Techniques for the interactive development of numerical linear algebra libraries for scientific computation.

摘要

The development of high-performance numerical algorithms and their effective use in application codes is an iterative process involving the refinement of the algorithms and their implementations that continues during the lifetime of the algorithm. Knowledge and expertise from the areas of numerical analysis, computer software, compilers, machine architecture, and applications are required during the development. To improve this process, the FALCON environment was developed to combine the analysis techniques from restructuring compilers with the algebraic techniques from numerical analysis. In this thesis, interactive techniques that were developed to extend the FALCON environment are described. These techniques allow the developer to improve the analysis of the algorithm, to restructure the algorithm using transformation patterns, to utilize additional information about structures within the data, and to control the generation of the target code. The experimental results show that the codes generated by the interactive techniques have better performance than those generated automatically. In addition, the environment was extended to support the generation of C++ code. When the C++ code generated by FALCON is compared to the code generated by other MATLAB translators, the C++ code is typically faster. However, when compared against the Fortran 90 code generated by FALCON, the C++ code is usually slower.