HLSPilot is the first automatic HLS code generation and optimization framework from sequential C/C++ code using LLM.
This framework investigates the use of LLM for HLS design strategy learning and tool learning, and build a complete hardware acceleration workflow ranging from runtime profiling, kernel identification, automatic HLS code generation, design space exploration, and HW/SW co-design on a hybrid CPU-FPGA computing architecture.
This project was tested on Xilinx Alveo U280 and Vitis HLS 2019.1 suite. The following python libraries are required:
- openai (v1.33.0)
- langchain (v0.2.3)
-
Place your code under benchmark directory
-
Run Code Analysis: This step will call the gprof tool to profile the code and extract performance bottlenecks, thereby helping to complete the HW-SW partition.
python3 src/hw_sw_partition/auto_analysis.py
-
Task pipeline: This step will split the kernel into multiple subtasks, allowing for pipeline execution between subtasks and improving program parallelism.
# update code_file_path in the script python3 src/task_pipeline/task_pipeline.py -
Task optimization: This step will optimize the subtask code by retrieving suitable pragma from the optimization strategy knowledge base and applying it to the code.
# update code_file_path in the script python3 src/task_opt/task_opt.py -
DSE: This step will call the DSE tool to complete the parameter tuning of pragma
cd src/auto_dse python3 aoto_dse.py ./exec.sh run APP_NAME .cpp
If you find this work useful in your research, please consider citing our work:
@article{xiong2024hlspilot,
title={HLSPilot: LLM-based High-Level Synthesis},
author={Xiong, Chenwei and Liu, Cheng and Li, Huawei and Li, Xiaowei},
journal={arXiv preprint arXiv:2408.06810},
year={2024}
}