关键词: 并行可视化, 并行图像合成, 图像压缩, 最小通信开销

Abstract: Sort-last is the most widely used method for large scale parallel visualization, and the bottleneck of sort-last method is the image compositing stage. Direct Send is the cornerstone for all other compositing algorithms, so it makes a lot of sense to improve its performance for accelerating image composition. To minimize the communication cost in image compositing, we propose a new type of Direct Send method. Compared with the static partition strategy of the traditional Direct Send, our method is dynamic and adaptive, and it is composed of two phrases: Firstly, we compute all active pixel prefix sums of each image by GPU multi-threads in parallel. This process can remove the background pixels tremendously, and the images are all compressed efficiently. Secondly, a dynamic programming model is built and solved to generate the optimal partitions of subimages for Direct Send image compositing, which ensures that the communication costs of Direct Send is minimal. In the experiments, we firstly measure the image compression ratios of our method, and obtain the optimal size of pixel blocks. Then, we compare the compositing time of our method with RLE and greedy algorithm on varying number of rendering nodes, showing that our method is more efficient than the existing two methods.

Key words: parallel visualization, parallel image compositing, image compression, minimal communication cost