I'd never written a ray tracer before, so the task description taught me a lot. And I came up with what I thought were some pretty clever tricks---but I'm sure they aren't original.
I decided to write this in Perl since that's the language I can prototype code in the fastest. I also decided to cut some corners; there was no way I could win given the time constraints I had (I have a job!) and that I was a one-man team and that I had never done anything in ray tracing before. So I didn't worry about error checking the input, and I didn't implement the strong typing (floats vs ints). These parts just weren't interesting to me.
Implementing the language was trivial; I simply translated the GML into Perl. Since Perl supports closures, lexical variables, and first-class functions, this was very easy, and it provided some speed since Perl is quite fast as an interpreter.
Implementing the renderer was more difficult. I am very familiar with affine transformations in general (from PostScript), but not with ray tracing or how angles get distorted, etc. So I ended up reinventing a lot of stuff, like how to map a normal. (I knew how to map a point and a vector, but was delighted with how elegant mapping a normal turns out to be; use the transpose of the inverse matrix!)
One neat trick I came up with was dealing with inverse transformations. I did not want to compute matrix inverses, even of 4x4 matrices, because it's difficult and ugly code. So instead, I computed both the forward and inverse matrix on each transformation step, and kept both with the transformed objects!
Surface acne just wasn't a problem at all because I only considered surfaces that were facing the proper way. No need for kludges like epsilons; just make sure the ray actually bounces off the visible surface of the object rather than the underside. You won't find any numerical kludges in my code.
Because my renderer was so slow (being in Perl and all), I made it generate intermediate files at 1/8th, 1/4th, and 1/2 the resolution so I could check on its progress and correctness without waiting for the full resolution.
Of course, I did the usual sphere bounding test on all objects, but other than that, didn't really do anything for performance. Because of Perl, I'm already about 100x slower than a compiled language so there's little hope of recovering sufficient speed.
At entry time, I thought I had a complete if slow Tier 2 entry (if you ignore the fact that I don't do strong typing); that is, each valid Tier 2 image should render correctly, but ones that cause runtime errors would not necessarily throw an error. Well I was wrong. There were at least five errors in my program:
All in all, I tremendously enjoyed this contest and I applaud everyone who entered and especially the judges!
I can be contacted at rokicki@cs.stanford.edu.