A. A. Berlin and R. J. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In PEPM'94, pp.133--141, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....transformation systems; i.e. they analyze programs in a high level language and generate specialized programs in the same language. They have been successful in the optimization of various programs, such as interpreters, scientific application programs, and graphical application programs[4, 13, 20]. Run time specialization (RTS) techniques[10, 12, 18, 22] e#ciently perform partial evaluation at run time (1) by constructing a specializer (or a generating extension) for each source program at compile time and (2) by directly generating native machine code at run time. The drastically ....

A. A. Berlin and R. J. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In PEPM'94, pp.133--141, 1994.

....transformation systems; i.e. they analyze programs in a high level language and generate specialized programs in the same language. They have been successful in the optimization of various programs, such as interpreters, scientific application programs, and graphical application programs[4, 11, 17]. Run time specialization (RTS) techniques[10, 15, 19] e#ciently perform partial evaluation at run time (1) by constructing a specializer (or a generating extension) for each source program at compile time and (2) by directly generating native machine code at run time. The drastically improved ....

A. A. Berlin and R. J. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In PEPM'94, pp.133--141, 1994.

....too complex, or because they depend on the input s, which is not available at compile time) and 2. remove administrative overhead introduced by modular or general programming Point 2 has been demonstrated in many applications of partial evaluation (e.g. Mogensen 86] Berlin 90b] Berlin 90a] Berlin 94] Jorgensen 91] Lisper 91] Mossin 93] Baier 94] but little attention has been paid to point 1. In this project we will focus on point 1, by applying partial evaluation to a realistic application that is already programmed for speed. Point 2 will be of no concern. Other goals for the ....

Berlin, Andrew A. and Surati, Rajeev J. Partial Evaluation for Scientific Computing: The Supercomputer Toolkit Experience. In ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based Program Manipulation. 1994. 20

....such as Smalltalk [20] and Self [12, 41] and simulators [74, 77] Recent just in time (JIT) compilers for mobile code such as Java [37] use dynamic compilation techniques to improve on the performance of interpreted code without incurring the overhead of a static compiler. Berlin and Surati [6] reported 40x performance improvements thanks to partial evaluation of data independent scientific Scheme code. The conversion from Scheme to a low level, special purpose program exposes vast amounts of fine grained parallelism, which creates the potential for additional orders of magnitude of ....

....unrolling and inlining. A detailed study of the effect of various architectural features on dynamic compilation might provide insights that would improve overall system performance. Similarly, it may be useful to explore the interaction of dynamic compilation and parallelism. Berlin and Surati [6] argue that specialization of high level code such as Scheme can expose vast amounts of fine grained parallelism. Such parallelism should become increasingly important as architectures evolve towards greater instruction issue widths and deeper pipelines. Dynamic specialization should be more ....

A. A. Berlin and R. J. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In Proceedings of the Workshop on Partial Evaluation and Semantics-Based Program Manipulation, Orlando, FL, June 1994. ACM.

....is then used to automatically produce specialised versions with the required efficiency. 2. 4 Partial Evaluation for Scientific Computation There has been a lot of work at MIT on the Supercomputer Toolkit, a system that uses PE to achieve very impressive speedups for scientific computations [6, 29, 30, 7]. There are three aspects to this. 1. Using PE to convert high level, abstract code to low level, efficient code The idea is to let scientists write programs in a high level language (Scheme) that allows them to express their abstract understanding of the problem. A partial evaluator is then ....

....write programs in a high level language (Scheme) that allows them to express their abstract understanding of the problem. A partial evaluator is then used to translate this program into a low level, purely numerical C program, where the control flow has been resolved at PE time. Berlin and Surati [7] report a speedup of 38 attributable to this step for five time steps of a simulation of the 9 body problem. 2. Using PE to make effective use of Pipelines By removing most control flow instructions, PE exposes vast amounts of fain grain parallelism that can be used to keep floating point ....

[Article contains additional citation context not shown here]

Andrew Berlin and Rajeev Surati. Partial evaluation for scientific computing: the super computer experience. Technical Report AI Memo 1487, Cambridge, MA: MIT Press, May 1994.

....evaluation was initially studied in the context of functional programming languages. Since then, partial evaluation techniques and tools have been developed for a large variety of languages [23, 9, 5, 20] logic, imperative, object oriented and have been applied with encouraging results [6, 16, 2]. Currently, there exist several tools able to treat real life, widely used languages such as C [4, 1, 11] These tools can perform specialization either at compile time (with respect to statically available constants) or even during program execution (with respect to run time constants) 12] 2 ....

A. Berlin and R. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In PEPM'94 [25], pages 133--141.

....as possible (using the static values) and outputs a specialized residual program. For example, a partial evaluator could take a dot product routine, and produce a specialized 8 version for a particular vector length. These techniques have been applied to scientific codes with promising results [2, 3, 22]. However, this just provides a taste of the field; partial evaluation has evolved into a comprehensive toolbox containing both theories and practical software. One of the most important theoretical contributions was that the concept of generating extensions [13] unifies a very wide category of ....

A. Berlin and R. Surati, Partial evaluation for scientific computing: the super computer experience, tech. rep., Artificial Intelligence Laboratory, Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts, May 1994.

....program specialization techniques generate an optimized program with respect to those parameter values. Partial evaluation[10, 13] which specializes programs at sourceto source level, has been shown to be useful to optimize various programs, such as interpreters[17 19] scientific applications[5], and graphical applications[11] Recently, run time specialization (RTS) 1 techniques have been actively studied[6 9, 14, 15, 21, 22] Those techniques e#ciently specialize programs at run time (1) by constructing a dedicated specializer for each target program at compile time, and (2) by ....

A. A. Berlin and R. J. Surati. Partial evaluation for scientific computing: The supercomputer toolkit experience. In Partial Evaluation and Semantics-Based Program Manipulation, pages 133--141, June 1994. Published as Technical Report 94/9, Department of Computer Science, University of Melbourne.

No context found.

A. Berlin and R. Surati, "Partial Evaluation for Scientific Computing: The Supercomputer Toolkit Experience," Proc of ACM SIGPLAN Workshop on Partial Evaluation and Semantics-Based Program Manipulation 1994, Orlando, FL 1994. Also available as MIT Artificial Intelligence Laboratory Memo No. 1487, May 1994

No context found.

) Berlin, A. A. and Surati, R. J. "Partial Evaluation for Scientific Computing: The Supercomputer Toolkit Experience". In Partial Evaluation and SemanticsBased Program Manipulation, pp. 133--141, Orlando, FL, June 1994. ACM SIGPLAN. Published as Technical Report 94/9, Department of Computer Science, University of Melbourne.

No context found.

A. A. Berlin, R. J. Surati, "Partial evaluation for Scientific Computing: The Supercomputer Toolkit experience", PEPM 1994 - Proceedings of the ACM SIGPLAN Workshop on Partial Evaluation and Program Transformation Techniques, Orlando, FL, 1994, pp. 133--141.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC