An Extended Volume Visualization Systelll

We present a special architecture for arbitrary parallel projection for visualization of volumetric data. Using a ray-casting technique, parallel memory access, and pipelined processing of rays in a composition tree, we can achieve interactive rendering rates for a 5123 dataset. 1.1 The Cube Architecture Cube is a special-purpose computer architecture for volume visualization [1]. The heart of the architecture is a Cubic Frame Buffer (CFB) , which is a la.rge (e.g., 128M voxels for a 5123 CFB) three-dimensional memory of voxels. The voxel is a quantum unit of volume, which has a value representing some measurable properties of the real object or phenomenon, such as the color, fluorescent level, material, and translucency ratio. Cube's processing speed is achieved by handling beams of voxels rather than single voxels. In order to access a full beam of voxels simultaneously, a 3D modular organization of the CFB has been designed [1]. A special 3D skewed organization of the CFB enables conflict-free access to a full beam (axial ray) of n voxels, in any orthographic direction. The SD Viewing P1'Ocessor (VPS) [2] generates 2D shaded orthographic projections of the CFB images. It casts rays into the CFB in the specified viewing direction, and utilizes the CFB parallel memory organization for conflict-free retrieval of a beam and then determines the pixel projection along that beam. It employs a sequence of n processing units which team up to generate the projection along a beam of n voxels in O(1og n) time for a CFB of n3 voxels. Consequently, the time necessary to generate an orthographic projection of n2 pixels is only 0(n2 log n), rather than the conventional 0(71,3) time. The VP3 "a.lso shades the projected pixels concurrently with the projection stage by employing the depth-gradient congradient shading technique [3]. Arbitrary parallel projections are currently created by first rotating the scene and then viewing it through a principal orthographic direction. However, the CFB image is distorted every time a. rotation is executed. A major goal of the extended Cube architecture project, presented in this papre, is to develop and prototype an alternative mechanism for pa.rallel viewing that supports real-time arbitrary viewing. 1.2 The Extended Cube Architecture The new architecture described here is an extension of the existing orthographic projection mechanism of Cube. It enahles arbitrary parallel projection in the same time complexity as orthographic projection, 0(n2 log n). Perspective projection can also be generated in a similar fashion [41, but requires a more complex architecture. A projection ray, originating at a pixel in the projection plane and cast through the CFB in an arbitrary direction, is the basic unit of projected data. Two processing stages are concerned with this data: first there is a need to retrieve the data from the CFB, and then to obtain the projection along the ray. In the original Cube architecture, where only orthographic viewing is supported, projection rays always coincide with orthogra.phic 64 R. Bakalash, A. Kaufman, R. Pacheco, H. Pfister beams and are fetched and processed by the beam projection mechanism. However, for arbitrary viewing there is no direct way to fetch arbitrary discrete rays from the eFB in paralleL The set of projection rays belonging to the same scan line of the projected 2D frame-buffer form a slanted plane, termed the Projection Ray Plane (PRP). For every parallel projection, all the PRPs can be made parallel to one major axis by fixing a degree of freedom in specifying the projection parameters, namely, by rotating the projection plane about the viewing axis. A whole PRP of beams (now parallel to an axis) is fetched in n memory cycles and stored in a 2D temporary buffer called the 2D Skewed BujJer (2DSB). The direction of the viewing ray within the original PRP depends on the observer's viewing direction. When a PRP is copied from the eFB to a 2D memory, it undergoes a 2D shearing to align all the viewing rays into beams along a direction parallel to a 2D axis (e.g., vertical). This step is a de-skewing step that is accomplished by a barrel shifter (see below). Once the viewing rays are aligned within the 2D memory, they can be individually fetched and treated by the ray projection mechanism. This imposes a basic structural condition on the 2D memory. It should be capable of parallel access for storing "horizontal" beams corning from the PRP, and for parallel retrieval of "vertical" viewing rays. The structure chosen for the 2D memory is a 2D Skewed Buffer (2DSB), described below. The retrieved "vertical" rays must pa.ss through another de-skewing process on their way to the ray projection mechanism in order to match the physical sequential order of the modules in the projection mechanism. The latter is a Ray Projection Tree (RPT), which is a hardware mechanism structured as a. hierarchical pipeline, capable of implementing a variety of projection functions (see below). The communication mechanism that bridges between the CFB and the 2DSB, and between the 2DSB and RPT and performs the de-skewing steps, is a unique beam-based barrel-shifting mechanism, termed the Conveyor [6], and is described below. Fig. 1.1 illustrates the genera.! architecture of the extended system for arbitrary parallel viewing comprising of the Cubic Frame Buffer (CFB), the 2D Skewed Buffer (2DSB), the Ray Processing Tree (RPT), and two Conveyors for ray de-skewing. 1.3 The 2D Skewed Buffer The 2DSB is used for storing the slanted PRP. The slanted PRP is loaded into the 2DSB one beam at a time, from the closest beam to t.he furthest. Each beam is shifted to the left or to the right within the 2DSB in order to align the viewing rays vertically. Since there may be 2n-l para.llel rays entering the slanted PRP (one for each voxel on the visible edges of the slanted PRP), the 2D memory must be at least 2n-l columns wide. Once the slanted PRP has been loaded onto the 2DSB one "horizonta1" beam at a time, each vertical ray is retrieved in turn, and transferred into the RPT. The rays are shifted as they are transferred in order to ensure that the closest "oxel in the ray appears at the desired position in the RPT. The 2DSB is physically divided into n modules and diagonally skewed to allow the writing of an entire horizontal beam simultaneously (conflict free), as well as the conflict free retrievaJ of an entire vertical ray (see Fig. 1.2 ). Each ray is processed in parallel to compute a pixel value to be displayed for that ray. Certain algorithms require that the values of the entire neighborhood of up to 26 voxels surrounding a central voxel be used in computing the pixel value. This requires information from slanted PRPs just above and just below the plane containing the ray currently being processed. This is accomplished by using multiple parallel 2DSBs and processing rays in a plane only after the succeeding