Multi level partitioning of signal processing algorithms for multiprocessor SoC

点击翻译按钮获取中文摘要

The growing demand for connectivity induces the wireless communication research community to constantlyinnovate. The novel techniques which are currently being explored allow to significantly increase spectralefficiency. However, a larger number of users and greater amounts of information on the same bandwidthrequire increased signal processing capacities. In this context, the computing power of programmable platformshas considerably increased, thereby requiring designers to deal with the exponential rise of their complexity.Due to the very intense competition in this market, the industry and designers try to get the most out of all thetechniques that are likely to speed up their design process and to optimize the performances. The currentresearch project conducted by Professors Thibeault and Gagnon in collaboration with Octasic since May 2006has allowed to demonstrate the importance of partitioning and that this crucial stage constitutes a bottleneckduring the design process.This research proposal aims to explore various ways to optimize partitioning. These approaches are comparedfor a multiprocessor platform with signal processing systems in order to achieve a judicious allocation ofavailable resources. The specific needs of the final application will also be considered. Constrained andunconstrained partitioning scenarios will be compared. The aim of the unconstrained scenario is to manage thepartitioning problem at a high level of abstraction with a minimum of initial constraints. It is based on ametaheuristic evolutionary algorithm which has the capacity to solve such complex problems. It will becombined to two other constrained partitioning scenarios in order to evaluate its full potential. The novelty ofthis project resides in the exploration of several partitioning techniques for a reduced set of criteria. They arespecifically evaluated for Software-Defined Radio applications with multi-user transmission and reception.These applications are available and continuously evaluated in our laboratories.