SSAO with reconstructed depths yields strip artifacts

Question

I wrote an ssao algorithm with compute shaders and view-position-reconstruction from the depth buffer. It works great for objects near by but results in strip or banding artifacts for large planes further away:

This does not happen if i use an explicit position buffer. I don't think this is because of floating point accuracy. My methods to reconstruct the position for the original pixel and the z coordinate for the pixels in its hemisphere:

vec3 calcViewPos(vec2 texCoords, float depth){
    vec3 clipPos = vec3(texCoords, depth) * 2.0 - vec3(1.0);
    vec4 viewPos = vec4(vec2(invProj[0][0], invProj[1][1]) * clipPos.xy, -1,
                                   invProj[2][3] * clipPos.z + invProj[3][3]);
    return(viewPos.xyz / viewPos.w);
}

I use a GL_DEPTH_COMPONENT32 depth texture. I could solve the problem by a little hack: Comparing the normal of the original pixel and the normal of the neighbourhood pixels and skipping those which have a too similar normal solves this problem (since this only occurs on those large plains), but causes an additional texture lookup in a for loop, which costs too much (and would make using a position buffer again more efficient).

Did any of you experience a simillar problem or have ideas of how to solve this smarter? Thank you very much :)

Answer 1

For anyone having the same problem: I calculated the normalized texture coordinates by

vec2 imgSize = imageSize(img_output);
vec2 normtc = vec2(gl_GlobalInvocationID.xy) / imgSize;

this is of course wrong, the last line should be:

vec2 normtc = (vec2(gl_GlobalInvocationID.xy) + vec2(0.5)) / imgSize;

since we want to sample the center of the pixel. This solved the problem for me.