Dxvk: Resident Evil 2 remake is too dark with DXVK 1.0

[SergeyLatyshev]

RESIDENT EVIL 2 / BIOHAZARD RE:2 (883710) worked fine with DXVK 0.96 but with DXVK 1.0 the image is very, very dark in-game. It reverts back to normal if I downgrade DXVK.

Software information

RESIDENT EVIL 2 / BIOHAZARD RE:2
https://store.steampowered.com/app/883710/

System information

• GPU: Nvidia GTX 1060 6GB
• Driver: 415.27
• Wine version: Proton 3.16-8 beta
• DXVK version: 1.0

Log files

steam-883710.zip

[doitsujin]

Still works fine on my RX 480. Please bisect the issue if you think this is a DXVK regression.

Please keep in mind that there is a game bug causing the image to be too dark on some systems when running above 60 FPS.

[SergeyLatyshev]

I don't have 60 FPS. What is "bisect the issue"?

[doitsujin]

https://git-scm.com/docs/git-bisect

In any case, I cannot reproduce the problem.

[SergeyLatyshev]

So, what should I do with 'git bisect'?

[L-as]

You need to bisect it and test different versions of DXVK.

[SergeyLatyshev]

So, I have to compile them? I tried to compile DXVK once and failed. Are there compiled builds between 0.96 and 1.0 available? I would test them.

[Riesi]

Are you using the dxvk version that is shipped with proton or have you tried dxvk by placing it next to the game? Just to eliminate proton in this equation.

[SergeyLatyshev]

Both. It started glitching when I tried DXVK 1.0 almost two weeks ago. I had Proton 3.16-6 or 3.16-7 back then. I've been using 0.96 since that. But now Proton got updated, got DXVK 1.0, and that happened again. If I replace Proton's DXVK with v0.96, the game works fine.

[doitsujin]

You could try setting d3d11.relaxedBarriers = False in the config file.

Although that doesn't really explain why it works fine with 1.0 for most people.

[Tk-Glitch]

I have bisected the issue and found https://github.com/doitsujin/dxvk/commit/915091b76b2de06d04a4de0b050001b8e7c88169 to be the breaking point. Reverting it fixes the issue on both RE2 and DMC5.

[doitsujin]

Should be fixed as of 032625882990ae843cbdcd526f3c72087ab31888.

I don't know why using shift instructions (even arithmetic shifts, which should be equal to signed division) breaks on Nvidia, but it does, and this commit works around the issue. Thanks to @Tk-Glitch for testing.

[mindinsomnia]

I can confirm I have the same issue as reported on a NVIDIA GTX 980 with latest Proton.

[doitsujin]

Yes, everyone on Nvidia has this issue, just read the comment above for a fix.

[jeffbolznv]

I've identified the compiler bug, and it is already fixed in the most recent Vulkan developer driver at https://developer.nvidia.com/vulkan-driver. It should be fixed in our next mainline driver release as well.

[Lastique]

arithmetic shifts, which should be equal to signed division

Arithmetic shift of negatives is not equal to division. Division rounds to zero, shift rounds to negative infinity.

[doitsujin]

The numbers in question are guaranteed to be a multiple of 4 (and technically also guaranteed to be non-negative).

[ValZapod]

@jeffbolznv I do not think you fixed it (in windows). See https://www.nvidia.com/en-us/geforce/forums/game-ready-drivers/13/333239/devil-may-cry-vresident-evil-2-biohazard-re2-brig/2341397/

Really, you should not play with llvm codebase if you do not know what you are doing ;)
In devil may cry 5:

[doitsujin]

Wasn't this found to be a game bug when running at low frame rates anyway?

[ValZapod]

Low? Is 100 Hz low? I do not think so. And I have 2080 Ti, so you know, framerate is good. Also, it is only in Dirèctx 11 (directx 12 is good) and I am pretty sure it shader compiler nvidia bug...

[doitsujin]

Well I'm sorry that I wasn't able to guess your hardware and framerate based on absolutely no information whatsoever...

[ValZapod]

@doitsujin You are only working with Linux Directx, right)) I was asking that to Nvidia guy because bug is in Windows...

[jeffbolznv]

The bugfix I made was only in the Vulkan SPIR-V compiler and had something to do with shared memory ordering. It's possible the DX compiler could have a similar issue, but it would be in a different part of the code base. But it's also likely it could be something completely different.

[doitsujin]

Was that a genuine driver bug though? The game is known to rely on undefined behaviour wrt workgroup shared memory, which has also caued issues with RADV in the past.

[jeffbolznv]

I don't actually know. The change I made was for a different app, where it was fixing a legitimate compiler bug, and the same fix also fixed this bug. So we never debugged this particular shader in detail.