Operator generation methods, apparatus, electronic devices and media
By acquiring structured operator description information and using automated methods for generating large language models, the problems of high development cost and low performance of BLAS operators are solved. High-performance BLAS operator generation across multiple hardware architectures is achieved, reducing the technical threshold and improving the reliability of the generated code.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF SOFTWARE - CHINESE ACAD OF SCI
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-30
AI Technical Summary
In existing technologies, BLAS operators have high development costs, high technical barriers, rely on experts and cannot be generated efficiently and automatically. Furthermore, the assembly code generated by LLM has low performance and is difficult to optimize across various hardware architectures.
By acquiring structured operator description information, including algorithm-level and register-level descriptions, it is decomposed into multiple functional blocks, generating kernel-level code, which is then combined with a runtime framework. Automated code generation is achieved using a large language model, and optimization is performed by combining compilation testing and runtime testing, ultimately generating a high-performance BLAS operator.
It achieves high-performance BLAS operator generation across multiple hardware architectures, significantly reducing development costs and time. Its performance is superior to direct generation by LLM and manual optimization by experts, and it has excellent cross-architecture capabilities and code generation reliability.
Smart Images

Figure CN122308835A_ABST