Done by TPathTracer class in RayTracer unit.

Surface lights are used: every shape with non-zero emissiveColor is considered a light emitter. For each image pixel many random paths are checked and final pixel color is the average color gathered from all paths.

Path length is determined by a given minimal path length and a Russian-roulette parameter. Every path will have at least the specified minimal length, and then Russian-roulette will be used to terminate the path. E.g. if you set minimal path length to 3 and Russian-roulette parameter to 0.5 then 1/2 of all paths will have length 3, 1/4 of all paths will have length 4, 1/8 of all paths will have length 5 etc.

Russian-roulette makes sure that the result is unbiased, i.e. the expected value is the correct result (the perfect beautiful realistic image). However, Russian-roulette introduces also a large variance, visible as a noise on the image. That's why forcing some minimal path length helps. Sensible values for minimal path length are around 1 or 2. Of course, the more the better, but it will also slow down the rendering. You can set minimal length to 0, then Russian-roulette will always be used to decide about path termination (expect a lot of noise on the image!).

Actually our path-tracer does something more than a normal path-tracer should: for every pixel it checks PrimarySamplesCount of primary rays, and then each primary ray that hits something splits into NonPrimarySamplesCount. So in total we check PrimarySamplesCount * NonPrimarySamplesCount paths. This optimization comes from the fact that there is no need to take many PrimarySamplesCount, because all primary rays hit more-or-less the same thing, since they have very similar direction.

To get really nice results path-tracer requires a different materials description. I added a couple of additional fields to Material node to describe physical material properties (for Phong's BRDF). If these fields are not specified in Material node, path-tracer tries to calculate them from normal material properties, although this may result in a poor-looking materials. There's also a program kambi_mgf2inv available that let's you convert MGF files to VRML 1.0 generating correct values for these additional Material fields.

Shadow cache is used, this makes path-tracer a little faster. Also you can generate the image pixels in more intelligent order than just line-by-line: you can use Hilbert or Peano space-filling curves. In combination with shadow cache this can make path-tracing faster (shadow cache should hit more often). Although in practice space-filling curves don't make any noticeable speed difference. Undoubtedly, there are many possibilities how to improve the speed of our path-tracer, and maybe one day space-filling curves will come to a real use.