A data flow multiprocessor ray tracing model
Ardeshir Pebdani · 1993
Photorealistic computer generated pictures, produced by ray tracing algorithms are by far the most realistic images. However the cost for this realism is at the expense of one of the most intensive calculations known today, which may take minutes to several days of computations. The core and big percentage of this costly execution time is in the ray-object intersection tests and calculation of the intersection points. Bounding volumes and spatial subdivision are among the more well known software or algorithmic strategies that improve the performance of this algorithm on a single processor configuration. Only few hardware approaches, multiprocessor configurations, have been employed for these algorithms. In these parallel processing techniques, scenes and rays are distributed among the processors so that they can be traced individually. The parallelism in these hardware acceleration stops at the distribution level and does not extend to a lower level so that the intersections instructions may be optimized within a processor. In this dissertation a pipelined data flow multiprocessor model is introduced in order to present a combined hardware and software ray-object intersection acceleration technique in ray tracing. This multiprocessor is a square array of N x N pipelined data flow processing elements. While selection of square array arrangement increases the global speed ups at the higher level, by assigning uniform spatial subdivisions of the scene to be traced to different processing elements, selecting pipelined data flow processor as the processing elements improves the local speed ups, by enhancing the ray-object intersection calculations at lower or instruction level. A complete functional description and theory of operation for this multiprocessor is given. Based on these descriptions and operations, a series of simulations have been carried out on a set of randomly populated scenes to evaluate the overall performance. The performance analysis and simulations on this system has confirmed its suitability for complex scenes in ray-tracing algorithms. At the end a set of enhancements are suggested to adapt the simulations for variety of multiprocessor array topologies and applications.