No rings in the visualization of the Saturn system

[typpo]

No rings in the visualization of the Saturn system

[typpo]

Thanks for the report. I'm on Mac Firefox and am able to see the rings. So I suspect it may be some combination of OS+GPU.

Here is an old but potentially related issue. However, my ring shader doesn't return any structs. So I'm not sure the fix is directly applicable.

[Postrediori]

So I suspect it may be some combination of OS+GPU.

Yes, this happens to be a Windows-specific issue. Ubuntu 18.04 on the same machine (using Live USB) renders rings as intended.

[Postrediori]

However, my ring shader doesn't return any structs.

It seems that the version of RING_SHADER_FRAGMENT from src/shaders.js is not called from anywhere in the Saturn demo. Instead the shader is in build/spacekit.js.

Warning: D3D shader compilation failed with default flags.

I've found out the similar error in the TensorFlow JS library (tensorflow/tfjs#1337). It's not directly aplicable to this case, but the solution for that bug (tensorflow/tfjs-core#1646) showed that there might be flaws hidden deep in a valid and working code because of the GPU drivers (like an inability to handle integer division correctly in TensorFlow case).

So I've checked the ring's frament shader for suspicious things. There is the following calculation of the pixel's colour:

  void main(void) {
	gl_FragColor = color() * vec4(lights() * shadow(), 1.0);
  }

The color() function contains discard option as the conditional path.

  vec4 color() {
    ...
    if (...) {
      discard;
    }
    ...
  }

Commenting this out (or replacing with return vec4(0.0)) fixes the problem (and renders rings wrong of course).

My guess is that the driver on Windows cannot handle variable deallocation correctly. When an operand of the multiplication operation (color() function) calls discard it does not discard the multiplication operation as a whole. Light and shadow components are calculated anyway and when the second operand vec4(...) is calculated it tries to multiply the result by the left operand that has already been discarded and the driver crashes (hence the vague 'unbound variable' text in the error log).

Switching the operands of the final multiplication allows rings to be rendered:

	gl_FragColor = vec4(lights() * shadow(), 1.0) * color();

Also, saving light&shadow components to a temporary vector also works fine as right value is already known when discard in color() is executed:

        vec3 lightAndShadow = lights() * shadow();
	gl_FragColor = color() * vec4(lightAndShadow , 1.0);

This is just my first look on the problem and I'm not sure this is the final or 'proper' way to fix this.

P.S. The following version also works for some mysterious reason:

  void main(void) {
        vec3 foo = lights();
        vec3 bar = shadow();
	gl_FragColor = color() * vec4(lights() * shadow(), 1.0);
  }

[typpo]

@Postrediori Thanks so much for the detailed debugging help! I've made one of your suggested changes:

gl_FragColor = vec4(lights() * shadow(), 1.0) * color();

This means the example should load correctly (you may have to hard refresh): https://typpo.github.io/spacekit/examples/saturn/index.html. How does it look?

[Postrediori]

Rings are rendered correctly:

Checked on Win10 & 8.1, Firefox and Chromium.

Thanks!

[typpo]

Thank you again for the help!

[Postrediori]

FYI Maybe a bit late, I reported this issue to Chromium bugtracker: https://bugs.chromium.org/p/chromium/issues/detail?id=1192924. It was labeled as Wont Fix though, because is is a bug somewhere in legacy DirectX dll.

[typpo]

This is interesting, thanks so much for following up. I've run into strange issues with ANGLE before.