Tune number of channels for gfx90a (#349)

src/graph/paths.cc

@@ -573,7 +573,7 @@ ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm) {
     }
   }
 
-  if (comm->topo->nodes[GPU].count == comm->topo->nRanks && (comm->topo->type & RCCL_TOPO_4P2H_ROME) && !(comm->topo->type & RCCL_TOPO_GDR_ALL)) {
+  if (comm->topo->nodes[GPU].count == comm->topo->nRanks && (comm->topo->type & RCCL_TOPO_4P2H_ROME) && !(comm->topo->type & RCCL_TOPO_GDR_ALL) && (comm->topo->nodes[GPU].nodes[0].gpu.gcn != 910)) {
     // Adjust P2P channels on Rome
     comm->p2pnChannelsPerPeer = 2;
     comm->p2pnChannels = 2;

src/init.cc

@@ -949,6 +949,8 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* comm
     allGather3Data[rank].nc = comm->nChannels*4;
   if (comm->topo->nodes[GPU].count != comm->topo->nRanks && comm->topo->nodes[NET].count && (comm->topo->type & RCCL_TOPO_4P2H_ROME))
     allGather3Data[rank].nc = (comm->topo->nodes[NET].count > 3 ? 2 : 4)*comm->topo->nodes[NET].count;
+  if (comm->topo->nodes[GPU].count == comm->topo->nRanks && comm->topo->nodes[GPU].nodes[idx].gpu.gcn == 910)
+    allGather3Data[rank].nc = comm->nChannels*6;
   allGather3Data[rank].tree.pattern = treeGraph.pattern;
   allGather3Data[rank].tree.sameChannels = treeGraph.sameChannels;
   allGather3Data[rank].tree.speedIntra = treeGraph.speedIntra;