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

34 Chapter 3: A Brief Survey of Ray Tracing Acceleration Methods
3.4 Summary
In this chapter we have given a brief overview over the different methods and
techniques for accelerating ray tracing that have been proposed over the last
two decades. As mentioned in the beginning, this list of techniques is not
complete, but should contain at least the most famous and most commonly
used techniques.
Among all these techniques, hierarchical subdivision methods are the
most important, as they allow for reducing the computational complexity
(from O(N) to O(log N) [Havran01]), whereas all other methods, roughly
speaking, only improve “the constants” of the ray tracing algorithm. As
such, any non-trivial ray tracing system today uses one or another form of
the previously discussed scene subdivision and traversal techniques.
While all of these acceleration methods are commonly agreed to have the
same computational complexity of O(NlogN), their respective performance
in practive varies from case to case, depending on scene, implementation,
hardware platform, and application (i.e. ray distribution). As such, it is
not possible to name one single method that is always best, though practice
has shown that kd-trees usually perform at least comparable to any
other technique [Havran01, Havran00]. Though much of the original ray
tracing literature covers Octrees, in practive today the most commonly used
techniques are either uniform grids or kd-trees. Grids (especially uniform
ones) are advantageous because of their simplicity. Because of this, they are
also well suited for implementation on hardware architectures with restricted
propgramming model (such as e.g. GPUs) 15 ), whereas kd-trees usually adapt
better to varying scene complexity, and often achieve superior performance
if used correctly.
Except for faster intersection tests and hierarchical subdivision, most of
the other previously discussed techniques (such as reducing the number of
rays in one or another form) are limited to special cases and to a restricted
set of applications. As such they are of limited use for a ray tracing system
that is to be used for as many applications as possible, and will thus not
receive any further attention in the remainder of this thesis. Even so, many
of those are orthogonal to the techniques used by RTRT/OpenRT, and can
still be combined with the RTRT/OpenRT system to achieve even higher
performance for those applications in which they are effective.
In order to be as wide a class of applications as possible the RTRT core
has concentrated mostly on those topics that influence the performance of
15 Uniform grids do not need to maintain a recursion stack, and can be implemented
with few floating point operations