Concerns regarding how MSAA depth is resolved

RiccardoBlb · February 20, 2019, 4:06pm

As far as i can tell, the depth is resolved the same way of the color textures by MultiSample.glsllib

jMonkeyEngine/jmonkeyengine/blob/master/jme3-core/src/main/resources/Common/ShaderLib/MultiSample.glsllib#L24


#else
    #define DEPTHTEXTURE sampler2D
#endif
// NOTE: Only define multisample functions if multisample is available
#if defined(GL_ARB_texture_multisample) || (defined GL_ES && __VERSION__>=310)
vec4 textureFetch(in sampler2DMS tex,in vec2 texC, in int numSamples){
      ivec2 iTexC = ivec2(texC * vec2(textureSize(tex)));
      vec4 color = vec4(0.0);
      for (int i = 0; i < numSamples; i++){
         color += texelFetch(tex, iTexC, i);
      }
      return color / float(numSamples);
}
vec4 fetchTextureSample(in sampler2DMS tex,in vec2 texC,in int sampleId){
    ivec2 iTexC = ivec2(texC * vec2(textureSize(tex)));
    return texelFetch(tex, iTexC, sampleId);
}
vec4 getColor(in sampler2DMS tex, in vec2 texC){

This doesn’t seem correct to me, since depth represents spatial data. Lets say for example we are at the edge between a cube and the skybox, some samples might end on the skybox and some on the cube, the averaged result will place the depth somewhere in between the cube and zFar, where in reality there should be nothing.

Please let me know if i am missing something.

pspeed · February 20, 2019, 7:24pm

There is no “right” answer… so what do you suggest instead?

RiccardoBlb · February 20, 2019, 7:31pm

I would use the first sample or the sample with the nearest depth

pspeed · February 20, 2019, 8:34pm

So, “random” or “nearest depth”. Random might be interesting (no way to tell for sure which is first, I guess). Nearest seems like it would also edge cause artifacts.

Since there is no right answer on how to mix depth I guess the code took the position that most things interested in depth might be gradient based and would be happier with an intermediate value than some hard incorrect cut-off.

RiccardoBlb · February 20, 2019, 8:50pm

First sample approach would de facto behave like if there was no multisampling for the depth.

actually the current code is potentially turning edge pixels into wrong values while the “cut-off” value would be the correct one.
Then we need to consider also that depth is non linear.