Craig M. Wittenbrink

Opacity-weighted color interpolation, for volume sampling

Volume rendering creates images from sampled volumetric data. The compute intensive nature of volume rendering has driven research in algorithm optimization. An important speed optimization is the use of preclassification and preshading. The authors demonstrate an artifact that results when interpolating from preclassified or preshaded colors and opacity values separately. This method is flawed, leading to visible artifacts. They present an improved technique, opacity-weighted color interpolation, evaluate the RMS error improvement, hardware and algorithm efficiency, and demonstrated improvements. They show analytically that opacity-weighted color interpolation exactly reproduces material based interpolation results for certain volume classifiers, with the efficiencies of preclassification. The proposed technique may also have broad impact on opacity-texture-mapped polygon rendering.

R-buffer: a pointerless A-buffer hardware architecture

We present a graphics hardware architecture that implements Carpenters A-buffer. The A-buffer is a software renderer that uses pointer based linked lists. Our pointerless approach computes order independent transparency for any number of layers with minimal hardware complexity. Statistics are shown for a variety of different scenes using a trace based methodology, with an instrumented Mesa OpenGL implementation. The architecture is shown to require from 2.1 to 3.6 times more memory than traditional Z-buffering. A detailed hardware design is provided. Order independent transparency is computed without application sorting and without artifacts. The architecture can also be used for antialiasing, and an example of Carpenters classical A-buffer antialiasing is shown.

An architecture for interactive tetrahedral volume rendering

We present a new architecture for interactive unstructured volume rendering. Our system moves all the computations necessary for order-independent transparency and volume scan conversion from the CPU to the graphics hardware, and it makes a software sorting pass unnecessary. It therefore provides the same advantages for volume data that triangle-processing hardware provides for surfaces. To address a remaining bottleneck — the bandwidth between main memory and the graphics processor — we introduce two new primitives, tetrahedral strips and tetrahedral fans. These primitives allow performance improvements in rendering tetrahedral meshes similar to the improvements triangle strips and fans allow in rendering triangle meshes. We provide new techniques for generating tetrahedral strips that achieve, on the average, strip lengths of 17 on representative datasets. The combined effect of our architecture and new primitives is a 72 to 85 times increase in performance over triangle graphics hardware approaches. These improvements make it possible to use volumetric tetrahedral meshes in interactive applications.

Irregular grid volume rendering with composition networks

Volumetric irregular grids are the next frontier to conquer in interactive 3D graphics. Visualization algorithms for rectilinear 2563 data volumes have been optimized to achieve one frame/second to 15 frames/second depending on the workstation. With equivalent computational resources, irregular grids with millions of cells may take minutes to render for a new viewpoint. The state of the art for graphics rendering, PixelFlow, provides screen and object space parallelism for polygonal rendering. Unfortunately volume rendering of irregular data is at odds with the sort last architecture. I investigate parallel algorithms for direct volume rendering on PixelFlow that generalize to other compositing architectures. Experiments are performed on the Nasa Langley fighter dataset, using the projected tetrahedra approach of Shirley and Tuchman. Tetrahedral sorting is done by the circumscribing sphere approach of Cignoni et al. Key approaches include sort-first on sort-last, world space subdivision by clipping, rearrangeable linear compositing for any view angle, and static load balancing. The new world space subdivision by clipping provides for efficient and correct rendering of unstructured data by using object space clipping planes. Research results include performance estimates on PixelFlow for irregular grid volume rendering. PixelFlow is estimated to achieve 30 frames/second on irregular grids of 300,00 tetrahedra or 10 million tetrahedra per second.

PermWeb: remote parallel and distributed-volume visualization

In this paper we present a system for visualizing volume data from remote supercomputers. We have developed both parallel volume rendering algorithms, and the World Wide Web (WWW) software for accessing the data at the remote sites. The implementation uses Hypertext Markup Language, Java, and Common Gateway Interface scripts to connect WWW servers/clients to our volume renderers. The front ends are interactive Java classes for specification of view, shading , and classification inputs. We present performance results, and implementation details for connections to our computing resources at the University of California Santa Cruz including a MasPar MP-2, SGI Reality Engine-RE2, and SGI Challenge machines. We apply the system to the task of visualizing trabecular bone from finite element simulations. Fast volume rendering on remote compute servers through a web interface allows us to increase the accessibility of the results to more users. User interface issues, overview of parallel algorithm developments, and overall system interfaces and protocols are presented. Access is available through Uniform Resource Locator http://www.cse.ucsc.edu/research/slvg/.

Extensions to permutation warping for parallel volume rendering

Abstract Biomedical volume visualization requires high quality and high performance, but the existing high performance solutions such as the Shear Warp algorithm, 3D texture mapping, and special purpose hardware have problems. Permutation warping achieves high fidelity for biomedical datasets of regular rectilinear volumes, using a one-to-one communication scheme for optimal O(1) communication on massively parallel computers. Extensions are presented including data dependent optimizations using octrees, arbitrary view angle flexibility, and multiple instruction stream multiple data stream (MIMD) implementation. A MasPar MP-2, single instruction stream multiple data stream (SIMD) (16,384 processor), implementation achieves 14 frames/s, using trilinear reconstruction on 128 3 volumes for 400% runtime improvement over our previous result. A Proteus MIMD (32 processor) implementation achieves 1 frame/s on the same data. Additionally the PermWeb software architecture is presented, that has been shown as a proof of concept means to provide wide shared access to a powerful centralized renderer. All of these improvements make permutation warping an effective solution for biomedial volume visualization.

DATA DEPENDENT OPTIMIZATIONS FOR PERMUTATION VOLUME RENDERING

We have developed a highly efficient, high fidelity approach for parallel volume rendering that is called permutation warping. Permutation warping may use any one pass filter kernel, an example of which is trilinear reconstruction, an advantage over the shear warp approach. This work discusses experiments in improving permutation warping using data dependent optimizations to make it more competitive in speed with the shear warp algorithm. We use a linear octree on each processor for collapsing homogeneous regions and eliminating empty space. Static load balancing is also used to redistribute nodes from a processors octree to achieve higher efficiencies. In studies on a 16384 processor MasPar MP-2, we have measured improvements of 3 to 5 times over our previous results. Run times are 73 milliseconds, 29 Mvoxels/second, or 14 frames/second for 1283 volumes, the fastest MasPar volume rendering numbers in the literature. Run times are 427 milliseconds, 39 Mvoxels/second, or 2 frames/second for 2563 volumes. The performance numbers show that coherency adaptations are effective for permutation warping. Because permutation warping has good scalability characteristics, it proves to be a superior approach for massively parallel computers when image fidelity is a required feature. We have provided further evidence for the utility of permutation warping as a scalable, high fidelity, and high performance approach to parallel volume visualization.

CellFast: Interactive Unstructured Volume Rendering and Classification

CellFast is an interactive system for unstructured volume visualization. Our CellFast system uses optimizations of OpenGL triangle fans, customized quicksort, memory organization for cache efficiency, display lists, tetrahedral culling, and multithreading. The optimizations improve the performance of an approach similar to Shirley and Tuchman’s projected tetrahedra rendering to provide 1 frame/second for 240,122 tetrahedral cells, 3 frames/second for 70,125 tetrahedral cells, and 15 frames/second for 12,936 tetrahedral cells. CellFast also performs fully automated classification to assist rendering. Demonstration of fluid flow, medical simulation, and medical imaging datasets are provided. CellFast provides for very high resolutions (up to 3840x1024), at frame rates that are orders of magnitude higher in performance compared to those in the literature. The CellFast system combines an interactive renderer and automatic classification to make unstructured volume rendering more useful.

Data level comparison of surface classification and gradient filters

Surface classification and shading of three dimensional scalar data sets are important enhancements for direct volume rendering (DVR). However, unlike conventional surface rendering, DVR algorithms do not have explicit geometry to shade, making it difficult to perform comparisons. Furthermore, DVR, in general, involves a complex set of parameters whose effects on a rendered image are hard to compare. Previous work uses analytical estimations of the quality of interpolation, gradient filters, and classification. Typical comparisons are done using side-by-side examination of rendered images. However, non-linear processes are involved in the rendering pipeline and thus the comparison becomes particularly difficult. In this paper, we present a data level methodology for analyzing volume surface classification and gradient filters. Users can more effectively estimate algorithmic differences by using intermediate information. Based on this methodology, we also present new data level metrics and examples of analyzing differences in surface classification and gradient calculation. Please refer to www.cse.ucsc.edu/research/avis/dvr.html for a full color version of this paper.

Road Map and Issues in Collaborative Visualization

A majority of the collaborative systems today are still in 2D. This paper highlights several issues and challenges in extending support to 3D collaborative workspaces. In particular, we focus on collaborative visualization systems that support shared 3D virtual workspaces. Several features unique to 3D collaborative visualization systems can be identified. For example, the number of participants in a session is usually relatively small; the data sets and the amount of data transferred in a session are typically large; and because each participant may be located anywhere within the 3D virtual workspace, each participant may have his own viewpoint or perspective. These properties help constrain and shape the design of 3D collaborative visualization systems. This paper presents two related 3D collaborative visualization systems called CSpray and PET SLUG, and discusses the particular issues and challenges shared by other 3D collaborative visualization systems. We provide a road map from the insights developed from our research, to the development of 3D collaborative visualization.