2.12.7-1

Add network communication through another GPU connected with NVLink
(PXN).
Add aggregation of messages coming from different local GPUs through
PXN and going to the same destination.
Add new v5 plugin API with grouped receives and tags.
Add compat for v4 plugins.
Add naming of NCCL threads to help debugging.
Fix NVLink detection and avoid data corruption when some NVLinks are
down.
Add support for Relaxed Ordering for IB.
Add profiling and timing infrastructure.


[ROCm/rccl commit: 3c223c105a]

projects/rccl/src/graph/xml.cc

@@ -1,5 +1,5 @@
 /*************************************************************************
- * Copyright (c) 2019-2021, NVIDIA CORPORATION. All rights reserved.
+ * Copyright (c) 2019-2022, NVIDIA CORPORATION. All rights reserved.
  *
  * See LICENSE.txt for license information
  ************************************************************************/
@@ -602,7 +602,7 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
       NCCLCHECK(xmlGetAttr(pciNode, "busid", &busId));
       if (busId == NULL || cudaDeviceGetByPCIBusId(&dev, busId) != cudaSuccess) dev = -1;
     } else {
-      NCCLCHECK(wrapNvmlDeviceGetIndex(nvmlDev, (unsigned int*)&dev));
+      NCCLCHECK(ncclNvmlDeviceGetIndex(nvmlDev, (unsigned int*)&dev));
     }
     NCCLCHECK(xmlSetAttrInt(gpuNode, "dev", dev));
   }
@@ -617,7 +617,7 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
       CUDACHECK(cudaGetDeviceProperties(&devProp, dev));
       cudaMajor = devProp.major; cudaMinor = devProp.minor;
     } else {
-      NCCLCHECK(wrapNvmlDeviceGetCudaComputeCapability(nvmlDev, &cudaMajor, &cudaMinor));
+      NCCLCHECK(ncclNvmlDeviceGetCudaComputeCapability(nvmlDev, &cudaMajor, &cudaMinor));
     }
     NCCLCHECK(xmlSetAttrInt(gpuNode, "sm", cudaMajor*10+cudaMinor));
   }
@@ -638,15 +638,15 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, nvmlDevice_t nvm
     for (int l=0; l<maxNvLinks; ++l) {
       // Check whether we can use this NVLink for P2P
       unsigned canP2P;
-      if ((wrapNvmlDeviceGetNvLinkCapability(nvmlDev, l, NVML_NVLINK_CAP_P2P_SUPPORTED, &canP2P) != ncclSuccess) || !canP2P) continue;
+      if ((ncclNvmlDeviceGetNvLinkCapability(nvmlDev, l, NVML_NVLINK_CAP_P2P_SUPPORTED, &canP2P) != ncclSuccess) || !canP2P) continue;
 
       // Make sure the Nvlink is up. The previous call should have trained the link.
       nvmlEnableState_t isActive;
-      if ((wrapNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive) != ncclSuccess) || (isActive != NVML_FEATURE_ENABLED)) continue;
+      if ((ncclNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive) != ncclSuccess) || (isActive != NVML_FEATURE_ENABLED)) continue;
 
       // Try to figure out what's on the other side of the NVLink
       nvmlPciInfo_t remoteProc;
-      if (wrapNvmlDeviceGetNvLinkRemotePciInfo(nvmlDev, l, &remoteProc) != ncclSuccess) continue;
+      if (ncclNvmlDeviceGetNvLinkRemotePciInfo(nvmlDev, l, &remoteProc) != ncclSuccess) continue;
 
       // Make a lower case copy of the bus ID for calling ncclDeviceType
       // PCI system path is in lower case
@@ -701,13 +701,7 @@ ncclResult_t ncclTopoFillGpu(struct ncclXml* xml, const char* busId, struct nccl
   NCCLCHECK(xmlSetAttrIfUnset(node, "class", "0x03"));
   NCCLCHECK(ncclTopoGetXmlFromSys(node, xml));
   nvmlDevice_t nvmlDev = NULL;
-  static int nvmlInit = 0;
-  if (nvmlInit == 0) {
-    nvmlInit = (wrapNvmlSymbols() != ncclSuccess || wrapNvmlInit() != ncclSuccess) ? 2 : 1;
-  }
-  if (nvmlInit == 1) {
-    if (wrapNvmlDeviceGetHandleByPciBusId(busId, &nvmlDev) != ncclSuccess) nvmlDev = NULL;
-  }
+  if (ncclNvmlDeviceGetHandleByPciBusId(busId, &nvmlDev) != ncclSuccess) nvmlDev = NULL;
   NCCLCHECK(ncclTopoGetXmlFromGpu(node, nvmlDev, xml, gpuNode));
   return ncclSuccess;
 }