Optimize CUDA graph launch; avoid launching a CPU callback for
intra-node operations.
Simplify kernel common code to improve the latency of send/recv
operations.
Strengthen CUDA streams semantics.
Change NET API to v6, to add dmabuf support.
Add ncclGetLastError() function.
Add ncclRemoteError code and use it for remote network errors.
Support the use of a different NCCL_NET parameter per communicator.
Add support for SHM and P2P transfers using cudaMemcpy.


[ROCm/rccl commit: 19ab67d172]
This commit is contained in:
Sylvain Jeaugey
2022-05-24 02:02:31 -07:00
parent 1c5734046d
commit 91154e8df9
62 changed files with 4787 additions and 2496 deletions
+2 -2
View File
@@ -23,7 +23,7 @@ ncclResult_t ncclTopoGetSystem(struct ncclComm* comm, struct ncclTopoSystem** sy
ncclResult_t ncclTopoSortSystem(struct ncclTopoSystem* system);
ncclResult_t ncclTopoPrint(struct ncclTopoSystem* system);
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclPeerInfo* info);
ncclResult_t ncclTopoComputePaths(struct ncclTopoSystem* system, struct ncclComm* comm);
void ncclTopoFree(struct ncclTopoSystem* system);
ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm* comm);
ncclResult_t ncclTopoComputeP2pChannels(struct ncclComm* comm);
@@ -33,7 +33,7 @@ ncclResult_t ncclTopoGetNvbGpus(struct ncclTopoSystem* system, int rank, int* nr
ncclResult_t ncclTopoGetNetDev(struct ncclComm* comm, int rank, struct ncclTopoGraph* graph, int channelId, int peerRank, int* net, int* proxyRank);
ncclResult_t ncclTopoCheckP2p(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* p2p, int *read, int* intermediateRank);
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int64_t busId, int netDev, int read, int* useGdr);
int ncclPxnDisable();
int ncclPxnDisable(struct ncclComm* comm);
ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks, int* nranks);
ncclResult_t ncclTopoGetLocalRank(struct ncclTopoSystem* system, int rank, int* localRank);