Review and Comparative Study of Ray Traversal Algorithms on a Modern GPU Architecture

Abstract

In this paper we present a chronological review of five distinct data structures commonly found in literature and ray tracing systems: Bounding Volume Hierarchies (BVH), Octrees, Uniform Grids, KD-Trees, and Bounding Interval Hierarchies (BIH). This review is then followed by an extensive comparative study of six different ray traversal algorithms implemented on a modern Kepler CUDA GPU architecture, to point out pros and cons regarding performance and memory consumption of such structures. We show that a GPU KD-Tree ray traversal based on ropes achieved the best performance results. It surpasses the BVH, often used as primary structure on state-of-the-art ray tracers. Compared to BVHs, a carefully well implemented ropes-based KD-Tree is from 13 to 71% faster. However, it may consume up to 9x more memory than other structures. This disadvantage can be a limiting factor on memory limited architectures.