Project
Pitch
I propose to develop a WebGL based real
time volumetric renderer and (if time permits) use it to render physically
based smoke simulation. Volume rendering is a well-known technique used for
rendering data that is typically stored in the form of a grid. Some problems
can be solved very efficiently and are naturally suited to be solved when space
is portioned into an imaginary uniform grid structure. Some common examples are
rendering clouds, smoke, fire, etc.
Generating images based on this method has
been traditionally done in an offline manner, i.e. images are not generated as
soon as data is available; putting it more precisely, images cannot be
generated at 30 fps or greater. However, volumetric rendering technique is
“embarrassingly parallel”, which means that the technique can be parallelized
with little effort. The first step in the process is to generate a ray for each
pixel of the image that we want to generate. Then we can gather densities along
each ray independently on separate fragment shaders. The process can be further
optimized if we only ray march within the region containing voxels.
I would be using 3D textures for storing the
voxel data if it is supported in the latest WebGL specification. If it is not,
then I would have to use 2D textures and simulate 3D textures.
The main challenge in the project is using
WebGL for doing all of the above. Being a graphics guy, my work has mainly been
in C++ and C++ like languages. It’s going to be a steep learning curve for me
to use JavaScript for doing graphics.
Once I am done with the rendering, I would
start working on creating physically based smoke simulation. I propose to use
the Semi-Lagrangian method for smoke simulation, meaning I would be using parts
of both a particle based (Lagrangian) approach and grid based (Eulerian)
approach. The Eulerian approach enables us to parallelize smoke simulation
conveniently on the GPU. Every grid element stores both scalars (like density, pressure,
temperature) and vectors (like velocity) which get modified during the
simulation. The densities computed in each step of the simulation would then be
transferred to the volumetric renderer for visualization.
I would be referring to Chapter 30 of the
book GPU Gems 3 for my implementation:
