272: Force the wasm32 types to match Send / Sync with the native types r=kvark a=kyren
This is potentially a controversial change so this is just a draft to discuss it.
Currently, when building wgpu on wasm32, none of the interface types are Send or Sync, whereas on any other platform they are generally Send and Sync and this can cause a lot of pain, especially in the presence of async.
This PR is the nuclear option: lie and forcibly declare nearly all interface types to be Send or Send + Sync.
I think I understand the situation with proposed wasm32 threading a bit better now, and as of right now at least with my current understanding, it seems to me unlikely that javascript values held by rust could *ever* be sent transparently to another worker. The mechanism that rust apparently will use to implement transparent threading will be to use `SharedArrayBuffer` to share memory between multiple workers. Javascript provides a way for certain (serializable) objects to be shared with with other workers through `worker.postMessage()`, so Javascript has *some* story for sending javascript values to other threads, but only plain rust values that exist solely in the `SharedArrayBuffer` memory are send-able under this scheme. I've looked briefly through some of the wasm proposals like the interface types proposal, but I don't have a good sense if there's any proposed mechanism at all for interface types and threads to interact at all.
If that's true that's really unfortunate because it seems like any thread safe API shared by native and web is doomed to mismatch. This is even more unfortunate because many of the webgpu types in the draft standard are marked as 'Serializable', meaning that they *are* actually safe to send to another worker via `postMessage`, but there does not appear to be a planned way for rust to do this transparently.
I definitely could be misunderstanding the situation though and even if I am understanding it, the situation may have since changed. I also don't know the timing here, I don't know what the eventual timing of wasm32 threading support will be vs webgpu support and interface types support. *Right now* this PR is almost certainly sound, but there's definitely a possible future where it becomes unsound, and even maybe a possible future where it becomes unsound but then could later be made sound again.
Co-authored-by: kyren <kerriganw@gmail.com>
There are several things in this commit:
1) Add `Send` or `Send + Sync` bounds to most of the associated types in `Context`.
2) Force most of the handle types and future types returned by the web backend to be Send / Sync
3) Make `BufferReadMapping` and `BufferWriteMapping` `Send` again on native by
making the associated detail types implement `Send`.
Main change here is the introduction of the Context trait.
The "direct" implementation of it goes straight to wgpu-core.
This commit also has small improvements:
- consuming command buffers on submit
- Instance type
- proper call to destructors
`File::create` isn't available on wasm, so we can't actually write the
output image anywhere unless we target something else as the file
system (e.g. local storage)
Remove odd patching matching
Replace std::sync::Mutex with spin::Mutex in GpuFuture
Reduce GpuFuture usage to one explicit allocation instead of two
Fix examples to poll the device in the background when using map_read or map_write
Remove device.poll from GpuFuture::poll and document future invariants
Massively simplify examples
Use Arc::clone(...) instead of arc.clone()
Switch println to log::info
210: Update WGPU deps to rev 08e8d406c1 r=kvark a=aloucks
I also moved the `path` chunks in `Cargo.toml` to a single `[patch]` section so that it's one block to commment/uncomment for local development.
Co-authored-by: Aaron Loucks <aloucks@cofront.net>
