Enable local sendrecv over network if GDR is available on all GPUs (#324)

[ROCm/rccl commit: c018edf0f2]

projects/rccl/src/include/comm.h

@@ -121,6 +121,9 @@ struct ncclComm {
   // Flag to ask NCCL kernels to abort
   volatile uint32_t *abortFlag;
 
+  // Flags for enable P2P NET
+  uint32_t *p2pNet;
+
   // Device side of the communicator
   struct ncclDevComm *devComm;
   // Host copy of the devComm (to free CUDA allocs)

projects/rccl/src/include/devcomm.h

@@ -150,6 +150,8 @@ struct ncclTree {
 struct ncclPeer {
   struct ncclConnector send;
   struct ncclConnector recv;
+  struct ncclConnector p2pSend;
+  struct ncclConnector p2pRecv;
 };
 
 struct ncclDevComm;
@@ -332,6 +334,9 @@ struct ncclDevComm {
   // Channels, device side
   struct ncclChannel* channels;
 
+  // Flags for enable P2P NET
+  uint32_t *p2pNet;
+
 #ifdef ENABLE_PROFILING
   // Profiling counters
   struct ncclProf* devProf;