Contact : Thomas Carle, Christine Rochange

Email : thomas.carle@irit.fr, christine.rochange@irit.fr

Organisation : IRIT

Année : Master 2

Description du stage

The ongoing adoption of Graphics Processing Units (GPUs) as accelerators for neural network (NN) inference computations in autonomous vehicles raises several challenges. One of them, and not the least, is the capacity of the system designer to guarantee the respect of the real-time constraints of the system. Indeed, mature timing analysis techniques target traditional processors (CPUs) which exhibit few instruction-level parallelism, whereas GPUs implement a completely different execution model, heavily relying on thread-level parallelism (a.k.a. SIMT for Single Instruction Multiple Threads).

As part of the AVATAr project (funded by Labex CIMI), our research team has started to develop a timing simulator for the Nvidia Pascal GPU architecture, which helps us to understand the low-level execution mechanisms of this GPU. Our objective is now to design models and analysis methods for this target in the OTAWA framework, a worst-case execution time (WCET) analysis tool developed in the team, in order to compute the WCET of neural networks (starting with basic fully-connected networks). The goal of this internship is to build upon the knowledge already acquired by the team on the Pascal GPU to design an OTAWA model for the static WCET analysis of programs running on GPU targets. We will start by modelling the mechanisms that handle execution divergence between parallel threads.

We are looking for a highly motivated F/M master 2 student with a background in at least one of the following areas: static analysis, computer architecture, GPU programming, WCET analysis, real-time systems.

Additionally, the applicant will be skilled in C/C++ programming, assembly languages and will have a working knowledge of Linux.

The internship can be followed by a 3 year PhD financed by the MeSCALiNe ANR project, which will aim at generalizing the results of the internship to build a complete and precise WCET analysis model for GPU targets.

The research will be conducted in team TRACES of the IRIT lab in Toulouse, France. TRACES is a research team specialized in embedded systems, computer architectures, compilers and static analysis, with a focus on static timing analysis of programs. It develops and maintains the open-source static analysis framework OTAWA. The team is located in Université Toulouse 3 Paul Sabatier. Toulouse is a very nice city located in the South West of France, and has just been elected best city for students. It offers a lot of opportunities for going out, practicing sports, travelling and eating great french food. Moreover it is the home of Airbus and the European capital for aerospace.

Mots-clé : Analyse de pire-temps d'exécution, GPU

Ressources supplémentaires

Stage M2 analyse statique de pire temps d'exécution de réseaux de neurones sur GPU.pdf

Article proposé par H. Cassé.