Please note: this technical report has been superseded by a newer one, available here.

Razor: An architecture for dynamic multiresolution ray tracing

Gordon Stoll, William R. Mark, Peter Djeu, Rui Wang, and Ikrima Elhassan

The University of Texas at Austin, Department of Computer Sciences.
Technical Report TR-06-21.

April 26, 2006


Rendering systems organized around the ray tracing visibility algorithm provide a powerful and general tool for generating realistic images. These systems are being rapidly adopted for offline rendering tasks, and there is increasing interest in utilizing ray tracing for interactive rendering as well. Unfortunately, standard ray tracing systems suffer from several fundamental problems that limit their flexibility and performance, and until these issues are addressed ray tracing will have no hope of replacing Z-buffer systems for most interactive graphics applications.

To realize the full potential of ray tracing, it is necessary to use variants such as distribution ray tracing and path tracing that can compute compelling visual effects: soft shadows, glossy reflections, ambient occlusion, and many others. Unfortunately, current distribution ray tracing systems are fundamentally inefficient. They have high overhead for rendering dynamic scenes, use excessively detailed geometry for secondary rays, perform redundant computations for shading and secondary rays, and have irregular data access and computation patterns that are a poor match for cost-effective hardware.

We describe Razor, a new software architecture for a distribution ray tracer that addresses these issues. Razor supports watertight multiresolution geometry using a novel interpolation technique and a multiresolution kD-tree acceleration structure built on-demand each frame from a tightly integrated application scene graph. This dramatically reduces the cost of supporting dynamic scenes and improves data access and computation patterns for secondary rays. The architecture also decouples shading computations from visibility computations using a two-phase shading scheme. It uses existing best-practice techniques including bundling rays into SIMD packets for efficient computation and memory access. We present an experimental system that implements these techniques, although not in real time. We present results from this system demonstrating the effectiveness of its software architecture and algorithms.