Visualization has become a very important research tool in the scientific community. Today the large number of graphics capable computers being sold and in use, make it possible for researchers to explore data with unprecedented detail.
The most common volume visualization method in use is volume rendering. This method evolved during the late 1980s and today several efficient and accurate visualization techniques are in wide spread use. Volume rendering consists of trying to ``synthesize'' all the information contained in a given dataset in a single 2D picture as if the user were looking at the real data. The key to the success of volume rendering as a visualization technique is to avoid binary decision during the rendering pipeline.
Previous techniques for displaying surfaces from volume data consisted of applying a surface detector to the sample array, fitting geometric primitives to the detected surfaces, then rendering these primitives using conventional surface-rendering algorithms. One common approach was to apply thresholding to the data and render by considering the remaining voxels as opaque cubes having six polygonal faces. For more information on volume rendering techniques see Section 3 and [35,1].
Volume rendering is a computationally intensive task. Current graphics computers are not fast enough to generate high quality images at interactive speeds. While several computer vendors have polygon based geometry engines, only a few very expensive machines support volume data and at a speed far from the ideal. Among the current available systems that support volume data are the IBM POWER Visualization System and the G2 by ISG Technologies Inc [36].
Several specialized machines for volume rendering have been designed and some are under construction. Among these are the CUBE architecture [2,3], the Voxel Processor [37], the PARCUM [38]. For a complete up to date survey on these architectures see [39], for a survey of volume rendering architectures for medical imaging [37]. The development of such machines may, in the future, lead to workstation graphics engines capable of achieving real-time volume rendering.
Until production versions of these machines are marketed, the trend in fast volume rendering will be to use commercially available parallel computers. In fact, some supercomputer centers are already offering packages for volume rendering for their parallel machines [40].
In order to make volume rendering feasible on the wide variety of architectures of parallel computer, one must understand how these machines work and how to program them efficiently. An introduction to general parallel computer architectures is given in Section 2.
There are several different algorithms for volume rendering, all of which have different strengths and are suitable for different architectures. A brief introduction to some of these algorithms is given in Section 3.
The other sections of this paper describe current implementations of the algorithms described in Section 3. Descriptions of the actual machines used in the implementation are given in the sections that reference them. Section 4 will describe shared memory MIMD implementations, Section 5 will describe message passing MIMD implementations and Section 6 will describe Data Parallel machine (SIMD) implementations. Parallel algorithms for irregular grid datasets are discussed in Section 7. A comparison of all the algorithms is given in the conclusions on Section 8.