Abstract
Photorealistic image synthesis relies heavily on the concept of ray tracing.
The problem of ray tracing is to find the nearest intersection between a given ray
and scene primitives. Although this problem is geometrically quite simple; in practice,
it is necessary to test millions of rays against millions of scene primitives.
To improve the efficiency of ray tracing, scene primitives are usually arranged
into various acceleration data structures. This thesis focuses on the bounding
volume hierarchy (BVH), which is one of the most popular acceleration data structures
for ray tracing.
The thesis is a compilation of research papers published in journals with impact factor
with the introductory text describing the contributions and putting the individual papers in a common context.
The work consists of five novel methods addressing the BVH for high-performance ray tracing. The first
one is a technique for handling dynamic geometry based on constructing a single BVH taking into account
geometric changes through the animation. The second one is a parallel BVH construction algorithm based
on a combination of the k-means algorithm and agglomerative clustering. The third one is a
parallel BVH construction algorithm based on the progressively refined cut of an existing BVH targeting
multi-core CPUs. The fourth one is a GPU-based algorithm based on locally-ordered clustering building
the BVH in a bottom-up manner by merging a batch of cluster pairs in each iteration. The last one is
a parallel insertion-based optimization for BVHs targeting contemporary GPU architectures. Each of
these methods to a certain extent advances the current state-of-the-art in ray tracing.
