Thesis defense of Matthieu Payet
The October 26, 2016
at 9:30 am
25 rue du Dr Remy Annino,
"Conception de systèmes programmables basés sur les NoC par synthèse de haut niveau : analyse symbolique et contrôle distribué"
Network-on-Chip (NoC) introduces parallelism in communications and emerges with the growing integration of circuits as large designs need scalable communication architectures. This introduces the separation between communication tasks and processing tasks, and makes the design with NoC more complex. High level synthesis (HLS) tools can help designers to quickly generate high quality HDL (Hardware Description Level) designs. But their control schemes are centralized, usually using finite state machines. To take benefit from parallel algorithms and the ever growing FPGAs, HLS tools must (1) properly extract the parallelism from the input representation and (2) use the available resources efficiently. Even if many HLS tools use data flow representations as input, algorithm designers are used with programming languages. But this behavioral specification has to be enriched with architectural details for a correct optimization of the generated design. The C to FPGA path is not straightforward, and the need for architectural knowledges limits the adoption of FPGAs, and more generally, parallel architecture. In this thesis, we present a method that uses a symbolic execution technique to extract the parallelism of an algorithmic specification written in a high level language. Parallelization skills are not required from the users. A methodology is then proposed for adding NoCs in the automatic design generation that takes the benefit of potential parallelizations. To dimension the design, we estimate the design resource consumption using a mathematical model for the NoC. A scalable application, hardware specific, is then generated using a High Level Synthesis flow. We provide a distributed mechanism for data path reconfiguration that allows different applications to run on the same set of processing elements. Thus, the output design is programmable and has a processor-less distributed control. This approach of using NoCs enables us to automatically design generic architectures that can be used on FPGA servers for High Performance Reconfigurable Computing. The generated design is programmable. This enable users to avoid the logic synthesis step when modifying the algorithm if a existing design provide the needed operators.
Le jury sera composé de :
- Loic Lagadec, Professeur, ENSTA Bretagne, président
- Amer Baghdadi, Professeur, Lab-STICC, rapporteur
- Camel Tanougast, Professeur, LCOMS, rapporteur
- Pascal Remy, Docteur, Adacsys, examinateur
- Frédéric Rousseau, Professeur, Lab.TIMA, directeur de thèse
- Virginie Fresse, Maître de conférence, LabHC, co-directrice de thèse
La soutenance sera en français.