Article Details

Abstract

In recent years, large language models have achieved remarkable progress in general-purpose code translation. However, when directly applied to embedded signal processing MATLAB-VSIPL code translation, these models suffer from poor accuracy and non-compliance with domain-specific norms. To address this issue, this paper proposes a domain-specific fine-tuning approach for intelligent MATLAB-VSIPL code translation in embedded signal processing. By mining and preprocessing historical project code repositories, high-quality MATLAB-VSIPL parallel training samples are constructed. Supervised fine-tuning is then employed to inject VSIPL programming specifications and transformation rules from the signal processing domain into a general-purpose large language model, enabling it to learn the mapping logic from high-level MATLAB to low-level VSIPL. Additionally, parameter-efficient fine-tuning and early stopping strategies are adopted to effectively mitigate overfitting under limited domain sample sizes. Experimental results demonstrate that the proposed method significantly improves both semantic similarity and functional correctness in MATLAB-VSIPL code translation.

Keywords

LLM; MATLAB-VSIPL; code translation; LoRA; parameter-efficient fine-tuning

References

1. [1] Liu Hao, Li Chao Design of Signal Processing Platform Based on Domestic High Performance CPU [J]. Modern Navigation, 2025, 16 (5): 380-385
2. [2] Deng Zhusha Software Design of Cross Platform Algorithm Middleware Based on VSIPL [J]. Telecommunications Technology, 2013, 53 (6): 773-776
3. [3] Qiu Jin Design and Implementation of Software based Radar Algorithm Components [D]. Xi'an: Xi'an University of Electronic Science and Technology, 2019
4. [4] Hui B Y，Yang J，Cui Z Y，et al. Qwen2.5-Coder Technical Report [J/OL].arXiv:2409.12186 [cs.CL].https://doi.org/10.48550/arXiv.2409.12186.
5. [5] Yang Z，Liu F，Yu Z X，et al. Exploring and Unleashing the Power of Large Language Models in Automated Code Translation [J].Proceedings of the ACM on Software Engineering，2024，1（FSE）：1585-1608.
6. [6] Vaswani A，Shazeer N，Parmar N，et al. Attention Is All You Need [J/OL].arXiv:1706.03762 [cs.CL].https://doi.org/10.48550/arXiv.1706.03762.
7. [7] Zhang Qintong, Wang Yuchao, Wang Hexi, etc A review of research on fine-tuning techniques for large language models [J]. Computer Engineering and Applications, 2024, 60 (17): 17-33
8. [8] Hu E J，Shen Y L，Wallis P，et al. LoRA: Low-Rank Adaptation of Large Language Models [J/OL].arXiv:2106.09685 [cs.CL].https://doi.org/10.48550/arXiv.2106.09685.
9. [9] Zheng Y W，Zhang R C，Zhang J H，et al. LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models [C]//Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)，2024：400-410.
10. [10] Ren S，Guo D Y，Lu S，et al. CodeBLEU: a Method for Automatic Evaluation of Code Synthesis [J/OL].arXiv:2009.10297 [cs.SE].https://doi.org/10.48550/arXiv.2009.10297.

About the authors

Li Weichao (1988-), male, Han ethnicity, born in Zhengzhou, Henan Province, is an engineer with a doctoral degree. His research direction is intelligent development of embedded software.