[cairo] Avoiding seams with CAIRO_OPERATOR_SATURATE (the OpenGL way)

[Robert O'Callahan]

On Sat, 2005-06-11 at 00:45 +0200, David Reveman wrote:
> So, OpenGL's GL_SRC_ALPHA_SATURATE blending factor can only be used with
> non-premultiplied images, that's why glitz doesn't accelerate
> CAIRO_OPERATOR_SATURATE right now. There's no way to accelerate RENDER's
> saturate operator with OpenGL, except for when using solid colors of
> course. I'll add saturate operator support for solid colors in glitz
> sometime soon but that's pretty much all we can do for acceleration of
> saturate operator with OpenGL.

That's not good.

There are a couple of other problems with this approach (although I do
find it intriguing):

Changing all our code to paint front to back would be really painful,
especially because we'd still have to support back to front painting
(for non-cairo targets in the interim, and maybe for cairo PDF, unless
we can work SATURATE support in there by some magic).

Suppose you start with a black background and fill it with 50%-alpha
white. Then you fill a white rectangle at (0,0,1,0.5). Using SATURATE
(and doing these in reverse, of course) the pixel (0,0)-(1,1) ends up
solid white. But the "true" result should, in this case, be 75% white.
The underlying problem is that alpha values are unable to distinguish
"partially translucent shape covering the entire pixel" from "shape
partially covering the pixel" so it seems to me we must accept either
this error or seams if our solutions are restricted to compositing
operators. Now it could well be that this is much less of a problem than
seams, in real usage, in which case this would be the right thing to do
if the other issues were fixed.

Mozilla's objects do have a uniform Z value, even with SVG, but we still
don't want to see seams along the edges of adjacent objects. For example
if the star-and-ring example were drawn with Mozilla SVG, each
individual primitive shape as seen by Mozilla would have a uniform Z
value, but to the author the combination is *logically* a shape with
nonuniform Z value so seams would be anomalous. The same thing can arise
even in HTML/CSS, where authors often position elements adjacent to each
other and expect continuity.

Rob