Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Princeton University, 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....inherent in rendering tasks makes multi processing a natural solution for graphics acceleration. However, the type of parallelism and architectures best suited to exploit it are very di#erent for scan conversion systems and ray tracers. Most extant work in the field has focused on scan converters [6, 12]. In scan conversion, parallelism is, in general, achieved by distributing primitives to di#erent processors. Several di#erent stages in the rendering pipeline have been proposed as the appropriate place to do the communication [7] but all su#er from very poor worst case performance. In addition, ....

....performance gains realized are significant. Parallel ray tracing is done either asynchronously using MIMD architectures or synchronously using SIMD architectures. MIMD computers are more general, and there is evidence that they are more e#cient for scan converting renderers (see, for example, [6, 12]) and so most research on parallel ray tracing has focused on MIMD architectures. We contend, however, that the hardware resources of MIMD machines are not used e#ciently for ray tracing. SIMD architectures, such as the one we propose, perform the task as e#ectively with much less hardware cost ....

Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Princeton University, 1995.

....inherent in rendering tasks makes multi processing a natural solution for graphics acceleration. However, the type of parallelism and architectures best suited to exploit it are very different for scan conversion systems and ray tracers. Most extant work in the field has focused on scan converters [2,7]. In scan conversion, parallelism is, in general, achieved by distributing primitives to different processors. Several different stages in the rendering pipeline have been proposed as the appropriate place to do the communication [3] but all suffer from very poor worst case performance. In ....

....gains realized are significant. Parallel ray tracing is done either asynchronously using MIMD architectures or synchronously using SIMD architectures. MIMD computers are more general, and there is evidence that they are more efficient for scan converting renderers (see, for example, [2,7]) and so most research on parallel ray tracing has focused on MIMD architectures. We contend, however, that the hardware resources of MIMD machines are not used efficiently for ray tracing. SIMD architectures, such as the one we propose, perform the task as effectively with much less hardware cost ....

Michael Cox, Algorithms for Parallel Rendering, PhD thesis, Princeton University, 1995.

.... work that is characterized by the overlap factor (the number of tiles a triangle covers) Most systems approximate the tiles that a triangle overlaps by computing the intersection of the triangle s bounding box with the grid of miltchen,gws,homan,kekoa,hanrahan graphics.stanford.edu tiles [1][3] 4] 10] 11] Molnar [8] presents an equation for the expected overlap of rectangular bounding boxes on rectangular tiles that is experimentally verified in both Molnar [9] and Cox [2] Sorting triangles into tiles exactly, rather than by boundingbox, is certainly possible and would result in ....

M. Cox. Algorithms for Parallel Rendering, Ph.D. thesis, Princeton University, May 1995.

....127 Center for Integrated Systems, Stanford CA 94305. hanrahan cs.stanford.edu Appeared in Parallel Rendering Symposium , Atalanta, Oct.30 31, 1995, pp.81 88. well as by the I O node which controls the HiPPI channel. For parallel rendering on existing multicomputers, this is a serious problem [8, 9, 11, 18, 23]. Connecting a frame buffer directly to the multicomputer s routing network can achieve better bandwidth. Scalability can be improved by connecting the frame buffer to the network using multiple ports. An important issue for multi port frame buffer designs is how to synchronize the frame buffer ....

Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Princeton University, May 1995.

....will have its own copy of the database to traverse. Given that such systems have limited traversal performance and offer limited or no load balancing capability among the different pipelines, we offer them no further consideration. The only other documented sort first system in evidence is [COX95]. Cox created a sortfirst implementation of the Renderman REYES graphics pipeline on the Connection Machine CM5. Because Renderman is aimed at realistic rendering, this was not an interactive system. It was based upon an immediate mode input stream, and thus did not make use of frame to frame ....

....processors. If such a clump of primitives occupies 1 16th of the screen, then a granularity ratio of at least 16 is needed to partition this clump. Since this clumping behavior is very scene dependent, determining the ideal granularity ratio to use is not easy, as many researchers have discovered [WHIT92, ELLS93, COX95]. We examine the effects of the granularity ratio in the experiments described below. 37 The question of region assignment can perhaps be addressed more simply. The goal is to scatter each processor s regions uniformly over the screen. This can be achieved readily enough using regular ....

M. Cox, Algorithms for Parallel Rendering, doctoral dissertation, Princeton University, May 1995.

....to render 30 48M triangles per second at interactive frame rates while storing only one fourth of the scene in memory on each PC. 2 Background and Related Work Parallel rendering systems are often classified by the stage in which primitives are partitioned: sort first, sort middle, or sort last [7, 17]. Most traditional systems have been based on the sort middle approach, in which graphics primitives are partitioned among geometry processors, screen tiles are partitioned among rasterization processors, and a fast communication network is used to send primitives from geometry processors to ....

Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Department of Computer Science, Princeton University, 1995.

.... parallelism is usually used; some systems use all four types (e.g. the Silicon Graphics SkyWriter [SGI92] Many systems use object parallelism for the front end tasks (transformation, clipping, and lighting) and image parallelism for the rasterization and shading (e.g. Whit92] Croc93] and [Cox95]) Hardware systems usually use these two types of parallelism for the different tasks, but perform them on different processors, and thus also use functional parallelism (e.g. Ghar88] the Silicon Graphics GTX [Akel88] PixelPlanes 5 [Fuch89] and the Silicon Graphics RealityEngine [Akel93] ....

....It is difficult to compare the performance of these earlier works. Comparing triangles per second figures does not take into account differences in databases, rendering quality, or speed of the underlying machine. 14 Image Parallel Algorithms Object Parallel Algorithms [Orte93] Kili94] [Cox95] Polygon Rendering Algorithms for General Purpose Systems No Load Balancing Static Load Balancing Dynamic Load Balancing [Croc93] Robl88] Cox95] Whit92,Whit94] Figure 2.2 Taxonomy of the parallel polygon rendering algorithms for general purpose systems covered in this section. ....

[Article contains additional citation context not shown here]

Cox, Michael B., Algorithms for Parallel Rendering, Ph.D. Dissertation, Department of Computer Science, Princeton University, Princeton, New Jersey, May 1995.

....a sort last image composition scheme to achieve scalable speedups for large clusters of PCs driving a single display. 3 Choosing a Partitioning Strategy The first challenge in implementing a parallel rendering system is to choose a partitioning strategy. Following the taxonomy of Molnar et al. [5, 17], we consider sort middle, sort first, and sort last approaches. In sort middle systems, processing of graphics primitives is partitioned equally among geometry processors, while processing of pixels is partitioned among rasterization processors according to overlaps with screen space tiles. This ....

Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Department of Computer Science, Princeton University, 1995.

....values to form the final image. The total polygon rendering time can be significantly reduced by exploiting the parallelism of the rendering pipeline. The parallelization strategies for polygon rendering can be classified as sort first, sort middle and sortlast [11] In the sort first strategy [2, 14, 15], the 3D primitives in the world coordinate space are partitioned depending on the image region to which they will get projected. Every 3D primitive is then distributed to the processor that is responsible for that image region, transformed and rasterized. Finally the image regions are combined to ....

M. Cox. Algorithms for Parallel Rendering. PhD thesis, Department of Computer Science, Princeton University, May 1995.

....interaction, and new content creation. Our research focus is on constructing a high performance parallel rendering system with a network of PCs. Parallel rendering systems are often classified according to the stage in which primitive partitioning is done sort first, sortmiddle, or sort last [7, 22]. Most are based on a sort middle architecture, in which processing of graphics primitives is partitioned equally among geometry processors, while processing of pixels is partitioned among rasterization processors according to overlaps with screen space tiles. This approach is best suited for ....

Michael Cox, Algorithms for Parallel Rendering, Ph.D. thesis, Department of Computer Science, Princeton University, May, 1995.

.... issue in 3D rendering is sorting the geometric primitives with respect to a given viewpoint, the parallelization strategies for polygon rendering can be classified as sort first, sort middle and sort last depending on where the sorting operation is performed [15] In the sort first strategy [2, 18, 19], the set of 3D primitives in the input model are partitioned into subsets according to fixed geometrical constructs in the world coordinate space, such as rectangular bounding boxes. A 3D primitive can belong to more than one subset because it crosses the boundaries of the partitioning bounding ....

M. Cox, Algorithms for Parallel Rendering, Ph.D. Thesis, Department of Computer Science, Princeton University, May 1995.

....factor. To our knowledge, there has not been previous work studying different overlap classification algorithms for bucket rendering systems. Our investigation is based on an immediate mode sort first approach [26] Sort first has been studied for retained mode systems by Mueller [28] and Cox [5], among others. Mueller simulated a real time sort first system, primarily studying load balancing [28] and dynamic distribution of hierarchical scene descriptions [29] while Cox [5] implemented a sortfirst version of Renderman [36] Relatively little work has been done on interactive polygon ....

....sort first approach [26] Sort first has been studied for retained mode systems by Mueller [28] and Cox [5] among others. Mueller simulated a real time sort first system, primarily studying load balancing [28] and dynamic distribution of hierarchical scene descriptions [29] while Cox [5] implemented a sortfirst version of Renderman [36] Relatively little work has been done on interactive polygon rendering using a cluster of networked PCs [17, 34] Prior distributed rendering systems have mostly focused on inter frame parallelism [17] rendering separate frames of a sequence on ....

Michael Cox, Algorithms for Parallel Rendering, Ph.D. thesis, Department of Computer Science, Princeton University, May, 1995.

No context found.

Michael Cox. Algorithms for Parallel Rendering. PhD thesis, Princeton University, 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC