Home // ACCSE 2018, The Third International Conference on Advances in Computation, Communications and Services // View article

Data Stream Optimization of Sum of Absolute Differences Algorithm on a Graphics Processing Unit

Authors:
Tom Pudpai
Tae Kyun Kim
Charles Liu

Keywords: - data streaming; Sum of Absolute Differences algorihtm; massive parallel architecture

Abstract:
This paper describes the data streaming approaches to performance optimization of the Sum of Absolute Differences (SAD) algorithm on an NVIDIA Graphics Processing Unit (GPU) using the OpenCL programming paradigm. The SAD algorithm forms one of several steps required to implement stereo vision. It creates pixel-based disparity maps from two concurrent images captured by a pair of cameras positioned with a distance in between. The disparity maps can be used to derive depths of objects in the scenes of interest. The massively parallel architecture of a GPU can take advantage of the highly parallelizable SAD algorithm. OpenCL programming framework was chosen to develop the parallel algorithm on the GPU. Performance gains are realized by explicitly mapping data from the slower global memory to the faster shared local memory of the GPU. Local memory is loaded by either a centralized or distributed approach from the OpenCL-defined work-items operating in a workgroup. The resulting performance improvements were discussed based on the architectural features of the GPU and the data streaming approaches used in this research work.

Pages: 22 to 27

Copyright: Copyright (c) IARIA, 2018

Publication date: July 22, 2018

Published in: conference

ISSN: 2519-8459

ISBN: 978-1-61208-658-3

Location: Barcelona, Spain

Dates: from July 22, 2018 to July 26, 2018