Neural network spatio-temporal scheduling for reconfigurable heterogeneous architecture

Constant evolution of applications, in terms of complexity or performance needs, makes necessary the development of new architectures. Among all proposed architectures, reconfigurable architectures (RAs) offer performances close to a dedicated circuit with more flexibility. This flexibility is supported by the ''dynamic reconfiguration'' mechanism which permits to multiplex temporally and spatially different applications. Similarly to a general purpose processor, this feature requires an operating system to be fully exploited. This thesis focuses on th e definition of schedulers designed for RAs. Our work targets execution of complex applications - composed by several tasks whose execution order is not known in advance where scheduling algorithms must be executed at run-time and with a fast generation of scheduling solutions . For this purpose, we developed our scheduling algorithms following the Hopfield Neural Networks model (HNNs). This kind of neural networks has already been used to solve optimization problems (such as scheduling problems) where it was found that they produced solutions quickly . Moreover, they can be efficiently implemented on a RA since their evaluation consists of simple operations requiring few control flow statements.