Optimisation du fonctionnement d'un générateur de hiérarchies mémoires pour les systèmes de vision embarquée

Abstract : The research of this thesis focuses on the application of the Operations Research (OR)methodology to design new optimization algorithms to enable low cost and efficient embeddedvision systems, or more generally devices for multimedia applications such as signal and imageprocessing.The design of embedded vision systems faces the “Memory Wall” challenge regarding thehigh latency of memories holding big image data. For the case of non-linear image accesses, onesolution has been proposed by Mancini et al. (Proc. DATE 2012) in the form of a software tool,called Memory Management Optimization (MMOpt), that creates an ad-hoc memory hierarchiesfor such a treatment. It creates a circuit called a Tile Processing Unit (TPU) that containsthe circuit for the treatment. In this context, we address the optimization challenge set by theefficient operation of the circuits produced by MMOpt to enhance the 3 main electronic designcharacteristics. They correspond to the energy consumption, performance and size/productioncost of the circuit.This electronic problem is formalized as a 3-objective scheduling problem, which is called3-objective Process Scheduling and Data Prefetching Problem (3-PSDPP), reflecting the 3 mainelectronic design characteristics under consideration. To the best of our knowledge, this problemhas not been studied before in the OR literature. A review of the state of the art, including theprevious work proposed by Mancini et al. (Proc.DATE, 2012) as well as a brief overview onrelated problems found in the OR literature, is then made. In addition, the complexity of someof the mono-objective sub-problems of 3-PSDPP problem is established. Several resolutionapproaches, including exact methods (ILP) and polynomial constructive heuristics, are thenproposed. Finally, the performance of these methods is compared, on benchmarks available inthe literature, as well as those provided by Mancini et al. (Proc.DATE, 2012), against the onecurrently in use in the MMOpt tool.The results show that our algorithms perform well in terms of computational efficiency andsolution quality. They present a promising track to optimize the performance of the TPUs producedby MMOpt. However, since the user’s needs of the MMOpt tool are contradictory, such aslow cost, low energy and high performance, it is difficult to find a unique and optimal solutionto optimize simultaneously the three criteria under consideration. A set of good compromisesolutions between these three criteria was provided. The MMOpt’s user can then choose thebest compromise solution he wants or needs.
Document type :
Theses
Complete list of metadatas

Cited literature [183 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01913131
Contributor : Abes Star <>
Submitted on : Tuesday, November 6, 2018 - 9:23:05 AM
Last modification on : Friday, February 22, 2019 - 5:18:11 PM
Long-term archiving on : Thursday, February 7, 2019 - 12:54:46 PM

File

HADJ_SALEM_2018_diffusion.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01913131, version 1

Collections

STAR | LCIS | UGA

Citation

Khadija Hadj Salem. Optimisation du fonctionnement d'un générateur de hiérarchies mémoires pour les systèmes de vision embarquée. Intelligence artificielle [cs.AI]. Université Grenoble Alpes, 2018. Français. ⟨NNT : 2018GREAM023⟩. ⟨tel-01913131⟩

Share

Metrics

Record views

96

Files downloads

76