Proposal to Extend the Linear Algebra Library LAPACK for High-Performance Computers

点击翻译按钮获取中文摘要

LAPACK is a linear algebra library for high-performance computers. The first release, targeted at shared memory machines, will be available at the end of 1990, and is already attracting significant interest both from the potential user community and from manufacturers. This research is concerned with extending still further the usefulness of LAPACK by increasing the functionality through additional linear algebra capability, and by increasing the user community through targeting a wider range of high-performance architectures, providing additional language versions, and exploiting clean computer arithmetic. The extensions include new subroutines useful in systems and control theory such as solving Sylvester equations and generalized singular value decomposition, new algorithms for certain eigenproblems which are much more accurate than any algorithms in EISPACK, accommodating distributed-memory architectures, versions of several subroutines to exploit special properties of some computer arithmetics (especially IEEE arithmetic), and Fortran 8X and C language versions of some of the most important subroutines. LAPACK includes a comprehensive set of testing and timing routines. These will be used as a basis for the design and development of a systematic performance analysis package for the common algorithms of linear algebra. This library will ease code development, make codes more portable among machines of different architectures, increase efficiency, provide tools for evaluating computer performance.