Setup collectives threshold for enabling intranet (#387)

* Setup collectives threshold for enabling intranet

* Use separate operation counters for coll and p2p

src/collectives/device/all_reduce.h

@@ -37,7 +37,7 @@ class ncclFunction<ncclFuncAllReduce, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE, FUNC, T
     T * __restrict__ thisOutput = (T*)args->recvbuff;
 
     ncclPrimitives<UNROLL, ALLREDUCE_CHUNKSTEPS/ALLREDUCE_SLICESTEPS, ALLREDUCE_SLICESTEPS, T, 1, 1, 1, FUNC>
-      prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm, ncclShmem->ptrs, 0);
+      prims(tid, nthreads, &ring->prev, &ring->next, thisOutput, stepSize, channel, comm, ncclShmem->ptrs, args->coll.connIndex);
 
     for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += nranks*loopSize) {
       ssize_t realChunkSize = min(chunkSize, DIVUP(size-gridOffset,nranks*nChannels));