Merge remote-tracking branch 'nccl-tests/master' into develop

[ROCm/rccl-tests commit: 5625599dda]

projects/rccl-tests/src/all_reduce.cu

@@ -8,7 +8,7 @@
 #include "cuda_runtime.h"
 #include "common.h"
 
-void AllReduceGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+void AllReduceGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, size_t eltSize, int nranks) {
   *sendcount = count;
   *recvcount = count;
   *sendInplaceOffset = 0;
@@ -56,7 +56,7 @@ struct testColl allReduceTest = {
 
 void AllReduceGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
   size_t paramcount, sendInplaceOffset, recvInplaceOffset;
-  AllReduceGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+  AllReduceGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, /*eltSize=*/1, nranks);
 }
 
 testResult_t AllReduceRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {