A raytracer calculates how a camera would observe a potentially complex scene consisting of numerous objects and light sources. If all the objects are modeled by n primitives, e.g., triangles, the runtime of a software raytracer scales at least logarithmically. This scaling behavior effectively imposes limitations on the scene's complexity, its size, and the raytracer's real-time capabilities. As an alternative, this paper proposes a parallel hardware raytracing machine. A prototypical implementation on a field-programmable gate array, as offered by markets today, validates that this machine achieves rendering in constant time, regardless of both the scene's size and its complexity. I. INTRODUCTIONRendering is a well-established technique to calculate how a camera would observe a scene consisting of several objects and light sources. Rather than solving complex equations exactly, virtually all state-of-the-art rendering machines employ approximation techniques, which is raytracing [1] in most cases. A raytracer, as the name suggest, shoots rays into the scene. In so doing, it traces their progress, and considers secondary rays in case they touch objects or run into light sources.In order to be efficient, a real-time raytracer assumes the following two simplifications: (1) it models all the present objects by a total of n primitives, which are simple triangles in most cases [2], and (2) light sources have the size of a mathematical point. Furthermore, it uses a simple and efficient shading model: in case a ray hits a primitive, this ray spawns at least three secondary rays to model the physical effects called reflection, refraction, and shadow.A software raytracer spends most of its processing time on the calculation of potential intersections of a ray with any of the n primitives. A naive implementation considers all primitives leading to a computational complexity of O(n) per ray. In order to relieve this bottleneck, previous research has developed acceleration data structures, such as regular grids [3], kd-trees [4], and bounding box hierarchies [5], to mention but a few. These data structures provide a criterion that allows the raytracer to stop the testing process. The pertinent literature [6] suggests that these enhancements reduce the rendering time to about Ω(log n) per ray under certain favorable conditions. However, previous research does not provide any performance guarantees or necessary preconditions. That is, it might as well happen that a scene of a certain complexity or size still requires O(n) intersection tests per ray.The time bound Ω(log n) mentioned above is notorious for single-processor software solutions [7], since they can access their memory and perform intersection tests only sequentially. To go beyond these limitations, Section II proposes a new

Under certain conditions, accelerator data structures significantly reduce a raytracer's runtime. But for most of them, unfortunately, the literature does not provide any a priori indication of whether or not they yield any speedup at all in a particular scene. For others though, the literature does provide lower bounds, but those accelerators are of rather theoretical utility. A straight forward analysis indicates that both the sequential processing and memory accesses constitute an inherent bottleneck for any of the software-centered raytracers. By contrast, this paper proposes a parallel hardware architecture, which renders a scene in constant time O(1) per ray, regardless of the scene's size and complexity. A prototypical implementation on a field-programmable gate array validates the architecture's feasibility and provides first laboratory results.

No abstract