基于CPU_MIC_GPU异构架构的Roberts算法优化探究

The Research of Optimized Roberts Algorithm Based on Heterogeneous CPU/MIC/GPU Architecture

在高分辨率图像日益普及的情况下,Roberts边缘检测的处理速度急需进一步提高。在CPU表现不尽如人意的情况下,基于CPU/GPU和CPU/MIC的高度并行运算的研究愈加深入。在分析Roberts算法特点的基础上,将能并行的部分移植到GPU和MIC上进行。完成基于CPU/GPU和CPU/MIC的异构架构上的Roberts算法实现,并针对CPU/MIC上将程序进行向量化优化。实验结果表明,在相同单精度浮点运算能力下,GPU处理低分辨率图像的速度更快、加速比更高,但处理高分辨率图像时MIC的加速比最高为23.52,高于GPU的21.43。

In the case of the growing popularity of high-resolution images, the Roberts edge detection processing speed urgent need to further improve. Because the disappointing performance of CPU, the height parallel computing research based on the CPU/GPU and CPU/MIC become deepened. On the basis of analysis of the characteristics of Roberts algorithm, the part of the parallel will be respectively transplanted into GPU and MIC. Accomplish the Ro- berts algorithm which is based on CPU/GPU and CPU/MIC in a Heterogeneous architecture, and optimization the procedure aim at CPU/MIC. Experimental results show that, in the same single-precision FLOPS, the GPU pro- cessing low resolution image than faster, higher, but the MIC when dealing with high resolution images faster, speed is higher.