Fast GPGPU Based Quadtree Construction

We introduce a method for fast quadtree construction on the Graphics Processing Unit (GPU) using a level-by-level approach to quadtree construction. The algorithm is designed to build each subsequent level from the parent nodes of the previous level, and is thus suitable for parallelization. Our work is motivated by the use of General Purpose GPU (GPGPU) techniques for large data sets generally, and for the use of quadtrees for spatial segmentation of lidar data points for grid digital elevation models (DEM) in particular. We introduce an algorithm suitable for quadtree construction on the GPU which reduces the construction problem to bucket sort. We then describe possible implementations and refinements to the algorithm: utilization of multiple threads on a single quadtree node, null-node pruning, and a hybrid CPU-GPU approach for extending our solution past the limits of GPU memory. We find that our fully implemented algorithm outperforms a CPU approach by a factor of between 5×−12× for sufficiently large datasets.