A parallel-pipelined multiprocessor system for the radiosity method

Abstract

Ray-tracing and radiosity algorithms can produce very realistic images, but they require a lot of computations which make them impractical for scenes of high complexity. Several attempts have been made to speed up computations through parallel processing. To get orders of magnitude speedup, massive parallelism involving multiple streams will be necessary. In this paper, a parallel-pipelined multiprocessor system is described, which is made of clusters of specialized computing modules, each constructed of an Intersection Computation Unit (ICU) and a number of Cell Traversal Units (CTUs). Both ICU and CTU are of type pipeline and with data-driven execution. A pseudodynamic scheduling is used to reconfigure the system at run time so that the workloads distributed over clusters can be more or less balanced. Furthermore, a hierarchical memory structure is proposed to reduce the average loading time of patches. Performance evaluation has been done and 15% more speedup can be obtained as observed by queueing network simulation. A complete system level simulation is under way by using BONeS which is a block oriented network simulator.