基于GPU的快速色阶映射算子实现

摘要

To improve the efficiency of tone mapping computing, this paper designed a fast implementation algorithm based on GPUs. First combined with basic reduce algorithm and the parallel computing characteristic of GPUs, the algorithm proposed two kernel functions to compute maximum luminance. Then it computed pixel-centered area average luminance using shared area intermediate result. Besides,for video stream processing, it proposed a solution to the unmatched problem of CPU reading and GPU computing by a texture pool and a heuristic method to update the global maximum luminance according to the value of a pixel set. Experiment proves that it obtains 4-5 efficiency improvement relative to CPU implementations. The algorithm can take full advantage of GPUs' parallel computing capability and reduce much duplicate computation. It's sufficient for realtime rendering and satiable for textures with different scales.%为了提高色阶映射计算的效率,设计了基于GPU的快速色阶映射算法.首先结合基本规约算法和GPU的并行运算特征设计了基于两个核函数的最大亮度计算方法,然后通过区域中间值共享计算以像素为中心的区域平均亮度,最后针对视屏处理,提出利用纹理缓存池解决CPU读数据和GPU处理数据速度不匹配的问题,并根据像素子集最大亮度自适应地更新全局最大亮度.实验结果相对相同算法的CPU实现得到了4～5倍的速度提升,表明所提出的算法能够充分利用GPU的并行性,并减少了大量重复运算,满足实时渲染的要求,并且对不同规模的纹理具有良好的适应性.