Multi-scale stream reduction for volume rendering on GPUs

详细信息    查看全文

• 作者：Yatong Jiang^aAuthor Vitae ; Seungmin Rho^b ; Yingping Zhang^c ; ^{yingpzhang@gmail.com}Author Vitae ; Feng Jiang^d ; Jian Yin ; ^a ; ^{jianyinsdu@126.com} ; ^{clarenceyan@126.com}Author Vitae
• 关键词：GPU ; Marching cubes ; Volume raycasting
• 刊名：Microprocessors and Microsystems
• 出版年：2016
• 出版时间：November 2016
• 年：2016
• 卷：47
• 期：part_PA
• 页码：133-141
• 全文大小：2720 K
• 卷排序：47

文摘

In this paper, we present a uniform acceleration framework for GPU-based interactive visualization of regular scalar fields. Firstly, in order to exploit the coherence of volume fields in both the object space and the image space, we propose a general bi-space rendering proxy (BSRP) to represent volume fields. These BSRP are organized into pointerless tree structures which can index voxels in a multi-scale manner. Based on BSRP, we present a novel multi-scale stream reduction (MSSR) algorithm to rapidly process BSRP-indexed valid voxels (i.e., active voxels in marching cubes or nonempty space in volume raycasting). In the object space, MSSR utilizes pre-computed tree structure to rapidly get rid of invalid voxels using multi-scale BSRP with minimal overhead, and thus can noticeably reduce the complexity of classification, scan and compaction for valid voxels. In the image space, given view parameters, the BSRP containing valid voxels are rasterized in a coarse-scale. Then, MSSR expands them as lossless ray segments for volume raycasting, where both the exterior and interior empty space are skipped. Our framework addresses the acceleration problem by decomposing volume rendering algorithm into several data-parallel stages processing multi-scale stream, which are mapped efficiently to the massively parallel architecture of modern GPUs. Thanks to the proposed MSSR algorithm, our framework is immune to the changes of iso-value, transfer function and view parameters, which is especially efficient in scenarios requiring frequently interactions. Experimental results demonstrate that the performance of our framework outperforms state-of-the-art algorithms.