2.7.3-1

Add support for A100 GPU and related platforms.
Add support for CUDA 11.
Add support for send/receive operations (beta).


[ROCm/rccl commit: 5949d96f36]

projects/rccl/src/misc/argcheck.cc

@@ -1,5 +1,5 @@
 /*************************************************************************
- * Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
+ * Copyright (c) 2019-2020, NVIDIA CORPORATION. All rights reserved.
  *
  * See LICENSE.txt for license information
  ************************************************************************/
@@ -34,7 +34,6 @@ ncclResult_t PtrCheck(void* ptr, const char* opname, const char* ptrname) {
 }
 
 ncclResult_t ArgsCheck(struct ncclInfo* info) {
-  NCCLCHECK(PtrCheck(info->comm, info->opName, "comm"));
   // First, the easy ones
   if (info->root < 0 || info->root >= info->comm->nRanks) {
     WARN("%s : invalid root %d (root should be in the 0..%d range)", info->opName, info->root, info->comm->nRanks);
@@ -44,7 +43,7 @@ ncclResult_t ArgsCheck(struct ncclInfo* info) {
     WARN("%s : invalid type %d", info->opName, info->datatype);
     return ncclInvalidArgument;
   }
-  // Type is OK, compute nbytes. Convert Allgather/Broadcast calls to chars.
+  // Type is OK, compute nbytes. Convert Allgather/Broadcast/P2P calls to chars.
   info->nBytes = info->count * ncclTypeSize(info->datatype);
   if (info->coll == ncclCollAllGather || info->coll == ncclCollBroadcast) {
     info->count = info->nBytes;
@@ -58,12 +57,20 @@ ncclResult_t ArgsCheck(struct ncclInfo* info) {
   }
 
   if (info->comm->checkPointers) {
-    // Check CUDA device pointers
-    if (info->coll != ncclCollBroadcast || info->comm->rank == info->root) {
-      NCCLCHECK(CudaPtrCheck(info->sendbuff, info->comm, "sendbuff", info->opName));
-    }
-    if (info->coll != ncclCollReduce || info->comm->rank == info->root) {
-      NCCLCHECK(CudaPtrCheck(info->recvbuff, info->comm, "recvbuff", info->opName));
+    if (info->coll == ncclCollSendRecv) {
+      if (strcmp(info->opName, "Send") == 0) {
+        NCCLCHECK(CudaPtrCheck(info->sendbuff, info->comm, "sendbuff", "Send"));
+      } else {
+        NCCLCHECK(CudaPtrCheck(info->recvbuff, info->comm, "recvbuff", "Recv"));
+      }
+    } else {
+      // Check CUDA device pointers
+      if (info->coll != ncclCollBroadcast || info->comm->rank == info->root) {
+        NCCLCHECK(CudaPtrCheck(info->sendbuff, info->comm, "sendbuff", info->opName));
+      }
+      if (info->coll != ncclCollReduce || info->comm->rank == info->root) {
+        NCCLCHECK(CudaPtrCheck(info->recvbuff, info->comm, "recvbuff", info->opName));
+      }
     }
   }
   return ncclSuccess;

projects/rccl/src/misc/utils.cc

@@ -93,6 +93,7 @@ uint64_t getHostHash(void) {
   int offset = strlen(hostHash);
 
   if ((hostId = getenv("NCCL_HOSTID")) != NULL) {
+    INFO(NCCL_ENV, "NCCL_HOSTID set by environment to %s", hostId);
     strncpy(hostHash, hostId, sizeof(hostHash));
   } else {
     FILE *file = fopen(HOSTID_FILE, "r");