[P2P] Have connIdx for both send and recv (#1524)

[ROCm/rccl commit: 387c973b5d]

projects/rccl/src/device/common.h

@@ -63,7 +63,8 @@
       collTrace->p2p.nSendChannels = p2pWork->nSendChannels; \
       collTrace->p2p.nRecvChannels = p2pWork->nRecvChannels; \
       collTrace->p2p.channelBase = p2pWork->channelBase; \
-      collTrace->p2p.connIndex = p2pWork->connIndex; \
+      collTrace->p2p.sendConnIndex = p2pWork->sendConnIndex; \
+      collTrace->p2p.recvConnIndex = p2pWork->recvConnIndex; \
       collTrace->p2p.sendProtoLL = p2pWork->sendProtoLL; \
       collTrace->p2p.recvProtoLL = p2pWork->recvProtoLL; \
       collTrace->p2pOpCount[0] = p2pWork->sendOpCount; \

projects/rccl/src/device/sendrecv.h

@@ -42,7 +42,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
 
     Primitives<T, RedOp, FanAsymmetric<0, 1>, 0, Proto, 1>
       prims(tid, tn, nullptr, &work->sendRank, work->sendAddr, nullptr,
-            /*redOpArg(ignored)=*/0, group, work->connIndex, work->connIndex, nullptr,
+            /*redOpArg(ignored)=*/0, group, work->sendConnIndex, work->sendConnIndex, nullptr,
             /*userBufferMode=*/work->sendRegistered, ncclShmem.comm.p2pChunkSize);
 
 #if defined(ENABLE_NPKIT)
@@ -100,7 +100,7 @@ struct RunWorkBatch<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
 
     Primitives<T, RedOp, FanAsymmetric<1, 0>, 0, Proto, 1>
       prims(tid, tn, &work->recvRank, nullptr, nullptr, work->recvAddr,
-            /*redOpArg(ignored)=*/0, group, work->connIndex, work->connIndex, nullptr,
+            /*redOpArg(ignored)=*/0, group, work->recvConnIndex, work->recvConnIndex, nullptr,
             /*userBufferMode=*/work->recvRegistered, ncclShmem.comm.p2pChunkSize);
 
 #if defined(ENABLE_NPKIT)