Update `GlobalTransform` as part of `FixedUpdate`

[jabuwu]

What problem does this solve or what need does it fill?

The purpose of FixedUpdate and fixed timesteps in general is to improve determinism in systems, but any fixed update system relying on GlobalTransform will currently suffer from a lack of determinism if it runs multiple times in 1 frame.

What solution would you like?

GlobalTransform should be updated between fixed updates.

What alternative(s) have you considered?

Games could add these systems to FixedUpdate themselves. It's actually fairly easy to do already.

use bevy::{
    prelude::*,
    transform::systems::{propagate_transforms, sync_simple_transforms},
};

#[derive(Debug, Hash, PartialEq, Eq, Clone, SystemSet)]
pub enum FixedTransformSystem {
    TransformPropagate,
}

// ...

app.add_system_to_schedule(
    CoreSchedule::FixedUpdate,
    sync_simple_transforms.in_set(FixedTransformSystem::TransformPropagate),
)
.add_system_to_schedule(
    CoreSchedule::FixedUpdate,
    propagate_transforms.in_set(FixedTransformSystem::TransformPropagate),
);

[alice-i-cecile]

WRT transform propagation, @inodentry has previously advocated for separating out a "physics transform" from a "rendering transform", and using interpolation to avoid visual hiccups.

This is pretty compelling to me, but definitely doesn't seem compatible with the approach proposed here.

[alice-i-cecile]

I would also be interested in seeing if #7838 could be used to resolve these problems more gracefully.

[jabuwu]

I would also be interested in seeing if #7838 could be used to resolve these problems more gracefully.

While writing this issue, I did think about suggesting a way to manually "flush" GlobalTransform the same way we do command buffers, but quickly realized this is already possible. It's just not very pretty. Your before_and_seperated_by syntax is a nice solution here. I think it needs to be before_and_seperated_by(b, (sync_simple_transforms, propagate_transforms)). With two systems?

[alice-i-cecile]

Yeah agreed: that's a nice catch that we may need to force separation by a collection of systems instead.

[james7132]

In it's current state, transform propagation requires a full hierarchy traversal, this can be quite expensive, even if nothing changes, on large worlds. While this is probably needed for ECS-native physics simulation to work, I'm hesitant to suggest adding it now when the FixedUpdate schedule is empty by default, and runs potentially multiple times per frame. Ideally we should have some optimizations like #7840 before we enable this.