Что такое findslide.org?

FindSlide.org - это сайт презентаций, докладов, шаблонов в формате PowerPoint.

Для правообладателей

Обратная связь

Email: Нажмите что бы посмотреть

Презентация на тему Mobile optimizations

Содержание

2. Mobile Optimizations
3. Corey JohnsonProduct Manager of Doom
4. Arm you with more possibilities for optimizations that you will be able to utilizeGoal
5. AgendaRecognizing Your Performance BottleneckProfiling in and out of UnityOptimizing Tips
6. What do we mean, Performance?FrametimeCPU usage (Gamecode,
7. Know Your BottlenecksQuestion: Why are we slow?
8. Know Your BottlenecksQuestion: Why are we slow?CPU or GPU Bound?Physics or Rendering?Update() or FixedUpdate() loop?
9. Know Your BottlenecksAnswer: Always start in the same place...ProfileProfileProfile
10. CPU-Heavy TasksPhysicsAnimation Gameplay code Runtime GI Reflection probes ParticlesCreating Batches
11. GPU-Heavy TasksSwitching BatchesGeometry/Pixel shadersCompute shadersSkinning
12. Profiling in UnityUnity ProfilerIn-EditorLive Builds on devicesRapid
13. Custom Profiler TagsDo this:Get This:
14. Unity Memory ProfilerOpen Source https://bitbucket.org/Unity-Technologies/memoryprofiler Profile memory
15. Profiling outside of Unity (iOS)InstrumentsProfile game running
16. Profiling outside of Unity (Android)Unity ProfileradblogcatGPUAdreno (Qualcomm)PVRTune, PVRUniSCo (PowerVR)Intel GPA
17. Garbage CollectionManaged MemorySize doubles when limit is
18. Garbage Collection - Stack vs HeapHeap ObjectsMemory
19. Data Layout Mattersstruct Stuff {int a;float b;bool c;string name;}Stuff[] arrayOfStuff; //
20. Object PoolingCreate a pool of objects to
21. Use System.Text.StringBuilder over stringstring str = “1
22. More Memory OptimizationsReuse temporary buffersIf buffers for
23. Other GC Optimizationsfor(;;) instead of foreachforeach on
24. Marshalling CostYou can write native pluginsCan be
25. Case Study - CachingGetting 20k matrices which
26. Case Study - CachingCache complex expressionsProperties can hide expensive operationsOptimized implementation: 33.5ms
27. Case study - CopyingCreate a method using
28. Optimizing GraphicsBake what can be bakedLightingShadowsBatch what can be batchedStatic MeshesMaterialsUI Canvas elements
29. Optimizing MeshesOnly use as many vertices as
30. Optimizing MeshesDisable “Import Blend Shapes” if none
31. Combine Meshes
32. Combine TexturesTexture Atlases can be made by artists too…
33. Optimizing TexturesiOS - Use PVRTCAndroidOpenGLES 2.0 devices:
34. Optimizing Textures - ExampleShadowgunUsed “Render to Texel”
35. Optimizing Textures - Example
36. Optimizing Audio“Force to Mono” if sounds don’t
37. Optimizing UIKeep UI elements at the same
38. Other OptimizationsLimiting Rigidbodies to 2 dimensions in
39. Script OptimizationsAvoid Find…() methodsCache a reference insteadFindWithTag()
40. Vector Math OptimizationsNormalize a vector once if
41. Shader OptimizationIn general, less instructions is better*Move
42. 10000 Objects Update() vs Update() 10000 ObjectsBlog
43. Скачать презентацию
44. Похожие презентации

Mobile Optimizations

Arm you with more possibilities for optimizations that you will be able to utilizeGoal

AgendaRecognizing Your Performance BottleneckProfiling in and out of UnityOptimizing Tips

What do we mean, Performance?FrametimeCPU usage (Gamecode, Physics, Skinning, Particles, …)GPU usage

Know Your BottlenecksQuestion: Why are we slow?

Know Your BottlenecksQuestion: Why are we slow?CPU or GPU Bound?Physics or Rendering?Update() or FixedUpdate() loop?

Know Your BottlenecksAnswer: Always start in the same place...ProfileProfileProfile

CPU-Heavy TasksPhysicsAnimation Gameplay code Runtime GI Reflection probes ParticlesCreating Batches

GPU-Heavy TasksSwitching BatchesGeometry/Pixel shadersCompute shadersSkinning

Profiling in UnityUnity ProfilerIn-EditorLive Builds on devicesRapid IterationMemory usage of individual

Profiling outside of Unity (iOS)InstrumentsProfile game running on iOS deviceMono & IL2CPP

Profiling outside of Unity (Android)Unity ProfileradblogcatGPUAdreno (Qualcomm)PVRTune, PVRUniSCo (PowerVR)Intel GPA

Garbage CollectionManaged MemorySize doubles when limit is hitNEVER SHRINKSCan stall when collectedCan

Garbage Collection - Stack vs HeapHeap ObjectsMemory block allocated on the Heap

Data Layout Mattersstruct Stuff {int a;float b;bool c;string name;}Stuff[] arrayOfStuff; //

Object PoolingCreate a pool of objects to reuseInstantiate as many objects as

Use System.Text.StringBuilder over stringstring str = “1 allocation” + “ 2 allocations”;Each

More Memory OptimizationsReuse temporary buffersIf buffers for data processing are needed every

Other GC Optimizationsfor(;;) instead of foreachforeach on anything but arrays allocates an

Marshalling CostYou can write native pluginsCan be super fast!Can be expensive!Design plugins

Case Study - CachingGetting 20k matrices which transform object from local to

Case Study - CachingCache complex expressionsProperties can hide expensive operationsOptimized implementation: 33.5ms

Case study - CopyingCreate a method using referencesWe had 3 redundant copies

Optimizing GraphicsBake what can be bakedLightingShadowsBatch what can be batchedStatic MeshesMaterialsUI Canvas elements

Optimizing MeshesOnly use as many vertices as you needSet “Read/Write” to false

Optimizing MeshesDisable “Import Blend Shapes” if none are usedDisable “Normals and Tangents”

Combine TexturesTexture Atlases can be made by artists too…

Optimizing TexturesiOS - Use PVRTCAndroidOpenGLES 2.0 devices: ETC1OpenGLES 3.0 devices: ETC2Specific GPUs

Optimizing Textures - ExampleShadowgunUsed “Render to Texel” tool to bake normal-mapped lighting

Optimizing Audio“Force to Mono” if sounds don’t require 3D/StereoLoad Type “Decompress on

Optimizing UIKeep UI elements at the same z-depthDifferent z-depths breaks batchingUse Sprite

Other OptimizationsLimiting Rigidbodies to 2 dimensions in a 2D gameUse Box2D or

Script OptimizationsAvoid Find…() methodsCache a reference insteadFindWithTag() is more optimized but still

Vector Math OptimizationsNormalize a vector once if used over and overNormalization function

Shader OptimizationIn general, less instructions is better*Move calculations to Vertex ShaderHigh DPI

10000 Objects Update() vs Update() 10000 ObjectsBlog Post - http://blogs.unity3d.com/2015/12/23/1k-update-calls/ With

Слайды презентации

Слайд 2 Mobile Optimizations

Слайд 3 Corey Johnson
Product Manager of Doom

Слайд 4 Arm you with more possibilities for optimizations that

you will be able to utilize
Goal

Слайд 5 Agenda

Recognizing Your Performance Bottleneck

Profiling in and out of

Unity

Optimizing Tips

Слайд 6 What do we mean, Performance?
Frametime
CPU usage (Gamecode, Physics,

What do we mean, Performance?FrametimeCPU usage (Gamecode, Physics, Skinning, Particles, …)GPU

Skinning, Particles, …)
GPU usage (Drawcalls, Shader usage, Image effects,

…)
Stalls
Spikes in framerate caused by heavy tasks (e.g. GC.Collect)
Physics world rebuild due to moved static colliders*
Memory
Optimizing memory is very important on device
Avoid GC Hickups by reducing memory activity
Leak detection

Слайд 7 Know Your Bottlenecks
Question: Why are we slow?

Слайд 8 Know Your Bottlenecks
Question: Why are we slow?

CPU or

GPU Bound?
Physics or Rendering?
Update() or FixedUpdate() loop?

Слайд 9 Know Your Bottlenecks
Answer: Always start in the same

place...

Profile
Profile
Profile

Слайд 10 CPU-Heavy Tasks
Physics
Animation
Gameplay code
Runtime GI
Reflection probes

Particles
Creating Batches

Слайд 11 GPU-Heavy Tasks
Switching Batches
Geometry/Pixel shaders
Compute shaders
Skinning

Слайд 12 Profiling in Unity
Unity Profiler
In-Editor
Live Builds on devices
Rapid Iteration
Memory

usage of individual assets
TIPS: Use Deep Profile to see

calls to all methods (including game code) Use BeginSample() EndSample() to minimize overhead

Слайд 13 Custom Profiler Tags
Do this:

Get This:

Слайд 14 Unity Memory Profiler
Open Source https://bitbucket.org/Unity-Technologies/memoryprofiler
Profile memory of games

running on device
TIPS: IL2CPP memory info is better than

Mono Under active development

Слайд 15 Profiling outside of Unity (iOS)
Instruments
Profile game running on iOS

Profiling outside of Unity (iOS)InstrumentsProfile game running on iOS deviceMono &

device
Mono & IL2CPP Builds
TIPS: Best for profiling on-device memory

usage Best for determining method CPU usage

Слайд 16 Profiling outside of Unity (Android)
Unity Profiler
adb
logcat
GPU
Adreno (Qualcomm)
PVRTune, PVRUniSCo (PowerVR)
Intel

Слайд 17 Garbage Collection
Managed Memory
Size doubles when limit is hit
NEVER

Garbage CollectionManaged MemorySize doubles when limit is hitNEVER SHRINKSCan stall when

SHRINKS
Can stall when collected

Can explicitly call System.GC.Collect() during breaks

in game

Слайд 18 Garbage Collection - Stack vs Heap
Heap Objects
Memory block

Garbage Collection - Stack vs HeapHeap ObjectsMemory block allocated on the

allocated on the Heap and must be Garbage Collected

when no longer in use
As Heap expands and contains more objects it takes longer for the GC to scan & clean
Classes, Strings, Arrays, Lists
Stack Objects
Only live within their scope and memory is freed when it goes out of scope
Structs, primitive types

Слайд 19 Data Layout Matters
struct Stuff {
int a;
float b;
bool c;
string

name;
}
Stuff[] arrayOfStuff; //

As;
float[] Bs;
bool[] Cs;
string[] names; //<< Only this is scanned. GC takes less time.

Слайд 20 Object Pooling
Create a pool of objects to reuse
Instantiate

Object PoolingCreate a pool of objects to reuseInstantiate as many objects

as many objects as you’ll need before you need

them
Enable as-needed
Disable, Reset when they’re done
No more Instantiate/Destroy cycle (expensive)
Saves GC from having to run as often
No new memory allocated
Allocate a sensible number of objects
Don’t allocate TOO many objects as they do take up their own memory in the Heap that can’t be reused

Слайд 21 Use System.Text.StringBuilder over string
string str = “1 allocation”

Use System.Text.StringBuilder over stringstring str = “1 allocation” + “ 2

+ “ 2 allocations”;

Each string concatenation allocates multiple objects
Plus

a 3rd for the actual result
Problematic if called in loops, Update(), FixedUpdate(), etc
Use System.Text.StringBuilder
.Append() is faster in loops
Starts with a capacity, increases when it is surpassed in an Append() call. Then it allocates more memory
Mecanim: Use Animator.StringToHash() for release
Can be used for custom code

Слайд 22 More Memory Optimizations
Reuse temporary buffers
If buffers for data

More Memory OptimizationsReuse temporary buffersIf buffers for data processing are needed

processing are needed every frame, allocate the buffer once

and reuse

Don’t use OnGUI
Even empty OnGUI calls are very memory intensive

Don’t .tag == .tag
Use CompareTag()

Слайд 23 Other GC Optimizations
for(;;) instead of foreach
foreach on anything

Other GC Optimizationsfor(;;) instead of foreachforeach on anything but arrays allocates

but arrays allocates an Enumerator (due to old Mono

implementation)
Avoid LINQ functions
Allocates memory for Enumerators, essentially a foreach
Avoid anonymous functions and lambda expressions
Allocates memory if needing to access variables outside its scope
Avoid Boxing value types
Converts them to reference types allocated on the Heap

Слайд 24 Marshalling Cost
You can write native plugins
Can be super

Marshalling CostYou can write native pluginsCan be super fast!Can be expensive!Design

fast!
Can be expensive!
Design plugins carefully to avoid marshalling cost

Can

sneak up on you
gameObject.GetComponent<...>()
Cache your components

Слайд 25 Case Study - Caching

Getting 20k matrices which transform

Case Study - CachingGetting 20k matrices which transform object from local

object from local to camera space

Naive implementation: 125 ms

Слайд 26 Case Study - Caching

Cache complex expressions
Properties can hide

expensive operations

Optimized implementation: 33.5ms

Слайд 27 Case study - Copying

Create a method using references
We

Case study - CopyingCreate a method using referencesWe had 3 redundant

had 3 redundant copies (2 inputs, 1 output)
Matrix4x4 is

a value-type

Optimized implementation: 21.5ms

Слайд 28 Optimizing Graphics
Bake what can be baked
Lighting
Shadows
Batch what can

be batched
Static Meshes
Materials
UI Canvas elements

Слайд 29 Optimizing Meshes
Only use as many vertices as you

Optimizing MeshesOnly use as many vertices as you needSet “Read/Write” to

need
Set “Read/Write” to false if not accessing vertices in

script
Enabled = extra copy in memory
Non-uniform scaling requires read/write
Enable “Optimize Mesh”
Reorder vertex info for fast reading
Always enable ‘Optimize Mesh Data’ in ‘Player Settings->Other Settings’
Removes redundant vertex attributes (tangents, normal, color, etc)

Слайд 30 Optimizing Meshes
Disable “Import Blend Shapes” if none are

Optimizing MeshesDisable “Import Blend Shapes” if none are usedDisable “Normals and

used
Disable “Normals and Tangents” if they won’t be used

by materials
Pre-transform static geometry to world space
Enable Static and Dynamic batching

Слайд 31 Combine Meshes

Слайд 32 Combine Textures
Texture Atlases can be made by artists

too…

Слайд 33 Optimizing Textures
iOS - Use PVRTC
Android
OpenGLES 2.0 devices: ETC1
OpenGLES

Optimizing TexturesiOS - Use PVRTCAndroidOpenGLES 2.0 devices: ETC1OpenGLES 3.0 devices: ETC2Specific

3.0 devices: ETC2
Specific GPUs might handle other formats more

efficiently
UI - for textures that can’t be compressed without fidelity loss use 16-bit texture instead of 32
16-bit Texture Formats
Gradient alpha - RGBA4444
Only cutout alpha - RGBA5551
No alpha - RGB565

Слайд 34 Optimizing Textures - Example
Shadowgun
Used “Render to Texel” tool

Optimizing Textures - ExampleShadowgunUsed “Render to Texel” tool to bake normal-mapped

to bake normal-mapped lighting into textures
https://www.assetstore.unity3d.com/en/#!/content/4153
Saved massive run

time calculations

Слайд 35 Optimizing Textures - Example

Слайд 36 Optimizing Audio
“Force to Mono” if sounds don’t require

Optimizing Audio“Force to Mono” if sounds don’t require 3D/StereoLoad Type “Decompress

3D/Stereo
Load Type “Decompress on Load” if clip smaller than

200kb
Unity uses 200kb playback buffer when decompressing audio so leave it decompressed. Saves memory when playing the sound.
“Stream to Disk” for long audio clips
Only 1 clip at a time
Buffers compressed data
Decodes on the fly
Uses minimal memory
“Compressed in Memory” for other clips

Слайд 37 Optimizing UI
Keep UI elements at the same z-depth
Different

Optimizing UIKeep UI elements at the same z-depthDifferent z-depths breaks batchingUse

z-depths breaks batching
Use Sprite Packer
Fewer draw calls for Sprites
Separate

UI into several Canvases (but not too many)
Batch time grows more than linearly by # of elements to sort, analyze
Combine UI that doesn’t change
Canvas won’t need to be rebatched
Reduce switching between overlapping Text and Sprites
Reduce text in UI if possible
Text is batched separately from Sprites

Слайд 38 Other Optimizations
Limiting Rigidbodies to 2 dimensions in a

Other OptimizationsLimiting Rigidbodies to 2 dimensions in a 2D gameUse Box2D

2D game
Use Box2D or roll your own
Doesn’t pull in

whole physics system(s)
Rigidbodies on projectiles
Calculate collision on your own
Lots of individual 3D objects for collectables or characters
Use animated sprites on particles to represent simple objects
Perform expensive calculations every few frames and cache the results
Coroutines (maybe)

Слайд 39 Script Optimizations
Avoid Find…() methods
Cache a reference instead
FindWithTag() is

Script OptimizationsAvoid Find…() methodsCache a reference insteadFindWithTag() is more optimized but

more optimized but still not as fast
Use Non-allocating functions
i.e.

pass array as parameter to fill instead of allocating and returning a new one
Unity’s Physics system has examples of non-allocating functions
i.e. Physics2D.RaycastNonAlloc()
public static int RaycastNonAlloc(Vector2 origin, Vector2 direction, RaycastHit2D[] results)

Слайд 40 Vector Math Optimizations
Normalize a vector once if used

Vector Math OptimizationsNormalize a vector once if used over and overNormalization

over and over
Normalization function takes longer than just storing

and accessing it
v.normalized slower than v * 1.0/v.length
Use Vector’s .sqrMagnitude to compare distances instead of getting the actual distance
Saves some calculations

Слайд 41 Shader Optimization
In general, less instructions is better*
Move calculations

Shader OptimizationIn general, less instructions is better*Move calculations to Vertex ShaderHigh

to Vertex Shader
High DPI devices make every pixel count
Simplify

math
Trig functions are super expensive
Bake into lookup textures
Reduce temporary registers used
Number of shader threads that can work simultaneously depends on this

Слайд 42 10000 Objects Update() vs Update() 10000 Objects
Blog Post

- http://blogs.unity3d.com/2015/12/23/1k-update-calls/ With Sample Project - https://github.com/valyard/Unity-Updates/commits/master By Unity’s Valentin

Simonov
Much faster to run a function on 10000 objects from a single manager GameObject’s Update() method
Due to remaining on the Managed side. Native → Managed call to Update and various safety checks Unity does internally makes Update() on 10000 objects slow