Implementing FastMap on the GPU: Considerations on General-Purpose Computation on Graphics Hardware

Abstract

In this paper we focus on the implications of implementing generic algorithms on graphics hardware. As an example, we ported the dimensionality reduction algorithm FastMap to fragment programs and thus accelerated it by orders of magnitude, allowing for interactive tweaking and evaluating of the algorithm parameters for datasets of several hundred thousand points and tens of dimensions; even the animation of structural changes in relation to parameters is possible. This allows to complement the algorithmic heuristic used by FastMap by explorative results from human interaction. Such an approach can be considered a heuristic in itself, but has the advantage of being based on visual feedback, therefore allowing for iterative improvement of the results. Thus we demonstrate how to benefit from the high execution parallelism on commodity graphics hardware as an alternative to making use of other, more costly, multiprocessing techniques. We discuss performance and bandwidth aspects as well as accuracy problems since these results are of more general interest and can be applied to general processing on graphics hardware as a whole.