performance bottlenecks in iOS and android

[huachangmiao]

Is there a better way to optimize?

[durswd]

-Program
I have four ideas.

SIMD. Some parts can be improved with NEON
Multicore. parallel computing with multi thread. It is required to many changes.
Reduce a calculation. If parameters are default, skip calculations.
GPU. Some calculations moves to GPU from CPU. It is required to many changes.

[durswd]

I saw cocos2d-x form. I think it is strange.

[huachangmiao]

sorry. what does form mean？

[durswd]

Cocos2d-x Forums. I'm sorry for typo

[huachangmiao]

i think so too.
the performance increase a lot of when i mark the drawElement function.

[huachangmiao]

maybe something wrong in rendering

[durswd]

I think the effects call 4 or 5 times drawElement by a frame.
Otherwise, it is bug in Effekseer for cocos2d-x or Effekseer1.4(this is beta version)
I'll check it after few days.

[durswd]

I'm checking how many times drawElement is called by a frame at first.
drawElement is called 3 or 5 times and vertices is not many.
Is it same on your environment?

I think that bottle neck is a sending time of vertecies data between gpu and cpu.

Next I check it on smartphone and find bottle necks.

Homing_Laser01
3

Simple_Distortion
5

Sword_Ember is only for PC.

[durswd]

And do you mean that "i mark" is comment out?

[huachangmiao]

yes.

[durswd]

I'm very sorry for my iPhone's connector is something wrong.
I changed a lightning cable, but it is not recognized.
So, I buy new smartphones.
Please wait for few days.

[durswd]

I bought iPhoneSE which has a same spec to iPhone6s.
I played homing and distortion. But it maintains 60 fps.
Would you please send your project to me?

[huachangmiao]

Add a few more items, how many can maintain fps60？

[durswd]

I try it

[durswd]

I'm sleeping. In japan, it is 1 o'clock. I return a result tommorow

[huachangmiao]

have a good dream

[durswd]

Three homings decreases FPS under 60fps (40～60fps)
But I got a good hint for improving.
I try to optimize it on this weekend.

[huachangmiao]

Looking forward to your good news.

[durswd]

Good news.
I realized to improve performance on iOS.
2 times faster at leaset.

This version has not be tested yet on Android, so this version is on branch optimized

https://github.com/effekseer/EffekseerForCocos2d-x/tree/optimized

And you need to change as follows.

manager = efk::EffectManager::create(rsize, 8000);

8000 is the maximum number of sprites generated in an application.

[huachangmiao]

How about the performance of the unity platform in iOS?
There are more than 100 roles in my game's battle scene.
I think play 10 homings maintains 60 fps is OK. now still 3.
T _ T

[durswd]

This is my personal opinion.

Effekseer is optimizing and update a few days after.
But even if Effekseer for Unity is updated, I think 4 or 5 is maximum homings maintaining 60 fps.

It is not opinion for effects.
Unity is difficult to maintain 60 fps than cocos2d-x because C# and GC (C++ is very fast)

I think that to realize 100 roles and rich effects at the same time is very difficult on current smartphone.
To reduce effect sprites and contrive roles is required.

[huachangmiao]

Still thank you for everything.
This is my game. Already released.
http://yxwd.qq.com
In this time, we want to make a 3D game.
Let's optimize together.

[durswd]

Looks very good.
I will help you.

To optimize it, I have any ideas.

1. Homing is not light effect.
   Homing is a sample for PC. So it is not optimized.
   Heavier than it looks.

We are implementing Effekseer 1.4, which is multi platform version and going to show draw call on the editor. It can easy to optimize.

2. Culling
   Effekseer has a culling system to hide effects on out of view.
   But it is not used current Effekseer for cocos2d-x.
   If this function is enabled, effects on the view is only shown.

3, Multithreading
I need to some time to implement it.

4. OpenGL ES3.0
   If We can use opengl ES 3.0, It may improve a performance.

I have questions.

1. Is your 3D game TPS, FPS or isometric?

2. Many large effects like Homing is shown at the same time?
   Or 2 or 3 large effects and many hit effects?

If you cannot show an information on a public, please send a mail.

[huachangmiao]

just like http://yxwd.qq.com this game.
but 3D scene, 3D roles, 3D effect and more roles.

[huachangmiao]

When does the 1.4 version be released?
There are still many users using android by opengl 2.0. so I have to use the opengl2.0.

[durswd]

Thank you. I'll play your game and check it.

I plan to release 1.4 beta in early May on Github.

[huachangmiao]

thx. I will follow it.

[huachangmiao]

It's can play 50 homings maintaining 60 fps.
You just need build the iOS project by release mode.
orz.
that's great.

[huachangmiao]

When the number of vertices exceeds a certain value, the display will be wrong.
maybe vertices > 65535.
You can try play 50 Laser03.efk.
8000 is the maximum number of sprites generated in an application.
may I set this param more than 8000?

[durswd]

1. performance
   I try it (I checked on release mode. But i check it up to 6, because a display is filled.)

2. vertices > 65535
   Because vertex id is managed as short.
   I need to sprite rendering.

You can set this param more than 8000.
But if this parameter is larger than 65535 / 6(the number of index on a sprite), rendering may be something wrong or invalid.

Thank you your information, I can fix these bugs.

[huachangmiao]

I almost completed the battle of my project.
I test the performance in some devices.

iPhone6/iPhone6s/iPhone7/iPhone8/iPhoneX : Perfect performance.
huawei CUN-AL00(GPU none): acceptable
huawei 荣耀V10(GPU mali): very bad

Very strange.
荣耀V10's hardware is higher than CUN-AL00, but performance is worse.

Is this article helpful?
https://community.arm.com/graphics/f/discussions/6657/how-to-gain-performance-through-pbo-pixel-buffer-object-on-mali-t-880

[durswd]

I feel very strange too
I read it.

I have questions.

1. How many is it FPS on CUN-AL00 and 荣耀V10?
2. 荣耀V10's display resolution is 4 times larger than CUN-AL00. Do you decrease a frame buffer's resolution?
3. Do you use maximum vertex count?

I trying to parallel computing now. I hope this optimization help you.

[huachangmiao]

1. I locked the FPS to 30/s.
   iOS: 30/s.
   CUN-AL00 20-25/s.
   荣耀V10 12/s.

2. I draw all objects in a 1136 * 640 renderTexture. then draw the renderTexture scale to device display resolution.

3. I set the maximum vertex count to 8000.

by the way.
I just play 2-3 effects at the same time.

[durswd]

Can you send me this effect?

[huachangmiao]

skills.zip

[durswd]

Thank you. I will check it. I'm sleeping.

[huachangmiao]

Thank you. ; )

[durswd]

I will try to use it.
https://developer.arm.com/products/software-development-tools/graphics-development-tools/mali-graphics-debugger

[huachangmiao]

I close the mapbuffer and bufferrange.
It‘s also very slow in huawei honor v10.

[durswd]

Does this change get enough performance?
I think mapbuffer and bufferrange are not good for current Android (error and performance)

[huachangmiao]

No improvement at all. Still 10-12fps/s.
I try to disable the NEON tomorrow.

[durswd]

OK.

[huachangmiao]

Finally, I found the reason.

void Renderer::setupVBO()
{
glGenBuffers(2, &_buffersVBO[0]);
// Issue #15652
// Should not initialize VBO with a large size (VBO_SIZE=65536),
// it may cause low FPS on some Android devices like LG G4 & Nexus 5X.
// It's probably because some implementations of OpenGLES driver will
// copy the whole memory of VBO which initialized at the first time
// once glBufferData/glBufferSubData is invoked.
// For more discussion, please refer to cocos2d/cocos2d-x#15652
// mapBuffers();
}
I set the maximum vertex count to 100.

It's running OK. 30fps/s

How to fix it?

[durswd]

Thank you !!!!!!!!!
I try to think how to fix. (perhaps divide two draw call)

By the way, may I remove division count parameter in ring and track?
I think it is already not needed.

-So I have a new idea.
-Can the Track and the Ribbon set the count of vertices? just like the Ring.

[huachangmiao]

I think it can be removed. : )

[durswd]

OK!! Please wait for few days.
Until I fix it, you set the maximum vertex count to 100

[durswd]

https://github.com/effekseer/EffekseerForCocos2d-x

I updated. Is it OK?

// large buffer make application slow on Android
int32_t spriteSize = 600;

#if (CC_TARGET_PLATFORM == CC_PLATFORM_IOS || CC_TARGET_PLATFORM == CC_PLATFORM_ANDROID)
renderer2d = ::EffekseerRendererGL::Renderer::Create(spriteSize, EffekseerRendererGL::OpenGLDeviceType::OpenGLES2);
#else
renderer2d = ::EffekseerRendererGL::Renderer::Create(spriteSize, EffekseerRendererGL::OpenGLDeviceType::OpenGL2);
#endif

where spriteSize is larger than 100, but VBO size is lower than 65536.

[huachangmiao]

I'm sorry my reply was late. I spent some time trying.

• When playing a lot of effect will be wrong if changed spriteSize to 600. And the performance also will be affected on the android devices like huawei honor v10.

I refer to the cocos rendering.
Because some implementations of OpenGLES driver will copy the whole memory of VBO which initialized at the first time.
I think glBufferData should be called before draw every time.
I comment the glBufferData's call in VertexBuffer's creator function. and add this function into VertexBuffer::Unlock().
I built it in huawei honor v10. It 's working well.
I think at least in the android platform should be modified like this.
What do you think?

[durswd]

I think it is OK.
But it should be included with #ifdef __ANDROID__
It seem that this implementation decrease a performance on iOS and PC.

[durswd]

And another question,
Is nBufferingMode true? If this option makes no sense, I want to remove this.

[huachangmiao]

OK. This is just a temporary code. I will include #ifdef ANDROID later.

nBufferingMode is true in android opengles2.0?

nBufferingMode = !GLExt::IsSupportedBufferRange() && GLExt::IsSupportedMapBuffer();

[durswd]

I think it is true.
If you have a time, please would you check a performance inserting nBufferingMode = false?

[huachangmiao]

OK, I will check it.

[huachangmiao]

I tested in
huawei CUN-AL00:
nBufferingMode = false; 38/fps
nBufferingMode = true; 41/fps

ZTE C880U:
nBufferingMode = false; 41/fps
nBufferingMode = true; 42/fps

I think you can remove Triple buffering. It 's using with mutiThread render. But cocos is simpleThread render.
mapBuffer shouldn 't be removed, because it can improve performance on android.

I optimize the code.

[durswd]

I fixed it.
(Removed nBuffering and add ANDORID)

https://github.com/effekseer/EffekseerForCocos2d-x

Is it OK?

[huachangmiao]

Thank you.
This is a little different.

[durswd]

Thank you, I will try to fix

[durswd]

I updated.
https://github.com/effekseer/EffekseerForCocos2d-x

[huachangmiao]

It's OK

[durswd]

Thank you for your help.