Overview

Introduction

An AI CPU operator is an operation of complete compute logic that runs on AI CPU, one of the compute engines in Ascend AI Processor. During neural network (NN) training or inference, if you find an unsupported operator when converting a third-party open-source network to adapt to Ascend AI Processor, a custom AI CPU operator can help you streamline the model execution process and improve the functionality debugging efficiency.

AI CPU provides a range of APIs for you to develop AI CPU operators. For details about the header files, see API Overview.

API Overview

**Table 1** API Overview
API	Description	Header File	Library File
Class CpuKernelContext	Describes the context definition and methods of CPU kernels, which can be used to obtain information such as inputs, outputs, and attributes.	include/aicpu/cpu_context.h under the CANN component directory	opp/built-in/op_impl/aicpu/aicpu_kernel/lib/libcpu_kernels_context.a under the CANN component directory
Class TensorShape	Describes the tensor shape definitions and methods of CPU kernels.	include/aicpu/cpu_tensor_shape.h under the CANN component directory
Class Tensor	Describes the tensor definitions and methods of the CPU kernels.	include/aicpu/cpu_tensor.h under the CANN component directory
Class AttrValue	Describes the attribute definition and method of CPU kernels.	include/aicpu/cpu_attr_value.h under the CANN component directory
Data types	Describes the data types and formats of CPU kernels.	include/aicpu/cpu_types.h under the CANN component directory
Macro definition	Describes the kernel base class definition and registration macros of CPU kernels.	include/aicpu/cpu_kernel.h under the CANN component directory
Dump log APIs	Describes the dump log API declaration.	include/aicpu/cust_cpu_utils.h under the CANN component directory
Class NodeDefBuilder	Describes NodeDef and methods of CPU kernels.	include/aicpu/cpu_nodedef_builder.h under the CANN component directory

Parent topic: AI CPU API