Volumetric point/spot light in-scatter vanishes in large `FogVolume`s

[blaind]

Bevy version and features

• v0.19.0 (default features, Linux x86_64, Vulkan)

What you did

Built a foggy street scene: a world-scale FogVolume with street lamps and headlights as volumetric point/spot lights. Minimal form (ambient at 0 so the point light is the only fog illumination):

commands.spawn((
    Camera3d::default(),
    Hdr,
    VolumetricFog {
        ambient_intensity: 0.0,
        ..default()
    },
    Transform::from_xyz(0.0, 4.0, 16.0).looking_at(Vec3::new(0.0, 3.0, 0.0), Vec3::Y),
));

// A world-scale fog volume...
commands.spawn((
    FogVolume {
        density_factor: 0.15,
        scattering: 0.6,
        absorption: 0.05,
        ..default()
    },
    Transform::from_xyz(0.0, 5.0, -22.0).with_scale(Vec3::new(100.0, 20.0, 60.0)),
));
// ...swap the scale for Vec3::splat(4.0) and the same light below produces
// the expected glow ball.

// Shadow-mapped volumetric point light inside the fog.
commands.spawn((
    PointLight {
        intensity: 300_000.0,
        range: 40.0,
        shadow_maps_enabled: true,
        ..default()
    },
    VolumetricLight,
    Transform::from_xyz(0.0, 2.5, 0.0),
));

Full runnable reproduction:
https://github.com/blaind/bevy-volumetric-ambient-repro (MODE=pointlight)

What went wrong

• Expected: a glow ball around the light in both volume sizes.
• Actually: the 4x4x4 volume shows the expected glow; the 100x20x60 volume crushes the same light, same position, same intensity, to a barely visible pinprick.

Cause, in crates/bevy_pbr/src/volumetric_fog/volumetric_fog.wgsl (applied per raymarch step in the point/spot loop, and in the directional loop too):

let light_attenuation = exp(-density * bounding_radius * (absorption + scattering));

bounding_radius is the radius of the sphere bounding the entire fog volume, a constant. Every light at every sample is charged the cost of crossing the whole volume regardless of the actual light-to-sample distance, and the local sample density is extrapolated over that constant distance. For the large volume here: density 0.15 x radius ~59 x extinction 0.65 gives exp(-5.75), 0.3% of the light's scatter. The falloff is exponential in volume size, so realistic-lumen punctual lights die completely at world scale; directional god rays survive in practice only because tens of thousands of lux have headroom.

Additional information

• Suggested fix included in the repro as fix-punctual-light-fog-attenuation.patch

4-unit volume: the expected glow ball:

100x20x60 volume, same light: the glow is crushed to a pinprick by the bounding-radius term:

same large volume with the fix: the light-to-sample distance replaces the bounding radius, and the glow returns

fix regression check: the small volume still renders a comparable glow ball:

[blaind]

More implications from this issue / proposed fix / various initial thoughts:

• in my understanding, the volume size should not really affect how the point/spotlights inside the volume effect. I.e. there is definitely some bug there
• the fix would change existing pointlight interactions with fog, i.e. bevy volumetric volume example shows a lot brighter lights
• the fix is probably not correct yet, i.e. now it estimates the length to the pointlight source within fog. But what if the light is outside the fog, how should that affect the attenuation?