Adaptive acceleration structures in perspective space

Traversal efficiency of ray tracing acceleration structures can be improved by specializing them, each frame, for the rays that are traced in that frame. A companion paper to this one demonstrates that extremely high traversal performance for eye and hard shadow rays can be obtained by transforming rays and geometry with a perspective transform, then using a grid acceleration structure in the perspective space. However, the performance of this perspective grid acceleration structure suffers for off-axis rays such as those used for soft shadows or depth of field. In this paper we address this shortcoming by exploring the use of the perspective transform with adaptive acceleration structures such as kd-trees. The problem of choosing optimal split planes for an acceleration structure is different in perspective space than it is in world space, so we introduce a new cost metric for estimating traversal cost of acceleration structures in perspective space. This metric is related to the traditional surface area metric but defined in perspective space. We implement a ray-tracing system with both traditional and perspective acceleration structures and demonstrate significant performance improvements using the perspective space structures even when build time is included. Additionally we evaluate the effectiveness of our new cost metric compared to traditional cost metrics. A key insight demonstrated by these results is that end-to-end rendering performance can be improved by building a different specialized acceleration structure for each light and camera in the scene instead of using a single non-specialized acceleration structure for all rays.