A Sorting Classification of Parallel Rendering

Reviewer: Thomas W. Crockett

In computer graphics, the process of converting an abstract description of a scene into a finished image is known as rendering. Because rendering of complex scenes at interactive rates is computationally demanding, the use of parallelism to speed up the process has received considerable attention. While ample parallelism is available, exploiting it effectively requires algorithms that are carefully matched to both the application requirements and the underlying hardware architecture. Inherent in the rendering process is a projection from three-dimensional object space to two-dimensional image space. Assuming that both the object data and the image buffer have been partitioned across processors, this projection implies a scene- and view-dependent remapping of intermediate results, and hence extensive communication. In this paper, the authors introduce a taxonomy of parallel rendering algorithms based on the point in the rendering process at which communication occurs. The emphasis is on conventional polygon rendering techniques, as opposed to ray-casting or radiosity methods. The authors propose three categories—sort-first, sort-middle, and sort-last—and examine the computational and communication costs of each approach using high-level analytical models. The analysis indicates that none of the methods is clearly superior in all circumstances; rather, the choice of algorithm depends on particular characteristics of the application and the target architecture. While this conclusion may not be very satisfying, it reflects the complex tradeoffs that occur in practice. Although the taxonomy is not comprehensive, it does provide important landmarks for navigating through the large algorithmic design spaces that characterize parallel rendering. The accompanying analysis and discussion offer valuable insights into the design of effective parallel rendering systems. This paper should be regarded as mandatory reading for anyone who wants to be well-versed in parallel rendering.