ForgeHLS

A Large-Scale, Open-Source Dataset for High-Level Synthesis

ForgeHLS Workflow

The following diagram provides a comprehensive overview of the ForgeHLS workflow and dataset construction process:

Repository Structure

source_code/
├── forgehls_dataset_json/          # ForgeHLS dataset in JSON format
│   ├── data_of_designs_forgehls_with_strategy_formatted.json.gz
│   └── README.md                   # Dataset format documentation
├── dataset_construction/           # Dataset construction scripts
│   ├── src/                       # Source code for dataset generation
│   ├── plot_dataset_QoR/          # Quality of Results plotting scripts
│   └── algo_list/                 # Algorithm configurations
└── downstream_task/               # Downstream task implementations
    ├── QoR_Prediction/           # Quality of Results prediction
    └── automatic_pragma_insertion/ # Automatic pragma insertion
README.md                          # This file

ForgeHLS Dataset

The json format dataset is located in code/forgehls_dataset_json/ and contains: For detailed dataset format information, see code/forgehls_dataset_json/README.md.

The full dataset and all related artifacts are available on huggingface/ForgeHLS, with official hosting support provided by the Hugging Face team upon request for extended storage to accommodate large-scale benchmarks.

ForgeHLS-benchmark

ForgeHLS-benchmark contains 80 representative and structurally diverse C++ files out of ForgeHLS.

Prerequisites

• HLS tools (Vitis HLS)
• LLVM and Clang

llvm-config --version
sudo apt install clang=x.x.x  libidn11-dev libclang # the clang version should be the same as llvm
find / -name "libclang.so" 2>/dev/null
export LD_LIBRARY_PATH=/usr/lib/llvm-14/lib:$LD_LIBRARY_PATH

• Required Python packages:

pip install openai anytree pandas matplotlib notebook seaborn scikit-learn scikit-optimize tiktoken

Citation

If you use ForgeHLS in your research, please cite our paper:

@misc{peng2025forgehls,
    title={ForgeHLS: A Large-Scale, Open-Source Dataset for High-Level Synthesis},
    author={Zedong Peng and Zeju Li and Mingzhe Gao and Qiang Xu and Chen Zhang and Jieru Zhao},
    year={2025},
    eprint={2507.03255},
    archivePrefix={arXiv},
    primaryClass={cs.AR}
}

Team's Other Work

For more information about ForgeHLS and other datasets from our team (ForgeEDA, DeepCircuitX, and ForgeHLS), please visit our team's GitBook.