Automated Code Generation for Lattice QCD Simulation
Résumé
Quantum Chromodynamics (QCD) is the theory of strong nuclear force, responsible of the interactions between sub-nuclear particles. QCD simulations are typically performed through the lattice gauge theory approach, which provides a discrete analytical formalism called LQCD (Lattice Quantum Chromodynamics). LQCD simulations usually involve generating and then processing data on petabyte scale which demands multiple teraflop-years on supercomputers. Large parts of both, generation and analysis, can be reduced to the inversion of an extremely large matrix, the so-called Wilson-Dirac operator. For this purpose, and because this matrix is always sparse and struc-tured, iterative methods are definitely considered. Therefore, the procedure of the application of this operator, resulting in a vector-matrix product, appears as a critical computation kernel that should be optimized as much as possible. Evaluating the Wilson-Dirac operator involves symmetric stencil calculations where each node has 8 neighbors. Such configuration is really hindering when it comes to memory accesses and data exchanges among processors. For current and future generation of supercomputers the hierarchical memory structure make it next to impossible for a physicist to write an efficient code. Addressing these issues in other to harvest an acceptable amount of computing cycles for the real need, which means reaching a good level of efficiency, is the main concern of this paper. We present here a Domain Specific Language and corresponding toolkit, called QIRAL, which is a complete solution from symbolic notation to simulation code.
Domaines
Théorie et langage formel [cs.FL]
Origine : Fichiers produits par l'(les) auteur(s)
Loading...