Efficient implementation of multilevel algorithms on hypercube supercomputers for computer vision

The authors deal with the problem of implementing multilevel algorithms on hypercube parallel computers for computer vision problems. The conventional pyramid is a multilevel structure with identical 2 × 2 reductions between all pairs of neighboring levels. The mapping algorithms proposed for source multilevel structures and target parallel computers implementing the hypercube topology are based on some of the most important pyramid mapping algorithms. Application algorithms are classified according to their performance requirements and the most appropriate mapping algorithms are then chosen for their implementation The new mapping algorithms are compared according to a set of graph embedding metrics and from results produced on a Connection Machine system CM-2 massively parallel computer. The comparison of the CM-2 results shows that the selection of multilevel structures other than the pyramid for the implementation of algorithms is a more efficient solution most of the time. In addition, the selection of the most appropriate mapping algorithm often becomes a very critical decision for high yields