[cairo] [RFC] Cairo GLES2 and GL_CLAMP_TO_BORDER

[Alexandros Frantzis]

On Fri, Feb 18, 2011 at 06:26:37PM +0200, Jonathan Morton wrote:
> On Fri, 2011-02-18 at 18:13 +0200, Alexandros Frantzis wrote:
> > create W+1 x H+1 textures
> 
> You need four borders, so: s/+1/+2/g.

Oops. Right, thanks!

> 
> Another alternative would be to compute the edge fade-to-transparent in
> the shader, using CLAMP_TO_EDGE on the texture itself.  The only extra
> information required by the shader for this would be the texture pixel
> dimensions, which would be required anyway if the texture is "manually"
> filtered.  I think these could be passed in uniforms.
> 
> As a first pass:
> 
> sampler2D tex;
> uniform vec2 dims;
> varying vec2 coords;
> vec2 fades = clamp(dims * (1.0 - abs(coords - 0.5)), 0.0, 1.0);
> vec4 colour = texture(tex, coords) * fades.x * fades.y;
> ...
> 

The expression I am currently experimenting with is:

clamp(-abs(dims * (coords - 0.5)) + (dims + vec2(1, 1)) * 0.5, 0.0, 1.0)

It works well but we still get some extra failures in the test suite (vs
using GL_CLAMP_TO_BORDER). The differences are small and are mostly
located near the edges. I am still investigating, but my initial guess
is that they are caused by the fact that we are trying to apply this
fade-to-transparency to values that have already been bilinearly
filtered. 

The only way around this is to use GL_NEAREST and perform the full
bilinear filtering in the shader ourselves. However, I doubt that this
is worth the computational cost. The current results, although not pixel
perfect, are a reasonable trade-off.

Thanks,
Alexandros