Automated Generation of Heterogeneous Multiprocessor Architectures: Software and Hardware Aspects
Youenn Corre · HAL (Le Centre pour la Communication Scientifique Directe) · 2013
Embedded systems are now ubiquitous and the increase in the in- tegration capacity allows for more features and capabilities. This trend has led to the emergence of Heterogeneous Multiprocessors Systems-on-Chip(H-MPSoC)whichprovideawaytorespectthecost and performance constraints inherent to embedded systems. How- ever they also make the task of designing and programming such systems a long and arduous process. The skills required along with the long development time are obstacles to their diffusion. It is thus necessary to develop tools that will free designers from architectural and programming details, so that they can focus on the tasks where they can bring added-value. The objective is thus to automatize the tedious tasks that burden the design of H-MPSoC, in particular on FPGA, by providing a higher-level of abstraction following a method that brings together High-Level Synthesis and hardware/software co- design beyond the existing solutions which are whether incomplete or unfit.The presented work aims at providing an answer to these problems. They introduce a design framework relying on the automation of te- dious tasks and allowing designers to express their expertise where they want to. For this, we rely on an architecture model defined with a high-level formalism independent from implementation details, pro- viding a solution to the lack of multiprocessor architecture in FPGAs. Thisspecificationmodelalsoallowsdesignerstoprovidedesigncon- straintsinaccordancewiththeirlevelofexpertiseorinvolvement.The design space exploration is implemented as a scalable algorithm re- lying on fast and accurate estimation techniques. A method for the explorationofhardwareacceleratorsbasedonhigh-levelsynthesis to provide fast cost estimations is introduced. Finally the integration of model-driven engineering methods enables portability and reuse by generating the final design implementation. The framework is val- idated through two case studies: an MJPEG video decoder and a morecomplexfacedetectionapplication.