Add MSCCL Support (#658)

* Add MSCCL support

* Add alignment and message size checking

* Fix nRanks checking, in-place and out-of-place tests and group call handling

* Fix hipGraph unit test

* Change MSCCL init warning to INFO

* Revise license info

[ROCm/rccl commit: adafc0f759]

projects/rccl/src/group.cc

@@ -1,6 +1,7 @@
 /*************************************************************************
  * Copyright (c) 2015-2022, NVIDIA CORPORATION. All rights reserved.
  * Modifications Copyright (c) 2019-2022 Advanced Micro Devices, Inc. All rights reserved.
+ * Modifications Copyright (c) Microsoft Corporation. Licensed under the MIT License.
  *
  * See LICENSE.txt for license information
  ************************************************************************/
@@ -12,6 +13,8 @@
 #include "channel.h"
 #include <assert.h>
 
+#include "msccl/msccl_lifecycle.h"
+
 __thread int ncclGroupDepth = 0; // depth of ncclGroupStart nesting
 __thread ncclResult_t ncclGroupError = ncclSuccess;
 __thread struct ncclComm* ncclGroupCommHead = nullptr;
@@ -100,6 +103,14 @@ exit:
   return ret;
 }
 
+ncclResult_t ncclGroupStartInternal() {
+  ncclGroupDepth++;
+  if (mscclAvailable() && !mscclIsCaller()) {
+    NCCLCHECK(mscclGroupStart());
+  }
+  return ncclSuccess;
+}
+
 NCCL_API(ncclResult_t, ncclGroupEnd);
 ncclResult_t ncclGroupEnd() {
   ncclResult_t ret = ncclSuccess;
@@ -386,6 +397,10 @@ ncclResult_t ncclGroupEndInternal() {
     goto exit;
   }
 
+  if (mscclAvailable() && !mscclIsCaller()) {
+    NCCLCHECK(mscclGroupEnd());
+  }
+
   if ((--ncclGroupDepth) > 0) goto exit;
 
   if ((ret = ncclGroupError) != ncclSuccess) goto fail;