I am not intending to go into high end GPUs here. That is too advanced. What tasks does a basic graphics hardward perform? By basic perhaps we can go back in time and see what it did in the early days before we got into all this 3D graphics and pixel shaders world. I understand that it should be able to draw lines and write text and "rotate things", but that is layman's language. For engineers that work with these things, how would they describe the functions of a basic graphics hardware in technical terms?

Now usually in digital circuits there is some state machine based controller which implements the hardware functionality. I curious about how a basic graphics hardware works and would be good to see a basic example of it.

And finally, have people had to write a graphics hardware to implement on FPGA for an application which has a rather complex LCD display?

  • \$\begingroup\$ OK, this question might be rather broad but I don't know where to ask \$\endgroup\$ – quantum231 Jan 7 '15 at 1:21
  • \$\begingroup\$ Well, if you have access to an academic library any book on graphics from the 1970s-1980s will cover all of that in detail... at a conceptual level. On the other hand if you want it down to gate-level implementations of the hardware units, I don't think such a book was ever written, even in the old days when it might have been possible to cover an entire graphics adapter in a book. You'd have to search the relevant academic papers and patents for the kind of sub-unit you want to research in-depth. \$\endgroup\$ – Fizz Jan 7 '15 at 1:55
  • \$\begingroup\$ Occasionally, you'll find some more recent text. To give a semi-random example: here's a presentation on Bresenham's algorithm in hardware. \$\endgroup\$ – Fizz Jan 7 '15 at 1:59
  • \$\begingroup\$ Wow, I am happy with these answers. If I can have a VHDL example of a rather simple graphics hardware for 2D display, that will make things very clear. \$\endgroup\$ – quantum231 Jan 8 '15 at 18:13
  • At first there was text display, which each character displayed from a character generator ROM, sometimes complemented with semigraphic characters (like Teletext, ZX-81 characters...)
  • Then came the raw framebuffer where the CPU had to paint each pixel (IBM : CGA, Hercules), sometimes with hardware assistance for overlay/masking colours (IBM : EGA, VGA multiplane), sometimes with an hardware cursor overlay (to avoid wasting time redrawing the cursor and erasing it each time the mouse moves)
  • Then came a few 2D graphic primitives :

    • BITBLT / Blitting. Copying rectangular areas on screen and/or to memory.
    • FILL : Filling rectangular areas with a solid colour.
    • Character display : Expanding 1bit per pixel character shape data to suitable colormaps.
    • Line drawing. The famous Brezenham algorithms.

These hardware acceleration features were usually done with fixed hardware logic. Sometimes, DSPs or special purpose CPUs (for example Texas Instruments TIGA chipsets) were used to offload the main CPU. These efforts were sometimes defeated when the cost of transferring data or programming the graphic accelerator exceeded the time required by the CPU for doing the work itself, especially when special effects (like transparency) were needed. 2D rotation is not used enough to reward hardware acceleration. In games, pre-rendered rotated sprites were stored in memory.

The hardware complexity and cost must be balanced with the time saved by the main CPU. Blitting and filling is very useful, drawing shapes like anti-aliased lines and Bezier curves are often beyond the capacity of 2D hardware.

(For a recent example of basic 2D acceleration in small ARM SOCs, look for "ST Chrom-ART Accelerator")

  • \$\begingroup\$ hmmm, so in actuallity it mainly concentrates on text rather than on scaling, rotating and transparenting things in an image. \$\endgroup\$ – quantum231 Jan 8 '15 at 18:12

A basic 2D graphics card has to have three basic things:

  • Some way to get data in
  • Some way to store or generate an image
  • Some way to get data to the screen

Looking at some traditional graphics systems we can see how these have often been done.

In the days of incredibly expensive memory some very clever tricks have been done to keep costs down. This typically means using very very small amounts of memory at the cost of increased CPU usage. For instance, on some early games consoles the image data was generated realtime by the game software itself, with the CPU knowing at all times just where on the screen the current scan line of the electron beam is, and generating just the graphic data needed for that.

As memory became cheaper, and demands for a better user interface became greater, the processing and generating of the image data moved out of the CPU and into the graphics chipset. So we get the more traditional environment we are more used to now.

A graphics card, primarily, needs somewhere to store the current image. Basically a block of memory. That memory could be memory mapped into the CPU's address space, or even be part of the CPU's main memory in some cases. But memory is needed however it's provided. How much memory very much depends on how the rest of the graphics system works.

The memory can basically be used in two different ways, depending on how much there is available and what you want to display. The two modes are Direct and Interpreted framebuffering.

Direct framebuffering is the most simple to understand. You have a block of memory where each address within it represents the colour of a point on the screen. The data is split into the three channels (for RGB) or treated as a single brightness value (monochrome) and fed into DACs to generate the video waveform. Pretty simple, but needs lots of memory.

In the time while memory was still in the process of getting cheaper interpreted framebuffering was more in use.

In this mode each address in the memory represents something other than a single pixel. It may be that it represents a cluster of pixels, with each value being used to look up the colour to use in a "CLUT" (Colour Look-Up Table). The PC's old CGA graphics mode worked like this, where you had the ability to display 4 colour pixels (including black) with a choice of 2 palettes to choose from. Those palettes were actually hard wired to give you Black/Red/Green/Yellow or Black/Magenta/Cyan/White (basically turning the blue channel on or off for each palette), but each pixel would then only take 2 bits, enabling you to fit 4 times as much video data into memory.

Another popular method for interpreted framebuffering was the ROM character set. In this mode each address in the memory represents a single character (often with attributes like colour or brightness) on the screen. The graphics engine then takes each pixel of the screen in turn, works out which character cell is under that pixel, gets the character from that cell, then looks up the colour of the pixel for that location within the character from ROM. This meant that you could have a much higher resolution text display than graphics display - ideal for word processing and similar applications where all you need is text.

For example, an 80x24 text display would use just bytes, but for the same display using pure graphics would take (say with 8x8 fonts) 15360 bytes.

Once memory got cheap, and computers started having hundreds of KB, or even megabytes of RAM in them, manufacturers started to get much more creative.

Being able to have blocks of RAM allocated purely to graphical work opened up lots of new doors.

Graphics chipsets started being able to maniuplate the graphics themselves offloading some of the processing from the CPU. Moving chunks of data around in RAM very fast ("blitting") allowed high speed display updates for more graphically rich games. Being able to get graphical data from multiple locations and combine them together at image output time ("sprites") allowed for much simpler interaction with the user. For instance, the player character in a game would be a sprite overlaid on the graphics "playfield" by the graphics hardware. The mouse cursor would be a sprite placed over the top of everything else. This separation of "foreground" sprites and "background" framebuffer meant that you didn't need to remember what was under the mouse cursor (or whatever was moving around) to re-draw it after movement.

Then you had the more advanced systems where you had a fully programmable controller chip - basically a graphical co-processor (such as the Amiga's "Copper" chip) which could perform all sorts of complex operations synchronised with the display timing.

But basically most accelerated graphical operations were about moving large amounts of data around, not just incredibly quickly, but also at the right time and in the right way.

  • \$\begingroup\$ so all we are doing is writing what images or parts of image exist at XY coordinates. The graphics chipset merely reads the data and puts it on screen pixel by pixel. hmmm \$\endgroup\$ – quantum231 Jan 8 '15 at 18:10
  • \$\begingroup\$ That is what the output stage of a direct framebuffer does. The GPU can manipulate the data at high speeds in different ways. \$\endgroup\$ – Majenko Jan 8 '15 at 18:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.