This tool is suitable for testing the numerical properties such as accuracy and stability, of the library software and application codes.
Work on examining the user interface and documentation for PRECISE has been conducted with the intention of producing a tool that will allow non-expert users to do useful work. The aim is to strike a balance between flexibility and ease of use, particularly for non-expert users.
Preliminary Fortran routines for the computation of spectral portraits for sparse matrices have been completed. One routine is for application to a sparse matrix A and the second is for sparse matrix pencils (A,B). Each routine handles both real and complex matrices. The programs have been tested on the Meiko-CS2, the IBM SP2, and clusters of SUN and RS6000 workstations. Presently, the routines can produce spectral portraits for sparse matrices ranging up to approximately 5000 non-zero elements, depending on the properties of the matrices.
The possibilities of using the NAG visualization system IRIS Explorer as a front end for PRECISE have been investigated. IRIS Explorer has the necessary data reading and visualization facilities that PRECISE needs and CERFACS have agreed that they will examine the possibilities it offers. A copy of IRIS Explorer has been provided to CERFACS for evaluation purposes under a NAG Software Collaboration Agreement.
Fortran routines have been developed and completed to perform statistical error analysis with dense, sequential and parallel linear system solvers; testing and evaluation of the new routines is currently in progress. Testing has centered on LAPACK and ScaLAPACK linear equation solvers.
In addition, further testing and experience with the Fortran spectral portrait codes has been gained by producing portraits for approximately 50 matrices in the Harwell-Boeing Sparse matrix collection.
For Module 2 of PRECISE, routines have also been developed to produce plots of perturbed spectra for dense matrices. A modified complex version of the NAG routine F01BRF (based on the Harwell Library routine MA28), which factorizes a real sparse matrix, has been integrated into the standard version of the spectral portrait code. The testing has been successfully completed.