ChatHLS: Towards Systematic Design Automation and Optimization for High-Level Synthesis

2025-07-01

AI Summary

To address the low design efficiency in high-level synthesis (HLS) caused by stringent coding constraints and complex hardware optimization, this paper proposes the first multi-agent collaborative framework tailored for HLS. The framework integrates a domain-specific fine-tuned large language model (LLM), an HLS-semantic-feedback-driven automatic error correction mechanism, and resource-aware performance optimization strategies, enabling end-to-end HLS design automation and deep hardware optimization. Evaluated on 612 test cases, it achieves an average repair pass rate of 82.7%, outperforming GPT-4o and Llama3-8B by 19.1% and 63.0%, respectively; for resource-constrained kernels, it delivers up to 14.8× speedup, with a geometric mean 4.9× higher than state-of-the-art DSL-based approaches. The core contribution lies in establishing a scalable, formally verifiable LLM-augmented HLS automation paradigm.

Technology Category

• Machine Learning: Hardware-aware ML
• Multiagent Systems: Multiagent Learning
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search

Abstract

The increasing complexity of computational demands has accelerated the adoption of domain-specific accelerators, yet traditional hardware design methodologies remain constrained by prolonged development and verification cycles. High-Level Synthesis (HLS) bridges the gap between software and hardware by enabling hardware design from high-level programming languages. However, its widespread adoption is hindered by strict coding constraints and intricate hardware-specific optimizations, creating significant obstacles for developers. Recent advancements in Large Language Models (LLMs) demonstrate substantial potential in hardware design automation. However, their effectiveness is limited by the scarcity of high-quality datasets, particularly in the context of HLS. To address these challenges, we introduce ChatHLS, an agile HLS design automation and optimization workflow that leverages fine-tuned LLMs integrated within a multi-agent framework for error correction and design optimization. Our extensive evaluations reveal that ChatHLS achieves an average repair pass rate of 82.7% over 612 test cases, outperforming the GPT-4o and Llama3-8B by 19.1% and 63.0%, respectively. Furthermore, ChatHLS delivers performance enhancements ranging from 1.9$ imes$ to 14.8$ imes$ upon resource-constrained kernels. By enabling sophisticated optimization reasoning within practical computational budgets, ChatHLS attains a 4.9$ imes$ geometric mean speedup compared to state-of-the-art DSL-based approaches. These results underscore the potential of ChatHLS in substantially expediting hardware development cycles while maintaining rigorous standards of design reliability and optimization quality.

Problem

• Overcoming strict coding constraints in High-Level Synthesis
• Addressing scarcity of high-quality datasets for HLS automation
• Reducing prolonged hardware development and verification cycles

Innovation

• Fine-tuned LLMs for HLS automation
• Multi-agent framework for error correction
• Resource-constrained kernel performance enhancement