John Nagle [email protected]
Reviving Rend3 after its abandonment.
Changes to WGPU have broken Rend3. The worst change is
which changes the lifetime of "RenderPass" and Encoder to 'static
which generates this problem:
RenderPass and Controller ownership is currently manipulated by unsafe code. That needs to be figured out.
###Trouble spot: take_rpass
pub(super) enum RenderGraphEncoderOrPassInner<'a, 'pass> {
Encoder(&'a mut CommandEncoder),
RenderPass(&'a mut RenderPass<'pass>),
#[default]
None,
}
pub struct RenderGraphEncoderOrPass<'a, 'pass>(pub(super) RenderGraphEncoderOrPassInner<'a, 'pass>);
...
pub fn take_rpass(&mut self, _handle: DeclaredDependency<RenderPassHandle>) -> &'a mut RenderPass<'pass> {
match mem::take(&mut self.0) {
This in itself seems legit, i.e. safe. Except that it's taking a mutable reference, not the object itself. There can only be one such mutable reference. Where does this come from?
Around here:
where there is unsafe code to seemingly allow two mutable references to the same object to exist simultaneously.
// SAFETY: There is no active renderpass to borrow this. This reference lasts for the duration of
// the call to exec.
None => RenderGraphEncoderOrPassInner::Encoder(unsafe { &mut *encoder_cell.get() }),
Note that this is not an assertion that the code here is safe. It is an unchecked constraint on the rest of the program.
Why is this not done with a RefCell and .borrow_mut?
Other than the mess around RenderPass, there's not that much unsafe code in rend3. This might be fixable.
Maybe if we pass around something that has a RefCell to everybody who needs that... But that encapsulates a mutable reference. Who actually owns the thing? Question: if we put RenderPass and Encoder inside an Rc<RefCell<>>, will that work? Or will we panic at borrows? Worth a try.
Who actually owns RenderPass and Encoder?
RenderGraphEncoderOrPassInner holds an &mut of CommandEncoder or RenderPass
Who actually creates RenderGraphEncoderOrPassInner?
-
Created in graph.rs execute
RenderGraphEncoderOrPassInner::RenderPass(rpass)None => RenderGraphEncoderOrPassInner::Encoder(unsafe { &mut *encoder_cell.get() }),Unsafe code. Is this really necessary?
Same situation.
That's a get from an UnsafeCell. That's probably what should be a RefCell. The UnsafeCell actually owns the encoder. Who owns the RenderPass? The caller of create_rpass_from_desc
create_rpass_from_desc
returns an Rc<RefCell>.
At line 391 in graph.rs, encoder_cell becomes an Rc> of a new command encoder.
Similarly for rpass_temps_cell.
RenderGraphEncoderOrPassInner must hold an Rc<RefCell<>> for each item.
So where do we do the borrows?
At the unsafe points? Probably.
What about take_rpass? Do the borrow at self.internal.render in lib.rs?
This will probably all compile and it's all safe code. If the comments about safety are true, the borrow_mut calls should not fail. We will see.
Fix all other compile errors from Winit changes. [DONE]
Completed fixing Winit problems, with event loop disabled. Time to do the above changes to Rend3.
Most of it now compiles.
Trouble at line 507 of graph.rs