Volumetric Fog Rendering

[soundform]

Does it support volumetric fog data with RGBA components? (color + density)

[emcfarlane]

@soundform not yet, but I do want to add support for it. Could you expand on an example model to test with? I know there is a fire.vdb and embergen have some animations available too.

[soundform]

The embergen samples are really good. My volumetric data is more plasma-like, meaning that it can be bright, but nearly transparent. In other words, its colors have premultiplied alpha, and the renderer does something like this:

vec4 sum = vec4(0);
float dt = 0.001; // approx voxel size

for (float t = dt*0.5; t < tmax; t += dt) {
  vec3 p = rayOrigin + t*rayDirection;
  vec4 col = texture3D(volumetricData, p);
  
  // shadowData[p] = integral(s=0..inf, volumetricData[p + s*sunDir].a)
  col.rgb *= texture3D(shadowData, p);

  // Don't do this, alpha is premultiplied.
  // col.rgb *= col.a;

  sum +=  dt*col*(1.0 - sum.a);
  if (sum.a > 0.995) break;
}

return sum;

[soundform]

Here is one of my samples - a 100 x 100 x 100 grid with 'density' and 'color' components: flame_100.zip

It would be good to support volumetric mipmaps for quick previews. For some reason, even Blender doesn't get this right.

Custom data compression would help too. Using 4 x float32 to store one RGBA voxel is often too much. For example, the density component can be represented with 1 byte and a non-linear mapping to float32: x -> 1.0 / (1.0 + exp(-x)).

A related nice feature would be block read-writes. Similar to OpenVDB's copyToArray/copyFromArray, your webgpu implementation could provide API to download or upload a block of the 3d texture: copyToArray(ijk_min, ijk_max, cpu_array) would download data to CPU, and copyFromArray would upload data to GPU.