fixed 2.8.1 still has serious issues with terrain objects shadows

[some1]

I used the Hornet ready on the ramp instant action mission comparing both the latest DCS 2.7 and 2.8.1 which was uploaded to "stable" branch yesterday. Using clean DCS "saved games" folder, "high" graphics preset, just changing the Terrain Objects Shadows option. So it's the same track, the same preset and the same conditions in both cases. The tests are on a PC with Ryzen 5800X3D, RTX4090 and 64GB ram, in this scenario the game is CPU limited all the time. The results, according to the built-in stats are:

 

Version	Objects Shadows	FPS	Objects	Triangles
2.7.18	Default	139	1511	1,5 milion
2.7.18	Flat	130	2689	2.8 milion
2.7.18	Off (doesn't work)	-	-	-
2.8.1	Default	131	1700	2,5 milion
2.8.1	Flat	111	3880	6,4 milion
2.8.1	Off	133	3880	6,4 milion

Several issues are clearly visible:

 1. Even with terrain shadows set to off, the game is preparing all the shadow related geometry like it would with Flat shadows. Which is A LOT, especially for something that is never displayed on the screen. It may be related to the fact, that this option still creates shadows in 2.7 even when OFF, and it might not have been correctly disabled in 2.8. 

2. Flat shadows generate a lot more geometry in 2.8. In 2.7 the increase between "default" and "flat" is 86%. In 2.8, the increase from "default" to "flat" is 155%. I'm not surprised that Flat shadows create extra geometry, as I understand they work by creating a "pancake" version of every model and add it to the scene as shadow. I'm not even surprised that this option in both 2.7 and 2.8 generates less FPS in this particular scene, as it doesn't work well when there's a lot of high poly models around. It still works better when there's a lot of shadows from a simple geometry, like flying over a city. But what's surprising, is the huge jump in generated geometry between 2.7 and 2.8, which is reflected in FPS hit: 7% in 2.7, but 15% in 2.8. 

3. Flat shadows look nothing like they did in 2.7. In the previous version they were sharp and aliased. In 2.8 they are smooth and blended. Don't think it comes for free.

4. Even in a "best case" scenario with "default" shadows, DCS 2.8 is creating 66% more triangles for no apparent reason at the same game settings, according to the metric in game. 

It looks like the poor performance in VR is just a reflection of worse 2.8 performance overall. It's just that people playing the game in VR are more susceptible to 10-20% drops in performance, while on a monitor it may not be noticeable if the game now makes 110 FPS instead of 130. Also with a weaker GPU the results will look different, as it will mask the problems with CPU being overloaded with extra geometry in 2.8.  

2.7 flat:

2.7 default:

2.8 default:

2.8 flat:

2.8 off:

hornetReadyRamp.trk DxDiag.txt

Edited December 28, 2022 by some1

[speed-of-heat]

Thanks for this. For what it is worth i am seeing the same behaviour in terms of scale of triangles and objects and broad frame rates in 2d.

[Radius]

Please keep up this test with every update and post the numbers. Lets use gauge the optimisation or lack there of.

Awesome to see how one update can axe 20 FPS like it's nothing.

Thanks.

[zildac]

Please keep up this test with every update and post the numbers. Lets use gauge the optimisation or lack there of.
Awesome to see how one update can axe 20 FPS like it's nothing.
Thanks.

I concur, 20fps loss with flat terrain shadows in 2.8 versus 2.7. Something is quite obviously very "broken *. Setting terrain shadows to *off" returns 2.7 VR FPS and frametimes, pretty much.

[Dangerzone]

7 hours ago, Бойовий Сокіл said:

Who asked ED developers to even touch perfectly working shadows? For the sake of improvement and optimisation?

Hypothesizing on possible reasons:

ED was saying that 2.8 would bring "Performance improvements to VR".  My best guess (and it's only that - a guess) is that the implemented changes are probably a necessary step towards the development of MT and/or Vulkan - and they have done so with the expectation at the same time we should see improvements in VR but instead we're seeing the opposite. 

Another possibility - given that other additional features have been implemented (such as rainbows) when we still don't have performance improvements - it's also possible that ED are focused on improving the visual aspect even though it continues to be more taxing on existing systems prior to actual improvements being released, because let's face it - those youtube video's make DCS look good - even if VR users are hurting more with this update. 

Both of these are just guesses though. I seriously hope my first guess is closer to the mark.

I'm disappointed to see Stable Release get elevated to 2.8.1 with this being an unaddressed ongoing issue. I hope both a fix, and a quick Stable promotion to said fix comes soon.

[TZeer]

What is causing the object count to skyrocket? Is it the shadows?

I think this is the main problem here. Object count has always been a big issue with DCS as it's heavily CPU bound.

[some1]

10 hours ago, TZeer said:

What is causing the object count to skyrocket? Is it the shadows?

Flat shadows are created by making a flattened version of each object and drawing it as shadow. This is an older method of creating shadows, not very efficient when there is a lot of high-poly objects around (like on a ramp) but it does not put as much load on the GPU as the GPU-based "default" shadows. For example when flying over a city where there are many simple buildings and trees, it will perform better. 

But why in 2.8 there's so much more geometry generated with flat/off shadows, and why even "default" option reports a lot of extra geometry for no apparent reason, I've no idea. That's the question to ED devs. 

Another example is the same mission tested in VR. Standing in the parking spot and looking to the right. You can see that "Default" shadows in 2.7 put a bit more pressure on the GPU, but less load on the CPU, so it's always a tradeoff. Either way, RTX4090 + 5800X3D is able to make 90 fps on "High" preset in this particular scene (just barely). That's simply not possible in 2.8. Even with "default" shadows, there's noticeably higher load on both CPU and GPU comapred to 2.7. "Flat" shadows in 2.8 are abysmal, with 30% increase in CPU frametimes compared to the same setting in 2.7. Only turning the terrain objects shadows OFF brings the performance back, but the CPU load is still suspiciously high, higher in fact than 2.7 with default shadows on. No wonder with all that extra geometry created for no reason, as explained in the first post. 

 

2.7 preset HIGH, terrain objects shadows default:

2.7 preset HIGH, terrain objects shadows Flat:

2.8 preset HIGH, terrain objects shadows default:

2.8 preset HIGH, terrain objects shadows Flat:

2.8 preset HIGH, terrain objects shadows OFF:

hornetReadyRampVR.trk

Edited December 29, 2022 by some1

[TZeer]

Thank you for the feedback.

I don't think the triangles are the problem here. I would put my money on the number of objects.

I did a test a few years ago with 3d models in DCS that I found was less optimized. I could find 3d models consisting of 40-150 objects. That was scattered around. In certain areas there was a high amount of these objects, which caused massive FPS drops when you came into those areas.

Construction cranes, containers, small sailboats, and some parts of the cargo trains.

Some did not have a "very" high amount of objects, but when there where large amount of small sailboats or containers, it caused the number of objects to double in certain areas.

Renaming these objects, causing them to not render in-game, doubled my FPS. As the number of objects would go from close to 2000 and down to 700ish. I remember one model alone added 300+ objects into the world when shadows was activated as well....

With CPU's getting better and better, I guess this issue was less of a problem. But getting close to 4000 objects I guess we are close to the limit again of what the CPU can handle without having problems.

The funny thing with the test back then, was that the quality of the model had nothing to do with number of objects. I found newer optimized objects only consisting of 2-3 objects, while older models had close to 20...

Newer models had less objects but more triangles. 

I also placed x number of units in a deserted area on the map and compared the FPS between them. "High object" 3d models gave crap performance, while newer "low object" 3d models gave superior performance.

I'm not home right now, but I will test this next week when I'm home again.

Edited December 29, 2022 by TZeer

[5ephir0th]

Since the release of 2.8 two months ago we are reporting problems with shadows and, above all, the flat ones and the only thing we get from ED is "we are looking at it"

I dont doubt that ED is working on it but an explanations would be great if are taking more than two months to fix it, we don't know what they changed and why about something working fine, we dont know if, for ED, this is a real bug or if its working as intended, just "we are looking at it" while the ones of us that spends thousands on a VR rig want to throw it through the window...

 

[WipeUout]

Feedback from ED staff stated that not everyone is experiencing a FPS drop with 2.8, which got me thinking at what could cause the issue in my setup and not in others.  In this case, then it could be related to certain configurations.  I took this as the main assumption for the test I ran.  I did not find any culprit but it may provide some good lead to look into to find the issue.

Also, looking at my dcs.log, I noticed several warning and errors which could have been related to mods installed/removed/installed over and over again.  So my first step was to un-install DCS completely and re-install fresh.  I also took the extra step of making sure all the sub-folders, other than skins, were emptied before re-install.

After a fresh stable 2.8.1 re-install, which took quite some time, I started to run my test track file vanilla and for each of the mods I used in hope I would find the culprit.  Unfortunately, no different results were observed, I still take a GPU Frate Time hit of 1.6ms from 2.7.18 to 2.8.1.  My results are in the table below.

Findings:

1. It seems that the drop is not related to a broken install or the mods I have been using.

2. The drop is significant also from med shadows to flat shadows.

3. I tried both SteamVR and OpenXR and there were no differences.

4. I use default settings for NCP, HAGS on/off with not difference.

Other possible causes would be driver/hardware issue that would be common to all experiencing the drop, which would not be impossible but unlikely.  It does point to the changes that were made between 2.7.18 and 2.8.0.  The question remains, why would some experience no change or improvement in performance?  Did ED received testing data to demonstrate this? 

After re-install dcs.log Before re-install dcs.log DxDiag.txt performance track Caucasus F-18.trk

Edited December 29, 2022 by WipeUout
Typo

[speed-of-heat]

I wonder if the "resolution" is a factor either through super sampling or PD, or SSVR etc...  so for example I know that BN runs his G2 in near native physical resolution or was (so 2600 x 2600 or so) ?  I may also be talking complete codswallop 

[WipeUout]

I forgot to mention that I also tried at different PD (1.1 vs 1.2) with no noticeable difference other than the one caused by the 121% vs 144% super sampling.  I even went to the extent of changing my HMD's refresh rate and switch my 8k+ with my 5K Super for no difference.  I would rule out also Win11 vs Win10 as both OS were used by people reporting the drop.  Men, this is not easy! 

[speed-of-heat]

15 minutes ago, WipeUout said:

I forgot to mention that I also tried at different PD (1.1 vs 1.2) with no noticeable difference other than the one caused by the 121% vs 144% super sampling.  I even went to the extent of changing my HMD's refresh rate and switch my 8k+ with my 5K Super for no difference.  I would rule out also Win11 vs Win10 as both OS were used by people reporting the drop.  Men, this is not easy! 

fair enough, and I agree I dont think it's likely to be a windows or patch problem for the same reason I have seen some fairly radical performance variations from using older drivers ... the only other explanation I have is ... people are not measuring correctly... so for example of those who arent having a problem how many of them are using ASW/MR/MS etc... and are still "over the frametime limit" and so are taking the hit, they just never notice it (and then not measuring without it, but, I know BN doesn't run MR on his G2) ... it could be a driver, but again that would be just masking the effect with driver performance improvements.

[Gripen 4-1]

6 minutes ago, speed-of-heat said:

fair enough, and I agree I dont think it's likely to be a windows or patch problem for the same reason I have seen some fairly radical performance variations from using older drivers ... the only other explanation I have is ... people are not measuring correctly... so for example of those who arent having a problem how many of them are using ASW/MR/MS etc... and are still "over the frametime limit" and so are taking the hit, they just never notice it (and then not measuring without it, but, I know BN doesn't run MR on his G2) ... it could be a driver, but again that would be just masking the effect with driver performance improvements.

My guess too - If you use ASW /Motion Reprojection you will never notice the performance loss

[slughead]

4 minutes ago, Gripen 4-1 said:

My guess too - If you use ASW /Motion Reprojection you will never notice the performance loss

I run with MR and I noticed the performance loss as I always run Marianas to get a sense if the performance has decreased. It becomes very evident in that arena.

[Gripen 4-1]

Yep and also in very CPU bound scenarios with many AI units.

[freehand]

So who has not had a decrease in performance or has more performance or the same performance as I read some posts from ED of this.

I have a decrease and not read anyone who has not perhaps I have missed them.

Edited December 29, 2022 by freehand

[5ephir0th]

But anyone here, not just increased performance but doesnt feel a noticiable worst performance?

I would love to see some test before and after from those who say they dont feel any difference, something has change for bad since 2.8 and all guns points to shadows

Edited December 29, 2022 by 5ephir0th

[MattCri]

On 12/28/2022 at 3:55 PM, some1 said:

Also with a weaker GPU the results will look different, as it will mask the problems with CPU being overloaded with extra geometry in 2.8.  

 

 

With my old amd rx580 the performance penality is huge, im losing 30 to 40% of the fps i had with 2.7

2.8 with no shadows is slower than 2.7 with flat shadows for me

Edited December 30, 2022 by MattCri

[RuskyV]

19 minutes ago, AirMeister said:

Maybe i'm paranoid but i'm beginning to suspect ED does not really know what they are doing and are just randomly doing stuff ?
Like trying to crack a safe by just punching in random numbers ?

That’s a little disrespectful don’t you think, ED and the community put in a lot of time and effort into DCS and if you had/have been playing this sim for a long time then you would appreciate just how far it has come. 
 

Be part of the solution not the problem

[NineLine]

Thanks I am looking into this now.

[YoYo]

On 12/28/2022 at 3:55 PM, some1 said:

I used the Hornet ready on the ramp instant action mission comparing both the latest DCS 2.7 and 2.8.1 which was uploaded to "stable" branch yesterday. Using clean DCS "saved games" folder, "high" graphics preset, just changing the Terrain Objects Shadows option. So it's the same track, the same preset and the same conditions in both cases. The tests are on a PC with Ryzen 5800X3D, RTX4090 and 64GB ram, in this scenario the game is CPU limited all the time. The results, according to the built-in stats are:

 

Version	Objects Shadows	FPS	Objects	Triangles
2.7.18	Default	139	1511	1,5 milion
2.7.18	Flat	130	2689	2.8 milion
2.7.18	Off (doesn't work)	-	-	-
2.8.1	Default	131	1700	2,5 milion
2.8.1	Flat	111	3880	6,4 milion
2.8.1	Off	133	3880	6,4 milion

Several issues are clearly visible:

 1. Even with terrain shadows set to off, the game is preparing all the shadow related geometry like it would with Flat shadows. Which is A LOT, especially for something that is never displayed on the screen. It may be related to the fact, that this option still creates shadows in 2.7 even when OFF, and it might not have been correctly disabled in 2.8. 

2. Flat shadows generate a lot more geometry in 2.8. In 2.7 the increase between "default" and "flat" is 86%. In 2.8, the increase from "default" to "flat" is 155%. I'm not surprised that Flat shadows create extra geometry, as I understand they work by creating a "pancake" version of every model and add it to the scene as shadow. I'm not even surprised that this option in both 2.7 and 2.8 generates less FPS in this particular scene, as it doesn't work well when there's a lot of high poly models around. It still works better when there's a lot of shadows from a simple geometry, like flying over a city. But what's surprising, is the huge jump in generated geometry between 2.7 and 2.8, which is reflected in FPS hit: 7% in 2.7, but 15% in 2.8. 

3. Flat shadows look nothing like they did in 2.7. In the previous version they were sharp and aliased. In 2.8 they are smooth and blended. Don't think it comes for free.

4. Even in a "best case" scenario with "default" shadows, DCS 2.8 is creating 66% more triangles for no apparent reason at the same game settings, according to the metric in game. 

It looks like the poor performance in VR is just a reflection of worse 2.8 performance overall. It's just that people playing the game in VR are more susceptible to 10-20% drops in performance, while on a monitor it may not be noticeable if the game now makes 110 FPS instead of 130. Also with a weaker GPU the results will look different, as it will mask the problems with CPU being overloaded with extra geometry in 2.8.  

2.7 flat:

2.7 default:

2.8 default:

2.8 flat:

2.8 off:

hornetReadyRamp.trk 735.86 kB · 4 downloads DxDiag.txt 103.06 kB · 2 downloads

 

Good comparisons, it shows clearly where is mistake. Now its a move from ED side. Personally, I will add that I do not see any big differences in graphics, but I also see a downgrade of FPS. I really don't know what it's for? Now this problem has been moved to Stable version which is also incomprehensible. Going in denial with a function that spoils a lot for DCS performance as it turns out. I hope that they'll fix it in January finally or move back to 2.7 effects (as option in settings pretty please). We probably won't see multithreading too soon, just like the promises with Vulcan did not come true. The simulator definitely needs a new engine, because it has not kept up with the development of graphics cards and processors.

[Gripen 4-1]

@some1

Your observations are accurate! I just tested with 2.8.1. and 2.7

In 2.8.1. the CPU framtimes for flat shadows are 3-4 ms higher than (up to 16 ms) on default and the GPU usage is 10-20 % lower because of the CPU limitation

So flat shadows in 2.8.1 still have a bug!

In 2.7. the difference is much smaller and flat is less demanding than default!

[huchanronaa]

Unless DCS WORLD fully supports multi-threaded processors, the above FPS drop problem cannot be solved
Obviously, the core of DCS WORLD needs to be greatly modified to support the new GPU and CPU

[5ephir0th]

4 hours ago, huchanronaa said:

Unless DCS WORLD fully supports multi-threaded processors, the above FPS drop problem cannot be solved
Obviously, the core of DCS WORLD needs to be greatly modified to support the new GPU and CPU

This is not something you fix with multi threading, thats like killing flys with a shotgun, this is just they changed something about flat shadows and they are broken, making it to increase CPU frametimes for no reason, with multi threading suppor they will still be broken...