Parallel construction and update of acceleration structures for ray tracing

Abstract

Image synthesis is the process of generating an image from a representation of a virtual scene. Ray tracing is used in a number of image synthesis techniques in order to determine the visibility of geometric primitives from the point of view of an observer, and requires considerable amounts of computational resources. Specific data structures are meant to accelerate this process by taking advantage of the spatial coherence of objects within a scene. In the case of scenes which do not exhibit change, acceleration structures are constructed once and used throughout. However, with deformable objects, these data structures can become invalid over time. In this work we have explored ways of constructing and updating acceleration data structures of scenes which exhibit change, by taking advantage of multiple processing elements and a shared memory architecture. We have introduced a novel acceleration data structure which we term the dual-acceleration structure and two metrics for evaluating the degradation of specific acceleration structures over time. Results show that our method performs better than the straightforward approach, especially on large datasets.