Profilage, caracterisation et partitionnement fonctionnel dans une plate-forme de conception de systemes embarques

Embedded systems have increasingly complex architectures and are now composed of several processors, buses, peripherals and hardware accelerators. Embedded system design methodologies currently used in industry are not keeping up with this evolution. System-level methodologies have been proposed in order to raise the level of abstraction of embedded system design. Such a methodology includes a virtual platform in which components can be allocated while application tasks can be bound to allocated components for a transaction-level simulation of the resulting architecture. A system-level methodology can accelerate embedded system design by using an executable specification, automating design space exploration and synthesizing an optimized architecture for the application.;However, current system-level methodologies have several shortcomings. They typically assume that the application is modeled with a restrictive model of computation and do not automate the synthesis of hardware blocks from application modules. They do not support a non-intrusive profiling of the application or of an architecture implementing the application. Their estimation methods do not automate the characterization of the application or of the platform. These methodologies consider processor allocation, task binding to processors and the choice of a communication topology to be separate problems instead of being different aspects of a single problem.;We present a system-level methodology for the design, architectural exploration and synthesis of embedded systems based on the Space Codesign. technology and its SPACE virtual platform. This methodology tackles these problems by combining a more expressive model of computation, a method for the automated synthesis of hardware blocks from a SystemC specification's modules, a non-intrusive system-level profiling, a method for the automated characterization of the application and of the real-time operating system (RTOS), as well as heuristics for a unified formulation of the architectural exploration problem.;We have thus defined for our methodology a novel model of computation, called real-time process networks (RTPN), which is an extension of Kahn process networks. This extension enables the modeling of important aspects of real-time processing, such as polling, sensor sampling, input/output peripherals and real-time constraints. We define the denotational semantics of RTPNs, which is used to verify the functional correctness of a refinement from a SystemC executable specification to a concrete implementation.;Our methodology includes an automated refinement from transaction-level communications to cycle- and pin-accurate protocols as well as an automated generation of hardware blocks from application modules. This method enables, when combined with an embedded vii software generation method which includes a RTOS, the generation of an implementation of the application, which can be simulated with the virtual platform or synthesized and executed on the final target. A novel profiling method is applied to such simulations in order to non-intrusively extract data on the performance of modules, processors, RTOS, buses and memories.;A novel automated method characterizes, through profiled simulations, both the application functionality and the software and hardware implementations of its modules. The devices and buses of the virtual platform have also been characterized and a novel method automates the characterization of the RTOS. These characterizations configure a high-level performance simulator for an accurate and very fast estimation of the performance of several candidate architectures for the application, taking into account bus contention and task scheduling on processors. This characterization also powers a fast and accurate estimation of required hardware resources.;A formulation of the architectural exploration problem is given such that it combines hardware/software partitioning, processor allocation, task binding on processors and the selection of a communication topology. This architectural exploration evaluates architectures for criteria of performance and hardware cost with our estimation method. We present for the first time a combinatorial analysis of this problem and its formulation as a local search problem, for which heuristics based on adaptative simulated annealing and reactive tabu search are defined. The architecture selected by the architectural exploration can then be synthesized towards a final implementation in a well-established RTL design flow. The methodology as a whole has been applied to three case studies: a rover guiding system, a JPEG decoder with skin detection and a WiMAX encoder/decoder.