Mesa RADV fails to run Basemark 1.2.1 with "High" pre-set
- Have Polaris AMD GPU (AMD RX 470 or 570 8 GiB, same results with both), Linux 5.9 rc5 (or any kernel, really) and Mesa 20.2 rc4 (or earlier Mesa, doesn't really matter).
- Get the "Ubuntu" stand-alone tar.gz of the proprietary Finnish Basemark GPU benchmark from https://www.basemark.com/benchmarks/basemark-gpu/ (Make sure you don't get the Flatpak or you will end up using some wildly outdated Flatpak Mesa/LLVM)
- Start Basemark GPU, change no settings, click big red "Run Official Test" button
- Watch half a benchmark followed by disaster.
The result is: The benchmark runs half-way through and then either X hangs and ssh'ing into the machine is required or X restarts (depends on kernel version).
Kernel ring buffer story when this happens is, on Linux 5.9 rc5,
[ 611.999724] [drm:amdgpu_dm_atomic_commit_tail] *ERROR* Waiting for fences timed out! [ 611.999734] [drm:amdgpu_dm_atomic_commit_tail] *ERROR* Waiting for fences timed out! [ 617.120300] [drm:amdgpu_job_timedout] *ERROR* ring gfx timeout, but soft recovered
- Running Basemark GPU Vulkan at "High" with Mesa RADV at any resolution results in the disaster.
- Running Basemark GPU Vulkan at "High" with competing AMDVLK 2020.Q3.5 Vulkan driver works fine at any resolution other than 4K. Basemark GPU does fail in a similar fashion with AMDVLK 2020.Q3.5 at 4K and "High".
- Running at resolutions from 1080p to 4K "Medium" with Mesa RADV is fine (also fine with AMDVLK 2020.Q3.5).