GAMMA

This is the implementation of the paper GAMMA, GAMMA is an autonomous framework for optimizing the HW mapping of DNN models on the DNN Accelerators. We use MAESTRO as our cost model.

Extended Works

This code repository also incorporates the code-bases of the extended works:

• DiGamma
• Formalism of Accelerator Flexibility

---

Sister Repo: Gamma-Timeloop

• We also have GAMMA supporting Timeloop as cost model. The sister repo can be found here -- Gamma-Timeloop.

  It enables using GAMMA algorithm to search through the design space of Timeloop, a DNN cost model from NVIDIA.

---

Installation

• Create virtual env

conda create --name gammaEnv python=3.6
conda activate gammaEnv

• Install requirement

pip install -r requirements.txt

• Download cost model and build symbolic link

python build.py

• Setup larger limitation for opened file if there is warning "Too many open files." (for threading)

ulimit -n 4096

Take a Trial Run

./run_gamma.sh

---

Different Map Space Exploration Scenarios

• Map space exploration for fully flexible accelerator (in a full search space): Related reading - GAMMA.
• Map space exploration for partially flexible accelerator (in a constrained search space): Related reading - Formalism of Accelerator Flexibility.
• PE(HW)-Mapping Co-exploration: Related reading - DiGamma.

More details can be found here

Resources

• Tutorial of GAMMA, in IEEE/ACM International Symposium on Microarchitecture (MICRO), 2020 [video]
• GAMMA paper presentation, in IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2020 [video]

Contributor

• Sheng-Chun (Felix) Kao
• Tushar Krishna

Pull Request

• We welcome pull request.
• If you have any question, please contact Felix chuchu40507@gmail.com.

Citation

@inproceedings{gamma,
    author       = {Kao, Sheng-Chun and Krishna, Tushar},
    title        = {GAMMA: Automating the HW Mapping of DNN Models on Accelerators via Genetic Algorithm},
    booktitle     = {ICCAD},
  year          = {2020}
}

@inproceedings{digamma,
title={DiGamma: Domain-aware Genetic Algorithm for HW-Mapping Co-optimization for DNN Accelerators},
author={Kao, Sheng-Chun and Pellauer, Michael and Parashar, Angshuman and Krishna, Tushar},
booktitle     = {DATE},
year={2022}
}

@inproceedings{kao2022formalism,
  title={A Formalism of DNN Accelerator Flexibility},
  author={Kao, Sheng-Chun and Kwon, Hyoukjun and Pellauer, Michael and Parashar, Angshuman and Krishna, Tushar},
  booktitle={Proceedings of the 2022 ACM SIGMETRICS/IFIP PERFORMANCE Joint International Conference on Measurement and Modeling of Computer Systems},
  year={2022}
}