Singular value decomposition on SIMD hypercube and shuffle-exchange computers (Q1205900)

The paper presents several implementations of the Hestenes method for the singular value decomposition of real matrices on the hypercube and shuffle-exchange computers of SIMD architecture. Unlike previous methods, the algorithms presented map a column pair onto a column of processors which reduces the time complexity. Another contribution is the proposal of a concept of high-dimensional shuffle-exchange networks. The time complexities achieved for one- and two-dimensional shuffle-exchange computers are of the same order as for their hypercube counterparts while the number of connections is lower/three for one-dimensional shuffle-exchange computers and four for two-dimensional ones/.