GPU hang in Marvel Rivals with Lumen GI on RDNA1

[runar-work]

As first reported here: ValveSoftware/Proton#8300 (comment). It takes me 1-3 minutes to reproduce in the Tokyo 2099 Spider-Islands map if I just start a custom game and run around in the starting area with the trees and lanterns.

I couldn't get it to hang with RADV_DEBUG=hang or syncshaders, and I didn't get any active waves from the hang with GPU recovery disabled.

I was able to get some breadcrumbs. The crash region contains 2-4 shaders with RADV, always two global illumination shaders and sometimes a pair of MainVS/MainPS shaders from the preceding draw calls. The two shader hashes that always appear are 70594334b6b0d4a7 (MarkRadianceProbesUsedByTranslucencyVolumeCS) and aec033de19e91d96 (MarkRadianceProbesUsedByScreenProbesCS).

It also hangs with AMDVLK. That adds some additional compute shader hashes to the breadcrumbs trace, but it's all related to the GI pass.

I didn't have any hangs when I tested with SSGI, so it seems specific to Lumen GI. I also didn't have any issues when I tested on RX 7600 with the same settings and VKD3D_CONFIG=nodxr or RADV_PERFTEST=emulate_rt.

Software information

Marvel Rivals, Lumen GI enabled (graphics preset high).

System information

• GPU: RX 5700 XT
• Driver: AMDVLK and Mesa (24.3.1 and main)
• Wine version: Proton Experimental bleeding-edge
• VKD3D-Proton version: 0e68113 from Proton 8 up to the recent c965c13

Log files

Proton logs with breadcrumbs and shaders: Marvel_Rivals_RDNA1_hang_logs.zip