12/01/1996 09:00 AM
Computer Engineering
Volume rendering or volume visualization is an algorithm for creating images from three-dimensional and four- dimensional data sets, without computing intermediate surface representations. Because of the inherent O(N^3) run time, numerous approximations are used to provide interactivity. As compute platforms have become more capable in their operations per second, and in their memory capacities, the requirements of volume rendering become more and more advanced. The goal of application users is high fidelity renderings of medical, simulation, and remotely sensed data sets. Interactivity provides a tractable means for setting the many input parameters, but interactivity is challenging because of the run time complexity. We investigate the further optimization of massively parallel algorithm solutions, in an effort to have the efficiency of the parallel approaches reach that of sequential ones, without sacrificing fidelity. This paper discusses experiments in extending permutation warping on the MasPar MP-2 implementation to include data dependent coherency optimizations, and object space to screen space precompositing communication savings. We start with a description of the base algorithm, and then describe the extensions, showing results of nearly 400% speed improvement or a factor of 5 speedup.
