BEGIN:VCALENDAR
PRODID:-//Microsoft Corporation//Outlook MIMEDIR//EN
VERSION:1.0
BEGIN:VEVENT
DTSTART:20131121T184500Z
DTEND:20131121T190000Z
LOCATION:601/603
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:ABSTRACT: High performance linear algebra is an important component to many numerical applications. Efficient matrix computation requires careful optimization based on application and hardware, typically a time consuming process. Today most scientific applications use BLAS (Basic Linear Algebra Subprograms) to provide high-performance linear algebra computation. This research presents an alternative approach, based on an empirical search strategy for linear algebra optimization. It uses Build-to-Order BLAS, a domain-specific language for linear algebra that produces highly-optimized multicore C code as output. This approach maintains a high-level understanding of the operations, which allows more aggressive code transformation. The search is based on a highly-tuned genetic algorithm, which performs global search over a transformation space while maintaining correctness and using application-specific heuristics. Our results show significant speedups compared to traditional library-based approaches for many matrix kernels.
SUMMARY:Empirical Search to Optimize Matrix Computation
PRIORITY:3
END:VEVENT
END:VCALENDAR