Harvard University
Computer Science Colloquium Series
33 Oxford St., Cambridge, MA 02138
Colloquium
Compilation of Matlab Programs for Uniprocesors and Parallel Machines
based on Telescoping Languages
Ken Kennedy
John and Ann Doerr University Professor
Director, Center for High Performance Software
Rice University
Thursday, October 26, 2006
4:00PM
Maxwell Dworkin G125
(Ice Cream at 3:30PM - Maxwell Dworkin 2nd Floor Lounge Area)
Abstract
One way to increase national productivity is to broaden the community
of programmers by making it possible for end users to develop
applications for themselves. Indeed, many users today are producing
highly functional applications using scripting languages and high-
level problem-solving systems such as Matlab, Visual Basic, and S-
PLUS. Unfortunately, the productivity gains are offset by the costs
of rewriting these applications in “production” programming languages
such as C or Fortran once they are determined to be useful.
Eliminating the need for this rewriting step would bring about a
dramatic increase in global programming productivity.
In this talk, I will describe Rice’s work on telescoping languages,
which is exploring ways to generate optimized high-level problem-
solving languages from annotated domain libraries. The strategy
involves an extensive, compute-intensive preliminary analysis of the
library, performed at language-generation time. The output of this
process, which could take many hours, or even days, to complete, will
be an efficient compiler for an extended scripting language in which
calls to the underlying domain library are recognized and optimized
as primitive operations.
The talk will focus on how this strategy is being applied to
compilation of Matlab and associated libraries for both sequential
and parallel machines. At the heart of this approach is a novel type
analysis system that can be applied to libraries in the absence of a
calling program to produce type jump functions, which can be used at
script compilation time to produce good code without requiring
unreasonable compile times. To produce efficient code for parallel
machines, we are leveraging Rice’s High Performance Fortran compiler
technology at the library preprocessing stage.
The long-term goal of this research is to make it possible for
ordinary users, particularly scientists and engineers, to build their
own high-performance applications, just as they were once able to do
in the early days of Fortran. If this effort succeeds, it could
facilitate a dramatic broadening of the community that can use high-
performance computing platforms for problem solving.
_______________________________________________
iic-seminars mailing list
iic-seminars(a)calists.harvard.edu
http://calists.harvard.edu/mailman/listinfo/iic-seminars