Re: Let's just build it and see how it works!

Around 7 o'clock on Nov 26, Brian Paul wrote:

> > In particular, I 
> > need per-component alpha compositing to display text on my LCD monitor, 
> > and OpenGL just doesn't do that (yet).
> 
> Keith, can you point me to the code which does this now?  I have a 
> strong hunch that you actually can pull this off with OpenGL as-is.

Yeah, I was reading through some of the GL compositing options wondering 
if they weren't able to do this, either with multiple passes or some other 
suitable kludge in the solid color case.

Looks like the current specification is a bit light on the details 
surrounding 'component-alpha', about the only information I could find was:

"When the mask contains separate alpha values for each channel, the
 alpha value resulting from the combination of that value with the source
 alpha channel is used in the final image composition."

Here's how it works.  The fundemental pixel primitive in Render is:

	dst = (src IN mask) OP dst

		(OP is one of the Porter-Duff compositing operators like OVER)

Normally, 'mask' is just a regular alpha channel, but with component-alpha 
enabled, 'mask' has four separate alpha channels, one for each component.  
This combined operation is not expressable as a sequence of primitive 
operations with intermediate buffers because the (src IN mask) actually 
has eight separate values (four components, plus four alpha values).

The code which implements this fine operation is split across a few 
functions in the 'fbcompose.c' file:

http://freedesktop.org/cgi-bin/viewcvs.cgi/xserver/fb/fbcompose.c?rev=HEAD&cvsroot=xserver&content-type=text/vnd.viewcvs-markup

The fbCombineMaskC function computes (src IN mask) value as two separate 
four-channel values and the fbCombineOverC procedure uses that to place 
that value OVER dst.

I think the total lack of optimization should make this code relatively 
easy to read, even as it sucks the life from my CPU as I type this email.

-keith