采用SIMT结构GPU的二维离散哈尔小波变换的优化

Optimization of two-dimensional discrete Haar wavelet transform using SIMT GPU

摘要:

为了解决CPU环境下小波变换在运行时对高分辨率图片处理速度较慢的问题，利用GPU有大量可编程核心的特点，针对二维离散哈尔小波变换进行了在SIMT(单指令多线程)体系结构GPU环境下的并行推导，同时调整GPU的逻辑布局，将数据分割，更改了数据同步方式，并且采用了虚拟寻址，将速度进一步提升到了0.92ms，比CPU环境下效率提升51.1%，比SIMD架构效率提升16.3%，效果显著，满足实时性要求。

Abstract:

Wavelet transform in order to solve the CPU environment at runtime for high resolution image processing speed slow problem, using the GPU has the characteristic of a large number of programmable core, in view of the two-dimensional discrete wavelet transform Hal in SIMT (single instruction multithreading) under the environment of architecture of GPU parallel deduction, at the same time, adjust the GPU logical layout, data segmentation, The data synchronization mode is changed, and virtual addressing is adopted, which further increases the speed to 0.92ms, 51.1% higher than that in CPU environment, and 16.3% higher than that in SIMD architecture. The effect is remarkable and meets the real-time requirements.