[cairo] new trapezoid code in glitz

[David Reveman]

On Thu, 2005-01-27 at 13:21 +0100, David Reveman wrote:
> I've now committed the new trapezoid code I've been working on for
> glitz.
> 
> Here's a short description of how this new code works:
> 
> An implicit mask surface for a single anti-aliased edge that intersect a
> rectangle in any direction, can be represented accurately and
> consistently with a set of one-dimensional texture coordinates in a
> fixed size texture. By using linear texture filtering this allows us to
> render perfectly anti-aliased geometry on a wide range of hardware.
> 
> Each trapezoid needs to be slit up into a set of rectangles along with
> appropriate horizontal texture coordinates for the specified mask
> surface. In general the mask is a 2x1 surface with pixel 0,0 "clear
> black" and pixel 1,0 "solid white". It could also be a Nx1 surface
> containing a pre-computed gamma curve with resolution N for gamma
> corrected anti-aliasing.
> 
> The algorithm that I've come up with, and the current implementation
> that will do this "trapezoid -> rectangles/texcoords" conversion can be
> found in glitz_trapimp.h. The current implementation is far from
> perfect, it could probably be a lot more efficient, and it's all in
> floating point arithmetics right now. However, the performance of the
> current implementation is still pretty good.
> 
> Rendering non-intersecting trapezoids this way is very fast. For
> intersecting trapezoids we need to use an intermediate mask and that
> really kills performance, but this is a completely different problem
> that should be solved by upcoming extensions to OpenGL.
> 
> However, non-intersecting trapezoids are always going to be faster. I'd
> like to see a trapezoid attribute in cairo that will indicate to the
> backends that the trapezoids are not intersecting. I shouldn't be too
> hard for the tessellator to keep track of this, right? 
> 
> How about making the tessellator snap intersection points to the pixel
> grid and produce non-intersecting trapezoids if the tolerance is greater
> than half a device unit?
> 
> 
> This new trapezoid code is not currently used by cairo's glitz backend.
> Xgl is using it and it works quite well in there. I'd like to switch to
> using this new trapezoid code in the glitz backend as well. However,
> until we have a good render-texture extension in OpenGL, I'm going to
> let libpixman handle trapezoid rasterization when trapezoids are
> intersecting. This will of course be a lot slower than the current glitz
> backend but the output quality will instead be perfect, and I think
> that's a lot more useful. If someone thinks this is a bad idea let me
> know.

"non-intersecting" was the wrong word. By "non-intersecting" I meant; no
pixel is intersected by more than one trapezoid.

So in my previous mail:
"non-intersecting" = no pixel is intersected by more than one trapezoid
"intersecting" = at least one pixel is intersected by more than one
trapezoid

Sorry for any confusion this might have caused.

-David