[WarpSpeed] Improve handling for auto and manual modes (#2125)

* Force ring in WarpSpeed manual mode and log event

* Skip usage for non-ring in WarpSpeed auto mode

* Enable WarpSpeed when its CU count is set

[ROCm/rccl commit: 93fdcb160c]

projects/rccl/src/init.cc

@@ -1078,6 +1078,7 @@ NCCL_PARAM(AllocP2pNetLLBuffers, "ALLOC_P2P_NET_LL_BUFFERS", 0);
 #ifdef ENABLE_WARP_SPEED
 extern int64_t rcclParamWarpSpeedEnable();
 extern int64_t rcclParamWarpSpeedAutoMode();
+extern int64_t rcclParamWarpSpeedCuCount();
 #endif
 // MNNVL: Flag to indicate whether to enable Multi-Node NVLink
 NCCL_PARAM(MNNVLEnable, "MNNVL_ENABLE", 2);
@@ -1458,7 +1459,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, struct ncclComm* p
     }
   }
 #ifdef ENABLE_WARP_SPEED
-  comm->topo->warpSpeedEnabled = (rcclParamWarpSpeedEnable() != 0 || rcclParamWarpSpeedAutoMode() != 0);
+  comm->topo->warpSpeedEnabled = (rcclParamWarpSpeedEnable() != 0 || rcclParamWarpSpeedAutoMode() != 0 || rcclParamWarpSpeedCuCount() > 0);
 #endif
 
   // For single node communicators that do not uses the full xgmi links per gpu, i.e., nranks < 8

projects/rccl/src/rccl_wrap.cc

@@ -51,11 +51,19 @@ void rcclRestrictMaxChannels(struct ncclComm* comm, int& nc ) {
     }
 }
 
+static inline bool rcclCollSupportsRing(ncclFunc_t func) {
+  return (func == ncclFuncAllReduce ||
+          func == ncclFuncAllGather ||
+          func == ncclFuncReduceScatter ||
+          func == ncclFuncBroadcast ||
+          func == ncclFuncReduce);
+}
+
 int32_t rcclGetProtoForGfx12(ncclFunc_t collectiveFunc, size_t sizePerRank){
   int returnVal = NCCL_PROTO_SIMPLE;
-  int SingleNodeLLCutoffs[] = { 
+  int SingleNodeLLCutoffs[] = {
     /*ncclFuncBroadcast*/     1536,
-    /*ncclFuncReduce*/        8192, 
+    /*ncclFuncReduce*/        8192,
     /*ncclFuncAllGather*/     98304,
     /*ncclFuncReduceScatter*/ 98304,
     /*ncclFuncAllReduce*/     913532,
@@ -513,13 +521,20 @@ void rcclSetWarpSpeedSupportAndFinalCuCount(struct ncclComm* comm, struct ncclKe
 
 void rcclSetWarpSpeedAuto(struct ncclComm* comm, struct ncclTaskColl* info, size_t nBytes) {
   info->useWarpSpeed = false;
-  if(rcclParamWarpSpeedAutoMode() != 0) { // auto mode
+  if(!rcclCollSupportsRing(info->func)) return;
+  if(rcclParamWarpSpeedAutoMode() != 0) { // Auto performance mode
     if(!IsArchMatch(comm->archName, "gfx950")) {
       // Auto mode only available for gfx950 currently, keep it to false
       return;
     }
     size_t minBytes = 0;
     commSetUnrollFactor(comm);  // TODO: reset unroll factor per task rather than per comm
+    // No early return based on the algorithm at the start of the function
+    // to allow unroll factor to be reverted to default.
+    // This can be changed once per-task unroll factor setting is implemented.
+    if(info->algorithm != NCCL_ALGO_RING) {
+      return; // If Ring is not selected, assume it is suboptimal and return
+    }
     if(info->func == ncclFuncAllReduce || info->func == ncclFuncAllGather) minBytes = RCCL_WARP_SPEED_MIN_BYTES;
     else if (info->func == ncclFuncReduceScatter) minBytes = RCCL_WARP_SPEED_MIN_BYTES << 2; // ReduceScatter requires higher message size to benefit from WarpSpeed
     if(comm->nNodes == 1) {
@@ -529,7 +544,11 @@ void rcclSetWarpSpeedAuto(struct ncclComm* comm, struct ncclTaskColl* info, size
         info->useWarpSpeed = true;
       }
     }
-  } else if (comm->topo->warpSpeedEnabled && info->algorithm == NCCL_ALGO_RING) {
+  } else if (comm->topo->warpSpeedEnabled) {
+    if(info->algorithm != NCCL_ALGO_RING) {
+      INFO(NCCL_TUNING, "Overriding %s algorithm with RING for nccl%s at %zu bytes as WarpSpeed is requested and only supports RING", ncclAlgoToString(info->algorithm), ncclFuncToString(info->func), nBytes);
+      info->algorithm = NCCL_ALGO_RING; // Force Ring when WarpSpeed is enabled in manual mode as it only supports Ring
+    }
     info->useWarpSpeed = true;
   }
 }