Methodes de conception et de partitionnement logiciel pour des architectures paralleles SIMD lineaires.

摘要

The growing demand for integration, sophistication, reliability and particularly performance, pushes to design microelectronics systems of increasing complexity which makes the development and the use of these systems more difficult. This motivates the need for new optimization methods allowing to increase the performances, reliability and applications portability on these systems. The search for such methods is the object of this thesis.;This thesis begins with a literature review of relevant methods. A suitable method is selected and adapted to be used in the framework of the PULSE project. This project was aimed at developing a SIMD multiprocessor dedicated to real time processing of digital signals, including video signals. A series of examples representative of typical applications are optimized according to the selected method. The obtained performances are analyzed and compared with manual practical applications, evaluating obtained gains and setting the method limits.;A tool for simulating the PULSE system was developed to allow the analysis and the estimation of dynamic performances. This tool also allows to visualize step by step, the simulation states. This characteristic was very much used by a wide number of development team members on the PULSE project. Which permitted the debugging of many applications and facilitated the training of users for the PULSE multiprocessor.;This thesis is a step toward the development of methods for optimizing and estimating performances. In spite of significant progress, the goal of developing an automatic parallelising compiler for PULSE remained elusive. Such a compiler, whose development appears feasible, but beyond the scope of this project, could handle the hard task of extracting and exploiting parallelism available in the applications.