A survey of architectures for volume rendering. IEEE Engineering in Medicine and Biology Magazine, 9(4):18{23, 1990. Lev89a] Marc Levoy. Design for a real-time high-quality volume rendering workstation. A three-dimensional shaded display method for voxel-based representations.planes 5: A heterogeneous multiprocessor graphics system using processor-enhanced memories. 7. Acknowledgements 13 structures such as the transfer functions or viewing parameters, followed by rerendering. In that case, some of the data structures and code will be in the local instruction or data caches, giving improved performance. The parallel raycasting method intuitively seems to be more easily extended to render other types of volumetric datasets. Those datasets in which neighboring cells are not implicit (such as nite element meshes) will require an explicit visibility graph as opposed to the implicit one used here for a regular rectilinear dataset. In particular, pointers to adjacent cells will be needed in addition to the counts. For curvilinear computational meshes, ordering constraints may need to be imposed between two cells that are not even neighbors since the mesh can be curved in < 3. These complications will result in even larger memory requirements for the storage of the visibility graph which may be prohibitive for large volumetric datasets. 6.4 Future Work This research is being continued along several directions: Development of schemes for local memory utilization will reduce the overhead incurred in remote references. The projection approach is an important method for volume rendering. It would be interesting and worthwhile to design a more eecient task generator or decomposition for this algorithm. In interactive use, two of the most likely updates will be to the transfer functions and to the viewing parameters. It may be possible to achieve very fast image update rates for changing transfer functions by obtaining, ordering, and caching all the samples for each pixel and requiring only compositing operations for each change to the transfer function. Changing the view would require the volumetric dataset to be resampled. Since the transfer function and viewing parameters are currently propagated to each processor's local memory, disseminating updates to these data structures will be an issue. It may be that they will need to be in shared memory which will increase contention. The tradeoos will need to be explored. Design of parallel algorithms which can utilize coherency in an eecient manner. This will be particularly important for eecient rendering of embedded geometric primitives. Extensions to the algorithms to handle curvilinear …