Tekken TT2 crash upon character select #1100
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Current Behavior
I compiled Cemu for Fedora 39 using my project:
https://github.com/leifliddy/podman-cemu-builder
Update: this issue also occurs with the
Cemu-2.0-66-x86_64.AppImage
as well.Cemu + Tekken has always worked perfectly. However, I recently changed the GPU on my system from an Nvidia RTX 3070 --> AMD RX 7900 XTX. So I believe this issue is related to the AMD card (with Cemu or Mesa)
I'm just using the open source drivers from the kernel and Mesa 23.3.5
And now the game crashes (consistently and always with the same error). Which is right after selecting a character.
I created a new Cemu env to test this with. So no shader caches existed....no existing game profiles, and just using the default cemu config.
**as a side note: Tekken 8 (Steam/Proton) runs perfectly on this machine
Expected Behavior
I expected the game to run and not exit with
Steps to Reproduce
Run Tekken TT2 on a F39 system with a AMD Radeon RX 7900 XTX card.
System Info (Optional)
OS: Fedora 39
GPU: AMD Radeon RX 7900 XTX (RADV NAVI31)
Emulation Settings (Optional)
No response
Logs (Optional)
Here's the full error log
This is a regression in the mesa driver and also affects vulkan games
https://github.com/ValveSoftware/csgo-osx-linux/issues/3656
@Squall-Leonhart I'm not sure that csgo issue lines up with that I'm experiencing with
Cemu
I don't have any issues running any Steam games with with running vulkan utilities....
Here's some additional information on this issue
It appears to be related to this line:
https://gitlab.freedesktop.org/mesa/mesa/-/blob/23.3/src/amd/vulkan/meta/radv_meta_clear.c?ref_type=heads#L325
Can you make sure you're on the amdgpu driver and not radeonsi
I filed a ticket with Mesa -- it has a bit more information in it.
Like the output of
inxi -GSC -xx
https://gitlab.freedesktop.org/mesa/mesa/-/issues/10677
It's definitely using
amdgpu
and notradeonsi
nvm, i see you are.
Valgrind output would be really beneficial here, I suspect a local variable might be getting freed to early though, as was the case with a similar Godot crash in 2020
You can find the valgrind output here:
https://leifliddy.com/.steam/valgrind.log
It never got past the initial loading screen -- valgrind stopped after logging more then 10000000 errors.
I might try again tomorrow and remove the error limit cutoff. \
Also, I want to try and get rid of this error:
If I don't strip the Cemu binary (after building it) -- can it get the debug info from the binary itself?
Looks like I'm not the only one experiencing this issue
https://www.reddit.com/r/EmuDeck/comments/x7g2lj/tekken_tag_2_through_cemu_crashes_on_character/
I have a better backtrace -- by using a non-stripped type=debug binary
Redoing the valgrind log with that debug build will get significantly better output.
Ok, I redid the valgrind log with the debug binary:
https://leifliddy.com/.cemu/valgrind.log
I also installed the
vulkan-validation-layers
package and setexport VK_INSTANCE_LAYERS=VK_LAYER_KHRONOS_validation
That log can be found here:
https://leifliddy.com/.cemu/cemu.vulkan.validation.log
The Mesa guys are that saying based on the validation log I just posted -- that it looks like it's an application bug.
FWIW, it was me that commented on the Mesa issue, but i'm no Mesa developer myself. However, this VVL error does make me think that this might be a CEMU bug:
Yeah, this looks like something the windows drivers just tolerate despite being a violation, many of the valgrind errors are luckily just libhid jank and some benign mutex stuff
I should have mentioned that I could never got passed the initial loading screen with
valgrind
(after running for several hours) so it never got to the point where the error occurred.Would using less options help?
I was using
No more info is needed, thanks. I will fix this when I get a chance.