GPU并行程序自动优化方法研究

As a general-purpose processor, GPU has been widely used. However, it is very difficult to design a GPU parallel program with high performance. The programer must be very familiar with the underlying architecture of GPU. Meanwhile the process of optimizing parallel codes is very time -consuming and error-prone. To address this problem, the project intends to study the method of automaticly generating parallel codes from Serial codes and also the method to optimize these codes automaticly. Our approach is to improve OpenACC’s functionality which can make its parallel strategies more diverse, and generate different parallel codes via variant parameters. Then these codes should be analyzed by evaluation model. Acording to this parallel program performance evaluation, parameter sets should be changed to generate a more optimal parallel program. Repeating this process until we find the most optimal parallel source codes. The study work of this project is list as follows: 1) How to increase the directive and API in OpenACC to generate codes that meet those GPU hardware constraints. For example, the parallel program can satisfy coaleced memory access requirment or effictively use hierarchy GPU memory; 2) Considering pipeline components and impact factors how to establish analytical model which can accurately evaluate the performance of GPU programs; 3)Based on parallel program evaluation results, how to search a better parameter sets to generate more optimal parallel codes.

GPU作为通用计算处理器已得到了广泛的应用，但设计高质量的GPU并行程序却非常困难，它要求程序开发者对GPU底层硬件非常熟悉，并且GPU并行程序优化过程耗时且易出错。针对这个问题，本项目拟研究从串行程序到GPU并行程序的自动生成和自动优化方法。我们设计的方法是扩展OpenACC功能，使其并行策略更为多样，且可以通过参数修改得到不同的并行程序，然后对生成的并行程序进行性能评价，最后依据评价的值改变参数生成更优的并行程序，如此迭代直到找到最优并行程序。项目拟研究的问题包括：1）如何通过增加指导语句和API使OpenACC生成的代码满足GPU的硬件约束，例如可以自动实现合并存取、实现多级存储器规划等；2）怎样综合考虑GPU中多部件流水线的工作方式和影响因子的作用，以建立能够精准评价GPU程序的性能分析模型；3）如何从已有并行程序的评价结果出发，在并行化参数空间中搜索出能产生更优源代码的参数。