Execution framework for mGEMM: Low-latency Convolution with Minimal Memory Overhead Optimized for Mobile Devices in MobiSys'22

This is the modified Darknet framework used in the evaluation of MobiSys'22 paper mGEMM: Low-latency Convolution with Minimal Memory Overhead Optimized for Mobile Devices [Paper Link]

Below is the teaser for mGEMM that I made for the MobiSys'22 conference. (Click the image to watch it on YouTube!)

You can download additional materials (Posters, Presentation Slides, Graphics) used in this project from my [Homepage].

---

Usage Instructions

mGEMM, amongst other convolution solutions from XNNPACK, ARMNN, etc., is included in this repository and can be selected as a drop-in replacement of the baseline convolution kernel from the Darknet framework.

1. Use the makefile to build the program. Please use ARMv8 & Aarch64 (64bit OS) based environments. Raspberry Pi 4 with 64-bit Raspberry Pi OS is known to work.

2. Use the command "wget https://pjreddie.com/media/files/yolov3-tiny.weights" to download the filter weights for the YoloV3-Tiny network.

3. Use the command "./darknet detector backend cfg/coco.data cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg <CONVOLUTION_BACKEND>" to run the YoloV3-Tiny with different convolution backend. <CONVOLUTION_BACKEND> can be either GEMMPLUS, ARMNN, XNNPACK, DEFAULT, or OPENBLAS, depending on your choice. GEMMPLUS refers to the mGEMM algorithm.

mGEMM is referred to as "PTMM" or "GEMM PLUS" in the code. These are just the development names of mGEMM.

Libraries (.a files) for mGEMM, XNNPACK, and ARMNN versions we used are included in the git. ARMNN library is modified to use the GEMM-based convolution method.

You'll have to install OPENBLAS on your own. "apt install libopenblas-dev" usually works. Please refer to the Makefile for other dependencies.

Please reach out to me at "cakeng at snu dot ac dot kr" if you have any questions!

---

Source code Links

The source code for the mGEMM library can be found at https://github.com/cakeng/GEMM_Plus

The source code for the ARMNN library can be found at https://github.com/ARM-software/armnn.

The source code for the XNNPACK library can be found at https://github.com/google/XNNPACK.

Original Darknet Repo can be found at https://github.com/pjreddie/darknet