high performance computing on graphics processing units: hgpu.org

Upcoming heterogeneous systems ask for new programming paradigms. Abstracting the underlying hardware architecture is desirable in order to support productive software development. This thesis proposes a design flow and runtime-system for executing process networks on heterogeneous systems using OpenCL. Process networks are a popular model of computation for deterministic parallel programming and OpenCL is a royalty-free standardised programming interface with a broad support in industry. The proposed design flow consists of a program code synthesis framework for building applications from a generic high-level process network specification. The synthesised application is targeted to a certain OpenCL architecture that was predefined by a high-level specification. The target code is built for this architecture specification integrating reusable building block primitives into it. Those primitives are location-based FIFO channels minimising the number of memory copy operations, a process wrapper that is interconnectable to channels and mirrors the process network functionality, and an extensible task activation framework responsible for interprocess synchronisation. Heterogeneous systems, being inherently parallel computer architectures, demand scalable and parallel applications. To simplify this, a notion of shadow copies was introduced to transparently abstract data-parallelism. Extensive evaluations on two heterogeneous systems have shown that the proposed design flow and runtime-system support a wide range of heterogeneous platforms and parallel applications. Furthermore, the evaluations have proved that the proposed framework is suitable for efficient and productive software development for heterogeneous systems and that it provides enough flexibility so that the programmer can efficiently exploit the parallelism offered by multicore CPUs and GPUs.