Computing parallel prefix and reduction using coterie structures

Abstract

The efficient computation of region parameters in image understanding by a SIMD (single-instruction multiple-data) array requires that those regions be processed simultaneously. The difficulty is in orchestrating nonuniform data-dependent communication using only a single thread of control. The authors have found that, on reconfigurable broadcast meshes, coterie structures can be used to overcome this problem. They present a deterministic algorithm to compute parallel prefix in O(log N) communication steps for a number of real images and sketch a randomized reduction algorithm based on graph contraction that has O(log N) complexity for all images.<>