Energy-Efficient Auto-Vectorization via Relational Graph Convolutional Networks

With the growing demand for energy efficiency in high-performance computing, traditional auto-vectorization heuristics face limitations due to their reliance on predefined rules and neglect of energy consumption, particularly for complex program structures and cross-platform adaptation. This paper proposes an energy-efficient auto-vectorization method using Relational Graph Convolutional Networks (RGCN) to address the limitations of heuristic compiler approaches in handling complex program structures and energy consumption optimization. By constructing code graph representations integrating control flow, data flow, and call flow, and training an RGCN model to predict Energy-Delay Product (EDP)-optimal vector widths, our method achieves significant improvements. Experimental results on two Intel Xeon platforms demonstrate EDP optimizations of 53.9% and 50.5% respectively compared to default configurations, outperforming rule-based models and validating its efficacy in handling complex code and cross-platform energy efficiency optimization.