Non-Self-Adjacent Ray Classes for Parallelizable Shooting–Bouncing Ray Tracing Double Count Removal

Abstract

We propose and experimentally validate a new ray spawning and associated double count removal (DCR) technique for shooting–bouncing ray tracing (SBR). This technique allows, for the first time, efficient parallelization of ray DCR, the major bottleneck and least parallel aspect of modern SBR ray-tracing relying on the ray-cone approximation (RCA). We define non-self-adjacent (NSA) ray classes on a recursively sampled icosahedron, guaranteeing removal of mutual adjacency data dependencies between rays that previously prevented efficient parallelization of ray DCR and, by extension, SBR. Using a GPU-parallelized implementation of the technique, we demonstrate speedups of DCR over 300×, limited in our testing only by the available hardware. As DCR is the asymptotically dominant contributor to the computation time of SBR-RCA, with respect to the number of parallel processes available, the achieved speedup applies to parallel SBR-RCA as a whole.