high performance computing on graphics processing units: hgpu.org

When implementing a function mapping on the contemporary GPU, several contradictory performance factors have to be balanced. Previously a decomposition-fusion scheme was devised to guide such an implementation and this work is here further elaborated. To ease this process, an automatic source-to-source compiler is presented, while the main subject of this thesis are the core algorithms for generation, pruning and search in the state-space of possible implementations of the mapped function. The performance of the generated implementation is evaluated together with the overall complexity of the optimization process.