The scene is compiled into a display list where each patch is given a different color. So a patch can be uniquely indexed using its color value. This is called an indexed model.
The hemicube is placed on a patch which is the current emitter. The hemicube has a resolution equal to the preview window. Form factors for each divison on the hemicube is precalculated and stored. Now a OpenGL camera is placed at the center of the patch in question and made to look along the hemicube axes. Now the indexed model is rendered. Whatever the camera sees (which gets recorded in the framebuffer) is the projection of the target patches onto the hemicube along the axes direction. Adding up the form factors for the pixels in the framebuffer which have the same color gives the form factor for the patch indexed by that color. So GL does nearly all the hard work. All is required is a bit of book-keeping and a few additions. The hemicube geometry looks something like this -
AT cse.iitd.ac.in
