high performance computing on graphics processing units: hgpu.org

Efficient parallel algorithms are often written with embedded knowledge of the back-end that they are meant to be executed on, and if they are not, the translation to target language often produces inefficient code. A concrete problem is space complexity in nested data parallel (NDP) languages such as NESL and Data Parallel Haskell, where large intermediate arrays are often allocated during execution. This thesis presents an NDP language with a streaming based model where the time complexity of programs is just as good as in traditional NDP languages, but the space complexity is significantly better in many cases. A minimal NDP language with semantics and a desirable cost model is defined, as well as a streaming based target language, and the two languages are related with a translation, a proof-of-concept implementation and a conjecture about value and cost preservation.