A2H-MAS: An Algorithm-to-HLS Multi-Agent System for Automated and Reliable FPGA Implementation

TL;DR

The paper addresses the challenge of translating MATLAB-based algorithm prototypes into FPGA-ready High-Level Synthesis (HLS) implementations under strict latency and resource constraints. It introduces A2H-MAS, a modular, hierarchical multi-agent system that uses standardized interfaces, execution-based validation, and dataflow-oriented algorithm decomposition, coupled with algorithm–hardware co-design to surpass single-agent LLM limitations. The approach is validated on representative wireless workloads (5G NR SSB detection and WLAN synchronization), showing functionally correct, resource-efficient, and latency-optimized hardware designs, with substantial gains demonstrated by ablation experiments (Adaption and Refinement stages). This work offers a robust, scalable workflow for end-to-end MATLAB-to-HLS-to-FPGA deployment, with potential applicability to broader domains such as vision and signal processing and opportunities for richer feedback and benchmarks.

Abstract

Bridging the gap between algorithm development and hardware realization remains a persistent challenge, particularly in latency- and resource-constrained domains such as wireless communication. While MATLAB provides a mature environment for algorithm prototyping, translating these models into efficient FPGA implementations via High-Level Synthesis (HLS) often requires expert tuning and lengthy iterations. Recent advances in large language models (LLMs) offer new opportunities for automating this process. However, existing approaches suffer from hallucinations, forgetting, limited domain expertise, and often overlook key performance metrics. To address these limitations, we present A2H-MAS, a modular and hierarchical multi-agent system. At the system level, A2H-MAS assigns clearly defined responsibilities to specialized agents and uses standardized interfaces and execution-based validation to ensure correctness and reproducibility. At the algorithmic level, it employs dataflow-oriented modular decomposition and algorithm-hardware co-design, recognizing that the choice of algorithm often has a larger impact on hardware efficiency than pragma-level optimization. Experiments on representative wireless communication algorithms show that A2H-MAS consistently produces functionally correct, resource-efficient, and latency-optimized HLS designs, demonstrating its effectiveness and robustness for complex hardware development workflows.

Figures (6)

• Figure 1: Compared with a single LLM agent that suffers from hallucination, forgetting, and unstable behavior, A2H-MAS decomposes the MATLAB-to-HLS-to-hardware flow into specialized agents with standardized interfaces. Agent outputs are guided by explicit rules, constrained and verified using deterministic tools, and refined through feedback, resulting in reliable and high-quality hardware implementations.
• Figure 2: Schematic of standardized input–output interfaces for agents, enabling seamless pipeline integration with minimal inter-agent coupling.
• Figure 3: Example of a rule-guided and tool-driven agent. Each agent follows a predefined workflow pattern and leverages deterministic tools for execution and validation, ensuring reliability and stability of the outputs.
• Figure 4: Illustration of the submodule configuration file, showing key fields and interface definitions.
• Figure 5: Overall workflow of the A2H-MAS framework. The process is organized into seven modular phases, including modularization, test data generation, function flattening, code optimization, code translation, refinement, and final integration & implementation. Each phase is handled by a dedicated Agent responsible for a specific function, with validation mechanisms ensuring correctness, ultimately achieving a stable and efficient FPGA implementation.