high performance computing on graphics processing units: hgpu.org

Heterogeneous architectures become dominant in high-performance computing platforms but programming them remains really hard, especially because high-performance programs are usually written using low-level languages (C, Fortran, OpenMP…) and frameworks (CUDA, OpenCL…). Mid-level frameworks have been introduced to automatically perform management of distributed memory and scheduling on different devices, allowing applications to only submit tasks and their dependencies (forming a task graph) to get them executed efficiently. However, they remain quite lowlevel and they don’t allow static checking and optimizations to be performed on dynamically constructed task graphs. We propose a high-level way to express task graphs inspired by concepts from functional programming and we describe how a runtime system supporting it should work. Our approach allows checks and transformations to be performed statically as well as dynamically on task graphs. In addition, we mention how functional paradigm concepts could be used to further enhance highperformance computing frameworks on heterogeneous and manycore architectures.