high performance computing on graphics processing units: hgpu.org

Graphics Processing units have become a valuable support for High Performance Computing (HPC) applications. However, despite the many improvements on the General Purpose GPU, there is still the need of a generic programming model adaptable to the many forms of parallelism that an application can express. The CUDA programming model is widely used on the GPGPU domain, but is very limited in aspects like load balancing and task parallelism. This work introduces a new programming model to be used on general purpose graphics processors. We propose an hybrid model combining tasks and data parallelism which extends the domain of applications that can efficiently make use of graphics processors. We implement a work stealing scheduler to efficiently schedule tasks inside a GPU keeping an even load balance between its multiprocessors. Finally, we evaluate the performance of our work stealing scheduler comparing it with static and list scheduling methods applied to the problems of array transformation and octree partitioning.