makepad/libs/render/ENTITY_LIGHTS.md

4.2 KiB

Entity lights

Sandbox Splash can attach named point lights and spotlights to any entity:

let car = game.car({pos: vec3(0, 1, 0)})
game.headlights(car, true)
game.light(car, {
    name: "headlight_left"
    kind: "spot"
    pos: vec3(-0.7, 0.1, -2.1)
    dir: vec3(0, -0.08, -1)
    color: #fff0d1
    intensity: 4
    range: 30
    inner_angle: 12
    outer_angle: 25
    shadows: true
})
game.headlights(car, false)
game.light_remove(car, "headlight_left")

game.light(owner, options) upserts one name, preserving omitted settings. The default name is light; default kind is point. It returns false and logs invalid data, without replacing the previous fixture. game.light_remove returns whether it removed one. game.headlights installs/toggles headlight_left and headlight_right without resetting custom tuning. Both game.car and game.racecar install enabled headlights by default; headlights: false creates them disabled.

Position is body-local and follows entity scale. Direction follows the visible entity rotation (including rigid-body pitch/roll), not scale. Forward is -Z. Range is world units and does not scale. Cone angles are half-angles in degrees; require 0 <= inner <= outer and 0.5 <= outer <= 85. Color is linear RGB in 0..1; intensity is a nonnegative artistic multiplier, not lumens/candelas. Intensity is limited to 10000 and range to 0.05..10000. Names are 1..64 bytes; there are at most 16 authored lights per entity.

Shadow casters include the owner: position the emitter outside its solid mesh. This is surface lighting and occlusion, not a volumetric beam/fog effect.

Lights live in Entity::lights, clone and disappear with their owner, and are transformed every frame. Hidden physics bodies may own visible fixtures. There is no geometry rebuild or light bake on a light edit. Authored values replicate reliably in protocol version 7; movement uses the existing entity pose stream. Every peer must run the same protocol version. This does not add new rigid-body orientation fields to the network protocol.

Named model/bone sockets are not part of this API. Use body-local offsets; imported model emitters are a separate modeling API.

Local shadows

In clustered mode, shadows: true requests realtime occlusion. Shadows are off by default for generic lights and on for the headlight helper. The renderer draws a depth atlas before the forward pass, including static models, primitives, terrain/voxels, moving models/parts and skinned casters. Transparent light transmission and alpha-cutout shadow silhouettes are not implemented; dynamic primitive casters retain the existing box approximation.

renderer.set_local_shadow_config(makepad_render::LocalShadowConfig {
    max_faces: 2,
    resolution: 256,
});

Default budget: eight 512px faces. Limits: 0..16 faces, 128..1024px per face (rounded up to a power of two). A spotlight uses one face; a point uses six. The allocator prioritizes active lights by brightness/range/distance to the camera. Allocation is all-or-nothing per light; an excluded shadow request is omitted, not changed into a wall-leaking unshadowed light. This can cause popping at the budget boundary. RenderStats::clustered.shadows reports selected lights, faces, omitted requests, caster draws and CPU encoding time, not GPU duration.

Use a small spotlight budget on constrained devices; no Quest GPU performance claim is implied by these defaults. Sun CSM remains independent. The legacy MAKEPAD_CLUSTERED=off path does not support these local shadow/cone semantics.

Verification fixture

After a release sandbox build, launch from the checkout root:

MAKEPAD_CLUSTER_DEMO_HEADLIGHTS=1 \
  ./apps/commercial/sandbox/target/release/makepad-sandbox --remote   # world setting `clustered`

A moving body carries two lights toward a freestanding blocker and a rear wall. MAKEPAD_CLUSTER_DEMO_FREEZE=1 pins the body for comparison; MAKEPAD_HEADLIGHT_UNSHADOWED=1 removes the two shadow requests for an A/B run. MAKEPAD_HEADLIGHT_POINT=1 adds a blue, six-face point light. MAKEPAD_HEADLIGHT_SHADOW_FACES=N overrides the atlas budget in this fixture. Finish owned test instances through their printed /gq endpoint.