An Analysis of Cache Awareness for Interactive Selective Rendering

Abstract

Interactive high-fidelity rendering is one of the major goals of computer graphics. Algorithms based on ray tracing are usually used to drive high-fidelity renderers. While ray tracing is often thought to be impractical for real time performance, recent algorithmic advances in the field have led to the development of interactive ray tracers which leverage the performance of modern parallel systems and cache awareness to obtain real-time rates for moderately complex scenes using Whitted-style ray tracing. Another method used for accelerating ray tracing is the use of selective rendering where only the pixels that need to be computed are traced and the remainder are computed by other means such as interpolation. The choice of which pixels to render may depend on a number of factors, from simple ones that just compare the radiance values of the rays within a certain area and shoot further rays recursively if the difference in radiance is above a certain threshold, to more complex ones based on visual attention. Selective rendering algorithms may not be perfectly compatible with current interactive ray tracing techniques since selective rendering methods tend to be naturally incoherent, and the computation of rays at a stride may result in expensive cache misses. In this paper we analyse the effect of selective rendering on interactive ray tracing and hint at possible solutions that would allow selective rendering to be compatible with interactive rendering methods to further improve rendering speeds.