Design of a fast voxel processor for parallel volume visualization

Abstract

The basics of a parallel real-time volume visualization architecture are introduced. Volume data is divided into subcubes that are distributed among multiple image processors and stored in their private voxel memories. Rays fall into ray segments at the subcube borders. Each image processor is responsible for the ray segments within its assigned subcubes. Results of the ray segments are passed to the image processor where the ray continues. The enumeration of resampling points on the ray segments and the interpolation at resampling points is accelerated by the voxel processor. The voxel processor can additionally compute a normalized gradient vector at a resampling point used as a surface normal estimation for shading calculations. In the paper the focus is on operation and hardware implementation of this pipeline processor and the organization of voxel memory. The instruction set of the voxel processor is explained. A performance of 20 images per second for a 2563 voxel volume and 16 image processors can be achieved.