Meta-programming and Multi-stage Programming for GPGPUs

Ian Masliah (1) , Marc Baboulin (1) , Joel Falcou (1)

1 ParSys - LRI - Systèmes parallèles (LRI)

GPGPUs and other accelerators are becoming a mainstream asset for high-performance computing. Raising the programmability of such hardware is essential to enable users to discover, master and subsequently use accelerators in day-today simulations. Furthermore, tools for high-level programming of parallel architectures are becoming a great way to simplify the exploitation of such systems. For this reason, we have extended NT2 (Numerical Template Toolbox), a C++ scientific computing library which can generate code for SIMD and multi-threading systems in a transparent way. In this paper, we study how to introduce an accelerator-based programming model into this library to allow developers to reap the benefits of such an architecture from a simple, MATLAB like code. After a brief description of the NT2 framework, we explain how our accelerator programming model has been designed and integrated in a pure C++ library. We conclude by showing the applicability and performance of this tool on some practical applications.

Domaines: Calcul parallèle, distribué et partagé [cs.DC]; Architectures Matérielles [cs.AR]

Ian Masliah, Marc Baboulin, Joel Falcou. Meta-programming and Multi-stage Programming for GPGPUs. 10th IEEE International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSOC 2016), IEEE Computer Society, Sep 2016, Lyon, France. pp.369 - 376, ⟨10.1109/MCSoC.2016.49⟩. ⟨hal-01416797⟩