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
Go言語でMac GPUプログラミング
Search
monochromegane
December 18, 2023
Programming
1
690
Go言語でMac GPUプログラミング
2023.12.18 Fukuoka.go#19 Reboot
https://fukuokago.connpass.com/event/302717/
monochromegane
December 18, 2023
Tweet
Share
More Decks by monochromegane
See All by monochromegane
Go言語での実装を通して学ぶLLMファインチューニングの仕組み / fukuokago22-llm-peft
monochromegane
0
130
不確実性下における目的と手段の統合的探索に向けた連続腕バンディットの応用 / iot70_gp_rff_mab
monochromegane
2
180
なめらかなシステムと運用維持の終わらぬ未来 / dicomo2025_coherently_fittable_system
monochromegane
0
3.6k
ベクトル検索システムの気持ち
monochromegane
37
11k
Go言語での実装を通して学ぶ、高速なベクトル検索を支えるクラスタリング技術/fukuokago-kmeans
monochromegane
1
210
Go言語でターミナルフレンドリーなAIコマンド、afaを作った/fukuokago20_afa
monochromegane
2
290
多様かつ継続的に変化する環境に適応する情報システム/thesis-defense-presentation
monochromegane
1
1k
Online Nonstationary and Nonlinear Bandits with Recursive Weighted Gaussian Process
monochromegane
0
700
AIを前提とした体験の実現に向けて/toward_ai_based_experiences
monochromegane
2
1k
Other Decks in Programming
See All in Programming
uniqueパッケージの内部実装を支えるweak pointerの話
magavel
0
930
After go func(): Goroutines Through a Beginner’s Eye
97vaibhav
0
240
タスクの特性や不確実性に応じた最適な作業スタイルの選択(ペアプロ・モブプロ・ソロプロ)と実践 / Optimal Work Style Selection: Pair, Mob, or Solo Programming.
honyanya
3
140
Domain-centric? Why Hexagonal, Onion, and Clean Architecture Are Answers to the Wrong Question
olivergierke
1
600
どの様にAIエージェントと 協業すべきだったのか?
takefumiyoshii
2
620
(Extension DC 2025) Actor境界を越える技術
teamhimeh
1
240
実践AIチャットボットUI実装入門
syumai
7
2.5k
GraphQL×Railsアプリのデータベース負荷分散 - 月間3,000万人利用サービスを無停止で
koxya
1
1.2k
チームの境界をブチ抜いていけ
tokai235
0
100
Back to the Future: Let me tell you about the ACP protocol
terhechte
0
130
なぜあの開発者はDevRelに伴走し続けるのか / Why Does That Developer Keep Running Alongside DevRel?
nrslib
3
380
育てるアーキテクチャ:戦い抜くPythonマイクロサービスの設計と進化戦略
fujidomoe
1
160
Featured
See All Featured
Building Adaptive Systems
keathley
43
2.8k
"I'm Feeling Lucky" - Building Great Search Experiences for Today's Users (#IAC19)
danielanewman
229
22k
Principles of Awesome APIs and How to Build Them.
keavy
127
17k
Design and Strategy: How to Deal with People Who Don’t "Get" Design
morganepeng
132
19k
Navigating Team Friction
lara
189
15k
Responsive Adventures: Dirty Tricks From The Dark Corners of Front-End
smashingmag
252
21k
Helping Users Find Their Own Way: Creating Modern Search Experiences
danielanewman
30
2.9k
Being A Developer After 40
akosma
91
590k
Documentation Writing (for coders)
carmenintech
75
5k
How STYLIGHT went responsive
nonsquared
100
5.8k
Put a Button on it: Removing Barriers to Going Fast.
kastner
60
4k
Faster Mobile Websites
deanohume
310
31k
Transcript
Lightning Talks ࡾ༔հ / Pepabo R&D Institute, GMO Pepabo, Inc.
2023.12.18 Fukuoka.go#19 Reboot GoݴޠͰMac GPUϓϩάϥϛϯά
ϓϦϯγύϧΤϯδχΞ ࡾ ༔հ / @monochromegane 2 https://blog.monochromegane.com Yusuke Miyake ϖύϘݚڀॴ
ݚڀһ
• ଟมྔਖ਼ن ʹै͏ཚੜʹ͕͔͔࣌ؒΔ • ͜ͷཚੜͷखॱʢͷҰͭʣҎԼͷ௨Γ 1. ֤ཁૉ͕ඪ४ਖ਼نʹै͏ཚ ΛಘΔ 2. ڞࢄߦྻ
ΛίϨεΩʔղʢ ʣͯ͠ࡾ֯ߦྻ ΛಘΔ 3. ΛٻΊΔ • ಛʹɺ֬ͷύϥϝʔλʢ ͱ ʣ͕ҟͳͬͨΓɺ࣍ݩ ͕େ͖͍ ߹ʹɺཚੜʹ͕͔͔࣌ؒͬͯ͠·͏ y ∼ 𝒩 (μ, Σ), μ ∈ ℝD, Σ ∈ ℝD×D z = {zi }1≤i≤D , zi ∼ 𝒩 (0,1) Σ Σ = LL⊤ L y = μ + Lz μ Σ D 3 ͡Ίʹ
• ߦྻܭࢉಠཱ͔ͭฒߦͨ͠λεΫΛଟؚ͘ΉͨΊɺߴԽʹฒྻԽ͕༗ޮ • ͢ͳΘͪɺSIMDɺCPUͷϚϧνίΞɺGPUͳͲʹΑΔฒྻԽ • CPUόϯυͰλεΫཻখ͍͞ͷͰgoroutine͔ͳ͍ʢͱࢥ͏ʣ • GoݴޠͰͷߦྻܭࢉϥΠϒϥϦGonumCPUͷϚϧνίΞΛαϙʔτ͢Δ BLASͷόΠϯσΟϯάΛఏڙ͍ͯ͠Δ •
Apple silicon (M1) ʹGPU͕ࡌ͞Ε͍ͯΔͷͰɺͦͪΒ׆༻͍ͨ͠ 4 ͡Ίʹ
• GPUͷΞΫηεΛఏڙ͢ΔOSඪ४ࡌͷϑϨʔϜϫʔΫ • άϥϑΟοΫεॲཧҎ֎ʹɺGPU্Ͱͷฒྻܭࢉॲཧѻ͑Δ • Objective-C·ͨSwift͔ΒɺGPU্ͷॲཧΛهड़ͨ͠γΣʔμʔؔΛݺͿ • γΣʔμʔؔC++ϕʔεͷMetal Shader Language
(MSL) Ͱهड़ • Metal Performance Shaders (MPS) ͱ͍͏γΣʔμʔؔ܈ఏڙ͞ΕΔ 5 Metal: MacͰGPUϓϩάϥϛϯά
6 Metal: MacͰGPUϓϩάϥϛϯά • جຊతͳྲྀΕɺσόΠεʢGPUʣͷίϚϯυΩϡʔʹର͠ɺίϚϯυόο ϑΝͱ͍͏୯ҐͰγΣʔμʔؔΛొ͠ɺ݁ՌΛड͚औΔͱ͍͏ͷ • ͳ͓ɺCPUͱGPUͷͷΓऔΓʹઐ༻ͷόοϑΝ͕༻ҙ͞Ε͍ͯΔ ίϚϯυΩϡʔ ͷ४උ
ΓऔΓ༻ͷ όοϑΝͷ४උ όοϑΝͷσʔλ͔Β ߦྻΠϯελϯεੜ .14ͷγΣʔμʔؔ ΛॳظԽɺίϚϯυ όοϑΝͱͯ͠Τϯ ίʔυɺΩϡʔʹొ όοϑΝ͔Β݁Ռͷड ͚औΓ 0CKFDUJW$Ͱͷ ࣮ྫ
• Goݴޠ͔ΒcgoΛ͑͜ΕΒͷObjective-CͷίʔυΛݺΔ༷ࢠ • https://github.com/a-h/gpu ϥΠϒϥϦͱͯ͠ར༻Ͱ͖Δ͕MPSʹରԠ͍ͯ͠ͳ͍ • https://github.com/mikecvet/go-mm MPSͷݺͼग़͠Λ࣮͍ͯ͠Δ͕ϕϯνϚʔΫͷίʔυͷΈ • ্هΛࢀߟʹͭͭ͠ɺGoݴޠ্ͰͷGPUΛ༻͍ͨଟมྔਖ਼نʹै͏ཚ
ੜ͕Ͱ͖ͦ͏ 7 Cgo: GoݴޠͰMac GPUϓϩάϥϛϯά
1. Objective-CͷϔομϑΝΠϧΛinclude͠ɺLDFLAGSʹMetalϑϨʔϜϫʔΫΛࢦఆ͢Δ 2. ʢඞཁʹԠͯ͡ʣࣗલͷγΣʔμʔؔΛgo:embedͰΈࠐΜͰ͓͘ 3. C.xxͱͯ͠Objective-CͰهड़ͨ͠ॳظԽγΣʔμʔؔΛ࣮ߦ͢ΔؔΛݺͿɻ࣮ߦ࣌ ͷύϥϝʔλ݁ՌunsafeύοέʔδΛͬͯΞΫηεɻ 8 Cgo: GoݴޠͰMac
GPUϓϩάϥϛϯά (PͰͷ࣮ྫ
• Goݴޠ্ͰͷGPUΛ༻͍ͨଟมྔਖ਼نʹै͏ཚੜ • Goͷίʔυ͔Β ΛcgoΛܦ༝ͯ͠Objective-Cͷؔʹ͢ • MPSͷMPSMatrixDecompositionCholeskyΛ༻͍ͯίϨεΩʔղ • ࣗલγΣʔμʔؔΛ༻͍ͯԼࡾ֯ߦྻҎ֎Λ0ʹຒΊΔ •
MPSͷMPSMatrixVectorMultiplicationΛ༻͍ͯ Λܭࢉ • MPSͷMPSMatrixSumΛ༻͍ͯ Λܭࢉ • GoͷίʔυͰ݁ՌΛड͚औΔ z, μ, Σ Lz μ + Lz 9 Cgo: GoݴޠͰMac GPUϓϩάϥϛϯά
• GonumͱMetal࣮ͷ࣮ߦΛൺֱʢ1000࣍ݩʣ 10 ඪ४ਖ਼نཚͷมͷൺֱ BenchmarkTransformNormMetal-8 9 117668310 ns/op BenchmarkTransformNormGonumBLAS-8 55
21494668 ns/op BenchmarkTransformNormGonum-8 21 54288034 ns/op BenchmarkTransformNormCholMetal-8 1140 1070094 ns/op BenchmarkTransformNormCholGonumBLAS-8 15124 78960 ns/op BenchmarkTransformNormCholGonum-8 6712 177368 ns/op • ίϨεΩʔղͷ݁ՌΛผ్͢Α͏ʹͨ͠߹ͷൺֱ • MPSͷίϨεΩʔղগ͍͔͠͠Εͳ͍͕ɺͦͷଞͷࠩԿ͔
• ߦྻʢ1000x1000ʣͱߦྻʢ1000x1000ʣͷࢉΛൺֱ 11 ߦྻࢉͷൺֱ BenchmarkMatrixMultipicationMetal-8 494 2222134 ns/op BenchmarkMatrixMultipicationGonumBLAS-8 60
22063894 ns/op BenchmarkMatrixMultipicationGonum-8 19 59507228 ns/op BenchmarkMatrixVectorMultipicationMetal-8 1497 792244 ns/op BenchmarkMatrixVectorMultipicationGonumBLAS-8 10000 114843 ns/op BenchmarkMatrixVectorMultipicationGonum-8 972 1239177 ns/op • ߦྻʢ1000x1000ʣͱϕΫτϧʢ1000x1ʣͷࢉΛൺֱ • ߦྻಉ࢜ͷΑ͏ͳܭࢉྔͰGPUͷํ͕ߴɻ ͷΑ͏ͳߦྻͱϕΫτϧͷࢉͰ͜ ͷ࣍ݩʹ͓͍ͯGPUҠৡͷΦʔόʔϔουͷํ͕େ͖͔ͬͨͱߟ͑ΒΕΔ Lz
• GoݴޠͰMac GPUϓϩάϥϛϯά͢Δํ๏Λհͨ͠ • ؆қతͳͷൺֱධՁΛ௨ͯ͠ɺ͍ॴͷഽײΛಘΔ͜ͱ͕Ͱ͖ͨ • MPSͷϚχϡΞϧΛಡΉͱχϡʔϥϧωοτϫʔΫͷαϙʔτ͋ΓɺΞΠ σΟΞ࣍ୈͰ໘ന͍͜ͱ͕Ͱ͖ͦ͏ • MPSͷݺͼग़͠ՄೳͳϥΠϒϥϦΛ࡞ͬͯΈ͍ͨ
• Ͳ͏ϝϞϦϦʔΫͯͦ͠͏ͳͷͰͦͷลΓվળ͍ͨ͠ • ࡾGo 12 ·ͱΊ
None