LAW: A Tool for Improved Productivity with High-Performance Linear Algebra Codes. Design and Applications

摘要

LAPACK and ScaLAPACK are arguably the defacto standard libraries among thescientific community for solving linear algebra problems on sequential,shared-memory and distributed-memory architectures. While ease of use was amajor design goal for the ScaLAPACK project; with respect to its predecessorLAPACK; it is still a non-trivial exercise to develop a new code or modify anexisting LAPACK code to exploit processor grids, distributed-array descriptorsand the associated distributed-memory ScaLAPACK/PBLAS routines. In this paper,we introduce what we believe will be an invaluable development tool for thescientific code developer, which exploits ad-hoc polymorphism, derived-types,optional arguments, overloaded operators and conditional compilation in Fortran95, to provide wrappers to a subset of common linear algebra kernels. Thesewrappers are introduced to facilitate the abstraction of low-level detailswhich are irrelevant to the science being performed, such as target platformand execution model. By exploiting this high-level library interface, only asingle code source is required with mapping onto a diverse range of executionmodels performed at link-time with no user modification. We conclude with acase study whereby we describe application of the LAW library in theimplementation of the well-known Chebyshev Matrix Exponentiation algorithm forHermitian matrices.