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

[ROCm/rccl commit: 4f0e223db4]
This commit is contained in:
Wenkai Du
2022-10-20 15:40:03 +00:00
51 changed files with 2139 additions and 1081 deletions
+2 -6
View File
@@ -335,12 +335,8 @@ ncclResult_t ncclGpuGdrSupport(struct ncclComm* comm, int* gdrSupport) {
ncclResult_t ret;
ncclDebugNoWarn = NCCL_NET;
NCCLCHECKGOTO(ncclNetListen(comm, dev, &handle, &lComm), ret, cleanup1);
while (sComm == NULL) {
NCCLWAITGOTO(ncclNetConnect(comm, dev, &handle, &sComm), sComm != NULL, comm->abortFlag, ret, cleanup2);
}
while (rComm == NULL) {
NCCLWAITGOTO(ncclNetAccept(comm, lComm, &rComm), rComm != NULL, comm->abortFlag, ret, cleanup3);
}
NCCLWAITGOTO(ncclNetConnect(comm, dev, &handle, &sComm), sComm != NULL, comm->abortFlag, ret, cleanup2);
NCCLWAITGOTO(ncclNetAccept(comm, lComm, &rComm), rComm != NULL, comm->abortFlag, ret, cleanup3);
CUDACHECKGOTO(hipMalloc(&gpuPtr, GPU_BUF_SIZE), ret, cleanup4);
if (ncclNetRegMr(comm, sComm, gpuPtr, GPU_BUF_SIZE, NCCL_PTR_CUDA, &mHandle) == ncclSuccess) {
NCCLCHECK(ncclNetDeregMr(comm, sComm, mHandle));