Large-scale construction of diverse HLS design datasets

Construct large, diverse datasets of high-level synthesis (HLS) designs that extend beyond common community benchmarks, addressing the dataset-scaling and diversity limitations that impede training quality-of-results models and benchmarking large language models on HLS tasks.

Background

The paper identifies the limited size and diversity of existing HLS benchmarks as a barrier to research on quality-of-results prediction, design-space exploration, and large-language-model-based HLS design. Although HLSFactory-Agent is proposed to automate the extraction of standalone synthesizable designs from academic and open-source repositories, the broader challenge of building sufficiently large and diverse datasets remains unresolved in the paper.

The proposed system addresses part of this challenge by automating repository analysis, design extraction, compilation checks, and Vitis HLS validation. However, the paper presents only initial results from 26 repositories, yielding 271 candidate designs, of which 130 pass synthesis, so it does not establish that the general dataset-construction problem has been solved at the scale or diversity required by the HLS research community.