Parallel Machines for Computer Graphics

Abstract

Computer graphics provides some ideal applications for the kind of highly parallel implementations made possible by advances in integrated circuit technology. Specifically, hidden line and hidden surface algorithms, while easily defined and simple in concept, entail a substantial amount of computation. This requirement fits the characteristics of integrated circuit technology, where modular designs involving regular communication between many concurrent operations are rewarded with high performance at an acceptable cost. Ray tracing is a very flexible technique that can be used to produce some of the most realistic of all computer generated images by simulating the interactions of light rays with surfaces in a modeled scene. Because light rays are mutually independent, many may be processed simultaneously, and the potential for concurrency is great. One architecture for expediting a ray tracing algorithm consists of a conventional computer equipped with a special purpose peripheral device for locating the intersections of ray8 and surfaces. This intersection computation is the most time consuming aspect of a ray tracing algorithm. Although the attached processor configuration can produce images more quickly than an unaided computer, its per- formance is limited. Alternatively, a pipeline of surface processors can replace the peripheral device. Each processor computes the intersections of its stored surface with rays that flow through the pipe. Such a machine machine can be quite fast, and its performance can be increased by lengthening the pipeline, but the component processors are not very effectively utilized. A third approach combines the advantages of the prior two machines by using an array of processors, each simulating a distinct subvolume of the modeled world by treating light rays traveling through space as messages flowing between processors. Local communication is sufficient because light rays travel continuously through space. In real time computer graphics, successive images must be produced in times that are imperceptible to a viewer. Although the ray tracing machines fall short of this performance, it is possible to compromise image quality in order to produce a highly parallel machine capable of real time operation. The processors in such a machine are organized to form a binary tree. Leaf processors scan-convert surfaces, producing a sequence of segments, where a segment is the portion of a surface that appears on a single scan line of the display. Processors towards the root of the tree accept two such segment sequences and produce a third in which all segment overlap has been resolved. The final image is available at the root of the tree. The communication bottleneck that would otherwise occur at the root can be eliminated by breaking out parallel roots, and the resulting tree may be extended to scenes of almost arbitrary complexity merely by increasing the supply of available processors. Massive parallelism can also be applied to the problem of removing hidden edges from line drawings. A suitable architecture takes the form of a pipeline in which each processor is dedicated to the handling of a single polygon edge. These processors successively clip line segments passing through the pipeline to eliminate portions hidden behind surfaces. Each edge processor can be constructed out of little more than three serial multipliers. The machines described here are varied in organization, and each functions differently, but their treatment of sorting is one ingredient common to all. Sorting is a key component of hidden surface algorithms running on conventional computers, but its extensive communication requirements make it costly for use in a highly integrated design. Consequently, the highly parallel machines described here operate largely without sorting. Instead, they maintain information in sorted order or make use of already sorted information to limit communication requirements.

Files

Additional details