex][NCCL_PROTO_LL][RCCL_PROTOCOL_MIN_IDX];