The mission has low fps. Searching for a solution.

[stgatilov]

Maybe some team members will join discussion ( @Dragofer, @kingsal ) and share their opinion.

In the worst instance of the issue (which I posted as screenshot), there are about 17 x 3 ~= 55 models at same location.
Maybe the author of the model did not anticipate such "massive" usage.

The fact that leaves are alpha-tested quads which overlap each other does not help, I suppose.

[Gadavre]

New important information

This information may be unpleasant for developers, but I can not remain silent....

I tested this mission in game versions 2.07 and 2.09

This mission version works with high good fps  in the game version 2.07. ddaazzaa is a professional  mapper and a very kind person.  May God grant him health and long life!  Now anyone can play his mission if he has weak hardware, but in version 2.07

 

 

In version 2.09, low fps in this mission. We, Players, would like to see this   perfomance issue resolved in version 2.10

Developers, you are the best! I am sure that you will succeed!

Edited April 27, 2021 by Gadavre

[stgatilov]

1 hour ago, Gadavre said:

I tested this mission in game versions 2.07 and 2.09

This mission version works with high good fps  in the game version 2.07.

Which scene do you test?

What is the difference in FPS numbers exactly?
Note that you can use "getviewpos" / "setviewpos" console commands to get into same position across several launches.

[stgatilov]

I tried on TDM 2.07, this FM does not start at all.

@Gadavre, are you sure you used version 2.07?
If you open game console, you will see the four-digit revision number in its bottom-right corner.
Please tell me what you see on the version which is supposedly 2.07.

[Gadavre]

2 hours ago, stgatilov said:

Which scene do you test?

What is the difference in FPS numbers exactly?
Note that you can use "getviewpos" / "setviewpos" console commands to get into same position across several launches.

 

tested the   exterior territory of the castle

Yes, I use " getviewpos"

In version 2.07   59 - 60 fps

in version 2.09, the lowest fps is  14-15

Edited April 27, 2021 by Gadavre

[Gadavre]

21 minutes ago, stgatilov said:

I tried on TDM 2.07, this FM does not start at all.

@Gadavre, are you sure you used version 2.07?
If you open game console, you will see the four-digit revision number in its bottom-right corner.
Please tell me what you see on the version which is supposedly 2.07.

TDM 2.07/64 #8079

ddaazzaa made a new version with compatibility for 2.07

https://drive.google.com/file/d/1i5XHI_CGqrVRLtT1LRLe4cjBaaLgkQ0u/view?usp=sharing

Edited April 27, 2021 by Gadavre

[Dragofer]

2 hours ago, stgatilov said:

Maybe some team members will join discussion ( @Dragofer, @kingsal ) and share their opinion.

My take would be that it may be worth combining models by selecting them and exporting them with DR's model exporter in order to reduce the number of draw calls. Especially the architecture modules, since each piece has a lot of materials that need to be drawn.

If a light hits any part of such a combined model, the whole model will be rendered again. So you need to be careful how you merge models that are around lights.

Combining LOD-enabled is a little trickier since you'd need to export the lower-detail models as groups, too, and update the LOD spawnargs.

For the hedges, it'd be good if someone made a medium detail stage. Could open a thread in the art assets forum, yes.

Generally speaking you can figure out what's dragging down your FPS by temporarily deleting certain pieces, i.e. the architecture modules. That'd show what things may be worth optimising.

[stgatilov]

43 minutes ago, Dragofer said:

My take would be that it may be worth combining models by selecting them and exporting them with DR's model exporter in order to reduce the number of draw calls. Especially the architecture modules, since each piece has a lot of materials that need to be drawn.

It won't help with the problematic model: 50 models is not many, it is 250K triangles which is too many

[stgatilov]

With the special version of FM, I can confirm that the bushes at "setviewpos -1500 -1500 250 50 -170 0" are rendered much faster in 2.07. For me, it is 82 FPS on current SVN vs 150+ FPS in 2.07. I used default/minimum settings, I tried to disable some new features, but did not achieve anything with it.

[cabalistic]

If that's true even with `r_useNewBackend 0`, then the most obvious candidates would either be changes to the model/materials itself in-between those versions, or changes in culling behaviour on the frontend. How does 2.08 perform?

[stgatilov]

16 minutes ago, cabalistic said:

If that's true even with `r_useNewBackend 0`, then the most obvious candidates would either be changes to the model/materials itself in-between those versions, or changes in culling behaviour on the frontend. How does 2.08 perform?

As far as I remember, 2.08 is as slow as 2.09.
New/old backend does not make any difference.
r_showPrimivites shows the same numbers in 2.07 and SVN, so I don't think culling is involved.

I have created issue: 5598
Also extracted an almost minimum test map, it is already in SVN (and attached to issue too).

[cabalistic]

Guess I'll fire up nSight later...

[cabalistic]

Well, this might get a little more complicated.

2.07:

2.09:

@GadavreI know this is probably not your intention, but your posts often read as if you accuse everyone around here of doing a bad job. But as you can see, issues are typically a little more complicated than they may appear at first sight. We do definitely strive to improve the mod, not make it worse. But we are only a handful of hobbyists working on the project in our spare time. We don't have the resources of commercial studios to test everything on every possible system. Just something to keep in mind

That being said, the profiling on my system is pretty clear: GPU time is approximately similar, but 2.07/2.08 are CPU bound due to draw call inefficiency. 2.09 focussed on improving that part specifically, and so the framerate improves quite a bit. Obviously, these improvements do not help if you are GPU bound, but there's also nothing here that would suggest that 2.09 should perform that much worse.

I'll try and test on a weaker GPU to see if I can reproduce the issue there. You might also try and update your drivers and system, if there are updates available. Just in case...

[duzenko]

 

2 hours ago, stgatilov said:

As far as I remember, 2.08 is as slow as 2.09.
New/old backend does not make any difference.
r_showPrimivites shows the same numbers in 2.07 and SVN, so I don't think culling is involved.

I have created issue: 5598
Also extracted an almost minimum test map, it is already in SVN (and attached to issue too).

Off the top of my head - VBO ends up in the system RAM?

[stgatilov]

52 minutes ago, cabalistic said:

That being said, the profiling on my system is pretty clear: GPU time is approximately similar, but 2.07/2.08 are CPU bound due to draw call inefficiency.

I suggest going to bugtracker for technical details.

I have Ryzen 1600 and GeForce 1050ti, and for me the difference is pretty clear.
Perhaps we should recheck with default configs on both versions, just to be sure...
 

[Gadavre]

@cabalistic  I have no doubt, that version 2.09 is better, superior to version 2.07 in some way....  You've done a great job.  Almost now there are no such free game projects in the world as this one.  your contribution to the development of the game is very significant.... And the project continues to develop...

But I'm asking you very much. Do not make new versions of this game available only for owners of powerful gaming computers. You can always leave ability  of playing with the performance as in 2.07. For those who wish it

 

P.S.

But it's up to you, the developers, to decide . The choice is yours. I will accept any choice you make. As a last resort, I will use version 2.07

Edited April 27, 2021 by Gadavre

[cabalistic]

1 hour ago, Gadavre said:

Do not make new versions of this game available only for owners of powerful gaming computers. You can always leave ability  of playing with the performance as in 2.07.

This is what I mean. You seem to think that's our goal, or simply a matter of choice. It is not. But we simply can't test on every possible hardware, we have neither the time nor the resources. The changes in 2.09 are designed to benefit more powerful hardware, yes, but they are not designed to make weaker hardware worse (except really old hardware, perhaps). And in general, they don't. This is a particular case, and we'll have to figure out why. And reporting performance regressions is very much appreciated. Just leave the analysis and assumptions to us

Btw, I did a brief test on my work laptop with a 1050 Ti. Performance does drop slightly in @stgatilov's scene, from about 200 fps to 180. So there is a regression, but it's nowhere near your magnitude. I haven't been able to analyse it further, because nSight for some reason refuses to access performance counters on the Laptop, will have to figure that out first.

[nbohr1more]

Hmm...

Are we doing a true apples to apples comparison ( 32-bit color on 2.08 / 2.09 )?

Both compared with shadow map vs shadow map ?

[stgatilov]

6 hours ago, nbohr1more said:

Are we doing a true apples to apples comparison ( 32-bit color on 2.08 / 2.09 )?
Both compared with shadow map vs shadow map ?

I tried to set 32-bit color, tried to disable tonemap, tried old backend --- made no difference for me.

I used stencil shadows, because 1) that's default, and 2) they are softer to GPU.
Anyway, this scene has no lights except ambient and no shadows.

I have posted my results in the issue.
It seems that rendering cost on GPU has doubled for me.
The only thing which comes to my mind is VertexCache changes...

[duzenko]

R550/64b: 100 vs. 60 fps, seemingly limited by backend/gpu

Afterburner reports lower VRAM usage by SVN vs. 2.07: about 1010 vs 1080MB. But I'm not sure how the texture compression changes affect this. Is there any way nSight can tell where the VBO is located?

Retested with uncompressed textures:

[AluminumHaste]

Are you sure that antialiasing is set the same for both? Are you sure that you haven't changed settings for each executable in your graphics control panel differently?

In 2.07 running 4xAA was doable with not that much performance penalty.

In 2.09 running 4xAA would cause an almost 4x decrease in performance, at least that's what I've experienced.

[AluminumHaste]

Well, now that I've said that....I went and checked on the latest SVN build, and the performance drop is not that bad anymore?

I guess something has changed.

0xAA:

2xAA:

4xAA:

 

180FPS is still very playable

 

[nbohr1more]

Yep, stgatilov improved shadow AA performance just before 2.10 beta.