The Fenix VDI (document v1.00beta, 971111)

Goals

• As compatible as possible with modern VDIs, but with new extensions.
• No worse performance wise than any competing VDI.
• Layered design to make porting to new hardware easy.
• As flexible and portable as possible.
• Implemented as a shared library, but still compatible with old software.

Design

All accesses to the VDI are made in the current process' context. This means that the VDI must be fully reentrant and that semaphore type locking must be used for shared resources.
Applications as well as internal routines should be able to send a bunch of drawing commands as one block. The main benefit appears when accelerated hardware which buffers commands are used. Internally this would for example be used when blitting text.
Any VDI function, at any level, should be exchangeable for a direct one in a specific device driver.
The VDI is used to emulate LineA calls.
Input/output as well as synchronization between different VDI workstations is handled by a separate server that is common for both the VDI and the AES.

Calling

The standard C bindings for VDI calls make much more sense than the way those actually take place (filling an array with the parameters and a function number to use). Non-Fenix-aware programs will, however, use the old method which must therefore still be supported.
For calls where it can provide a noticeable performance increase, the library should have native support for both methods, but otherwise it will suffice to convert to the new format.
If the new format is used, multiple commands can be sent at once by providing pointers to properly setup 'stack frames', kept on a linked list. The suggested new calling convention is described in the 'implementation thoughts' section below.

Device drivers

The VDI is built up of a number of layers which can, but are not required to, rely on lower level ones. There are also some things that stand completely outside they layered structure, such as set/query, workstation control, printer/meta-file control, input and vector redirection.
For new hardware its enough to write the 'primitive' routines below.
All more complex routines can be built from these.
The 'acceleration' layer should always have a native implementation.
Without this layer things will be very slow, but they'll still work.
The operations here are often accelerated by graphics hardware.

• Layer 0, primitive
  • Get pixel, set pixel.
  • Set graphics mode, set palette colours.
  • Hardware initialization/shutdown/control.
• Layer 1, accelerated
  • Line drawing, blitting (copy/expand/fill).
• Layer 2, standard
  • Arc drawing, bitmap text drawing.
• Layer 3, specialized
  • Filled shape (other than box) drawing.
  • Screen escape functions.
  • Outline font drawing.

Development schedule

Step one

The things here are needed before anything useful at all can be done. For some software, this might actually be enough for normal operation. Unless otherwise noted (due to existing code), this step is monochrome only. That includes plane-skipping mono, though.
Given that quite a bit of code already exists, it should be possible to finish this during November.

• Query - Only enough to make sure programs don't give up. [done+]
• Set - At least colour. [done+]
• Setup - The most necessary workstation stuff. [done]
• Input - Only simple mouse and keyboard read.
• Pixel - (Get and set) Any colour, no mode. [done]
  Simple changes to existing line routines for set, but get is new.
• Line - Any colour, no mode/pattern/width/style. [done]
  Can be implemented via pixel. Routines exist for all bitplane modes.
• Fill - (Rectangle) Any colour, no mode/pattern/outline. [done+]
  Can be implemented via pixel or line.
  Since there are line routines, this can be considered finished.
• Blit - At least mono, no modes. [done+]
  Can be implemented via pixel.
  Some of these things are already done.
• Expand - This is the single to multiplane copy. [done]
  Can be implemented via pixel.
  This is not useful without the modes, but it should be easy enough.
• Text - At least monochrome, one font, no modes, no effects. [done+]
  Can be implemented via pixel or expand.
  Some routines exist already.

Step two

When this step is completed, it should be possible to use graphics cards to some degree.

Implement the things from step one for 16 bit true colour modes.
Quite a bit o code should exist here already as well.

Step three

This is supposed to make the VDI really useful for most purposes.
Most of this is relatively straightforward, but since it's not as 'interesting' as step one, I don't think we can count on it being finished before the end of January.

Implement the rest of the workstation stuff, including off screen bitmaps.
Make everything work in all graphics modes (bitplane, and chunky 8/24/32).
Implement remaining (non-complex) query/set.
Make all drawing modes operational. [done-]
Implement standard<->device format transformation.
Implement the most important text effects.
Implement box outline and hollow fill.
Add support for markers.
Add the screen escape stuff.
Add the vector redirection stuff.

Step four

When this step is completed, all standard VDI screen output should work!
I won't hazard a guess as to when that might happen, though.

Add patterns for fills and lines. [done-]
Implement the remaining text effects.
Implement line width and styles.
Implement the remaining graphics primitives (circles etc).
Implement polygon fill.
Implement the remaining input stuff.
Implement the printer and meta file stuff.

Step five

Perhaps this can be done in parallel. I have no idea about how long it might take, though.

Add outline font support.
Add whatever else is new in NVDI.

New ideas

These are not given in any specific order.

• Set pixel
  A low overhead function that can be used for quick porting to new hardware.
• Plane skipping
  Ignoring some image planes for efficiency reasons.
• Dual source raster operations
  The current VDI often need to do separate mask and draw passes. These could be combined as they are on the Amiga if two sources are used.
• Pattern alignment
  Current VDIs always align fill patterns to the top left of the output area. It would be nice to be more flexible about this.
• Chunky MFDB format
  The standard format for VDI rasters is bitplane based. This is a very bad idea with modern graphics hardware.
• Selectable background colour
  Current VDIs can only draw in one colour. The background is always pen 0. Especially for text it would be useful to be able to draw in for example blue on grey in one step.
• Character width setting
  Fonts are normally designed to be scaled proportionally, but it would be nice if this could be circumvented.
• Text antialiasing
  This is simple enough to do if there are enough colours available and it can look a lot better than standard text output.
• Correct clipping
  All current VDIs are braindead when it comes to clipping lines. This is very obvious when a 'bouncing lines' program is running in a window and you move another over it.
  There would probably not be any incompatibilities if this was fixed.
• Image drawing in any graphics mode
  The current standard->device format transformation is only supposed to skip planes when there are too many.
  It would be nice if images could be dithered to the current palette.
• Colour dithering
  It's not really nice to just drop bits when too few are available. This is a similar problem to the one above, but the idea here is that anything, for example fills, should be possible in dithered colour.
• Virtual screen support
  A new call is needed to reinitialize a workstation.
  Also, the VDI has to make sure that it's drawing in the correct mode.

Implementation thoughts

Just a collection of random thoughts.

• New internal bitmap font format
  Fonts should not be stored internally in the standard format, since this is very inefficient. Instead, make sure each character starts on a byte (or even word) boundary and that all the lines follow each other directly.
• Optimized text output
  Do all simple things directly to the screen memory.
  More complex things should be done to a monochrome buffer and blitted.
• Quick despatch
  Optimize despatcher for quick functions. Probably mostly useful for GEMBench testing, but it wouldn't hurt the slower functions anyway.
• Vector font support
  Code fot Truetype and PostScript font rendering is available, but Speedo is unfortunately proprietary. Perhaps via NVDI?
• Accelerated hardware
  XFree86 has code that can be used.
• VDI function calling
  Both single and blocked calls are supported and are done as follows:
  • Call to a single VDI function:
    • Push parameters (or place them somewhere) in standard C VDI call order
    • Set a1 to point to the parameters
    • Set a0 to point to the workstation structure
    • Do whatever might be necessary to get the shared lib function address
    • Call the function
  • Blocked call:
    • Set next_call to NULL.
    • For all calls
      • 'Push' parameters (see above)
      • 'Push' workstation structure address
      • Get function address (see above) and 'push' that as well
      • 'Push' next_call
      • Set next_call to current 'stack' address
    • Call the 'block execution' function
  • The 'block execution' function would then:
    • Get link and if it's not NULL
      • Get function address
      • Get strucutre address for a0
      • Point a1 to the parameters
      • Call the function
      • Follow the link
    • else
      • return