From 04cd446d89e4092a8dda9da91494fac87f35e9c3 Mon Sep 17 00:00:00 2001 From: Wenkai Du Date: Fri, 16 Aug 2019 16:34:14 -0700 Subject: [PATCH] Assign unused nmvlDev to avoid random number [ROCm/rccl commit: 86efdfc3b55ec231c213266be88ac619efbadbf1] --- projects/rccl/src/misc/utils.cc | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/projects/rccl/src/misc/utils.cc b/projects/rccl/src/misc/utils.cc index 614c78b936..ec0c0502c6 100644 --- a/projects/rccl/src/misc/utils.cc +++ b/projects/rccl/src/misc/utils.cc @@ -17,6 +17,10 @@ // Convert a logical cudaDev index to the NVML device minor number ncclResult_t getNvmlDevice(int cudaDev, int *nvmlDev) { +#if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__) + // assign nmvlDev to be same as cudaDev to avoid garbage numbers + *nvmlDev = cudaDev; +#else char busId[NVML_DEVICE_PCI_BUS_ID_BUFFER_SIZE]; nvmlDevice_t nvmlDevice; unsigned int dev; @@ -26,6 +30,7 @@ ncclResult_t getNvmlDevice(int cudaDev, int *nvmlDev) { NCCLCHECK(wrapNvmlDeviceGetMinorNumber(nvmlDevice, &dev)); *nvmlDev = dev; +#endif return ncclSuccess; }