protonscr

Fallout 3 sporadic big white dots on MSAA since v3.0

dxvkclosed d3d9
doitsujin/dxvk#5758 · opened 2026-06-30 by FranTesla · updated 2026-07-08 · 20 comments · github
FFranTesla 2026-06-30 github

Fallout 3

  • Driver: 582.66
  • DXVK version: 3.0 vs 2.7.1

Sporadic big white random dots apear with MSAA using DXVK 3.0 but that doesn't happen with 2.7.1
default settings, nvidia profile deleted to test, no change.
This happen in the past with early versions, but something changed from 2.7 to 3.0 that this happens again.

Screenshots of MSAA 4X, is even more noticiable with 8X and less with 2X.
Image

Image
KK0bin maintainer 2026-06-30 github

Please make an apitrace.

Driver: 582.66

Also for the record, I recognize the Nvidia driver version but you should really mention what GPU you have too...

BBlisto91 2026-06-30 github

If this is the same issue as in New Vegas then it is technically a game bug that i have also reproduced on native Windows AMD in the past. Nvidia works around it with a game profile.

@K0bin As a reminder https://github.com/doitsujin/dxvk/issues/2319#issuecomment-1283072644

WWallSoGB 2026-07-01 github

I've been noticing this issue as well.

I fixed the decal issue talked about here, some time ago, as it was occurring in D3D9 on Intel and AMD.

This one however is new and seems to stem from normal object shaders.

Ddoitsujin maintainer 2026-07-01 github

Need an apitrace of sorts to actually be able to look into it.

Ddoitsujin maintainer 2026-07-01 github

nvm, got a repro. Should be fixed with https://github.com/doitsujin/dxvk/pull/5761.

AALTVRA 2026-07-08 github

Still getting a very similar issue with bright, blurry dots / bright lines in distant LOD (in Nuclear Sunset modlist for TTW mind you-- in Capital Wasteland) on DXVK 3.0.1. Issue isn't present in 2.7.1.

Image Image Image

AMD Radeon RX 6600
Driver Version: 26.6.2

BBlisto91 2026-07-08 github

If it's somewhat easy to reproduce can you make a apitrace of where it usually occurs?

AALTVRA 2026-07-08 github

I have made the apitrace. Whatever is happening with LOD should be present later in. I loaded into the game save from the first screenshot, used TCL to go towards and away from the LOD issue shown in that screenshot, and then ALT + F4'd out of the game after about 15 seconds of doing that. The issue wasn't present during the apitrace (which I read is normal when trying to reproduce the issue with the new d3d9, etc). Here is the apitrace:

https://mega.nz/file/lv01iBTI#tKAigDjKByUdWaBbFN6pIvLplm7AEI7zHDM-m68kuic

For further information, I am using the Nuclear Sunset modlist for TTW with the Hikikomori Performance preset. This also seems to be occurring with only a few structures in the game, and only in their distant LOD form, this is one of those structures:

Image

I have told the developers of the modlist about the issue as well. It does not occur on DXVK 2.7.1 and before, and so does not appear to be a problem with the mods themselves.

Ddoitsujin maintainer 2026-07-08 github

Yay, more float emulation memes. Doesn't happen on any reasonable driver of course.

Also completely unrelated to the original issue.

Ddoitsujin maintainer 2026-07-08 github

Can't even debug this properly because AMD's driver just crashes as soon as I try to look into this further.

Just throw d3d9.floatEmulation=Strict at it and move on I guess.

AALTVRA 2026-07-08 github

Can't even debug this properly because AMD's driver just crashes as soon as I try to look into this further.

Just throw d3d9.floatEmulation=Strict at it and move on I guess.

I've had d3d9.floatEmulation set to Strict for some time, because of another interesting bug in Nuclear Sunset with rain looking like this:

Image

But this issue with some distant LOD being really bright spots still happens with that set to Strict.

KK0bin maintainer 2026-07-08 github

Please also make an apitrace of the rain.

AALTVRA 2026-07-08 github

Please also make an apitrace of the rain.

I haven't personally been able to get this to happen again since when I first set floatEmulation to Strict, and even when floatEmulation had been set to Auto, I only saw this once across maybe ~20+ hours IGT on the modlist. I'll try to the next time it crops up though.

Ddoitsujin maintainer 2026-07-08 github

I've had d3d9.floatEmulation set to Strict for some time, because of another interesting bug in Nuclear Sunset with rain looking like this:

Please post a log to confirm that it actually applies, because it just doesn't reproduce with Strict here (but does with True even on RADV, so I guess that makes it more debuggable).

I swear to god 90% of the post-3.0 work has been wasting time on broken Windows drivers.

AALTVRA 2026-07-08 github

I've had d3d9.floatEmulation set to Strict for some time, because of another interesting bug in Nuclear Sunset with rain looking like this:

Please post a log to confirm that it actually applies, because it just doesn't reproduce with Strict here (but does with True even on RADV, so I guess that makes it more debuggable).

I swear to god 90% of the post-3.0 work has been wasting time on broken Windows drivers.

If you can point me to what sort of log I need to produce, I can do that now. Sorry, not at all knowledgeable with this stuff.

Ddoitsujin maintainer 2026-07-08 github

There should be a FalloutNV_d3d9.log (or whatever the game exe is called) somewhere (usually next to the game exe itself).

AALTVRA 2026-07-08 github

There should be a FalloutNV_d3d9.log (or whatever the game exe is called) somewhere (usually next to the game exe itself).

I checked the effective configuration in there and realised that the freaking d3d9.floatEmulation = Strict has been commented out the entire goddamn time in dxvk.conf... Everything normal in game with that now active. Apologies for wasting your time here.

Ddoitsujin maintainer 2026-07-08 github

well i mean there's still an actual regression here somewhere given that 2.7.1 works fine without strict (and strict is pretty bad for perf because AMD chose not to implement our fmulz pattern despite the hardware natively supporting it), but yeah.

AALTVRA 2026-07-08 github

Yeah I'm thinking I'm probably gonna stay away from AMD for my next PC upgrade...

Ddoitsujin maintainer 2026-07-08 github

ok, turns out it's not really a bug after all, just slightly different handling on vector normalization where the vector input is already NaN (even on the working build). This really needs strict float emu.

Upstream links