GNΩSIS: Lessons Learned in Generating a High-Level Synthesis Dataset

High-Level Synthesis (HLS) streamlines FPGA programming by abstracting low-level hardware complexities, facilitating rapid microarchitecture customization through the use of directives. However, identifying optimal directives remains a significant challenge, particularly for software developers without extensive hardware expertise. HLS-driven Design Space Exploration (DSE) addresses this challenge by automating the generation of directive configurations through a variety of techniques. The effectiveness of these methods heavily relies on the quality of the data on which they are built. Unfortunately, existing datasets are often constrained in scope, complexity, and device diversity. To address these limitations, we present GNΩSIS, 1 the largest open-source HLS dataset, to the best of our knowledge, containing almost 219K design points generated from prominent benchmark suites and public repositories. The dataset covers two FPGAs with varying resource characteristics and three clock frequencies. Additionally, we introduce a versatile framework to automate design point generation. We evaluate the impact of frequency scaling, FPGA architecture, and optimization targets on Quality of Result (QoR) metrics, analyze HLS directive effectiveness across code constructs, and study design transferability across FPGAs and frequencies. These analyses provide practical insights into QoR and HLS behavior, offering guidance for designers with limited optimization experience. We believe GNΩSIS will enhance FPGA accessibility and act as an enabler for innovation in HLS-driven DSE research. The dataset is publicly accessible on Hugging Face, 2 and the framework is available on GitHub. 3