基于雷达资料的外推是临近预报中重要的方法之一，随着全国气象雷达网络建设规模的不断提高以及观测资料精细化程度的提升，基于区域乃至全国雷达拼图的外推预报，每次计算都需花费大量时间，甚至滞后于每6分钟一次的资料观测频次。为解决传统外推算法运算复杂度高，实时性差的问题，运用OpenCL构建基于GPU的异构计算模型对外推算法进行并行化改进。然后逐步分析影响算法性能的瓶颈，并通过改进和测试数据比对，阐述算法优化的过程。其中，内存与线程的映射优化、合理利用局部存储器作为高速缓存以及隐藏CPU执行时间等方法不仅对本算法的执行效率带来显著提升，也可为其他基于OpenCL异构计算的优化提供参考。以AMD Graphic Core Next 和Northern Islands二代GPU架构作为测试平台，并以Intel CPU并行计算作为测试参考，测试结果表明，改进后的算法在硬件同等功耗的情况下，计算性能提升15～22倍。%Extrapolation based on radar data is one of the important methods for weather nowcasting .With the increasing scale of the national weather radar network construction , and the enhancing about the refinement of meteorological observational data , the extrapolation forecast based on regional and even national radar puzzle , the computation time is very long .The time waiting for extrapolation computation usually lags behind the data ’ s observation frequency which is once every six minutes .To solve the problem that the traditional extrapolation algorithm is of high computational complexity and poor real -time, this paper discusses the heterogeneous computing model based on GPU , and presents a parallel algorithm with OpenCL to achieve high performance , then analyzes the bottlenecks of this application , and discusses how to bring up the computation speed by algorithm process im-provement and test data comparison .Some methods such as optimizing the mapping relationship of memory and threads , utilizing local memory as high speed cache , and hiding CPU execution time , not only bring the efficiency of the algorithm significantly im-proved, but also provide a reference for other optimization based on OpenCL heterogeneous computing .Using AMD Graphic Core Next and Northern Islands which are two generation GPU architectures as test platforms , and using Intel CPU parallel computing as a test reference , the test results show that the improved algorithm consuming the same power dissipation under different hard -ware, the computing performance is improved 15-22 times.
展开▼
