high performance computing on graphics processing units: hgpu.org

We show that case-based reasoning (CBR) and deterministic code analysis can be successfully used in optimizing compilers of unstructured mesh applications to obtain better execution times. With the recent shift of CPU architectures towards SIMD capabilities, and of GPU architectures towards general purpose computing, it is no longer clear what optimizations are optimal given a particular problem and target architecture. As a result, we explore the use of machine learning and deterministic algorithms on OP2 C++ Airfoil variations to determine whether such methods can provide optimal or near-optimal results. Our choice of optimizations are loop fusion and runtime parameter variation (block size, partition size and warpsize). The new perspectives we are exploring in this project are determining optimisations by looking at OP2 code and user kernel complexity, irrespective of low-level architecture details, the integration of a CBR system and deterministic methods to significantly prune our search space and our focus on multiple heterogeneous architectures (CPUs, GPUs).