Description
Date depot: 1 janvier 1900
Titre: Compilation de réseaux de processus de Kahn dans le modèle polyédrique
Directeur de thèse:
Albert COHEN (Google)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
On modern processor architectures (multicores with deep cache hierarchies), parallelization and optimization of computationally-intensive programs have the potential of increasing performance by several orders of magnitude. Starting from naive, algorithmic-level source code, current compilers have a hard time achieving the required transformations to reach such levels of performance. The so-called « polyhedral model » is one of the main research directions to address this challenge. It is dedicated to the optimization and automatic parallelization of regular, loop- and array-centric applications. It is a geometrical representation and transformation model of programs, where the program semantics is captured as affine functions and systems of affine inequalities.
This thesis proposes to study and extend the polyhedral model to operate on concurrent programs built of continuous functions defined on unbounded data streams, also known as Kahn process networks. We will then integrate these methods within the GNU Compiler Collection, leveraging an extension of the OpenMP language supporting data-flow concurrency and data streams.
The thesis will be partially funded by the TERAFLUX and CARP European Projects. We will consider shared-memory multicore architectures as well as future and emerging manycore designs, including accelerator hardware abstracted through an OpenCL software stack.
The this will take place at ENS (45 rue d'Ulm), in the PARKAS INRIA team, and possibly at IRILL (Initiative pour la Recherche et l'Innovation sur le Logiciel Libre, 23 avenue d'Italie), where an activity around the GCC platform is being established. The PhD student will collaborate with two PhD students, Feng Li and Tobias Grosser, and two postdocs, Sven Verdoolaege and Antoniu Pop.
Doctorant.e: Baghdadi Riyadh