Vectorization of Visualization Algorithms - A Practical Example

In scientific visualization, interactive image generation times are needed. Yet typical algorithms are computationally expensive, e.g. image generation algorithms for volume rendering, which is an important and widely used technique. Since vector computers are still among the most powerful machines available, we propose a vectorized variant of a volume rendering algorithm. Vectorization has a long research tradition. Many of the obstacles encountered in transforming loops of sequential programs into vector operations can be handled automatically by state of the art compilers. However, experiences in the course of our work revealed that there are still some shortcomings. It turned out that even for relatively simple optimization transformations it can be necessary to resort to assembler level programming. This is exemplified by a detailed analysis of the proposed algorithm. Results of the final implementation on a mini supercomputer conclude the presentation.