Multi-adapter support #227
Conversation
rosefromthedead
force-pushed
the
multiadapter
branch
from
60ae7bf
to
1d33988
Compare
|
Would just like to note that this branch fixes the issue for me as well; I am running on X11 (since I still can't screenshare using Zoom on Wayland), and have both integrated and discrete graphics. Some more detail on Zulip. |
|
Re: the discussion on Zulip - for what it's worth, I think this is complexity that has to happen somewhere. Picking an adapter with no criteria won't work, as we've seen for nvidia laptops. I tried to balance vello complexity and API consumer complexity, by exposing the fact that multiple adapters are in use but leaving the selection and reuse logic on our side. That logic could be left to the user, or the whole business could be hidden away, but whatever happens, there's gonna be more work somewhere. |
|
To add some context, I agree that we’ll need extra complexity somewhere. The open question for me is whether it’s actually necessary to maintain a collection of devices or just choose the appropriate one. In the former case, we’ll also need to duplicate the Renderer state and all associated resources per device. I also wonder what happens if you drag a window to another screen that may be attached to a different device. For the latter case, I’d rather add the logic to select a single device that supports the surface. |
|
I do appreciate the fix and this does address the immediate problem allowing people with dual GPUs to actually run the code so if you fix the conflicts I’ll be happy to approve this and get it merged. We can revisit to address the full implications later. |
some `RenderContext` stuff is now `DeviceHandle` stuff, and device handles are created alongside surfaces to ensure compatibility, but are reused if possible. Fixes linebender#224 wgpu Adapter and Surface might not be compatible
rosefromthedead
force-pushed
the
multiadapter
branch
from
1d33988
to
c5401e7
Compare
|
That's weird, the merge conflicts weren't showing up locally, so I just rebased and hoped for the best. I'd always imagined that dragging windows across screens should force apps to make a new surface for each new screen the window is on; maybe because I've been reading too many discussions about fractional scaling. In the absence of this, I see that things could get messy with my code. Good point about duplication. Looking at wgpu again, I think it's possible to minimise the number of devices in use, but it would mean that a surface needs to switch device if a new surface is created. Room for improvement, I guess |
Now Adapter, Device, Queue are bundled into DeviceHandle and created alongside Surface, or reused if we already created a compatible handle.
I'm open to design complaints, just wanted to prove that the logic would solve my problem from #224, and it does :). Now on x11 (through xwayland) the winit example works fine, and on wayland I get a different error message than before 🥳
All of this comes at the cost of a little extra user-side juggling, but in the winit example it's not much extra effort.