search

gfx-rs / wgpu

A cross-platform, safe, pure-Rust graphics API.
https://wgpu.rs
Apache License 2.0
12.4k stars 908 forks source link

Support extern Context implementations #6330

Open sagudev opened 3 weeks ago

sagudev commented 3 weeks ago

In servo we want to use vello for 2d canvas and we want it in content process (where it shouldn't access GPU), so we cannot run wgpu in that process. Our initial plan was to rewrite part of vello that uses wgpu to reuse webgpu implementation already present in content process (mostly just wrappers around wgpu's IDs and that sends IPC messages to wgpu thread) or abstract vello's wgpu usage behind something we can implement with our webgpu implementation. Ironically, wgpu is already an abstraction, so instead of another level of abstracting we could just implement our own wgpu::context::Context using servo's webgpu types and provide custom instance (that has our own context inside) to vello and let it do its 🪄 .

I already have dummy impl of how that would look like altogether: https://github.com/sagudev/wgpuidex/tree/dummy-vello and changes needed in wgpu to support external context impl: https://github.com/gfx-rs/wgpu/compare/v22...sagudev:wgpu:extern-impl-experiment.

Describe alternatives you've considered Alternative is that we would have ipc based backend in tree and consumers would provide ipc sender on instance creation (and we could also provide impl needed for receivers), that could be useful for those who wants to do process isolation. I wonder if it would be possible for servo/firefox to use such impl on content process, or is wgpu abstraction to much high level 🤔.

Or simply do nothing as this is really obscure request.

Additional context This will fight with https://github.com/gfx-rs/wgpu/issues/6324

We have chat about servo using vello on zulip: https://servo.zulipchat.com/#narrow/stream/263398-general/topic/Using.20Vello.20in.20Servo

Wumpf commented 2 weeks ago

This will fight with https://github.com/gfx-rs/wgpu/issues/6324

Very true unfortunately so far! @cwfitzgerald and me were hoping to simplify the wgpu forwarding a lot by no longer assuming arbitrary context trait impls. The motivation is to make it much easier to follow the path of a wgpu call to wgpu-core internals (hard & annoying even for active maintainers) and allow the indirection to be optimized out completely for native and webgl/webgpu-only cases.

Or simply do nothing as this is really obscure request.

still pondering this. I don't think it's unreasonable to be able to redirect the api. Usually the requests we get about this are for wgpu-hal which is harder still. It's also super interesting how little changes you needed right now for your dummy impl on wgpu.

Maybe we can extend the #6324 proposal to make space for a custom context next to wgpucore & webgpu. The fact that we do already have slight api differences for native & wasm doesn't make this easier, but as I understand you'd be only interested in native.

sagudev commented 2 weeks ago

The motivation is to make it much easier to follow the path of a wgpu call to wgpu-core internals (hard & annoying even for active maintainers) and allow the indirection to be optimized out completely for native and webgl/webgpu-only cases.

I agree with motivation, the indirection really makes it hard.

still pondering this. I don't think it's unreasonable to be able to redirect the api. Usually the requests we get about this are for wgpu-hal which is harder still.

IIRC somebody did try to do something like this to be able to trace wgpu or smth like that, but I do not recall at which level (but probably hal).

Maybe we can extend the https://github.com/gfx-rs/wgpu/issues/6324 proposal to make space for a custom context

That would be ideal.

cwfitzgerald commented 2 weeks ago

We could have a variant hold a Box<dyn Trait>