LaunchWithZeroEleOutputTensors

Function

When the operator output is an all-empty tensor, you can configure the operator to still be executed on the NPU.

Prototype

OpAICoreDef &OpAICoreDef::LaunchWithZeroEleOutputTensors(bool launchFlag)

Parameters

Parameter	Input/Output	Description
launchFlag	Input	For a custom operator developed by the user, if the operator needs to be executed on the NPU when all outputs are empty tensors, set this parameter to true. Otherwise, the operator will not be executed.

Returns

OpAICoreDef operator definition. For details, see OpAICoreDef.

Restrictions

None

Example

class AddCustom : public OpDef {
public:
    AddCustom(const char* name) : OpDef(name)
    {
        this->Input("x")
            .ParamType(REQUIRED);
        this->Input("y")
            .ParamType(REQUIRED);
        this->Output("z")
            .ParamType(REQUIRED);
        this->SetInferShape(ge::InferShape);

        this->AICore()
            .SetTiling(optiling::TilingFunc)
            .SetTilingParse(optiling::TilingPrepare)
            .SetOpSelectFormat(optiling::OpSelectFormat)
            .SetCheckSupport(optiling::CheckSupported)
            .LaunchWithZeroEleOutputTensors(true);

        OpAICoreConfig aicConfig;
        aicConfig.DynamicCompileStaticFlag(true)
            .DynamicFormatFlag(true)
            .DynamicRankSupportFlag(true)
            .DynamicShapeSupportFlag(true)
            .NeedCheckSupportFlag(true)
            .PrecisionReduceFlag(true);
        // Note: Replace soc_version with the actual AI Processor version.
        this->AICore().AddConfig("soc_version", aicConfig);
    }
};

Parent topic: OpAICoreDef