G2, supersampling, MSAA performance face-off

[DeltaMike]

OK I set up this test.  Obviously these tanks are a lot closer than an aircraft would be.  The pic is taken at 100/100/1.0 with no anti-aliasing, and it's actually a fairly good representation of what I see through the G2.  The third tank is shimmery but identifiable as a tank.  The fourth, I can't consistently see the gun so I might know it's armor, but wouldn't know it was a tank.  5 and 6 are just blobs, I couldn't tell 6 from any other vehicle.

SSAAx2 and MSAAx2 look about the same.  

Running a pixel density of 1.4 (equivalent to Steam at 196%), I think tank #4 looks a little clearer.  Shimmery but I think I could identify it as a tank.

To my surprise setting steam resolution at 200%, with SuperDooperSampling turned on, doesn't look as good as PD 1.4.  

I think the supersampling techniques (SSAA, MSAA) have some potential here, as you are increasing the amount of information available to your brain, averaged over time.  FXAA of course does not increase information available and seems to be reducing it actually.  Unclear to me which way TAA swings.

Biggest concern I have is the resolution of objects as a function of draw distance.  I think it's possible the information just isn't there. 

 

Edited March 21, 2022 by DeltaMike

[Al-Azraq]

34 minutes ago, DeltaMike said:

OK I set up this test.  Obviously these tanks are a lot closer than an aircraft would be.  The pic is taken at 100/100/1.0 with no anti-aliasing, and it's actually a fairly good representation of what I see through the G2.  The third tank is shimmery but identifiable as a tank.  The fourth, I can't consistently see the gun so I might know it's armor, but wouldn't know it was a tank.  5 and 6 are just blobs, I couldn't tell 6 from any other vehicle.

SSAAx2 and MSAAx2 look about the same.  

Running a pixel density of 1.4 (equivalent to Steam at 196%), I think tank #4 looks a little clearer.  Shimmery but I think I could identify it as a tank.

To my surprise setting steam resolution at 200%, with SuperDooperSampling turned on, doesn't look as good as PD 1.4.  

I think the supersampling techniques (SSAA, MSAA) have some potential here, as you are increasing the amount of information available to your brain, averaged over time.  FXAA of course does not increase information available and seems to be reducing it actually.  Unclear to me which way TAA swings.

Biggest concern I have is the resolution of objects as a function of draw distance.  I think it's possible the information just isn't there. 

 

 

Thanks a lot for more testing! A couple of notes though:

SSAA doesn't affect the VR image, it is supersampling but for a 2D image. The way of supersampling in VR in DCS is going to SteamVR and setting higher resolutions but even if you put some crazy resolution you won't get rid of the shimmering because even in that case, the upscaled image will have jaggies which will not be entirely removed when downscaled to the headset resolution. To clear the image we will also need an AA solution as well.

SSAA is a very expensive AA solution because it is not selective and it creates extra pixels for the whole frame. Techniques like MSAA generate extra pixels but only in the relevant areas such as the object borders, that's why MSAA sometimes doesn't apply to every object. So MSAA is sort of supersampling but only in the relevant areas.

TAA being the most modern one and less taxing is perfect for VR because it uses information from past and present frames without the need of upsampling or creating new pixels (contrary to what MSAA does). Then it reconstructs the image using these pixels knowing where they were placed in past frames and doing some fancy calculations, that's why it needs motion vectors introduced by the developer.

Edited March 21, 2022 by Al-Azraq

[DeltaMike]

Interesting, so we could do both then.  Supersample via steam or DCS, and then do a pass with TAA.  Interesting.  I'd try that in MSFS, except I'm scared my GPU will catch on fire

Regarding draw distance, I guess that's been a topic of discussion for years.  Good discussion here for example.  

When I talk about spotting distance in my tests, I'm referring to the pixel-sized sprites.  Way I look at it, in VR, we have the same probletunity as people with low-res flat screen monitors.  Even though you might have high-res monitors in your headset, when you put em right on your eyeball and blow them up to iMax proportions, each pixel looks pretty big, and I think we do come out ahead by turning AA off.  

As for once you're close enough to get a 3D object, I still wonder about level of detail as a function of draw distance, but I wouldn't know one way or the other

[Al-Azraq]

28 minutes ago, DeltaMike said:

Interesting, so we could do both then.  Supersample via steam or DCS, and then do a pass with TAA.  Interesting.  I'd try that in MSFS, except I'm scared my GPU will catch on fire

Regarding draw distance, I guess that's been a topic of discussion for years.  Good discussion here for example.  

When I talk about spotting distance in my tests, I'm referring to the pixel-sized sprites.  Way I look at it, in VR, we have the same probletunity as people with low-res flat screen monitors.  Even though you might have high-res monitors in your headset, when you put em right on your eyeball and blow them up to iMax proportions, each pixel looks pretty big, and I think we do come out ahead by turning AA off.  

As for once you're close enough to get a 3D object, I still wonder about level of detail as a function of draw distance, but I wouldn't know one way or the other

Yes sure, you can do both at the same time. TAA as it doesn't render additional pixels it should have a really minimal computational cost but it is not easy to demonstrate because the games that introduce this technique, usually don't offer the option to turn it off because otherwise it breaks lighting, shadows, etc.

I don't think LOD has an effect for spotting distance as the object gets rendered at low poly, but the shape remains the same if the developer did its job right.

[DeltaMike]

I think I see what you're saying.  You don't need a bajillion triangles to define an object's profile; just enough.  If you blur the edges, eg with FXAA, you would need more pixels.  Likewise, with lower pixels-per-degree, you would need more pixels to resolve the object, as if you were looking at a low-res monitor.  

Now I feel guilty about casting aspersions on ED.   I don't normally do that.  I think they do good work.  

Edited March 21, 2022 by DeltaMike

[DerekSpeare]

A bit off topic:  Which gives better performance, setting PD in DCS directly or setting the equivalent PD resolution in Steam VR and set DCS PD at 1.0?
Is super sampling better on the performance than FXAA?  MSAA?  Which form of AA in DCS has the least impact on frames that gives "some" improvement in visuals?

Really, ED has optimizing to do with DCS in VR so it is optimized for the VR player...it is the future, believe it or not. 

[DeltaMike]

1.. Neither.  They both do the same thing.

The big picture is, performance correlates with pixel count.  Pixel count increases linearly with Steam resolution, and with the square of pixel density.  For a G2, the formula is 

(3100x3100)*g*a*p^2 

Where g = Steam General Resolution, a = Per App resolution, p = pixel density

Doesn't matter how you get there, long as the numbers add up.  Best to just adjust one, and set the others to 1.0 or 100%.  With Steam resolution being the more granular parameter. 

There's some debate as to whether adjusting one is better than the other.  If there's a difference in quality, it is subtle indeed.  It's OK to ignore that debate until we come up with some solid data.  Meanwhile, pick one. 

Resolution is king in DCS. Get a high res headset, and shoot for native. 

2.  In terms of antialiasing, in decreasing order of performance hit: 

• Supersampling has theoretical benefits, especially for low resolution headsets, but in DCS at least, eliminating jaggies and shimmering isn't one of them.  Flogs the daylights out of your GPU. 
• MSAA looks great but it crushes your framerate.  Ditto.  I think MSAA is quite a bit more efficient than supersampling at getting rid of jaggies and shimmering.  
• FXAA and SMAA accomplish something at least, at about the third of the cost of MSAA.  (SMAA is probably the better technology and the ReShade version of it is pretty impressive imo)

 

3. It all sucks.  People were kind of butt-hurt that DCS went to deferred rendering, which killed MSAA. But time marches on, and there's new technology we are begging DCS for.  Meanwhile, take a ride around NTTR and then run the same mission in MSFS. Which one is better? I imagine your answer is like mine, "it depends."  Do you want to look at the scenery, or blow it up?  Valid question I think

Edited March 22, 2022 by DeltaMike

[Sr.]

1 minute ago, DeltaMike said:

1.. Neither.  They both do the same thing.

The big picture is, performance correlates with pixel count.  Pixel count increases linearly with Steam resolution, and with the square of pixel density.  For a G2, the formula is 

(3100x3100)*g*a*p^2 

Where g = Steam General Resolution, a = Per App resolution, p = pixel density

Doesn't matter how you get there, long as the numbers add up.  Best to just adjust one, and set the others to 1.0 or 100%.  With Steam resolution being the more granular parameter. 

There's some debate as to whether adjusting one is better than the other.  If there's a difference in quality, it is subtle indeed.  It's OK to ignore that debate until we come up with some solid data.  Meanwhile, pick one. 

Resolution is king in DCS. Get a high res headset, and shoot for native. 

2.  In terms of antialiasing, in decreasing order of performance hit: 

• Supersampling has theoretical benefits, especially for low resolution headsets, but in DCS at least, eliminating jaggies and shimmering isn't one of them.  Flogs the daylights out of your GPU. 
• MSAA looks great but it crushes your framerate.  Ditto.  I think MSAA is quite a bit more efficient than supersampling at getting rid of jaggies and shimmering.  
• FXAA and SMAA accomplish something at least, at about the third of the cost of MSAA.  (SMAA is probably the better technology and the ReShade version of it is pretty impressive imo)

 

3. It all sucks.  People were kind of butt-hurt that DCS went to deferred rendering, which killed MSAA. But time marches on, and there's new technology we are begging DCS for.  Meanwhile, take a ride around NTTR and then run the same mission in MSFS. Which one is better? I imagine your answer is like mine, "it depends."  Do you want to look at the scenery, or blow it up?  Valid question I think

 

Can you off some input on settings to start with using the ReShade SMAA? Like a middle of the road. I have a hard time determining if the slider is having an effect.

[DeltaMike]

16 minutes ago, Sr. said:

Can you off some input on settings to start with using the ReShade SMAA? Like a middle of the road. I have a hard time determining if the slider is having an effect.

Based on what I've been able to surmise, the default settings are thought to be optimal.  Ton of discussion in the ReShade forums if you're interested. Tough crowd over there. 

Edited March 22, 2022 by DeltaMike

[Sr.]

1 hour ago, DeltaMike said:

Based on what I've been able to surmise, the default settings are thought to be optimal.  Ton of discussion in the ReShade forums if you're interested. Tough crowd over there. 

 

If it's anything like Reddit...I'll pass

[Al-Azraq]

5 hours ago, DeltaMike said:

1.. Neither.  They both do the same thing.

The big picture is, performance correlates with pixel count.  Pixel count increases linearly with Steam resolution, and with the square of pixel density.  For a G2, the formula is 

(3100x3100)*g*a*p^2 

Where g = Steam General Resolution, a = Per App resolution, p = pixel density

Doesn't matter how you get there, long as the numbers add up.  Best to just adjust one, and set the others to 1.0 or 100%.  With Steam resolution being the more granular parameter. 

There's some debate as to whether adjusting one is better than the other.  If there's a difference in quality, it is subtle indeed.  It's OK to ignore that debate until we come up with some solid data.  Meanwhile, pick one. 

Resolution is king in DCS. Get a high res headset, and shoot for native. 

2.  In terms of antialiasing, in decreasing order of performance hit: 

• Supersampling has theoretical benefits, especially for low resolution headsets, but in DCS at least, eliminating jaggies and shimmering isn't one of them.  Flogs the daylights out of your GPU. 
• MSAA looks great but it crushes your framerate.  Ditto.  I think MSAA is quite a bit more efficient than supersampling at getting rid of jaggies and shimmering.  
• FXAA and SMAA accomplish something at least, at about the third of the cost of MSAA.  (SMAA is probably the better technology and the ReShade version of it is pretty impressive imo)

 

3. It all sucks.  People were kind of butt-hurt that DCS went to deferred rendering, which killed MSAA. But time marches on, and there's new technology we are begging DCS for.  Meanwhile, take a ride around NTTR and then run the same mission in MSFS. Which one is better? I imagine your answer is like mine, "it depends."  Do you want to look at the scenery, or blow it up?  Valid question I think

 

Good summary! I’m going to update MSFS 2020 today and try it on VR, see how it looks on Nevada.

And yeah, resolution is what impacts performance more. At first I had almost everything on low but then I noticed that I could increase settings without performance impact as I had some headroom and I’m locked at 45 fps due to reprojection anyways.

With the PC at my signature, I can run DCS at 80% resolution, medium settings, terrain shadows flat, MSAA off. As we talked before MSAA is increasing pixels which paired with deferred rendering it destroys performance.

[DeltaMike]

12 hours ago, Sr. said:

If it's anything like Reddit...I'll pass

Seems like a lot of really bright people, the developers post frequently.  not a lot of hand holding  

[DeltaMike]

Here's a tip from a Discord member:  set resolution at 200% and downscale using FSR at a scale factor of 0.59.

Supersampling to the tune of 200% is good anti-aliasing in anybody's book.  But it's expensive.  Will this strategy work?  Maybe.  

Here is a good discussion of how scaling works.  +

Let's pretend the runtime is the boss, DCS is the worker, and FSR is the union representative.  The boss calls down and says, "I want 200% this cycle!"  The union representative tells DCS, "Don't listen to him.  Just work at about 70% of your normal pace, and we'll make up numbers to make it look like you worked at 200%."  The question is, what are we going to do with those made-up numbers?  And the answer is, we are going to average them all together to smooth things out.

I like running at native resolution, so I made one pass at 200%, scale factor 0.83.  That means I'm rendering at 98% and I'm hoping to get a free pass of antialiasing.  

Here's the results:

1.  Native resolution, no scaling.  FPS 46.  I'm using OpenXR now and it actually doesn't have a terrible amount of shimmering, so let's say it's "ok."
2. 200% + scale factor 0.7.  Performance should be the same; instead, it dropped.  FPS 40, which is really bad, this introduces a lot of stuttering.  Shimmering is about the same.
3. 200% + 0.59.  (recommended setting).  FPS showing 47 but there's a lot of stuttering.  Image looks terrible with ghosting, double images, blurry textures. And the shimmering was even worse than running at native resolution with no AA.
4. 100% x 0.83.  Although theoretically this should result in the same performance, FPS went up to 54.  Image quality is actually quite a bit better but the shimmering was still worse than baseline.
5. 200% with no scaling.  This looks nice!  Very little shimmering.  But my FPS tanked to 32.

There's still no such thing as a free lunch.  

ETA

To be fair, I went back through some of these runs in SVR.  In SVR my baseline performance is 45fps.  I did get the performance gains from FSR, into the 53FPS range.   But, the more scaling involved, the worse the shimmering.  200/.59 had worse shimmering than 100/.83, even though the performance was the same.  

ETA2:

FFS made a math error.  Turns out, FSR scales differently in OXR than it does in SVR.  So, in OXR, running at 200% with a scale factor of 60%, you're actually supersampling, ie increasing your DCS frame size to about 120% net of everything.  So it should run slower.  

Conclusion is the same.  Supersampling doesn't erase the jaggies you get from scaling algo's like FSR.  Suspect nothing does. 

Edited April 22, 2022 by DeltaMike

[greenbelt]

I cant find NFAA or DLAA in Reshade options.

How can i activate them?

[jnr4817]

Very interesting topic and great research. I've got a Rift S, so low resolution. I am running OOT with PD=1.4. I am using DCS PD=1.0 and MSAA=2x. I am happy how it looks, but wish I could get solid FPS. My GFX is 1070ti. I'm getting 30FPS at about 23-24ms. I am not sure what to adjust to get better frames, but I also don't want to have a worse view.

I also agree, DCS needs VR optimization bad.

[edmuss]

Try having a play with OC_ACC, the openxr mod that a lot of people have used to improve their DCS VR experience.

Whilst it's designed for the reverb G2 it has been used with the quest so should be workable for the rift s as well.

[DeltaMike]

On 4/23/2022 at 5:30 PM, greenbelt said:

I cant find NFAA or DLAA in Reshade options.

How can i activate them?

Go back and run the install program, this time select all the available in packages

[greenbelt]

8 hours ago, DeltaMike said:

Go back and run the install program, this time select all the available in packages

I have the lastest version installed and all the packages, by still can´t find those two.

I have found 4 AAs but none of those.

What do I have to look for exactly in Reshade?

Thank you.

[DeltaMike]

I uninstalled ReShade several weeks ago and lost interest in it, things have been moving fast around here last few months.  If you have a WMR headset anyway.  Might need to check the ReShade forums.   

 

Edited April 26, 2022 by DeltaMike