19:27 Ermine: Dear friends, my sloppy vulkan code made GPU reset. Is it to be expected? And should I file bug report against amdgpu or radv?
19:27 Ermine: GPU in question is rx 470
19:49 Remco: Are you using the Vulkan Validation Layers?
19:51 Remco: Vulkan doesn't do a lot of error checking so it's easy to crash. Use the layers to verify your code is right
20:06 Ermine: I do
20:06 Ermine: And my code is wrong
20:08 Ermine: i expected my code to crash and not to get system in a state where it needs to be rebooted from ssh
20:09 fililip: gpu resets should work fine though
20:09 fililip: are you using something other than kwin?
20:09 Ermine: sway
20:09 fililip: hm, that should support gpu resets, weird
20:09 fililip: wlroots got support for it recently
20:10 Ermine: screen got distorted, and some kernel workqueue got hung
20:10 fililip: but yeah, other gpu vendors somehow manage to just kick the faulty job out and leave everything else running
20:11 fililip: maybe try it on kwin and if it doesn't work there, file a bug report
20:12 Ermine: oh, it worked on kwin
20:13 Ermine: ... or not
20:13 Ermine: no
20:16 Ermine: it tries hard to restore...
20:17 fililip: if latest kernel and mesa from master don't work, submit a report here: https://gitlab.freedesktop.org/drm/amd/-/issues/new
20:17 fililip: a timeout is not supposed to destroy your session like that
20:21 Ermine: but, generally speaking, should I expect my faulty code to crash GPU?
20:22 fililip: crash gpu and time it out, yes; make the session unusable, no
20:22 HdkR: amdgpu kernel recovery definitely isn't 100% reliable, so it's not entirely unexpected.
20:23 fililip: for instance, you can just write a fragment shader that executes over a very long period of time, that is enough to trigger the timeout mechanism
20:25 Ermine: okay, will keep that in mind