A task and data balanced distributed photon mapping method

Abstract

Photon mapping is an important global illumination method of simulating some visual effects, such as caustic. However, photon mapping is difficult to be parallelized efficiently, due to unbalanced task distribution and memory management. In this paper, we present a combined data-parallel and task-parallel approach to improve performance across multiple rendering clients in a distributed system. The shading points (receivers) from pixels are divided into groups. We associate the photons with each receiver group by computing a tree cut of the photon hierarchy, thus both the data and the task are tightly coupled. Furthermore, we organize both the photon hierarchy and the receivers with Morton code for efficient tree construction. In the end, our method balances both the illumination task of receivers and related photons, resulting in better load balancing and higher performance without any quality degradation.