Manually computed FragCoord wrong? (small error)

[_x57]

Hi all - i hope you can help me again

I am trying to manually compute gl_FragCoord to do a lookup in a depth texture. I want to do this manually since i later need to alter gl_Position in the vertex shader but still want to perform depth lookup at the old fragment position. (its a gpgpu combined with normal pipeline thing…).

Ok, here is what i doin the VS:

varying vec4 myPos;

void main( void )
{
gl_Position = ftransform(); //changed later
myPos=ftransform();
}

and in the FS:

uniform vec2 viewportSize; // size of viewport in pixels

varying vec4 myPos;

void main( void ) {
vec4 myFragCoord;
myFragCoord.xy = (0.5+(0.5*p.xy/p.w))viewportSize;
myFragCoord.z = 0.5+(0.5p.z/p.w);
myFragCoord.w = 1.0/p.w;
}

Unfortunately my computed myFragCoord slightly differs from the true one. For example, when rendering a plane slightly inclined in z direction (viewing direction), i get for some fragment

gl_FragCoord = [ 8.5, 8.5, 0.999943077564239501953125, 0.005894298665225505828857421875]
myFragCoord = [ 8.50000762939453125, 8.50000762939453125, 0.9999430179595947265625, 0.005894298665225505828857421875]

The w component is ok, but not xyz. Why is that? (drives me crazy all day since i cannot perform depth comparison with these deviations…)

Note: i also read this post but i don’t understand the part of “interpolating w_clip after division, e.g. gl_FragCoord.w=(1/w_c)_i” -> this seems no good idea and doesn’t work for me - or am i missing s.th.?

[Ilian_Dinev]

gl_FragCoord.xy is always “floor(coord)+0.5” .
But generally those errors you meet are precision-errors.

[_x57]

Is this “floor(coord)+0.5” somewhere in the specs or is it just a usual thing to do to cut off the floating point error?

I think you may be right saying this is in the range of floating point precision error. Still, i’d like to know if my implementation is the best way to emulate gl_FragCoord or if there is a better way in the sense of beeing closer to the spec or resulting in greater numerical stability and less error …?

(just to think it through to the end and then get rid of this topic once and for all in my head … to make room for more problems to solve that will most likely appear out of nowhere once this one is out )