Design space exploration for optimal memory mapping of data and instructions in multimedia applications to Scratch-Pad Memories

Abstract

In this paper, we propose a new methodology for optimal memory mapping of data and instructions to Scratch-Pad Memories (SPM). In the mapping process, we optimize, as the main priority, the number of memory accesses to minimize power consumption. Minimization of external memory accesses lowers switching activity and therefore power consumption. The optimization is done by finding Pareto-points, using multi-objective optimization that combines different cost functions. Our methodology is intended to be used in real-life situations in industry where there is often a need for mapping third party applications to a specific architecture. For evaluating our methodology, we also use commercial video H.264 and audio eAAC+ applications. Our experiments show that SPM is well suited for these applications for reducing external accesses to reduce power consumption but has limited significance on overall performance improvements. The proposed methodology provides a way to combine SPMs with caches to optimally use this memory architecture. Our experiments indicate high accuracy of our methodology for predicting SPM and external memory accesses. We have obtained 90% accuracy between results of our methodology and results for executing applications on a given architecture.