Performance issue with DXGI swapchain and recent NVIDIA drivers in combination with unlocked UI thread (0.8.144 and later)

[abaire]

Bug Description

Forking the nvidia-specific workaround from #2790.

Initially I found that this setting affected my machine on 0.8.133 and later (I did not test earlier builds). After some time and various debugging attempts, I can no longer reproduce this problem myself even when forcing DXGI swapchain use.

@aitorciki discovered that setting the Vulkan/OpenGL present method to Prefer layered on DXGI Swapchain in the NVIDIA control panel causes guest framerate to be roughly halved when using the Vulkan renderer. In response, xemu was updated to forcibly disable the DXGI swapchain option using the nvapi. Unfortunately this is problematic for certain situations with HDR enabled.

@Toff-kun found that using the undocumented Vulkan/OpenGL Present Method - Flags (OGL_DX_PRESENT_DEBUG_ID, 0x20324987 from Orbmu2k/nvidiaProfileInspector ) and setting the value 0x000802A5 Treat DXVK as Native allows use of DXGI without perf degradation.

Both were able to verify that this does not affect 0.8.133 for them, and using a debug build that adds back the 0.8.133 UI-loop 60fps throttle fixes the issue regardless of the state of the present method settings in the NV profile. A general solution for #2790 will likely solve this issue as well.

Expected Behavior

Users should be able to enable DXGI without having to set up undocumented flags via additional configuration tooling. While arguably this should be fixed in the NV driver directly, xemu already performs NVIDIA-specific configuration and can hopefully set the DXGI flag on the user's behalf.

xemu Version

Affects at least 0.8.133 and 0.8.134, presumably tied to contemporary (late 2025/early 2026) NVIDIA drivers.

System Information

Windows 11, possibly other Windows versions

Additional Context

No response

[abaire]

Hey @aitorciki and @Toff-kun - thanks again for your help tracking this down so far. Unfortunately I've lost the ability to reproduce the original bug; even with my initial fix to force DXGI off removed, I only see framerate halving on the official release builds, not my dev build.

I'm hoping you can help me narrow down exactly what flags resolve the issue on your machines. To that end, I've created a number of builds that set various permutations of the flag bits, linked below. The first build in the list explicitly does not set any flags, and will hopefully have poor performance again.

Could you try each build in order and let me know which ones have the perf issues and which do not? Please make sure that your NVIDIA profile enables DXGI and ideally that you're not setting the Vulkan/OpenGL Present Method - Flags setting via nvidiaProfileInspector or some other means.

• Allow DXGI but do not set any flags: https://github.com/abaire/xemu/actions/runs/25325986486
• Set just OGL_DX_PRESENT_DEBUG_ALLOW_DXVK_PROMOTION: https://github.com/abaire/xemu/actions/runs/25326741545
• ... with OGL_DX_PRESENT_DEBUG_ENABLE_DFLIP_ALWAYS: https://github.com/abaire/xemu/actions/runs/25326827183
• ... with OGL_DX_PRESENT_DEBUG_DISABLE_FULLSCREEN_OPT: https://github.com/abaire/xemu/actions/runs/25326961196
• ... with 0x00000200: https://github.com/abaire/xemu/actions/runs/25326989269
• Using the magic 0x000802A5 value: https://github.com/abaire/xemu/actions/runs/25326886775

Sorry that it's a bunch of things to test; I want to minimize use of magic undocumented values while still resolving this performance issue for other less technically savvy users.

[Toff-kun]

I tested all of those builds blindly, all had poor performance, including the "magic" one. So I checked Nvidia Profile Inspector, and the 0x000802A5 flag wasn't set. I then tried the Canary build I had that force-disabled DXGI - the Present method was correctly changed to Prefer native. Then I tried the builds from the above post again, but the Present method was still set to Prefer native.
It seems like none of the above builds set the Nvidia profile flags correctly, at least on my end.
Maybe it requires deleting the xemu driver profile entirely which I'm hesitant to try.

[aitorciki]

Forking the nvidia-specific issues from #2790 since they are reproducible in 0.8.133 and not related to the 134 perf regression affecting AMD devices.

In my case the DXGI perf issues started with 0.8.134, 0.8.133 and prior perform as expected when forcing DXGI in the NVIDIA driver profile.

Will test the builds asap and share my results.

[abaire]

Thanks for checking!

I'm not entirely confident that Vulkan/OpenGL Present Method - Flags maps to 0x20324987 (though it appears to be the case from the nvidiaProfileInspector source.

@Toff-kun - can you point me at where you found the 0x000802A5 magic value and link me to whatever tool you're using to set it? Hopefully I can double check how I'm setting the value via the API against the source for that version of the utility.

In my case the DXGI perf issues started with 0.8.134, 0.8.133 and prior perform as expected when forcing DXGI in the NVIDIA driver profile.

@aitorciki are you certain that it didn't affect 0.8.133? When you first reported that disabling DXGI fixed the issue I immediately checked on 133 and found that forcing the mode to native had a significant affect on the framerate tester results. I think if forcing the flag value fixes it in 144+ it doesn't really matter, but it would be strange for us to see very different results with 133.

[Toff-kun]

@Toff-kun - can you point me at where you found the 0x000802A5 magic value and link me to whatever tool you're using to set it? Hopefully I can double check how I'm setting the value via the API against the source for that version of the utility.

I just experimented with the latest Nvidia Profile Inspector pre-release build trying out different flags from its dropdown list, found that this particular one fixes performance while others don't seem to affect anything.

[aitorciki]

@aitorciki are you certain that it didn't affect 0.8.133?

Yes, I've just re-tested. Here are some recordings. In all cases "Prefer layered on DXGI Swapchain" is enabled and the graphics backend is Vulkan.

0.8.133 +DXGI + "Treat DXVK as Native" NOT enabled (set to driver default):
https://www.youtube.com/watch?v=PPPcY1qy9WY

0.8.134 +DXGI + "Treat DXVK as Native" NOT enabled (set to driver default):
https://www.youtube.com/watch?v=opJ_lQJh6-8

0.8.134 +DXGI + "Treat DXVK as Native" ENABLED:
https://www.youtube.com/watch?v=R0swcGa-gt4

The main difference between 0.8.133 and 0.8.134 (other than the performance regression) is the UI resolution: while I have internal resolution scale set to 3x and the game is rendering at 4K, 0.8.133 renders the UI at 720p, and the video capture from the NVIDIA overlay is also 720p (as seen in the Youtube upload), while 0.8.134 renders the UI at 4K.

[aitorciki]

can you point me at where you found the 0x000802A5 magic value and link me to whatever tool you're using to set it? Hopefully I can double check how I'm setting the value via the API against the source for that version of the utility.

If it helps, here's how NVIDIA Profile Inspector exports the value (second ProfileSetting):

<Profile>
  <ProfileName>c:\users\aitorciki\emulators\xemu\xemu.exe</ProfileName>
  <Executeables>
    <string>c:/users/aitorciki/emulators/xemu/xemu.exe</string>
  </Executeables>
  <Settings>
    <ProfileSetting>
      <SettingNameInfo>Vulkan/OpenGL present method</SettingNameInfo>
      <SettingID>550932728</SettingID>
      <SettingValue>1</SettingValue>
      <ValueType>Dword</ValueType>
    </ProfileSetting>
    <ProfileSetting>
      <SettingNameInfo />
      <SettingID>540166535</SettingID>
      <SettingValue>524965</SettingValue>
      <ValueType>Dword</ValueType>
    </ProfileSetting>
  </Settings>
</Profile>

[aitorciki]

I have tested many of the builds, and the results are very different from 0.8.134, but I don't see clear differences between them. I have tested DGXI + no flags, OGL_DX_PRESENT_DEBUG_ALLOW_DXVK_PROMOTION and 0x000802A5. In all cases my NVIDIA profile is set to forced DXGI without any flags.

In 0.8.134, as shown in the videos I've linked before, JSRF runs at around 10fps, with very high mspf (easily over 50ms, reaching 100ms). In all the test builds, fps jump between 60-ish fps (not locked, drops a few frames) for many seconds, then 30fps for a while, then back. And mspf is pretty stable around 14/15/16ms.

As soon as I enable "Treat DXKV as Native" manually, all the builds perform like 0.8.134 with that same setting enabled: fps close to locked 60, mspf below 6ms (like in my previously linked video for that scenario).

[abaire]

Thanks!

0x20324987 is the ID I'm using when attempting to set the flag so that matches your profile @aitorciki. I'll take a closer look to see if I broke this somewhere else.

Since you're able to set this via the profile inspector, could you try a few different flag values for me to confirm if they work or not?

• We know 0x802A5 works
• 0x80000 - OGL_DX_PRESENT_DEBUG_ALLOW_DXVK_PROMOTION alone
• 0x80004 - promo + DFLIP_ALWAYS
• 0x80005 - promo + DFLIP_ALWAYS + disable fullscreen opt
• 0x800A5 - The above plus some undocumented flag(s)?
• 0x80205 - The above plus another undocumented flag?

I'm curious if there's a subset that still works or if we really need the full thing. Given that one of the flags is for fullscreen optimization, I'm also curious if setting the flag has a negative effect on fullscreen performance (esp. w/ DXGI turned off, and this might be exclusive fullscreen, which needs to be toggled in the xemu settings).

@Toff-kun can you screenshot the dropdown that ended up setting the flag to 0x802A5? When I look at the source I don't see entries for the 0x2A0 part of the mask so I'm confused where that came from.

[aitorciki]

I've tested all these. 0x80205 behaves like 0x802A5, it fixes the performance regression on my setup. The rest of them by their own have no impact and the game keeps running at 10fps.

As for where 0x802A5 comes from, this is all I've found: Orbmu2k/nvidiaProfileInspector@49c451a

[abaire]

I've tested all these. 0x80205 behaves like 0x802A5, it fixes the performance regression on my setup. The rest of them by their own have no impact and the game keeps running at 10fps.

As for where 0x802A5 comes from, this is all I've found: Orbmu2k/nvidiaProfileInspector@49c451a

Thanks, that's extremely helpful!

Can you try 0x80204 to see if that works?

[aitorciki]

Can you try 0x80204 to see if that works?

0x80204 does also work.

[abaire]

Perfect, I feel somewhat better about proposing that combination, despite not knowing what 0x200 does. Disabling a fullscreen optimization makes me nervous that it'll just break some other usecase.

I'll look into fixing the xemu-built in profile and will post a new build here for verification.