100% CPU usage since the update to wgpu 0.20

[EmbersArc]

Describe the bug

Any egui app uses 100% CPU when redrawing since commit 814ad07, the wgpu 0.20 update.

To Reproduce
Steps to reproduce the behavior:

1. Run e.g. cargo run --bin hello_world
2. Move the cursor
3. The CPU usage of the egui process goes up to 100%

Expected behavior

Only a fraction of the CPU usage as observed before the wgpu update.

Screenshots

Desktop

• OS: Arch Linux 6.10.8-arch1-1 (64-bit)
• X11 with NVIDIA 560.35.03

[EmbersArc]

I did some bisecting and the offending wgpu PR seems to be gfx-rs/wgpu#4967.

[Wumpf]

egui is already on wgpu 22.0, so unless there is reason to believe this is fixed on wgpu main (which pr fixes it?) this is completely unrelated

[EmbersArc]

Switching to wgpu trunk does not fix it yet unfortunately.

[jleibs]

@jleibs I can try to repro this locally

[EmbersArc]

For what it's worth, this is not related to egui directly. I can also reproduce it with the wgpu boids example, i.e. with cargo run --release --bin wgpu-examples boids. gfx-rs/wgpu#6067 seems related although it mentions lag and not CPU usage.

[jleibs]

I can also repro this on Arch w/ 560.35.03.

I am running through Wayland. On wayland native I don't see the issue, but I do when using xorg-xwayland.

@EmbersArc can you confirm you're running X11 native and not wayland+xwayland?

[jleibs]

On wayland, the boids example runs at 4400 fps using 30% CPU.
On X-wayland, the boids example runs at 80 fps using 100% CPU.

[EmbersArc]

@EmbersArc can you confirm you're running X11 native and not wayland+xwayland?

Yes this is on X11 native.

[cwfitzgerald]

Howdy, wgpu maintainer here. We really will need a bit more information on the performance usage here. Would it be possible to take a performance profile to understand where the cpu is spending its time? Vulkan swapchain sync is mildly cursed, so even after the rewrite we may be still doing it subtly wrong.

[Wumpf]

@jleibs and me will dive a bit deeper into this either this or next week (we work at the same company and while I'm not as knowledgeable as @cwfitzgerald in this area, I know a few things about wgpu from being a maintainer myself 😄, making this easier)
From what we learned so far this is very related to the latest Nvidia drivers as there has been a lot of changes on how they handle vsync. But I also suspect that there's some subtleties in wgpu still.

[jleibs]

Also because I happened to turn on the other computer I'll note that I also see this on:

• OS: Ubuntu 22.10
• X11 with Nvidia 535.54.03

So, it does not appear unique to the more recent 560-series drivers.

[Wumpf]

Apparently the validation layer has some interaction with this as well, we should make sure that we try release versions and debug versions with disabled validation, see
gfx-rs/wgpu#5756 (comment)