Cross Domain Multi-Acceleration Benchmarks

FPGA-based domain-specific accelerators are increasingly becoming popular, but currently there is a lack of systems support to utilize multiple accelerators across various application domains. Moreoever, currently there does not exist a widely established benchmark suite that can be used to measure the benefits of cross-domain multi-acceleration. To address this issue, we provide a benchmark suite that consists of six real-life applications ranging from deep brain stimulation, film captioning, medical imaging, and more. This benchmark suite is developed at the Alternative Computing Technologies (ACT) Laboratory, University of California, San Diego.

Applications

1. ** Memory Enhance [DSP, Data Analytics, Optimized Control] **: In this application, brain electrophysiological activity is collected in real-time and passed through Fast Fourier Transform. Then, logistic regression is used to decode and classify the signals to be used as biomarkers. Lastly, Model Predictive Control is used configure the synthesized signals depending on the classification output from the previous step. This pipeline can be executed repeatedly to enhance memory capacity of small lab rats.

2. ** Robot Explorer [Robotics, Computer Vision] **: This application consists of a four-sheeled autonomous robot equipped with a Kinect sesnsor to navigate its route through small, narrow tunnels using the Model Predictive Control algorithm. Furthermore, it reconstructs a 3D map of the surrounding environment using the KinectFusion algorithm.

3. ** Video Sync [DSP] **: In this application, given a video file and a corresponding subtitle text file, it synchronizes the subtitles with speech segments using MPEG-decoding and Fast Fourier Transform to boost the speech-text pattern matching process.

4. ** Stock Market [Data Analytics, Finance] **: This application performs a sentiment analysis on news article texts using logistic regression, and uses Black-Scholes to predict call option pricing.

5. ** Leukocyte [Computer Vision] **: This application detects leukocytes from video microscopy of blood vessels, where detection in the frame is done via Gradient Inverse Coefficient of Variation scores, and leukocytes are detected via Motion Gradient Vector Flow matrix.

6. ** Security Camera [Deep Learning, DSP] **: This application decoddes MPEG encoded input video stream, and performs an object detection task using the Tiny-YOLO-v2 model.

System Requirements

This benchmark suite requires the following systen environment parameters and tools in order to succesfully compile and run.

• Ubuntu 16.04.5 LTS (Xenial Xerus) or newer
• gcc (and g++) version 5.4.0
• CMake version 3.5.1 or newer
• Python version 2.7 or newer
• LAPACK version 3.6.0
• LAPACKE version 3.6.0

Build and Run Instructions

We provide two options to build the benchmark suite: you can either build all the benchmark applications at once, or build each benchmark application individually. To build all applications at once, we provide a utility script build_all.sh located in scripts/ directory. You can simply run the script like so: ./scripts/build_all.sh. If you would like to build each benchmark individually, go to its respective directory under applications/ and follow the build instructions on the README file.

To run the benchmarks, go to each benchmark application directory and follow the directions on the respective README file.