Note: hfalk, ESD

Abstract: Usually, optimizations are applied by compilers which translate source code into machine code for a specific processor. Recently, transformations taking place before compilation are attracting interest. Such source code optimizations have the following advantages over compiler-integrated approaches:<br /> First, source code optimizations are inherently portable, since the optimized source code can be compiled by any compiler for the considered source language. Second, the correctness of source code transformations is often easier to validate, since faster servers can be used due to the portability, instead of slow evaluation hardware or instruction set simulators. Third, source code optimizations are easy to integrate into existing industrial tool chains. Fourth, source code transformations are easier to understand due to their high abstraction level as compared to machine code optimizations.<br /> The examination of the source codes of embedded multimedia applications in this Ph. D. thesis revealed that such software only uses a small fraction of its execution time to compute audio or video data. Most of the execution time is used to evaluate and execute the complex control flow inside these multimedia algorithms. The source code optimizations proposed in this Ph. D. thesis thus focus on the analysis of the control flow of multimedia applications, and on its improvement. All proposed techniques are jointly able to reduce the execution times of realistic applications on average over nine different processor architectures by 69%. Furthermore, cumulative energy reductions by 83% have been observed.