Optimal Design Parameters for Tiny-YOLO2 Implementation with Light-Weight Embedded GPGPU Environment

摘要

The aim of this paper is to find the optimal design for tiny YOLO2 in an embedded system environment. Our focus is to rebuild the given YOLO2 code and to find the optimal design parameters in order to maximize the speed using the light-weight GPGPU in a target embedded environment. To maximize the YOLO2 performance we used OpenCL framework while utilizing the embedded GPGPU and tried various aspects of OpenCL design parameters such work item, work group, and resulting in reducing the global memory access overhead and maximizing computing load balancing between computing units under constraints including local memory resources and computing resources. Our experimental results show that the overall performance enhancement is 18.2 times compared to the naive implementation.