M4BRAM-Artifacts

This github repository contains the necessary artifacts for the paper:

Y. Chen, J. Dotzel and M. Abdelfattah, "M4BRAM: Mixed-Precision Matrix-Matrix Multiplication in FPGA Block RAMs", to appear in Proceedings of the IEEE International Conference on Field Programmable Technology (FPT), December 2023.

 

Overview

The scripts folder contains all scripts to reproduce the results of Section-V.D, Section-V.E, Section-V.F of the paper.

The COFFE folder is a modified version of the COFFE simulator (https://github.com/vaughnbetz/COFFE), with additional features to describe and model the circuits of M4BRAM. We use COFFE to evaluate the area and frequency of M4BRAM (described in Section V-B of the paper).

 

Tested Environment & Dependencies

OS requirement:

• Any OS with python3 to run files in the "scripts" directory.
• CentOS Linux 7 (or above) to run files in the "COFFE" directory.
• Ubuntu 20.04.6 (or above) to run DNN quantization experiments.

To run "scripts" and "COFFE", the following software is required.

• Python 2.7
• Synopsys Hspice 2013.03 (or above)
• Python 3.9 (require numpy, pandas, matplotlib packages)
• Bash

 

Summary of Artifact Executions

Run the following command in Terminal to clone this github repository:

git clone https://github.com/abdelfattah-lab/M4BRAM.git

Then go to the "M4BRAM" repository in your terminal by running.

cd M4BRAM

1. Running the "scripts" folder

Run the following commands after going to the "M4BRAM" repository in your terminal.

cd scripts
bash run_experiments.sh

It will take several minutes to finish. A new directory "results" under "scripts" will be created and contain experiment results in .txt format. To view the results, go to the "results" directory. There are totally 4 .txt files:

• Section_VD_mixed_activation_results.txt: This file contains the results shown in Fig. 9 of the M4BRAM paper.
• Section_VD_mixed_weight_results.txt: This file contains the results shown in Table III of the M4BRAM paper.
• Section_VE_m4bram_vs_bramac_results.txt: This file contains the results shown in Fig. 10 of the M4BRAM paper.
• Section_VE_m4bram_vs_bramac_ablation_results.txt: This file contains the results shown in Fig. 11 of the M4BRAM paper.
• Section_VF_no_dsp_results.txt: This file contains the results shown in Fig. 12 of the M4BRAM paper.

We also provide the scripts to reproduce the figures in the M4BRAM paper. Run the following command in your terminal.

bash plot_results.sh

Then 4 figures in .svg format will be generated under the "results" directory. The 4 figures' names contain the corresponding figure numbers in the M4BRAM paper.

2. Running the "dnn-compression" experiments

Please go to the "dnn-compression" folder and follow the instructions to run the experiments.

3. Running the "COFFE" simulator for M4BRAM

Before running this experiment, please source the necessary bash files in your terminal to enable Synopsys HSpice. Make sure that the following command:

hspice

can be run successfully in the current terminal.

Run the following commands (assuming still in the "scripts" directory after running the first experiments):

cd ../COFFE
python2 coffe.py -i 6 input_files/M4BRAM/M4BRAM_S_16k.txt
python2 coffe.py -i 6 input_files/M4BRAM/M4BRAM_L_16k.txt

The two python2 commands evaluate M4BRAM-S and M4BRAM-L, respectively. Each of the two COFFE experiments takes roughly 12 hours to complete (depending on the computer's performance). After completing, users can view the reported delay and area by running the following commands inside the COFFE repository:

cd input_files/M4BRAM/M4BRAM_S_16k
open report.txt 
cd input_files/M4BRAM/M4BRAM_L_16k
open report.txt 

The first report.txt file shows the COFFE results for M4BRAM-S, while the second report.txt shows the COFFE results for M4BRAM-L. The area and delay information of M4BRAM is inside the report.txt file, under DUMMY ARRAY AREA CONTRIBUTIONS.

---

This work is subject to a patent application filed by Cornell University.