Systematic preprocessing of data dependent constructs for embedded systems

Data transfers and storage are dominating contributors to the area and power consumption for all modern multimedia applications. A cost-efficient realisation of these systems can be obtained by using high-level memory optimisations. This paper demonstrates that the state-of-the-art memory optimisation techniques only partly can deal with code from real-life multimedia applications. We propose a systematic preprocessing methodology that can be applied on top of the existing work. This opens more opportunities for existing memory optimisation techniques. Our methodology is complemented with a postprocessing step, which eliminates the negative effects of preprocessing and may further improve the code quality [7,10]. Our methodology has been applied on several real-life multimedia applications. Results show a decrease in the number of main memory accesses up to 45.8% compared to applying only state-of-the-art techniques.