Reflex seems to not work in dx12 games making terrible input lag under 100% GPU.

[1Laggy1]

Software information

• Name of the application or game: THE FINALS, Cyberpunk 2077

System information

• GPU: RTX 3070
• Driver: 595.45.04
• Launcher (e.g Proton or Bottles): Proton
• Proton/Wine version: 10.0-4

Log files

nvapi64.log

steam-1091500.log

I changed my resolution to 4k to get maximum gpu bottleneck Using smartphone 480 fps camera, I have recorded my mouse click, and monitor. Using Openshot I've calculated how much time is spent from clicking, and character reacting in game.

The Measurements (Average of 5 clicks per scenario): I tested both DX11 and DX12 to see if the translation layer makes a difference.

DX11 Game (Overwatch 2 via DXVK) - 100% GPU Load:

Reflex OFF: ~68 ms

Reflex ON: ~11 ms (Reflex works perfectly here, completely bypassing the render queue).

2. DX12 Game (The Finals via VKD3D) - 100% GPU Load:

Reflex OFF: ~70 ms

Reflex ON: ~70 ms (Reflex seems to do absolutely nothing here).

Capped FPS IN GAME (GPU at ~85%): ~15 ms (The only way to get a responsive mouse in this game)

Cyberpunk 2077 seems to have the same problem.

This warning in proton log could be related?

22828.824:050c:0550:warn:vkd3d-proton:d3d12_low_latency_device_SetLatencyMarker: PRESENT_START_NV is non-monotonic 16470000 < 16480000.
Fossilize WARN: Graphics pipeline handle 0x00007f9550460410 is not registered.

[Saancreed]

First things first, please try again with Proton Experimental, 10-4 is too old/too stable to have this fix specifically targetting THE FINALS: HansKristian-Work/vkd3d-proton#2804

On the other hand, Cyberpunk 2077 in general has some issues around its Streamline integration, so it wouldn't surprise me if they affected Reflex as well. While they were most visible with framegen enabled, maybe the solution from #245 (comment) will help on your Ampere too.

That said, I've noticed that your CP log comes from cachyos-10.0-20260228-native, which should be fine for THE FINALS. Are you perhaps using different distributions of Proton per game?

[1Laggy1]

Thanks for the quick reply!

Just to clarify, I actually tested several Proton versions (including cachyos-native) before reporting, but I specifically tested 10.0-4 just to provide a stable baseline.

Following your advice, I just re-tested THE FINALS using the latest Proton Experimental, but unfortunately, the result is exactly the same: I'm still getting ~70 ms of latency at 100% GPU load with Reflex ON. The render queue is still not being bypassed.

I only included Cyberpunk 2077 to see if this was a universal DX12 issue. Since Cyberpunk has its own specific quirks with the Streamline integration, I am more than happy to drop it and focus entirely on debugging THE FINALS, as this is the game where I originally noticed the massive input lag or test other dx12 games if you will.

Here is the new log from THE FINALS using Proton Experimental

steam-2073850.zip

nvapi64.log

[Saancreed]

Hm, I see.

Unfortunately, debugging Reflex issues is always quite an adventure because it ultimately plugs into Nvidia's proprietary driver, and I don't have the equipment available to verify exactly how well it works on my own machine (other than looking at the logs and seeing how much time passes in each latency sleep call). I'll try to think of something that maybe will make issues like that easier to pinpoint in the future, but for now I have two requests for you:

1. Try running https://github.com/NVIDIA-RTX/RTXPT/releases/tag/v.1.7.0 and see if Reflex works there. You'll have to enable it in the dropdown menu first, but you should hopefully also see a lower-level dump of Reflex latency stats there.
2. Gather logs from the sample above and again from THE FINALS but with DXVK_NVAPI_LOG_LEVEL=trace set in the environment. This should let us see the timings, and whether the latency sleep calls actually insert necessary delays.

[1Laggy1]

I tested the RTXPT demo you provided using Proton Experimental

Result: Reflex definitely works in the RTXPT demo. The responsiveness drastically improves, which I also verified with the high-speed camera:

Reflex off: ~77 ms

Reflex on: ~16ms

I have collected the DXVK_NVAPI_LOG_LEVEL=trace logs for both applications under 100% GPU load with Reflex ON. Hopefully, comparing the timing stats from the working demo against the broken game will help pinpoint why the sleep calls are failing.

Logs attached below:
steam-.zip

[Saancreed]

Thank you, this at least confirms that there's no systemwide nvapi/vkd3d-proton/driver Reflex breakage in d3d12 applications on your setup if the demo works fine with it. Apparently only some games are cursed, and you managed to run into a few of them.

😩

In your RTXPT log I can see calls to latency sleep sleeping correctly. The last two made are:

607085:17637.170:0138:013c:trace:nvapi64:NvAPI_D3D_Sleep (ptr=0x3560080)
607086:17637.198:0138:013c:info:nvapi64:<-NvAPI_D3D_Sleep: OK
607139:17637.201:0138:013c:trace:nvapi64:NvAPI_D3D_Sleep (ptr=0x3560080)
607140:17637.230:0138:013c:info:nvapi64:<-NvAPI_D3D_Sleep: OK

… which is roughly 28-29 ms spent in each of them, which is great. On the other hand, THE FINALS' calls to latency sleep work for around the first 16 rendered frames, sometimes reaching 7 ms of sleep per frame (at which point FPS are limited to 100), then it breaks and future calls finish in less than 1 ms each. At a glance, latency markers look fine and having multiple frames in flight shouldn't confuse the driver… At least these calls return success, so we know they arrive to vkd3d-proton, which is the next stop where we can get booted off the bus, so to speak.

I have one more request: run THE FINALS with vkd3d-proton from here: https://github.com/Saancreed/vkd3d-proton/actions/runs/23317558956 (or compile your own with changes from Saancreed/vkd3d-proton@56137e3 applied), dropping 64-bit d3d12core.dll into files/lib/wine/vkd3d-proton/x86_64-windows subdirectory of your Proton installation, then get me another Proton log with DXVK_NVAPI_LOG_LEVEL=trace.

This should give us an idea of whether vkd3d-proton is passing the calls correctly to the NV Vulkan driver or internally dropping them. And once we know that, I'll be able to forward this to appropriate people to diagnose it further, because myself I have little idea what's happening, and why 😅

[1Laggy1]

Very interesting. I can confirm that first fps limit was set to 100, I then changed it to unlimited. In the new log fps limit was set to unlimited from the start.

I hope I installed your d3d12core.dll correctly.

Here is a new log:

steam-2073850.zip

[Saancreed]

Thanks. Yes, you installed it correctly, the reported revision of vkd3d-proton changed to my commit hash. The lack of any new log messages leads me to believe that at least latency sleep calls (and probably everything else too) does arrive to the driver, so I guess it's up to NV to figure this out now.

@wcampbell-nv You might be interested in this ☝️

[netborg-afps]

Hello, I've just tested input lag in the Finals via my monitor's built-in Reflex Analyzer.
Using plasma-x11-session, no AA ingame, high settings at 1080p.

vkd3d + reflex on: 19 ms at 195 fps
vkd3d + reflex off: 22 ms at 195 fps
dxvk-low-latency: 14 ms at 235 fps (+ shader compiler stutters)

Also subjectively, I consider the mouse input via vkd3d+Reflex currently unplayable in this game (I've not tested any others).
Reflex was working well though when I tested it via Proton+Vulkan in Deadlock.

[1Laggy1]

Hello, thank you very much for testing.

However, to truly understand whether Reflex is not working on your side as well, you need to have a much lower FPS.

With my configuration, I intentionally set a higher resolution then my monitor, in my case, 4K resolution gives me 60 FPS, and that’s where I experienced a 70 ms delay

[netborg-afps]

The 70 ms you get suggests that something is broken on top of what's already not working correctly. My GPU was definitely maxed out during the tests.

I suggest you can try out PROTON_ENABLE_WAYLAND=1 (supported in proton-GE and proton-cachyos) because xWayland is known to buffer frames. It usually does that when the frame rate is higher than your display refresh. 3 frames at 16.6 ms each would be ~50 ms plus,

[wcampbell-nv]

I gave this a try myself and am having trouble reproducing. Reflex successfully reduces latency on my system in both The Finals and Cyberpunk 2077. I measured latency from the SIMULATION_START marker submission to the corresponding QUEUE_OPERATIONS_END timestamp produced by VK_EXT_present_timing, with The Finals running GPU bound at 60fps:

Reflex off: 64ms latency
Reflex on: 26ms latency

I did however reproduce similar symptoms by launching the game with DLSS-FG enabled, and then turning off frame generation after entering gameplay. In this case Reflex did not improve latency. I was stuck at 64ms latency whether or not Reflex was enabled. This is probably a separate issue from what you describe, but I'll investigate why this is.

I suggest you can try out PROTON_ENABLE_WAYLAND=1 (supported in proton-GE and proton-cachyos) because xWayland is known to buffer frames.

Reflex actually does not currently work in native Wayland applications due to a driver issue. There's a chance the Wayland backend could improve latency over XWayland+Reflex off, but it still won't be as good as correctly-working Reflex. We have a fix for this in development.

[JustCauseWhyNot]

I'm getting this same issue but not in all games. In cyberpunk 2077 reflex is working correctly reducing latency by 5ms when 100% gpu bound from 20 to 15ms. This was tested using a osltt device. But in ready or not dx12 reflex didn't work in reducing the latency even though in dx11 reflex was working for ready or not with pretty much the same latency reduction as cyberpunk 2077 20ms -> 15ms.