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/include/comm.h

@@ -1,5 +1,5 @@
 /*************************************************************************
- * Copyright (c) 2015-2021, NVIDIA CORPORATION. All rights reserved.
+ * Copyright (c) 2015-2022, NVIDIA CORPORATION. All rights reserved.
  *
  * See LICENSE.txt for license information
  ************************************************************************/
@@ -31,8 +31,6 @@ struct cudaLaunchParams {
 #define NCCL_LL128_THREAD_THRESHOLD 8
 #define NCCL_SIMPLE_THREAD_THRESHOLD 64
 
-#define NCCL_MAX_INTRA_RANKS 32
-
 struct ncclSendMem {
   union {
     struct {
@@ -41,10 +39,10 @@ struct ncclSendMem {
       void* ptrExchange;
       uint64_t redOpArgExchange[2];
       char pad2[CACHE_LINE_SIZE-sizeof(void*)-2*sizeof(uint64_t)];
+      int offsFifo[NCCL_STEPS];
     };
     char pad3[MEM_ALIGN];
   };
-  char buff[1]; // Actually larger than that
 };
 
 struct ncclRecvMem {
@@ -53,18 +51,18 @@ struct ncclRecvMem {
       uint64_t tail;
       char pad1[CACHE_LINE_SIZE-sizeof(uint64_t)];
       int sizesFifo[NCCL_STEPS];
-      void* ptrsFifo[NCCL_STEPS];
+      int offsFifo[NCCL_STEPS];
+      int flush; // For GDRCopy-based flush
     };
     char pad4[MEM_ALIGN];
   };
-  char buff[1]; // Actually larger than that
 };
 
 typedef cudaError_t(*pfn_cuMemGetAddressRange_t)(void**, size_t*, void*);
 
 enum helperThreadState {ThreadStart, ThreadStop};
 
-#define NCCL_IPC_POOL_SIZE (2*NCCL_MAX_INTRA_RANKS*NCCL_MAX_OPS)
+#define NCCL_IPC_POOL_SIZE (2*NCCL_MAX_LOCAL_RANKS*NCCL_MAX_OPS)
 
 struct ncclGraphHelperResources {
   ncclComm* comm;
@@ -82,6 +80,11 @@ struct ncclUserRedOp {
   ncclDevRedOpFull opFull;
 };
 
+struct ncclNodeRanks {
+  int localRanks;
+  int* localRankToRank;
+};
+
 struct ncclComm {
   struct ncclChannel channels[MAXCHANNELS];
 
@@ -102,12 +105,14 @@ struct ncclComm {
 
   int node;
   int nNodes;
-
-  // Intra-node rank info
-  int intraNodeGlobalRanks[NCCL_MAX_INTRA_RANKS];
+  int localRank;
   int localRanks;
-  int intraNodeRank;
-  int8_t* rankToIntraNodeRank;
+  int maxLocalRanks;
+  int* rankToNode;
+  int* rankToLocalRank;
+  int* localRankToRank;
+  // localRanks and localRanktoRank for all nodes
+  struct ncclNodeRanks* nodeRanks;
 
   enum { GROUP, PARALLEL, GROUP_GRAPH } launchMode;
   cudaStream_t userStream;
@@ -161,14 +166,13 @@ struct ncclComm {
   // Storage for deferred intra-process launch
   struct cudaLaunchParams * intraParams;
   struct cudaLaunchParams *myParams;
+  pthread_t* intraThreads;
   int* intraCudaDevs;
   int* intraCGMode; // Whether we can use CUDA9 CGMD or not
   int* intraCC; // Only to check all have the same ComputeCap and disable CGMode if not
   struct ncclWorkElem args;
-  void* argsptr;
+  void* argsptrs[2];
 
-  // Global proxy thread
-  pthread_t proxyThread;
   struct ncclProxyState proxyState;
 
   // Whether this communicator uses collNet