| Abstract |
|---|
| This article introduces a method for using redundant computations in automatic compilation and optimization of SIMD programs for distributed memory machines. This method is based on a generalized definition of parametrized data distributions which allows a more flexible use of redundancies. In contrast to former (scheduling) methods using redundant computations, this method uses parametrized data distributions. This makes the usage of redundancies manageable in practice. Two examples show the benefits of this optimization method compared to other methods using parametrized data distributions: a wave simulation algorithm and the convolution. |
| Contact |
| Joern Eisenbiegler Universitaet Karlsruhe,Inst. f. Programmstrukturen und Datenorganisation,Lehrstuhl Prof. Goos,Zirkel 2,76128 Karlsruhe,Germany, eisen@ipd.info.uni-karlsruhe.de |