Improving Ray Traversal by Using Several Specialized Kd-Trees

Abstract

In this paper, we present several variants of the Surface Area Heuristics (SAH) to build kd-trees for specific sets of rays' directions. In order to cover the whole space of directions, several sets of directions are considered and each of them leads to a different specialized kd-tree. We call Multi-kd-tree to the set of these kd-trees. During rendering, each ray will traverse the kd-tree associated with the set containing its direction. In order to evaluate the efficiency of our proposal, we have implemented a Path Tracing and an Ambient Occlusion renderer on GPU with CUDA. A SAH-based kd-tree has been compared to a Multi-kd-tree and we show that all the new heuristics exhibit a better performance than SAH over usual scenes.