ACCO: Automated Causal CNN Scheduling Optimizer for Real-Time Edge Accelerators

Paper:

Yin J, Mei L, Guntoro A, et al. "ACCO: Automated Causal CNN Scheduling Optimizer for Real-Time Edge Accelerators." 2023 IEEE 41st International Conference on Computer Design (ICCD). IEEE, 2023.

Abstract

ACCO is an automated optimizer that explores efficient Causal CNN transformation and DF scheduling for ST-CNNs on edge hardware accelerators. It is built upon our previous depth-first DNN scheduling DSE framework DeFiNES, HPCA, 2023.

This repository contains the basic scripts to perform the cost-model-based DSE and the result analysis for the experiments in the paper. For more low-level implementation details and explanations, please refer to the DeFiNES repository.

Installation

1. Install Anaconda environment

2. Download ACCO (git clone this repo)

3. Use a terminal or an Anaconda Prompt for the following steps:

   • cd into the ACCO repo
   • Create the conda environment from the environment.yml file

     conda env create -f environment.yml
     

   • Activate the new environment:

     conda activate ACCOenv
     

Run

The exploration includes the following steps. Please refer to each scripts' comments for further explanation and available tweaks.

Step 1

Run

python DfStackCutAndXform_mp_misc.py

What does this script do?

It triggers the main DSE execution across the combination of assigned hardware model, DNN workload model, tiling options, depth-first fusion ranges, and etc. The output of this script is the cost model ensemble of each scenario stored in .pkl files.

Please note that if using a broad range of workloads and DSE options in one launch, the execution time and disk usage would significantly grow.

Step 2

Run

python explore_pickle_compile_readings.py

What does this script do?

It parses the cost model ensembles from the previous step's output and compiles reports in terms of the frame-wise, layer-wise, and total hardware cost of different depth-first scheduling flavours. These reports are on the level of each fused depth-first layer trunk, so that this is the preparation step for the final exploration of the best strategy across the entire DNN workload.

Step 3

Run

python explore_pickle_hierarchy_cache_allocation.py

What does this script do?

This script leverages the distilled fuse-layer-level cost model reports to find the best overall scheduling strategy across the entire DNN workload.

---

Citation

@inproceedings{yin2023acco,
  title={ACCO: Automated Causal CNN Scheduling Optimizer for Real-Time Edge Accelerators},
  author={Yin, Jun and Mei, Linyan and Guntoro, Andre and Verhelst, Marian},
  booktitle={2023 IEEE 41st International Conference on Computer Design (ICCD)},
  pages={391--398},
  year={2023},
  organization={IEEE}
}

We are continually improving the framework, and welcoming all questions and feedback.

We hope our tool can help other researchers to better explore and understand the vast DNN accelerator architecture-and-scheduling design space and can offer the best design solutions.

For the latest features of this framework, please refer to the main repository ZigZag and Stream.