Efficient Radiosity Simulation using Perceptual Metrics and Parallel Processing

Simon Gibson.
Ph.D. Thesis, University of Manchester, September 1998.

Abstract

The generation of realistically illuminated synthetic images has been at the forefront of computer graphics research since the early 1970s. The core problem that must be addressed in order to achieve this goal is the accurate simulation of light transport throughout an environment. As compute resources have developed, rendering techniques have also progressed from basic local illumination models to physical simulations that consider not only the local interaction of light with a surface, but also important global effects such as shadows and inter-reflections.

The radiosity method is one of the most important global illumination algorithms, and the generation of geometrically complex, view-independent radiosity solutions has received much attention in recent years. This thesis provides two contributions that enable such radiosity solutions to be generated quickly and efficiently.

Firstly, a new framework for radiosity simulation (based on the progressive refinement method) is presented that uses perceptually-based calculations to control the simulation process. It is shown how such an approach enables computational effort to be focussed into those areas that have a visually significant effect on the appearance of the solution, leading to considerable savings in computation time.

Secondly, a novel algorithm for computing radiosity solutions on distributed shared-memory parallel architectures is described. Our approach is based on the notion of a {\it sub-iteration\/} being the transfer of energy from a single source to a subset of the scene's receiver patches. By using an efficient queue-based scheduling system to process these sub-iterations, we show how radiosity solutions can be generated without the need for processor synchronisation between iterations of the progressive refinement algorithm. This free-running method of computation exhibits excellent data locality, thereby limiting the amount of inter-processor communication that occurs, and eliminates the need for load-balancing until the end of the solution process. Results are presented indicating an almost linear speedup in the time required to generate a solution.

Awards: Best thesis, Department of Computer Science Research Student Symposium, University of Manchester, 1998-1999.