The tree is built by combining all of the objects into one bounding box. The objects are then sorted by their center value (either by X,Y, or Z coordinate). Then half of the objects are placed into a new bounding box and the other half are put into a seperate bounding box. This division process continues until each object is in its own box. The critera to select whether to sort in X,Y, or Z is determined by minimizing surface area of the resulting bounding box.

The following scenes were used to test this structure:
N spheres were created, with the center of the sphere as <drand48()*2-1,drand48()*2-1,drand48()*2-1> and a radius of 0.1/(N^0.3333). The spheres were colored based on the absolute value of their position.

The eye was located at <0,0,2> and was looking at <0,0,0> with a fov of 90 degrees. The scene was rendered at a resolution of 512x512 with one ray being fired per pixel. Notice that there were no shadow rays cast, although a simple phong shading model was applied to the spheres.

Click here to view the scene when N=1, N=10, N=100, N=1,000, N=10,000, N=100,000, and N=1,000,000.

Timings were performed on an SGI R10K ("Raptor") and on an UltraSparc 30.

Notice that as scene complexity increased there was a much more significant overhead in creating the tree structure.
Compilition was performed with the native CC compiler on each platform with the optimization flags -Ofast and -fast respectively.

(All times are in seconds)