A Cellular Architecture for Ray Tracing

Abstract

We propose in this paper a massively parallel machine dedicated to image synthesis by discrete ray tracing techniques. This machine is a four-stage pipeline, the last stage being a bidimensional cellular array with one cell per pixel. Two main phases describe its behaviour: · Loading into the cellular array of the objects of the scene to be displayed, after having been transformed into sets of planar polygons, and then into voxels. · Cellular ray tracing over the fully distributed scene. The first phase allows us to see this machine as a massively parallel (not realistic) rendering unit: at the end of the loading phase: objects are fully identified pixel per pixel in the cellular array. Then, we have only to display the computed visual features (by means of Gouraud or Phong-like incremental methods during the loading phase). The second phase increases the image quality by executing the ray tracing algorithm in a very special way, i.e., completely distributed all over the many cells of the array. In that phase, objects are seen as split into voxels into a virtual 3D memory space. The machine is an attempt to bring a dramatic answer to the problem of performance, taking into account not only the computational power required for image synthesis by using a massive parallelism, but also the realization costs by using very regular structures, which make it a VLSI-oriented architecture.