Function: reduce_async

Applicability

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

Description

Performs a Reduce operation, including SUM, MIN, and MAX.

This API is asynchronous. The API call delivers a task rather than executes a task. After this API is called, you must call a synchronization API (for example, synchronize_stream) to ensure that the memory initialization task is complete. Otherwise, unknown problems may occur, such as service exceptions (for example, training or inference exceptions) or device disconnection.

Prototype

C Prototype

aclError aclrtReduceAsync(void *dst, const void *src, uint64_t count, aclrtReduceKind kind, aclDataType type, aclrtStream stream, void *reserve)

Python Function

ret = acl.rt.reduce_async(dst, src, count, kind, type, stream, reserve)

Parameters

Parameter	Description
dst	Int, pointer to the destination memory address.
src	int: pointer to the source memory address.
count	int, size of the source memory, in bytes.
kind	int: operation type. For details, see aclrtReduceKind.
type	int: data type. Atlas A3 training product/Atlas A3 inference productThe following types are supported: int8, int16, int32, fp16, fp32 and bf16. Atlas A2 training product/Atlas A2 inference productThe following types are supported: int8, int16, int32, fp16, fp32 and bf16. The Atlas 200I/500 A2 inference product supports the following types: int8, int16, int32, fp16 and fp32. Atlas inference productThe following data types are supported: fp32, fp16, and int16. The Atlas training product supports only the fp32 type. For details, see aclDataType.
stream	Int, stream. If the default stream is used, set the parameter to 0.
reserve	Int, reserved parameter. Currently, the value is fixed to NULL.

Return Value

Return Value	Description
ret	Int, error code. 0 on success; else, failure.

Restrictions

dts and src must be on the same device as stream.

Parent topic: Execution Control