ebugger / optimus

This is the implementation of the paper [Optimus: Towards Optimal Layer-Fusion on Deep Learning Processors].

Optimus

This is the implementation of the paper [Optimus: Towards Optimal Layer-Fusion on Deep Learning Processors], which has accepted by LCTES2021.

Getting Started Guide

Start from Source Code

1. Create virtual env

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

2. Install requirement

pip install -r requirements.txt

3. Run a test

./test.sh

4. To find out all the options

python ./fusion/tools/optimal_schedule_search.py --help

Experiments

1. Run overall experiment to get the memory access and energy over multiple models The result will be stored in result/overall_experiment/. It will take more than ten minutes to complete this experiment.

python ./fusion/experiment/overall_experiment.py

2. Run memory access analysis over multiple models The result will be stored in result/analysis/. It will take more than ten minutes to complete this experiment.

python ./fusion/experiment/analysis.py

3. Evaluate the Impact of Batch Size The result will be stored in result/batch_size/.

python ./fusion/experiment/batch_size.py

4. Evaluate the impact of on-chip memory space The result will be stored in result/buffer_size/.

python ./fusion/experiment/buffer_size.py

5. Evaluate the impact of Dataflow This experiment supports the experiment results of section 4.2.5 in our paper, and the result will be stored in result/dataflow/.

python ./fusion/experiment/dataflow.py

6. Evaluate the impact of PE-array and buffer The result will be stored in result/pe_array/.

python ./fusion/experiment/pe_array.py

7. Evaluate the performance on different processors The result will be stored in result/processor/.

python ./fusion/experiment/processor.py