PhD graduated Team : APR
Localisation : JussieuUPMC - LIP6 Boîte courrier 169 Couloir 26-00, Étage 3, Bureau 307 4 place Jussieu 75252 PARIS CEDEX 05 FRANCE Tel: +33 1 44 27 62 42,
Mathias.Bourgoin (at) nulllip6.fr
Supervision : Emmanuel CHAILLOUX
Opengpu Project - Efficient abstractions for graphics cards
Graphics Processing Units (GPUs) are complex devices with many computation units. Dedicated to display management and 3D processing, they are very efficient, but also highly specialized. Since recent years, it is possible to divert their use to enable them to perform general computations normally performed by the CPU of the computer. This programming model, GPGPU (General Purpose GPU) programming is mainly based on two frameworks : Cuda and OpenCL. Both are very low-level and demands explicit management of hardware parameters such as the memory or the placement of computations on the various computation units.
The goal of this thesis is the study of solutions of higher level of abstraction for GPGPU programming, in order to make it more accessible and safer.
After an introduction to the context of GPGPU programming, we presnet two programming languages dedicated to GPGPU programming, SPML and Sarek. Through their operationnal semantics, we discuss their properties and the guarantees they offer. Then, we present an implementation of these languages with OCaml through the SPOC library and the domain specific language, Sarek. Performance tests show that our solution achieves a high level of performance for simple examples, as well as with the translation of a realistic numerical application from Fortran and Cuda, to OCaml. We also show how our solutions allow to define algorithmic skeletons that offer more abstractions. Through an example, we present how these skeletons eases GPGPU programming and offers additional automatic optimizations. Finally we discuss how the current hardware and software evolution can help providing a unified solution for GPGPU programming.
Defence : 12/11/2013 Jury members : M. Jocelyn Sérot (Professeur, Université Blaise Pascal) [Rapporteur]
M. Marco Danelutto (Associate Professor, Universita di Pisa, Italie) [Rapporteur]
M. Emmanuel Chailloux (Professeur, UPMC)
M. Jean-Luc Lamotte (Professeur, UPMC)
M. Jocelyn Sérot (Professeur, Université Blaise Pascal)
M. Carlos Agon (Professeur, IRCAM)
M. Joël Falcou (Maître de Conférences, Université Paris-Sud)
M. Gaétan Hains (Professeur, Université Paris-Est Créteil Val de Marne)
M. Stéphane Vialle (Professeur, Supélec)
M. Bourgoin, B. Canou, E. Chailloux, A. Jonquet, Ph. Wang : “OC4MC: Objective Caml for Multicore Architectures”, Draft of 21st Symposium on Implementation and Application of Functional Languages, vol. SHU-TR-CS-2009-09-1, South Orange États-Unis, pp. 24-41 (2009)