Collective Communication Functions — NCCL 2.26.2 documentation (original) (raw)

The following NCCL APIs provide some commonly used collective operations.

ncclAllReduce¶

ncclResult_t ncclAllReduce(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, ncclRedOp_t op, ncclComm_t comm, cudaStream_t stream)¶

Reduces data arrays of length count in sendbuff using the op operation and leaves identical copies of the result in each recvbuff.

In-place operation will happen if sendbuff == recvbuff.

ncclBroadcast¶

ncclResult_t ncclBroadcast(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, int root, ncclComm_t comm, cudaStream_t stream)¶

Copies count elements from sendbuff on the root rank to all ranks’ recvbuff.sendbuff is only used on rank root and ignored for other ranks.

In-place operation will happen if sendbuff == recvbuff.

ncclResult_t ncclBcast(void* buff, size_t count, ncclDataType_t datatype, int root, ncclComm_t comm, cudaStream_t stream)¶

Legacy in-place version of ncclBroadcast in a similar fashion to MPI_Bcast. A call to

ncclBcast(buff, count, datatype, root, comm, stream)

is equivalent to

ncclBroadcast(buff, buff, count, datatype, root, comm, stream)

ncclReduce¶

ncclResult_t ncclReduce(const void* sendbuff, void* recvbuff, size_t count, ncclDataType_t datatype, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream)¶

Reduce data arrays of length count in sendbuff into recvbuff on the root rank using the op operation.recvbuff is only used on rank root and ignored for other ranks.

In-place operation will happen if sendbuff == recvbuff.

ncclAllGather¶

ncclResult_t ncclAllGather(const void* sendbuff, void* recvbuff, size_t sendcount, ncclDataType_t datatype, ncclComm_t comm, cudaStream_t stream)¶

Gathers sendcount values from all GPUs and leaves identical copies of the result in each recvbuff, receiving data from rank i at offset i*sendcount.

Note: This assumes the receive count is equal to nranks*sendcount, which means that recvbuff should have a size of at least nranks*sendcount elements.

In-place operation will happen if sendbuff == recvbuff + rank * sendcount.

Related links: AllGather, In-place Operations.

ncclReduceScatter¶

ncclResult_t ncclReduceScatter(const void* sendbuff, void* recvbuff, size_t recvcount, ncclDataType_t datatype, ncclRedOp_t op, ncclComm_t comm, cudaStream_t stream)¶

Reduce data in sendbuff from all GPUs using the op operation and leave the reduced result scattered over the devices so that the recvbuff on rank i will contain the i-th block of the result.

Note: This assumes the send count is equal to nranks*recvcount, which means that sendbuff should have a size of at least nranks*recvcount elements.

In-place operation will happen if recvbuff == sendbuff + rank * recvcount.

Related links: ReduceScatter, In-place Operations.