Out-of-core Global Illumination
I'm currently working on a system which performs data
management for out-of-core, hybrid computation. Out-of-core
means that the data required for the computation is too
large to fit in system memory, often requiring a large
number of slow disk reads to accomplish the task of
accessing data. By hybrid I mean that we aren't restricted
to CPUs for our computation, but can use CPUs, GPUs, and
theoretically Cell processors and custom ASICs.
This has been applied to rendering synthetic images via path
tracing. Path tracing tends to be expensive, and has
difficult access patterns for out-of-core algorithm design,
however, with the use of our data management system, and
some additional intelligent design choices, we now have a
highly efficient out-of-core path tracer which can utilize
the powerful computational resources offered by modern GPUs,
as well as multiple multi-core CPUs.
Improved Convergence of Caustics
Caustics tend to be a computational burden when computing
realistic images via unbiased means. Several methods have
made strides in computational efficiency for diffusive types
of materials, however, caustics are generally calculated via
a biased method such as photon mapping because of the
difference in render speed.
We split the integral of the rendering equation, and utilize
previous methods for computing the "diffusive component",
while applying our new method to computing the "specular
component".
Real-time Ray Tracing on GPUs
I have implemented a real-time ray tracer running on NVIDIA
GPUs which achieves over 50 frames/second for small
scenes like the Stanford Bunny, while reaching nearly 12
frames/second on a seven million triangle scene, the XYZ_RGB
Dragon, also from Stanford's scanning repository.
Better RBSP Trees
We are also working on improving restricted BSP trees, both
from a build and a render aspect. Our build shows asymptotic
improvement, and in practical terms is hundreds of times
faster than previous work. Our render times are also an
order of magnitude faster than previous work.