GPU hang on Intel Core i7-7500U CPU

[eaglgenes101]

With integrated gpu cycles free to burn on my Ubuntu system, I decided to run https://github.com/gcp/leela-zero with this driver. In the middle of the self-tuning process, it hung, with the following dmesg messages shortly after about 6 seconds of screen freeze:

[ 6583.292439] [drm] GPU HANG: ecode 9:0:0x8ed9fff2, in leelaz [9165], reason: Hang on rcs0, action: reset
[ 6583.292441] [drm] GPU hangs can indicate a bug anywhere in the entire gfx stack, including userspace.
[ 6583.292442] [drm] Please file a _new_ bug report on bugs.freedesktop.org against DRI -> DRM/Intel
[ 6583.292442] [drm] drm/i915 developers can then reassign to the right component if it's not a kernel issue.
[ 6583.292442] [drm] The gpu crash dump is required to analyze gpu hangs, so please always attach it.
[ 6583.292443] [drm] GPU crash dump saved to /sys/class/drm/card0/error
[ 6583.292448] i915 0000:00:02.0: Resetting rcs0 after gpu hang

The contents of /sys/class/drm/card0/error I copypasted to https://gist.github.com/eaglgenes101/2c30c93d15953c75d476c2486a0b608a .

When I run the self-tuning process with beignet, it runs fine, but with repeated messages of "Beignet: "Work group size exceed Kernel's work group size."" during the tuning process, increasing with frequency as the tuning process goes on.

Where should I start digging deeper about this problem? And what further information should I provide to help solve this crash?

[MichalMrozek]

Hello,
I have reproduce this problem and root caused it.
This is due to the fact that application is passing greater local work group size than device supports.
Driver should report an error here, while it doesn't.
I am currently fixing this problem.

[MichalMrozek]

Problem fixed with commit ce8c44c.
Can you check on your side?

[eaglgenes101]

Yes, it works. There's a performance reduction of about 10% compared to beignet, but that's another issue for another time. Actually, a repeated tune is back to parity with Beignet, indicating what's probably a time/performance variation during the first tuning process.

[MichalMrozek]

Glad to hear that problem is fixed.