HcclReduce

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	☓
Atlas inference product	☓
Atlas training product	√

Function

Performs the sum operation (or other reduction operations) on the data of all ranks and sends the result to the specified position on the root rank.

Prototype

HcclResult HcclReduce(void *sendBuf, void *recvBuf, uint64_t count, HcclDataType dataType, HcclReduceOp op, uint32_t root, HcclComm comm, aclrtStream stream)

Parameters

Parameter	Input/Output	Description
sendBuf	Input	Address of the send buffer.
recvBuf	Output	Address of the buffer to receive collective communication result.
count	Input	Number of data records to perform Reduce operation. For example, if only one int32 data record is involved, then count=1.
dataType	Input	Data type of the Reduce operation, which is of the HcclDataType type. Atlas 350 Accelerator Card: The supported data types are int8, int16, int32, int64, uint64, float16, float32, float64, and bfp16. Data types int64, uint64, and float64 supports only intra-node communication. Atlas A3 training product/Atlas A3 inference product: The supported data types are int8, int16, int32, int64, float16, float32, and bfp16. Atlas A2 training product/Atlas A2 inference product: The supported data types are int8, int16, int32, int64, float16, float32, and bfp16. Note that the performance will degrade for the int64 data type. Atlas training product: The supported data types are int8, int32, int64, float16, and float32.
op	Input	Reduction operation type. Currently, the following operation types are supported: sum, prod, max, and min. NOTE: Atlas 350 Accelerator Card: The supported operation types are sum, max, and min. Atlas A3 Training Series Product: In the current version, the prod operation does not support the int16 and bfp16 data types. Atlas A2 training product/Atlas A2 inference product: In the current version, the prod operation does not support the int16 and bfp16 data types.
root	Input	ID of the root rank used for the reduction operation.
comm	Input	Communicator where the operation is performed.
stream	Input	Stream of the rank.

Returns

HcclResult: HCCL_SUCCESS on success, or else failure.

Constraints

All ranks must have the same count, dataType, and op.
The input and output addresses (sendBuf and recvBuf) of the operator must meet the following alignment requirements based on different data types:
- int8: 1-byte aligned
- int16, float16, bfp16: 2-byte aligned
- int32 and float32: 4-byte aligned
- int64, uint64, float64: 8-byte aligned

Call Example

// Allocate device memory for collective communication.
void *sendBuf = nullptr;
void *recvBuf = nullptr;
uint64_t count = 8;
size_t mallocSize = count * sizeof(float);
aclrtMalloc((void **)&sendBuf, mallocSize, ACL_MEM_MALLOC_HUGE_ONLY);
aclrtMalloc((void **)&recvBuf, mallocSize, ACL_MEM_MALLOC_HUGE_ONLY);

// Initialize the communicator.
uint32_t rankSize = 8;
HcclComm hcclComm;
HcclCommInitRootInfo(rankSize, &rootInfo, deviceId, &hcclComm);

// Create a task flow.
aclrtStream stream;
aclrtCreateStream(&stream);

// Execute Reduce to sum up sendBuf of all ranks at the corresponding location and then send the result to recvBuf of the root rank.
HcclReduce(sendBuf, recvBuf, count, HCCL_DATA_TYPE_FP32, HCCL_REDUCE_SUM, rootRank, hcclComm, stream);
// Wait until the collective communication task in the task flow is complete.
aclrtSynchronizeStream(stream);

// Free resources.
aclrtFree(sendBuf);          // Free the device memory.
aclrtFree(recvBuf);          // Free the device memory.
aclrtDestroyStream(stream);  // Destroy the task flow.
HcclCommDestroy(hcclComm);   // Destroy the communicator.

Parent topic: Collective Communication