[Michael] Victory!

After a sleep that did not last long enough, I quickly refactored my code to compute and store only the form factors for a single patch at a time (which requires O(n) space as opposed to the old O(n^2)). This slows down computation, since we have to recompute form factors every bounce, but makes large scenes able to not take over my entire hard drive.

I then spent some time fixing bugs and debugging scenes. Note to everyone: Dabrovic Sponza has pretty awful tesselation, don't use it at all possible. But anyway, I changed one of the doors in Dabrovic Sponza to emit lots of light and let my radiosity solver run on a course tesselation for a few dozen seconds (it did direct + some of the first bounce indirect in that time frame), giving me:

I'm going to try on Crytek Sponza; then profile and do the easiest optimizations possible before 3. But I ran radiosity on Sponza and had a soft lighting image in <1 minute; so I win!