From owner-virtu-l@VMD.CSO.UIUC.EDU Sat Jun 24 04:48:48 1995 Return-Path: Received: from mx4.u.washington.edu by saul5.u.washington.edu (5.65+UW95.05/UW-NDC Revision: 2.33 ) id AA29221; Sat, 24 Jun 95 04:48:48 -0700 Received: from vmd.cso.uiuc.edu by mx4.u.washington.edu (5.65+UW95.05/UW-NDC Revision: 2.31 ) id AA16200; Sat, 24 Jun 95 04:48:46 -0700 Received: from VMD.CSO.UIUC.EDU by vmd.cso.uiuc.edu (IBM VM SMTP V2R2) with BSMTP id 9847; Sat, 24 Jun 95 06:51:39 CDT Received: from VMD.CSO.UIUC.EDU (NJE origin LISTSERV@UIUCVMD) by VMD.CSO.UIUC.EDU (LMail V1.2a/1.8a) with BSMTP id 1894; Sat, 24 Jun 1995 06:51:35 -0500 Received: from VMD.CSO.UIUC.EDU by VMD.CSO.UIUC.EDU (LISTSERV release 1.8b) with NJE id 5313 for VIRTU-L@VMD.CSO.UIUC.EDU; Sat, 24 Jun 1995 06:51:25 -0500 Received: from UIUCVMD (NJE origin SMTP@UIUCVMD) by VMD.CSO.UIUC.EDU (LMail V1.2a/1.8a) with BSMTP id 1890; Sat, 24 Jun 1995 06:51:25 -0500 Received: from relay3.UU.NET by vmd.cso.uiuc.edu (IBM VM SMTP V2R2) with TCP; Sat, 24 Jun 95 06:51:22 CDT Received: from uucp2.UU.NET by relay3.UU.NET with SMTP id QQyvmt05970; Sat, 24 Jun 1995 07:48:19 -0400 Received: from uunet-in.UUCP by uucp2.UU.NET with UUCP/RMAIL ; Sat, 24 Jun 1995 07:48:20 -0400 Received: by relay.UUNET.IN (Smail-N3.28-UUI) Message Id: ; Sat, 24 Jun 95 17:10 IST Received: by bangalore.UUNET.IN (Smail-N3.28-UUI) Message Id: ; Sat, 24 Jun 95 17:02 IST Received: by vreal.uunet.in (UUPC/extended 1.11x); Sat, 24 Jun 1995 15:46:03 IST Message-Id: <2febecec.vreal@vreal.uunet.in> Date: Sat, 24 Jun 1995 15:45:58 IST Reply-To: Rawla Sender: "VR / sci.virtual-worlds" From: Rawla Subject: Article in Windows Sources To: Multiple recipients of list VIRTU-L Status: OR Hi ! The Following is an interesting article on 3D Graphic accelerators for the PC. It's *long* but interesting. Regards Anuj __________________________ ARTICLE_________________________________ ***** Computer Select, April 1995 : Articles ***** Journal: Windows Sources April 1995 v00000003 n4 p144(6) COPYRIGHT Ziff-Davis Publishing Company 1995 ------------------------------------------------------------------------- Title: A new dimension (Graphics Cards Up the Ante) Author: Plain, Stephen W. Abstract: Three-dimensional software interfaces will create a demand for an entirely new breed of video accelerators capable of handling elaborate 3-D graphics. Even Pentium-based systems require hardware acceleration for acceptable 3-D performance because 3-D acceleration involves detailed, multi-step processes that are inherently resource-intensive, resulting in a tradeoff between performance and realism. Solutions that implement much of the 3-D video pipeline in software use more CPU cycles than hard-wired solutions but are much more commonly used today. Rendering systems use algorithms that break down 3-D objects into polygons. The simplest use triangles, while faster and more sophisticated algorithms build polygons with four or more sides. Texture mapping molds bitmapped images to 3-D objects to give them surfaces, and a variety of techniques can be used to add lighting and shading effects. Microsoft Windows 95 will support the new OpenGL standard, an API developed by Silicon Graphics that offers sophisticated and comprehensive 3-D support. ------------------------------------------------------------------------- Full Text: Pipelines, lighting effects, and API madness 3-D software adds the illusions of depth, lighting, and movement through three-dimensional space to deliver a look and feel that borders on the ultimate interface--real life. Though traditionally the domain of CAD and scientific applications, 3-D will no doubt eventually find its way into even the average user's computing experience. 3-D games are already starting to appear, and any significant stab at virtual reality--the holy grail of many interface designers--must build on 3-D graphics. Building and tracking 3-D objects in virtual space requires an enormous amount of processing power and relies heavily on floating-point calculations, as does portraying 3-D objects on a flat display. The widespread use of Pentium-based machines, which boast relatively powerful floating-point processing capabilities, provides an expanding hardware base finally capable of handling the rigors of 3-D. Even with a Pentium system, however, you'll need a graphics adapter capable of 3-D acceleration if you want acceptable 3-D performance. Unlike 2-D acceleration, which simply speeds relatively simple operations that draw 2-D objects on the screen, 3-D acceleration involves a detailed, multiple-step process that forces trade-offs between realism and performance. 3-D products implement this process, commonly known as the 3-D rendering pipeline, through software, hardware, or a combination of both. Accelerators can implement the rendering pipeline in various ways. Obviously, solutions that implement a significant part of the pipeline in software--those that depend on the video driver to perform a great number of computations rather than on the accelerator chip itself--will chew up more CPU cycles. Unfortunately, most of the accelerators now in the works relegate the initial steps of the process to the video driver, though ideally the accelerator itself would handle these functions. THE PIPELINE The application generating the 3-D image generally works with geometric objects such as spheres, cubes, and cones. To display an image, the application calls geometric functions analogous to the GDI calls in a traditional 2-D scenario. The driver (or accelerator in a hardware-only design) first breaks each geometric object these functions describe into polygons that describe its surface. It then sends each polygon through the rest of the rendering pipeline. Accelerators vary in how they create these polygons. Algorithms that use convex polygons to break down 3-D objects prove the easiest to implement, so vendors can most easily design an accelerator that breaks objects into triangles, which are inherently convex. On the other hand, performance decreases as the number of polygons increases, and triangularization often produces more polygons than absolutely necessary. For this reason, the more-sophisticated rendering systems support polygons with four or more sides. This approach adds speed and flexibility to the accelerator but adds cost in R&D. As the accelerator passes each polygon through the pipeline, it rotates, scales, or zooms the shape if necessary. The accelerator then performs the clipping stage, calculating which objects fall within the current viewing angle and eliminating those outside it. Well-designed accelerators also remove hidden surfaces to reduce the number of polygons they must process. LIGHTING and SHADING Next, the accelerator typically applies lighting to the scene, a task key to rendering objects realistically. The subtle nuances of shadows often determine the perceived quality of a rendering. To be optically correct, objects must appear to cast shadows both on themselves and on other objects. To achieve this effect, the accelerator applies lighting equations to 3-D objects and creates the shading by adjusting the colors on the various object faces. The simplest shading technique, flat shading, applies the appropriate lighting equation to a single point on an object's face and then shades the entire face with the resulting color. Although flat shading adds a degree of realism, it usually results in an artificial look. The more established technique, Gouraud shading, provides much better results. Gouraud shading applies the lighting equation to all the vertices of an object and then interpolates color gradually between each vertex. Color interpolation is most effective in 24-bit color modes, but most 3-D accelerators will dither colors to boost realism when 24-bit modes are unavailable. A third shading method, Phong shading, produces even more spectacular results than Gouraud shading, but its complexity makes it too expensive to implement on most graphics accelerators today. After the accelerator fixes each object in three-dimensional coordinates and applies shading, it must map each surface to the 2-D coordinates it can display on the screen. Working with 3-D graphics requires more information than the frame buffer traditionally stores, of course, primarily because the accelerator must store an indication of depth. Accelerators achieve this through the z-buffer, a separate section of video memory. The z-buffer stores a value that corresponds to the depth of each pixel instead of the color. Just as the data width of the frame buffer determines the granularity of color, the data width of the z-buffer determines the accuracy with which the chip can store depth, and therefore absolute position. Most applications call for no more than 16-bit z-buffering, but precise modeling applications may require a wider data path. The most-ambitious PC accelerators now slated use a z-buffer width of 32 bits. SURFACE APPEAL Rendering objects to their proper shape is one thing; making them look real is quite another. Given that the real world does not comprise a set of flat, smooth, homogeneous objects, the accelerator must also apply texture to objects to complete the illusion. Most accelerators use an approach called texture mapping to achieve this result. Texture mapping molds a two-dimensional bitmapped image to the shape of the 3-D object, wrapping it with the appearance of a particular texture. To achieve even more-realistic results, high-end accelerators apply the appropriate perspective to the texture image as they apply the texture. Texture mapping requires tremendous processing power, and having to rely on the CPU to perform this task can impose serious limits on software development. Conversely, the complicated nature of texture mapping makes it expensive to achieve in hardware. To implement texture mapping in hardware, an accelerator must dedicate areas of local memory to holding the texture map itself. The accelerator chip itself must calculate the appropriate addresses for the texture map from pixel coordinates. The accelerator might also perform additional color interpolation and blending operations as part of the texture-mapping process. If rendering 3-D objects realistically is a Herculean task, setting 3-D objects into motion proves even more demanding. To help maintain frame rates during 3-D animations, most 3-D accelerators implement double-buffering, a technique in which they display one scene while at the same time rendering the next in excess off-screen memory. API MADNESS A confusing number of 3-D API's are available or under development for the PC, but APIs are beginning to segregate into those suitable for game development and those that provide a fuller feature set. OpenGL carries the most promise at the high end, where it and Autodesk's HOOPS and HDI now constitute the major contenders. Numerous vendors including Microsoft have already endorsed OpenGL, which accounts for a good bit of its promise. Windows NT 3.5 supports OpenGL, as will a future version of Windows 95. Developed by Silicon Graphics, OpenGL provides a high degree of accuracy and 3-D sophistication, and compliance testing ensures consistency among OpenGL products. At the other end of the spectrum--the games market--three APIs vie for dominance: Reality Lab, from RenderMorphics; bRender, from Argonaut; and RenderWare, from Criterion. These rendering APIs are slimmer than OpenGL and better suited to game development. Intel also developed an API called 3DR that provides the low-level rasterization operations for 3-D, but 3DR lacks support for high-level geometric objects, such as cones and spheres. 3DR will appeal to game vendors who have already written their geometry routines and want to avoid rewriting their code from scratch. Microsoft is also developing 3D-DDI, a specification that will provide a low-level driver interface to OpenGL for chip and API vendors. Although the low-level interface is attractive to programmers, Microsoft is discouraging application developers from writing directly to it, positioning it instead as a low-level glue to provide OpenGL compatibility. DCI will play a role in 3-D as well, because OpenGL will provide direct DCI support in a manner similar to a GDI. 3-D accelerators capable of handling OpenGL commands directly will be able to use DCI to bypass the software version of OpenGL, which runs as a DLL under Windows. Vendors prepare a range of 3-D offerings Most accelerator vendors are developing boards that accelerate 3-D graphics, but only Matrox was shipping 3-D accelerators boards as we closed this project: the MGA Impression Plus and MGA Impression Lite, both based on the company's 64-bit MGA Athena chip. The Athena provides a balance of 2-D and 3-D acceleration at about the price you'd pay for a high-end 2-D accelerator. The Lite version of the Impression differs from the Plus only by its inability to accept Matrox's plug-in upgrade modules. The PCI version of the Impression Plus sells for $449 with 2MB of VRAM, expandable to 4MB. The Impression Plus delivers excellent quality 3-D graphics, but many of its 3-D capabilities leave a heavy processing burden on the system CPU. Matrox's offerings break 3-D objects into triangles, a simpler but somewhat slower approach than breaking objects into four-sided polygons. The application and driver software handle the initial stages of the rendering pipeline in main memory, courtesy of the CPU, and the Athena kicks in only after the driver has prepared each 3-D triangle for display on the screen. The Athena then performs depth interpolation in hardware by calculating the color and intensity of each pixel based on the 3-D coordinates the driver calculates. The chip also takes care of filling a 16-bit VRAM z-buffer with depth information for each pixel and writing information to the frame buffer when appropriate. Vendors and advocates of specific 3-D APIs have rushed to support the MGA boards, currently the only mainstream 3-D offerings available. The MGA consequently enjoys compatibility with OpenGL, 3D-DDI, HOOPS, 3DR, RenderWare, Reality Lab, and Matrox SXCI, the company's proprietary API. The Matrox boards ship with the MGA 3D-SuperPack, a CD collection that contains 3-D games, Criterion's RenderWare, a software developer's kit, and a demo version of Caligari's trueSpace rendering software. The 3D-SuperPack is not a comprehensive set of 3-D programs by any means, but it does provide a glimpse into the future of desktop 3-D. It also clearly illustrates that the MGA is capable of 3-D animations fluid enough for arcade style games and that it is suitable for CAD applications. However, the Impression lacks hardware-assisted texture mapping, making it ill-suited for applications that require high levels of realism. In the Works A more ambitious 3-D accelerator chip, 3D Labs' GLiNT, has also started to appear. As we went to press, 3D Labs was providing the first samples of the GLiNT 300SX chip to board vendors, though at least one company was already advertising boards based on the chip. The GLiNT devotes all its silicon to accelerating the complex calculations 3-D rendering requires. The chip has 2-D capabilities, but it uses these mostly to provide the building blocks necessary for 3-D rendering. 3D Labs wired the entire OpenGL command set into the GLiNT, ensuring that all calculations will occur in hardware. With Windows NT 3.5 and presumably a future version of Windows 95, Microsoft will include a runtime DLL that provides software support for OpenGL calls. GLiNT-based boards will ship with a replacement for this DLL that redirects the OpenGL calls through the video driver to the chip, offloading the processing from the system CPU. 3D Labs predicts the GLiNT will speed OpenGL processing by as much as a factor of 5 to 10 relative to unassisted processing on a 90-MHz Pentium. Boards that employ the GLiNT will need loads of RAM to support the chip's extensive hardware acceleration options. A typical GLiNT configuration will include a 4MB of VRAM frame buffer plus 4MB of z-buffer memory, either DRAM or VRAM. Unlike the MGA Athena, the GLiNT sports an adjustable 32-bit z-buffer. The chip can use a 16-, 24-, or 32-bit data path to the z-buffer, allowing it to provide more accurate depth information than the standard 16-bit z-buffer. The GLiNT also takes advantage of the alpha channel, the 8 bits left over within a 32-bit word when you use a color depth of 24 bits. It uses this space to store information it uses to determine a rendered image's translucency, from opaque to transparent, as laid out in OpenGL. Also unlike the MGA Athena, the GLiNT provides hardware acceleration of texture mapping and implements perspective on textured objects. A future version of the chip--the 300TX--will provide complete texture mapping for OpenGL in hardware. Several vendors are preparing to release accelerators based on the GLiNT. For instance, S3 has been working on a board design that couples its Vision968 and the GLiNT 300SX to provide 2-D, 3-D, and full-motion video acceleration on a single board, and STB plans to begin shipping in the near future a board that employs the two chips. Creative Labs is implementing a scaled-down games -oriented version of the GLiNT in hopes of establishing a 3-D standard as successful as its SoundBlaster audio standard. The company is reportedly planning to release a combination 3-D and sound card by Christmas 1995 that will sell for $250. Cirrus Logic is also beginning to sample its upcoming $100 2-D/3-D offering--the Mondello chip set. Cirrus based this three-chip set, also called the GD547X series, on the 5470 rendering engine the company acquired from Austek. The 5470 chip provides 3-D acceleration features such as Gouraud shading and alpha transparency. The Mondello will accelerate the rendering portion of the pipeline in hardware, but it will still rely on the system's CPU for the up-front floating point calculations and geometric processing. Mondello-based boards will support three separate arrays of on-board memory--the VRAM frame buffer, a DRAM z-buffer, and a DRAM private buffer. The 5470 uses the private memory to process special objects such as texture maps, icons, fonts, and display lists. The chip passes these objects to the frame buffer when it has finished processing them. Each separate memory array will accommodate from 1MB to 16MB, making the Mondello quite extensive. Cirrus Logic claims the Mondello will deliver workstation performance at a low cost and provide a better price/performance combination than the Glint 300SX, but given that neither chip is available for testing, we could not verify these claims. At the time of this writing, board vendors were evaluating the Mondello, but none had announced plans to use it in their designs. A Different Angle VideoLogic claims a 3-D chip set it is developing in cooperation with NEC Technologies will revolutionize 3-D technology. Instead of breaking 3-D objects into polygons and storing depth information in z-buffers, the PowerVR accelerator will treat the 3-D objects as entire entities and will manage them directly, according to VideoLogic. The company is counting on this approach's relatively low memory requirements and more-streamlined processing of objects to give it a cost/performance advantage. VideoLogic claims the PowerVR will deliver as much as a 10:1 cost/performance advantage over traditional z-buffered chips. The PowerVR will appear first in arcade game systems, but the company is banking on the flexibility of the chip set to take it into different markets as well, including PC graphics accelerators. Because its 3-D approach is so different, VideoLogic has defined its own API, called SGL. This API will provide ultra-high-level functions aimed at game design, such as functions that simulate smoke and fire. Possibly more imminent in the PC market will be the introduction of a new chip this year from a company called NVidea. The company is saying little about the chip at this point but reports that it will integrate 3-D and video acceleration in a single package and will target the low-end multimedia and games markets. Choose your cost performance trade-off The migration of 3-D graphics software from ultra-expensive workstations to mainstream PCs comes as almost as big a technology shift as the arrival of the GUI itself. Traditionally, sophisticated 3-D software demanded expensive graphics workstations by companies such as SGI. These workstations might dedicate as many as 20 banks of memory to a single graphics function, such as texture mapping. The relatively frugal PC market will tolerate only a fraction of the cost of such systems, a fact that has encouraged the development of innovative, less expensive designs. Even so, you'll have to make some tough choices when buying a 3-D-capable accelerator. The more highly integrated chips with more hardwired capabilities will lead in performance and rendering quality, but chips that integrate 2-D and video acceleration along with 3-D will maintain a cost advantage. Whichever route you choose, you'll need the fastest system you can manage to achieve adequate performance--a Pentium-based system at least, and very possibly a machine based on one of the coming next- generation processors from AMD, Cyrix, NexGen, or Intel itself. As the 3-D accelerator market continues to define itself, two tiers of accelerator chips seem to be emerging: highly integrated chips like the GLiNT 300SX will probably target high-end applications, such as CAD, while designs that devote proportionally less silicon to 3-D, such as the Matrox and NVidea offerings, will prove more suitable for accelerating 3-D games. -- "If you aim for the stars you just might make it to the moon." Anuj Rawla Telephone: +91-80-6634775 VR REAL TECHNOLOGIES (P) LTD. Fax +91-80-5585964 rawla@vreal.uunet.in Address: Casa Ansal, No. 18 N.S. Palya, Bannerghatta Rd. Bangalore. India