ant_simulator

ANT Simulator

This repository contains the code for the ANT simulator based on DNNWeaver and BitFusion.

Prerequisite

• Ubuntu 18.04.5 LTS
• Andconda 4.10.1
• Python 3.8
• gcc 7.5.0

Getting Started

$ # Environment.
$ conda create -n ant_sim python=3.8
$ conda activate ant_sim  
$ pip install -r  requirements.txt
$ # Cacti for the memory simulation.
$ git clone https://github.com/HewlettPackard/cacti ./bitfusion/sram/cacti/
$ make -C ./bitfusion/sram/cacti/
$ # Run ANT simulation.
$ python ./run_ant.py       # About 15 minutes

Evaluation

The script run_ant.py generates statistic data and stores it in file ./result/ant_res.csv. Note that BiScaled only test on VGG16 and ResNet50.

In ./result/ant_res.csv, Line 3 shows the cycle data that normalized with AdaFloat. Line 7-10 shows the energy data that normalized with AdaFloat.

As shown below, the ./result/ANT-simulator.xlsx provides the template. You can fill it with the numbers of ./result/ant_res.csv to generate Figure 13 in the paper.