Faster incoherent rays: Multi-BVH ray stream tracing

Abstract

High fidelity rendering via ray tracing requires tracing incoherent rays for global illumination and other secondary effects. Recent research show that the performance benefits from fast packet traversal schemes that exploit high coherence are lost when coherency is low due to inefficient use of the CPU's SIMD units. In an effort to solve this problem, methods have been proposed which try to extract the remaining coherency from secondary rays through ray sorting, reordering and streaming. Another category of traversal methods have also been proposed which ignore coherency altogether and use a higher order tree branching factor while tracing single rays at a time. These single ray methods not only target applications with incoherent rays but are also scalable with larger SIMD widths. This paper combines ideas from both categories to form a new traversal method which extracts coherency from a group of rays through simple filtering while still providing a fast single ray traversal in cases where there is no coherency present. This new algorithm does not depend on the use of packets which cleanly decouples traversal from shading and is scalable for larger SIMD widths. Results show that overall performance benefits are obtained on a current generation CPU architecture.