Mobile Game Optimization with Vulkan and Unity Adaptive Performance

Powered by Samsung GameSDK

Presented by Samsung and Unity

Vulkan Optimization

GameDev Engineer

Samsung Electronics

shawn1.lee@samsung.com

Samsung

Adaptive Performance

David Berger

Software Developer

Unity Technologies

db@unity3d.com

GameSDK

GameDev Engineer

Samsung Electronics

junsik.kong@samsung.com3

GameDevGALAXY GAMEDEV

What Samsung did…

• Helping developers with adoption of new technologies

GameDev

Team GameDev

•→

•→

•→

•→

•→

Engine/Game

GPU Driver

GameEngines

DeveloperSupport

Tools & SDKs

Vulkan Grahpics API

• Way of Developing High Fidelity Android Games

Vulkan Grahpics API

•

•

•

•

Application

OpenGL Driver

GPU

Vulkan Driver

GPU

Application

Vulkan

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

• GameEngine Optimize

• Contents Optimize

• Trouble Support

• Pipeline Barrier Optimize

• Pending RenderPass

• Primitive Buffer

Before After

42 FPS 60 FPS

S8 Mali — Within Sustainable Power

• GameEngine Optimize

• Contents Optimize

• Trouble Support

• Pipeline Barrier Optimize

• Pending RenderPass

• Primitive Buffer

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

Before After

42 FPS 60 FPS

S8 Mali — Within Sustainable Power

• GameEngine Optimize

• Contents Optimize

• Trouble Support

• Pipeline Barrier Optimize

• Pending RenderPass

• Primitive Buffer

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

Before After

42 FPS 60 FPS

S8 Mali — Within Sustainable Power

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

Vertex Shader

…

Fragment Shader

SRC VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT

DST VK_PIPELINE_STAGE_VERTEX_SHADER_BIT VK_PIPELINE_STAGE_VERTEX_INPUT_BIT

Pipeline Barrier, Wait - !

Vertex Shader

…

Fragment Shader

Vertex Shader

…

Fragment Shader

• Pipeline Barrier Optimize

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

• Vulkan with proper barriers

• Vulkan with wrong barriers

• GameEngine Optimize

• Contents Optimize

• Trouble Support

• Pipeline Barrier Optimize

• Pending RenderPass

• Primitive Buffer

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

Before After

42 FPS 60 FPS

S8 Mali — Within Sustainable Power

FINAL FANTASY XV POCKET EDITIONSquare Enix / Summertime Studio

• Pending RenderPass

Hundred SoulHOUND 13

• Bloom Optimize

• RenderPass Load/Store Optimize

• Paging Allocate Buffer

• Batching UpdateDescriptorSet

• Shader Module Cache

Before After

41 FPS 52 FPS

Note9 Adreno — Within Sustainable Power

Hundred SoulHOUND 13

• Bloom Optimize

• RenderPass Load/Store Optimize

• Paging Allocate Buffer

• Batching UpdateDescriptorSet

• Shader Module Cache

Before After

41 FPS 52 FPS

Note9 Adreno — Within Sustainable Power

Hundred SoulHOUND 13

• Bloom Optimize

• RenderPass Load/Store Optimize

• Paging Allocate Buffer

• Batching UpdateDescriptorSet

• Shader Module Cache

Before After

41 FPS 52 FPS

Note9 Adreno — Within Sustainable Power

Hundred SoulHOUND 13

• RenderPass Load/Store Optimize

https://community.arm.com/developer/tools-software/graphics/b/blog/posts/the-mali-gpu-an-abstract-machine-part-4---the-bifrost-shader-core

Honor of KingsTencent Timi Studio – L1

• Single Scratch Buffer

• Low Priority Destroy Thread

• Batching UpdateDescriptorSet

• RenderPass Load/Store Optimize

• Shader Module Cache

Before After

41 FPS 53 FPS

S7 Adreno — Within Sustainable Power

• Single Scratch Buffer

• Low Priority Destroy Thread

• Batching UpdateDescriptorSet

• RenderPass Load/Store Optimize

• Shader Module Cache

Before After

41 FPS 53 FPS

S7 Adreno — Within Sustainable Power

Honor of KingsTencent Timi Studio – L1

Honor of KingsTencent Timi Studio – L1

• Only recreate during 1st loading

• 100% hit rate during playingoffset offset offset

Free Memory

Not enough memory

Larger buffer

Honor of KingsTencent Timi Studio – L1

• Single Scratch Buffer

• Low Priority Destroy Thread

• Batching UpdateDescriptorSet

• RenderPass Load/Store Optimize

• Shader Module Cache

Before After

41 FPS 53 FPS

S7 Adreno — Within Sustainable Power

Honor of KingsTencent Timi Studio – L1

• Single Scratch Buffer

• Low Priority Destroy Thread

• Batching UpdateDescriptorSet

• RenderPass Load/Store Optimize

• Shader Module Cache

Before After

41 FPS 53 FPS

S7 Adreno — Within Sustainable Power

Honor of KingsTencent Timi Studio – L1

● 매 draw 마다 UpdateDescriptorSets호출● Texture, Buffer, Buffer Offset, Buffer View 변경을위해호출

vkUpdateDescriptorSets

vkCmdDraw

vkUpdateDescriptorSets

vkCmdDraw

...

vkQueueSummit

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

• VB, IB binding optimization

• Direct buffer access

• Cache optimization

• Async texture load

• Single buffer for image upload

Before After

49 FPS 59 FPS

Note8 Mali — Within Sustainable Power

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

• VB, IB binding optimization

• Direct buffer access

• Cache optimization

• Async texture load

• Single buffer for image upload

Before After

49 FPS 59 FPS

Note8 Mali — Within Sustainable Power

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

– Run thread for vkQueuePresent()

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

• VB, IB binding optimization

• Direct buffer access

• Cache optimization

• Async texture load

• Single buffer for image upload

Before After

49 FPS 59 FPS

Note8 Mali — Within Sustainable Power

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

Skip binding buffers, just draw with bound

buffers from previous draw call.

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

• VB, IB binding optimization

• Direct buffer access

• Cache optimization

• Async texture load

• Single buffer for image upload

Before After

49 FPS 59 FPS

Note8 Mali — Within Sustainable Power

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Direct Buffer Access

Data(vertex, index,

uniform…)

Main Thread

Temp Buffercopyupload

Render Thread

Real Buffer

Before

Data(vertex, index,

uniform…)

Main Thread

upload

Render Thread

Real Buffersync

After

Honor of Kings – The King’s ExpeditionTencent Timi Studio – L1

• Threaded queue present

• VB, IB binding optimization

• Direct buffer access

• Cache optimization

• Async texture load

• Single buffer for image upload

Before After

49 FPS 59 FPS

Note8 Mali — Within Sustainable Power

Collaboration Titles

• 2017 GameLoft, Asphalt 8

• 2017 Cornfox&Bros, Oceanhorn: Monster of Uncharted Seas

• 2017 Deep Silver, Galaxy On Fire 3: Manticore!

• 2017 Digital Legends Entertainment, Afterpulse

• 2017 First Touch Games, Score! Hero / Dream League Soccer

• 2018 Tencent, Honor of Kings

• 2018 Pearl Abyss, BlackDesert Mobile

• 2018 Epic Games, Fortnite Battle Royale

• 2018 Amazon Lumberyard, Bistro

• 2019 Moai Games, TRAHA

• 2019 PUBG corporation / Tencent, PUBG MOBILE

• 2019 Tencent, QQ Speed

• 2016 Epic Games, ProtoStar Galaxy S7 Collaboration

• 2016 NetGames, HIT

• 2016 Super Evil MegaCorp, VainGlory

• 2017 433, HeroDC

• 2017 Netmarble, Lineage 2: Revolution

• 2017 Nexon, AxE

• 2017 XL Games, Archeage: Begins

• 2017 Action Square, Blade II

• 2017 Hound13, HundredSoul

• 2017 Square Enix, FinalFantasy XV Pocket Edition

• 2017 Croteam, Talos Principle

• 2017 Roblox Corporation, Roblox

• 2017 Doragon Entertainment, Danmaku Unlimited 3

Contact

gamedev@samsung.com

https://developer.samsung.com/game

Samsung GameSDK

Presented by Junsik Kong

Samsung Galaxy GameDev

Limitation of Mobile

—

– No active cooling system

– Performance Throttling

—

– Limitation of battery

– Use more Power than necessary

— Developer can’t control

FPS

Core Freq

Temperature After Throttling

43

GameApplicationunity

Galaxy GameSDK

44

Thermal Throttling

40 fps @ end

60 fps @ start

Throttling ?

45

— Performance drops and shakes

Thermal Throttling

— onHighTempWarning(int warningLevel)

warningLevel 0 warningLevel 1 warningLevel 2

Throttling46

Thermal Throttling

— getTempLevel(), getCPU/GPUJTLevel(), getSkinTempLevel()

01

23

45

6 7

47

Thermal Throttling

— Now you can do !

– Thermal Throttling 시점예측가능

– Thermal Throttling 방지

– Adjust Target FPS, Quality or etc!

50 fps steady

48

Power control

CPU Freq

GPU Freq

— Frequency over reat

49

Power control

— setLevelWithScene(String scene, int cpuLevel, int gpuLevel)

CPU Freq

GPU Freq

Level 2

Level 1

50

Bottleneck Check

— Power level

Target FPS 30

@FPS 28

setlevel….(1,1)

setlevel….(2 ,1)

setlevel….(1 ,2)

51

Bottleneck Check

— getGpuFrameTime()

– #n frame’s GpuFrameTime

GPU Bottleneck! GPU Level 증가

16.00 ms10.23ms

#n frame Time

#n gpuframeTime

52

Goal #1 Balancing Quality & Performance

40 fps @ end

60 fps @ start

50 fps steady

53

—

Goal #2 Controlling Power Budge

—

CPU Freq

GPU Freq

Level 2

Level 1

54

Goal #3 Maximizing Peformance

—

–

setlevel….(1,1)55

16.00 ms10.23ms

#n frame Time

#n gpuframeTime

setlevel….(2 ,1)

setlevel….(1 ,2)

55

Adaptive Performance in Megacity

56

Presented by David BergerUnity Technologies

Adaptive Performance

Get device performance status and thermal trends

Proactively adjust performance and quality settings on the fly

Smoother Gameplay

58

59

8MTriangles

6MEntities

2KCars

60

30FPS - stable over time

61

30

15

200 Time [m]

Framerate

62

Adaptive Performance

• Start with Low CPU/GPU levels (Menu->Level)

– Increase levels for CPU or GPU bottleneck respectively

– Keeps energy consumption low

• Not hitting target frame rate

– Decrease the LOD bias

– Decreases GFX load (triangles, memory, draw calls)

• Decrease target framerate when close to throttling

63

30

15

200 10

Time [m]

64

Adaptive Performance: How it works

1-Add the package

66

Adaptive Performance: How it works

2-Register to OnThermalEvent, Set initial target level

67

Adaptive Performance: How it works

3-Manage your quality settings when Thermal status changes

68

Adaptive Performance: How it works

4-You can subscribe to Performance bottleneck events

69

Adaptive Performance: How it works

4-You can subscribe to performance bottleneck events

70

Adaptive Performance Preview

available now!

71

Adaptive Performance Blost Posthttps://blogs.unity3d.com/2019/04/01/higher-fidelity-and-smoother-

frame-rates-with-adaptive-performance/

72

73

Roadmap

• Preview

• 2019.1+

• Verified

• 2019.3

• Future

• 2019.3+

74

Questions?db@unity3d.com

75