GAMMA-TimeLoop

This is the implementation of the mapper GAMMA using Timeloop as cost model.

GAMMA-Timeloop searches through the design space modeled by Timeloop and proposes an optimized mapping. It was introduced in this IISWC 2022 paper and talk.

Sister Repo: Gamma-Maestro

We also have GAMMA supporting MAESTRO as cost model. The sister repo can be found here Gamma-Maestro. It searches through the design space of MAESTRO and proposes an optimized mapping.

---

Install Dependency

Install Timeloop

Timeloop documentation is hosted at https://timeloop.csail.mit.edu/timeloop. The guides there cover detailed installation steps.

Install Timeloop-python

python build_pytimeloop.py

For more installation detail, please visit https://timeloop.csail.mit.edu/timeloop.

---

Take a Trial Run

Run GAMMA-Timeloop

./run_gamma_timeloop.sh

Run GAMMA-Timeloop with multi-objective

./run_gamma_timeloop_multiObjective.sh

For more detail, please look at ./src

---

Citation

@inproceedings{gamma_timeloop,
    author       = {Kao, Sheng-Chun and Parashar, Angshuman and Tsai Po-An and Krishna, Tushar},
    title        = {Demystifying Map Space Exploration for NPUs},
    booktitle     = {IISWC},
  year          = {2022}
}

@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}
}