[RADV] [Navi] Shenmue I & II (758330) - Graphical glitch (DXVK)
System information
- OS: Gentoo
- GPU: [AMD/ATI] Navi 10 [Radeon RX 5600 OEM/5600 XT / 5700/5700 XT] (rev c1)
- Kernel version: 5.7.8 with may amd-staging-drm-next updates
- Mesa version: OpenGL version string: 4.6 (Compatibility Profile) Mesa 20.2.0-devel (git-41c7bb6e)
- Xserver version (if applicable): XWayland git
- Desktop manager and compositor: Sway/Wayland
If applicable
- DXVK version: git
- Wine/Proton version: 4.11-13, 5.0.9
Describe the issue
Okay, so I found a regression in this game with DXVK, resulting in minor graphical glitches. I tracked down the regression in DXVK to:
68be040f4aec52d39f0f3b77236215e2385424e7 is the first bad commit commit 68be040f4aec52d39f0f3b77236215e2385424e7 Author: Philip Rebohle philip.rebohle@tu-dortmund.de Date: Tue May 5 12:49:13 2020 +0200
[dxvk] Use compute queue for transfers if DMA queue is not available
Affects RADV. Let's see if this works and does anything for performance.
I then tested older versions of Mesa, and at least from 20.1.3 earlier, the bug doesn't exist. If I revert to a DXVK version before that regression, the bug also doesn't exist. It affects both ACO and LLVM.
I've had trouble with getting this particular game to adhere to VK_ICD_FILENAMES (likely due to needing a custom launcher to workaround dotnet issues) so doing a mesa regression test is a bit painful.
Here's a screenshot of the issue:
Note the red lines, although they are sometimes black, and tend to pop in and out.
I've had trouble getting an apitrace or renderdoc capture on Linux, but I'll try and get something from Windows a bit later.
Regression
Somewhere after 20.1.3 and before mesa-git