Compressed Volume Rendering Based On GPU General-purpose Computing

GPU based direct volume rendering for large-scale volume data is often limitedby the amount of available video memory and the bandwidth between main memoryand video memory. This causes the frequent data interaction from main memory tovideo memory, and becomes the bottleneck of increasing the drawing speed. CVR(compressed volume rendering) which combines the data compressing and renderingprocess can solve the problem effectively. In this paper, we focus on CVR algorithmof large-scale volume data that is suitable for GPU and CVR accelerating technology.The main contributions of this paper can be outlined as follows.Firstly, in order to solve the growing capacity contradiction between videomemory and large-scale volume data, we propose a strategy of CVR for large-scalevolume data. First of all, apply3D wavelet transform to volume data. Then, thesub-bands of transformed data are analyzed by histogram and divided into severalgroups. We use classified vector quantization to encode and compress the data. Whenrendering, a GPU-based ray casting algorithm is adopted and only few current neededdata is decompressed and transformed, which saves the video memory. Compared toonly introduction of the vector quantization algorithm, this strategy can get betterimage quality. But the usage of wavelet transform consumes a certain amount of GPUcomputing resources and limits the rendering speed. Other accelerating strategies forthe rendering process should be further presented.Additionally, through comprehensively taking the factors that affect therendering speed in the compressing and rendering strategy into account, two kinds ofLOD model are proposed. Object space based LOD (OSLOD) use the waveletcoefficients of low frequency sub-band as low level of detail of the original data.During the decompressing and rendering process, OSLOD can only do the vectorquantization decoding without wavelet inverse transformation, which improve therendering speed. Image space based LOD (ISLOD) get low resolution of image bydecrease the number of rays, and then use nearest neighbor interpolation algorithm toenlarge the image in real time. At last, manage LOD by detecting the Perspectivetransformation and rendering frame rate.Finally, the CVR and LOD accelerating strategy are implemented and tested on avariety of experimental platforms. The results show that we can obtain better image quality and realize interactive rendering of large scale volume data while the capacityof compressed data was a little smaller than the video memory. Applying the CVRstrategy into a flexible and compatible seismic data processing system, theexperimental results show that the method can preload multi-data and speed up therendering of multi-data in multi-window.