最小限の実装で試す MetalによるGPUコンピューティング

by yuky_az

Slide 1

Slide 1 text

࠷খݶͷ࣮૷Ͱࢼ͢ .FUBMʹΑΔ(16ίϯϐϡʔςΟϯά :VLJOBHB"[VNB !ZVLZ@B[ QPUBUPUJQT

Slide 2

Slide 2 text

ࣗݾ঺հ :VLJOBHB"[VNB  ʢզ࠺޾௕ʣ J04ΤϯδχΞϓϩάϥϛϯάΠϯετϥΫλʔ !ZVLZ@B[ J0%$Ͱ͸εϚϗ޲͚73ΞϓϦͷ։ൃʹ͍ͭͯൃද 6EFNZͰʮΈΜͳͷ"*ߨ࠲ʯ։ߨத

Slide 3

Slide 3 text

Metalͱ͸ʁ Φʔόʔϔουͷখ͍͞ϩʔϨϕϧͳ ίϯϐϡʔλάϥϑΟοΫε"1* 0QFO(-ͱ0QFO$-ʹࣅͨػೳΛҰͭͷ"1*ʹ౷߹

Slide 4

Slide 4 text

ࠓճͷൃද಺༰ 1. ඳըॲཧ 2. ฒྻԋࢉ MetalʹΑΔ

Slide 5

Slide 5 text

GPUฒྻԋࢉͱ͸ʁ GPUͷԋࢉϦιʔεΛը૾ॲཧҎ֎ʹར༻

Slide 6

Slide 6 text

CPUͱGPUͷ࿈ܞ EFWFMPQFSBQQMFDPNNFUBM

Slide 7

Slide 7 text

Metal CPUଆͷొ৔ਓ෺ 1 .5-%FWJDF୯Ұͷ(16ͷΠϯλʔϑΣΠε .FUBMؔ࿈ͷ"1*͸ଟ͕͘ϓϩτίϧͰهࡌ EFWJDF.5-$SFBUF4ZTUFN%FGBVMU%FWJDF QSJOUEFWJDFOBNF"QQMF"(16

Slide 8

Slide 8 text

Metal CPUଆͷొ৔ਓ෺ 2 .5--JCSBSZγΣʔμʔ΁ͷΠϯλʔϑΣΠε .5-'VODUJPO.FUBM4IBEJOH-BOHVBHFͰॻ͔Εͨؔ਺ .5-$PNQVUF1JQFMJOF4UBUF.5-'VODUJPOΛίϯύΠϧ͞Εͨίʔυʹม׵ .5-$PNQVUF$PNNBOE&ODPEFSίϚϯυΛ(16༻ʹΤϯίʔυ .5-$PNNBOE#V⒎FS(16Ͱ࣮ߦ͞ΕΔίϚϯυΛ֨ೲ .5-$PNNBOE2VFVF্هͷΩϡʔɺίϚϯυͷ࣮ߦॱΛ؅ཧ .5-%FWJDF ੜ੒

Slide 9

Slide 9 text

Metal Shading Language JODMVEFNFUBM@TUEMJC VTJOHOBNFTQBDFNFUBM LFSOFMWPJEBEE0OF DPOTUEFWJDFqPBUJOQVU%BUB<<CV⒎FS >> EFWJDFqPBUPVUQVU%BUB<<CV⒎FS >> VJOUJE<<UISFBE@QPTJUJPO@JO@HSJE>> \ qPBUSFTVMUJOQVU%BUB SFTVMUG PVUQVU%BUBSFTVMU ^

Slide 10

Slide 10 text

GPU vs CPU ɾೖྗʹΛ଍ͯ͠Ҿ͘Λ܁Γฦ͢ ɾ$16ଆͰ͸ɺ্هΛສY/ճ ɾ(16ଆͰ͸ສεϨουY/ճ LFSOFMWPJEBEE"OE4VCUSBDU DPOTUEFWJDFqPBUJOQVU%BUB<<CV⒎FS >> EFWJDFqPBUPVUQVU%BUB<<CV⒎FS >> VJOUJE<<UISFBE@QPTJUJPO@JO@HSJE>> \ qPBUSFTVMUJOQVU%BUB GPS VJOUJJ/J \ SFTVMUG SFTVMUG ^ PVUQVU%BUBSFTVMU ^

Slide 11

Slide 11 text

CPU vs GPU ഒఔ౓

Slide 12

Slide 12 text

CPU vs GPU ʢର਺ʣ $16(16ؒͷ΍ΓͱΓΛ࠷খݶʹ͢Δ͜ͱ͕େࣄ

Slide 13

Slide 13 text

࣮૷ํ๏ɺಛੑ͕෼͔͖ͬͯͨͱ͜ΖͰ Կ͔Ԡ༻ͯ͠ΈΑ͏ʂ

Slide 14

Slide 14 text

܈஌ೳͱ͸ʁ ਓ޻஌ೳͷҰछɺूஂͱͯ͠ߴ౓ͳৼΔ෣͍

Slide 15

Slide 15 text

܈Εͷ3ͭͷϧʔϧ B౳ڑ཭֤ݸମ͕౳ڑ཭ΛอͭΑ͏ʹಈ͘ Cฒߦ֤ݸମ͕ଞͷݸମͱಉ͡޲͖ʹͳΔΑ͏ʹಈ͘ D౳଎֤ݸମ͸ଞͷݸମͱಉ͡଎౓ʹͳΔΑ͏ʹಈ͘

Slide 16

Slide 16 text

਺ࣜԽ B౳ڑ཭ Cฒߦ D౳଎౓

Slide 17

Slide 17 text

γΣʔμʔͷίʔυ LFSOFMWPJENPWF ʜ GPS VJOUJJOPEF$PVOUJ \ ʜ ౳ڑ཭ qPBUBUUSBDUJPOFYQ C EJTUBODFB EJTUBODFB qPBUSFQVMTJPOFYQ CEJTUBODFEJTUBODF E"OHMFBMQIB OFBS"OHMFBUUSBDUJPOGBS"OHMFSFQVMTJPO JOUFSWBM ฒߦ qPBUQBSBMMFM"OHMF%JGHFU3BOHFE"OHMF OPEFBOHMFDVSSFOU/PEFBOHMF E"OHMFCFUBQBSBMMFM"OHMF%JGFYQ CEJTUBODFEJTUBODF JOUFSWBM ౳଎౓ qPBUOPEF7FMPDJUZTRSU OPEFWFMPDJUZ9OPEFWFMPDJUZ9OPEFWFMPDJUZ:OPEFWFMPDJUZ: WFMPDJUZHBNNB OPEF7FMPDJUZWFMPDJUZ FYQ CEJTUBODFEJTUBODF ^

Slide 18

Slide 18 text

ߏ੒ ʢ6**NBHF7JFXͷαϒΫϥεʣ 7JFX$POUSPMMFSTXJGU .FUBM.BOBHFSTXJGU 4IBEFSNFUBM

Slide 19

Slide 19 text

࣮ԋ case1 IUUQTZPVUVCFD"60+XNF6" ЋЌЍ

Slide 20

Slide 20 text

࣮ԋ case2 IUUQTZPVUVCF(KES:&3:,7D ЋЌЍ

Slide 21

Slide 21 text

࣮ԋ case3 IUUQTZPVUVCF8M.J+%[$G ЋЌЍ

Slide 22

Slide 22 text

ύϑΥʔϚϯεʹ͍ͭͯ ͷݸମΛ࢖༻ɺͦΕͧΕͷ૬ޓ࡞༻Λߟ͑ͨͷͰ YͰճͷԋࢉΛຖϑϨʔϜ͝ͱ ඵؒʹճ࣮ߦ GQTΛΩʔϓɺ$16઎༗཰ఔ౓ FYQؔ਺͚ͩͰສճඵʢʣ

Slide 23

Slide 23 text

·ͱΊ ຊ೔ͷൃද಺༰ɺৄࡉ͸2JJUBͰ .FUBMͰϋΠύϑΥʔϚϯεͳ(16ฒྻԋࢉ͕Մೳ ෳࡶͳݱ৅ΛJ1IPOF಺ʹ࠶ݱͰ͖Δ ਓ޻஌ೳͳͲ΁ͷԠ༻ɺظ଴େ IUUQRJJUBDPNZVLZ@B[