gfx11: don't use LL for sendrecv (#853)

* gfx11: don't use LL for sendrecv

* Use builtin instead of inline asm

src/enqueue.cc

@@ -318,7 +318,8 @@ static ncclResult_t addP2pToPlan(
   // 1 is connIndex
   struct ncclConnInfo* conn = isSendNotRecv ?
     &comm->channels[channelId].peers[peer]->send[1].conn : &comm->channels[channelId].peers[peer]->recv[1].conn;
-  info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold()) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
+  // do not use LL on gfx11
+  info.protocol = ((conn->buffs[NCCL_PROTO_LL] != nullptr) && bytes <= ncclParamP2pLLThreshold() && comm->topo->nodes[GPU].nodes[0].gpu.gcn/100 != 11) ? NCCL_PROTO_LL : NCCL_PROTO_SIMPLE;
 
   struct ncclProxyOp proxyOp = {};
   NCCLCHECK(ncclProxyComputeP2p(&info, &proxyOp));