Higher Performance Traversal and Construction of Tree-Based Raytracing Acceleration Structures

Ray tracing is an important computational primitive used in different algorithms including collision detection, line-of-sight computations, ray tracing-based sound propagation, and most prominently light transport algorithms. It computes the closest intersections for a given set of rays and geometry. The geometry is usually modeled with a set of geometric primitives such as triangles or quadrangles which define a scene. An efficient ray tracing implementation needs to rely on an acceleration structure to decouple ray tracing complexity from scene complexity as far as possible. The most common ray tracing acceleration structures are kd-trees and bounding volume hierarchies (BVHs) which have an O(log n) ray tracing complexity in the number of scene primitives. Both structures offer similar ray tracing performance in practice. This thesis presents theoretical insights and practical approaches for higher quality, improved graphics processing unit (GPU) ray tracing performance, and faster construction of BVHs and kd-trees, where the focus is on BVHs. The chosen construction strategy for BVHs and kd-trees has a significant impact on final ray tracing performance. The most common measure for the quality of BVHs and kd-trees is the surface area metric (SAM). Using assumptions on the distribution of ray origins and directions the SAM gives an approximation for the cost of traversing an acceleration structure without having to trace a single ray. High quality construction algorithms aim at reducing the SAM cost. The most widespread high quality greedy plane-sweep algorithm applies the surface area heuristic (SAH) which is a simplification of the SAM. Advances in research on quality metrics for BVHs have shown that greedy SAH-based plane-sweep builders often construct BVHs with superior traversal performance despite the fact that the resulting SAM costs are higher than those created by more sophisticated builders. Motivated by this observation we examine different construction algorithms that use the SAM cost of temporarily constructed SAH-built BVHs to guide the construction to higher quality BVHs. An extensive evaluation reveals that the resulting BVHs indeed achieve significantly higher trace performance for primary and secondary diffuse rays compared to BVHs constructed with standard plane-sweeping. Compared to the Spatial-BVH, a kd-tree/BVH hybrid, we still achieve an acceptable increase in performance. We show that the proposed algorithm has subquadratic computational complexity in the number of primitives, which renders it usable in practical applications. An alternative construction algorithm to the plane-sweep BVH builder is agglomerative clustering, which constructs BVHs in a bottom-up fashion. It clusters primitives with a SAM-inspired heuristic and gives mixed quality BVHs compared to standard plane-sweeping construction. While related work only focused on the construction speed of this algorithm we examine clustering heuristics, which aim at higher hierarchy quality. We propose a fully SAM-based clustering heuristic which on average produces better performing BVHs compared to original agglomerative clustering. The definitions of SAM and SAH are based on assumptions on the distribution of ray origins and directions to define a conditional geometric probability for intersecting nodes in kd-trees and BVHs. We analyze the probability function definition and show that the assumptions allow for an alternative probability definition. Unlike the conventional probability, our definition accounts for directional variation in the likelihood of intersecting objects from different directions. While the new probability does not result in improved practical tracing performance, we are able to provide an interesting insight on the conventional probability. We show that the conventional probability function is directly linked to our examined probability function and can be interpreted as covertly accounting for directional variation. The path tracing light transport algorithm can require tracing of billions of rays. Thus, it can pay off to construct high quality acceleration structures to reduce the ray tracing cost of each ray. At the same time, the arising number of trace operations offers a tremendous amount of data parallelism. With CPUs moving towards many-core architectures and GPUs becoming more general purpose architectures, path tracing can now be well parallelized on commodity hardware. While parallelization is trivial in theory, properties of real hardware make efficient parallelization difficult, especially when tracing so called incoherent rays. These rays cause execution flow divergence, which reduces efficiency of SIMD-based parallelism and memory read efficiency due to incoherent memory access. We investigate how different BVH and node memory layouts as well as storing the BVH in different memory areas impacts the ray tracing performance of a GPU path tracer. We also optimize the BVH layout using information gathered in a pre-processing pass by applying a number of different BVH reordering techniques. This results in increased ray tracing performance. Our final contribution is in the field of fast high quality BVH and kd-tree construction. Increased quality usually comes at the cost of higher construction time. To reduce construction time several algorithms have been proposed to construct acceleration structures in parallel on GPUs. These are able to perform full rebuilds in realtime for moderate scene sizes if all data completely fits into GPU memory. The sheer amount of data arising from geometric detail used in production rendering makes construction on GPUs, however, infeasible due to GPU memory limitations. Existing out-of-core GPU approaches perform hybrid bottom-up top-down construction which suffers from reduced acceleration structure quality in the critical upper levels of the tree. We present an out-of-core multi-GPU approach for full top-down SAH-based BVH and kd-tree construction, which is designed to work on larger scenes than conventional approaches and yields high quality trees. The algorithm is evaluated for scenes consisting of up to 1 billion triangles and performance scales with an increasing number of GPUs.