Ray tracing again

I recently started working on a new ray tracer. I've started from scratch redesigning everything to create a core that's both efficient and flexible enough to allow me to experiment with various algorithms and techniques. I created this blog to act as a journal on my progress, as well as a scratch-pad for random thoughts and ideas. This is also my first attempt at blogging, so if it shines through bear with me :)

The status quo is a ray tracer which supports the obj and ply model formats. I've already done some ray tracing and simple path tracing with it, so here are some obligatory pictures:

The acceleration structures I have implemented currently are a uniform grid and the bounding interval hierarchy. I was really amazed by how simple the latter was to implement, while still being both efficient and flexible. The in-place quick sort-like construction algorithm with the global heuristic is especially elegant. I'm looking forward to implementing a kd-tree to see how the BIH compares with it at ray traversal.

As I mentioned I have implemented a simple, bare-bone unbiased path tracer already (very simple once you have the ray-model intersection code). However, it's quite noisy and requires an intimidating amount of samples to get a decent result. Eventually I'm planning to implement bidirectional path tracing and metropolis light transport to try to mitigate that. But first I want to optimize the acceleration structures so that they can act as a core for both stochastic and real-time ray tracing. This will probably involve implementing and SIMD-izing both single-ray and coherent traversal routines.

Other stuff I want to experiment with: terrain, volumetric ray tracing, sub-surface scattering, etc... I'll see where things lead and how long until interest fades :)

Instant Ray Tracing: The Bounding Interval Hierarchy
The Utah 3D Animation Repository
The Stanford 3D Scanning Repository


Targh said...

Hi capisce!!!

Kevin said...

like the fairy image :) how long did that take?

capisce said...

kevin: that was done doing pure brute force path tracing with no direct light samples, so with 25000 samples per pixel it took 74 hours :)

the bih and traversal has been optimized a bit since I rendered that image though...