high performance computing on graphics processing units: hgpu.org

Heterogeneous multi-core architectures have a higher performance/power ratio than traditional homogeneous architectures. Due to their heterogeneity, these architectures support diverse applications but developing parallel algorithms on these architectures can be difficult. In implementing algorithms for heterogeneous systems, proprietary languages are often required, limiting portability. Although general purpose graphics processing units (GPUs) have shown great promise in accelerating the performance of throughput computing applications, it is still limited by the memory wall. The memory wall can greatly affect application performance for problems that incorporate amorphous parallelism or irregular workload. Now, AMD’s Fusion series of Accelerated Processing Units (APUs) attempts to solve this problem. The APU is a radical change from the traditional systems of a few years ago. This design change enables consumers to have a capable CPU connected to a powerful, compute-capable GPU using a Very Long Instruction Word (VLIW) architecture. In this thesis, I address the suitability of irregular workload problems on APU architectures. I consider four scientific computing problems of varying characteristics and map them onto the architectural features of the APU. I develop several software optimizations for each problem by making effective use of VLIW static scheduling through techniques such as loop unrolling and vectorization. Using AMD’s OpenCL profiler, I analyze the execution of the various optimizations and provide an in-depth performance analysis using metrics such as kernel occupancy, ALUFetchRatio, ALUBusy Percentage and ALUPacking. Finally, I show the effect of register pressure due to vectorization and the limitations associated with the APU architecture for irregular workloads.