Analysis for supporting real-time computer vision workloads using OpenVX on Multicore+GPU platforms

Abstract

In the automotive industry, there is currently great interest in utilizing computer vision algorithms to support driver-assist and autonomous-control features. OpenVX is an emerging standard for supporting workloads in which such algorithms are applied. OpenVX uses a graph-based software architecture designed to enable efficient computation on heterogeneous platforms that may include CPUs, graphics processing units (GPUs), digital signal processors (DSPs), and other accelerators. Unfortunately, in settings where real-time constraints exist, the usage of OpenVX poses certain challenges. In a recent paper, the authors presented a new implementation of OpenVX directed at platforms comprised of CPUs and GPUs that leverages various analytical techniques to enable these challenges to be addressed. In this paper, these analytical techniques are presented and discussed in detail. These techniques enable end-to-end frame processing times to be analytically bounded under OpenVX while encouraging parallelism through pipelining. Additionally, they enable bounds on frame buffering requirements to be determined.