Heterogeneous Platform Research Based On Streamit Compiler

The heterogenous platform based on manycore and multicore becomes prevalent.The efcient parallel programming model which can support heterogenous platformbecomes an important research. Streamit is a kind of parallel programming languagedeveloped by MIT which can support multicore or cluster. This thesis extends Streamitto heterogenous platform. First, we deep analyze Streamit benchmarks and the loopstatements. Second, we conduct not only the scheme feasibility analysis, but also datadependency analysis. Third, we study the data transmission strategy and GPU opti-mization. Based on above research, The below results are obtained.(1) The heterogenous programming model(Streamit+GPU) is proposed. Thismodel aims mainly at the highly parallel loop statements.(2) Separation strategy for computing and data operation is proposed. This sepa-rates a loop block into two independent parts: one is a parallel loop block and the otheris a sequential data operation.(3) The communication pattern between CPU and GPU is proposed in heteroge-nous environment. This pattern can better support data transmission.(4) The optimization strategy for GPU is proposed. More research about GPUfocuses on optimization because of its complex memory hierarchy.(5) Several optimization methods are proposed to improve the pipeline perfor-mance in Streamit program. The first method removes pipeline bottleneck by not usingshare operation. The second method proposes multiple output and multiple input tomake Streamit more efcient in heterogonous environment. This breaks out the restric-tion of the filter port number.