ParaC:面向GPU平台的图像处理领域的编程框架

发布时间：2019-01-04 08:00

【摘要】：GPGPU加速器是当前提高图像处理算法性能的主流加速平台,但在GPGPU平台上,同一个程序充分利用硬件体系结构特征和软件特征的优化版本与简单实现版本在性能上会有数量级的差异.GPGPU加速器具有多维多层的大量执行线程和层次化存储体系结构,后者的不同层次具有不同的容量、带宽、延迟和访问权限.同时,图像处理应用程序具有复杂的计算操作、边界处理规则和数据访问特性.因此,任务的并发执行模式、线程的组织方式和并发任务到设备的映射不仅影响到程序的并发度、调度、通信和同步等特性,而且也会影响到访存的带宽、延迟等.因此,GPGPU平台上的程序优化是一个困难、复杂且效率较低的过程.提出基于语言扩展的领域编程模型:Para C.Para C编程环境利用高层语言扩展描述的程序语义信息,自动分析获取应用程序的操作信息、并发任务间的数据重用信息和访存信息等程序特征,同时结合硬件平台特征,利用基于领域先验知识驱动的编译优化模型自动生成GPGPU平台上的优化代码,最后,利用源源变换编译器生成标准Open CL程序.在测试用例上的实验结果表明,ParaC在GPGPU平台上自动生成的优化版本相对于手工优化版本的加速比最高达到3.22倍,但代码行数只是后者的1.2%~39.68%.
[Abstract]:GPGPU accelerator is the main acceleration platform to improve the performance of image processing algorithms, but on the GPGPU platform, The optimized version of the same program makes full use of the hardware architecture features and software features and the simple implementation version will have several orders of magnitude in performance. The GPGPU accelerator has multi-dimensional and multi-layer execution threads and hierarchical storage architecture. The latter has different levels of capacity, bandwidth, latency, and access rights. At the same time, image processing applications have complex computing operations, boundary processing rules and data access characteristics. Therefore, the concurrent execution mode of tasks, the organization of threads and the mapping of concurrent tasks to devices not only affect the concurrency degree, scheduling, communication and synchronization of programs, but also affect the bandwidth and latency of visiting memory. Therefore, program optimization on GPGPU platform is a difficult, complex and inefficient process. This paper presents a domain programming model based on language extension: Para C.Para C programming environment uses program semantic information described by high-level language extensions to automatically analyze and obtain application operation information. The program features such as data reuse information and memory access information among concurrent tasks, combined with hardware platform features, are automatically generated on GPGPU platform using domain-based prior knowledge driven compilation optimization model. The standard Open CL program is generated by the source conversion compiler. The experimental results on the test cases show that the speedup of the optimized version automatically generated by ParaC on GPGPU platform is 3.22 times higher than that of the manual optimized version, but the number of lines of code is only 1.2% and 39.68% of the latter.
【作者单位】：体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学;
【基金】：国家自然科学基金(61432018,61402445,61502452,61602443,61432018) 国家重点研发计划(2016YFB1000402) 数学工程与先进计算国家重点实验室开放基金(2016A03) 北京市科委计划(D161100001216002)~~
【分类号】：TP391.41

【相似文献】