Created attachment 102222 [details]
luxrays diff/patch to disable image support while reproducing bug.
While working on getting the Luxrays slg4 program running (which is invoked as part of the luxmark CL-based raytracing benchmark), I encountered some GPU hangs on the evergreen card I'm using (CEDAR, Radeon 5400-series).
Radeon 5400-series PCI express w/ dual 1080p monitors
Mesa: Git master
drm: fairly recent git
libclc: git master + a few extra patches to implement missing built-ins
luxrays: From http://src.luxrender.net/luxrays, with the attached patch.
1) Checkout luxrays source from the mercurial repository
2) Apply patch to disable image support in the renderer
3) Build luxrays from source (cmake).
4) run bin/slg4
5) Watch the UI start to pop-up and the GPU reset.
dmesg, kernel source, llvm bitcode to be attached. rlockup_*.c cs trace available as a 35MB gzipped C file (>300MB uncompressed).
Created attachment 102223 [details]
Kernel source for luxrays
Created attachment 102224 [details]
dmesg from luxrays GPU hang/reset
Grumble.. I was just attempting to get the bitcode since I had lost my previous dump of that... and now for some reason the GPU's not locking and I'm instead getting an LLVM diagnostic handler and a segfault due to invalid reads (according to valgrind).
I'm pretty sure I'll be able to reproduce this, but part of my software stack needs to be rebuild first.
Created attachment 102226 [details]
bitcode, assembly, and rest of stderr from slg4 run
This contains the llvm bitcode, R600 assembly, and the rest of the stdout as generated from:
R600_DUMP_SHADERS=1 RADEON_DUMP_SHADERS=1 R600_DEBUG=cs,compute,trace_cs bin/slg4 2> bitcode.ll
The bitcode and hangs were all reproduced with luxrays commit 2099 (hg up -r 2099) and the image support disabling patch applied. Commit 2099 is the same as was used for the luxrender_v1.3.1 tag.
-- GitLab Migration Automatic Message --
This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity.
You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/mesa/mesa/issues/515.