Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MPM-LLM4DSE: Reaching the Pareto Frontier in HLS with Multimodal Learning and LLM-Driven Exploration

Lei Xu, Shanshan Wang, Chenglong Xiao

TL;DR

The paper tackles the challenge of rapid QoR prediction and Pareto-front design space exploration in high-level synthesis (HLS) with vast pragma spaces. It introduces MPM-LLM4DSE, a framework that fuses graph-based CDFG features with semantic text embeddings from source code using an Enhanced-CoGNN and CodeBERT-based LM, and leverages an LLM-driven optimizer guided by the PEODSE prompting strategy. Empirical results show substantial improvements: QoR-prediction RMSE improvements up to 10.25× over ProgSG and average DSE gains of ~39.9% over prior methods, with further ADRS gains in prompting strategies for large design spaces. This work demonstrates the potential of multimodal representations and LLM-driven exploration to accelerate discovery of Pareto-optimal HLS configurations, with code and models available on GitHub.

Abstract

High-Level Synthesis (HLS) design space exploration (DSE) seeks Pareto-optimal designs within expansive pragma configuration spaces. To accelerate HLS DSE, graph neural networks (GNNs) are commonly employed as surrogates for HLS tools to predict quality of results (QoR) metrics, while multi-objective optimization algorithms expedite the exploration. However, GNN-based prediction methods may not fully capture the rich semantic features inherent in behavioral descriptions, and conventional multi-objective optimization algorithms often do not explicitly account for the domain-specific knowledge regarding how pragma directives influence QoR. To address these limitations, this paper proposes the MPM-LLM4DSE framework, which incorporates a multimodal prediction model (MPM) that simultaneously fuses features from behavioral descriptions and control and data flow graphs. Furthermore, the framework employs a large language model (LLM) as an optimizer, accompanied by a tailored prompt engineering methodology. This methodology incorporates pragma impact analysis on QoR to guide the LLM in generating high-quality configurations (LLM4DSE). Experimental results demonstrate that our multimodal predictive model significantly outperforms state-of-the-art work ProgSG by up to 10.25$\times$. Furthermore, in DSE tasks, the proposed LLM4DSE achieves an average performance gain of 39.90\% over prior methods, validating the effectiveness of our prompting methodology. Code and models are available at https://github.com/wslcccc/MPM-LLM4DSE.

MPM-LLM4DSE: Reaching the Pareto Frontier in HLS with Multimodal Learning and LLM-Driven Exploration

TL;DR

The paper tackles the challenge of rapid QoR prediction and Pareto-front design space exploration in high-level synthesis (HLS) with vast pragma spaces. It introduces MPM-LLM4DSE, a framework that fuses graph-based CDFG features with semantic text embeddings from source code using an Enhanced-CoGNN and CodeBERT-based LM, and leverages an LLM-driven optimizer guided by the PEODSE prompting strategy. Empirical results show substantial improvements: QoR-prediction RMSE improvements up to 10.25× over ProgSG and average DSE gains of ~39.9% over prior methods, with further ADRS gains in prompting strategies for large design spaces. This work demonstrates the potential of multimodal representations and LLM-driven exploration to accelerate discovery of Pareto-optimal HLS configurations, with code and models available on GitHub.

Abstract

High-Level Synthesis (HLS) design space exploration (DSE) seeks Pareto-optimal designs within expansive pragma configuration spaces. To accelerate HLS DSE, graph neural networks (GNNs) are commonly employed as surrogates for HLS tools to predict quality of results (QoR) metrics, while multi-objective optimization algorithms expedite the exploration. However, GNN-based prediction methods may not fully capture the rich semantic features inherent in behavioral descriptions, and conventional multi-objective optimization algorithms often do not explicitly account for the domain-specific knowledge regarding how pragma directives influence QoR. To address these limitations, this paper proposes the MPM-LLM4DSE framework, which incorporates a multimodal prediction model (MPM) that simultaneously fuses features from behavioral descriptions and control and data flow graphs. Furthermore, the framework employs a large language model (LLM) as an optimizer, accompanied by a tailored prompt engineering methodology. This methodology incorporates pragma impact analysis on QoR to guide the LLM in generating high-quality configurations (LLM4DSE). Experimental results demonstrate that our multimodal predictive model significantly outperforms state-of-the-art work ProgSG by up to 10.25. Furthermore, in DSE tasks, the proposed LLM4DSE achieves an average performance gain of 39.90\% over prior methods, validating the effectiveness of our prompting methodology. Code and models are available at https://github.com/wslcccc/MPM-LLM4DSE.
Paper Structure (13 sections, 6 equations, 7 figures, 6 tables)

This paper contains 13 sections, 6 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. PRELIMINARIES
  3. Multimodal Graph Learning In HLS QoR Prediction
  4. Multimodal Dataset In HLS QoR Prediction
  5. Methodology
  6. Data Generator
  7. Multimodal Predictive Model
  8. Design Space Exploration
  9. Experiments
  10. Experimental Setup
  11. Evaluation of HLS QoR Prediction Accuracy
  12. Evaluation of Design Space Exploration
  13. Conclusion

Figures (7)

  • Figure 1: An overview of HLS DSE.
  • Figure 2: The framework of MPM-LLM4DSE.
  • Figure 3: The workflow of Dataset Generator.
  • Figure 4: The architecture of the proposed predictive model.
  • Figure 5: Multimodal feature fusion for QoR prediction.
  • ...and 2 more figures