Time-Optimal Multiple Rank Computations on Meshes with Multiple Broadcasting

Abstract

Consider arbitrary collections A = a_1,a_2,.. .,a_n of items and Q = q_1,q_2,...,q_m (1 leqslant mn leqslant n) of queries from a totally ordered universe. The multiple rank problem involves computing for every query qi the number of items in A that have a lesser value. Our contribution is to show that the problem at hand can be solved time-optimally on meshes with multiple broadcasting. More specifically, if the collection A is siored in some order one item per processor and if Q is stored one query per processor in the leftmost frac{m} {{sqrt n }} columns of a mesh with multiple broadcasting of size sqrt n x /sqrt n, the corresponding instance of the multiple rank problem can be solved in Theta left( {m^{frac{1} {3}} n^{frac{1} {6}} } right) time. As an application we present a time-optimal algorithm to compute the histogram of a m-level gray image of size sqrt n x sqrt n in Theta left( {m^{frac{1} {3}} n^{frac{1} {6}} } right) time.