Dresden 2014 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
TT: Fachverband Tiefe Temperaturen
TT 79: Correlated Electrons - Poster Session
TT 79.30: Poster
Mittwoch, 2. April 2014, 15:00–19:00, P2
Optimizing Matrix Product State Codes Using SciPAL: Parallelisation and GPU Portability — •Thomas Köhler1, Johannes Hagemann2, Salvatore R. Manmana1, and Stephan C. Kramer3 — 1Institut f. Theoretische Physik, Universität Göttingen — 2Institut f. Röntgenphysik, Universität Göttingen — 3Max-Planck-Institut f. biophysikalische Chemie, Göttingen
We apply the SciPAL library [1] to matrix product states (MPS), as used in the context of density-matrix renormalization group (DMRG) methods, which are important tools for treating low-dimensional strongly correlated quantum systems as realized in certain materials like quantum magnets (e.g. Azurite) and typically described by Hubbard-like or Heisenberg models. SciPAL (scientific parallel algorithms library) is a C++-based, hardware-independent open-source library, compatible with the widely used finite element library deal.II. By its link to CUDA, it provides the flexibility to extend existing codes to work on graphics processors. SciPAL’s core asset is a user-friendly API to BLAS and NVidia’s CUBLAS which allows to use an operator-based formulation of typical linear algebra operations. We discuss the extension of SciPAL to a domain-specific formulation of the numerical methods for computing the properties of MPS and their performance on graphics cards. A. K. Mitchell et al. [1] SciPAL: Expression Templates and Composition Closure Objects for High Performance Computational Physics with CUDA and OpenMP, S. C. Kramer and J. Hagemann, submitted to ACM TOPC.