LLMA4OpenMP: A Large Language Model Based Agent for Automatic OpenMP Parallelization

In the field of parallel computing, efficiently leveraging multi-core processors is essential. OpenMP, despite its popularity, poses challenges due to data-sharing and synchronization issues, making manual parallelization both time-consuming and prone to errors. Traditional compiler-based automation tools could be constrained to the final development stages. Recent advancements in large language models offer new opportunities for automation, with models like HPC-Coder demonstrating potential through fine-tuning. However, fine-tuning alone does not fully address the complexities of code comprehension. We introduce a novel agent that dynamically orchestrates static analysis tools to analyze code dependency and leverages in-context learning to generate OpenMP directives. Then, we implement this approach into a user-friendly plugin, namely LLMA4OpenMP, for Neovim and VS Code, enabling one-click OpenMP parallelization. We test our approach on the sophisticated AutoParBench and obtain an accuracy of$\mathbf{9 3. 1 \%}$, outperforming the state-of-the-art LLM-based method. The source code for testing is available on the following GitHub repository: https://github.com/LLMA4OpenMP/LLMA4OpenMP_VScode