The International Conference for High Performance Computing, Networking, Storage and Analysis
Creating High-Performance Linear Algebra Using Lighthouse and Build-to-Order BLAS.
Authors: Jeffrey J. Cook (University of Colorado), Elizabeth R. Jessup (University of Colorado), Sa-Lin C. Bernstein (Argonne National Laboratory), Boyana Norris (University of Oregon)
Abstract: Creating efficient linear algebra software is essential to HPC, but it is a time-consuming task. Typically, scientists use a Basic Linear Algebra Subprograms (BLAS) as the mathematical building-blocks for complete algorithms. They can present a steep learning curve, however, and require extensive knowledge to optimize. In order to make BLAS and derivative libraries such as LAPACK more accessible, we present Lighthouse, a web-based taxonomy of dense and sparse linear system and eigenvalue solvers . With its new interface to the Build-to-Order BLAS compiler (BTO) , which generates custom linear algebra kernels from MATLAB-like equations, it is easy for users to create high-performance code.