high performance computing on graphics processing units: hgpu.org

This thesis intends to show how to efficiently exploit the parallelism present in applications in order to enjoy the performance benefits that multiprocessors can provide, using a new automatic task parallelization methodology for compilers. The key characteristics we focus on are resource constraints and static scheduling. This methodology includes the techniques required to decompose applications into tasks and generate equivalent parallel code, using a generic approach that targets both different parallel languages and architectures. We apply this methodology in the existing tool PIPS, a comprehensive source-to-source compilation platform. This thesis mainly focuses on three issues. First, since extracting task parallelism from sequential codes is a scheduling problem, we design and implement an efficient, automatic scheduling algorithm called BDSC for parallelism detection; the result is a scheduled SDG, a new task graph data structure. In a second step, we design a new generic parallel intermediate representation extension called SPIRE, in which parallelized code may be expressed. Finally, we wrap up our goal of automatic parallelization in a new BDSC- and SPIRE-based parallel code generator, which is integrated within the PIPS compiler framework. It targets both shared and distributed memory systems using automatically generated OpenMP and MPI code.