Device API

Host-Side Setup

ncclDevComm

type ncclDevComm

A structure describing a device communicator, as created on the host side using ncclDevCommCreate(). The structure is used primarily on the device side; elements that could be of particular interest include:

int rank

The rank within the communicator.

int nRanks

The size of the communicator.

int lsaRank

int lsaSize

Rank within the local LSA team and its size (see Teams).

uint8_t ginContextCount

The number of supported GIN contexts (see ncclGin; available since NCCL 2.28.7).

ncclDevCommCreate

ncclResult_t ncclDevCommCreate(ncclComm_t comm, struct ncclDevCommRequirements const *reqs, struct ncclDevComm *outDevComm)

Creates a new device communicator (see ncclDevComm) corresponding to the supplied host-side communicator comm. The result is returned in the outDevComm buffer (which needs to be supplied by the caller). The caller needs to also provide a filled-in list of requirements via the reqs argument (see ncclDevCommRequirements); the function will allocate any necessary resources to meet them. It is recommended to call ncclCommQueryProperties() before calling the function; the function will fail if the specified requirements are not supported. Since this is a collective call, every rank in the communicator needs to participate. If called within a group, outDevComm may not be filled in until ncclGroupEnd() has completed.

Note that this is a host-side function.

ncclDevCommDestroy

ncclResult_t ncclDevCommDestroy(ncclComm_t comm, struct ncclDevComm const *devComm)

Destroys a device communicator (see ncclDevComm) previously created using ncclDevCommCreate() and releases any allocated resources. The caller must ensure that no device kernel that uses this device communicator could be running at the time this function is invoked.

Note that this is a host-side function.

ncclDevCommRequirements

type ncclDevCommRequirements

A host-side structure specifying the list of requirements when creating device communicators (see ncclDevComm). Since NCCL 2.29, this struct must be initialized using NCCL_DEV_COMM_REQUIREMENTS_INITIALIZER.

bool lsaMultimem

Specifies whether multimem support is required for all LSA ranks.

int lsaBarrierCount

Specifies the number of memory barriers to allocate (see ncclLsaBarrierSession).

int railGinBarrierCount

Specifies the number of network barriers to allocate (see ncclGinBarrierSession; available since NCCL 2.28.7).

int barrierCount

Specifies the minimum number for both the memory and network barriers (see above; available since NCCL 2.28.7).

int ginSignalCount

Specifies the number of network signals to allocate (see ncclGinSignal_t; available since NCCL 2.28.7).

int ginCounterCount

Specifies the number of network counters to allocate (see ncclGinCounter_t; available since NCCL 2.28.7).

ncclDevResourceRequirements_t *resourceRequirementsList

Specifies a list of resource requirements. This is best set to NULL for now.

ncclTeamRequirements_t *teamRequirementsList

Specifies a list of requirements for particular teams. This is best set to NULL for now.

ncclCommQueryProperties

ncclResult_t ncclCommQueryProperties(ncclComm_t comm, ncclCommProperties_t *props)

Exposes communicator properties by filling in props. Before calling this function, props must be initialized using NCCL_COMM_PROPERTIES_INITIALIZER. Introduced in NCCL 2.29.

Note that this is a host-side function.

ncclCommProperties_t

type ncclCommProperties_t

A structure describing the properties of the communicator. Introduced in NCCL 2.29. Properties include:

int rank

Rank within the communicator.

int nRanks

Size of the communicator.

int cudaDev

CUDA device index.

int nvmlDev

NVML device index.

bool deviceApiSupport

Whether the device API is supported. If false, a ncclDevComm cannot be created.

bool multimemSupport

Whether ranks in the same LSA team can communicate using multimem. If false, a ncclDevComm cannot be created with multimem resources.

ncclGinType_t ginType

The GIN type supported by the communicator. If equal to NCCL_GIN_TYPE_NONE, a ncclDevComm cannot be created with GIN resources.

ncclGinType_t

type ncclGinType_t

GIN type. Communication between different GIN types is not supported. Possible values include:

NCCL_GIN_TYPE_NONE

GIN is not supported.

NCCL_GIN_TYPE_PROXY

Host Proxy GIN type.

NCCL_GIN_TYPE_GDAKI

GPUDirect Async Kernel-Initiated (GDAKI) GIN type.

LSA

All functionality described from this point on is available on the device side only.

ncclLsaBarrierSession

template<typename Coop> class ncclLsaBarrierSession

A class representing a memory barrier session.

ncclLsaBarrierSession(Coop coop, ncclDevComm const &comm, ncclTeamTagLsa tag, uint32_t index, bool multimem = false): Initializes a new memory barrier session. coop represents a cooperative group (see Teams). comm is the device communicator created using ncclDevCommCreate(). ncclTeamTagLsa is here to indicate which subset of ranks the barrier will apply to. The identifier of the underlying barrier to use is provided by index (it should be different for each coop; typically set to blockIdx.x to ensure uniqueness between CTAs). multimem requests a hardware-accelerated implementation using memory multicast.

void arrive(Coop, cuda::memory_order order)

Signals the arrival of the thread at the barrier session.

void wait(Coop, cuda::memory_order order)

Blocks until all threads of all team members arrive at the barrier session.

void sync(Coop, cuda::memory_order order)

Synchronizes all threads of all team members that participate in the barrier session (combines arrive and wait).

ncclGetPeerPointer

void *ncclGetPeerPointer(ncclWindow_t w, size_t offset, int peer)

Returns a load/store accessible pointer to the memory buffer of device peer within the window w. offset is byte-based. peer is a rank index within the world team (see Teams). This function will return NULL if the peer is not within the LSA team.

ncclGetLsaPointer

void *ncclGetLsaPointer(ncclWindow_t w, size_t offset, int lsaPeer)

Returns a load/store accessible pointer to the memory buffer of device lsaPeer within the window w. offset is byte-based. This is similar to ncclGetPeerPointer, but here lsaPeer is a rank index with the LSA team (see Teams).

ncclGetLocalPointer

void *ncclGetLocalPointer(ncclWindow_t w, size_t offset)

Returns a load-store accessible pointer to the memory buffer of the current device within the window w. offset is byte-based. This is just a shortcut version of ncclGetPeerPointer with devComm.rank as peer, or ncclGetLsaPointer with devComm.lsaRank as lsaPeer.

Multimem

ncclGetLsaMultimemPointer

void *ncclGetLsaMultimemPointer(ncclWindow_t w, size_t offset, ncclDevComm const &devComm)

Returns a multicast memory pointer associated with the window w and device communicator devComm. offset is byte-based. Availability of multicast memory is hardware-dependent.

Host-Accessible Device Pointer Functions

The following functions provide host-side access to device pointer functionality, enabling host code to obtain pointers to LSA memory regions.

All functions return ncclResult_t error codes. On success, ncclSuccess is returned. On failure, appropriate error codes are returned (e.g., ncclInvalidArgument for invalid parameters, ncclInternalError for internal failures), unless otherwise specified.

The returned pointers are valid for the lifetime of the window. Pointers should not be used after either the window or communicator is destroyed. Obtained pointers are device pointers.

ncclGetLsaMultimemDevicePointer

ncclResult_t ncclGetLsaMultimemDevicePointer(ncclWindow_t window, size_t offset, void **outPtr)

Returns a multimem base pointer for the LSA team associated with the given window. This function provides host-side access to the multimem memory functionality.

window is the NCCL window object (must not be NULL). offset is the byte offset within the window. outPtr is the output parameter for the multimem pointer (must not be NULL).

This function requires LSA multimem support (multicast capability on the system). The window must be registered with a communicator that supports symmetric memory, and the hardware must support NVLink SHARP multicast functionality.

Note

If the system does not support multimem, the function returns ncclSuccess with *outPtr set to nullptr. This allows applications to gracefully detect and handle the absence of multimem support without breaking the communicator. Users should check if the returned pointer is nullptr to determine availability.

Example:

void* multimemPtr;
ncclResult_t result = ncclGetLsaMultimemDevicePointer(window, 0, &multimemPtr);
if (result == ncclSuccess) {
    if (multimemPtr != nullptr) {
        // Use multimemPtr for multimem operations
    } else {
        // Multimem not supported, use fallback approach
    }
}

ncclGetMultimemDevicePointer

ncclResult_t ncclGetMultimemDevicePointer(ncclWindow_t window, size_t offset, ncclMultimemHandle multimem, void **outPtr)

Returns a multimem base pointer using a provided multimem handle instead of the window’s internal multimem. This function enables using external or custom multimem handles for pointer calculation.

window is the NCCL window object (must not be NULL). offset is the byte offset within the window. multimem is the multimem handle containing the multimem base pointer (multimem.mcBasePtr must not be NULL). outPtr is the output parameter for the multimem pointer (must not be NULL).

This function requires LSA multimem support (multicast capability on the system).

Note

If the system does not support multimem, the function returns ncclSuccess with *outPtr set to nullptr. The function validates that multimem.mcBasePtr is not nullptr before proceeding.

Example:

// Get multimem handle from device communicator setup
ncclMultimemHandle customHandle;
// ... (obtain handle)

void* multimemPtr;
ncclResult_t result = ncclGetMultimemDevicePointer(window, 0, customHandle, &multimemPtr);
if (result == ncclSuccess) {
    if (multimemPtr != nullptr) {
        // Use multimemPtr for multimem operations with custom handle
    } else {
        // Multimem not supported, use fallback approach
    }
}

ncclGetLsaDevicePointer

ncclResult_t ncclGetLsaDevicePointer(ncclWindow_t window, size_t offset, int lsaRank, void **outPtr)

Returns a load/store accessible pointer to the memory buffer of a specific LSA peer within the window. This function provides host-side access to LSA pointer functionality using LSA rank directly.

window is the NCCL window object (must not be NULL). offset is the byte offset within the window (must be >= 0 and < window size). lsaRank is the LSA rank of the target peer (must be >= 0 and < LSA team size). outPtr is the output parameter for the LSA pointer (must not be NULL).

On success, ncclSuccess is returned and the LSA pointer is returned in outPtr.

The window must be registered with a communicator that supports LSA. The LSA rank must be within the valid range for the LSA team, and the target peer must be load/store accessible (P2P connectivity required).

Example:

void* lsaPtr;
ncclResult_t result = ncclGetLsaDevicePointer(window, 0, 1, &lsaPtr);
if (result == ncclSuccess) {
    // Use lsaPtr to access LSA peer 1's memory
}

ncclGetPeerDevicePointer

ncclResult_t ncclGetPeerDevicePointer(ncclWindow_t window, size_t offset, int peer, void **outPtr)

Returns a load/store accessible pointer to the memory buffer of a specific world rank peer within the window. This function converts world rank to LSA rank internally and provides host-side access to peer pointer functionality.

window is the NCCL window object (must not be NULL). offset is the byte offset within the window. peer is the world rank of the target peer (must be >= 0 and < communicator size). outPtr is the output parameter for the peer pointer (must not be NULL).

On success, ncclSuccess is returned and the peer pointer is returned in outPtr.

If the peer is not reachable via LSA (not in LSA team), outPtr is set to NULL and ncclSuccess is returned. This matches the behavior of the device-side ncclGetPeerPointer function.

The window must be registered with a communicator that supports LSA. The peer rank must be within the valid range for the communicator, and the target peer must be load/store accessible (P2P connectivity required).

Example:

void* peerPtr;
ncclResult_t result = ncclGetPeerDevicePointer(window, 0, 2, &peerPtr);
if (result == ncclSuccess) {
    if (peerPtr != NULL) {
        // Use peerPtr to access world rank 2's memory
    } else {
        // Peer 2 is not reachable via LSA
    }
}

GIN

GIN is supported since NCCL 2.28.7.

ncclGin

class ncclGin

A class encompassing major elements of the GIN support.

ncclGin(ncclDevComm const &comm, int contextIndex): Initializes a new ncclGin object. comm is the device communicator created using ncclDevCommCreate(). contextIndex is the index of the GIN context – a network communication channel. Using multiple GIN contexts allows the implementation to spread traffic onto multiple connections, avoiding locking and bottlenecks. Therefore, performance-oriented kernels should cycle among the available contexts to improve resource utilization (the number of available contexts is available via ginContextCount).

void put(ncclTeam team, int peer, ncclWindow_t dstWnd, size_t dstOffset, ncclWindow_t srcWnd, size_t srcOffset, size_t bytes, RemoteAction remoteAction, LocalAction localAction, Coop coop, DescriptorSmem descriptor, cuda::thread_scope alreadyReleased, cuda::thread_scope expected_scope)

Schedules a device-initiated, one-sided data transfer operation from a local buffer to a remote buffer on a peer.

peer is a rank within team (see Teams); it may refer to the local rank (a loopback). The destination and source buffers are each specified using the window (dstWnd, srcWnd) and a byte-based offset (dstOffset, srcOffset). bytes specifies the data transfer count in bytes.

Arguments beyond the first seven are optional. remoteAction and localAction specify actions to undertake on the destination peer and on the local rank when the payload has been settled and the input has been consumed (respectively). They default to ncclGin_None (no action); other options include ncclGin_Signal{Inc|Add} (for remoteAction) and ncclGin_CounterInc (for localAction); see Signals and Counters below for more details. coop indicates the set of threads participating in this operation (see Thread Groups); it defaults to ncclCoopThread (a single device thread), which is the recommended model.

The visibility of the signal on the destination peer implies the visibility of the put data it is attached to and all the preceding puts to the same peer, provided that they were issued using the same GIN context.

The API also defines an alternative, “convenience” variant of this method that uses ncclSymPtr types to specify the buffers and expects size to be conveyed in terms of the number of elements instead of the byte count. There are also two putValue variants that take a single element at a time (no greater than eight bytes), passed by value.

void flush(Coop coop, cuda::memory_order ord = cuda::memory_order_acquire): Ensures that all the pending transfer operations scheduled by any threads of coop are locally consumed, meaning that their source buffers are safe to reuse. Makes no claims regarding the completion status on the remote peer(s).

Signals and Counters

type ncclGinSignal_t

Signals are used to trigger actions on remote peers, most commonly on the completion of a ncclGin::put() operation. They each have a 64-bit integer value associated with them that can be manipulated atomically.

class ncclGin_SignalAdd

ncclGinSignal_t signal

uint64_t value

class ncclGin_SignalInc

ncclGinSignal_t signal

These objects can be passed as the remoteAction arguments of methods such as ncclGin::put() and ncclGin::signal() to describe the actions to perform on the peer on receipt – in this case, increase the value of a signal specified by index. ncclGin_SignalInc{signalIdx} is functionally equivalent to ncclGin_SignalAdd{signalIdx, 1}; however, it may not be mixed with other signal-modifying operations without an intervening signal reset (see below). Signal values use “rolling” comparison logic to ensure that an unsigned overflow maintains the property of x < x + 1.

Signal methods of ncclGin:

void ncclGin::signal(ncclTeam team, int peer, RemoteAction remoteAction, Coop coop, DescriptorSmem descriptor, cuda::thread_scope alreadyReleased, cuda::thread_scope expected_scope)

uint64_t ncclGin::readSignal(ncclGinSignal_t signal, int bits = 64, cuda::memory_order ord = cuda::memory_order_acquire)

void ncclGin::waitSignal(Coop coop, ncclGinSignal_t signal, uint64_t least, int bits = 64, cuda::memory_order ord = cuda::memory_order_acquire)

void ncclGin::resetSignal(ncclGinSignal_t signal)

These are signal-specific methods of ncclGin. ncclGin::signal() implements an explicit signal notification without an accompanying data transfer operation; it takes a subset of arguments of ncclGin::put(). ncclGin::readSignal() returns the bottom bits of the value of the signal. ncclGin::waitSignal() waits for the bottom bits of the signal value to meet or exceed least. Finally, ncclGin::resetSignal() resets the signal value to 0 (this method may not race with concurrent modifications to the signal).

type ncclGinCounter_t

Counters are used to trigger actions on the local rank; as such, they are complementary to signals, which are meant for remote actions. Like signals, they use “rolling” comparison logic, but they are limited to storing values of at most 56 bits.

class ncclGin_CounterInc

ncclGinCounter_t counter

This object can be passed as the localAction argument of methods such as ncclGin::put(). It is the only action defined for counters.

Counter methods of ncclGin:

uint64_t ncclGin::readCounter(ncclGinCounter_t counter, int bits = 56, cuda::memory_order ord = cuda::memory_order_acquire)

void ncclGin::waitCounter(Coop coop, ncclGinCounter_t counter, uint64_t least, int bits = 56, cuda::memory_order ord = cuda::memory_order_acquire)

void ncclGin::resetCounter(ncclGinCounter_t counter)

These are counter-specific methods of ncclGin and they are functionally equivalent to their signal counterparts discussed above.

ncclGinBarrierSession

template<typename Coop> class ncclGinBarrierSession

A class representing a network barrier session.

ncclGinBarrierSession(Coop coop, ncclGin gin, ncclTeamTagRail tag, uint32_t index): Initializes a new network barrier session. coop represents a cooperative group (see Thread Groups). gin is a previously initialized ncclGin object. ncclTeamTagRail indicates that the barrier will apply to all peers on the same rail as the local rank (see Teams). index identifies the underlying barrier to use (it should be different for each coop; typically set to blockIdx.x to ensure uniqueness between CTAs).

ncclGinBarrierSession(Coop coop, ncclGin gin, ncclTeam team, ncclGinBarrierHandle handle, uint32_t index): Initializes a new network barrier session. This is the general-purpose variant to be used, e.g., when communicating with ranks from the world team (see Teams), whereas the previous variant was specific to the rail team. This variant expects team to be passed as an argument, and also takes an extra handle argument indicating the location of the underlying barriers (typically set to the railGinBarrier field of the device communicator).

void sync(Coop coop, cuda::memory_order order, ncclGinFenceLevel fence): Synchronizes all threads of all team members that participate in the barrier session. ncclGinFenceLevel::Relaxed is the only defined value for fence for now.