Merge remote-tracking branch 'nccl/master' into HEAD

[ROCm/rccl-tests commit: 5361dd8177]

projects/rccl-tests/.gitignore

@@ -0,0 +1,4 @@
+# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+#
+# See LICENCE.txt for license information
+/build

projects/rccl-tests/README.md

@@ -29,9 +29,9 @@ Run on 8 GPUs (`-g 8`), scanning from 8 Bytes to 128MBytes :
 $ ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 8
 ```
 
-Run with MPI on 40 processes (potentially on multiple nodes) with 4 GPUs each, disabling checks :
+Run with MPI on 40 processes (potentially on multiple nodes) with 4 GPUs each :
 ```shell
-$ mpirun -np 40 ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 4 -c 0
+$ mpirun -np 40 ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 4
 ```
 
 ### Performance

projects/rccl-tests/src/all_gather.cu

@@ -49,10 +49,10 @@ testResult_t AllGatherInitData(struct threadArgs* args, ncclDataType_t type, ncc
 }
 
 void AllGatherGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
-  double baseBw = (double)(count * typesize * (nranks - 1)) / 1.0E9 / sec;
+  double baseBw = (double)(count * typesize * nranks) / 1.0E9 / sec;
 
   *algBw = baseBw;
-  double factor = 1;
+  double factor = ((double)(nranks - 1))/((double)nranks);
   *busBw = baseBw * factor;
 }
 

projects/rccl-tests/src/alltoall.cu

@@ -67,7 +67,10 @@ testResult_t AlltoAllRunColl(void* sendbuff, void* recvbuff, size_t count, ncclD
   size_t rankOffset = count * wordSize(type);
   if (count == 0) return testSuccess;
 
-#if NCCL_MAJOR >= 2 && NCCL_MINOR >= 7
+#if NCCL_MAJOR < 2 || NCCL_MINOR < 7
+  printf("NCCL 2.7 or later is needed for alltoall. This test was compiled with %d.%d.\n", NCCL_MAJOR, NCCL_MINOR);
+  return testNcclError;
+#else
 #if defined(RCCL_GATHER_SCATTER) && defined(USE_RCCL_GATHER_SCATTER)
   NCCLCHECK(ncclAllToAll(sendbuff, recvbuff, count, type, comm, stream));
 #else
@@ -78,9 +81,8 @@ testResult_t AlltoAllRunColl(void* sendbuff, void* recvbuff, size_t count, ncclD
   }
   NCCLCHECK(ncclGroupEnd());
 #endif
-#endif
-
   return testSuccess;
+#endif
 }
 
 struct testColl alltoAllTest = {

projects/rccl-tests/src/common.cu

@@ -352,9 +352,9 @@ testResult_t CheckData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t
          printf("%d:%d ", j, dataHost[j]);
        }
        printf("\n");
-       free(dataHost);
        free(expectedHost);
-    //}
+       free(dataHost);
+    }
 #endif
   }
   double nranks = args->nProcs*args->nThreads*args->nGpus;

projects/rccl-tests/src/reduce_scatter.cu

@@ -48,10 +48,10 @@ testResult_t ReduceScatterInitData(struct threadArgs* args, ncclDataType_t type,
 }
 
 void ReduceScatterGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
-  double baseBw = (double)(count * typesize * (nranks - 1)) / 1.0E9 / sec;
+  double baseBw = (double)(count * typesize * nranks) / 1.0E9 / sec;
 
   *algBw = baseBw;
-  double factor = 1;
+  double factor = ((double)(nranks - 1))/((double)nranks);
   *busBw = baseBw * factor;
 }