Odyssey

Publication

• Suhail Basalama, Jie Wang, Jason Cong. [A Comprehensive Automated Exploration Framework for Systolic Array Designs]. In DAC, 2023.

About

This repo contains the source code of Odyssey, an automated framework incorporating a hyprid genetic-mathematical programming search method, a padding-based search algorithm, and other randomized search methods. It can be used to explore the huge design space of systolic arrays. The framework is tested using the open-source AutoSA framework, but the methodology of our paper is applicable for other systolic array designs.

Content

1. Requirements and Dependencies

3. Running the Project

Requirements and Dependencies

The hybrid genetic/mathematical-programming method requires the installation of the ampl solver.

All other dependencies are python packages found in the requirement.txt

Running the Project

To perform the tests, run the python 'run_tests.py' script under the 'tests' directory. For example:

source env.sh
cd tests
python run_tests.py --workload=mm

To test other problem sizes, please add your tests to a json file similar to 'tests/mm.json', and rerun the previous commands.

Citation

If you find any of the ideas/codes useful for your research, please cite our paper:

@inproceedings{basalamadac2023,
    title={A Comprehensive Automated Exploration Framework for Systolic Array Designs},
    author={Basalama, Suhail and Wang, Jia and Cong, Jason},
    booktitle={2023 60th ACM/IEEE Design Automation Conference (DAC)},
    year={2023}
}