Henrik Wann Jensen
We present a technique for improving the speed of multi-pass GPU computations by using adaptive refinement. We tile the screen and use occlusion queries to adaptively cull inactive parts of the computation. An implementation of this technique in a photon map renderer and a Mandelbrot fractal has resulted in speedups of up to one order of magnitude. Our technique is applicable to many of the recently developed multi-pass algorithms running on GPUs. It is easy to implement and often provides significant speedups by exploiting computational similarity, coherence, and locality.