| Volume visualization has emerged as a prominent off-shoot of graphics for viewing and manipulating scientific datasets, such as those obtained from MRI, CT-scans, and CFD, or volumes which are generated by voxelizing geometric models. Direct volume rendering has the mechanisms for allowing simple manipulation techniques and easy viewing of the inside of objects. However, the size of these volumes tends to be so large that even the most powerful uniprocessor machines are unable to provide the much desired interactivity in such environments.; In this dissertation, we have designed and implemented three scalable parallel volume rendering algorithms on the Cray T3D. Our methods suggest new paradigms and alternatives to traditional ways of parallel rendering in general, and parallel volume rendering in particular. They are designed in such a way that the complete local memory is used only for that data which are required by a processor. Most earlier algorithms made use of the hardware cache only for exploiting spatial coherency; the performance of these algorithms is prone to degrade with increasing dataset sizes or decreasing cache sizes. Utilizing the local memory as another level of cache, i.e. software cache, was not explored for parallel rendering. In addition, ways to hide latency and minimize network congestion in each of these algorithms are novel approaches in themselves. In addition, previous algorithms are not applicable to render colossal datasets (possibly in compressed form), as they are prone to thrashing. We propose a new algorithm that combines the advantages of both the object-order algorithms and image-order algorithms to solve the problem of thrashing and provide the most coherent screen traversal scheme.; In summary, this research has primarily focussed on some of the as yet unexplored problems in parallel volume rendering, e.g., latency hiding, optimal local memory utilization, optimal screen traversal, reducing network congestion, and portability, and has suggested exclusive ways to eliminate some or all of these. Our coherent algorithms demonstrate scalability to a very high degree, with potential to improve even further. With the advent of high-resolution scanners, the dataset sizes approach the limits of disk storage. The coherent algorithms developed in this dissertation will provide state-of-the-art methods for visualizing such large datasets in the future. |
