Parallel processing for progressive refinement radiosity

Abstract

Progressive refinement radiosity is an increasingly popular method for realistic image synthesis of non-existing environments. The method successfully approximates the light distribution in an environment, however it requires excessive amount of computation. In this thesis, the progressive refinement method is investigated for parallelization on ring and hypercube-connected multicomputers. Two different approaches for parallelization, based on synchronous parallelism witli static task assignment, are proposed, in order to achieve better coherence in parallel light distributions and obtain good performance on simple topologies. Efficient global circulation schemes are proposed in order to decrease the total volume of communication by asymptotical factors. The first scheme for parallelization is a modification of the sequential algorithm in that several patches shoot their energy at a time, while the second scheme is based on the parallelism level of one shooting patch at a time. The proposed parallel algorithms are evaluated theoretically and implemented for ring and hypercube-connected topologies on Intel’s iPSC72 multicomputer. Load balance quality of the proposed schemes are evaluated experimentally