Memory size estimation for multimedia applications

Memory modules dominate the cost, performance, and power of embedded systems that process multidimensional signals, typically present in image and video processing. Therefore, studying the impact of parallelism on memory size is crucial for trading off system performance against area cost to enable intelligent system partitioning and exploration. We propose a memory size estimation method for algorithmic specifications containing multidimensional arrays and parallel constructs, intended as part of a high-level partitioning and exploration methodology. The system designer can trade-off estimation accuracy for increased run time. We present the results of our estimation approach on a number of image and video processing kernels, and discuss some preliminary results on the influence of parallelism on storage requirement.