high performance computing on graphics processing units: hgpu.org

We present a framework for parallel programming. It consists of a distributed shared memory based simplified programming model, which leaves the application developer to focus mainly on task decomposition. This is a unified model for many-core processors (e.g., CPUs and GPUs), multiple processors on a system, as well as multiple systems. We also present a library implementation as a proof of concept of the model. It efficiently maps tasks to multiple compute engines, performs the required communication and schedules tasks to completion. In addition to convenience, the framework provides a race free programming environment by letting tasks own a partition of the memory. This simplifies programming significantly. We report a number of experiments.