Add support for external network.

Dynamically load external network from libnccl-net.so.
Add init function in networks.
Move PCI scoring to net.cu, only ask transport to provide a path.
Simplify CUDA PCI path detection.
Add dummy external network

src/collectives/all_gather.cu

@@ -12,7 +12,7 @@
 ncclResult_t ncclAllGatherFunc(const void* sendbuff, void* recvbuff, size_t count,
     ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
   size_t nbytes = count*ncclTypeSize(datatype);
-  INFO(COLL,"AllGather: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
+  INFO(NCCL_COLL,"AllGather: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
   if (comm->nRanks == 1) {
     if (sendbuff != recvbuff)
       CUDACHECK(cudaMemcpyAsync(recvbuff, sendbuff, nbytes, cudaMemcpyDeviceToDevice, stream));

src/collectives/all_reduce.cu

@@ -12,7 +12,7 @@
 ncclResult_t ncclAllReduceFunc(const void* sendbuff, void* recvbuff, size_t count,
     ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
   size_t nbytes = count*ncclTypeSize(datatype);
-  INFO(COLL,"AllReduce: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
+  INFO(NCCL_COLL,"AllReduce: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
   if (comm->nRanks == 1) {
     if (sendbuff != recvbuff)
       CUDACHECK(cudaMemcpyAsync(recvbuff, sendbuff, nbytes, cudaMemcpyDeviceToDevice, stream));

src/collectives/broadcast.cu

@@ -12,7 +12,7 @@
 ncclResult_t ncclBroadcastFunc(const void* sendbuff, void* recvbuff, const size_t count,
     ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
   size_t nbytes = count*ncclTypeSize(datatype);
-  INFO(COLL,"Broadcast: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
+  INFO(NCCL_COLL,"Broadcast: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
   if (comm->nRanks == 1) {
     if (sendbuff != recvbuff)
       CUDACHECK(cudaMemcpyAsync(recvbuff, sendbuff, nbytes, cudaMemcpyDeviceToDevice, stream));

src/collectives/reduce.cu

@@ -12,7 +12,7 @@
 ncclResult_t ncclReduceFunc(const void* sendbuff, void* recvbuff, const size_t count,
     ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
   size_t nbytes = count*ncclTypeSize(datatype);
-  INFO(COLL,"Reduce: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
+  INFO(NCCL_COLL,"Reduce: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
   if (comm->nRanks == 1) {
     if (sendbuff != recvbuff)
       CUDACHECK(cudaMemcpyAsync(recvbuff, sendbuff, nbytes, cudaMemcpyDeviceToDevice, stream));

src/collectives/reduce_scatter.cu

@@ -12,7 +12,7 @@
 ncclResult_t ncclReduceScatterFunc(const void* sendbuff, void* recvbuff, size_t count,
     ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream) {
   size_t nbytes = count*ncclTypeSize(datatype);
-  INFO(COLL,"ReduceScatter: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
+  INFO(NCCL_COLL,"ReduceScatter: opCount %lx sendbuff %p recvbuff %p count %zi datatype %d op %d root %d comm %p [nranks=%d] stream %p", comm->opCount, sendbuff, recvbuff, count, datatype, op, root, comm, comm->nRanks, stream);
   if (comm->nRanks == 1) {
     if (sendbuff != recvbuff)
       CUDACHECK(cudaMemcpyAsync(recvbuff, sendbuff, nbytes, cudaMemcpyDeviceToDevice, stream));