GPGPU性能模型及应用实例分析  

A Performance Model for General-Purpose Computation on GPU

作　　者：韩博[1] 周秉锋[1] 

机构地区：[1]北京大学计算机科学与技术研究所,北京100871

出　　处：《计算机辅助设计与图形学学报》2009年第9期1219-1226,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(60573149)

摘　　要：现代图形处理器(GPU)的高性能吸引了大量非图形应用,为了有效地进行性能预测和优化,提出一种GPU处理通用计算问题的性能模型.通过分析现代GPU并行架构和工作原理,将GPU的通用计算过程划分为数据获取、计算、输出和传输4个并列的阶段,结合程序特点和硬件规格对各阶段进行量化分析,完成性能预测.通过实验分析得出两大性能影响要素:计算强度和访问密度,并将其作为性能优化的基本准则.该模型被用于分析几种常见的图像和视频处理算法在GPU上的实现,包括高斯卷积、离散余弦变换和运动估计.实验结果表明,通过增大计算强度和访问密度,文中优化方案显著地降低了GPU上的执行时间,使得计算效率提升了4～10倍,充分说明了该模型在性能预测和优化方面的有效性.We present an efficient model to analyze and improve the performance of general-purpose computation on graphics processing units （GPGPU）. After analyzing the parallel architecture and massive-thread working mechanism of modern GPUs, we build a performance model on a four-level stream processing pipeline, including data input, computation, output and transfer. We further conclude two key factors for GPGPU applications to achieve high performance： high intensity in both computation and memory access. To demonstrate the effectiveness of our model, we highlight its performance on three typical image processing applications, including Gaussian convolution, DCT and motion estimation. Guided by our performance model, some techniques are put forward to remove the performance bottleneck and achieve 4-10 times performance improvement. In practice, the experimental results fit well with the estimations of our proposed model.

关 键 词：GPU GPGPU 图像处理 性能模型 DCT 卷积 运动估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]