Novel Parallelization Techniques for Computer Graphics Applications

Abstract

Increasingly complex and data-intensive algorithms in computer graphics applications require software engineers to find ways of improving performance and scalability to satisfy the requirements of customers and users. Parallelizing and tailoring each algorithm of each specific application is a time-consuming task and its implementation is domain-specific because it can not be reused outside the specific problem in which the algorithm is defined. Identifying reusable parallelization patterns that can be extrapolated and applied to other different algorithms is an essential task needed in order to provide consistent parallelization improvements and reduce the development time of evolving a sequential algorithm into a parallel one.

This thesis focuses on defining general and efficient parallelization techniques and approaches that can be followed in order to parallelize complex 3D graphic algorithms. These parallelization patterns can be easily applied in order to convert most kinds of sequential complex and data-intensive algorithms to parallel ones obtaining consistent optimization results.

The main idea in the thesis is to use multi-threading techniques to improve the parallelization and core utilization of 3D algorithms. Most of the 3D algorithms apply similar repetitive independent operations on a vast amount of 3D data. These application characteristics bring the opportunity of applying multi-thread parallelization techniques on such applications. The efficiency of the proposed idea is tested on two common computer graphics algorithms: hidden-line removal and collision detection. Both algorithms are data-intensive algorithms, whose conversions from a sequential to a multithread implementation introduce challenges, due to their complexities and the fact that elements in their data have different sizes and complexities, producing work-load imbalances and asymmetries between processing elements.

The results show that the proposed principles and patterns can be easily applied to both algorithms, transforming their sequential to multithread implementations, obtaining consistent optimization results proportional to the number of processing elements. From the work done in this thesis, it is concluded that the suggested parallelization warrants further study and development in order to extend its usage to heterogeneous platforms such as a Graphical Processing Unit (GPU). OpenCL is the most feasible framework to explore in the future due to its interoperability among different platforms.