Meeting the Challenges of Optimized Memory Management in Embedded Vision Systems Using Operations Research

The ever growing complexity of signal and image processing applications, and the stringent constraints related to their implementation makes their design, simulation, and implementation more and more challenging. Memory management is among the main challenge that electronic designers have to face. In fact, it impacts heavily the main cost metrics, including area, performance (real-time aspect) and energy consumption, of modern-day electronic devices. For some particular cases of image treatments, with non-linear access patterns to the memory addresses, a co-designed architectural solution and its optimization process, called Memory Management Optimization (MMOpt), was proposed by Mancini et al. (Proc. DATE, 2012). It creates an ad-hoc memory hierarchy for accelerating the accesses to the memories holding large image data. This chapter studies the optimization challenge reflecting the efficient operation of the MMOpt tool, which is formalized as a 3-objective scheduling problem. New algorithms are proposed for producing efficient solutions, leading to enhance the run-time performance and reduce both energy consumption and cost of the circuits produced by MMOpt. The performance of these algorithms is compared, on the same real-world data set as used by Mancini et al. [14], against the one currently in use in the MMOpt tool. The results show that our algorithms perform well in terms of computational efficiency and solution quality.