AgRefactor Automates HLS Code Refactoring with Self-Evolving Agents.

This paper introduces AgRefactor, an innovative multi-agent system powered by Large Language Models (LLMs) designed to automate the challenging process of refactoring software into High-Level Synthesis (HLS)-compatible code. HLS is crucial for accelerating the path from software concepts to silicon, but converting real-world software often faces hurdles due to restrictive language support and the inherent differences between software and hardware programming paradigms. Existing automated and LLM-based refactoring methods have limitations in flexibility, scalability, and computational cost. AgRefactor addresses these issues by incorporating a self-evolving memory system that continuously accumulates and retrieves both factual and strategic knowledge across various tasks, significantly enhancing its robustness and efficiency when encountering new programs. To further optimize cost and scalability, the system intelligently integrates automated refactoring tools, allowing agents to judiciously balance LLM-driven code rewrites with more efficient tool-based transformations. Evaluated on 9 out of 11 complex real-world benchmarks, which are substantially larger than those typically studied, AgRefactor demonstrated performance that either matched or surpassed state-of-the-art automated refactoring tools and strong LLM-based baselines. Furthermore, its agentic performance optimization led to a 6.51x geometric mean speedup over leading pragma tuning tools and a 1.20x speedup over optimized open-source designs, all while using less than 20% additional resources. The system is fully automated and has been open-sourced, offering a significant advancement for hardware design and optimization.