HGBO-DSE: Hierarchical GNN and Bayesian Optimization based HLS Design Space Exploration
Huizhen Kuang, Xianfeng Cao, Jingyuan Li, Lingli Wang
Abstract
High-Level Synthesis (HLS) design space exploration aims to find Pareto-optimal designs in the vast directive configuration space. This paper proposes an automatic framework, HGBO-DSE, which consists of a Hierarchical Graph neural network Predictor (HGP) to estimate post-implementation PPA accurately, a Tree-structured Design space Modeler (TDM) to remove the invalid configurations, and a Bayesian Optimization based Multi-objective Exploration engine (BOME) to search Pareto solutions efficiently at function/loop/array/operator-level. A standard dataset is constructed to facilitate AI EDA-related research. The experimental results demonstrate that our HGP can reduce the prediction error of power, critical path delay and resource utilization to 4.21%~7.72%, which outperforms the state-of-the-art works significantly. BOME integrated with our novel algorithm MOTPE-FL can achieve better Pareto fronts than meta-heuristic algorithms SA and NSGA-II, with PPA gains of 72.00% and 30.47% respectively. BOME with HGP can accelerate the DSE process by up to 24 × with an average speedup of 14×.