This repository presents the novel version of our tried-and-tested HW Architecture-Mapping Design Space Exploration (DSE) Framework for Deep Learning (DL) accelerators. ZigZag bridges the gap between algorithmic DL decisions and their acceleration cost on specialized accelerators through a fast and accurate HW cost estimation.
A crucial part in this is the mapping of the algorithmic computations onto the computational HW resources and memories. In the framework, multiple engines are provided that can automatically find optimal mapping points in this search space.
In this novel version, we have:
- Added an interface with ONNX to directly parse ONNX models
- Overhauled our HW architecture definition to:
- include multi-dimensional (>2D) MAC arrays.
- include accurate interconnection patterns.
- include multiple flexible accelerator cores.
- Enhanced the cost model to support complex memories with variable port structures.
- Revamped the whole project structure to be more modular.
- Written the project with OOP paradigms to facilitate user-friendly extensions and interfaces.
We recommend setting up an anaconda environment.
A conda-spec-file-{platform}.txt is provided to set up the environment for linux64 and win64.
$ conda create --name myzigzagenv --file conda-spec-file-{platform}.txt
Alternatively, you can also install all packages directly through pip using the pip-requirements.txt with the command:
$ pip install -r requirements.txt
The inputs folder provides some example accelerator/workload/settings.
For example, the xception snippet running on a single core accelerator using loma:
$ python main.py --workload inputs.workshop.workload --accelerator inputs.workshop.accelerator