Renderpass take resource ownership

[Wumpf]

Connections

• Fixes Lifetimes on RenderPass make it difficult to use.  #1453
  • 🥳🥳🥳
• Follow-up to Separate renderpass arc resolve & renderpass consume on end #5794
• Equivalent to previous compute pass prs:
  • Make compute pass end consume the pass #5575
  • Remove lifetime dependency of ComputePass to its parent command encoder #5620
  • Fix QuerySet ownership of ComputePass #5671
  • Reintroduce computepass->encoder lifetime constraint and make it opt-out via wgpu::ComputePass::forget_lifetime #5768

Description
Just like ComputePass before, we now take ownership of all ingoing resources to a RenderPass right away. We also take ownership of the parent encoder, making it possible to convert RenderPass<'a>->RenderPass<'static> (as opt-in)

Testing
New ownership tests added (very similar to what we added for compute pass previously)

Checklist

• Run cargo fmt.
• Run cargo clippy. If applicable, add:
  • --target wasm32-unknown-unknown
  • --target wasm32-unknown-emscripten
• Run cargo xtask test to run tests.
• Add change to CHANGELOG.md. See simple instructions inside file.

[John-Nagle]

It turns out that this change badly broke Rend3. See

https://users.rust-lang.org/t/unexpected-borrowed-data-escapes-outside-of-closure-for-generic/119501/2

(Rend3 was abandoned by its author around the time this went in. But I'm stuck using it, and now I have to fix it.)

[grovesNL]

@John-Nagle This PR loosens the lifetime requirements on the wgpu side. The change to 'static lifetime in egui makes you require forget_lifetime otherwise everything would work the same as usual.

You'll need to call forget_lifetime before you share renderpasses with egui. See emilk/egui#5149 for background and examples.

[John-Nagle]

I get that, mostly. The worst trouble is near here:

https://github.com/BVE-Reborn/rend3/blob/trunk/rend3/src/graph/graph.rs#L403

This code is trying to manage the lifetime of the render pass and encoder from unsafe code.
I'm still trying to figure this out. See the "SAFETY" comment:

// SAFETY: There are two things which borrow this encoder: the render pass and the node's
// encoder reference. Both of these have died by this point.

That's not an assurance that this is safe code. It's a warning that there are lifetime requirements imposed on the caller. Ones which are changing now. The code in that function does considerable unsafe manipulation of encoder and render pass lifetimes. So changes in this area are likely to cause undefined behavior.

I really don't want to touch that code if I can possibly avoid doing so. Is that possible?

[Wumpf]

@John-Nagle lifetime requirements only got relaxed and more things are now handled as a runtime error rather than a compile time error which is what allows forget_lifetime to be a safe operation. My expectation is that everything still works fine except that there's some lifetime annotations that can just go away - any lifetime annotation that ties a resource (e.g. pipeline) to the lifetime of a render pass should no longer be needed (and in some cases may cause compile errors).
From the little I know about rend3 via its author the unsafe code in this area is exactly because previously the ownership management was only via lifetime annotation & there was no forget_lifetime. But keeping the old unsafe code should still work fine afaik

[John-Nagle]

The unsafe code in this area is exactly because previously the ownership management was only via lifetime annotation & there was no forget_lifetime.

I think that's about right. It looks to me like the old WGPU lifetime system forced complex unsafe ownership management. I'm trying to figure that out now. I think ordinary Rc and RefCell can do the job now. Thanks.