Review: Kd‐tree Traversal Algorithms for Ray Tracing

Abstract

In this paper we review the traversal algorithms for kd-trees for ray tracing. Ordinary traversal algorithms such as sequential, recursive, and those with neighbour-links have different limitations, which led to several new developments within the last decade. We describe algorithms exploiting ray coherence and algorithms designed with specific hardware architecture limitations such as memory latency and consumption in mind. We also discuss the robustness of traversal algorithms as one issue that has been neglected in previous research.

Improving Memory Space Efficiency of Kd‐tree for Real‐time Ray Tracing

“…However, fast kd‐tree construction is now possible via the effective parallelization of the construction process on modern CPUs and GPUs, as reported previously [ZHWG08, CKL*10, WZL11]. Furthermore, considerable attention has been paid to the development of memory‐efficient kd‐tree layouts and the optimization of the traversal algorithm for specific hardware (for example, refer to a recent review article [HH11]).…”

Optimal Spatial Subdivision method for improving geometry navigation performance in Monte Carlo particle transport simulation

“…The advantage of uniform scheme is that it is easy to be implemented and provides O(1) complexity for grid location, but it is an object-independent method which still suffers a low-efficient problem when the objects in the 3D space are not distributed uniformly. Non-uniform scheme is an object-dependent method which usually uses an octree or kd-tree (Hapala and Havran, 2011) due to its ability of matching the subdivided grids with the distribution of geometric objects in the 3D space. However, the non-uniform scheme is more complex and time-consumed than uniform scheme for grid location.…”

Efficiency of k-d Tree Ray-Traversal Algorithms

“…While the author needs an efficient way to store nodes. Beside, in order to reduce memory latency when traversing through the k-d tree, many methods of memory layout [6] has been raised, which have a common concept that after traversing a parent node, people want its left or right child node can be in same cache as itself, because memory layout in same cache have an advantage of data locality. In our implementation, the author uses breadth-first tree node memory layout which depicted in Figure 4 In this section, the writer will use an example illustrated in Figure 5 to tell the difference between empty node and nonempty node.…”