multiple_windows does not render in secondary window

[aevyrie]

Bevy version

main

Operating system & version

Win11

What you did

cargo run --example multiple_windows

What you expected to happen

The gltf should render in both windows

What actually happened

The model only shows in one window, but not consistently the same one

Additional information

[DJMcNab]

I can reproduce. Interestingly, it seems to arbitrarily select one camera to work; I imagine that there's some hashmap somewhere causing this behaviour.

I'm assuming it's a camera issue, since the clear pass is running successfully.

[aevyrie]

Yeah, I just noticed the same thing and updated the ticket. 👍

[mockersf]

the randomness probably comes from

bevy/crates/bevy_render/src/camera/camera.rs

Lines 270 to 281 in 4480b36

	pub fn set_active_camera<T: Component>(
	mut active_camera: ResMut<ActiveCamera<T>>,
	cameras: Query<Entity, With<T>>,
	) {
	if active_camera.get().is_some() {
	return;
	}
	if let Some(camera) = cameras.iter().next() {
	active_camera.camera = Some(camera);
	}
	}

[cart]

Yeah we changed from rendering all cameras of a given type to just the active one (which is why this used to work).
Historically we did these:

1. Prior to this pr: created a brand new camera type and manually updated the render graph. Each type (the default and the secondary camera) had its own active camera, which we selected.
2. Old version of this pr: rendered every camera with the same type (indiscriminately). Not the direction we should be headed in imo.
3. Merged version of this pr: render active cameras of a given type.

The missing piece here is choosing the "right" way to drive rendering. This is the "camera driven" vs "render target driven" discussion started in #4191.

This is next on my todo list, but I've been focused on merging other peoples' work for 0.7 for awhile now. I'm pretty close to being out of that hole though.

I think we "need" active cameras with the current paradigm (and RenderTarget driven rendering). There are flavors of camera-driven where we don't (see #4191 for more info), but our current impl isn't that.

If we can't merge a full solution before 0.7, I think the move is to re-add the "second window camera type / render graph" pattern to this example, but I really want higher level render targets that "just work" asap.

[cart]

Also sorry for the example regression 😄