engleski

Parallelizing the Kogbetliantz Method: A First Attempt

The paper investigates a way how can thet two- sided Jacobi-type method for computing the singular value decomposition of triangular matrices, known as Kogbetliantz method, be adapted for use with parallel computers with shared memory. The slower row operations can be replaced, at low extra cost, by the faster column operations. It is shown how can the method be further modified to work with blocks. In any case, the initial triangular or rectangular matrix has to be brought to a special, butterfly-like form. In the iterative part of the algorithm, this special form gradually changes, but after a fixed number of parallel steps, which corresponds to two standard sweeps, the initial butterfly-like form is retained. This property simplifies the algorithm and enhances its performance.

singularvalue decomposition; Kogbetliantz method; parallel algorithm; block algorithm

nije evidentirano