Device objects are not cleared on wgpu-native

[almarklein]

This is a potential memory leak. Not a severe one, because people should probably not create many devices. See the memtests we added in #403. It looks like this is the only object that leaks.

I've looked into this, but have not been able to find a way to fix this in wgpu-py.

Plan / ideas:

• Wait until we update to wgpu-native 0.18.0.1, with luck its fixed in a newer version of wgpu-core.
• Check if this issue is already known. -> There are a few reports of memory leaks in wgpu-core, but not this one particularly.
• Create a minimal example to reproduce this using wgpu-native directly, and create an issue for it.

[Vipitis]

I think I am running into a memory leak as well, after running something between 50 and 100 different shadertoys through the utility (just creating the class, not even rendering a single frame), I can see my vram usage grow to about 4GB and then the python process exits. It doesn't throw any exceptions, nor are there any rust panics and it seems like it just finished (it's a kernel crash when when doing this inside a jupyter notebook).

It could be due to my Intel A750 GPU and their clearly buggy compute runtime, so I will double check the behavior on my laptop with nvidia GPU tomorrow.

Is there a some method to clear up vram apart from killing the python process?

[Korijn]

Is the CPU ram also growing? Or just the VRAM?

[almarklein]

You could try to modify the shadertoy.py in _prepare_render(). Where it requests an adapter and then a device, it could use get_default_device() (from wgpu.utils.device). I think that would make sense anyway, and it might just solve your leak?

[Vipitis]

Is the CPU ram also growing? Or just the VRAM?

there was no increase in CPU RAM.

You could try to modify the shadertoy.py in _prepare_render(). Where it requests an adapter and then a device, it could use get_default_device() (from wgpu.utils.device). I think that would make sense anyway, and it might just solve your leak?

I changed this specific line with your recommendation and it seems to be an improvement. There is no longer any uncontrolled vram rise. It looks steady for the most part. I still encountered a crash after 450 shaders, but will investigate a bit more.

I can make a PR for that (if there isn't any more changes given the other PRs right now).

[almarklein]

That's good to hear! Perhaps the crash after 450 shaders is not related to the leak. Still odd though. A PR for this part of the fix would be great!

[almarklein]

Still leaking in wgpu-native 0.18.1.1

[almarklein]

This will be fixed by #458

[Korijn]

This will be fixed by #458

Really!! Fantastic news 🎉