OpenVX Graph Optimization for Visual Processor Units

OpenVX is a standardized, cross-platform software framework to aid in development of accelerated computer vision, machine learning, and other signal processing applications. Designed for performance optimization, OpenVX allows the programmer to define an application using a graph-based programming model, where the nodes are selected from a repertoire of pre-defined kernels and the edges represent the flow of successive images between pairs of kernels. The graph-based representation exposes spatial and temporal concurrency and provides tuning opportunities to the managing runtime library. In this paper, we present a performance model-based approach for optimizing the execution of OpenVX graphs on the Texas Instruments C66x Digital Signal Processor (DSP), which has similar characteristics to other widespread DSPs such as the Qualcomm Hexagon, Nvidia Programmable Vision Accelerator, and Google Visual Pixel Core. Our approach involves training performance models to predict the impact of tile size and node merging on performance and DRAM utilization. We evaluate our models against randomly-generated, valid, and executable OpenVX graphs.