API

This section provides details of the library API

Communicator Functions

ncclResult_t ncclGetUniqueId(ncclUniqueId *uniqueId)

Generates an ID for ncclCommInitRank.

Generates an ID to be used in ncclCommInitRank. ncclGetUniqueId should be called once and the Id should be distributed to all ranks in the communicator before calling ncclCommInitRank.

Parameters

uniqueId[in] ncclUniqueId* pointer to uniqueId

ncclResult_t ncclCommInitRank(ncclComm_t *comm, int nranks, ncclUniqueId commId, int rank)

Creates a new communicator (multi thread/process version).

rank must be between 0 and nranks-1 and unique within a communicator clique. Each rank is associated to a CUDA device, which has to be set before calling ncclCommInitRank. ncclCommInitRank implicitly syncronizes with other ranks, so it must be called by different threads/processes or use ncclGroupStart/ncclGroupEnd.

Parameters

comm[in] ncclComm_t* communicator struct pointer

ncclResult_t ncclCommInitAll(ncclComm_t *comm, int ndev, const int *devlist)

Creates a clique of communicators (single process version).

This is a convenience function to create a single-process communicator clique. Returns an array of ndev newly initialized communicators in comm. comm should be pre-allocated with size at least ndev*sizeof(ncclComm_t). If devlist is NULL, the first ndev HIP devices are used. Order of devlist defines user-order of processors within the communicator.

ncclResult_t ncclCommDestroy(ncclComm_t comm)

Frees local resources associated with communicator object.

ncclResult_t ncclCommAbort(ncclComm_t comm)

Frees resources associated with communicator object and aborts any operations that might still be running on the device.

ncclResult_t ncclCommCount(const ncclComm_t comm, int *count)

Gets the number of ranks in the communicator clique.

ncclResult_t ncclCommCuDevice(const ncclComm_t comm, int *device)

Returns the rocm device number associated with the communicator.

ncclResult_t ncclCommUserRank(const ncclComm_t comm, int *rank)

Returns the user-ordered “rank” associated with the communicator.

Collective Communication Operations

Collective communication operations must be called separately for each communicator in a communicator clique.

They return when operations have been enqueued on the hipstream.

Since they may perform inter-CPU synchronization, each call has to be done from a different thread or process, or need to use Group Semantics (see below).

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

Reduce.

Reduces data arrays of length count in sendbuff into recvbuff using op operation. recvbuff may be NULL on all calls except for root device. root is the rank (not the CUDA device) where data will reside after the operation is complete.

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

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

(deprecated) Broadcast (in-place)

Copies count values from root to all other devices. root is the rank (not the CUDA device) where data resides before the operation is started.

This operation is implicitely in place.

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

Broadcast.

Copies count values from root to all other devices. root is the rank (not the HIP device) where data resides before the operation is started.

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

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

All-Reduce.

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

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

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

Reduce-Scatter.

Reduces data in sendbuff using op operation and leaves reduced result scattered over the devices so that recvbuff on rank i will contain the i-th block of the result. Assumes sendcount is equal to nranks*recvcount, which means that sendbuff should have a size of at least nranks*recvcount elements.

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

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

All-Gather.

Each device gathers sendcount values from other GPUs into recvbuff, receiving data from rank i at offset i*sendcount. Assumes recvcount is equal to nranks*sendcount, which means that recvbuff should have a size of at least nranks*sendcount elements.

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

ncclResult_t ncclSend(const void *sendbuff, size_t count, ncclDataType_t datatype, int peer, ncclComm_t comm, hipStream_t stream)

Send.

Send data from sendbuff to rank peer. Rank peer needs to call ncclRecv with the same datatype and the same count from this rank.

This operation is blocking for the GPU. If multiple ncclSend and ncclRecv operations need to progress concurrently to complete, they must be fused within a ncclGroupStart/ ncclGroupEnd section.

ncclResult_t ncclRecv(void *recvbuff, size_t count, ncclDataType_t datatype, int peer, ncclComm_t comm, hipStream_t stream)

Receive.

Receive data from rank peer into recvbuff. Rank peer needs to call ncclSend with the same datatype and the same count to this rank.

This operation is blocking for the GPU. If multiple ncclSend and ncclRecv operations need to progress concurrently to complete, they must be fused within a ncclGroupStart/ ncclGroupEnd section.

ncclResult_t ncclGather(const void *sendbuff, void *recvbuff, size_t sendcount, ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream)

Gather.

Root device gathers sendcount values from other GPUs into recvbuff, receiving data from rank i at offset i*sendcount.

Assumes recvcount is equal to nranks*sendcount, which means that recvbuff should have a size of at least nranks*sendcount elements.

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

ncclResult_t ncclScatter(const void *sendbuff, void *recvbuff, size_t recvcount, ncclDataType_t datatype, int root, ncclComm_t comm, hipStream_t stream)

Scatter.

Scattered over the devices so that recvbuff on rank i will contain the i-th block of the data on root.

Assumes sendcount is equal to nranks*recvcount, which means that sendbuff should have a size of at least nranks*recvcount elements.

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

ncclResult_t ncclAllToAll(const void *sendbuff, void *recvbuff, size_t count, ncclDataType_t datatype, ncclComm_t comm, hipStream_t stream)

All-To-All.

Device (i) send (j)th block of data to device (j) and be placed as (i)th block. Each block for sending/receiving has count elements, which means that recvbuff and sendbuff should have a size of nranks*count elements.

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

Group Semantics

When managing multiple GPUs from a single thread, and since NCCL collective calls may perform inter-CPU synchronization, we need to “group” calls for different ranks/devices into a single call.

Grouping NCCL calls as being part of the same collective operation is done using ncclGroupStart and ncclGroupEnd. ncclGroupStart will enqueue all collective calls until the ncclGroupEnd call, which will wait for all calls to be complete. Note that for collective communication, ncclGroupEnd only guarantees that the operations are enqueued on the streams, not that the operation is effectively done.

Both collective communication and ncclCommInitRank can be used in conjunction of ncclGroupStart/ncclGroupEnd.

ncclResult_t ncclGroupStart()

Group Start.

Start a group call. All calls to NCCL until ncclGroupEnd will be fused into a single NCCL operation. Nothing will be started on the CUDA stream until ncclGroupEnd.

ncclResult_t ncclGroupEnd()

Group End.

End a group call. Start a fused NCCL operation consisting of all calls since ncclGroupStart. Operations on the CUDA stream depending on the NCCL operations need to be called after ncclGroupEnd.

Library Functions

ncclResult_t ncclGetVersion(int *version)

Return the NCCL_VERSION_CODE of the NCCL library in the supplied integer.

This integer is coded with the MAJOR, MINOR and PATCH level of the NCCL library

const char *ncclGetErrorString(ncclResult_t result)

Returns a string for each error code.

Types

There are few data structures that are internal to the library. The pointer types to these structures are given below. The user would need to use these types to create handles and pass them between different library functions.

typedef struct ncclComm *ncclComm_t

Opaque handle to communicator.

struct ncclUniqueId

Enumerations

This section provides all the enumerations used.

enum ncclResult_t

Error type.

Values:

enumerator ncclSuccess
enumerator ncclUnhandledCudaError
enumerator ncclSystemError
enumerator ncclInternalError
enumerator ncclInvalidArgument
enumerator ncclInvalidUsage
enumerator ncclRemoteError
enumerator ncclInProgress
enumerator ncclNumResults
enum ncclRedOp_t

Values:

enumerator ncclSum
enumerator ncclProd
enumerator ncclMax
enumerator ncclMin
enumerator ncclAvg
enumerator ncclNumOps
enumerator ncclMaxRedOp
enum ncclDataType_t

Data types.

Values:

enumerator ncclInt8
enumerator ncclChar
enumerator ncclUint8
enumerator ncclInt32
enumerator ncclInt
enumerator ncclUint32
enumerator ncclInt64
enumerator ncclUint64
enumerator ncclFloat16
enumerator ncclHalf
enumerator ncclFloat32
enumerator ncclFloat
enumerator ncclFloat64
enumerator ncclDouble
enumerator ncclBfloat16
enumerator ncclNumTypes