Realtime Ray Tracing and Interactive Global Illumination - Scientific ...

32 Chapter 3: A Brief Survey of Ray Tracing Acceleration Methods
Different Traversal Algorithms: Though a different traversal algorithm
should still traverse exactly the same voxels (and thus intersect the same
primitives), the traversal algorithm often consumes a significant fraction of
compute time 12 . As such, a different traversal algorithm can significantly
affect performance.
Today, there exist several different traversal algorithms for each of the different
kind of acceleration structure, each of which have different advantages
and disadvantages.
While most of these alternative traversal algorithms do not actually change
the underlying data structure, some methods require to slightly modify the
data structure itself, e.g. by adding neighbour links [Samet89, Havran98] for
faster “horizontal” traversal from one voxel to its neighbour.
Variants of Building the Hierarchy: Apart from the traversal algorithm,
the eventual performance of an acceleration structure is significantly affected
by the algorithm for constructing the data structure. As the traversal algorithm
itself usually does not change either the number nor order of voxels
(and thus primitives) encountered by the ray, the number of traversal steps
and primitive intersections is mostly determined by the way that the acceleration
data structure has been built (i.e. by the number, size, and organization
of the voxels).
Even for uniform subdivision (whose structure is determined entirely by
the grid resolution, there usually is a tradeoff between reducing the number of
primitive intersections and performing too many traversal steps. In practice
the optimal grid resolution is hard to determine, and can only be guessed
approximately using some heuristics. For hierarchical data structures, finding
“the best” organization is even more complex 13 .
Though there are several groundbreaking papers on the optimal construction
of ray tracing hierarchies (e.g. by Goldsmith and Salmon [Goldsmith87],
Subramanian [Subramanian90b, Subramanian90a, Cassen95], or MacDonald
and Booth [MacDonald89, MacDonald90]; also see [Havran01] for an
overview), the importance of using these techniques is often underestimated
in practice. Even the RTRT/OpenRT system at its original publication only
considered a naive kd-tree construction algorithm. Havind added a Surface
Area Heuristic [Havran01] for building the kd-tree afterwards since then has
roughly doubled performance (see Section 7.3).
12 In practice (i.e. for realistically complex scenes and relatively simple primitives),
traversal is usually several times as costly as the ray-primitive intersections.
13 For hierarchical subdivision, the decision whether to continue subdivion or not, as well
as the exact position of the split have to be performed anew in each construction step.