Optimizing Ultra High-resolution Video Processing on Mobile Architecture with Massively Parallel Processing

Abstract

This paper introduces an optimized video frame pre-processing scheme for UHD video with up to 8K resolution using OpenCL for mobile architectures, particularly for convolution. The introduced scheme can fully utilize the maximum computational resources of the mobile architecture, with an adaptive work-group size adjustment. As a prototype, a simple video player with a Sobel kernel was implemented as an example of a convolution kernel. The prototype implementation showed a better video frame processing time than the de-facto image-processing library, OpenCV. On the other hand, the decoding time of the video increased because the OpenCL kernel utilizes GPU (Graphics Processing Unit) resources to almost its maximum. In the future, the processing workload will be distributed between the CPU (Central Processing Unit) and GPU to achieve higher performance.