Bounding Volume Hierarchies and Iterated Function Systems

2010, Dr. Lawlor, CS 481/681, CS, UAF

Basically, every raytracer looks like this code:
	for (each pixel) {
for (each ray bounce) {
for (each chunk of geometry) {
if (ray from pixel hits geometry)
shade pixel with geometry
The performance is clearly going to be O(pixels*bounces*geometry).  The number of pixels is in the millions, the amount of geometry is in the thousands, although luckily we can get away with only a few bounces (on average).  Searching thousands of chunks of geometry for each of the millions of pixels per frame is not going to be efficient.

One simple way to improve performance above is to reduce the number of pixels--that is, reduce the screen resolution.  Another simple trick is to reduce the geometric complexity of the scene.

But one surprising way to reduce the average geometric complexity per ray without changing the final image is to use a "bounding volume hierarchy".  The basic idea behind bounding volumes is before digging into a complicated block of geometry, we first intersect our ray with a simple bounding volume like a sphere.  If the ray misses the sphere, then we can skip the entire block of geometry.  Because misses are common (a ray can only hit one object in the end!), this can be a huge savings in complex scenes.  You can even have a nested tree of bounding volumes, and raytracing then boils down to something like a binary search: does the ray hit anything on the left?  No, how about the right?  Yes!  Then is it the left-right, or the right-right?  And so on.

Common bounding volumes include:
In pseudocode, you traverse the bounding volume tree to find all the geometry that intersects a ray like this:
ray_intersection intersect(ray r,geometry_tree_node g) {
ray_intersection ret=miss;
for (c in child geometry of g) {
if (r hits c's bounding volume)
return ret;
Of course, "intersect" is a recursive function, and recursion isn't supported on the GPU yet (no hardware stack).  Instead of recursing, you can:
Building a bounding volume hierarchy requires the scene geometry to be pre-processed into a search tree.  But we can see much of the same behavior with "recursive geometry" such as an iterated function system.

Iterated Function Systems: Recursively Defined Geometry

An iterated function system (IFS) is a set of "map" functions that transform a parent bounding volume into a set of child bounding volumes.  You can draw an IFS by recursively applying the map functions, which gives smaller and smaller bounding volumes.  For example, here's a Sierpinski Triangle, which is defined as three smaller Sierpinski triangles:
fractal triangle

You can read much more about iterated function systems online, for example the Wikipedia Sierpinksi page.

I wrote a little 2D IFS manipulation program back in 2003.