Performance!?

[Davision]

It seems to me the performance gets worse with every UDK release instead of better! Are you haveing problems with performance too?

Im trying to get my game run with 30 fps on a ipod touch 4 and its currently around 20fps. Tried a lot of different things already to improve it but had not get it much better. Also setting most of the graphic settings in the ini to something basic has no real impact on the performance for me. Even when no code is running and the camera has almost only the 4200 tris character in the view its around 25fps.

[darthviper107]

It depends on the type of stuff you have in the game--also, the iPod Touch isn't very powerful.

[cwill]

After updating my UDK I copyed over a project from the September update to the October update, gutted everything out of the map and installed to my iPod touch 4. I also noticed a massive fps loss, and this was on a map that had all it's static meshes removed versus the one that was covered in them in my September build.

[Tom P]

Yes, there is really a performance loss in the last beta. For me it's a fall from 30fps to around 20. Same content.
I'm using an ipod touch 4 and the ipad 1. The Ipod Touch is by far faster but both suffer from the last update.

[Davision]

After updating my UDK I copyed over a project from the September update to the October update, gutted everything out of the map and installed to my iPod touch 4. I also noticed a massive fps loss, and this was on a map that had all it's static meshes removed versus the one that was covered in them in my September build.

Yes, there is really a performance loss in the last beta. For me it's a fall from 30fps to around 20. Same content.
I'm using an ipod touch 4 and the ipad 1. The Ipod Touch is by far faster but both suffer from the last update.

wow, i had no idea it it is this extreme from september to october udk! I hope EPIC notices this thread somehow and does something about the performance for one time instead of only adding more and more performance heavy features.

It depends on the type of stuff you have in the game--also, the iPod Touch isn't very powerful.

The ipad 1 and iphone 3gs is les powerful then the ipod touch and the iphone 4 has only more ram which shouldnt mean much for the performance. When it gets already under 30fps with almost only a 4200 tris char in view on the ipod 4 then it cant be the type of stuff. IMO it cant be that the UDK performance is so slow that a decent game doesnt run with 30 fps on the ipod and that it gets even worse with new releases of the UDK!

[Biino]

I haven't upgraded to the Oct release. But I am curious if you guys have the new features enabled? Namely shadows and godrays? I'm sure those feature don't come for free performance wise.

[cwill]

Aren't those features only enabled for iPad2?

(by default of course)

[Showster]

Is it just that the new builds of UDK have ini's that enable too many features on devices like the Ipod 4???

Or is the exact same code with the exact same ini settings getting bad FPS drops? If so that really stinks!!!

You can find a ton of settings in Iphone_UDk_Engine.ini

[SystemSettingsIPhone3GS]
BasedOn=SystemSettingsMobileTextureBias
MobileEnableMSAA=True

[SystemSettingsIPhone4]
BasedOn=SystemSettingsMobile
MobileContentScaleFactor=2.0

[SystemSettingsIPodTouch4]
BasedOn=SystemSettingsMobileTextureBias
MobileContentScaleFactor=2.0

[SystemSettingsIPad]
BasedOn=SystemSettingsMobileTextureBias
MobileFeatureLevel=1
MobileFog=False
MobileSpecular=False
MobileBumpOffset=False
MobileNormalMapping=False
MobileEnvMapping=False
MobileRimLighting=False
MobileContentScaleFactor=0.9375

[SystemSettingsIPad2]
BasedOn=SystemSettingsMobile
MobileEnableMSAA=True

[reopard]

Just to be sure, can you please check your World Properties whether Gamma correction is enabled or not?

[Showster]

Yer I second what Reopard says, this changes fps on my IPod from 20fps to 55-60fps in some cases makes a big difference .

[Davision]

godrays are default off, shadows are default off ...i think they dont even work on the ipod anyway. realtime shadows and godrays would be of course very performance heavy on the ipod. Gamma correction is also default off and totally not worth the performance drop on mobile. Already tried all kinds off low settings with the ini but its not really helping.

@ Showster
What kind of game are you working on?

[Showster]

@ Davision,

Couple of things driving game and this old school shooter



Can you share pictures of your project maybe it'll help us see where the problem is coming from?

[Davision]

Looks like a sidescroller, the advantage you have with 2d gameplay is that you dont need much 3d/much depth. So you dont have much in the view.

My game is basicly a 3rd person "runner" game where you ride along a path and can strafe and jump. A bit like the first crash bandicoot games but you constantly ride and the world is randomly generated from level chunks. Its already very optimized with having the level chunks being combined single meshes and useing texture sheets.





As you can see the camera points down quite a lot so there is not so much in view. But the problem is already, like i stated in the first post, that the performance is low even on the start when there is no code running and only a few polys of the world are visible:



All the shader features are off here and the skin bone influences are set to 2 thats why parts of the knight and horse look screwed. The textures resolution is also pretty low there. Still, it runs with 26 fps as you can see in the upper right.

[Showster]

Looks cool, what happens if you turn alpha off on your materials?

I started to model some of my vegetation to see if alpha was killing some FPS.

Saved a few FPS here and there, I still use blend masked materials on the grass in the foreground and some vines in the background.

Not sure you can do that with yours but still could be worth just turning alpha off and checking the FPS.

Erm otherwise dunno what to suggest, if you have access to a Mac could try the profiling tools http://udn.epicgames.com/Three/MobileProfilingHome.html

Ooh these are handy http://udn.epicgames.com/Three/StatsDescriptions.html , I use STAT ES2 to get a bit of info on whats being drawn.

I'll shout if I find anything interesting

[Davision]

Tried that with alpha off already, didnt changed it noticeable. Im also using blend mask for all my alpha. I tried several stats commands but found not much useful there. Never used STAT ES2 though, its not in the list of stats commands, sounds like its a good help. Is there a description somewhere that explain all the things in the es2 stats list?

[Davision]

The gameplay profiler is very nice, with the help of it i managed to get 1 fps better performance at the start view, unfortunately not that much.

I made some screens with ES2 Stats:
Start view with no code running and all kinds of ini settings set low:


and with playing it:


Anything that should give me a idea what kills the performance?

I have no idea what that (UP) means but i know there is les then 10000 verts in the start view. The fps in the upper right are only that lower because the es2 stats are running.

[Lexluthor1]

Looks like a sidescroller, the advantage you have with 2d gameplay is that you dont need much 3d/much depth. So you dont have much in the view.

My game is basicly a 3rd person "runner" game where you ride along a path and can strafe and jump. A bit like the first crash bandicoot games but you constantly ride and the world is randomly generated from level chunks. Its already very optimized with having the level chunks being combined single meshes and useing texture sheets.





As you can see the camera points down quite a lot so there is not so much in view. But the problem is already, like i stated in the first post, that the performance is low even on the start when there is no code running and only a few polys of the world are visible:



All the shader features are off here and the skin bone influences are set to 2 thats why parts of the knight and horse look screwed. The textures resolution is also pretty low there. Still, it runs with 26 fps as you can see in the upper right.

Have you used Cull distance volumes?
I devices will render everything in the map every tick if you haven't added occlusion zones!


Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

[Showster]

Have u tried the environment with simple game type instead of your own to see if it is the art causing the problem or the code?

[Davision]

I have found something! On level start start i now destroy all actors that dont have a collision and have phys_none. This improves the performance quite a lot. My level is made with interp actors since the world is randomly generated on the fly. When i destroy those with "destroy" in kismet the performance get better significantly. Though useing destroy with "based actors" strangely didnt do it like that, i had it like that before, must be a bug somehow. After destroying i force a garbage collections which also makes the performance a little better and then i unhide all the interp actors so they are visible again. This improvement is also cleary visible in the gameplay profiler, the interp actors made up the most of time now they are down with 0.02. Its strange since i already turned the light component and collision off for the interp actors. For collision i use dynamic blocking volumes, collisions on the level chunks interp actors would be too complex, i need a few dynamic blocking volumes for most level chunks. Dynamic blocking volumes make up the most time now with 0.11 but i guess i will not able to improve that. Long story short, i have now around 24 fps when the gameplay is running and 31fps in the start view. Still not great but a significant improvement.

I think this trick will not work with static meshes and it will not work with actors that need to be really moved with interpolation or have collision since that gets deleted on destroy.


Here is how it looks now in the gameplay profiler, its from running on the pc of course. This is in the start view:



This is while the gameplay is running, the teleporting of the level chunks while playing takes quite a lot (the level chunks are attached to the target points). But i have no idea what i can do about that and i need to do it while playing so i can reuse level chunks:



@Showster
Yea, i tried to delete all my kismet before, that wasnt it. I basicly have everything made in kismet. But i now made my own more simpliefied gametype that got me 0.03 time improvement in the profiler.

@ Lexluthor1
Sry, but that would be crazy if the device renders everything in the map without cull distance volumes! Just look at the castle map, there are none of such volumes. I think what you got confused with is that there is no occlusion working on IOS, that meshes that are hidden by other meshes in front get not rendered. But thats not really a issue for me because the camera points so low and the level chunks are as small as possible because the camera moves on a fixed path. I also tried a low max draw distance on all my level chunks before which didnt changed anything.

[Lexluthor1]

I have found something! On level start start i now destroy all actors that dont have a collision and have phys_none. This improves the performance quite a lot. My level is made with interp actors since the world is randomly generated on the fly. When i destroy those with "destroy" in kismet the performance get better significantly. Though useing destroy with "based actors" strangely didnt do it like that, i had it like that before, must be a bug somehow. After destroying i force a garbage collections which also makes the performance a little better and then i unhide all the interp actors so they are visible again. This improvement is also cleary visible in the gameplay profiler, the interp actors made up the most of time now they are down with 0.02. Its strange since i already turned the light component and collision off for the interp actors. For collision i use dynamic blocking volumes, collisions on the level chunks interp actors would be too complex, i need a few dynamic blocking volumes for most level chunks. Dynamic blocking volumes make up the most time now with 0.11 but i guess i will not able to improve that. Long story short, i have now around 24 fps when the gameplay is running and 31fps in the start view. Still not great but a significant improvement.

I think this trick will not work with static meshes and it will not work with actors that need to be really moved with interpolation or have collision since that gets deleted on destroy.


Here is how it looks now in the gameplay profiler, its from running on the pc of course. This is in the start view:



This is while the gameplay is running, the teleporting of the level chunks while playing takes quite a lot (the level chunks are attached to the target points). But i have no idea what i can do about that and i need to do it while playing so i can reuse level chunks:



@Showster
Yea, i tried to delete all my kismet before, that wasnt it. I basicly have everything made in kismet. But i now made my own more simpliefied gametype that got me 0.03 time improvement in the profiler.

@ Lexluthor1
Sry, but that would be crazy if the device renders everything in the map without cull distance volumes! Just look at the castle map, there are none of such volumes. I think what you got confused with is that there is no occlusion working on IOS, that meshes that are hidden by other meshes in front get not rendered. But thats not really a issue for me because the camera points so low and the level chunks are as small as possible because the camera moves on a fixed path. I also tried a low max draw distance on all my level chunks before which didnt changed anything.

Try running the castle map on an iphone or ipad.....Its terrible........

What I suggested before was suggested to me by a proven iOS developer, "Sjoerd De Jong" Hourences, so are you saying that someone who has worked with Unreal Engine since 2004 is wrong?
Lol


Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

[Davision]

Try running the castle map on an iphone or ipad.....Its terrible........

What I suggested before was suggested to me by a proven iOS developer, "Sjoerd De Jong" Hourences, so are you saying that someone who has worked with Unreal Engine since 2004 is wrong?
Lol


Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

Are you saying that EPIC the maker of the unreal engine and epic citadel did it wrong!? :P
I know hourences, i think you just understood him wrong. Sure, having a cull distance can improve performance when you have a game where you can look far in the distance, the meshes will also noticeable pop in in that distance though. But its not like everything in a iOS map gets rendered all the time without it, its still the cameraview that defines what gets rendered. There are millions of verts in the castle map, no way that that gets all rendered at once on iOS. I know the performance of the castle demo on the ipod is shockingly bad but if it would get rendered all at the same time it would just crash.

[Lexluthor1]

Are you saying that EPIC the maker of the unreal engine and epic citadel did it wrong!? :P
I know hourences, i think you just understood him wrong. Sure, having a cull distance can improve performance when you have a game where you can look far in the distance, the meshes will also noticeable pop in in that distance though. But its not like everything in a iOS map gets rendered all the time without it, its still the cameraview that defines what gets rendered. There are millions of verts in the castle map, no way that that gets all rendered at once on iOS. I know the performance of the castle demo on the ipod is shockingly bad but if it would get rendered all at the same time it would just crash.

Yes I am saying Epic got it wrong and it does crash all the time..
It was thrown together in 9 weeks .....
Looks great but runs s**t



Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

[ffejnosliw]

iOS does not provide automatic occlusion culling as Lexluthor1 mentioned. So if no other methods of culling are used, the only culling performed would be frustum culling. This does mean that everything in the view of the camera, regardless of whether it is actually visible or not would be rendered.

EpicCitadel uses PrecomputedVisibility to cull geometry that is not visible and not culled by frustum culling. CullDistanceVolumes would just be another layer on top of that, but would probably only really provide much benefit in situations where the environment is open and the player can see for long distances.

[Lexluthor1]

iOS does not provide automatic occlusion culling as Lexluthor1 mentioned. So if no other methods of culling are used, the only culling performed would be frustum culling. This does mean that everything in the view of the camera, regardless of whether it is actually visible or not would be rendered.

EpicCitadel uses PrecomputedVisibility to cull geometry that is not visible and not culled by frustum culling. CullDistanceVolumes would just be another layer on top of that, but would probably only really provide much benefit in situations where the environment is open and the player can see for long distances.

Thank you Jeff, and forgive my arrogance...
I was convinced Hourences was right but I didn't mean to say Epic were wrong!

You know I can sometimes be unreasonable




Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

[Davision]

Well, i thought you meant there is no frustum culling on iOS with this:

I devices will render everything in the map every tick if you haven't added occlusion zones!

Anyway, good to have it heard now with the technical right words. PrecomputedVisibility is also tricky since it doesnt work with dynamic stuff and it decreases rendering thread time in game at the cost of increasing runtime memory.

[Lexluthor1]

Well, i thought you meant there is no frustum culling on iOS with this:

Anyway, good to have it heard now with the technical right words. PrecomputedVisibility is also tricky since it doesnt work with dynamic stuff and it decreases rendering thread time in game at the cost of increasing runtime memory.

Now there's something we agree on lol....



Seek And Destroy Games
Tutorials for UDK PC,iOS and 3ds Max
http://www.youtube.com/user/lexluthornumber1

[AaronJones]

I'm just bumping this thread to see if everyone is doing ok with their performance. Are you still having issues or did you all find the solutions you needed?

[Davision]

I'm just bumping this thread to see if everyone is doing ok with their performance. Are you still having issues or did you all find the solutions you needed?

Still have lots of problems with performance. I optimized my game for 4th gen a lot but only got a slight performance improvement out of it. I think a performance improvement on the actual engine is long overdue, especially because i never saw one, in fact it is only getting worse as it can be seen from the september to the october udk version! I also keep hearing from other developers that they have problems with performance aswell as seeing bad performance in UDK apps that are already in the appstore.

What i did to improve my game even more is that i have my already combined meshes now seperated in alpha and non alpha, for 4th gen the alpha stuff gets deleted (its mainly foliage like grass). Though I still have to find a proper way detect which device the game is running on to be able to delete it only for 4th gen.
I had also reduced the polycount of my environments aswell as made it delete all particles. But its still traveling from ca 28 to 20 fps. It can also drop down lower when the player picksup a coin or a enemy appears. I noticed the sound playing takes also quite a lot. The traveling of the fps also comes a lot from touching or not touching the screen. I noticed all input options that require to know the touch position take a lot of performance. Like 5fps in my case. And that is not related to my code nor does a retrigger delay help. So only the simple buttons take no performance. Also of course a big problem are the draw calls, when you need to aim for like 10 draw calls its hard to get any dynamic stuff in the game. Like in my case coins pickups and other elements that can appear. If the engine would be able to do dynamic batching this would be not much of a problem anymore. Also static batching would be a huge improvement, then we would be not required to merge the environment meshes. I have gone also so far to improve a engine script, the target point actors had sprites which showed up in the gameplay profiler.

Currently i think i will not be able to release my game on 4th gen since the performance is even with so many improvements i already made too bad and the lags are also pretty obvious. But without 4th gen support there is only a very slim change that we can ever make enough money with our game.

I think what would really help is making the engine perform overall better. And a solution that allows to detect the device so we can delete certain stuff in the maps for 4th gen devices.

I was hoping now that IB dungeons is in the works that there might be a good chance now for a actual engine improvement for iOS? If it should be really like diablo it will have to be more complex with various enemys at once on screen aswell as dynamic objects like loot.

[AaronJones]

Have you turned on/off features in the Mobile-UDKEngine.ini? There are device-specific options in there to enable/disable certain features.

[SystemSettingsIPhone3GS]
[SystemSettingsIPhone4]
[SystemSettingsIPhone4S]
[SystemSettingsIPodTouch4]
[SystemSettingsIPad]
[SystemSettingsIPad2]
[SystemSettingsIPad2_Detail]

[AaronJones]

Also, you mention 4th gen device; which devices specifically are giving you trouble, and are you having the same issues on all others?

[Davision]

Have you turned on/off features in the Mobile-UDKEngine.ini? There are device-specific options in there to enable/disable certain features.

[SystemSettingsIPhone3GS]
[SystemSettingsIPhone4]
[SystemSettingsIPhone4S]
[SystemSettingsIPodTouch4]
[SystemSettingsIPad]
[SystemSettingsIPad2]
[SystemSettingsIPad2_Detail]

OF course! thats the most basic way to optimize, but except the fancy features like shadows you can only control the shaders with it. I have it all to the most basic set. Also, it doesnt matter really since my game is CPU bound. The texture resolutions that can also be defined with the inis should only matter really if the ram would be a problem.
iPod Touch 4th gen, Iphone 4th gen and iPad 1. The issues are the same on these devices. Execept of the ram the hardware is also very similiar of these devices.

[TheBuilder]

Have you turned on/off features in the Mobile-UDKEngine.ini? There are device-specific options in there to enable/disable certain features.

[SystemSettingsIPhone3GS]
[SystemSettingsIPhone4]
[SystemSettingsIPhone4S]
[SystemSettingsIPodTouch4]
[SystemSettingsIPad]
[SystemSettingsIPad2]
[SystemSettingsIPad2_Detail]

We want to able to do something like this:

If (device == iphone4)
{
     DoSomething();
}
else if (device == iphone4s)
{
    doSomethingElse();
}

For my own case, I want to be able to chose what particle effects to use based on the device. For example having a simpler explosion and smoke particle on weaker devices.

[Creo]

We want to able to do something like this:

If (device == iphone4)
{
     DoSomething();
}
else if (device == iphone4s)
{
    doSomethingElse();
}

For my own case, I want to be able to chose what particle effects to use based on the device. For example having a simpler explosion and smoke particle on weaker devices.

I'd like to have a device type info right through the unrealscript, too. By now, we are using screen resolution to detect device that game running on. But it's not as good approach as we want to see.

[Davision]

I'd like to have a device type info right through the unrealscript, too. By now, we are using screen resolution to detect device that game running on. But it's not as good approach as we want to see.

Currently you can only distinquish from ipad to iphone/ipod with the resolution trick. You could change the resolutions of the different devices with the ini but that would of course result in interpolation of the pixels.

I still would really like to know if there is any chance of a performance update of the engine or if i have to live with what it is?

[AaronJones]

Have you guys tried using LODGroups? I believe that to be the best method as it already does autodetail for low end devices.

Davidson, I don't really have any problems with performance when developing for mobile until I get to older devices which are significantly less powerful. Still though, I am investigating the issues mentioned on performance.

[IamDouglas]

I have to agree with Aaron, I haven't seen a performance related issue as yet, and im testing mostly on 3GS for that reason.
Might it be an issue with something else? Settings or Assets maybe?

[Beezy8669]

Not no problem over here

[Davision]

Have you guys tried using LODGroups? I believe that to be the best method as it already does autodetail for low end devices.

Davidson, I don't really have any problems with performance when developing for mobile until I get to older devices which are significantly less powerful. Still though, I am investigating the issues mentioned on performance.

I think when the unreal engine supports the older devices then it should also easily run smooth on these devices. And its not like they are that old, the ipod 4 is the newest ipod and thats the only ipod the unreal engine supports! Also, the citadel demo is currently running 24-30 on a ipod and that is without skeletal meshes and gameplay, there are also very obvious lags where it goes under 20 fps. The jazz mini game currently goes down to 10 fps, though there seems also something got messed up with that in one of the newer UDK versions. You might want to investigate that too since that could give a very bad impression to new developers trying it.
I dont see how LOD groups have something to do with the performance? they only change the textures resolution. They of course have an impact on how much ram is used which i have already set lower for the 256mb ram devices so it doesnt crash.

I have to agree with Aaron, I haven't seen a performance related issue as yet, and im testing mostly on 3GS for that reason.
Might it be an issue with something else? Settings or Assets maybe?

What are you working on?

[TheBuilder]

Is it possible to somehow run a benchmark test in the game, and detect the device based on the performance?

[Xenobite]

how about adding some ingame gfx settings menu where you can tweak stuff same way as you would on pc?
so you can pick from low quality settings for better frame rates and better playing or high settings with low framerates but you would be able to see all eyecandy
(there should be an option to reset settings to default in case of game crashing)