Optimized two-level parallelization for GPU accelerators using the polyhedral model

While GPUs play an increasingly important role in today's high-performance computers, optimizing GPU performance continues to impose large burdens upon programmers. A major challenge in optimizing codes for GPUs stems from the two levels of hardware parallelism, blocks and threads; each of these levels has significantly different characteristics, requiring different optimization strategies. In this paper, we propose a novel compiler optimization algorithm for GPU parallelism. Our approach is based on the polyhedral model, which has enabled significant advances in program analysis and transformation compared to traditional AST-based frameworks. We extend polyhedral schedules to enable two-level parallelization through the idea of superposition, which integrates separate schedules for block-level and thread-level parallelism. Our experimental results demonstrate that our proposed compiler optimization framework can deliver 1.8x and 2.1x geometric mean improvements on NVIDIA Tesla M2050 and K80 GPUs, compared to a state-of-the-art polyhedral parallel code generator (PPCG) for GPGPUs.