Custom Operators

When built-in operators cannot meet requirements during network development, you can use MindSpore's custom operator functionality to integrate your operators. Currently, MindSpore provides two approaches for integrating custom operators:

• Custom Primitive-Based Custom Operators

• CustomOpBuilder-Based Custom Operators

Interface Comparison	Custom Primitive	CustomOpBuilder
Supported Modes	Graph Mode and PyNative Mode	PyNative Mode
Interface Functions	Provides a unified Custom Primitive that calls user interfaces at various stages of operator execution.	Compiles and loads custom operator modules online, which can be directly applied to networks.
Advantages	Supports both Graph and PyNative mode , with operator scheduling and execution processes consistent with built-in operators, ensuring high performance.	Enables operator development based on C++ tensors, offering a more intuitive custom execution process.
Disadvantages	Has more interface restrictions, and the operator execution process is not visible to users.	Involves multiple interfaces for operator development; currently lacks a concise and efficient C++ API, making the development of high-performance operators challenging.
Feature Level	STABLE	BETA