Bringing ray tracing to games has been a hot topic of
conversation lately. I know that I've had several conversations
over the last few years with co-workers about when the first ray traced games
would appear. The idea of using ray tracing isn't original or really
unique as we use rays for all kinds of things in modern games. What
I'm talking about here though is specifically a renderer that uses a typical
ray tracing algorithm to generate images instead of a rasterization based
"traditional" renderer. I thought it might be
interesting to think about some of the implications of this, the results of
which you see in front of you.
There are a lot of interesting things that using a ray
tracer allows. One of the biggest advantages in my opinion should
come from the authoring pipeline. Current raster based
engines are very tweaky when it comes to certain things. For
instance having different types of shaders work together can be tricky when
transparency, refraction, reflections and shadows are
involved. A ray tracer would allow reflections,
refractions, order independent transparency, volumetrics, better shadowing and
implicit and procedural surfaces. It would also allow all of these
things to interact together pretty much seamlessly. This will
really make it easier and more creative to build virtual environments and
starts removing some of the limitations that make content creation more
difficult. This is probably the best improvement that this kind of
technique offers us although there are other interesting benefits as
well.
One of the funnier objections that I've heard to the idea is
that ray tracing doesn't solve the global illumination problem. Well of
course it doesn't! But does rasterization actually do this
any better? Global illumination is an extremely hairy problem
that is probably going to require a real breakthrough to solve
effectively
(unless we crush it with pure compute horsepower at some
point). Ray tracing does however give you an excellent
playground to deal with global illumination issues. It simply
requires the kind of data structures that also lend themselves to
helping to
accelerate energy transfer (more on this later).
Of course the big elephant in the room that I haven't
mentioned is performance. At the end of the day ray tracing has to be
at
least somewhat competitive with rasterization. I would argue
that there are several things that make this not as important as it
seems in
the longer term. First of all we are starting to top out on
resolution. Yes this will slowly increase but we are at
the point now where more resolution per inch isn't going to terribly
helpful. TV is going to stick with 1920×1080 for a long time if the
past
is any indication. This implies that we might want to start scaling up
in image quality and possibly even geometric complexity where ray
tracing has
an edge. Isn't it time to start spending cycles on something more
interesting than resolution? In addition ray tracing opens up a
whole new set of techniques like implicit surfaces and certain types of
procedural geometry that just get hackier and hackier in a
rasterization
setup. One other thing to note is that most of the time spent
in a rasterizer these days is in the shading of the individual pixels.
The ray tracer won't have any advantage here but this implies that you
could
use one of several techniques to figure out what to shade if all of
them are
only a fraction of the shading time they might end up being very
similar in
performance. This will only become more true as shaders get more
complex and we want more complex interactions between them. Overall I
think it's clear that ray tracing is within spitting distance of being
fast
enough to use.
Earlier I mentioned that ray tracing requires the
same kind of data structures needed for global illumination
computations. Sound energy can also be
calculated using these data structures. Sound is one of the areas
that games could be vastly improved. Using the same hardware and
data structures to calculate a much more realistic sound-scape seems
like a
real tangible benefit that could make a game stand out.
A set of hybrid techniques using elements of both ray tracing
and rasterization has been tossed around by quite a few people. This
has a lot of drawbacks in my opinion although I could see it being used
in
certain cases. The general idea here is to fill in your initial frame
buffer using rasterization and then generate secondary rays during the
shading
of that first pass. I'm sure there are other ideas as
well. The main drawback is that now your scaling is limited by what
you can rasterize instead of what you can ray trace. You also need
to maintain two sets of data structures. One is some sort of scene
used for ray tracing and the other is some sort of scene that you can
feed into
the rasterizer. You could probably share things like vertex buffers
and textures easily. However, you would still need to write and
maintain two separate code bases to traverse and render these data
structures. You would also be eliminating certain classes of ray
tracing techniques like
implicit surfaces. Some people may use techniques like this to
partially start the transition to new techniques or in limited cases
(e.g.
using ray tracing for volumetric fog in part of the scene).
I believe that building worlds procedurally is going to be
commonplace in the coming years (I will write down some thoughts on this in the future). Modern ray tracing algorithms seem
to map well to the upcoming generation of GPU's that are extremely
programmable. Taking advantage of the natural coherence in scenes vastly
reduces the memory bandwidth requirements. Given the advances in
tool sets, the advance of hardware speed and parallel processors and level of
innovation we consistently see in games I think it will be feasible and
profitable to ship a ray traced game in the near future.
