ATTENTION/WARNING - NE PAS DÉPOSER ICI/DO NOT SUBMIT HERE

Ceci est la version de TEST de DIAL.mem. Veuillez ne pas soumettre votre mémoire sur ce site mais bien à l'URL suivante: 'https://thesis.dial.uclouvain.be'.
This is the TEST version of DIAL.mem. Please use the following URL to submit your master thesis: 'https://thesis.dial.uclouvain.be'.
 

High-level synthesis of TensorFlow neural networks

Van Brandt, Carl-Louis
(2020)

Files

VanBrandt_42111400_2020.pdf
  • UCLouvain restricted access
  • Adobe PDF
  • 2.71 MB

Details

Supervisors
Legat, Jean-Didier ; De Vleeschouwer, Christophe
Faculty
Ecole polytechnique de Louvain
Degree label
Master [120] : ingénieur civil électricien
Abstract
High-level synthesis tools provide high-performance RTL design of high-level code with a low time-to-market. However, the limited support of their input language and third-party libraries makes their use limited in the context of neural network design. These neural network models are usually designed using specialised frameworks that allow for quick prototyping and fast design space exploration but that do not often directly support hardware synthesis. This work studies LeFlow, a workflow that combines these two worlds by translating models written with the TensorFlow library into a RTL representation using the LegUp HLS tools. This work attempts to modify LeFlow to work with the latest commercial version of LegUp in order to profit from its long-term support. Despite technical issues and limitations, the modified version of LeFlow already reports a speed-up factor of up to 45 for the execution of standard neural networks layers compared to its previous implementation.