Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Evaluation of Antialiasing Techniques on Mobile...
Search
Tiago Almeida
November 21, 2019
Programming
0
19
Evaluation of Antialiasing Techniques on Mobile Devices
Tiago Almeida
November 21, 2019
Tweet
Share
More Decks by Tiago Almeida
See All by Tiago Almeida
Awesome features of RecyclerView
talmeida99
1
120
Other Decks in Programming
See All in Programming
ソフトウェア設計の実践的な考え方
masuda220
PRO
3
500
XP, Testing and ninja testing ZOZ5
m_seki
3
330
10年もののAPIサーバーにおけるCI/CDの改善の奮闘
mbook
0
790
Advance Your Career with Open Source
ivargrimstad
0
370
iOSアプリの信頼性を向上させる取り組み/ios-app-improve-reliability
shino8rayu9
0
150
AI Coding Meetup #3 - 導入セッション / ai-coding-meetup-3
izumin5210
0
610
Goで実践するドメイン駆動開発 AIと歩み始めた新規プロダクト開発の現在地
imkaoru
4
700
非同期jobをtransaction内で 呼ぶなよ!絶対に呼ぶなよ!
alstrocrack
0
550
The Past, Present, and Future of Enterprise Java
ivargrimstad
0
170
Building, Deploying, and Monitoring Ruby Web Applications with Falcon (Kaigi on Rails 2025)
ioquatix
1
270
そのpreloadは必要?見過ごされたpreloadが技術的負債として爆発した日
mugitti9
2
3.1k
階層構造を表現するデータ構造とリファクタリング 〜1年で10倍成長したプロダクトの変化と課題〜
yuhisatoxxx
3
930
Featured
See All Featured
Sharpening the Axe: The Primacy of Toolmaking
bcantrill
45
2.5k
Context Engineering - Making Every Token Count
addyosmani
5
190
Code Reviewing Like a Champion
maltzj
525
40k
ピンチをチャンスに:未来をつくるプロダクトロードマップ #pmconf2020
aki_iinuma
127
53k
Dealing with People You Can't Stand - Big Design 2015
cassininazir
367
27k
ReactJS: Keep Simple. Everything can be a component!
pedronauck
667
120k
Facilitating Awesome Meetings
lara
56
6.6k
YesSQL, Process and Tooling at Scale
rocio
173
14k
Making the Leap to Tech Lead
cromwellryan
135
9.5k
Measuring & Analyzing Core Web Vitals
bluesmoon
9
610
How STYLIGHT went responsive
nonsquared
100
5.8k
Build The Right Thing And Hit Your Dates
maggiecrowley
37
2.9k
Transcript
Evaluation of Antialiasing Techniques on Mobile Devices Tiago Almeida, João
Madeiras and João Jorge ICGI 2019
Motivation 2 Horizon Zero Dawn
Motivation 3 Horizon Zero Dawn Source: VentureBeat
Outline • Goals • Antialiasing techniques • Methodology • Results
• Conclusions • Demo 4
Goals • Study state of the art antialiasing solutions •
Compare the techniques on a mobile device • Develop a mobile application for benchmarks 5
Goals • Which antialiasing techniques are faster? • Which one
obtains the best image quality? • What should be explored on mobile antialiasing? 6
Aliasing 7
Antialiasing techniques • Multisampling • FXAA • Morphological AA •
Temporal AA 8
Multisampling Antialiasing (MSAA) 9
Fast Approximate Antialiasing (FXAA) • Post-processing technique • Based on
the contrast among neighbors • Blend colors to prevent discontinuities 10
Morphological Antialiasing (SMAA) • Identify perceptually visible edges • Re-vectorize
edges to calculate area • Blend with neighbors according to area 11
Temporal Antialiasing 12 History Temporal pass • Use frame history
information • Camera frustum jitter • Reprojection using velocity buffer
Methodology 13
Methodology 14 150k vertices Textures 1 million vertices Crytek Sponza
Hairball
Methodology • Collect performance metrics using Snapdragon Profiler ◦ GPU
utilization, memory, temperature • Compare image results with Structural Similarity Index (SSIM) • Analyze image details produced by each technique 15
Results - Framerate 16 Crytek Sponza (FPS)
Results - Framerate 17 Hairball (FPS)
Results - GPU Memory 18 Average Memory Read per Frame
(MB)
Performance results • FXAA obtained best results • MSAA slower
on high-polygon • SMAA with temporal reprojection is the slowest 19
Results - Image Quality 20 MSAA 4x FXAA SMAA 1x
SMAA T2x TAA No AA
Results - Structural Similarity Index (SSIM) 21
Image quality results • SMAA and MSAA with best quality
• FXAA blurs the details 22
Demo 23
Conclusions • Multisampling and FXAA have better performance • SMAA
with temporal reprojection has better image quality • MSAA provides the best balance 24
Thank you Tiago Almeida
[email protected]
25
Results - GPU Shader Processing 26 Fragment Instructions per Frame
(in millions)
Results - GPU Shader Processing 27 No AA MSAA 4x
FXAA SMAA 1x SMAA T2x TAA % Shaders Busy 91.4 79.3 92.1 92.3 87.1 85.6 % Time ALU working 53.6 43.6 53.6 50.7 49.5 52.3 % Time Shading Fragments 37.4 38.7 41.0 48.8 53.2 44.8 Mean values on Crytek Sponza
Summary • Research novel antialiasing techniques • Created an Android
application to test them • SMAA obtains the best image quality • MSAA and FXAA best performance 28