Institute of Computer Languages
Compilers and Languages Group

Talks 2004 - Dan Quinlan

The Compilers and Languages Group invites you to a talk given by

Dr. Dan Quinlan

(Lawrence Livermore National Laboratory) on

Telescoping Languages: From Domain Specific Libraries to Domain Specific Languages

Date:	Monday, June 7, 2004
Time:	16:15
Location:	TU Wien, building at the Freihaus, FH HS 2, Wiedner Hauptstraße 8, Turm B, yellow area, 2nd upper floor

Abstract:

The development of large scientific applications is particularly tedious, slow, and expensive given the combination of algorithm, application, and architecture complexities. Telescoping languages define a mechanism for building new languages from existing base languages by providing incrementally more optimization tailored to abstractions present in specific applications or libraries. A goal of the approach is to make it cheap and practical to construct efficient domain specific languages to service the requirements of small, but high specialized domains of computational science; or computational science generally. Any common language can form the base language for this concept; separately constructed optimizers scan the application to recognize user-defined abstractions expressed in the base language. The union of the grammar to define the base level language with the additional product rules to define user-defined abstractions outside the base language defines a new domain specific language built on top of the base language. Only optimizations on the domain specific abstractions are required in this incremental approach to define a new language. The dividing line between library and compiler optimizations may be adjusted as required. No additional syntax is possible within this approach, but this yields simpler new domain specific languages which only include new domain specific abstractions and are more acceptable to end users. This talk will introduce the ROSE project, current work is focused on development of "Telescoping Languages" based on source-to-source translators in C++ which has rich support for the development of object-oriented abstractions. Less powerful support for C is also of interest and is related to the topic of "Interface Compilation." Future work will also include F90, this work is a collaboration with Rice University.

This approach can be summarized as formally turning domain specific libraries into domain specific languages. And if pushed to the limits, defines a mechanism to turn any application into an implementation in its own language; supported with highly application specific optimizations, where automatically generated code is expressed in the base language and can be targeted at a specific computer architecture. Many parts of the construction of such specialized optimizing compilers could be automated from a specific library's interface. It is likely that work on the classification of the semantics of library abstractions could automate many of the optimizing transformations. This approach also feeds back into the better design of domain specific abstractions yielding abstractions that can be more efficiently optimized. Telescoping Languages is a rich area for research, the approach could well be a significant mechanism to addressing the poor economic characteristics related to the support of computer languages within scientific computing.

About Dr. Dan Quinlan:

Dan Quinlan works on the Overture Project in the Computer Science Group in CASC at Lawrence Livermore National Laboratory, CA, U.S.A. His research is in numerous areas that intersect computer science and numerical analysis. Research interests include object-oriented numerical frameworks, parallel adaptive mesh refinement, parallel multigrid algorithms, semantics-based source code transformations, C++ compiler tools/infrastructure/design, cache-based optimizations, parallel array classes, parallel data distribution mechanisms, and parallel load balancing algorithms.

Dr. Quinlan earned his Ph.D. in Computational Mathematics from University of Colorado.