A unified triangle/voxel structure for GPUs and its applications

We present a system that combines voxel and polygonal representations into a single octree acceleration structure that can be used for ray tracing. Voxels are well-suited to create good level-of-detail for high-frequency models where polygonal simplifications usually fail due to the complex structure of the model. However, polygonal descriptions provide the higher visual fidelity. In addition, voxel representations often oversample the geometric domain especially for large triangles, whereas a few polygons can be tested for intersection more quickly. We show how to combine the advantages of both into a unified acceleration structure allowing for blending between the different representations. A combination of both representations results in an acceleration structure that compares well in performance in construction and traversal to current state-of-the art acceleration structures. The voxelization and octree construction are performed entirely on the GPU. Since a single or two non-isolated triangles do not generate severe aliasing in the geometric domain when they are projected to a single pixel, we can stop constructing the octree early for nodes that contain a maximum of two triangles, further saving construction time and storage. In addition, intersecting two triangles is cheaper than traversing the octree deeper. We present three different use-cases for our acceleration structure, from LoD for complex models to a view-direction based approach in front of a large display wall