节点内无冗余图像合成方法  

Composition without Inactive Pixels in Cluster Node

作　　者：刘华海[1] 王攀[1] 李思昆[1] 蔡勋[1] 王文珂[1] 曾亮[1] 

机构地区：[1]国防科学技术大学计算机学院,长沙410073

出　　处：《计算机辅助设计与图形学学报》2013年第5期646-652,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"九七三"重点基础研究发展计划项目(2009CB723803);国家自然科学基金(61170157);xxx项目(2011AA7123022)

摘　　要：为了提高图像合成性能,提出一种基于GPU加速的节点内多GPU图像高效合成方法.首先通过判断帧缓存是否更新生成有效像素表,然后采用GPGPU计算生成有效像素合成索引列表,最后用其指导CPU端图像合成操作,从而完全避免了CPU端的冗余合成计算.理论分析表明,在理想负载平衡条件下,该方法的加速比为图像有效像素百分比与节点内GPU数量的比值.实验结果表明,在节点内配置4个GPU时,针对有效像素比为12%～76%的高分辨率图像,该方法比全图像合成方法的性能提高了3～5倍.A novel GPU acceleration method was proposed in this paper to improve performance of image composition in multi-GPU cluster node. The proposed method constructed an active pixel list by judging whether frame buffer pixels were updated. An index list was computed from the active pixel list using the GPGPU technology. Then the index list was used to conduct pixels composition on CPU. Given these operations, inactive pixel compositions could be totally avoided. Theoretical analysis show that the speedup of full image composition is determined by how many GPUs are configured in node and the proportion of active pixels to the full image. In the node with 4 GPUs, the speedup was about 3-5 times when composited different images whose active pixels was about 12% to 76%.

关 键 词：图像合成 多GPU GPGPU 并行绘制 sort—last 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]