Performance Optimization of Computer Vision Applications: A Case Study on STHORM Architecture

Abstract

In the realm of embedded many-core architectures, computer vision applications are a major motivator. A harmony between computation and communication is vital for such systems to reach their full potential, however many computer vision techniques provide a highly data dependent actions that make the work harder. The developer needs access to tools for rapid and accurate application-level performance analysis in order to optimize application performance. Here, we go through the steps required to migrate and fine-tune a face recognition program for the STHORM many-core accelerator by way of the STHORM OpenCL software development kit. We isolate the key bottlenecks and separate the effects of the application, the openly programming paradigm, and the STHORM Openly software development kit (SDK). Finally, we demonstrate how these problems might be fixed in the near future to provide programmers even more leeway to enhance the efficiency of their applications.