A toolbox for computing the singular value decomposition on distributed memory computers

We present a parallel software implementation for computing the singular va lue decomposition (SVD) of general, banded or bidiagonal matrices. First, t he matrix is reduced to bidiagonal form. This reduction can be rearranged i n a way that allows heavy use of matrix-matrix operations. Then the singula r values are computed in an iterative process. Finally the singular vectors are computed independently. The methods are compared to the ScaLAPACK libr ary with respect to accuracy, stability and performance on parallel compute rs.