AutoShrink 2.0: Automating Large-Scale Pareto Optimization in HLS via Adaptive and Cost-Aware Search Space Shrinkage

High-level synthesis (HLS) expedites accelerator customization by delivering a high-level hardware programming paradigm and offering a rich set of optimization directives. However, the quality of HLS designs is highly dependent on the designers’ selection of directives in the trade-off between different design metrics, which is non-trivial and can lead to a large design turnaround time. Design space exploration (DSE) serves as a promising solution to this problem, but existing studies on DSE suffer from a lack of scalability for large-scale applications and often spend a substantial portion of the exploration budget on costly design points. To address this problem, this paper proposes AutoShrink 2.0 , a DSE engine that adaptively shrinks the large search space of an HLS design to gradually retain only high-quality solutions. AutoShrink 2.0 incorporates (1) a comprehensive design space pruning strategy that integrates the joint effect of directives and domain knowledge, (2) an importance-guided Pareto optimization algorithm which dynamically traces the importance ranking of the applied directives and leverages this ranking to navigate the search direction of Pareto frontiers, and (3) a cost-aware post-filtering mechanism to screen out candidates that are computationally expensive yet of limited gain, reducing runtime overhead within a single evaluation. Experimental results on Vitis HLS demonstrate that AutoShrink 2.0 achieves an average distance to reference set (ADRS) of 8.40% across multiple applications with design spaces containing up to \(10^16 \) design points, outperforming three heuristic methods and three state-of-the-art approaches by 8.01 × and 6.24 ×, respectively. On timeout-prone benchmarks, the cost-aware post-filtering strategy reduces the average timeout ratio by 66.1% compared with the variant without this strategy. Furthermore, design tool transfer experiments from Vitis HLS to Catapult HLS demonstrate the promising cross-tool applicability of AutoShrink 2.0.