Final Gathering on GPU

1
Final Gathering on GPU
Demo application is available at
http//www.bee-www.com/parthenon/
Toshiya Hachisuka (toshiya_at_bee-www.com) The
University of Tokyo
Introduction
However, Final Gathering typically needs
thousands of samplings to achieve noiseless
result. For this reason, Final Gathering casts
many more rays compared to other rendering
process (i.e. photon tracing) using ray-casting
and it often becomes a bottleneck of all
rendering process. Irradiance Cache applies Final
Gathering to sparsely sampled pixels and
interpolates this data to produce entire
illumination data. Irradiance Cache can greatly
reduce number of sampling rays, but its parameter
is not intuitive (e.g. a maximum valid distance
current cache points can be used for
interpolation within this distance), so it is
difficult to control rendering time and the error
due to interpolation.
Proposed method uses inverse of Depth Peeling
which stacks each depth layer, whereas original
process strips away each depth layer. Sampling
point along ray direction on depth layer is
projected to screen space pixel during successive
pass. If sampling point on depth layer is behind
a ray origin, it is culled by a KILL operation.
Since common graphics hardware is based on Z
buffer algorithm, parallel projection in Depth
Peeling is straightforward process and it can be
computed quite efficiently. Although Ray tracing
on GPU can be used to cast these rays, its
performance is not sufficient for per-pixel Final
Gathering.
Producing global illumination image without any
noise is still difficult because it requires a
large number of samplings. Some approaches, such
as Photon Mapping Jensen 1996, can render
noiseless image but often results in biased
image. To produce high quality image, most of
current global illumination renderer uses
additional process called Final Gathering.
Final Gathering uses rough global illumination
solution such as Photon Map or Radiosity, and
gathers this data by casting rays to obtain
definitive incident radiance distribution. Since
Final Gathering requires a large number of rays
for producing noise-free result, it is very
costly to process all visible pixels. Irradiance
Gradients Wald and Heckbert 1992 can reduce
number of processing pixels by interpolation, but
its behavior is difficult to control. Instead of
reducing number of processing pixels by
interpolation, I propose an accelerated method of
the process itself by using a common graphics
hardware.
Photon Tracing rays Ray Tracing from Camera rays Final Gathering (1024 samples) rays
Cornell Box (512512, 500k photons) 607,002 262,144 268,435,456
Buddha (512512, 800k photons) 29,886,408 323,212 101,261,312
Teapot (512512, 300k photons) 368,793 379,452 253,227,008
Two-Pass Method
Most of existing global illumination renderer
implements two-pass method. The first pass is
making rough global illumination data, and the
second pass is actual rendering using this
solution. Rough global illumination solution can
be obtained by any method that can solve some
part (or all) of rendering equation. One of
popular and robust method for this first pass is
Photon Mapping. Since the resulting data of a
first pass itself is complete global illumination
solution, one can render a image by directly
visualizing this data. However, first pass result
is only approximated solution and its quality is
not sufficient for a final image, so second pass
is usually applied by using first pass result
indirectly. Whereas first pass result can be used
as a approximated probability density function
for Path-tracing, the most popular second pass
process is separated indirect illumination
computation by Final Gathering.
Successive Pass
Inverse process of Depth Peeling
Theoretical ray number of typical scenes
Result
Note that Final Gathering process casts many more
number of rays compared to other process.
I implemented global illumination renderer using
the proposed method using DirectX 9.0 on
WindowsXP. These images were rendered by this
program on Pentium4 2.8GHz and Radeon 9700 Pro
(Video memory is 128MB). Note that in spite of
processing all pixels, rendering time is still in
practical range.
Proposed Method
Proposed method uses a randomly sampled global
ray direction and regards ray-casting as a
multiple parallel projection. By choosing global
ray direction, many sampling rays of Final
Gathering can be viewed as a number of sets of
rays with same direction and different origin.
Rather than using ray casting, using parallel
projection is suitable for a common graphics
hardware based on Z buffering algorithm. Multiple
parallel projections are effectively done by
Depth Peeling using programmable shader. Since
proposed method simply samples all visible
points, it doesn't have non-intuitive parameters
like Irradiance Cache. Current performance of
proposed method is comparable to efficient
implementation of Final Gathering with Irradiance
Cache on CPU. However, by considering growing
speed of GPU, this method has potential to
outperform the CPU based methods.
Total sec Final Gathering only sec Average number of Depth Peeling iteration
Stanford's (512512, 1024samples) 175 142 3.156
Global illumination test (512512, 2048samples) 92 80 4.235
Performance of sample application using proposed
method
Note that the object in directly visualized image
seems like floating from floor due to
insufficient accuracy of global illumination
solution in Photon Map.
Conceptual diagram of proposed method
Final Gathering
Final Gathering can be done by iterate this
process for random global ray direction.
Conclusion
Ray Cast by Parallel Projection
To get N times reflected illumination, Final
Gathering gathers illumination data which is N-1
times reflected illumination and generates Nth
reflected illumination. Final Gathering process
is essentially the same for casting a large
number rays on upper hemisphere of visible
points. It can produce high quality result form
rough first pass illumination data by reusing
this solution all over the visible pixels.
Final Gathering can be effectively done by using
multiple parallel projection on GPU. Visual
quality is nearly equivalent of per-pixel Final
Gathering by ray casing. Current performance is
comparable to interpolation method (e.g.
Irradiance Cache) but it doesnt have
non-intuitive parameters like maximum distance of
Irradiance Cache. Since proposed method is
per-pixel method, it can be used to deal with any
set of points including ray-traced pixels for
image rendering, vertex data for pre-computed
radiance transfer, textures for light maps.
Whereas original Final Gathering processes a set
of rays with same origin, the proposed method
uses a set of rays with same direction (i.e.
global ray direction). Thinking from this
viewpoint, ray casting is the same as sampling a
point on the nearest depth layer from the ray
origin along global ray direction. Each depth
layer can be extracted by using inverse of Depth
Peeling Everitt 2001.
Reference
Jensen, H. W. 1996. Global illumination using
photon maps. Rendering Techniques 96
(Proceedings of the Seventh EurographicsWorkshop
on Rendering), pages 2130. Everitt, C. 2001.
Interactive order-independent transparency. Techn
ical report, NVIDIA Corporation. Szirmay-Kalos,
L. 1998. Global Ray-bundle Tracing.
Technical Report/TR-186-2-98-18. Ward, G. and
Heckbert, P. 1992. Irradiance gradients. Eurogra
phics Rendering Workshop 92, 85-98.
Ray distribution of Final Gathering
Final Gathering casts many rays on upper
hemisphere to gather the illumination data
obtained by fist pass.
Same color rays are in same set of rays.