Rendering Architecture With Selectable Processing of Multi-Pixel Spans

The present application relates to computer graphics and animation systems, and particularly to 3D graphics rendering hardware. Background of the art and the prior embodiment, according to the parent application, is described below. Some of the distinctions of the presently preferred embodiment are particularly noted beginning on page 10.

Modern computer systems normally manipulate graphical objects as high-level entities. For example, a solid body may be described as a collection of triangles with specified vertices, or a straight line segment may be described by listing its two endpoints with three-dimensional or two-dimensional coordinates. Such high-level descriptions are a necessary basis for high-level geometric manipulations, and also have the advantage of providing a compact format which does not consume memory space unnecessarily.

Such higher-level representations are very convenient for performing the many required computations. For example, ray-tracing or other lighting calculations may be performed, and a projective transformation can be used to reduce a three-dimensional scene to its two-dimensional appearance from a given viewpoint. However, when an image containing graphical objects is to be displayed, a very low-level description is needed. For example, in a conventional CRT display, a "flying spot" is moved across the screen (one line at a time), and the beam from each of three electron guns is switched to a desired level of intensity as the flying spot passes each pixel location. Thus at some point the image model must be translated into a data set which can be used by a conventional display. This operation is known as "rendering."

The graphics-processing system typically interfaces to the display controller through a "frame store" or "frame buffer" of special two-port memory, which can be written to randomly by the graphics processing system, but also provides the synchronous data output needed by the video output driver. (Digital-to-analog conversion is also provided after the frame buffer.) Such a frame buffer is usually implemented using VRAM memory chips (or sometimes with DRAM and special DRAM controllers). This interface relieves the graphics-processing system of most of the burden of synchronization for video output. Nevertheless, the amounts of data which must be moved around are very sizable, and the computational and data-transfer burden of placing the correct data into the frame buffer can still be very large.

Even if the computational operations required are quite simple, they must be performed repeatedly on a large number of datapoints. For example, in a typical 1995 high-end configuration, a display of 1280.times.1024 elements may need to be refreshed at 72 Hz, with a color resolution of 24 bits per pixel. If blending is desired, additional bits (e.g. another 8 bits per pixel) will be required to store an "alpha" or transparency value for each pixel. This implies manipulation of more than 3 billion bits per second, without allowing for any of the actual computations being performed. Thus it may be seen that this is an environment with unique data manipulation requirements.

If the display is unchanging, no demand is placed on the rendering operations. However, some common operations (such as zooming or rotation) will require every object in the image space to be re-rendered. Slow rendering will make the rotation or zoom appear jerky. This is highly undesirable. Thus efficient rendering is an essential step in translating an image representation into the correct pixel values. This is particularly true in animation applications, where newly rendered updates to a computer graphics display must be generated at regular intervals.

The rendering requirements of three-dimensional graphics are particularly heavy. One reason for this is that, even after the three-dimensional model has been translated to a two-dimensional model, some computational tasks may be bequeathed to the rendering process. (For example, color values will need to be interpolated across a triangle or other primitive.) These computational tasks tend to burden the rendering process. Another reason is that since three-dimensional graphics are much more lifelike, users are more likely to demand a fully rendered image. (By contrast, in the two-dimensional images created e.g. by a GUI or simple game, users will learn not to expect all areas of the scene to be active or filled with information.)

FIG. 1A is a very high-level view of other processes performed in a 3D graphics computer system. A three dimensional image which is defined in some fixed 3D coordinate system (a "world" coordinate system) is transformed into a viewing volume (determined by a view position and direction), and the parts of the image which fall outside the viewing volume are discarded. The visible portion of the image volume is then projected onto a viewing plane, in accordance with the familiar rules of perspective. This produces a two-dimensional image, which is now mapped into device coordinates. It is important to understand that all of these operations occur prior to the operations performed by the rendering subsystem of the present invention. FIG. 1B is an expanded version of FIG. 1A, and shows the flow of operations defined by the OpenGL standard.

A vast amount of engineering effort has been invested in computer graphics systems, and this area is one of increasing activity and demands. Numerous books have discussed the requirements of this area; see, e.g., ADVANCES IN COMPUTER GRAPHICS (ed. Enderle 1990-); Chellappa and Sawchuk, DIGITAL IMAGE PROCESSING AND ANALYSIS (1985); COMPUTER GRAPHICS HARDWARE (ed. Reghbati and Lee 1988); COMPUTER GRAPHICS: IMAGE SYNTHESIS (ed. Joy et al.); Foley et al., FUNDAMENTALS OF INTERACTIVE COMPUTER GRAPHICS (2.ed. 1984); Foley, COMPUTER GRAPHICS PRINCIPLES & PRACTICE (2.ed. 1990); Foley, INTRODUCTION TO COMPUTER GRAPHICS (1994); Giloi, Interactive Computer Graphics (1978); Hearn and Baker, COMPUTER GRAPHICS (2.ed. 1994); Hill, COMPUTER GRAPHICS (1990); Latham, DICTIONARY OF COMPUTER GRAPHICS (1991); Magnenat-Thalma, IMAGE SYNTHESIS THEORY & PRACTICE (1988); Newman and Sproull, PRINCIPLES OF INTERACTIVE COMPUTER GRAPHICS (2.ed. 1979); PICTURE ENGINEERING (ed. Fu and Kunii 1982); PICTURE PROCESSING & DIGITAL FILTERING (2.ed. Huang 1979); Prosise, How COMPUTER GRAPHICS WORK (1994); Rimmer, BIT MAPPED GRAPHICS (2.ed. 1993); Salmon, COMPUTER GRAPHICS SYSTEMS & CONCEPTS (1987); Schachter, COMPUTER IMAGE GENERATION (1990); Watt, THREE-DIMENSIONAL COMPUTER GRAPHICS (2.ed. 1994); Scott Whitman, MULTIPROCESSOR METHODS FOR COMPUTER GRAPHICS RENDERING; the SIGGRAPH PROCEEDINGS for the years 1980-1994; and the IEEE Computer Graphics and Applications magazine for the years 1990-1994.

In many areas of computer graphics a succession of slowly changing pictures are displayed rapidly one after the other, to give the impression of smooth movement, in much the same way as for cartoon animation. In general the higher the speed of the animation, the smoother (and better) the result.

When an application is generating animation images, it is normally necessary not only to draw each picture into the frame buffer, but also to first clear down the frame buffer, and to clear down auxiliary buffers such as depth (Z) buffers, stencil buffers, alpha buffers and others. A good treatment of the general principles may be found in Computer Graphics: Principles and Practice, James D. Foley et al., Reading MA: Addison-Wesley. A specific description of the various auxiliary buffers may be found in The OpenGL Graphics System: A Specification (Version 1.0), Mark Segal and Kurt Akeley, SGI.

In most applications the value written, when clearing any given buffer, is the same at every pixel location, though different values may be used in different auxiliary buffers. Thus the frame buffer is often cleared to the value which corresponds to black, while the depth (Z) buffer is typically cleared to a value corresponding to infinity.

The time taken to clear down the buffers is often a significant portion of the total time taken to draw a frame, so it is important to minimize it.

Due to the large number of at least partially independent operations which are performed in rendering, many proposals have been made to use some form of parallel architecture for graphics (and particularly for rendering). See, for example, the special issue of Computer Graphics on parallel rendering (September 1994). Other approaches may be found in earlier patent filings by the assignee of the present application and its predecessors, e.g. U.S. Pat. No. 5,195,186, and published PCT applications PCT/GB90/00987, PCT/GB90/01209, PCT/GB90/01210, PCT/GB90/01212, PCT/GB90/01213, PCT/GB90/01214, PCT/GB90/01215, and PCT/GB90/01216.

There are several general approaches to parallel processing. One of the basic approaches to achieving parallelism in computer processing is a technique known as pipelining. In this technique the individual processors are, in effect, connected in series in an assembly-line configuration: one processor performs a first set of operations on one chunk of data, and then passes that chunk along to another processor which performs a second set of operations, while at the same time the first processor performs the first set operations again on another chunk of data. Such architectures are generally discussed in Kogge, THE ARCHITECTURE OF PIPELINED COMPUTERS (1981).

The "OpenGL" standard is a very important software standard for graphics applications. In any computer system which supports this standard, the operating system(s) and application software programs can make calls according to the OpenGL standards, without knowing exactly what the hardware configuration of the system is.

The OpenGL standard provides a complete library of low-level graphics manipulation commands, which can be used to implement three-dimensional graphics operations. This standard was originally based on the proprietary standards of Silicon Graphics, Inc., but was later transformed into an open standard. It is now becoming extremely important, not only in high-end graphics-intensive workstations, but also in high-end PCs. OpenGL is supported by Windows NT.TM., which makes it accessible to many PC applications.

The OpenGL specification provides some constraints on the sequence of operations. For instance, the color DDA operations must be performed before the texturing operations, which must be performed before the alpha operations. (A "DDA" or digital differential analyzer, is a conventional piece of hardware used to produce linear gradation of color (or other) values over an image area.)

Other graphics interfaces (or "APIs"), such as PHIGS or XGL, are also current as of 1995; but at the lowest level, OpenGL is a superset of most of these.

The OpenGL standard is described in the OPENGL PROGRAMMING GUIDE (1993), the OPENGL REFERENCE MANUAL (1993), and a book by Segal and Akeley (of SGI) entitled THE OPENGL GRAPHICS SYSTEM: A SPECIFICATION (Version 1.0).

FIG. 1B is an expanded version of FIG. 1A, and shows the flow of operations defined by the OpenGL standard. Note that the most basic model is carried in terms of vertices, and these vertices are then assembled into primitives (such as triangles, lines, etc.). After all manipulation of the primitives has been completed, the rendering operations will translate each primitive into a set of "fragments." (A fragment is the portion of a primitive which affects a single pixel.) Again, it should be noted that all operations above the block marked "Rasterization" would be performed by a host processor, or possibly by a "geometry engine" (i.e. a dedicated processor which performs rapid matrix multiplies and related data manipulations), but would normally not be performed by a dedicated rendering processor such as that of the presently preferred embodiment. GLINT SYSTEM AND PREFERRED CONTEXT

The GLINT system provides a new approach to these needs. The preferred embodiment provides a graphics processing chip which uses a deep pipeline of multiple asynchronous units, separated by FIFOs, to achieve a high net throughput in 3D rendering. Besides the output interface to the frame buffer, a separate interface is to a local buffer which can be used for data manipulation (such as Z-buffering). Preferably reads and writes to the local buffer are provided by separate stages of the pipeline. Preferably some of the individual units include parallel paths internally. Preferably some of the individual units are connected to look ahead by more than one stage, to keep the pipeline filled while minimizing the use of expensive deep FIFOs.

The graphics management chip provided by the presently preferred embodiment implements the low-level rasterizing functions of OpenGL, together with some additional functions which aid in management of two-dimensional rendering to serve the graphical user interface.

The message-passing architecture of the presently preferred embodiment provides a long pipeline, in which the individual stages of the pipeline operate asynchronously. To optimize performance, stages of the pipeline may have internally parallel structure. (However, this is a basically quite different processing paradigm from the parallel rendering environments being explored by many workers.)

Where possible, data is kept on chip (registered) between blocks. However, of course, memory access is sometimes necessary. Thus, although most of the blocks are two-port blocks, some are multi-port to permit memory access. FIFO buffering is typically used for interface between the blocks. In many cases, one-deep FIFO's can be used, with appropriate look-ahead connections for timing control. However, in other stages, significantly deeper FIFO's are used, to avoid "bubbles" in the pipeline and optimize processor utilization.

The overall architecture of this innovative chip is best viewed using the software paradigm of a message passing system. In this system all the processing blocks are connected in a long pipeline with communication with the adjacent blocks being done through message passing. Between each block there is a small amount of buffering, the size being specific to the local communications requirements and speed of the two blocks.

The message rate is variable and depends on the rendering mode. The messages do not propagate through the system at a fixed rate typical of a more traditional pipeline system. If the receiving block can not accept a message, because its input buffer is full, then the sending block stalls until space is available.

The message structure is fundamental to the whole system as the messages are used to control, synchronize and inform each block about the processing it is to undertake. Each message has two fields--a data field and a tag field. The data field will hold color information, coordinate information, local state information, etc. The tag field is used by each block to identify the message type so it knows how to act on it.

A particular advantage of this architecture is that it inherently provides a very high degree of design for testability. Moreover, this is achieved without adding any special diagnostic hardware paths or registers. By providing appropriate commands to the chip, any desired input can be sent to any block within the pipeline. Thus modifications to the architecture can be tested very rapidly, and debugging can rapidly pinpoint any faults which may be present.

A particular advantage of this architecture is that it permits a very efficient test strategy: each unit can be taken out of the message stream and tested in isolation. This is possible because the interactions are all though the messages, and each unit does not know or care where the messages come from. Thus testing software can generate streams of messages as stimulus, and can check the resulting messages coming out against what the specified behavioral model defines. The input and output timings are varied to force the internal states to run in a blocked or non-blocking modes to further increase the test coverage. Moreover, the test coverage can be ascertained (both at the C statement level in the simulator and at the VHDL level), so that the comprehensiveness of the tests is not an unknown.

The parent application disclosed a graphics rendering chip with a pipelined message-passing architecture.

The present application discloses a new embodiment with many improvements over that disclosed in the parent application. Among these is the capability to decompose primitives into multi-pixel fragments. The multi-pixel "span" fragment, used in 2D rendering, includes many pixels in a row (32, in the presently preferred embodiment). When span fragments are enabled, the rasterizer decomposes objects into spans, to the extent possible. A mask is included with each span, so that the pixels of a span need not be continuous. These multi-pixel fragments provide some very significant economies at various downstream blocks, particularly where memory accesses are involved.

The disclosed inventions will be described with reference to the accompanying drawings, which show important sample embodiments of the invention and which are incorporated in the specification hereof by reference, wherein:

The disclosed inventions will be described with reference to the accompanying drawings, which show important sample embodiments of the invention and which are incorporated in the specification hereof by reference, wherein:

FIG. 1A, described above, is an overview of key elements and processes in a 3D graphics computer system.

FIG. 1B is an expanded version of FIG. 1A, and shows the flow of operations defined by the OpenGL standard.

FIG. 2A is an overview of the graphics rendering chip of the preferred embodiment of the parent case.

FIG. 2B is an overview of the graphics rendering chip of the presently preferred embodiment.

FIG. 2C is a more schematic view of the sequence of operations performed in the graphics rendering chip of FIG. 2B, when operating in a first mode.

FIG. 2D is a different view of the graphics rendering chip of FIG. 2B, showing the connections of a readback bus which provides a diagnostic pathway.

FIG. 2E is yet another view of the graphics rendering chip of FIG. 2B, showing how the functions of the core pipeline of FIG. 2C are combined with various external interface functions.

FIG. 2F is yet another view of the graphics rendering chip of FIG. 2B, showing how the details of FIFO depth and lookahead are implemented, in the presently preferred embodiment.

FIG. 3A shows a sample graphics board which incorporates the chip of FIG. 2B.

FIG. 3B shows another sample graphics board implementation, which differs from the board of FIG. 3A in that more memory and an additional component is used to achieve higher performance.

FIG. 3C shows another graphics board, in which the chip of FIG. 2B shares access to a common frame store with GUI accelerator chip.

FIG. 3D shows another graphics board, in which the chip of FIG. 2B shares access to a common frame store with a video coprocessor (which may be used for video capture and playback functions.

FIG. 4A illustrates the definition of the dominant side and the subordinate sides of a triangle.

FIG. 4B illustrates the sequence of rendering an Antialiased Line primitive.

The numerous innovative teachings of the present application will be described with particular reference to the presently preferred embodiment (by way of example, and not of limitation). The presently preferred embodiment is a GLINT.TM. 400.TM. 3D rendering chip. The Hardware Reference Manual and Programmer's Reference Manual for this chip describe further details of this sample embodiment, and are both hereby incorporated by reference. Both are available, as of the effective filing date of this application, from 3Dlabs Inc. Ltd., 181 Metro Drive, Suite 520, San Jose Calif. 95110.

The following definitions may help in understanding the exact meaning of terms used in the text of this application:

application: a computer program which uses graphics animation.

depth (Z) buffer: A memory buffer containing the depth component of a pixel. Used to, for example, eliminate hidden surfaces.

blt double-buffering: A technique for achieving smooth animation, by rendering only to an undisplayed back buffer, and then copying the back buffer to the front once drawing is complete.

FrameCount Planes: Used to allow higher animation rates by enabling DRAM local buffer pixel data, such as depth (Z), to be cleared down quickly.

frame buffer: An area of memory containing the displayable color buffers (front, back, left, right, overlay, underlay). This memory is typically separate from the local buffer.

local buffer: An area of memory which may be used to store non-displayable pixel information: depth(Z), stencil, FrameCount and GID planes. This memory is typically separate from the framebuffer.

pixel: Picture element. A pixel comprises the bits in all the buffers (whether stored in the local buffer or framebuffer), corresponding to a particular location in the framebuffer.

stencil buffer: A buffer used to store information about a pixel which controls how subsequent stencilled pixels at the same location may be combined with the current value in the framebuffer. Typically used to mask complex two-dimensional shapes.

Preferred Chip Embodiment--Overview

The GLINT.TM. high performance graphics processors combine workstation class 3D graphics acceleration, and state-of-the-art 2D performance in a single chip. All 3D rendering operations are accelerated by GLINT, including Gouraud shading, texture mapping, depth buffering, anti-aliasing, and alpha blending.

The scalable memory architecture of GLINT makes it ideal for a wide range of graphics products, from PC boards to high-end workstation accelerators.

There will be several of the GLINT family of graphics processors: the GLINT 300SX.TM. is the preferred embodiment of the parent application, and the GLINT 400TX.TM. is the presently preferred embodiment which is described herein in great detail. The two devices are generally compatible, with the 400TX adding local texture storage and texel address generation for all texture modes.

FIG. 2B is an overview of the graphics rendering chip of the presently preferred embodiment (i.e. the GLINT 400TX.TM.).

General Concept

The overall architecture of the GLINT chip is best viewed using the software paradigm of a message passing system. In this system all the processing blocks are connected in a long pipeline with communication with the adjacent blocks being done through message passing. Between each block there is a small amount of buffering, the size being specific to the local communications requirements and speed of the two blocks.

The message rate is variable and depends on the rendering mode. The messages do not propagate through the system at a fixed rate typical of a more traditional pipeline system. If the receiving block can not accept a message, because its input buffer is full, then the sending block stalls until space is available.

The message structure is fundamental to the whole system as the messages are used to control, synchronize and inform each block about the processing it is to undertake. Each message has two fields--a 32 bit data field and a 9 bit tag field. (This is the minimum width guaranteed, but some local block to block connections may be wider to accommodate more data.) The data field will hold color information, coordinate information, local state information, etc. The tag field is used by each block to identify the message type so it knows how to act on it.

Each block, on receiving a message, can do one of several things:

Not recognize the message so it just passes it on to the next block.

Recognize it as updating some local state (to the block) so the local state is updated and the message terminated, i.e. not passed on to the next block.

Recognize it as a processing action, and if appropriate to the unit, the processing work specific to the unit is done. This may entail sending out new messages such as Color and/or modifying the initial message before sending it on. Any new messages are injected into the message stream before the initial message is forwarded on. Some examples will clarify this.

When the Depth Block receives a message `new fragment`, it will calculate the corresponding depth and do the depth test. If the test passes then the `new fragment` message is passed to the next unit. If the test fails then the message is modified and passed on. The temptation is not to pass the message on when the test fails (because the pixel is not going to be updated), but other units downstream need to keep their local DDA units in step.

(In the present application, the messages are being described in general terms so as not to be bogged down in detail at this stage. The details of what a `new fragment` message actually specifies (i.e. coordinate, color information) is left till later. In general, the term "pixel" is used to describe the picture element on the screen or in memory. The term "fragment" is used to describe the part of a polygon or other primitive which projects onto a pixel. Note that a fragment may only cover a part of a pixel.)

When the Texture Read Unit (if enabled) gets a `new fragment` message, it will calculate the texture map addresses, and will accordingly provide 1, 2, 4 or 8 texels to the next unit together with the appropriate number of interpolation coefficients.

Each unit and the message passing are conceptually running asynchronous to all the others. However, in the presently preferred embodiment there is considerable synchrony because of the common clock.

How does the host process send messages? The message data field is the 32 bit data written by the host, and the message tag is the bottom 9 bits of the address (excluding the byte resolution address lines). Writing to a specific address causes the message type associated with that address to be inserted into the message queue. Alternatively, the on-chip DMA controller may fetch the messages from the host's memory.

The message throughput, in the presently preferred embodiment, is 50M messages per second and this gives a fragment throughput of up to 50M per second, depending on what is being rendered. Of course, this rate will predictably be further increased over time, with advances in process technology and clock rates.

Linkage

The block diagram of FIG. 2A shows how the units are connected together in the GLINT 300SX embodiment, and the block diagram of FIG. 2B shows how the units are connected together in the presently preferred GLINT 400TX embodiment. Some general points are:

The following functionality is present in the 400TX , but missing from the 300TX: The Texture Address (TAddr) and Texture Read (TRd) Units are missing. Also, the router and multiplexer are missing from this section, so the unit ordering is Scissor/Stipple, Color DDA, Texture Fog Color, Alpha Test, LB Rd, etc.

In the embodiment of FIG. 2B, the order of the units can be configured in two ways. The most general order (Router, Color DDA, Texture Unit, Alpha Test, LB Rd, GID/Z/Stencil, LB Wr, Multiplexer) and will work in all modes of OpenGL. However, when the alpha test is disabled it is much better to do the Graphics ID, depth and stencil tests before the texture operations rather than after. This is because the texture operations have a high processing cost and this should not be spent on fragments which are later rejected because of window, depth or stencil tests.

The loop back to the host at the bottom provides a simple synchronization mechanism. The host can insert a Sync command and when all the preceding rendering has finished the sync command will reach the bottom host interface which will notify the host the sync event has occurred.

Benefits

The very modular nature of this architecture gives great benefits. Each unit lives in isolation from all the others and has a very well defined set of input and output messages. This allows the internal structure of a unit (or group of units) to be changed to make algorithmic/speed/gate count trade-offs.

The isolation and well defined logical and behavioral interface to each unit allows much better testing and verification of the correctness of a unit.

The message passing paradigm is easy to simulate with software, and the hardware design is nicely partitioned. The architecture is self synchronizing for mode or primitive changes.

The host can mimic any block in the chain by inserting messages which that block would normally generate. These message would pass through the earlier blocks to the mimicked block unchanged and from then onwards to the rest of the blocks which cannot tell the message did not originate from the expected block. This allows for an easy work around mechanism to correct any flaws in the chip. It also allows other rasterization paradigms to be implemented outside of the chip, while still using the chip for the low level pixel operations.

"A Day in the Life of a Triangle"

Before we get too detailed in what each unit does it is worth while looking in general terms at how a primitive (e.g. triangle) passes through the pipeline, what messages are generated, and what happens in each unit. Some simplifications have been made in the description to avoid detail which would otherwise complicate what is really a very simple process. The primitive we are going to look at is the familiar Gouraud shaded Z buffered triangle, with dithering. It is assumed any other state (i.e. depth compare mode) has been set up, but (for simplicity) such other states will be mentioned as they become relevant.

The application generates the triangle vertex information and makes the necessary OpenGL calls to draw it.

The OpenGL server/library gets the vertex information, transforms, clips and lights it. It calculates the initial values and derivatives for the values to interpolate (X.sub.left, X.sub.right, red, green, blue and depth) for unit change in dx and dxdy.sub.left. All these values are in fixed point integer and have unique message tags. Some of the values (the depth derivatives) have more than 32 bits to cope with the dynamic range and resolution so are sent in two halves Finally, once the derivatives, start and end values have been sent to GLINT the `render triangle` message is sent.

On GLINT: The derivative, start and end parameter messages are received and filter down the message stream to the appropriate blocks. The depth parameters and derivatives to the Depth Unit; the RGB parameters and derivative to the Color DDA Unit; the edge values and derivatives to the Rasterizer Unit. * The `render triangle` message is received by the rasterizer unit and all subsequent messages (from the host) are blocked until the triangle has been rasterized (but not necessarily written to the frame store). A `prepare to render` message is passed on so any other blocks can prepare themselves.

The Rasterizer Unit walks the left and right edges of the triangle and fills in the spans between. As the walk progresses messages are send to indicate the direction of the next step: StepX or StepYDomEdge. The data field holds the current (x, y) coordinate. One message is sent per pixel within the triangle boundary. The step messages are duplicated into two groups: an active group and a passive group. The messages always start off in the active group but may be changed to the passive group if this pixel fails one of the tests (e.g. depth) on its path down the message stream. The two groups are distinguished by a single bit in the message tag. The step messages (in either form) are always passed throughout the length of the message stream, and are used by all the DDA units to keep their interpolation values in step. The step message effectively identifies the fragment and any other messages pertaining to this fragment will always precede the step message in the message stream.

The Scissor and Stipple Unit. This unit does 4 tests on the fragment (as embodied by the active step message). The screen scissor test takes the coordinates associated with the step message, converts them to be screen relative (if necessary) and compares them against the screen boundaries. The other three tests (user scissor, line stipple and area stipple) are disabled for this example. If the enabled tests pass then the active step is forwarded onto the next unit, otherwise it is changed into a passive step and then forwarded.

The Color DDA unit responds to an active step message by generating a Color message and sending this onto the next unit. The active step message is then forwarded to the next unit. The Color message holds, in the data field, the current RGBA value from the DDA. If the step message is passive then no Color message is generated. After the Color message is sent (or would have been sent) the step message is acted on to increment the DDA in the correct direction, ready for the next pixel.

Texturing, Fog and Alpha Tests Units are disabled so the messages just pass through these blocks.

In general terms the Local Buffer Read Unit reads the Graphic ID, Stencil and Depth information from the Local Buffer and passes it onto the next unit. More specifically it does:

1. If the step message is passive then no further action occurs.

2. On an active step message it calculates the linear address in the local buffer of the required data. This is done using the (X, Y) position recorded in the step message and locally stored information on the `screen width`0 and window base address. Separate read and write addresses are calculated.

3. The addresses are passed to the Local Buffer Interface Unit and the identified local buffer location read. The write address is held for use later.

4. Sometime later the local buffer data is returned and is formatted into a consistent internal format and inserted into a `Local Buffer Data` message and passed on to the next unit. The message data field is made wider to accommodate the maximum Local Buffer width of 52 bits (32 depth, 8 stencil, 4 graphic ID, 8 frame count) and this extra width just extends to the Local Buffer Write block.

The actual data read from the local buffer can be in several formats to allow narrower width memories to be used in cost sensitive systems. The narrower data is formatted into a consistent internal format in this block.

The Graphic ID, Stencil and Depth Unit just passes the Color message through and stores the LBData message until the step message arrives. A passive step message would just pass straight through.

When the active step message is received the internal Graphic ID, stencil and depth values are compared with the ones in the LBData message as specified by this unit's mode information. If the enabled tests pass then the new local buffer data is sent in the LBWriteData message to the next unit and the active step message forwarded. If any of the enabled tests fail then an LBCancelWrite message is sent followed by the equivalent passive step message. The depth DDA is stepped to update the local depth value.

The Local Buffer Write Unit performs any writes which are necessary. The LBWriteData message has its data formatted into the external local buffer format and this is posted to the Local Buffer Interface Unit to be written into the memory (the write address is already waiting in the Local Buffer Interface Unit). The LBWriteCancel message just informs the Local Buffer Interface Unit that the pending write address is no longer needed and can be discarded. The step message is just passed through.

In general terms the Framebuffer Read Unit reads the color information from the framebuffer and passes it onto the next unit. More specifically it does:

1. If the step message is passive then no further action occurs.

2. On an active step message it calculates the linear address in the framebuffer of the required data. This is done using the (X, Y) position recorded in the step message and locally stored information on the `screen width` and window base address. Separate read and write addresses are calculated.

3. The addresses are passed to the Framebuffer Interface Unit and the identified framebuffer location read. The write address is held for use later.

4. Sometime later the color data is returned and inserted into a `Frame Buffer Data` message and passed on to the next unit. The actual data read from the framestore can be in several formats to allow narrower width memories to be used in cost sensitive systems. The formatting of the data is deferred until the Alpha Blend Unit as it is the only unit which needs to match it up with the internal formats. In this example no alpha blending or logical operations are taking place, so reads are disabled and hence no read address is sent to the Framebuffer Interface Unit. The Color and step messages just pass through.

The Alpha Blend Unit is disabled so just passes the messages through.

The Dither Unit stores the Color message internally until an active step is received. On receiving this it uses the least significant bits of the (X, Y) coordinate information to dither the contents of the Color message. Part of the dithering process is to convert from the internal color format into the format of the framebuffer. The new color is inserted into the Color message and passed on, followed by the step message.

The Logical Operations are disabled so the Color message is just converted into the FBWriteData message Oust the tag changes) and forwarded on to the next unit. The step message just passes through.

The Framebuffer Write Unit performs any writes which are necessary.

The FBWriteData message has its data posted to the Framebuffer Interface Unit to be written into the memory (the write address is already waiting in the Framebuffer Interface Unit).

The step message is just passed through.

The Host Out Unit is mainly concerned with synchronization with the host so for this example will just consume any messages which reach this point in the message stream.

This description has concentrated on what happens as one fragment flows down the message stream. It is important to remember that at any instant in time there are many fragments flowing down the message stream and the further down they reach the more processing has occurred.

Interfacing Between Blocks

FIG. 2B shows the FIFO buffering and lookahead connections which are used in the presently preferred embodiment. The FIFOs are used to provide an asynchronous interface between blocks, but are expensive in terms of gate count. Note that most of these FIFOs are only one stage deep (except where indicated), which reduces their area. To maintain performance, lookahead connections are used to accelerate the "startup" of the pipeline. For example, when the Local-Buffer-Read block issues a data request, the Texture/Fog/Color blocks also receive this, and begin to transfer data accordingly. Normally a single-entry deep FIFO cannot be read and written in the same cycle, as the writing side doesn't know that the FIFO is going to be read in that cycle (and hence become eligible to be written). The look-ahead feature give the writing side this insight, so that single-cycle transfer can be achieved. This accelerates the throughput of the pipeline.

Programming Model

The following text describes the programming model for GLINT.

GLINT as at Register file

The simplest way to view the interface to GLINT is as a flat block of memory-mapped registers (i.e. a register file). This register file appears as part of Region 0 of the PCI address map for GLINT. See the GLINT Hardware Reference Manual for details of this address map.

When a GLINT host software driver is initialized it can map the register file into its address space. Each register has an associated address tag, giving its offset from the base of the register file (since all registers reside on a 64-bit boundary, the tag offset is measured in multiples of 8 bytes). The most straightforward way to load a value into a register is to write the data to its mapped address. In reality the chip interface comprises a 16 entry deep FIFO, and each write to a register causes the written value and the register's address tag to be written as a new entry in the FIFO.

Programming GLINT to draw a primitive consists of writing initial values to the appropriate registers followed by a write to a command register. The last write triggers the start of rendering.

GLINT has approximately 200 registers. All registers are 32 bits wide and should be 32-bit addressed. Many registers are split into bit fields, and it should be noted that bit 0 is the least significant bit.

Register Types

GLINT has three main types of register:

Control Registers

Command Registers

Internal Registers

Control Registers are updated only by the host--the chip effectively uses them as read-only registers. Examples of control registers are the Scissor Clip unit min and max registers. Once initialized by the host, the chip only reads these registers to determine the scissor clip extents.

Command Registers are those which, when written to, typically cause the chip to start rendering (some command registers such as ResetPickResult or Sync do not initiate rendering). Normally, the host will initialize the appropriate control registers and then write to a command register to initiate drawing. There are two types of command registers: begin-draw and continue-draw. Begin-draw commands cause rendering to start with those values specified by the control registers. Continue-draw commands cause drawing to continue with internal register values as they were when the previous drawing operation completed. Making use of continue-draw commands can significantly reduce the amount of data that has to be loaded into GLINT when drawing multiple connected objects such as polylines. Examples of command registers include the Render and ContinueNewLine registers.

For convenience this application will usually refer to "sending a Render command to GLINT" rather than saying (more precisely) "the Render Command register is written to, which initiates drawing".

Internal Registers are not accessible to host software. They are used internally by the chip to keep track of changing values. Some control registers have corresponding internal registers. When a begin-draw command is sent and before rendering starts, the internal registers are updated with the values in the corresponding control registers. If a continue-draw command is sent then this update does not happen and drawing continues with the current values in the internal registers. For example, if a line is being drawn then the StartXDom and StartY control registers specify the (x, y) coordinates of the first point in the line. When a begin-draw command is sent these values are copied into internal registers. As the line drawing progresses these internal registers are updated to contain the (x, y) coordinates of the pixel being drawn. When drawing has completed the internal registers contain the (x, y) coordinates of the next point that would have been drawn. If a continue-draw command is now given these final (x, y) internal values are not modified and further drawing uses these values. If a begin-draw command had been used the internal registers would have been reloaded from the StartXDom and StartY registers.

For the most part internal registers can be ignored. It is helpful to appreciate that they exist in order to understand the continue-draw commands.

GLINT I/O Interface

There are a number of ways of loading GLINT registers for a given context:

The host writes a value to the mapped address of the register

The host writes address-tag/data pairs into a host memory buffer and uses the on-chip DMA to transfer this data to the FIFO.

The host can perform a Block Command Transfer by writing address and data values to the FIFO interface registers.

In all cases where the host writes data values directly to the chip (via the register file) it has to worry about FIFO overflow. The InFIFOSpace register indicates how many free entries remain in the FIFO. Before writing to any register the host must ensure that there is enough space left in the FIFO. The values in this register can be read at any time. When using DMA, the DMA controller will automatically ensure that there is room in the FIFO before it performs further transfers. Thus a buffer of any size can be passed to the DMA controller.

The description above considered the GLINT interface to be a register file. More precisely, when a data value is written to a register this value and the address tag for that register are combined and put into the FIFO as a new entry. The actual register is not updated until GLINT processes this entry. In the case where GLINT is busy performing a time consuming operation (e.g. drawing a large texture mapped polygon), and not draining the FIFO very quickly, it is possible for the FIFO to become full. If a write to a register is performed when the FIFO is full no entry is put into the FIFO and that write is effectively lost.

The input FIFO is 16 entries deep and each entry consists of a tag/data pair. The InFIFOSpace register can be read to determine how many entries are free. The value returned by this register will never be greater than 16.

To check the status of the FIFO before every write is very inefficient, so it is preferably checked before loading the data for each rectangle. Since the FIFO is 16 entries deep, a further optimization is to wait for all 16 entries to be free after every second rectangle. Further optimizations can be made by moving dXDom, dXSub and dY outside the loop (as they are constant for each rectangle) and doing the FIFO wait after every third rectangle.

The InFIFOSpace FIFO control register contains a count of the number of entries currently free in the FIFO. The chip increments this register for each entry it removes from the FIFO and decrements it every time the host puts an entry in the FIFO.

Loading registers directly via the FIFO is often an inefficient way to download data to GLINT. Given that the FIFO can accommodate only a small number of entries, GLINT has to be frequently interrogated to determine how much space is left. Also, consider the situation where a given API function requires a large amount of data to be sent to GLINT. If the FIFO is written directly then a return from this function is not possible until almost all the data has been consumed by GLINT. This may take some time depending on the types of primitives being drawn.

To avoid these problems GLINT provides an on-chip DMA controller which can be used to load data from arbitrary sized (<64K 32-bit words) host buffers into the FIFO. In its simplest form the host software has to prepare a host buffer containing register address tag descriptions and data values. It then writes the base address of this buffer to the DMAAddress register and the count of the number of words to transfer to the DMACount register. Writing to the DMACount register starts the DMA transfer and the host can now perform other work. In general, if the complete set of rendering commands required by a given call to a driver function can be loaded into a single DMA buffer then the driver function can return. Meanwhile, in parallel, GLINT is reading data from the host buffer and loading it into its FIFO. FIFO overflow never occurs since the DMA controller automatically waits until there is room in the FIFO before doing any transfers.

The only restriction on the use of DMA control registers is that before attempting to reload the DMACount register the host software must wait until previous DMA has completed. It is valid to load the DMAAddress register while the previous DMA is in progress since the address is latched internally at the start of the DMA transfer.

Using DMA leaves the host free to return to the application, while in parallel, GLINT is performing the DMA and drawing. This can increase performance significantly over loading a FIFO directly. In addition, some algorithms require that data be loaded multiple times (e.g. drawing the same object across multiple clipping rectangles). Since the GLINT DMA only reads the buffer data, it can be downloaded many times simply by restarting the DMA. This can be very beneficial if composing the buffer data is a time consuming task.

The host can use this hardware capability in various ways. For example, a further optional optimization is to use a double buffered mechanism with two DMA buffers. This allows the second buffer to be filled before waiting for the previous DMA to complete, thus further improving the parallelism between host and GLINT processing. Thus, this optimization is dependent on the allocation of the host memory. If there is only one DMA host buffer then either it is being filled or it is being emptied--it cannot be filled and emptied at the same time, since there is no way for the host and DMA to interact once the DMA transfer has started. The host is at liberty to allocate as many DMA buffers as it wants; two is the minimum to do double buffering, but allocating many small buffers is generally better, as it gives the benefits of double buffering together with low latency time, so GLINT is not idle while large buffer is being filled up. However, use of many small buffers is of course more complicated.

In general the DMA buffer format consists of a 32-bit address tag description word followed by one or more data words. The DMA buffer consists of one or more sets of these formats. The following paragraphs describe the different types of tag description words that can be used.

There are 3 different tag addressing modes for DMA: hold, increment and indexed. The different DMA modes are provided to reduce the amount of data which needs to be transferred, hence making better use of the available DMA bandwidth. Each of these is described in the following sections.

In this format the 32-bit tag description contains a tag value and a count specifying the number of data words following in the buffer. The DMA controller writes each of the data words to the same address tag. For example, this is useful for image download where pixel data is continuously written to the Color register. The bottom 9 bits specify the register to which the data should be written; the high-order 16 bits specify the number of data words (minus 1) which follow in the buffer and which should be written to the address tag (note that the 2-bit mode field for this format is zero so a given tag value can simply be loaded into the low order 16 bits).

A special case of this format is where the top 16 bits are zero indicating that a single data value follows the tag (i.e. the 32-bit tag description is simply the address tag value itself). This allows simple DMA buffers to be constructed which consist of tag/data pairs.

This format is similar to the hold format except that as each data value is loaded the address tag is incremented (the value in the DMA buffer is not changed; GLINT updates an internal copy). Thus, this mode allows contiguous GLINT registers to be loaded by specifying a single 32-bit tag value followed by a data word for each register. The low-order 9 bits specify the address tag of the first register to be loaded. The 2 bit mode field is set to 1 and the high-order 16 bits are set to the count (minus 1) of the number of registers to update. To enable use of this format, the GLINT register file has been organized so that registers which are frequently loaded together have adjacent address tags. For example, the 32 AreaStipplePattern registers can be loaded as follows:

GLINT address tags are 9 bit values. For the purposes of the Indexed DMA Format they are organized into major groups and within each group there are up to 16 tags. The low-order 4 bits of a tag give its offset within the group. The high-order 5 bits give the major group number.

The following Register Table lists the individual registers with their Major Group and Offset in the presently preferred embodiment:

The following table lists registers by group, giving their tag values and indicating their type. The register groups may be used to improve data transfer rates to GLINT when using DMA.

This format allows up to 16 registers within a group to be loaded while still only specifying a single address tag description word.

If the Mode of the address tag description word is set to indexed mode, then the high-order 16 bits are used as a mask to indicate which registers within the group are to be used. The bottom 4 bits of the address tag description word are unused. The group is specified by bits 4 to 8. Each bit in the mask is used to represent a unique tag within the group. If a bit is set then the corresponding register will be loaded. The number of bits set in the mask determines the number of data words that should be following the tag description word in the DMA buffer. The data is stored in order of increasing corresponding address tag.

Host software must generate the correct DMA buffer address for the GLINT DMA controller. Normally, this means that the address passed to GLINT must be the physical address of the DMA buffer in host memory. The buffer must also reside at contiguous physical addresses as accessed by GLINT. On a system which uses virtual memory for the address space of a task, some method of allocating contiguous physical memory, and mapping this into the address space of a task, must be used.

If the virtual memory buffer maps to non-contiguous physical memory, then the buffer must be divided into sets of contiguous physical memory pages and each of these sets transferred separately. In such a situation the whole DMA buffer cannot be transferred in one go; the host software must wait for each set to be transferred. Often the best way to handle these fragmented transfers is via an interrupt handler.

GLINT provides interrupt support, as an alternative means of determining when a DMA transfer is complete. If enabled, the interrupt is generated whenever the DMACount register changes from having a non-zero to having a zero value. Since the DMACount register is decremented every time a data item is transferred from the DMA buffer this happens when the last data item is transferred from the DMA buffer.

To enable the DMA interrupt, the DMAInterruptEnable bit must be set in the IntEnable register. The interrupt handler should check the DMAFlag bit in the IntFlags register to determine that a DMA interrupt has actually occurred. To clear the interrupt a word should be written to the IntFlags register with the DMAFlag bit set to one.

This scheme frees the processor for other work while DMA is being completed. Since the overhead of handling an interrupt is often quite high for the host processor, the scheme should be tuned to allow a period of polling before sleeping on the interrupt.

To read data back from GLINT an output FIFO is provided. Each entry in this FIFO is 32-bits wide and it can hold tag or data values. Thus its format is unlike the input FIFO whose entries are always tag/data pairs (we can think of each entry in the input FIFO as being 41 bits wide: 9 bits for the tag and 32 bits for the data). The type of data written by GLINT to the output FIFO is controlled by the FilterMode register. This register allows filtering of output data in various categories including the following:

Depth: output in this category results from an image upload of the Depth buffer.

Stencil: output in this category results from an image upload of the Stencil buffer.

Color: output in this category results from an image upload of the framebuffer.

Synchronization: synchronization data is sent in response to a Sync command.

The data for the FilterMode register consists of 2 bits per category. If the least significant of these two bits is set (0.times.1) then output of the register tag for that category is enabled; if the most significant bit is set (0.times.2) then output of the data for that category is enabled. Both tag and data output can be enabled at the same time. In this case the tag is written first to the FIFO followed by the data.

For example, to perform an image upload from the framebuffer, the FilterMode register should have data output enabled for the Color category. Then, the rectangular area to be uploaded should be described to the rasterizer. Each pixel that is read from the framebuffer will then be placed into the output FIFO. If the output FIFO becomes full, then GLINT will block internally until space becomes available. It is the programmer's responsibility to read all data from the output FIFO. For example, it is important to know how many pixels should result from an image upload and to read exactly this many from the FIFO.

To read data from the output FIFO the OutputFIFOWords register should first be read to determine the number of entries in the FIFO (reading from the FIFO when it is empty returns undefined data). Then this many 32-bit data items are read from the FIFO. This procedure is repeated until all the expected data or tag items have been read. The address of the output FIFO is described below.

Note that all expected data must be read back. GLINT will block if the FIFO becomes full. Programmers must be careful to avoid the deadlock condition that will result if the host is waiting for space to become free in the input FIFO while GLINT is waiting for the host to read data from the output FIFO.

GLINT has a sequence of 1K.times.32 bit addresses in the PCI Region 0 address map called the Graphics Processor FIFO Interface. To read from the output FIFO any address in this range can be read (normally a program will choose the first address and use this as the address for the output FIFO). All 32-bit addresses in this region perform the same function: the range of addresses is provided for data transfer schemes which force the use of incrementing addresses.

Writing to a location in this address range provides raw access to the input FIFO. Again, the first address is normally chosen. Thus the same address can be used for both input and output FIFOs. Reading gives access to the output FIFO; writing gives access to the input FIFO.

Writing to the input FIFO by this method is different from writing to the memory mapped register file. Since the register file has a unique address for each register, writing to this unique address allows GLINT to determine the register for which the write is intended. This allows a tag/data pair to be constructed and inserted into the input FIFO. When writing to the raw FIFO address an address tag description must first be written followed by the associated data. In fact, the format of the tag descriptions and the data that follows is identical to that described above for DMA buffers. Instead of using the GLINT DMA it is possible to transfer data to GLINT by constructing a DMA-style buffer of data and then copying each item in this buffer to the raw input FIFO address. Based on the tag descriptions and data written GLINT constructs tag/data pairs to enter as real FIFO entries. The DMA mechanism can be thought of as an automatic way of writing to the raw input FIFO address.

Note, that when writing to the raw FIFO address the FIFO full condition must still be checked by reading the InFIFOSpace register. However, writing tag descriptions does not cause any entries to be entered into the FIFO: such a write simply establishes a set of tags to be paired with the subsequent data. Thus, free space need be ensured only for actual data items that are written (not the tag values). For example, in the simplest case where each tag is followed by a single data item, assuming that the FIFO is empty, then 32 writes are possible before checking again for free space.

GLINT also provides interrupt facilities for the following:

Sync: If a Sync command is sent and the Sync interrupt has been enabled then once all rendering has been completed, a data value is entered into the Host Out FIFO, and a Sync interrupt is generated when this value reaches the output end of the FIFO. Synchronization is described further in the next section.

External: this provides the capability for external hardware on a GLINT board (such as an external video timing generator) to generate interrupts to the host processor.

Error: if enabled the error interrupt will occur when GLINT detects certain error conditions, such as an attempt to write to a full FIFO.

Vertical Retrace: if enabled a vertical retrace interrupt is generated at the start of the video blank period.

Each of these are enabled and cleared in a similar way to the DMA interrupt.

There are three main cases where the host must synchronize with GLINT:

before reading back from registers

before directly accessing the framebuffer or the localbuffer via the bypass mechanism

framebuffer management tasks such as double buffering

Synchronizing with GLINT implies waiting for any pending DMA to complete and waiting for the chip to complete any processing currently being performed. The following pseudo-code shows the general scheme:

Initially, we wait for DMA to complete as normal. We then have to wait for space to become free in the FIFO (since the DMA controller actually loads the FIFO). We need space for 2 registers: one to enable generation of an output sync value, and the Sync command itself. The enable flag can be set at initialization time. The output value will be generated only when a Sync command has actually been sent, and GLINT has then completed all processing.

Rather than polling it is possible to use a Sync interrupt as mentioned in the previous section. As well as enabling the interrupt and setting the filter mode, the data sent in the Sync command must have the most significant bit set in order to generate the interrupt. The interrupt is generated when the tag or data reaches the output end of the Host Out FIFO. Use of the Sync interrupt has to be considered carefully as GLINT will generally empty the FIFO more quickly than it takes to set up and handle the interrupt.

Normally, the host will access the framebuffer indirectly via commands sent to the GLINT FIFO interface. However, GLINT does provide the whole framebuffer as part of its address space so that it can be memory mapped by an application. Access to the framebuffer via this memory mapped route is independent of the GLINT FIFO.

Drivers may choose to use direct access to the framebuffer for algorithms which are not supported by GLINT. The framebuffer bypass supports big-endian, little-endian and GIB-endian formats.

A driver making use of the framebuffer bypass mechanism should synchronize framebuffer accesses made through the FIFO with those made directly through the memory map. If data is written to the FIFO and then an access is made to the framebuffer, it is possible that the framebuffer access will occur before the commands in the FIFO have been fully processed. This lack of temporal ordering is generally not desirable.

At reset time the hardware stores the size of the framebuffer in the FBMemoryControl register. This register can be read by software to determine the amount of VRAM on the display adapter. For a given amount of VRAM, software can configure different screen resolutions and off-screen memory regions.

The framebuffer width must be set up in the FBReadMode register. The first 9 bits of this register define 3 partial products which determine the offset in pixels from one scanline to the next. Typically, these values will be worked out at initialization time and a copy kept in software. When this register needs to be modified the software copy is retrieved and any other bits modified before writing to the register.

Once the offset from one scanline to the next has been established, determining the visible screen width and height becomes a clipping issue. The visible screen width and height are set up in the ScreenSize register and enabled by setting the ScreenScissorEnable bit in the ScissorMode register.

The framebuffer depth (8, 16 or 32-bit) is controlled by the FBModeSel register. This register provides a 2 bit field to control which of the three pixel depths is being used. The pixel depth can be changed at any time but this should not be attempted without first synchronizing with GLINT. The FBModeSel register is not a FIFO register and is updated immediately it is written. If GLINT is busy performing rendering operations, changing the pixel depth will corrupt that rendering.

Normally, the pixel depth is set at initialization time. To optimize certain 2D rendering operations it may be desirable to change it at other times. For example, if the pixel depth is normally 8 (or 16) bits, changing the pixel depth to 32 bits for the duration of a bitblt can quadruple (or double) the bit speed, when the bit source and destination edges are aligned on 32 bit boundaries. Once such a bit sequence has been set up the host software must wait and synchronize with GLINT and then reset the pixel depth before continuing with further rendering. It is not possible to change the pixel depth via the FIFO, thus explicit synchronization must always be used.

As with the framebuffer, the localbuffer can be mapped in and accessed directly. The host should synchronize with GLINT before making any direct access to the localbuffer.

At reset time the hardware saves the size of the localbuffer in the LBMemoryControl register (localbuffer visible region size). In bypass mode the number of bits per pixel is either 32 or 64. This information is also set in the LBMemoryControl register (localbuffer bypass packing). This pixel packing defines the memory offset between one pixel and the next. A further set of 3 bits (localbuffer width) in the LBMemoryControl register defines the number of valid bits per pixel. A typical localbuffer configuration might be 48 bits per pixel but in bypass mode the data for each pixel starts on a 64-bit boundary. In this case valid pixel data will be contained in bits 0 to 47. Software must set the LBReadFormat register to tell GLINT how to interpret these valid bits.

Host software must set the width in pixels of each scanline of the localbuffer in the LBReadMode FIFO register. The first 9 bits of this register define 3 partial products which determine the offset in pixels from one scanline to the next. As with the framebuffer partial products, these values will usually be worked out at initialization time and a copy kept in software. When this register needs to be modified the software copy is retrieved and any other bits modified before writing to the register. If the system is set up so that each pixel in the framebuffer has a corresponding pixel in the localbuffer then this width will be the same as that set for the framebuffer.

The localbuffer is accessible via Regions 1 and 3 of the PCI address map for GLINT. The localbuffer bypass supports big-endian and little-endian formats. These are described in a later section.

Under some operating environments, multiple tasks will want access to the GLINT chip. Sometimes a server task or driver will want to arbitrate access to GLINT on behalf of multiple applications. In these circumstances, the state of the GLINT chip may need to be saved and restored on each context switch. To facilitate this, the GLINT control registers can be read back. (However, internal and command registers cannot be read back.)

To perform a context switch the host must first synchronize with GLINT. This means waiting for outstanding DMA to complete, sending a Sync command and waiting for the sync output data to appear in the output FIFO. After this the registers can be read back.

To read a GLINT register the host reads the same address which would be used for a write, i.e. the base address of the register file plus the offset value for the register.

Note that since internal registers cannot be read back care must be taken when context switching a task which is making use of continue-draw commands. Continue-draw commands rely on the internal registers maintaining previous state. This state will be destroyed by any rendering work done by a new task. To prevent this, continue-draw commands should be performed via DMA since the context switch code has to wait for outstanding DMA to complete. Alternatively, continue-draw commands can be performed in a non-preemptable code segment.

Normally, reading back individual registers should be avoided. The need to synchronize with the chip can adversely affect performance. It is usually more appropriate to keep a software copy of the register which is updated when the actual register is updated.

Internally GLINT operates in little-endian mode. However, GLINT is designed to work with both big- and little-endian host processors. Since the PCIBus specification defines that byte ordering is preserved regardless of the size of the transfer operation, GLINT provides facilities to handle byte swapping. Each of the Configuration Space, Control Space, Framebuffer Bypass and Localbuffer Bypass memory areas have both big and little endian mappings available. The mapping to use typically depends on the endian ordering of the host processor.

The Configuration Space may be set by a resistor in the board design to be either little endian or big endian.

The Control Space in PCI address region 0, is 128K bytes in size, and consists of two 64K sized spaces. The first 64K provides little endian access to the control space registers; the second 64K provides big endian access to the same registers.

The framebuffer bypass consists of two PCI address regions: Region 2 and Region 4. Each is independently configurable to by the Aperture0 and Aperture 1 control registers respectively, to one of three modes: no byte swap, 16-bit swap, full byte swap. Note that the 16 bit mode is needed for the following reason. If the framebuffer is configured for 16-bit pixels and the host is big-endian then simply byte swapping is not enough when a 32-bit access is made (to write two pixels). In this case, the required effect is that the bytes are swapped within each 16-bit word, but the two 16-bit halves of the 32-bit word are not swapped. This preserves the order of the pixels that are written as well as the byte ordering within each pixel. The 16 bit mode is referred to as GIB-endian in the PCI Multimedia Design Guide, version 1.0.

The localbuffer bypass consists of two PCI address regions: Region 1 and Region 3. Each is independently configurable to by the Aperture0 and Aperture1 control registers respectively, to one of two modes: no byte swap, full byte swap.

To save on the size of the address space required for GLINT, board vendors may choose to turn off access to the big endian regions (3 and 4) by the use of resistors on the board.

There is a bit available in the DMAControl control register to enable byte swapping of DMA data. Thus for big-endian hosts, this control bit would normally be enabled.

For a given graphics board the RAMDAC and/or API will usually force a given interpretation for true color pixel values. For example, 32-bit pixels will be interpreted as either ARGB (alpha at byte 3, red at byte 2, green at byte 1 and blue at byte 0) or ABGR (blue at byte 2 and red at byte 0). The byte position for red and blue may be important for software which has been written to expect one byte order or the other, in particular when handling image data stored in a file.

GLINT provides two registers to specify the byte positions of blue and red internally. In the Alpha Blend Unit the AlphaBlendMode register contains a 1-bit field called ColorOrder. If this bit is set to zero then the byte ordering is ABGR; if the bit is set to one then the ordering is ARGB. As well as setting this bit in the Alpha Blend unit, it must also be set in the Color Formatting unit. In this unit the DitherMode register contains a Color Order bit with the same interpretation. The order applies to all of the true color pixel formats, regardless of the pixel depth.

Hardware Data Structures

Some of the hardware data structure implementations used in the presently preferred embodiment will now be described in detail. Of course these examples are provided merely to illustrate the presently preferred embodiment in great detail, and do not necessarily delimit any of the claimed inventions.

The localbuffer holds the per pixel information corresponding to each displayed pixel and any texture maps. The per pixel information held in the localbuffer are Graphic ID (GID), Depth, Stencil and Frame Count Planes (FCP). The possible formats for each of these fields, and their use are covered individually in the following sections.

The maximum width of the localbuffer is 48 bits, but this can be reduced by changing the external memory configuration, albeit at the expense of reducing the functionality or dynamic range of one or more of the fields.

The localbuffer memory can be from 16 bits (assuming a depth buffer is always needed) to 48 bits wide in steps of 4 bits. The four fields supported in the localbuffer, their allowed lengths and positions are shown in the following table:

The order of the fields is as shown with the depth field at the least significant end and GID field at the most significant end. The GID is at the most significant end so that various combinations of the Stencil and FrameCount field widths can be used on a per window basis without the position of the GID fields moving. If the GID field is in a different positions in different windows then the ownership tests become impossible to do.

The GID, FrameCount, Stencil and Depth fields in the localbuffer are converted into the internal format by right justification if they are less than their internal widths, i.e. the unused bits are the most significant bits and they are set to 0.

The format of the localbuffer is specified in two places: the LBReadFormat register and the LBWriteFormat register.

It is still possible to part populate the localbuffer so other combinations of the field widths are possible (i.e. depth field width of 0), but this may give problems if texture maps are to be stored in the localbuffer as well.

Any non-bypass read or write to the localbuffer always reads or writes all 48 bits simultaneously.

The 4 bit GID field is used for pixel ownership tests to allow per pixel window clipping. Each window using this facility is assigned one of the GID values, and the visible pixels in the window have their GID field set to this value. If the test is enabled the current GID (set to correspond with the current window) is compared with the GID in the localbuffer for each fragment. If they are equal this pixel belongs to the window so the localbuffer and framebuffer at this coordinate may be updated.

Using the GID field for pixel ownership tests is optional and other methods of achieving the same result are:

clip the primitive to the window's boundary (or rectangular tiles which make up the window's area) and render only the visible parts of the primitive

use the scissor test to define the rectangular tiles which make up the window's visible area and render the primitive once per tile (This may be limited to only those tiles which the primitive intersects).

The depth field holds the depth (Z) value associated with a pixel and can be 16, 24 or 32 bits wide.

The stencil field holds the stencil value associated with a pixel and can be 0, 4 or 8 bits wide.

The width of the stencil buffer is also stored in the StencilMode register and is needed for clamping and masking during the update methods. The stencil compare mask should be set up to exclude any absent bits from the stencil compare operation.

The Frame Count Field holds the frame count value associated with a pixel and can be 0, 4 or 8 bits wide. It is used during animation to support a fast clear mechanism to aid the rapid clearing of the depth and/or stencil fields needed at the start of each frame.

In addition to the fast clear mechanism the extent of all updates to the localbuffer and framebuffer can be recorded (MinRegion and MaxRegion registers) and read back (MinHitRegion and MaxHitRegion commands) to give the bounding box of the smallest area to clear. For some applications this will be significantly smaller than the whole window or screen, and hence faster.

The fast clear mechanism provides a method where the cost of clearing the depth and stencil buffers can be amortized over a number of clear operations issued by the application. This works as follows:

The window is divided up into n regions, where n is the range of the frame counter (16 or 256). Every time the application issues a clear command the reference frame counter is incremented (and allowed to roll over if it exceeds its maximum value) and the nth region is cleared only. The clear updates the depth and/or stencil buffers to the new values and the frame count buffer with the reference value. This region is much smaller than the full window and hence takes less time to clear.

When the localbuffer is subsequently read and the frame count is found to be the same as the reference frame count (held in the Window register) the localbuffer data is used directly. However, if the frame count is found to be different from the reference frame count (held in the Window register) the data which would have been written, if the localbuffer had been cleared properly, is substituted for the stale data returned from the read. Any new writes to the localbuffer will set the frame count to the reference value so the next read on this pixel works normally without the substitution. The depth data to substitute is held in the FastClearDepth register and the stencil data to substitute is held in the StencilData register (along with other stencil information).

The fast clear mechanism does not present a total solution as the user can elect to clear just the stencil planes or just the depth planes, or both. The situation where the stencil planes only are `cleared` using the fast clear method, then some rendering is done and then the depth planes are `cleared` using the fast clear will leave ambiguous pixels in the localbuffer. The driver software will need to catch this situation, and fall back to using a per pixel write to do the second clear. Which field(s) the frame count plane refers to is recorded in the Window register.

When clear data is substituted for real memory data (during normal rendering operations) the depth write mask and stencil write masks are ignored to mimic the OpenGL operation when a buffer is cleared.

The coordinates generated by the rasterizer are 16 bit 2's complement numbers, and so have the range +32767 to -32768. The rasterizer will produce values in this range, but any which have a negative coordinate, or exceed the screen width or height (as programmed into the ScreenSize register) are discarded.

Coordinates can be defined window relative or screen relative and this is only relevant when the coordinate gets converted to an actual physical address in the localbuffer. In general it is expected that the windowing system will use absolute coordinates and the graphics system will use relative coordinates (to be independent of where the window really is).

GUI systems (such as Windows, Windows NT and X) usually have the origin of the coordinate system at the top left corner of the screen but this is not true for all graphics systems. For instance OpenGL uses the bottom left corner as its origin. The WindowOrigin bit in the LBReadMode register selects the top left (0) or bottom left (1) as the origin.

The actual equations used to calculate the localbuffer address to read and write are:

These address calculations translate a 2D address into a linear address.

The Screen width is specified as the sum of selected partial products so a full multiply operation is not needed. The partial products are selected by the fields PP0, PP1 and PP2 in the LBReadMode register.

For arbitrary width screens, for instance bitmaps in `off screen` memory, the next largest width from the table must be chosen. The difference between the table width and the bitmap width will be an unused strip of pixels down the right hand side of the bitmap.

Note that such bitmaps can be copied to the screen only as a series of scanlines rather than as a rectangular block. However, often windowing systems store offscreen bitmaps in rectangular regions which use the same stride as the screen. In this case normal bitblts can be used.

The localbuffer is used to hold textures in the GLINT 400TX variant. In the GLINT 300TX variant the texture information is supplied by the host.

The framebuffer is a region of memory where the information produced during rasterization is written prior to being displayed. This information is not restricted to color but can include window control data for LUT management and double buffering.

The framebuffer region can hold up to 32 MBytes and there are very few restrictions on the format and size of the individual buffers which make up the video stream. Typical buffers include:

True color or color index main planes,

Overlay planes,

Underlay planes,

Window ID planes for LUT and double buffer management,

Cursor planes.

Any combination of these planes can be supported up to a maximum of 32 MBytes, but usually it is the video level processing which is the limiting factor. The following text examines the options and choices available from GLINT for rendering, copying, etc. data to these buffers.

To access alternative buffers either the FBPixelOffset register can be loaded, or the base address of the window held in the FBWindow-Base register can be redefined. This is described in more detail below.

Each buffer resides at an address in the framebuffer memory map. For rendering and copying operations the actual buffer addresses can be on any pixel boundary. Display hardware will place some restrictions on this as it will need to access the multiple buffers in parallel to mix the buffers together depending on their relative priority, opacity and double buffer selection. For instance, visible buffers (rather than offscreen bitmaps) will typically need to be on a page boundary.

Consider the following highly configured example with a 1280.times.1024 double buffered system with 32 bit main planes (RGBA), 8 bit overlay and 4 bits of window control information (WID).

Combining the WID and overlay planes in the same 32 bit pixel has the advantage of reducing the amount of data to copy when a window moves, as only two copies are required--one for the main planes and one for the overlay and WID planes.

Note the position of the overlay and WID planes. This was not an arbitrary choice but one imposed by the (presumed) desire to use the color processing capabilities of GLINT (dither and interpolation) in the overlay planes. The conversion of the internal color format to the external one stored in the framebuffer depends on the size and position of the component. Note that GLINT does not support all possible configurations. For example; if the overlay and WID bits were swapped, then eight bit color index starting at bit 4 would be required to render to the overlay, but this is not supported.

Coordinate generation for the framebuffer is similar to that for the localbuffer, but there are some key differences.

As was mentioned before, the coordinates generated by the rasterizer are 16 bit 2's complement numbers. Coordinates can be defined as window relative or screen relative, though this is only relevant when the coordinate gets converted to an actual physical address in the framebuffer. The WindowOrigin bit in the FBReadMode register selects top left (0) or bottom left (1) as the origin for the framebuffer.

The actual equations used to calculate the framebuffer address to read and write are:

These address calculations translate a 2D address into a linear address, so non power of two framebuffer widths (i.e. 1280) are economical in memory.

The width is specified as the sum of selected partial products so a full multiply operation is not needed. The partial products are selected by the fields PP0, PP1 and PP2 in the FBReadMode register. This is the same mechanism as is used to set the width of the localbuffer, but the widths may be set independently. For arbitrary screen sizes, for instance when rendering to `off screen` memory such as bitmaps the next largest width from the table must be chosen. The difference between the table width and the bitmap width will be an unused strip of pixels down the right hand side of the bitmap.

Note that such bitmaps can be copied to the screen only as a series of scanlines rather than as a rectangular block. However, often windowing systems store offscreen bitmaps in rectangular regions which use the same stride as the screen. In this case normal bitblts can be used.

The contents of the framebuffer can be regarded in two ways:

As a collection of fields of up to 32 bits with no meaning or assumed format as far as GLINT is concerned. Bit planes may be allocated to control cursor, LUT, multi-buffer visibility or priority functions. In this case GLINT will be used to set and clear bit planes quickly but not perform any color processing such as interpolation or dithering. All the color processing can be disabled so that raw reads and writes are done and the only operations are write masking and logical ops. This allows the control planes to be updated and modified as necessary. Obviously this technique can also be used for overlay buffers, etc. providing color processing is not required.

As a collection of one or more color components. All the processing of color components, except for the final write mask and logical ops are done using the internal color format of 8 bits per red, green, blue and alpha color channels. The final stage before write mask and logical ops processing converts the internal color format to that required by the physical configuration of the framebuffer and video logic. The nomenclature n@m means this component is n bits wide and starts at bit position m in the framebuffer. The least significant bit position is 0 and a dash in a column indicates that this component does not exist for this mode. The ColorOrder is specified by a bit in the DitherMode register.

Some important points to note:

The alpha channel is always associated with the RGB color channels rather than being a separate buffer. This allows it to be moved in parallel and to work correctly in multi-buffer updates and double buffering. If the framebuffer is not configured with an alpha channel (e.g. 24 bit framebuffer width with 8:8:8:8 RGB format) then some of the rendering modes which use the retained alpha buffer cannot be used. In these cases the NoAlphaBuffer bit in the AlphaBlendMode register should be set so that an alpha value of 255 is substituted. For the RGB modes where no alpha channel is present (e.g. 3:3:2) then this substitution is done automatically.

For the Front and Back modes the data value is replicated into both buffers.

All writes to the framebuffer try to update all 32 bits irrespective of the color format. This may not matter if the memory planes don't exist, but if they are being used (as overlay planes, for example) then the write masks (FBSoftwareWriteMask or FBHardwareWriteMask) must be set up to protect the alternative planes.

When reading the framebuffer RGBA components are scaled to their internal width of 8 bits, if needed for alpha blending.

CI values are left justified with the unused bits (if any) set to zero and are subsequently processed as the red component. The result is replicated into each of the streams G,B and A giving four copies for CI8 and eight copies for CI4.

The 4:4:4:4 Front and Back formats are designed to support 12 bit double buffering with 4 bit Alpha, in a 32 bit system.

The 3:3:2 Front and Back formats are designed to support 8 bit double buffering in a 16 bit system.

The 1:2:1 Front and Back formats are designed to support 4 bit double buffering in an 8 bit system.

It is possible to have a color index buffer at other positions as long as reduced functionality is acceptable. For example a 4 bit CI buffer at bit position 16 can be achieved using write masking and 4:4:4:4 Front format with color interpolation, but dithering is lost.

The format information needs to be stored in two places: the DitherMode register and the AlphaBlendMode register.

In a GUI system there are two possible relationships between the overlay planes (or underlay) and the main planes.

The overlay planes are fixed to the main planes, so that if the window is moved then both the data in the main planes and overlay planes move together.

The overlay planes are not fixed to the main planes but floating, so that moving a window only moves the associated main or overlay planes.

In the fixed case both planes can share the same GID. The pixel offsets is used redirect the reads and writes between the main planes and the underlay) buffer. The pixel ownership tests using the GID field in the localbuffer work as expected.

In the floating case different GIDs are the best choice, because the same GID planes in the localbuffer can not be used for pixel ownership tests. The alternatives are not to use the GID based pixel ownership tests for one of the buffers but rely on the scissor clipping, or to install a second set of GID planes so each buffer has it's own set. GLINT allows either approach.

If rendering operations to the main and overlay planes both need the depth or stencil buffers, and the windows in each overlap then each buffer will need its own exclusive depth and/or stencil buffers. This is easily achieved with GLINT by assigning different regions in the localbuffer to each of the buffers. Typically this would double the localbuffer memory requirements.

One scenario where the above two considerations do not cause problems, is when the overlay planes are used exclusively by the GUI system, and the main planes are used for the 3D graphics.

High performance systems will typically use VRAM for the framebuffer and the extended functionality of VRAM over DRAM can be used to enhance performance for many rendering tasks.

These allow write masking in the framebuffer without incurring a performance penalty. If hardware write masks are not available, GLINT must be programmed to read the memory, merge the value with the new value using the write mask, and write it back.

To use hardware write masking, the required write mask is written to the FBHardwareWriteMask register, the FBSoftwareWriteMask register should be set to all 1's, and the number of framebuffer reads is set to 0 (for normal rendering). This is achieved by clearing the ReadSource and ReadDestination enables in the FBReadMode register.

To use software write masking, the required write mask is written to the FBSoftwareWriteMask register and the number of framebuffer reads is set to 1 (for normal rendering). This is achieved by setting the ReadDestination enable in the FBReadMode register.

Block writes cause consecutive pixels in the framebuffer to be written simultaneously. This is useful when filling large areas but does have some restrictions:

No pixel level clipping is available;

No depth or stencil testing can be done;

All the pixels must be written with the same value so no color interpolation, blending, dithering or logical ops can be done; and

The area is defined in screen relative coordinates.

Block writes are not restricted to rectangular areas and can be used for any trapezoid. Hardware write masking is available during block writes.

The following registers need to be set up before block fills can be used:

FBBlockColor register with the value to write to each pixel; and

FBWriteMode register with the block width field.

Sending a Render command with the PrimitiveType field set to "trapezoid" and the FastFillEnable and FastFillIncrement fields set up will then cause block filling of the area. Note that during a block fill of a trapezoid any inappropriate state is ignored so even if color interpolation, depth testing and logical ops, for example, are enabled they have no effect.

The block sizes supported are 8, 16 and 32 pixels. GLINT takes care of filling any partial blocks at the end of spans.

Graphics Programming

GLINT provides a rich variety of operations for 2D and 3D graphics supported by its Pipelined architecture.

This section describes each of the units in the graphics Pipeline. FIG. 2C shows a schematic of the pipeline. In this diagram, the localbuffer contains the pixel ownership values (known as Graphic IDs), the FrameCount Planes (FCP), Depth (Z) and Stencil buffer. The framebuffer contains the Red, Green, Blue and Alpha bitplanes. The operations in the Pipeline include:

Rasterizer scan converts the given primitive into a series of fragments for processing by the rest of the pipeline.

Scissor Test clips out fragments that lie outside the bounds of a user defined scissor rectangle and also performs screen clipping to stop illegal access outside the screen memory.

Stipple Test masks out certain fragments according to a specified pattern. Line and area stipples are available.

Color DDA is responsible for generating the color information (True Color RGBA or Color Index(CI)) associated with a fragment.

Texture is concerned with mapping a portion of a specified image (texture) onto a fragment. The process involves filtering to calculate the texture color, and application which applies the texture color to the fragment color.

Fog blends a fog color with a fragment's color according to a given fog factor. Fogging is used for depth cuing images and to simulate atmospheric fogging.

Antialias Application combines the incoming fragment's alpha value with its coverage value when antialiasing is enabled.

Alpha Test conditionally discards a fragment based on the outcome of a comparison between the fragments alpha value and a reference alpha value.

Pixel Ownership is concerned with ensuring that the location in the framebuffer for the current fragment is owned by the current visual. Comparison occurs between the given fragment and the Graphic ID value in the localbuffer, at the corresponding location, to determine whether the fragment should be discarded.

Stencil Test conditionally discards a fragment based on the outcome of a test between the given fragment and the value in the stencil buffer at the corresponding location. The stencil buffer is updated dependent on the result of the stencil test and the depth test.

Depth Test conditionally discards a fragment based on the outcome of a test between the depth value for the given fragment and the value in the depth buffer at the corresponding location. The result of the depth test can be used to control the updating of the stencil buffer.

Alpha Blending combines the incoming fragment's color with the color in the framebuffer at the corresponding location.

Color Formatting converts the fragment's color into the format in which the color information is stored in the framebuffer. This may optionally involve dithering.

The Pipeline structure of GLINT is very efficient at processing fragments, for example, texture mapping calculations are not actually performed on fragments that get clipped out by scissor testing. This approach saves substantial computational effort. The pipelined nature does however mean that when programming GLINT one should be aware of what all the pipeline stages are doing at any time. For example, many operations require both a read and/or write to the localbuffer and framebuffer; in this case it is not sufficient to set a logical operation to XOR and enable logical operations, but it is also necessary to enable the reading/writing of data from/to the framebuffer.

We may now revisit the "day in the life of a triangle" example given above, and review the actions taken in greater detail. Again, the primitive being rendered will be a Gouraud shaded, depth buffered triangle. For this example assume that the triangle is to be drawn into a window which has its colormap set for RGB as opposed to color index operation. This means that all three color components; red, green and blue, must be handled. Also, assume the coordinate origin is bottom left of the window and drawing will be from top to bottom. GLINT can draw from top to bottom or bottom to top.

Consider a triangle with vertices, v.sub.1, v.sub.2 and V.sub.3 where each vertex comprises X, Y and Z coordinates. Each vertex has a different color made up of red, green and blue (R, G and B) components. The alpha component will be omitted for this example.

GLINT requires many of its registers to be initialized in a particular way, regardless of what is to be drawn, for instance, the screen size and appropriate clipping must be set up. Normally this only needs to be done once and for clarity this example assumes that all initialization has already been done.

Other state will change occasionally, though not usually on a per primitive basis, for instance enabling Gouraud shading and depth buffering.

As shown in FIG. 4A, the dominant side of a triangle is that with the greatest range of Y values. The choice of dominant side is optional when the triangle is either flat bottomed or flat topped.

GLINT always draws triangles starting from the dominant edge towards the subordinate edges. This simplifies the calculation of set up parameters as will be seen below.

These values allow the color of each fragment in the triangle to be determined by linear interpolation. For example, the red component color value of a fragment at X.sub.n,)Y.sub.m could be calculated by:

adding dRdy.sub.13, for each scanline between Y.sub.1 and Y.sub.n, to R.sub.1.

then adding dRdx for each fragment along scanline Y.sub.n from the left edge to X.sub.n.

The example chosen has the `knee,`0 i.e. vertex 2, on the right hand side, and drawing is from left to right. If the knee were on the left side (or drawing was from right to left), then the Y deltas for both the subordinate sides would be needed to interpolate the start values for each color component (and the depth value) on each scanline. For this reason GLINT always draws triangles starting from the dominant edge and towards the subordinate edges. For the example triangle, this means left to right.

For the example triangle, the GLINT registers must be set as follows, for color interpolation. Note that the format for color values is 24 bit, fixed point 2's complement.

To draw from left to right and top to bottom, the depth gradients (or deltas) required for interpolation are: ##EQU1## And from the plane equation: ##EQU2## where

The divisor, shown here as c, is the same as for color gradient values. The two deltas dZdy.sub.13 and dZdx allow the Z value of each fragment in the triangle to be determined by linear interpolation, just as for the color interpolation.

Internally GLINT uses fixed point arithmetic. Each depth value must be converted into a 2's complement 32.16 bit fixed point number and then loaded into the appropriate pair of 32 bit registers. The `Upper` or `U` registers store the integer portion, whilst the `Lower` or `L` registers store the 16 fractional bits, left justified and zero filled.

For the example triangle, GLINT would need its registers set up as follows:

GLINT draws filled shapes such as triangles as a series of spans with one span per scanline. Therefore it needs to know the start and end X coordinate of each span. These are determined by `edge walking`. This process involves adding one delta value to the previous span's start X coordinate and another delta value to the previous span's end x coordinate to determine the X coordinates of the new span. These delta values are in effect the slopes of the triangle sides. To draw from left to right and top to bottom, the slopes of the three sides are calculated as: ##EQU3##

This triangle will be drawn in two parts, top down to the `knee` (i.e. vertex 2), and then from there to the bottom. The dominant side is the left side so for the top half:

The start X,Y, the number of scanlines, and the above deltas give GLINT enough information to edge walk the top half of the triangle. However, to indicate that this is not a flat topped triangle (GLINT is designed to rasterize screen aligned trapezoids and flat topped triangles), the same start position in terms of X must be given twice as StartXDom and StartXSub.

To edge walk the lower half of the triangle, selected additional information is required. The slope of the dominant edge remains unchanged, but the subordinate edge slope needs to be set to:

Also the number of scanlines to be covered from Y.sub.2 to Y.sub.3 needs to be given. Finally to avoid any rounding errors accumulated in edge walking to X.sub.2 (which can lead to pixel errors), StartXSub must be set to X.sub.2.

The GLINT rasterizer has a number of modes which have effect from the time they are set until they are modified and can thus affect many primitives. In the case of the Gouraud shaded triangle the default value for these modes are suitable.

GLINT can perform subpixel correction of all interpolated values when rendering aliased trapezoids. This correction ensures that any parameter (color/depth/texture/fog) is correctly sampled at the center of a fragment. Subpixel correction will generally always be enabled when rendering any trapezoid which is smooth shaded, textured, fogged or depth buffered. Control of subpixel correction is in the Render command register described in the next section, and is selectable on a per primitive basis.

GLINT is almost ready to draw the triangle. Setting up the registers as described here and sending the Render command will cause the top half of the example triangle to be drawn.

For drawing the example triangle, all the bit fields within the Render command should be set to 0 except the PrimitiveType which should be set to trapezoid and the SubPixelCorrectionEnable bit which should be set to TRUE.

After the Render command has been issued, the registers in GLINT can immediately be altered to draw the lower half of the triangle. Note that only two registers need be loaded and the command ContinueNewSub sent. Once GLINT has received ContinueNewSub, drawing of this sub-triangle will begin.

The rasterizer decomposes a given primitive into a series of fragments for processing by the rest of the Pipeline.

GLINT can directly rasterize:

aliased screen aligned trapezoids

aliased single pixel wide lines

aliased single pixel points

antialiased screen aligned trapezoids

antialiased circular points

All other primitives are treated as one or more of the above, for example an antialiased line is drawn as a series of antialiased trapezoids.

GLINT's basic area primitives are screen aligned trapezoids. These are characterized by having top and bottom edges parallel to the X axis. The side edges may be vertical (a rectangle), but in general will be diagonal. The top or bottom edges can degenerate into points in which case we are left with either flat topped or flat bottomed triangles. Any polygon can be decomposed into screen aligned trapezoids or triangles. Usually, polygons are decomposed into triangles because the interpolation of values over non-triangular polygons is ill defined. The rasterizer does handle flat topped and flat bottomed `bow tie` polygons which are a special case of screen aligned trapezoids.

To render a triangle, the approach adopted to determine which fragments are to be drawn is known as `edge walking`. Suppose the aliased triangle shown in FIG. 4A was to be rendered from top to bottom and the origin was bottom left of the window. Starting at (X1, Y1) then decrementing Y and using the slope equations for edges 1-2 and 1-3, the intersection of each edge on each scanline can be calculated. This results in a span of fragments per scanline for the top trapezoid. The same method can be used for the bottom trapezoid using slopes 2-3 and 1-3.

It is usually required that adjacent triangles or polygons which share an edge or vertex are drawn such that pixels which make up the edge or vertex get drawn exactly once. This may be achieved by omitting the pixels down the left or the right sides and the pixels along the top or lower sides. GLINT has adopted the convention of omitting the pixels down the right hand edge. Control of whether the pixels along the top or lower sides are omitted depends on the start Y value and the number of scanlines to be covered. With the example, if StartY=Y1 and the number of scanlines is set to Y1-Y2, the lower edge of the top half of the triangle will be excluded. This excluded edge will get drawn as part of the lower half of the triangle.

To minimize delta calculations, triangles may be scan converted from left to right or from right to left. The direction depends on the dominant edge, that is the edge which has the maximum range of Y values. Rendering always proceeds from the dominant edge towards the relevant subordinate edge. In the example above, the dominant edge is 1-3 so rendering will be from right to left.

The sequence of actions required to render a triangle (with a `knee`) is:

Load the edge parameters and derivatives for the dominant edge and the first subordinate edges in the first triangle.

Send the Render command. This starts the scan conversion of the first triangle, working from the dominant edge. This means that for triangles where the knee is on the left we are scanning right to left, and vice versa for triangles where the knee is on the right.

Load the edge parameters and derivatives for the remaining subordinate edge in the second triangle.

Send the ContinueNewSub command. This starts the scan conversion of the second triangle.

Pseudocode for the above example is:

After the Render command has been sent, the registers in GLINT can immediately be altered to draw the second half of the triangle. For this, note that only two registers need be loaded and the command ContinueNewSub be sent. Once drawing of the first triangle is complete and GLINT has received the ContinueNewSub command, drawing of this sub-triangle will start. The ContinueNewSub command register is loaded with the remaining number of scanlines to be rendered.

Single pixel wide aliased lines are drawn using a DDA algorithm, so all GLINT needs by way of input data is StartX, StartY, dX, dY and length.

For polylines, a ContinueNewLine command (analogous to the Continue command used at the knee of a triangle) is used at vertices.

When a Continue command is issued some error will be propagated along the line. To minimize this, a choice of actions are available as to how the DDA units are restarted on the receipt of a Continue command. It is recommended that for OpenGL rendering the ContinueNewLine command is not used and individual segments are rendered.

Antialiased lines, of any width, are rendered as antialiased screen-aligned trapezoids.

GLINT supports a single pixel aliased point primitive. For points larger than one pixel trapezoids should be used. In this case the PrimitiveType field in the Render command should be set to equal GLINT.sub.-- POINT.sub.-- PRIMITIVE.

GLINT uses a subpixel point sampling algorithm to antialias primitives. GLINT can directly rasterize antialiased trapezoids and points. Other primitives are composed from these base primitives.

The rasterizer associates a coverage value with each fragment produced when antialiasing. This value represents the percentage coverage of the pixel by the fragment. GLINT supports two levels of antialiasing quality:

normal, which represents 4.times.4 pixel subsampling

high, which represents 8.times.8 pixel subsampling.

Selection between these two is made by the AntialiasingQuality bit within the Render command register.

When rendering antialiased primitives with GLINT the FlushSpan command is used to terminate rendering of a primitive. This is due to the nature of GLINT antialiasing. When a primitive is rendered which does not happen to complete on a scanline boundary, GLINT retains antialiasing information about the last sub-scanline(s) it has processed, but does not generate fragments for them unless a FlushSpan command is received. The commands ContinueNewSub, ContinueNewDom or Continue can then be used, as appropriate, to maintain continuity between adjacent trapezoids. This allows complex antialiased primitives to be built up from simple trapezoids or points.

To illustrate this consider using screen aligned trapezoids to render an antialiased line. The line will in general consist of three screen aligned trapezoids as shown in FIG. 4B. This Figure illustrates the sequence of rendering an Antialiased Line primitive. Note that the line has finite width.

The procedure to render the line is as follows:

Note that when rendering antialiased primitives, any count values should be given in subscanlines, for example if the quality is 4.times.4 then any scanline count must be multiplied by 4 to convert it into a subscanline count. Similarly, any delta value must be divided by 4.

When rendering, AntialiasEnable must be set in the Antialias-Mode register to scale the fragments color by the coverage value. An appropriate blending function should also be enabled.

Note, when rendering antialiased bow-ties, the coverage value on the cross-over scanline may be incorrect.

GLINT can render small antialiased points. Antialiased points are treated as circles, with the coverage of the boundary fragments ranging from 0% to 100%. GLINT supports:

point radii of 0.5 to 16.0 in steps of 0.25 for 4.times.4 antialiasing

point radii of 0.25 to 8.0 in steps of 0.125 for 8.times.8 antialiasing

To scan convert an antialiased point as a circle, GLINT traverses the boundary in sub scanline steps to calculate the coverage value. For this, the sub-scanline intersections are calculated incrementally using a small table. The table holds the change in X for a step in Y. Symmetry is used so the table only holds the delta values for one quadrant.

StartXDom, StartXSub and StartY are set to the top or bottom of the circle and dY set to the subscanline step. In the case of an even diameter, the last of the required entries in the table is set to zero.

Since the table is configurable, point shapes other than circles can be rendered. Also if the StartXDom and StartXSub values are not coincident then horizontal thick lines with rounded ends, can be rendered.

GLINT supports VRAM block writes with block sizes of 8, 16 and 32 pixels. The block write method does have some restrictions: None of the per pixel clipping, stipple, or fragment operations are available with the exception of write masks. One subtle restriction is that the block coordinates will be interpreted as screen relative and not window relative when the pixel mask is calculated in the Framebuffer Units.

Any screen aligned trapezoid can be filled using block writes, not just rectangles.

The use of block writes is enabled by setting the FastFillEnable and FastFilllncrement fields in the Render command register. The framebuffer write unit must also be configured.

Note only the Rasterizer, Framebuffer Read and Framebuffer Write units are involved in block filling. The other units will ignore block write fragments, so it is not necessary to disable them.

As the rasterizer has 16 bits of fraction precision, and the screen width used is typically less than 216 wide a number of bits called subpixel precision bits, are available. Consider a screen width of 4096 pixels. This figure gives a subpixel precision of 4 bits (4096=2.sup.12). The extra bits are required for a number of reasons:

antialiasing (where vertex start positions can be supplied to subpixel precision)

when using an accumulation buffer (where scans are rendered multiple times with jittered input vertices)

for correct interpolation of parameters to give high quality shading as described below

GLINT supports subpixel correction of interpolated values when rendering aliased trapezoids. Subpixel correction ensures that all interpolated parameters associated with a fragment (color, depth, fog, texture) are correctly sampled at the fragment's center. This correction is required to ensure consistent shading of objects made from many primitives. It should generally be enabled for all aliased rendering which uses interpolated parameters.

Subpixel correction is not applied to antialiased primitives.

A Bitmap primitive is a trapezoid or line of ones and zeros which control which fragments are generated by the rasterizer. Only fragments where the corresponding Bitmap bit is set are submitted for drawing. The normal use for this is in drawing characters, although the mechanism is available for all primitives. The Bitmap data is packed contiguously into 32 bit words so that rows are packed adjacent to each other. Bits in the mask word are by default used from the least significant end towards the most significant end and are applied to pixels in the order they are generated in.

The rasterizer scans through the bits in each word of the Bitmap data and increments the X,Y coordinates to trace out the rectangle of the given width and height. By default, any set bits (1) in the Bitmap cause a fragment to be generated, any reset bits (0) cause the fragment to be rejected.

The selection of bits from the BitMaskPattern register can be mirrored, that is, the pattern is traversed from MSB to LSB rather than LSB to MSB. Also, the sense of the test can be reversed such that a set bit causes a fragment to be rejected and vice versa. This control is found in the RasterirerMode register.

When one Bitmap word has been exhausted and pixels in the rectangle still remain then rasterization is suspended until the next write to the BitMaskPattern register. Any unused bits in the last Bitmap word are discarded.

GLINT supports three "pixel rectangle" operations: copy, upload and download. These can apply to the Depth or Stencil Buffers (held within the localbuffer) or the framebuffer.

It should be emphasized that the GLINT copy operation moves RAW blocks of data around buffers. To zoom or re-format data, in the presently preferred embodiment, external software must upload the data, process it and then download it again.

To copy a rectangular area, the rasterizer would be configured to render the destination rectangle, thus generating fragments for the area to be copied. GLINT copy works by adding a linear offset to the destination fragment's address to find the source fragment's address.

Note that the offset is independent of the origin of the buffer or window, as it is added to the destination address. Care must be taken when the source and destination overlap to choose the source scanning direction so that the overlapping area is not overwritten before it has been moved. This may be done by swapping the values written to the StartXDom and StartXSub, or by changing the sign of dY and setting StartY to be the opposite side of the rectangle.

Localbuffer copy operations are correctly tested for pixel ownership. Note that this implies two reads of the localbuffer, one to collect the source data, and one to get the destination GID for the pixel ownership test.

GLINT buffer upload/downloads are very similar to copies in that the region of interest is generated in the rasterizer. However, the localbuffer and framebuffer are generally configured to read or to write only, rather than both read and write. The exception is that an image load may use pixel ownership tests, in which case the localbuffer destination read must be enabled.

Units which can generate fragment values, the color DDA unit for example, should generally be disabled for any copy/upload/download operations.

Warning: During image upload, all the returned fragments must be read from the Host Out FIFO, otherwise the GLINT pipeline will stall. In addition it is strongly recommended that any units which can discard fragments (for instance the following tests: bitmask, alpha, user scissor, screen scissor, stipple, pixel ownership, depth, stencil), are disabled otherwise a shortfall in pixels returned may occur, also leading to deadlock.

Note that because the area of interest in copy/upload/download operations is defined by the rasterizer, it is not limited to rectangular regions.

Color formatting can be used when performing image copies, uploads and downloads. This allows data to be formatted from, or to, any of the supported GLINT color formats.

A number of long-term modes can be set using the Rasterizer-Mode register, these are:

Mirror BitMask: This is a single bit flag which specifies the direction bits are checked in the BitMask register. If the bit is reset, the direction is from least significant to most significant (bit 0 to bit 31), if the bit is set, it is from most significant to least significant (from bit 31 to bit 0).

Invert BitMask: This is a single bit which controls the sense of the accept/reject test when using a Bitmask. If the bit is reset then when the BitMask bit is set the fragment is accepted and when it is reset the fragment is rejected. When the bit is set the sense of the test is reversed.

Fraction Adjust: These 2 bits control the action taken by the rasterizer on receiving a ContinueNewLine command. As GLINT uses a DDA algorithm to render lines, an error accumulates in the DDA value. GLINT provides for greater control of the error by doing one of the following:

leaving the DDA running, which means errors will be propagated along a line.

or setting the fraction bits to either zero, a half or almost a half (0.times.7FFF).

Bias Coordinates: Only the integer portion of the values in the DDAs are used to generate fragment addresses. Often the actual action required is a rounding of values, this can be achieved by setting the bias coordinate bit to true which will automatically add almost a half (0.times.7FFF) to all input coordinates.

Real coordinates with fractional parts are provided to the rasterizer in 2's complement 16 bit integer, 16 bit fraction format. The following Table lists the command registers which control the rasterizer unit:

The following Table shows the control registers of the rasterizer, in the presently preferred embodiment:

For efficiency, the Render command register has a number of bit fields that can be set or cleared per render operation, and which qualify other state information within GLINT. These bits are AreaStippleEnable, LineStippleEnable, ResetLineStipple, TextureEnable FogEnable, CoverageEnable and SubpixelCorrection.

One use of this feature can occur when a window is cleared to a background color. For normal 3D primitives, stippling and fog operations may have been enabled, but these are to be ignored for window clears. Initially the FogMode, AreaStippleMode and LineStippleMode registers are enabled through the UnitEnable bits. Now bits need only be set or cleared within the Render command to achieve the required result, removing the need for the FogMode, AreaStippleMode and LineStippleMode registers to be loaded for every render operation.

The bitfields of the Render command register, in the presently preferred embodiment, are detailed below:

A number of long-term rasterizer modes are stored in the RasterizerMode register as shown below:

Two scissor tests are provided in GLINT, the User Scissor test and the Screen Scissor test. The user scissor checks each fragment against a user supplied scissor region; the screen scissor checks that the fragment lies within the screen.

This test may reject fragments if some part of a window has been moved off the screen. It will not reject fragments if part of a window is simply overlapped by another window (GID testing can be used to detect this).

Stippling is a process whereby each fragment is checked against a bit in a defined pattern, and is rejected or accepted depending on the result of the stipple test. If it is rejected it undergoes no further processing; otherwise it proceeds down the pipeline. GLINT supports two types of stippling, line and area.

A 32.times.32 bit area stipple pattern can be applied to fragments. The least significant n bits of the fragment's (X,Y) coordinates, index into a 2D stipple pattern. If the selected bit in the pattern is set, then the fragment passes the test, otherwise it is rejected. The number of address bits used, allow regions of 1,2,4,8,16 and 32 pixels to be stippled. The address selection can be controlled independently in the X and Y directions. In addition the bit pattern can be inverted or mirrored. Inverting the bit pattern has the effect of changing the sense of the accept/reject test. If the mirror bit is set the most significant bit of the pattern is towards the left of the window, the default is the converse.

In some situations window relative stippling is required but coordinates are only available screen relative. To allow window relative stippling, an offset is available which is added to the coordinates before indexing the stipple table. X and Y offsets can be controlled independently.

In this test, fragments are conditionally rejected on the outcome of testing a linear stipple mask. If the bit is zero then the test fails, otherwise it passes. The line stipple pattern is 16 bits in length and is scaled by a repeat factor r (in the range 1 to 512). The stipple mask bit b which controls the acceptance or rejection of a fragment is determined using:

b=(floor (s/r)) mod 16 where s is the stipple counter which is incremented for every fragment (normally along the line). This counter may be reset at the start of a polyline, but between segments it continues as if there were no break.

The stipple pattern can be optionally mirrored, that is the bit pattern is traversed from most significant to least significant bits, rather than the default, from least significant to most significant.

Color DDA Unit

The color DDA unit is used to associate a color with a fragment produced by the rasterizer. This unit should be enabled for rendering operations and disabled for pixel rectangle operations (i.e. copies, uploads and downloads). Two color modes are supported by GLINT, true color RGBA and color index (CI).

When in Gouraud shading mode, the color DDA unit performs linear interpolation given a set of start and increment values. Clamping is used to ensure that the interpolated value does not underflow or overflow the permitted color range.

For a Gouraud shaded trapezoid, GLINT interpolates from the dominant edge of a trapezoid to the subordinate edges. This means that two increment values are required per color component, one to move along the dominant edge and one to move across the span to the subordinate edge.

Note that if one is rendering to multiple buffers and has initialized the start and increment values in the color DDA unit, then any subsequent Render command will cause the start values to be reloaded.

If subpixel correction has been enabled for a primitive, then any correction required will be applied to the color components.

In flat shading mode, a constant color is associated with each fragment. This color is loaded into the ConstantColor register.

Texture Unit

The texture unit combines the incoming fragment's color (generated in the color DDA unit) with a value derived from interpolating texture map values (texels).

Texture application consists of two stages; derivation of the texture color from the texels (a filtering process) and then application of the texture color to the fragment's color, which is dependent on the application mode (Decal, Blend or Modulate).

Both the GLINT 300TX and GLINT 400TX support all the filtering and application modes described in this section. However, when using the GLINT 300TX , texel values, interpolants and texture filter selections are supplied by the host. This implies that texture coordinate interpolation and texel extraction are performed by the host using texture maps resident on the host. The recommended technique for performing texture mapping using the GLINT 300TX is to scan convert primitives on the host and render fragments as GLINT point primitives.

The GLINT 400TX automatically generates all data required for texture application as textures are stored in the localbuffer and texture parameter interpolation with full perspective correction takes place within the processor. Thus the GLINT 400TX is the processor of choice when full texture mapping acceleration is desired, the GLINT 300TX is more suitable in applications where the performance of texture mapping is not critical.

Texture color generation supports all the filter modes of OpenGL, that is:

Minification is the name given to the filtering process used whereby multiple texels map to a fragment, while magnification is the name given to the filtering process whereby only a portion of a single texel maps to a single fragment.

Nearest is the simplest form of texture mapping where the nearest texel to the sample location is selected with no filtering applied.

Linear is a more sophisticated algorithm which is dependent on the type of primitive. For lines (which are ID), it involves linear interpolation between the two nearest texels, for polygons and points which are considered to have finite area, linear is in fact bi-linear interpolation which interpolates between the nearest 4 texels.

Mip Mapping is a technique to allow the efficient filtering of texture maps when the projected area of the fragment covers more than one texel (ie. minification). A hierarchy of texture maps is held with each one being half the size (or one quarter the area) of the preceding one. A pair of maps are selected, based on the projected area of the texture. In terms of filtering this means that three filter operations are performed: one on the first map, one on the second map and one between the maps. The first filter name (Nearest or Linear) in the MipMap name specifies the filtering to do on the two maps, and the second filter name specifies the filtering to do between maps. So for instance, linear mapping between two maps, with linear interpolation between the results is supported (LinearMipMapLinear), but linear interpolation on one map, nearest on the other map, and linear interpolation between the two is not supported.

The filtering process takes a number of texels and interpolants, and with the current texture filter mode produces a texture color.

The fog unit is used to blend the incoming fragment's color (generated by the color DDA unit, and potentially modified by the texture unit) with a predefined fog color. Fogging can be used to simulate atmospheric fogging, and also to depth cue images.

Fog application has two stages; derivation of the fog index for a fragment, and application of the fogging effect. The fog index is a value which is interpolated over the primitive using a DDA in the same way color and depth are interpolated. The fogging effect is applied to each fragment using one of the equations described below.

Note that although the fog values are linearly interpolated over a primitive the fog values can be calculated on the host using a linear fog function (typically for simple fog effects and depth cuing) or a more complex function to model atmospheric attenuation. This would typically be an exponential function.

Fog Index Calculation--The Fog DDA

The fog DDA is used to interpolate the fog index (f) across a primitive. The mechanics are similar to those of the other DDA units, and horizontal scanning proceeds from dominant to subordinate edge as discussed above.

The DDA has an internal range of approximately +511 to -512, so in some cases primitives may exceed these bounds. This problem typically occurs for very large polygons which span the whole depth of a scene. The correct solution is to tessellate the polygon until polygons lie within the acceptable range, but the visual effect is frequently negligible and can often be ignored.

The fog DDA calculates a fog index value which is clamped to lie in the range 0.0 to 1.0 before it is used in the appropriate fogging equation. (Fogging is applied differently depending on the color mode.)

Antialias Application Unit

Antialias application controls the combining of the coverage value generated by the rasterizer with the color generated in the color DDA units. The application depends on the color mode, either RGBA or Color Index (CI).

When antialiasing is enabled this unit is used to combine the coverage value calculated for each fragment with the fragment's alpha value. In RGBA mode the alpha value is multiplied by the coverage value calculated in the rasterizer (its range is 0% to 100%). The RGB values remain unchanged and these are modified later in the Alpha Blend unit which must be set up appropriately. In CI mode the coverage value is placed in the lower 4 bits of the color field. The Color Look Up Table is assumed to be set up such that each color has 16 intensities associated with it, one per coverage entry.

When using GLINT to render antialiased polygons, depth buffering cannot be used. This is because the order the fragments are combined in is critical in producing the correct final color. Polygons should therefore be depth sorted, and rendered front to back, using the alpha blend modes: SourceAlphaSaturate for the source blend function and One for the destination blend function. In this way the alpha component of a fragment represents the percentage pixel coverage, and the blend function accumulates coverage until the value in the alpha buffer equals one, at which point no further contributions can made to a pixel.

For the antialiasing of general scenes, with no restrictions on rendering order, the accumulation buffer is the preferred choice. This is indirectly supported by GLINT via image uploading and downloading, with the accumulation buffer residing on the host.

When antialiasing, interpolated parameters which are sampled within a fragment (color, fog and texture), will sometimes be unrepresentative of a continuous sampling of a surface, and care should be taken when rendering smooth shaded antialiased primitives. This problem does not occur in aliased rendering, as the sample point is consistently at the center of a pixel.

Alpha Test Unit

The alpha test compares a fragment's alpha value with a reference value. Alpha testing is not available in color index (CI) mode. The alpha test conditionally rejects a fragment based on the comparison between a reference alpha value and one associated with the fragment.

Localbuffer Read/Write Unit

The localbuffer holds the Graphic ID, FrameCount, Stencil and Depth data associated with a fragment. The localbuffer read/write unit controls the operation of GID testing, depth testing and stencil testing.

The LBReadMode register can be configured to make 0, 1 or 2 reads of the localbuffer. The following are the most common modes of access to the localbuffer:

Normal rendering without depth, stencil or GID testing. This requires no localbuffer reads or writes.

Normal rendering without depth or stencil testing and with GID testing. This requires a localbuffer read to get the GID from the localbuffer.

Normal rendering with depth and/or stencil testing required which conditionally requires the localbuffer to be updated. This requires localbuffer reads and writes to be enabled.

Copy operations. Operations which copy all or part of the localbuffer with or without GID testing. This requires reads and writes enabled.

Image upload/download operations. Operations which download depth or stencil information to the local buffer or read depth, stencil fast clear or GID from the localbuffer.

Writes to the localbuffer must be enabled to allow any update of the localbuffer to take place. The LBWriteMode register is a single bit flag which controls updating of the buffer.

Pixel Ownership (GID) Test Unit

Any fragment generated by the rasterizer may undergo a pixel ownership test. This test establishes the current fragment's write permission to the localbuffer and framebuffer.

The ownership of a pixel is established by testing the GID of the current window against the GID of a fragment's destination in the GID buffer. If the test passes, then a write can take place, otherwise the write is discarded. The sense of the test can be set to one of: always pass, always fail, pass if equal, or pass if not equal. Pass if equal is the normal mode. In GLINT the GID planes, if present, are 4 bits deep allowing 16 possible Graphic ID's. The current GID is established by setting the Window register.

If the unit is disabled fragments pass through undisturbed.

Stencil Test Unit

The stencil test conditionally rejects fragments based on the outcome of a comparison between the value in the stencil buffer and a reference value. The stencil buffer is updated according to the current stencil update mode which depends on the result of the stencil test and the depth test.

This test only occurs if all the preceding tests (bitmask, scissor, stipple, alpha, pixel ownership) have passed. The stencil test is controlled by the stencil function and the stencil operation. The stencil function controls the test between the reference stencil value and the value held in the stencil buffer. The stencil operation controls the updating of the stencil buffer, and is dependent on the result of the stencil and depth tests.

If the stencil test is enabled then the stencil buffer will be updated depending on the outcome of both the stencil and the depth tests (if the depth test is not enabled the depth result is set to pass).

In addition a comparison bit mask is supplied in the StencilData register. This is used to establish which bits of the source and reference value are used in the stencil function test. In addition it should normally be set to exclude the top four bits when the stencil width has been set to 4 bits in the StencilMode register.

The source stencil value can be from a number of places as controlled by a field in the StencilMode register:

Depth Test Unit

The depth (Z) test, if enabled, compares a fragment's depth against the corresponding depth in the depth buffer. The result of the depth test can effect the updating of the stencil buffer if stencil testing is enabled. This test is only performed if all the preceding tests (bitmask, scissor, stipple, alpha, pixel ownership, stencil) have passed. The source value can be obtained from a number of places as controlled by a field in the DepthMode register:

When using the depth DDA for normal depth buffered rendering operations the depth values required are similar to those required for the color values in the color DDA unit:

The dZdX value is not required for Z-buffered lines.

The depth unit must be enabled to update the depth buffer. If it is disabled then the depth buffer will only be updated if ForceL-BUpdate is set in the Window register.

Framebuffer Read/Write Unit

Before rendering can take place GLINT must be configured to perform the correct framebuffer read and write operations. Framebuffer read and write modes effect the operation of alpha blending, logic ops, write masks, image upload/download operations and the updating of pixels in the framebuffer.

The FBReadMode register allows GLINT to be configured to make 0, 1 or 2 reads of the framebuffer. The following are the most common modes of access to the framebuffer: Note that avoiding unnecessary additional reads will enhance performance.

Rendering operations with no logical operations, software write-masking or alpha blending. In this case no read of the framebuffer is required and framebuffer writes should be enabled.

Rendering operations which use logical ops, software write masks or alpha blending. In these cases the destination pixel must be read from the framebuffer and framebuffer writes must be enabled.

Image copy operations. Here setup varies depending on whether hardware or software write masks are used. For software write masks, the framebuffer needs two reads, one for the source and one for the destination. When hardware write masks are used (or when the software write mask allows updating of all bits in a pixel) then only one read is required.

Image upload. This requires reading of the destination framebuffer reads to be enabled and framebuffer writes to be disabled.

Image download. In this case no framebuffer read is required (as long as software writemasking and logic ops are disabled) and the write must be enabled.

For both the read and the write operations, an offset is added to the calculated address. The source offset (FBSourceOffset) is used for copy operations. The pixel offset (FBPixelOffset) can be used to allow multi-buffer updates. The offsets should be set to zero for normal rendering.

The data read from the framebuffer may be tagged either FBDefault (data which may be written back into the framebuffer or used in some manner to modify the fragment color) or FBColor (data which will be uploaded to the host). The table below summarizes the framebuffer read/write control for common rendering operations:

Framebuffer writes must be enabled to allow the framebuffer to be updated. A single 1 bit flag controls this operation.

The framebuffer write unit is also used to control the operation of fast block fills, if supported by the framebuffer. Fast fill rendering is enabled via the FastFillEnable bit in the Render command register, the framebuffer fast block size must be configured to the same value as the FastFillIncrement in the Render command register. The FBBlockColor register holds the data written to the framebuffer during a block fill operation and should be formatted to the `raw` framebuffer format. When using the framebuffer in 8 bit packed mode the data should be replicated into each byte. When using the framebuffer in packed 16 bit mode the data should be replicated into the top 16 bits.

When uploading images the UpLoadData bit can be set to allow color formatting (which takes place in the Alpha Blend unit).

It should be noted that the block write capability provided by the chip of the presently preferred embodiment is itself believed to be novel. According to this new approach, a graphics system can do masked block writes of variable length (e.g. 8, 16, or 32 pixels, in the presently preferred embodiment). The rasterizer defines the limits of the block to be written, and hardware masking logic in the frame-buffer interface permits the block to be filled in, with a specified primitive, only up to the limits of the object being rendered. Thus the rasterizer can step by the Block Fill increment. This permits the block-write capabilities of the VRAM chips to be used optimally, to minimize the length which must be written by separate writes per pixel.

Alpha Blend Unit

Alpha blending combines a fragment's color with those of the corresponding pixel in the framebuffer. Blending is supported in RGBA mode only.

The alpha blend unit combines the fragment's color value with that stored in the framebuffer, using the blend equation:

where: C.sub.o is the output color; C.sub.s is the source color (calculated internally); C.sub.d is the destination color read from the framebuffer; S is the source blending weight; and D is the destination blending weight. S and D are not limited to linear combinations; lookup functions can be used to implement other combining relations.

If the blend operations require any destination color components then the framebuffer read mode must be set appropriately.

The alpha blend and color formatting units can be used to format image data into any of the supported GLINT framebuffer formats.

Consider the case where the framebuffer is in RGBA 4:4:4:4 mode, and an area of the screen is to be uploaded and stored in an 8 bit RGB 3:3:2 format. The sequence of operations is:

Set the rasterizer as appropriate

Enable framebuffer reads

Disable framebuffer writes and set the UpLoadData bit in the FBWriteMode register

Enable the alpha blend unit with a blend function which passes the destination value and ignores the source value (source blend Zero, destination blend One) and set the color mode to RGBA 4:4:4:4

Set the color formatting unit to format the color of incoming fragments to an 8 bit RGB 3:3:2 framebuffer format.

The upload now proceeds as normal. This technique can be used to upload data in any supported format.

The same technique can be used to download data which is in any supported framebuffer format, in this case the rasterizer is set to sync with FBColor, rather than Color. In this case framebuffer writes are enabled, and the UpLoadData bit cleared.

Color Formatting Unit

The color formatting unit converts from GLINT's internal color representation to a format suitable to be written into the framebuffer. This process may optionally include dithering of the color values for framebuffers with less than 8 bits width per color component. If the unit is disabled then the color is not modified in any way.

As noted above, the framebuffer may be configured to be RGBA or Color Index (CI).

GLINT uses an ordered dither algorithm to implement color dithering. Several types of dithering can be selected.

If the color formatting unit is disabled, the color components RGBA are not modified and will be truncated when placed in the framebuffer. In CI mode the value is rounded to the nearest integer. In both cases the result is clamped to a maximum value to prevent overflow.

In some situations only screen coordinates are available, but window relative dithering is required. This can be implemented by adding an optional offset to the coordinates before indexing the dither tables. The offset is a two bit number which is supplied for each coordinate, X and Y. The XOffset, YOffset fields in the DitherMode register control this operation, if window relative coordinates are used they should be set to zero.

Logical Op Unit

The logical op unit performs two functions; logic operations between the fragment color (source color) and a value from the framebuffer (destination color); and, optionally, control of a special GLINT mode which allows high performance flat shaded rendering.

A special GLINT rendering mode is available which allows high speed rendering of unshaded images. To use the mode the following constraints must be satisfied:

Flat shaded aliased primitive

No dithering required

No logical ops

No stencil, depth or GID testing required

No alpha blending

The following are available:

Bit masking in the rasterizer

Area and line stippling

User and Screen Scissor test

If all the conditions are met then high speed rendering can be achieved by setting the FBWriteData register to hold the framebuffer data (formatted appropriately for the framebuffer in use) and setting the UseConstantFBWriteData bit in the LogicalOpMode register. All unused units should be disabled.

This mode is most useful for 2D applications or for clearing the framebuffer when the memory does not support block writes. Note that FBWriteData register should be considered volatile when context switching.

The logical operations supported by GLINT are:

Where:

S=Source (fragment) Color, D=Destination (framebuffer) Color.

For correct operation of this unit in a mode which takes the destination color, GLINT must be configured to allow reads from the framebuffer using the FBReadMode register.

GLINT makes no distinction between RGBA and CI modes when performing logical operations. However, logical operations are generally only used in CI mode.

Framebuffer Write Masks

Two types of framebuffer write masking are supported by GLINT, software and hardware. Software write masking requires a read from the framebuffer to combine the fragment color with the framebuffer color, before checking the bits in the mask to see which planes are writeable. Hardware write masking is implemented using VRAM write masks and no framebuffer read is required.

Software write masking is controlled by the FBSoftwareWriteMask register. The data field has one bit per framebuffer bit which when set, allows the corresponding framebuffer bit to be updated. When reset it disables writing to that bit. Software write masking is applied to all fragments and is not controlled by an enable/disable bit. However it may effectively be disabled by setting the mask to all U's. Note that the ReadDestination bit must be enabled in the FBReadMode register when using software write masks, in which some of the bits are zero.

Hardware write masks, if available, are controlled using the FBHardwareWriteMask register. If the framebuffer supports hardware write masks, and they are to be used, then software write masking should be disabled (by setting all the bits in the FBSoftwareWriteMask register). This will result in fewer framebuffer reads when no logical operations or alpha blending is needed.

If the framebuffer is used in 8 bit packed mode, then an 8 bit hardware write mask must be replicated to all 4 bytes of the FBHardwareWriteMask register. If the framebuffer is in 16 bit packed mode then the 16 bit hardware write mask must be replicated to both halves of the FBHardwareWriteMask register.

Host Out Unit controls which registers are available at the output FIFO, gathering statistics about the rendering operations (picking and extent testing) and the synchronization of GLINT via the Sync register. These three functions are as follows:

Message filtering. This unit is the last unit in the core so any message not consumed by a preceding unit will end up here. These messages will fall in to three classifications: Rasterizer messages which are never consumed by the earlier units, messages associated with image uploads, and finally programmer mistakes where an invalid message was written to the input FIFO. Synchronization messages are a special category and are dealt with later. Any messages not filtered out are passed on the output FIFO.

Statistic Collection. Here the active step messages are used to record the extent of the rectangular region where rasterization has been occurring, or if rasterization has occurred inside a specific rectangular region. These facilities are useful for picking and debug activities.

Synchronization. It is often useful for the controlling software to find out when some rendering activity has finished, to allow the timely swapping or sharing of buffers, reading back of state, etc. To achieve this the software would send a Sync message and when this reached this unit any preceding messages or their actions are guaranteed to have finished. On receiving the Sync message it is entered into the FIFO and optionally generates an interrupt.

Sample Board-Level Embodiment

A sample board incorporating the GLINT chip may include simply:

the GLINT chip itself, which incorporates a PCI interface;

Video RAM (VRAM), to which the chip has read-write access through its frame buffer (FB) port;

DRAM, which provides a local buffer then made for such purposes as Z buffering; and

a RAMDAC, which provides analog color values in accordance with the color values read out from the VRAM.

Thus one of the advantages of the chip of the presently preferred embodiment is that a minimal board implementation is a trivial task.

FIG. 3A shows a sample graphics board which incorporates the chip of FIG. 2B.

FIG. 3B shows another sample graphics board implementation, which differs from the board of FIG. 3A in that more memory and an additional component is used to achieve higher performance.

FIG. 3C shows another graphics board, in which the chip of FIG. 2B shares access to a common frame store with GUI accelerator chip.

FIG. 3D shows another graphics board, in which the chip of FIG. 2B shares access to a common frame store with a video coprocessor (which may be used for video capture and playback functions (e.g. frame grabbing).

In the presently preferred embodiment, the frame buffer interface of the GLINT chip contains additional simple interface logic, so that two chips can both access the same frame buffer memory. This permits the GLINT chip to be combined with an additional chip for management to the graphics produced by the graphical user interface. This provides a migration path for users and applications who need to take advantage of the existing software investment and device drivers for various other graphics chips.

FIG. 3C shows another graphics board, in which the chip of FIG. 2A shares access to a common frame store with a GUI accelerator chip (such as an S3 chip). This provides a path for software migration., and also provides a way to separate 3D rendering tasks from 2D rendering.

In this embodiment, a shared framebuffer is used to enable multiple devices to read or write data to the same physical framebuffer memory. Example applications using the GLINT 300TX:

Using a video device as a coprocessor to GLINT, to grab live video into the framebuffer, for displaying video in a window or acquiring a video sequence;

Using GLINT as a 3D coprocessor to a 2D GUI accelerator, preserving an existing investment in 2D driver software.

In a coprocessor system, the framebuffer is a shared resource, and so access to the resource needs to be arbitrated. There are also other aspects of sharing a framebuffer that need to be considered:

Memory refreshing;

Transfer of data from the memory cells into the shift registers of the VRAM;

Control of writemasks and color registers.

GLINT uses the S3 Shared Frame Buffer Interface (SFBI) to share a framebuffer. This interface is able to handle all of the above aspects for two devices sharing a frame buffer, with the GLINT acting as an arbitration master or slave.

The Control Signals used in the Shared Framebuffer interface, in the presently preferred embodiment, are as follows:

FBReqN is internally re-synchronized to System Clock.

FBSelOEN remains negated.

FBGntN is asserted an unspecified amount of time after FBReqN is asserted.--Framebuffer Address, Data and Control lines are tri-stated by GLINT (the control lines should be held high by external pull-up resistors). The secondary controller is now free to drive the Framebuffer lines and access the memory.

FBGntN remains asserted until GLINT requires a framebuffer access, or a refresh or transfer cycle.

FBReqN must remain asserted while FBGntN is asserted.

When FBGntN is removed, the secondary controller must relinquish the address, data and control bus in a graceful manner--i.e. RAS, CAS, WE and OE must all be driven high before being tri-stated.

The secondary controller must relinquish the bus and negate FBReqN within 500 ns of FBGntN being negated.

Once FBReqN has been negated, it must remain inactive for at least 2 system clocks (40 ns at 50 MHz).

Framebuffer Refresh and VRAM transfer cycles by GLINT are turned off when GLINT is a secondary framebuffer controller.

GLINT asserts FBReqN whenever is requires a framebuffer access.

FBGntN is internally re-synchronized to system clock.

When FBGntN is asserted, GLINT drives FBselOEN to enable any external buffers used to drive the control signals, and then drives the framebuffer address, data and control lines to perform the memory access. FBReqN remains asserted while FBGntN is asserted.

When FBGntN is negated, GLINT finishes any outstanding memory cycles, drives the control lines inactive, negates FBselOEN and then tri-states the address, data and control lines, then releases FBReqN. GLINT guarantees to release FBReqN within 500 ns of FBGntN being negated.

GLINT will not reassert FBReqN within 4 system clock cycles (80 ns @ 50 MHz).

The following are some points to be noted when implementing a shared framebuffer design with a GLINT 300TX:

Some 2D GUI Accelerators such as the S3 Vision964, and GLINT use configuration resistors on the framebuffer databus at reset. In this case care should be taken with the configuration setup where it effects read only registers inside either device. If conflicts exist that can not be resolved by the board initialization software, then the conflicts should be resolved by isolating the two devices from each other at reset so they can read the correct configuration information. This isolation need only be done for the framebuffer databus lines that cause problems;

GLINT should be configured as the secondary controller when used with an S3 GUI accelerator, as the S3 devices can only be primary controllers;

GLINT cannot be used on the daughter card interface as described in the S3 documentation, because this gives no access to the PCI bus. A suitable PCI bridge should be used in a design with a PCI 2D GUI accelerator and GLINT so they can both have access to the PCI bus;

The use of ribbon cable to carry the framebuffer signals between two PCI boards is not recommended, because of noise problems and the extra buffering required would impact performance;

The GLINT 300TX does not provide a way of sharing its localbuffer.

In addition to the above capabilities of the 300TX , the 400TX also allows grabbing of live video into the localbuffer and real-time texture mapping of that video into the framebuffer for video manipulation effects.

This technical note describes some system design issues on how multiple GLINT devices can be used in parallel to achieve higher performance. The main driving force for higher performance is the simulation market which, at the low end, demands somewhere between 25-30M texture mapped pixels per second.

There are some key points before we look at different parallel organizations:

To gain any benefit from running multiple GLINTs in parallel, the overall system must be rendering bound. If the system is host bound or geometry bound, then adding in more GLINTs will not improve the systems performance.

The memory systems (i.e. local buffer and framebuffer) are duplicated for each GLINT. Recall that the texture maps are stored in the local buffer. A single GLINT places very high demands on the memory systems, and it would be very difficult to share them between multiple GLINTs. In the presently preferred embodiment there are no provisions for sharing the local buffer, so if this is necessary it would have to be done behind GLINT's back and transparently. The framebuffer can be shared (since GLINT has a SFB interface), but this is likely to be a bottle neck if shared between GLINTs.

Broadcast. In some parallel systems each GLINT will get the same (or mostly the same) primitive data and just render those pixels assigned to it. It is very desirable that this data is written by the host only once, or fetched from the host address space once if DMA is being used. This presents two issues: Firstly the PCI bus does not have any concept of broadcasting to multiple devices, and secondly GLINT does not have a dedicated FIFO status signal pin an external controller can use. Neither of these issues are insurmountable, but will require hardware to solve.

However, if the application only uses a `few` large texture mapped primitives so repeatedly sending or fetching the parameters for each GLINT will not be a problem.

To avoid problems with Antialiasing, Bitmasks for characters, or Line stipple, the area stipple table can be used to reserve scanlines to a processor.

This section looks at some of the common ways of applying parallelism to the rendering operation. The list is not exhaustive and an interested reader is directed to the book by Whitman cited above. No one paradigm is best and the choice is very application or market dependent.

Frame Interleaving is where a GLINT works on frame n, the next GLINT works on frame n+1, etc. Each GLINT does everything or its own frame and the video is sourced from each GLINT's framebuffer in turn. This paradigm is perhaps the simplest one with very little hardware overhead and none of the above complications regarding antialiasing, block copies, bitmasks and line stipples.

This scheme only works when the image is double buffered (normal for simulation systems) and where the increase in transport delay is acceptable. Transport delay is the time it takes for a user to see a visual change after new input stimulus to the system has occurred. With 4 GLINTs this will be 4 frame times attributable to the rendering system, plus whatever else the whole system adds.

The cost of this method is also one of the highest, as ALL the memory has to be duplicated. By contrast, the schemes where the screen is divided up can save depth and color buffer memory (but not texture memory).

Sequential frames will usually have very similar amounts of rendering, unless there is a discontinuity in the viewing position and/or orientation, so load balancing is generally good.

Frame merging is a similar technique to frame interleaving where each GLINT has a full local buffer and framebuffer. In this case the primitives are distributed amongst the GLINTs and the resultant partial images composited using the depth information to control which fragment from the multiple buffers is displayed in each pixel position.

GLINT has not been designed to share the local buffer (where the depth information is held) so the compositing is not readily supported. Also the composition frequently needs to be done at video rate so requires some fast hardware.

Alpha blending and Antialiasing presents some problems but the bitmask, block copies and line stipple are easily accommodated. Good load balancing depends on even distribution of primitives. Not all primitives will take the same amount of time to process so a round robin distribution scheme, or a heuristic one with takes into account the expected processing time for each primitive will be needed.

Here the screen is divided up into large contiguous regions and a GLINT looks after each region. Primitives which overlap between regions are sent to both regions and scissor clipping used. Primitives contained wholly in one region are ideally just sent to the one GLINT.

The number of regions and the horizontal and/or vertical division of the screen can be chosen as appropriate, but horizontal bands are usually easier for the video hardware to cope with. Each GLINT only needs enough local buffer and frame buffer to cover the pixels in its own region, but texture maps are duplicated in full. Block copies are a problem when the block, or part block is moved between regions. Bit masking and line stipples can be solved with some careful clipping.

Load balancing is very poor in this paradigm, since most of the scene complexity can be concentrated into one region. Dynamically changing the size of the regions based on expected scene complexity (maybe measured from the previous frame) can alleviate the poor load balancing to some extent.

The interleave factor is every other n.sup.th scanline where n is the number of GLINTs. Vertical interleaves are possible, but not supported by the GLINT rasterizer. Nearly all primitives will overlap multiple scanlines so are ideally broadcast to all GLINTs. Each GLINT will have different start values for the rasterization and interpolation parameters.

Each GLINT only needs enough local buffer and frame buffer to cover the pixels in its own region, but texture maps are duplicated in full.

Some block copies are a problem when the block is moved between non nth scanlines, but horizontal moves are available with any alignment. Bit masking can be solved with some careful clipping, but line stipples have no easy solution. Antialiasing is not normally a problem but with GLINT 300TX there is no provision for sub scanline steps as well as nth scanline steps. Load balancing is excellent in this paradigm which is the main reason it features prominently in the literature.

Thus the simplest and lowest risk method of using multiple GLINTs is Frame Interleaving, but if this is not an option, e.g. because of the transport delay or the amount of memory needed, then the next best choice is the Interleaved Scanline.

Span Operations in the GLINT Processor

The particular innovative teachinss of the present application are described with relation to several of the particular GLINT processing units.

Span Operations in the Rasterizer Unit

Many 2D rendering operations can be implemented more efficiently using span operations. A span is always 32 pixels long and any combination of pixels within the span can be read and/or written. Pixels with a width of 8 or 16 bits are processed 4 or 2 pixels at a time respectively and all read and write alignment is handled in hardware. The span mechanism can be used for image upload, image download, filling with constant color, filling with a pattern, characters (i.e. bit masks), copies and copies with logical ops. Any trapezoid can be used and the scanning direction can be left-to-right or right-to-left.

If the span is being written with a constant color value and the VRAM supports block fills (where a number of pixels can be written simultaneously) then span filling will use this mode of operation and give a very much faster filling rate.

The number of pixels in a span is always 32 from the core's view point. The Framebuffer Interface Unit takes care of mapping this logical configuration on to the actual VRAM configuration where the VRAM chips may have less pixels in a block, the framebuffer may be interleaved and/or hold packed pixels.

This span operation does have some restrictions:

No accesses to the local buffer are made so GID, Stencil and Depth tests are not available. These units just ignore the messages associated with fast span fills.

Gouraud shading, alpha tests, alpha blend and dither operations are not available.

To support fast fills the rasterizer operates in the following way:

1. For left to right scanning direction: Calculate the visibility of the 32 pixels starting at the left hand edge of the trapezoid, and successive groups of pixels until the right hand edge is found.

For right to left scanning direction: Calculate the visibility of the 32 pixels ending at the right hand edge of the trapezoid, and preceding groups of pixels until the left hand edge is found.

No alignment of the 32 pixel group is assumed with the organization of the framebuffer. No bit mask it taken into account in this stage. The pixel on the left edge is included in the mask while the pixel on the right hand edge is not.

2. If SyncOnBitMask is enabled then the pixel mask is anded with the BitMaskPattern message to further constrain the pixels to be written to.

3. Information about each group of 32 pixels is sent in two messages. The SpanStepX or SpanStepYDomEdge message holds the X, Y coordinate address of the first pixel in the group. The SpanMask message holds the write status of the 32 pixels, with bit 0 assigned to the first pixel, bit 1 to the next (i.e. X+1) pixel, etc.

The two SpanStep messages are used to keep the Texture Address DDAs in step when the texture memory holds the bitmask values.

4. If SyncOnHostData is enabled then the number of words of data is calculated from the span mask (prior to any bitmask being anded in) calculated by the rasterizer and the pixel size. This number of Color or FBData messages are forwarded before the next 32 pixel span is dealt with.

Any trapezoid can be fast filled and not just rectangles. Span fills are enabled by setting the FastFillEnable bit in the Render message. The SpanOperation bit determines the if the span writes use constant color data or variable color data. All other trapezoid parameters are the same.

Span Operations in the Framebuffer Read Unit

The span mask is passed to the FBIU in the Ra FIFO for read masks and in the Dout FIFO for write masks. The Dout FIFO is connected to the FB Write Unit and is used in preference the to the Wa FIFO to allow for the possibility that a later unit from the FB Read Unit may change which pixels get written.

The general operation described here covers reads and writes. The actual span write operations are controlled in the FB Write Unit but are documented here to present more of the overall picture.

The amount of data associated with a span mask (if the modes require any data at all) depends on the pixel size. This is tabulated below:

Read Span Packed

When reading a span of pixels the FBIU only needs to return data for the pixels with their corresponding span mask bit set. The data is returned in the order of lowest address first. During a span read the Pattern RAM is used to temporary extend the apparent depth of the Din FIFO. This is essential when the source and destination spans overlap to prevent any writes occurring before the full span has been read.

The ReadSource and ReadDestination mode bits in the FBReadMode message determine the interpretation of the SpanMask message when it arrives, and what, if any, reads need to be done. The processing in the output stage is controlled by the ReadSource and ReadDestination mode bits and PatternEnable bit in the PatternRamMode message.

Write Span variable data

When writing a span of pixels the FB Write Unit will only supply pixel data for the pixels with their corresponding span mask bit set. The data is supplied in the order of lowest address first. The format of the data sent to the FBIU is identical to that received from the FBIU during a packed span read as documented above.

Write Span constant data

In this mode every pixel has the same data and this is taken from the FBColourRegister which will have previously been loaded up with the required color in the raw framebuffer format. If the pixel width is 8 or 16 bits then the color information will have been replicated into each pixel position in the FBColourRegister by software before the FBColourRegister message is sent. Ideally this mode is implemented using the block write capabilities of VRAM.

Span operations with ROP2 blits

The combination of Read Span Packed and Write Span with variable data are ideally suited to copying data around in the framebuffer. The source and destination addresses can have any alignment and the pixel size can be 8, 16 or 32 with the smaller pixel sizes yielding higher pixels per second performance figures.

When both source and destination reads are enabled the source span is read first and the data stored in the Pattern RAM. The source data is not forwarded on at this point in time. Next the destination span is read (using the same SpanMask) and as each word of data is forwarded on in an FBData message one word of data is taken from the Pattern RAM and forwarded on in an FBSourceData message. The FBSourceData message triggers the LogicalOp Unit to combine the preceding FBData and the FBSourceData messages as necessary and generate the FBWriteData message for the destination data to be written.

Span Operations with Pattern Fills.

The span write with constant color can be used to do pattern fills with the performance advantages of the span mechanism. This may at first sight look like a contradiction in that constant color does not equal pattern, however if we look at the very common case of an 8.times.8 pattern used in Windows then all will become clear. The constant color is in fact a 64 bit register and for 8, 16 and 32 bit pixels represents 8, 4 and 2 pixels worth of data respectively hence adjacent pixels don't have to be the same color.

The mode bits are set up as follows:

8 bit pixels Load the pattern into the FBBlockColour register for the first row of the pattern and span fill the region with a dy value of 8 (i.e. every 8th line is filled. Repeat this for each of the remaining 7 rows and offset the startY value by 1 for each row. This will give a very high pattern fill rate. The disadvantages are that it may have more page misses if the screen resolution permits two or more scanlines to fit a VRAM page, involve the host in more work over just using the texture mapping facilities and may give an odd visual effect.

16 bit pixels As for the 8 bit pixel method but now the process is repeated 16 times--twice per row to cover the first 4 pixels in the pattern and then the second 4 pixels in the pattern. The area stipple is set up to select the appropriate 4 pixels for the two passes per row of the pattern. The same set of observations apply but now there will be twice as many page breaks as each line is visited twice.

32 bit pixels As for the 8 bit pixel method but now the process is repeated 32 times--four per row to cover each of the 4 pair of pixels in the pattern. The area stipple is set up to select the appropriate 2 pixels for the four passes per row of the pattern. The same set of observations apply but now there will be four times as many page breaks.

This technique is not a general solution to doing pattern fills--using the Texture Units is much more flexible--but it does fill a niche for the 8.times.8 patterns used by Windows at the smaller pixel sizes, and maybe even the 32 bit case. It is limited to rectangular regions as the pattern is aligned to the fist pixel in a span and not a fixed X coordinate.

Span Operations with ROP2 Pattern Fills

The ROP2 operation prevents the previous method of using span fill with constant color to be used as the pattern is combined with the destination using a logical operation. In this case the pattern ram is used to source the pattern data. Again consider the more restricted case of a Windows 8.times.8 pattern the pattern ram can hold a the entire 8 bit and 16 bit patterns. There is only enough room to hold half the pattern for 32 bit pixels so two passes over the region picking out even and odd lines would be needed with the pattern ram being loaded for each pass.

The start position in the pattern ram where a spans worth of pattern data is read from is determined from the Y coordinate associated with the span mask. The start address is given by:

where Yshift and Ymask are in the PatternRamMode message. Only the least significant 5 bits of the Y address are of interest. The X offset is similarly given:

where the offset is the bit, pair or nibble offset for 32, 16 and 8 bit pixels respectively. The pattern ram address is then:

For an 8.times.8 bit pattern the values of the X and Y shift and masks are as follows:

This Y shift and mask values can be set up for different pattern sizes other than the 8.times.8 outlined here. The pattern must have a width of 2, 4, 8, 16 or 32, but can have any height. Also in a multi GLINT system there can be one pattern common to all GLINTs or the pattern can be divided up so each GLINT only gets the parts of the pattern it needs for its scanlines.

If a solid color is needed to be sourced from the pattern ram then this 32 bit value can be stored at offset 0 and the X and Y masks both set to 0.times.00.

This technique still has an advantage over using the texture units in that 2 or 4 pixels are processed at a time where as using the texture method on one pixel at a time is dealt with. Also the span method makes better use of the message stream bandwidth. This method can also be used as an alternative method to using span fill with constant color.

Selecting the span operation.

The span operation is controlled by two bits in the FBReadMode message and one in the PatternRamMode message as follows. Note that the assumption is that a span write is always required (it can be disabled in the FB Write Unit) and the type of write is controlled by a bit in the Render message. Setting this bit and these three bits into an inconsistent state may cause loss of performance. For example a span write with constant color does not expect any color data to follow the SpanMask message so any that does follow just ends up in registers and remains unused, but consequently has used message and/or framebuffer bandwidth.

Span Mask Operation in the Framebuffer Write Unit

The span mask operation, as far as this unit is concerned, operates in two modes:

Constant Color

Each pixel with its corresponding bit in the SpanMask is written with constant pixel data (previously loaded by the FBBlockColour message). This write ideally uses the block write capability of VRAM so very high fill rates can be achieved. The SpanMask is forwarded to the Framebuffer Interface Unit via the Write Data FIFO. The Framebuffer Interface Unit uses the SpanMask, together with the size of a pixel (previously sent in the PixelSize message) and local knowledge of the memory configuration to effect the optimum write strategy.

Variable Color

Each pixel with its corresponding bit in the SpanMask is written with variable pixel data. The SpanMask is forwarded to the Framebuffer Interface Unit via the Write Data FIFO and then followed by enough pixel data values to provide the variable pixel data for the given SpanMask and pixel size. The amount of data sent is a function of the pixel size:

32 bits One word of data is sent for each bit set (starting at bit 0). The number of words is found using a population count on the 32 bit SpanMask.

16 bits One word of data is sent for each pair of bits which are non zero. The number of words is found by first reducing the SpanMask to 16 bits by ORing pairs of bits (0, 1), (2, 3), etc. together and doing a population count on the reduced mask.

8 bits One word of data is sent for each nibble of bits which are non zero. The number of words is found by first reducing the SpanMask to 8 bits by oring nibbles of bits (0, 1, 2, 3), (4, 5, 6, 7), etc. together and doing a population count on the reduced mask.

When forwarding the pixel data this unit waits for the appropriate number of FBWriteData messages while writing their data into the Write Data FIFO. If a non FBWriteData message is found while forwarding data something has gone wrong (should never happen) and the span operation is aborted and the Framebuffer Interface Unit informed. The goal here is to prevent the chip from hanging rather than to prevent corruption of the framebuffer.

The data written to the framebuffer can be optionally uploaded (as in the active step case).

The FB Read Unit will have prepared the Framebuffer Interface Unit with the span address.

Particular Disclosed Embodiments

Among the disclosed classes of preferred embodiments, there is provided: A rendering system, comprising: an external data input connection; a rasterizer connected to receive and process 2D and 3D primitives from said external data input connections, and accordingly to decompose said primitives into fragments; wherein in a externally selected first mode of operation of said rasterizer, but not in a second mode, said rasterizer decomposes at least some of the 2D primitives into multi-pixel fragments which each include multiple pixels; and rendering hardware connected to receive said fragments from said rasterizer, and to render said fragments; whereby the rendering system provides accelerated performance of 2D rendering tasks.

Among the disclosed classes of preferred embodiments, there is also provided: A rendering system, comprising: an external data input connection; a rasterizer connected to receive and process 2D and 3D primitives from said external data input connections, and accordingly to decompose said primitives into fragments; wherein in a externally selected first mode of operation of said rasterizer, but not in a second mode, said rasterizer decomposes at least some of the 2D primitives into multi-pixel fragments which each include multiple pixels; and rendering hardware connected to receive said fragments from said rasterizer, and to render said fragments; wherein said rendering hardware uses a fast block fill on said multi-pixel fragments which are of a constant color; whereby the rendering system provides accelerated performance of 2D rendering tasks. Among the disclosed classes of preferred embodiments, there is also provided: A graphics processing subsystem, comprising: an input buffer; a rasterizer, connected to read messages from said input buffer and, in a first mode of operation, to generate a sequence of pixel locations accordingly, and in a second mode of operation, to generate multi-pixel sets and corresponding pixel location pointers accordingly; one or more differential data analyzers, operatively connected to receive messages from said rasterizer, and connected to provide a sequence of incrementally changing parameter values for sequences of pixel locations or pixel sets defined by messages received from said rasterizer; a plurality of additional processing units, operatively connected in series in a pipeline relationship, and jointly operatively connected to receive messages from said differential data analyzers; plural ones of said additional processing units each having customized hardware which is different from one another and from said color differential data analyzers; said processing units being mutually interconnected in a pipeline relationship, such that individual ones of said processing units receive messages and, in accordance with the content of each said respective message, programmably perform a respective graphics processing operation, and selectably transmit said respective message to a succeeding one of said processing units; whereby said plural processing units jointly provide a pipelined multiple-instruction-multiple-device (MIMD) graphics processing architecture.

Among the disclosed classes of preferred embodiments, there is also provided: A rendering method, comprising the steps of: receiving primitive data; if a first mode is selected, then decomposing said primitive data into multi-pixel fragments; performing a set of pre-rendering graphics processes on said fragments; rendering said fragments and storing the result in a framebuffer; if a second mode is selected, then decomposing said primitive data into single-pixel fragments; performing a set of pre-rendering graphics processes on said fragments; rendering said fragments and storing the result in a framebuffer; and displaying the contents of said framebuffer; wherein said multi-pixel fragments allow multiple pixels to be processed in parallel.

Modifications and Variations

As will be recognized by those skilled in the art, the innovative concepts described in the present application can be modified and varied over a tremendous range of applications, and accordingly the scope of patented subject matter is not limited by any of the specific exemplary teachings given.

The foregoing text has indicated a large number of alternative implementations, particularly at the higher levels, but these are merely a few examples of the huge range of possible variations.

For example, the preferred chip context can be combined with other functions, or distributed among other chips, as will be apparent to those of ordinary skill in the art.

For another example, the described graphics systems and subsystems can be used, in various adaptations, not only in high-end PC's, but also in workstations, arcade games, and high-end simulators.

For another example, the described graphics systems and subsystems are not necessarily limited to color displays, but can be used with monochrome systems.

For another example, the described graphics systems and subsystems are not necessarily limited to displays, but also can be used in printer drivers.