Fast and Simple Agglomerative LBVH Construction

Abstract

This paper continues the long-standing tradition of gradually improving the construction speed of spatial acceleration structures using sorted Morton codes. Previous work on this topic forms a clear sequence where each new paper sheds more light on the nature of the problem and improves the hierarchy generation phase in terms of performance, simplicity, parallelism and generality. Previous approaches constructed the tree by firstly generating the hierarchy and then calculating the bounding boxes of each node by using a bottom-up traversal. Continuing the work, we present an improvement by providing a bottom-up method that finds each node's parent while assigning bounding boxes, thus constructing the tree in linear time in a single kernel launch. Also, our method allows clustering the sorted points using an user-defined distance metric function.