A Modern Framework for Portable High Performance Numerical Linear Algebra

byJeremy G. Siek This thesis describes a generic programming methodology for expressing data struc-tures, algorithms, and optimizations for numerical linear algebra. A high-performanceimplementation of this approach, the Matrix Template Library (MTL), is also described.The goal of the MTL is to facilitate development of higher-level libraries and applicationsfor scientific computing. In addition, the programming techniques developed in this thesisare widely applicable and can be used to reduce development costs, improve readability,and improve the performance of many kinds of software. Portable high performance is aparticular focus of the MTL. Flexible kernels were constructed that provide an automatedtool for cross architecture performance portability.