Optimizing Ray-Triangle Intersection via Automated Search

In this paper, we examine existing direct 3D ray-triangle intersection tests (i.e., those that do not first do a ray-plane test followed by a 2D test) for ray tracing triangles and show how the majority of them are mathematically equivalent. We then use these equivalencies to attempt faster intersection tests for single rays, ray packets with common origins, and general ray packets. We use two approaches, the first of which counts operations, and the second of which uses benchmarking on various processors as the fitness function of an optimization procedure. Finally, the operation-counting method is used to further optimize the code produced via the fitness function