Bug 106980

Summary:	Basemark GPU vulkan benchmark hangs on GFX9
Product:	Mesa	Reporter:	Bas Nieuwenhuizen <bas>
Component:	Drivers/Vulkan/radeon	Assignee:	mesa-dev
Status:	RESOLVED FIXED	QA Contact:	mesa-dev
Severity:	normal
Priority:	medium	CC:	0xe2.0x9a.0x9b, airlied
Version:	git
Hardware:	Other
OS:	All
Whiteboard:
i915 platform:		i915 features:
Attachments:	Miscompiled shader

Description Bas Nieuwenhuizen 2018-06-20 22:45:28 UTC

basemark: https://www.basemark.com/products/basemark-gpu/

to run you need some old libraries:
https://packages.ubuntu.com/bionic/amd64/libcurl3/download
https://packages.ubuntu.com/bionic/amd64/libssl1.0.0/download

just download, extract and use LD_LIBRARY_PATH

it aborts with

extra entries in SSA def uses:
0x7ff0281861a0


Christoph Haag tested without nir_validate and for him it hung.

Comment 1 Bas Nieuwenhuizen 2018-06-20 23:05:54 UTC

with some print debugging, it looks like the failing validate is here:

https://gitlab.freedesktop.org/mesa/mesa/blob/master/src/amd/vulkan/radv_shader.c#L233

which is directly after spirv_to_nir. I think it is a reasonable assumption that this is a spirv_to_nir bug.

Comment 2 Bas Nieuwenhuizen 2018-06-20 23:23:21 UTC

Created attachment 140254 [details]
Miscompiled shader

Attached is the failing shader. spirv2nir is enough to trigger if you add a nir_validate_shader call at the end.

Comment 3 Bas Nieuwenhuizen 2018-06-23 10:50:28 UTC

Seems like the validation issue was fixed by the deref rewrite, and that has been upstreamed now.

However, now I get the hang too ...

Moving back to the radv-component since there is no hint at the moment that it is an issue on anv too.

Comment 4 Jason Ekstrand 2018-06-23 16:26:05 UTC

(In reply to Bas Nieuwenhuizen from comment #3)
> Seems like the validation issue was fixed by the deref rewrite, and that has
> been upstreamed now.

That makes me suspicious.  If we were getting a validation error before, likely something is still going wrong.

Comment 5 Jason Ekstrand 2018-06-25 16:15:42 UTC

I just sent two patches for stable which fix the validation issues.  Turns out there's nothing wrong with spirv_to_nir, just the validator having issues.  Sadly, that means the hang is likely real. :-(

Comment 6 Samuel Pitoiset 2018-08-22 12:41:46 UTC

The following patch fixes a crash on GFX8:
https://patchwork.freedesktop.org/series/48560/

Comment 7 Samuel Pitoiset 2018-08-22 13:36:18 UTC

The crash on GFX8 should be fixed with https://cgit.freedesktop.org/mesa/mesa/commit/?id=4c43ec461de4f122d5d6566361d064c816e4ef69

Next step is to test with GFX9.

Comment 8 Samuel Pitoiset 2018-10-31 15:21:56 UTC

I can't reproduce a GPU hang on GFX9 with latest mesa/llvm. Closing.

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.