Automatic Source Transformation and Performance Evaluation for Optimized Utilization of the Matlab Runtime System

The ADiMat software is a tool that offers Automatic Differentiation of any Matlab function using a hybrid approach. Depending on the configuration, the resulting derivative function is executed using a derivative container class. Earlier performance tests showed a weakness that was attributed to the object oriented class system of Matlab. In a first step to validate this assumption, the Matlab runtime environment is tested regarding function and method call overheads as well as property access overhead with Matlabs objects. Furthermore, using a simple test function, a specific derivative class is compared to a set of manually created Matlab functions offering the same functionality as the class. It was shown that the resulting speedup is substantial. An automatic method, transforming the XML based abstract syntax tree created with the existing ADiMat toolchain, was developed. The process completely removes the derivative object usage from any derivative function created with ADiMat. It is implemented as a set of XSLT stylesheets. The removal of object orientation works by resolving the class operators and methods statically. The necessary type inference process is based on the identification of the prefix of all derivative variables introduced by the differentiation process. Additionally, the respective classes are semi automatically transformed to a set of functions based on heuristics. A final performance assessment utilizing a PDE function is done comparing the benefits of the implemented transformation. It was shown that the improvement depend on the way the classes store the directional derivatives