GPU-based Ray Tracing of Dynamic Scenes

This paper presents the design and implementation of a GPU-based ray tracing system for dynamic scenes consisting of a set of individual, non-deformable objects. The triangles of each object are organized in a separate Kd-tree. A bounding volume hierarchy (BVH) is built on top of these Kd-trees. The BVH is updated and uploaded into GPU memory on a frame-by-frame basis, whereas the Kd-trees are uploaded only once. We use a single ray tracing kernel for handling all ray generations. Code execution path divergence is limited by the use of a ray stack, which treats all ray types in the same way. We make effective use of the very limited high bandwidth memory of the GPU’s multiprocessors by using a smart stack for the acceleration structure traversals. The results show that our GPU implementation of a two-level acceleration approach performs between 40 and 105 percent as fast as a single Kd-tree containing the entire scene.