A GPU implementation of tiled belief propagation on Markov random fields

In this work, we present a parallelized version of tiled belief propagation for stereo matching. The proposed algorithm is implemented in CUDA to leverage parallel processing capabilities of GPUs. In our solution, the original tiled BP algorithm is combined with a number of optimizations specific to parallel programs in CUDA. For the given test inputs, the proposed solution runs in 7.96 milliseconds on Nvidia Tesla C2050, achieving acceptable accuracy with respect to the reference code. This work has been published in 2013 Eleventh ACM/IEEE International Conference on Formal Methods and Models for Codesign (MEMOCODE 2013), winning the MEMOCODE Design Contest 2013 in the adjusted cost-accuracy category. To the best of authors knowledge, this represented the first work in optimizing a parallelized version of the tiled BP algorithm. After presenting our approach, at selecting an appropriate candidate algorithm for parallelization and implementing in on GPU by applying a series of appropriate optimizations, we discuss the current state of the art on stereo matching, that has been presented since publishing this work.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Kotsifakou, Maria

Contributors dc:contributor

• Adve, Vikram S.

• stereo matching
• tiled belief propagation
• GPUs

Statement dc:rights

• Copyright 2019 Maria Kotsifakou

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/104846

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/104846