Parallel Volume Rendering Under Multi-GPU Environment

In Scientific visualization field, volume rendering provides method to observe and analyze data. Volume rendering can visualize obscure data into image representation, and user can analyze internal and implied information through interaction. However, one processor can’t afford enough computing ability when facing large data sets. In order to render large data sets in real time, we have studied parallel volume rendering for large data sets under multi-GPU environment. Our goal is to take advantage of multiple processors’ computational capabilities, solve large data sets’real time rendering problem on high-resolution.Parallel volume rendering technology solve large data sets’ visualization problem through distribute data bricks among each processor in cluster. We use CUDA based Sort-last parallel volume rendering technique, cutting large data sets into small bricks using KD-Tree, each brick of data become a rendering task. Then we distribute these tasks among GPUs and try to balance the workload between them. Finally, we use CPU or GPU method to blend sub-images into final image.We have analyzed the cost of each step of the parallel volume rendering procedure, including sub-image computing time, image data copy time and sub-image blending time. For the workload between GPUs will influence the rendering performance, we use KD-tree to cut data and dynamically distribute data bricks to solve workload balancing problem. For the long time of data copy, we exploit asynchronous parallel between data copy and image computing. To accelerate blending progress, we take advantage of GPU parallel blending algorithm to improve sub-image blending efficiency.Based on these works, we have integrated parallel volume rendering method into a volume data visualization platform, Voxer. Voxer (Volume Explorer) is a general volume data visualization platform, based on which to construct visualization tools specially targeted to medical, geology and meteorology. To extend the platform’s generality and capacitate Voxer to handle large data sets, we add large data sets’ functionality support for Voxer’s data management and flow management module when we implement our prototype system.By testing several large data sets, we show our methods can efficiently render large scale data in real time, and extend the application spectrum of our general visualization platform.