Introduce multi-rank support per device.

This is a single commit of the source code changes required to
introduce support for multiple ranks per device.
A new interface (ncclCommRankInitMulti) has to be used to make use of
this new feature.

src/include/comm.h

@@ -112,6 +112,7 @@ struct ncclComm {
   int64_t busId;   // my PCI bus ID in int format
   cpu_set_t cpuAffinity; // CPU affinity of the GPU
   int WarpSize;
+  int virtualId;
 
   int node;
   int nNodes;

src/include/group.h

@@ -14,9 +14,9 @@
 bool ncclAsyncMode();
 ncclResult_t ncclAsyncErrCheck(ncclResult_t ret);
 
-typedef ncclResult_t(*ncclInitFunc_t)(ncclComm_t* newcomm, int ndev, ncclUniqueId commId, int myrank, int cudaDev);
+typedef ncclResult_t(*ncclInitFunc_t)(ncclComm_t* newcomm, int ndev, ncclUniqueId commId, int myrank, int cudaDev, int virtualId);
 
-ncclResult_t ncclAsyncInit(ncclInitFunc_t func, ncclComm_t* newcomm, int ndev, ncclUniqueId commId, int myrank, int cudaDev);
+ncclResult_t ncclAsyncInit(ncclInitFunc_t func, ncclComm_t* newcomm, int ndev, ncclUniqueId commId, int myrank, int cudaDev, int virtualId);
 
 typedef ncclResult_t(*ncclCollFunc_t)(const void* sendbuff, void* recvbuff, size_t count,
     ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, hipStream_t stream);

src/include/transport.h

@@ -40,6 +40,7 @@ struct ncclPeerInfo {
   int64_t busId;
   struct ncclComm* comm;
   int cudaCompCap;
+  int virtualId;
 };
 
 #define CONNECT_SIZE 128