Finalize

Applicability

Product	Supported
Atlas A3 training products/Atlas A3 inference products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference products	x
Atlas inference product's AI Core	x
Atlas inference product's Vector Core	x
Atlas training products	x

Function

Instructs the server to exit after the current communication task is complete.

Prototype

template <bool sync = true>
__aicore__ inline void Finalize()

Parameters

Parameter	Input/Output	Description
sync	Input	Whether to wait for the communication on the server to complete. The value is of the bool type. The options are as follows: true: default value. The client detects and waits until the last communication task is complete. false: the client does not wait for the communication task to complete. It exits directly. Atlas A3 training products/Atlas A3 inference products. This parameter supports the default value true. Only when the communication task is BatchWrite, the value false is supported. Atlas A2 training products/Atlas A2 inference products. This parameter supports only the default value true.

Parameter

Input/Output

Description

sync

Input

Whether to wait for the communication on the server to complete. The value is of the bool type. The options are as follows:

true: default value. The client detects and waits until the last communication task is complete.
false: the client does not wait for the communication task to complete. It exits directly.

Atlas A3 training products/Atlas A3 inference products. This parameter supports the default value true. Only when the communication task is BatchWrite, the value false is supported.

Atlas A2 training products/Atlas A2 inference products. This parameter supports only the default value true.

Returns

None

Restrictions

Before calling this API, ensure that the InitV2 and SetCcTilingV2 APIs have been called.
When this API is called on the AIC or AIV core, the calling core must be the same as that of the corresponding Prepare API.

Example

REGISTER_TILING_DEFAULT(ReduceScatterCustomTilingData); // ReduceScatterCustomTilingData is a structure defined in the operator header file.
GET_TILING_DATA_WITH_STRUCT(ReduceScatterCustomTilingData, tilingData, tilingGM);
Hccl hccl;
GM_ADDR contextGM = AscendC::GetHcclContext<0>();  // In the kernel of the AscendC custom operator, obtain the HCCL context through this method.
hccl.InitV2(contextGM, &tilingData);
hccl.SetCcTilingV2(offsetof(ReduceScatterCustomTilingData, mc2CcTiling));
if (AscendC::g_coreType == AIC) {
    HcclHandle handleId = hccl.ReduceScatter(sendBuf, recvBuf, 100, HcclDataType::HCCL_DATA_TYPE_INT8, HcclReduceOp::HCCL_REDUCE_SUM, 10);
hccl.Commit(handleId ); //Notify the server that the ReduceScatter task can be executed.
    for (uint8_t i=0; i<10; i++) {
        hccl.Wait(handleId); // Blocking API. Wait until the ReduceScatter task is complete.
    }
    hccl.Finalize<true>(); // If there are no other communication tasks to follow, notify the server to exit after executing the above ReduceScatter task.
}

Parent topic: HCCL Kernel APIs