1764953 - Issue with tf.js webgl based video processing demo

Status: RESOLVED WORKSFORME

(Core :: Graphics, defect, P3)

Description

[Jim Mathies [:jimm]]

Reported to us by folks at Zoom. I was finally able to reproduce this on an Intel UHD Graphics 620 while running swiggle.

I had to run the demo for about 30 minutes before I hit this fault.

Comment 1

[Jim Mathies [:jimm]]

Attachment: about:support software backend

Comment 2

[Jim Mathies [:jimm]]

Attachment: console output

Comment 3

[Jim Mathies [:jimm]]

Note the 'High memory usage in GPU: 1898.45 MB, most likely due to a memory leak' message happened much earlier than the fault.

Comment 4

[Jim Mathies [:jimm]]

Oddly enough, no gpu process crashes recorded. I'm going to enable the accelerated backend on this device and try again.

Comment 5

[Jim Mathies [:jimm]]

This does look like some sort of leak, maybe in the intel driver - monitoring the gpu process memory consumption I can see it slowly climbing from around 18GB up. A majority of this falls in heap-unclassified according to the memory report.

Comment 6

[Jim Mathies [:jimm]]

Reproduced with webrender as well. After the fault occurred (which took about 45 minutes to trigger) and closing the tensor tab, memory use in the gpu process returned to normal.

Comment 7

[Kelsey Gilbert [:jgilbert]]

Yeah, sounds like it might be a webgl leak?

Comment 8

[Jim Mathies [:jimm]]

Attachment: chrome console

I reproduced this on Chrome as well on the same intel device -

Comment 9

[Jim Mathies [:jimm]]

Zoom was able to address this on their end. Leaving this bug around as a reference since that test case is publicly accessible.

Comment 10

[BugBot [:suhaib / :marco/ :calixte]]

The severity field is not set for this bug.
:bhood, could you have a look please?

For more information, please visit auto_nag documentation.

Comment 11

[Jeff Muizelaar [:jrmuizel]]

I believe there was a misunderstanding and this problem is still an issue.

Comment 12

[Jeff Muizelaar [:jrmuizel]]

It looks like we're leaking gpu memory on this demo. gpu-committed goes up and up until it runs out.

Comment 13

[Jeff Muizelaar [:jrmuizel]]

It look like the leaks are also happening in Chrome. However, I don't see the same leaks happening when using a more recent version of tfjs (3.21.0 vs 3.13.0) so perhaps the problem has already been fixed upstream.

Comment 14

[Jeff Muizelaar [:jrmuizel]]

Since this also seems to happen in Chrome I think we can resolve it as WFM