Flexible Data Movement API for GPUs

The Graphics Processing Unit is designed to manipulate plenty of memory fast. To use its full capacity, a deeper understanding of the underlying architecture is required. This thesis presents a simple but still flexible Copy API to move N-dimensional data fragments between memory spaces in a GPU efficiently. We introduce different strategies to divide fine-grained parallelism over a user given workload. These strategies are then benchmarked to show their possible performance variety. In a last step, we display the use of the Copy API on different algebraic applications, highlighting the advantages of access to simple and flexible data movement functions