DCS has a performance problem with OpenXR and HP Reverb G2!

[zerO_crash]

After having issues with DCS in VR since 2.7, and having planned to upgrade from Windows 10 Pro to Windows 11 Pro, I finally decided to migrate. I have reinstalled a clean installation of Windows 11, with the most recent updates as per 30.03.2023. I have all drivers updated to the most recent revisions. I have went through BIOS and system, and changed everything to performance (in case of the CPU, an XMP-overclock - 4.7GHz). HAGS is disabled, Nvidia has "maximum performance" and "Shader buffer - 10GB" set, and more. All settings are set globally, so everything affects DCS as well. Here are my observations from the current MT-version:

 

Using my settings (mix of highest and lower for best framerate and visuals) in 2D (non-VR), I get 80-90fps in the most complex cockpits. In F2-view, I get anywhere from 110-160fps, sometimes more. My CPU-load is appropriate, my GPU is maxed out. Using the framerate-checker tool within DCS, I get overall really good performance-metrics, and a message that I´m GPU-bound (i.e. GPU is the limiting factor of my system). So far, so good.

 

When connecting the HP Reverb G2, and enabling the VR-option in DCS, it´s a whole different story. While everything in the main menu works smooth, as soon as I start an empty mission, I get problems right away. During the first 2-3 seconds (while the textures load, and shortly thereafter), I get high CPU-/GPU-load, with low frametimes on the GPU and good fps. However, as soon as those first few seconds pass, my CPU-load goes down, and my GPU-frametimes surge up to +-50 microseconds. This happens instantaneously! I have tried running DCS MT as a standalone, with native OpenXR on HP Reverb G2. I have also added the DCS-shortcut to the Steam library, and run it through SteamVR. There are no differences, I am down to 18-20fps! (SteamVR & OpenXR tweaked for performance).

 

From my observations then, I infer the following:

- The fault is not at my end, as 2D gives me superb performance.

- The problem occurs when DCS is put in VR, and attempted run with a HP Reverb G2 (could affect other HMD´s as well, not sure).

- A clan member of mine has an Nvidia 1080 (non-ti, weaker than mine), a 7th generation intel CPU (one generation older than mine) and the original HTC Vive (as opposed to my HP Reverb G2). He has splendid performance, better than he had before. For me, it´s the exact opposite.

 

As such, I am at this point sure, that the problem is in your integration of the OpenXR. This is the one constant I cannot remove from the equation, so as to eliminate it as a probable cause. A HP Reverb G2 cannot be run without OpenXR. I can change between WMR and SteamVR, but OpenXR will always be used by HP Reverb G2. I have checked every setting within DCS individually, I have made sure that my Windows 11 Pro is geared for maximum performance. The team must check the OpenXR-integration, as that is where the problem lies.

 

(At first, I thought it might be a general DCS VR issue, however as mentioned, my friend with a similar setup, but HTC Vive, runs DCS silkysmooth. He also doesn´t get the "CPU-bound" flickering-message. Everything points clearly at OpenXR from stand, and that is having tested it with two different operating systems, a multitude of drivers, 2D vs VR, etc...)

 

@Flappie @BIGNEWY @NineLine

[edmuss]

The problem is with your VR configuration somewhere.  Reverb G2 here and I get between 9 and 16ms frametimes, map and module dependant.

More information on your hardware, configuration and VR setup is needed
 

[Eugel]

6 minutes ago, zerO_crash said:

- A clan member of mine has an Nvidia 1080 (non-ti, weaker than mine), a 7th generation intel CPU (one generation older than mine) and the original HTC Vive (as opposed to my HP Reverb G2). He has splendid performance, better than he had before. For me, it´s the exact opposite.

You can´t really compare the Original Vive to a G2. That´s like saying "My friend runs the game fine in 720p, but on my 4K monitor it stutters, so it has to be a 2D Problem"

[zerO_crash]

Ok, this is interesting. Especially since I get superb 2D performance. It´s possible the issue is not a persistent one then, but still there.

 

My system:

- Asus ROG Z370-E

- Intel i7-8700k (OC´ed to stable 4.7GHz)

- Nvidia 1080Ti OC (Evga OC3)

- 64GB G.skill 3200MHz RAM

- M.2 SSD Samsung 980 Pro Evo

- XFX 850W PSU

- Secondary M.2 Samsung 980 Pro Evo (only for DCS)

 

As to the map and module, I´m doing the simplest test, Ka-50 BSIII at Senaki-Kolkhi. Clouds, weather-effects, otherwise empty.

 

Again, great performance in 2D, however terrible in VR. Also, that CPU-bound flickering bug (I presume).

 

7 minutes ago, Eugel said:

You can´t really compare the Original Vive to a G2. That´s like saying "My friend runs the game fine in 720p, but on my 4K monitor it stutters, so it has to be a 2D Problem"

I am aware of that, but as a supplement to a previous thread, I have lowered the projected resolution and I barely get a difference in framerate. Furthermore, as observed, the difference between headsets that natively use WMR + OpenXR, vs ones that don´t. That´s the point of the comparison, not the resolution.

Edited March 30, 2023 by zerO_crash

[edmuss]

Unfortunately your GPU is in no way powerful enough to run a G2 at anything other than absolute minimum resolution and settings.
For reference, at 100% resolution, the G2 requires around 19.5 million pixels to be rendered, in comparison a 4k monitor is only 8.29 million pixels; the HTC vive need 2.6 million pixels.

I run the G2 at 75% resolution (give or take 2400x2400 per eye) and with the 3080ti I get great performance, in 1080p 2D I get 200+ fps with my normal VR settings.

Unfortunately you simply do not have a big enough machine for the G2, even if you were to upgrade the GPU to a 3080 or above you would then be CPU bottlenecked pretty quickly (although MT is going some way to alleviate this).

I would suggest the following to attempt to use the G2 with your hardware: -

• Reduce resolution to around 75%
• Enable FSR or NIS upscaling provided by openxr toolkit.
• Enable foveated rendering.
• Enable reprojection with an unlocked framerate (default setting - this will allow openxr to reproject down to 30hz without much hassle).
• Disable shadows.
• Textures you can probably run high with the 1080ti as it has good bandwidth.

This should give you a resonable base on which you can tweak settings but it's not going to look great I'm afraid

edit: oh and BS3 is a performance hog no matter the settings.

 

Edited March 30, 2023 by edmuss

[zerO_crash]

14 minutes ago, edmuss said:

Unfortunately your GPU is in no way powerful enough to run a G2 at anything other than absolute minimum resolution and settings.
For reference, at 100% resolution, the G2 requires areound 19.5 million pixels to be rendered, in comparison a 4k monitor is only 8.29 million pixels; the HTC vive need 2.6 million pixels.

I run the G2 at 75% resolution (give or take 2400x2400 per eye) and with the 3080ti I get great performance, in 1080p 2D I get 200+ fps with my normal VR settings.

Unfortunately you simply do not have a big enough machine for the G2, even if you were to upgrade the GPU to a 3080 or above you would then be CPU bottlenecked pretty quickly (although MT is going some way to alleviate this).

I would suggest the following to attempt to use the G2 with your hardware: -

• Reduce resolution to around 75%
• Enable FSR or NIS upscaling provided by openxr toolkit.
• Enable foveated rendering.
• Enable reprojection with an unlocked framerate (default setting - this will allow openxr to reproject down to 30hz without much hassle).
• Disable shadows.
• Textures you can probably run high with the 1080ti as it has good bandwidth.

This should give you a resonable base on which you can tweak settings but it's not going to look great I'm afraid

 

 

Before the later iterations of 2.7, I was running the same settings at framerates between 50-75fps. I will be getting a new pc in the nearest future, however there is no explanation for why the performance has degraded in this manner. Again, in the initial couple of seconds, I get really good performance, that until the CPU bug kicks in and throttles me back. This only happens in VR. In 2D, the CPU works as intended. Due to my build, I often get better framerate than many people with with newer pc´s and unoptimized systems. There is a problem with the VR implementation.

 

As to resolution, I am capable, and have been running the HP Reverb G2 at "native" resolution. That is 2160x2160 (approximately 50%). You are upscaling the default resolution. That is not needed. You are better off with native resolution and add 2 or 4 times MSAA. Trust me, you will have much better visuals, without jagged edges. 

 

Mind you, I used SteamVR up until now, which is not the most efficient, yet has yielded great results.

Edited March 30, 2023 by zerO_crash

[edmuss]

2.8 introduced a lot of additional lighting which caused a large performance hit, especially to the ka50.  I too see the lower GPU frametimes whilst the textures are loading, typically I'll be at around 8-9ms which then jumps up to 12-14 in the ka50.

I'm not upscaling by running more than 2160, full render resolution as defined by HP on the G2 is around 3160x3100 per eye, this accounts for lense distortion. I run 75% resolution with 2xMSAA, it gives less aliasing and better performance than 100% without MSAA but the image is not as crisp; I find that if I drop the resolution much more then the reaibility of the MFDs starts to suffer.

[zerO_crash]

29 minutes ago, edmuss said:

2.8 introduced a lot of additional lighting which caused a large performance hit, especially to the ka50.  I too see the lower GPU frametimes whilst the textures are loading, typically I'll be at around 8-9ms which then jumps up to 12-14 in the ka50.

I'm not upscaling by running more than 2160, full render resolution as defined by HP on the G2 is around 3160x3100 per eye, this accounts for lense distortion. I run 75% resolution with 2xMSAA, it gives less aliasing and better performance than 100% without MSAA but the image is not as crisp; I find that if I drop the resolution much more then the reaibility of the MFDs starts to suffer.

Blurred shadows was one of the things introduced in 2.8 that initally lowered fps without anyone knowing about it, however soon after, it was separated and is now a separate setting. What else changed, I´m not sure (been afk for half a year - travelling), but I certainly get less than half of the fps that I used to get. I see the primary reason in the CPU-bound "bug" that occurs only when DCS is run in VR. Not really sure what happened, but this needs to be fixed.

 

You are indeed upscaling, the pixel resolution of the two displays in the HP Reverb G2 is indeed 2160x2160. How the lens warps the image and what HP recommends is another topic, but on purely technical level, anything above 2160x2160 are "rendered" pixels, as you are producing more pixels than the panels physically have. Beyond that, I am very satisfied with the default resolution of HP Reverb G2, if only the framerate was back to the early 2.7.

 

I hope the devs can really chime in on this and give a couple of definite answers on the follwing:

- Is the CPU-bound flickering (only in VR) recognised as a "bug", and if so, will it be fixed?

- Is there anything added graphics-wise (beside "Blurred shadows") that has impacted performance in any significant way and is there to stay? Or is this a matter of time and optimization? Any idea?

[edmuss]

The CPU bound thing I think is a bug or misnomer, I'm now far less CPU bound than I was pre-MT update (ignoring what the graph says) in that I can now run much heavier AI missions without the CPU dragging the GPU through the dirt - getting appGPU frametimes good for 80fps but only getting 50.

I get CPU bound render thread on the menu and it flickers between CPU bound main thread and render thread when in flight.  The GPU is typically around 95-100% utilisation (depending on module) and is my bottleneck.

The flat terrain shadows in their entirety were bugged for performance with the introduction of 2.8, these are now fixed with the latest MT update.

With 2.8, prior to the release of BS3 there was an engine lighting update with the things like canopy glare and shadows being added, additionally the new clouds/shadow engine was introduced.  The performance hit is the sum of a multitude of little tweaks, it's not openxr specific, simply DCS 2.8 itself

[dock999]

10 hours ago, edmuss said:

2.8 introduced a lot of additional lighting which caused a large performance hit, especially to the ka50.  I too see the lower GPU frametimes whilst the textures are loading, typically I'll be at around 8-9ms which then jumps up to 12-14 in the ka50.

I'm not upscaling by running more than 2160, full render resolution as defined by HP on the G2 is around 3160x3100 per eye, this accounts for lense distortion. I run 75% resolution with 2xMSAA, it gives less aliasing and better performance than 100% without MSAA but the image is not as crisp; I find that if I drop the resolution much more then the reaibility of the MFDs starts to suffer.

just have a quick question . why do you say the G2 resolution is actually around 3160 x 3100 rather than 2160 x2160 which is defined by manufacturer?  To set resolution 3160 x3100 the openXR tool has to be set up to something like 130% rather than 100%.

[edmuss]

44 minutes ago, dock999 said:

To set resolution 3160 x3100 the openXR tool has to be set up to something like 130% rather than 100%.

Incorrect, set openxr resolution to 100% and the render resolution will display on the second tab.

It was confirmed by HP tech when the G2 came out that ~3160 was the default resolution.  I don't know the reason why it's set up the way it is but that's how it works

 

[dock999]

6 hours ago, edmuss said:

Incorrect, set openxr resolution to 100% and the render resolution will display on the second tab.

It was confirmed by HP tech when the G2 came out that ~3160 was the default resolution.  I don't know the reason why it's set up the way it is but that's how it works

 

That is very interesting. thanks. Do you feel setting up above 100% in the  OpenXR tool increases the clarity of YOUR VR headset, which is actually beyond the native resolution of 3160 x3190 if the figure is true.  The very popular figure of the open XR tool in the forum is 150% if you have a 4090 or 4080 card.

[edmuss]

It does increase the clarity somewhat but there comes a point of diminishing returns.

I've never run more than 100% on the G2 outside of testing because I don't have the horsepower

[zerO_crash]

19 hours ago, edmuss said:

The CPU bound thing I think is a bug or misnomer, I'm now far less CPU bound than I was pre-MT update (ignoring what the graph says) in that I can now run much heavier AI missions without the CPU dragging the GPU through the dirt - getting appGPU frametimes good for 80fps but only getting 50.

I get CPU bound render thread on the menu and it flickers between CPU bound main thread and render thread when in flight.  The GPU is typically around 95-100% utilisation (depending on module) and is my bottleneck.

The flat terrain shadows in their entirety were bugged for performance with the introduction of 2.8, these are now fixed with the latest MT update.

With 2.8, prior to the release of BS3 there was an engine lighting update with the things like canopy glare and shadows being added, additionally the new clouds/shadow engine was introduced.  The performance hit is the sum of a multitude of little tweaks, it's not openxr specific, simply DCS 2.8 itself

I expect so, still, changing the overall settings yields very little fps-difference. I can add 2x MSAA at the cost of 1-2 fps for example. Changing coulds from Ultra to low, results in no change, same as turning off absolutely all shadows. That’s why I attribute the issue to the one thing I cannot avoid using - OpenXR. Furthermore, I’ve been using purely SteamVR before, and never had a problem. I can increase fps by lowering the initial resolution, however at a very bad rate (gain few fps for halving the total amount of pixels). There is something bugged with VR, more specifically, OpenXR. It might be even something else, but in 2D, this is not a problem. Also, it affects both ST and MT performance-wise.

 

I’ll have to wait for an update, but quite a few people with HP Reverb G2 are getting the same results as me. For me, MT runs equally bad as ST. I’ve been wondering why ST of all things runs bad as well, but I imagine it has to do with the implementation of MT (we are talking about VR all the time, 2D is working splendid).

 

Let’s not the derail the thread. I presume that the reason why OpenXR targets a much greater resolution at 1oo% setting, is the same as SteamVR - the first generations of VR headsets had so bad resolutions, that in order to have any decent image fidelity, it was common to upscale. As to why a HP-employee would claim that it’s recommended with upscaled resolution, that’s beyond me. HP Reverb G2 is quality-wise on the level of Varjo Aero, when looking at the sweetspot. It is absolutely fantastic, even at default resolution.

Edited March 31, 2023 by zerO_crash

[backspace340]

8 minutes ago, zerO_crash said:

I expect so, still, changing the overall settings yields very little fps-difference. I can add 2x MSAA at fhe cost of 1-2 fps for example. That’s why I attribute the issue to the one thing I cannot avoid using - OpenXR. Furthermore, I’ve been using purely SteamVR before, and never had a problem.

 

I’ll have to wait for an update, but quite a few people with HP Reverb G2 are getting the same results as me. For me, MT runs equally bad as ST. I’ve been wondering why ST of all things runs bad as well, but I imagine it has to do with the implementation of MT (we are talking about VR all the time, 2D is working splendid).

 

Let’s not the derail the thread. I presume that the reason why OpenXR targets a much greater resolution at 1oo% setting, is the same as SteamVR - the first generations of VR headsets had so bad resolutions, that in order to have any decent image fidelity, it was common to upscale. As to why a HP-employee would claim that it’s recommended with upscaled resolution, that’s beyond me. HP Reverb G2 is quality-wise on the level of Varjo Aero, when looking at the sweetspot. It is absolutely fantastic, even at default resolution.

You can fix this pretty easily - just set the Pixel Density in the in-game VR settings menu to 0.6 or so, that'll get you back down to where you were on SteamVR. I've got a G2 & 1080ti and had to do the same thing when switching to OpenXR - my Steam VR resolution was set closer to the actual G2 resolution, but OpenXR is running way more pixels than that by default. I dropped the pixel density, didn't see any drop in visual fidelity (because it was rendering more pixels than I could even see, wasting performance) and my frames recovered.

Edited March 31, 2023 by backspace340

[zerO_crash]

4 hours ago, backspace340 said:

You can fix this pretty easily - just set the Pixel Density in the in-game VR settings menu to 0.6 or so, that'll get you back down to where you were on SteamVR. I've got a G2 & 1080ti and had to do the same thing when switching to OpenXR - my Steam VR resolution was set closer to the actual G2 resolution, but OpenXR is running way more pixels than that by default. I dropped the pixel density, didn't see any drop in visual fidelity (because it was rendering more pixels than I could even see, wasting performance) and my frames recovered.

 

I’ll try it, although I already lowered the OpenXR resolution to 50% to accommodate the 1.0 Pixel Density in DCS. Essentially, I did it in the app, instead of DCS. The problem persisted. I’ll reverse it, and see if it helps.

Edited March 31, 2023 by zerO_crash

[edmuss]

Shadows high to off gives me an additional 3-5ms, it's a massive boost, similarly clouds low to ultra is 3-4ms - I suspect that you might simply be CPU bound which is why you're not seeing big changes in GPU settings.  The graphics options don't always make a huge difference and DCS doesn't scale well to lower hardware (never has done) but the difference between low and high with shadows specifically is pretty apparent.  The MT update has unloaded some of the workload onto more cores, however if your cores simply aren't very fast (as yours aren't in terms of throughput compared to a modern CPU) then they can become overwhelmed.

Think of a tank full of water, you drill 6 x 10mm holes in it to drain the water out (this is your CPU), now take an identical tank and drill 6 x 20mm holes in it to drain the water out (this is a modern CPU); the 8 holes will pass the water far, far quicker, reducing the bottleneck.

edit: just run the hydraulic calculations through, a 20mm hole will flow about 3.8 times the amount of water that a 10mm hole will - obviously it's not directly related but that's what DCS is doing to your CPU.  Trying to push too much data through too small a hole.

I don't think the high render resolution of the G2 is because of low panel resolution of previous generations of headset, but it's blatently obvious that if you run it at 47% of the total pixel count to achieve the 2160 widre resolution then it gives a rather mushy image.  The 100% setting of 3160 is driven by the headset drivers and WMR, HP themselves have set it like that.

No one is actually forcing you to use openxr, you're free to use openbeta via openvr but I suspect that your performance is going to be just about the same as via openxr

edit:

Some cloud performance testing screenshots I took the other night. Reverb G2 @ 2400 resolution, mostly high across the board settings.

Low clouds - 9ms = 111fps

Ultra clouds - 15ms = 67fps

The ultra clouds wasn't hugely consistent, sometimes it was 13ms, sometimes 15ms depending on which clouds I was looking at.  However there is a huge gulf in performance between the two.

Edited March 31, 2023 by edmuss

[St4rgun]

Before MT came out I used OpenXR @ 300% (!!!) with my G2. The clarity is MUCH better and the sweetspot is wider with those settings.

But of course, that resolution is so huge, that I had to apply OpenXR Toolkit's FSR rescaler with 60%. With this combination the visuals were extremely good.

Unfortunatelly this is not the case with the current OB (MT or ST doesn't matter), because these settings would give bad performance. Now I reverted back to "only" 100% and disabled FSR. That's why I'm eagerly waiting for DLSS to be able to properly upscale and have a nice, temporal anti-aliasing at the same time.

I already asekd @BIGNEWY about the possibility of the early DLSS implementation, but I think he was too busy to answer.

[BIGNEWY]

48 minutes ago, St4rgun said:

Before MT came out I used OpenXR @ 300% (!!!) with my G2. The clarity is MUCH better and the sweetspot is wider with those settings.

But of course, that resolution is so huge, that I had to apply OpenXR Toolkit's FSR rescaler with 60%. With this combination the visuals were extremely good.

Unfortunatelly this is not the case with the current OB (MT or ST doesn't matter), because these settings would give bad performance. Now I reverted back to "only" 100% and disabled FSR. That's why I'm eagerly waiting for DLSS to be able to properly upscale and have a nice, temporal anti-aliasing at the same time.

I already asekd @BIGNEWY about the possibility of the early DLSS implementation, but I think he was too busy to answer.

Hi DLSS is planned, but we are not ready to share a date for its public implementation in DCS. We will share more news in the future. 

thank you

[edmuss]

1 hour ago, St4rgun said:

Before MT came out I used OpenXR @ 300% (!!!) with my G2. The clarity is MUCH better and the sweetspot is wider with those settings.

But of course, that resolution is so huge, that I had to apply OpenXR Toolkit's FSR rescaler with 60%. With this combination the visuals were extremely good.

Unfortunatelly this is not the case with the current OB (MT or ST doesn't matter), because these settings would give bad performance. Now I reverted back to "only" 100% and disabled FSR. That's why I'm eagerly waiting for DLSS to be able to properly upscale and have a nice, temporal anti-aliasing at the same time.

Out of interest by some rough calculations: -

100% render resolution being approximately 3160x3090 per eye resulting in 9.76 million pixels per eye to render.

300% render resolution would be approximately 5470x5350 per eye resulting in 29.2 million pixels per eye to render.

If you then apply 60% upscaling to that resulting resolution you end up with a render resolution of around 3283x3211 in addition to the upscaling overhead which would be around 0.75-1ms.

Basically running 100% will always perform better than overriding 300% @60% upscale.

If you want to still override massively and upscale 60% to get the image quality then try 275% instead of 300%, this would give a resultant render resolution of around 3145x3075 and should give very similar performance (possibly slightly worse) to 100% once the upscaling overhead is factored in