Abstract: |
|
Programming models based on directives or language extensions or the evolution of the language itself offer hope to domain experts wishing to utilize the massive parallelism inherent in accelerators. Several models have emerged which layer on top of standard C, C++ and Fortran through either standards committees or the introduction of proposed de facto standard solutions by large industry players. Each panel participant will discuss the status and merits of their approach to heterogeneous parallel programming, to better manage the complexity, scalability and portability challenges, that current and future HPC architectures will present to us. Each speaker will be given ~6 minutes to provide status of their X and any proof points that it delivers performance portability across the major processor architectures (multicore, manycore, GPU) and how the mass of existing applications must adapt, while avoiding custom code for each architecture. The balance of the time will be reserved for audience questions and participation. The list of speakers includes Michael Wolfe (OpenACC), Bronis de Supinski (OpenMP), Christian Trott (KOKKOS), Thomas Schulthess (C++17), and Peter Meßmer (CUDA).
Targeted Audience This BoF targets a large variety in audience, ranging for programmers all the way to decision makers. The BoF addresses the pressing question of what to use for intra-node parallelization for large projects where porting/refactoring is a multi-year effort. The BoF also targets educators who want to know what to teach to students to prepare them for work. |
|