Numerical Software Libraries



LAPACK/ScaLAPACK

We have preparing a new release of LAPACK in mid year. This new release includes the divide-and-conquer algorithm for SVD, new QR factorization with column pivoting, modifying the pivoting of the symmetric indefinite factorization, and the correction of a number of errors in the existing software. We have been working on a version of LAPACK that can be used from Fortran 90. The software currently include only a subset of the LAPACK functionality, and is in the process of being extended. In addition to developing new algorithms and software our activities with LAPACK center on interaction with the user community via support for the LAPACK mailing alias and public relations for the commercial use of LAPACK; maintenance of the software, documentation, and errata files on netlib; maintenance of LAPACK WWW homepages; software quality control via the coordination and integration of new software with the authors; testing of new software across multiple architectures; and maintaining and updating the LAPACK Users' Guide. The research project ``A scalable parallel library for numerical linear algebra'' consisted of a number of closely related topics which involves researchers at a number of institutions. This research led to a number of important new software tools and standards. The initial public release of ScaLAPACK occurred last year, and included routines for the solution of a general system of linear equations via LU and Cholesky factorizations, orthogonal factorizations (QR, RQ, LQ, and QL), reduction to condensed form (upper Hessenberg, tridiagonal form, and bidiagonal), and the symmetric eigenproblem. Since this initial public release, many improvements have been underway to increase the flexibility and functionality of the library.

Click here for comment from users.


Linear Algebra Templates

The linear algebra work on Templates was begun after numerous discussions with members of the CRPC applications group. The Templates work is having an impact on the way the Center approaches the implementation of parallel algorithms. This represents a clear path in helping to communicate our parallel algorithms and applications between users. This effort is serving a valuable pedagogical role in teaching state-of-the-art numerical and parallel programming and instilling a better understanding of the algorithms employed and results obtained. The iterative methods templates work is already becoming a ``handbook'' for computation scientists interested in solving large non-symmetric systems of equations.

Click here for comment from users.


BLAS Workshop

The existing BLAS have proven to be very effective in assisting portable, efficient software for sequential and some of the current class of high-performance computers. We organized a Forum and series of workshops along with colleagues from the Rutherford Appleton Laboratory and Cray Research. The purpose of the workshops are to investigate the possibility of extending the currently accepted standards, to provide greater coverage of sparse matrices, and provide additional facilities for parallel computing. In particular to consider standardizing a set of Parallel BLAS along the lines of the existing BLAS for the dense and sparse cases.

  • The audience for the BLAS has expanded to encompass application scientists in general.
  • Similar meeting style as in HPF and MPI forum should be adopted.
  • Lower-level routines are needed to insure high-performance on small problems.
  • Interoperability between programming languages needs to be stressed.
  • Extensions and alternative entry points (to allow different levels of error checking and profiling) to the BLAS seem appropriate.

    Additional information, including many of the presentations for the workshop, can be found on the URL:

    http://www.netlib.org/utk/papers/blast-forum.html

    As a result of this meeting and a follow up meeting held at Supercomputer '95 in San Diego it was decided to organize a series of meetings similar to the MPI Forum to continue the BLAS standardization work. We expect this activity to be on going through FY 1997.