Closed HansKristian-Work closed 1 month ago
No issues related to this MR so far. Tested F1 2021 on 6.9.1-arch1-1, RX 7600 with RADV 24.0.7.
On RTX 3070 with 555.42.02:
Also on 555.42.02:
No issues apart from some glitchy shadow rendering in Deathloop, but that seems to be driver specific, unrelated to this MR.
We can't test TLOU with Nvidia on that list because it currently still crashes (Ada architecture, not sure about the others). (it's also the only game I have on the list)
I didn't have any issues when I tested TLOU1 this time, still on RTX 3070 and 555.42.02. Tested for about half an hour in a few different locations with various graphics settings, like outside the capitol building and in the slums where I know I've had a hang in the past. Maybe it was just chance, some difference in graphics settings, or maybe it's worse on Ada.
@runar-work Might be an Ada only issue, I don't know. I know for a fact I did not experience any issues with my 3060 when I had that. But I HAVE played TLOU with the 4070 without crashes too... I once thought it was fixed, then the second time I started it took ~1 minute then the game froze and I got XID 109.
For reference https://github.com/HansKristian-Work/vkd3d-proton/issues/1925
I tried to test Starfield, but the game won't launch with this MR and 555.42.02. Seems like it fails in a shader compile thread. I also launched it on my iGPU with similar results. Here's a proton log from the RTX 3070 with +vulkan and trace debug level: steam-1716740_launch_fail_MR_trace.log
I also tested Dragon's Dogma 2, which seemed to be equally glitchy with this MR and without it. I did get a Xid 109 with the MR right after the medusa fight in the beginning, but wasn't able to reproduce that afterwards. I tried to vary the graphics settings a bit, but I probably should've gone a bit lower to see if the glitchiness improved (occasional vertex explosions and such).
@runar-work Starfield should work now on updated PR.
Yeah, Starfield works fine now. Tested for a little while without issues.
Enough test coverage now, so going ahead.
Improves some behavior for strict helper lanes in subgroup ops. It turns out that RADV and NV can emit helper lanes even it makes no sense for the shader to do so, so be conservative and always consider helper lanes potentially active in fragment shaders.
Also consider helper lane behavior in SPV_NV_subgroup_partitioned.
This affects various games, so this needs some smoke testing on various games:
Should be tested on NV and RADV.