This paper presents an algorithm for the glossy global illumination problem, which runs on the Graphics Processing Unit (GPU). In order to meet the architectural limitations of the GPU, we apply randomization in the iteration scheme. Randomization allows to use that set of the possible light interactions, which can be efficiently computed by the GPU, and makes it unnecessary to read back the result to the CPU. Instead of tessellating the surface geometry, the radiance is stored in texture space, and is updated in each iteration. The visibility problem is solved by hardware shadow mapping after hemicube projection. The shooter of the iteration step is selected by a custom mipmapping scheme, realizing approximate importance sampling. The variance is further reduced by partial analytic integration.
GPU global illumination, stochastic glossy global illumination, Monte Carlo method.