Département d'informatique

Permanent URI for this community

http://dspace.univ-batna2.dz/handle/123456789/6

Browse

Browsing Département d'informatique by Author "BAROUDI Toufik"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    OPTIMISATION DES ACCÈS MÉMOIRE POUR LES ARCHITECTURES MULTI-COEURS.
    (2021-01-03) BAROUDI Toufik
    With the aim of increasing performance, the architecture of processors has evolved into multicore and manycore platforms composed of multiple processing units. However, the difficulty of programming these architectures often hinders taking full advantage of the potential of its processing units. This has motivated the compilation research community to develop alternative solutions, the goal of which is to free the programmer from the details of target architectures, while generating programs as efficient as possible. This is the area of research in parallelization and automatic code optimization. In almost all programs, loop nests account for most of the computing time. To optimize the execution of these code parties, the compilers apply transformations to them in order to improve the spatial and temporal locality of the memory accesses. The majority of these transformations are based on the polyhedral model which is a formalism allowing to represent iterations and references to arrays by points with integer coordinates of bounded polyhedra. Optimizing linear algebra programs has been one of the issues that has caught the attention of the code optimization research community for many years. In particular, operations on sparse matrices are part of the key calculation codes in many scientific and engineering applications. Consequently, several alternative storage formats have been proposed in order to store these matrices and to calculate only the non-zero elements. In this thesis, we first introduce a new approach to optimize matrix operations on particular sparse matrices using a dense data structure called the 2d-Packed format. The basic idea is that matrix operations using this new data structure can be automatically optimized and parallelized using source-to-source compilers based on the polyhedral model, such as Pluto. This work has improved the performance of a number of benchmarks in linear algebra. Second, we presented a study on the effect of allocation types, static and dynamic, on benchmark performance using the proposed data structure and certain linear algebra kernels of the PolyBench suite. Keywords : Multicore architecture, Optimization and parallelization of codes, Sparse matrices, Polyhedral model, 2d-Packed format.