A shortest path algorithm for banded matrices by a mesh connection without processor penalty

Abstract

We give an efficient shortest path algorithm on a mesh-connected processor array for n/spl times/n banded matrices with bandwidth b. We use a [b/2]/spl times/[b/2] semisystolic processor array. The input data is supplied to the processors array from the host computer. The output from the processor array can be also supplied to itself through the host computer. This algorithm computes all pair shortest distances within the band in 7n-4[b/2]-1 steps.<>