* Deallocate buffer * * @param buf Buffer to deallocate * @param buffer_type NCCL_PTR_HOST for host memory, NCCL_PTR_CUDA for device memory * @ret

* Allocate buffer (host or CUDA device memory) * * @param buf Output pointer to allocated buffer * @param size Size of buffer in bytes * @param bu

* Initialize buffer with a pattern * * @param buf Buffer to initialize * @param size Size of buffer in bytes * @param buffer_type NCCL_PTR_HOST fo

* Given 2 points a and b, find the line connecting them. * Then find the farthest point on that line with either the same x or same y coordinate * o

check that we get the expected response for all our known platforms */

* Post receive operation * * Posts an asynchronous receive operation using the plugin's irecv function. * Supports grouped receives (multiple buffe

* Validate received data against expected data * * @param recv_buf Received buffer to validate * @param expected_buf Expected buffer to compare aga

Allocate and copy regions */

Allocate a new freelist item * * Return pointer to memory of size entry_size (provided to init) from * the given freelist. If required, the fre

* Post send operation * * Posts an asynchronous send operation using the plugin's isend function. * The operation completes asynchronously and must

* @brief Return rx data struct of rx request */

* Helper: poll a single request to completion, return size. */

Release a freelist item * * Return a freelist item to the freelist. After calling this * function, the user should not read from or write to me

* @brief Ray-casting algorithm to check if a pointer is inside, outside or * on the edge of a region. * For a detailed explanation, check: *

* Helper: poll a grouped recv request, get per-sub sizes. */

* @brief Decide whether an eager entry is the next in sequence and can be * processed by the receiver eager drain. * * Eager batches are chained b

* @brief Returns the base domain's device back-pointer. */

Remove and return the first node, or nullptr if empty. */

* @brief Return recv data struct of recv request */

* Translate an errno (returned by Libfabric and internal plugin functions) to a * ncclResult_t (returned to NCCL per net API) * * This is used to t

* @brief Return send data struct of send request */

* Create memory registration iovec cache key of a memory region * * Default expectation on this cache key is that memory region covers * full pages

Returns a raw pointer for downcasting to the transport-specific * domain type. Safe because the ep holds shared_ptr<domain>, * so the domain is al

* @brief Update NCCL's system topology using static pre-configured topology * files for supported EC2 platform types. * * @return 0, when we are

* @brief Return endpoint rail with index `rail_id` */

* @brief Assign an allocated rdma request buffer */

* @brief Free request by returning request back into freelist */

* @brief Determines GDR (GPUDirect RDMA) support for all available network devices * * This function queries all available network devices and check

* @brief Non-blocking connect which returns sComm as nullptr * with an expectation that it will be called again until * sComm != nullptr * * Th

* @brief Return user data struct * * @return user data, if available * NULL, if iterator has reached end of vector */

* @brief Increment user data iterator */

* @brief Return control endpoint rail with index `rail_id` */

* Close recv communicator. This function will add the given communicator * to the deferred close list. When pending close actions (send_close message

Return the first node, or nullptr if empty. */

* @brief Return send communicator rail with index `rail_id` */

* Progress a request associated with recv * * Post request associated with a receive. If `add_to_pending` is true * and request could not be posted

Replicate the tag-matching search from get_ctrl_msg_buff_len / update_send_data_from_remote */

* @brief Return flush data struct of flush request */

* @brief Process completion entries for the given completion quque. * This also updates several request fileds like size, status, etc * * @return

* Add a request which returned -FI_EAGAIN, to be retried later */