既存の並列化手法を用いたGPGPUプログラミング

طଘͷฒྻԽख๏Λ༻͍ͨ

GPGPU

ϓϩάϥϛϯά

େ ౡ ૱ ࢙

†

ฏ ᖒ ক Ұ

†

ຊ ଟ ߂ थ

†

GPU ͷੑೳ޲্ʹ൐͍ɼGPU ͷੑೳΛ༷ʑͳ༻్ʹ׆༻͢Δ GPGPU ͕஫໨͞Ε͍ͯΔɽGPGPU ͸ಛʹฒྻϓϩάϥϜʹ͓͍ͯ CPU Λ௒͑Δߴ͍ੑೳ͕ظ଴͞ΕΔҰํɼGPGPU ϓϩάϥϛϯ άಛ༗ͷख๏Λ༻͍Δඞཁ͕͋ΔͨΊιϑτ΢ΣΞͷ࡞੒͕༰қͰ͸ͳ͍ɽຊߘͰ͸ GPGPU ϓϩ άϥϛϯάΛ༰қʹ͢Δ̍ͭͷख๏ͱͯ͠طଘͷฒྻԽख๏Λར༻͢Δ͜ͱΛఏҊ͢Δɽ·ͨ۩ମత ͳ࣮૷ʹ޲͚ͯɼۙ೥ར༻͞Ε࢝Ίͨ GPGPU ޲͚ϓϩάϥϛϯάݴޠ CUDA Λར༻͠ɼGPU ্ ͷॲཧΛطଘͷฒྻԽख๏Ͱ͋Δ SIMD ໋ྩɼOpenMPɼMPI Λ༻͍ͯهड़͢Δํ๏Λݕ౼͢Δɽ

GPGPU Programming

Using Existing Parallelizing Method

Satoshi OHSHIMA,

†

_{Shoichi HIRASAWA}

†

_{and Hiroki HONDA}

† GPGPU utilizing GPU’s performance for general-purpose computation is attracting much attention. GPGPU is expected to effect higher performance than CPU. However, creating GPGPU programming is not easy because programming methods peculiar to GPGPU pro-gramming are needed. In this paper, we propose to use existing parallellizing method as one of a new method making GPGPU programming easier. Also we consider writing programs running on GPUs with SIMD instruction, OpenMP and MPI based on the new GPGPU programming language of CUDA.

1. ͸ ͡ Ί ʹ

ۙ೥ɼߴ౓ͳը૾ॲཧͷཁٻʹ൐͍GPU(Graphics

Processing Unit) ͷੑೳ͕ஶ͘͠޲্͍ͯ͠Δ1)_ɽ

GPU ͸ CPU(Central Processing Unit) ͱ ൺ ΂ ͯฒྻॲཧ΍ϕΫτϧॲཧʹదͨ͠ϋʔυ΢ΣΞ ߏ ੒ Ͱ ͋ Δ ͜ ͱ ͔ Β ɼGPU Λ ൚ ༻ ԋ ࢉ ʹ ར ༻ ͢ Δ GPGPU(General-Purpose computation using GPUs)2)΁ͷ஫໨͕ߴ·͍ͬͯΔɽ ࠓ೔Ͱ͸ίϯγϡʔϚPCͷଟ͘ʹGPU͕౥ࡌ͞ Ε͍ͯΔͷͱൺֱ͢ΔͱɼGPGPUͷ׆༻ࣄྫ͸ݶΒ Ε͍ͯΔɽͦͷେ͖ͳཧ༝ͷ̍ͭʹGPGPUϓϩά ϥϛϯάͷ೉͠͞ɼ͢ͳΘͪGPU͸ฒྻॲཧʹద͠ ͨϋʔυ΢ΣΞͰ͋Γͳ͕ΒطଘͷฒྻԽख๏Λ༰қ ʹར༻Ͱ͖ͳ͍͜ͱ͕ڍ͛ΒΕΔɽैདྷͷGPGPU ϓϩάϥϛϯάʹ͓͍ͯ͸ɼάϥϑΟοΫεAPIͱ γΣʔμݴޠΛ༻͍ͨάϥϑΟοΫεϓϩάϥϛϯά ͷٕज़͕ඞཁͱ͞Ε͖ͯͨɽ͜͏ͨ͠ϓϩάϥϜ࡞੒ ख๏͸GPGPUϓϩάϥϛϯάʹಛ༗ͷ΋ͷͰ͋Δ † ిؾ௨৴େֶ େֶ Ӄ৘ใγεςϜֶ ݚڀՊ

Graduate School of Information Systems, The Univer-sity of Electro-Communications ͨΊɼଞͷ෼໺ͷϢʔβʹͱͬͯGPGPUͷ׆༻͸ ༰қͰ͸ͳ͍ɽݱࡏͰ͸CUDA3)ͷΑ͏ʹGPGPU ϓϩάϥϛϯάͷಛघੑΛӅณ͢Δϓϩάϥϛϯάݴ ޠͳͲ΋ొ৔͍ͯ͠Δ͕ɼಛఆͷGPUͰͷΈར༻Մ ೳͳ͏͑ʹɼGPUΞʔΩςΫνϟͷཧղͱ৽͍͠ݴ ޠͷशಘ͕ඞཁͰ͋Δɽ ͦ ͜ Ͱ ຊ ߘ Ͱ ͸ ɼط ଘ ͷ ฒ ྻ Խ ख ๏ Λ ༻ ͍ ͨ GPGPUϓϩάϥϛϯάΛఏҊ͢ΔɽGPU͸ฒྻ ੑͷߴ͍ϓϩάϥϜͷ࣮ߦʹద͍ͯ͠ΔͨΊɼطଘͷ ฒྻԽख๏Λ༰қʹGPU΁ద༻͢Δखஈ͕͋Ε͹ɼ GPUΛฒྻϓϩάϥϜͷ࣮ߦ؀ڥͱͯ͠༗ޮ׆༻Մ ೳͱͳΔ͜ͱ͕ظ଴Ͱ͖Δɽ ຊߘͷߏ੒ΛҎԼʹࣔ͢ɽ2ষͰ͸ຊݚڀͷഎܠʹ ͋ͨΔGPUͱGPGPUʹ͍ͭͯ؆୯ʹઆ໌͠ɼݱঢ় ͷ໰୊఺Λ໌Β͔ʹ͢Δɽ3ষͰ͸ղܾҊͱͯ͠طଘ ͷฒྻԽख๏Λར༻ͨ͠GPGPUϓϩάϥϛϯάΛఏ Ҋ͢Δɽ4ষͰ͸ఏҊʹର͢Δ࣮૷ྫͱͯ͠ɼCUDA Λ༻͍ͯGPU޲͚ʹطଘͷฒྻԽख๏Λ࣮૷͢Δํ ๏Λݕ౼͢Δɽ5ষͰ͸ؔ࿈ݚڀʹ৮Εɼ6ষ͸·ͱ Ίͱࠓޙͷ՝୊ͷষͱ͢Δɽ

2. GPU ͷΞʔΩςΫνϟͱطଘͷ GPGPU

ϓϩάϥϛϯάख๏

GPU͸ຊདྷɼߴ଎ʹը૾ඳըΛߦ͏ͨΊʹൃల͠ ͨϋʔυ΢ΣΞͰ͋ΔɽݱࡏͰ͸ಈը࠶ੜࢧԉͳͲͷ ॲཧ΋౷߹͞ΕΔ܏޲ʹ͋ΔͨΊɼੑೳͷࠩҟ͸େ͖ ͍΋ͷͷɼҰൠతͳPCͷଟ͘ʹ౥ࡌ͞Ε͍ͯΔɽ ਤ1ʹ఻౷తͳGPUͷϋʔυ΢ΣΞߏ੒ͱը૾ ඳըͷͨΊͷओͳॲཧͷରԠΛࣔ͢ɽGPU͕ߦ͏ը ૾ඳըͷͨΊͷओͳॲཧ͸ɼฒྻܭࢉ΍ϕΫτϧܭࢉ ʹΑΔߴ଎Խʹద͍ͯ͠ΔɽͦͷͨΊGPU্ͷॲཧ Ϣχοτ͸ฒྻԽ͕ਐΜͰ͓ΓɼϕΫτϧܭࢉʹద͠ ͨ಺෦ߏ଄ͱͳ͍ͬͯΔɽ·ͨߴ౓ͳը૾දݱʹ͸ඳ ը಺༰ʹԠͨ͡ෳࡶͳܭࢉ͕ඞཁͳͨΊɼॲཧϢχο τͷϓϩάϥϚϒϧԽ͕ਐΜͰ͍ΔɽGPUΛ༻͍ͨ ϓϩάϥϜΛ࡞੒͢Δʹ͸ɼDirectX΍OpenGLͱ ͍ͬͨάϥϑΟοΫεAPIΛ༻͍ͯGPUͷಈ࡞λΠ ϛϯά੍ޚ΍CPU-GPUؒͷσʔλ௨৴ͳͲΛߦ͍ɼ HLSLɼGLSLɼCgͱ͍ͬͨγΣʔμݴޠΛ༻͍ͯܭ ࢉϢχοτͷߦ͏ॲཧΛهड़͢Δඞཁ͕͋Δɽ ਤ1 ఻౷తͳ GPU ͷϋʔυ΢ΣΞߏ੒ͱը૾ඳըॲཧͷରԠ

Fig. 1 Hardware construction of traditional GPU and relationship with graphics processing

GPGPU͸ɼGPUͷಛ௃Λར༻༷ͯ͠ʑͳॲཧ(൚ ༻ܭࢉɼGeneral-Purpose computation)Λߦ͏΋ͷ Ͱ͋ΔɽGPUͷϋʔυ΢ΣΞੑೳΛ׆͔ͤΔ༻్ɼ͢ ͳΘͪฒྻܭࢉ΍ϕΫτϧܭࢉʹదͨ͠ର৅໰୊ʹର ͯ͠͸ɼCPUΛѹ౗͢ΔԋࢉੑೳΛಘΔ͜ͱ͕Ͱ͖ Δɽ͔͠͠ͳ͕ΒɼGPGPUΞϓϦέʔγϣϯͷ࣮૷ ͸༰қͰ͸ͳ͍ɽྫ͑͹GPUΛ༻͍ͯ਺஋ܭࢉΛߦ ͏ͨΊʹ͸ɼ਺஋ܭࢉΞϧΰϦζϜΛը૾ඳըͷγε ςϜʹରԠ࣮ͤͯ͞૷͢Δඞཁ͕͋Δ(ਤ2)ɽͦͷͨ Ίʹ͸਺஋ܭࢉϓϩάϥϛϯάͱάϥϑΟοΫεϓϩ άϥϛϯάͷ྆ํʹਫ਼௨͍ͯ͠Δඞཁ͕͋Γɼߴ͍ੑ ೳΛಘΔͨΊʹ͸ฒྻԽख๏΍GPUΞʔΩςΫνϟ ʹ͍ͭͯͷ஌ࣝ΋ॏཁͰ͋ΔɽݱࡏͰ͸CUDAͳͲ άϥϑΟοΫεϓϩάϥϛϯάͷ஌ࣝΛඞཁͱ͠ͳ͍ GPGPUϓϩάϥϛϯάͷͨΊͷݴޠ΍ϥΠϒϥϦ ΋͋Δ͕ɼ͜ΕΒΛ༻͍ͯΞϓϦέʔγϣϯΛ࡞੒͢ Δʹ͸GPUΞʔΩςΫνϟͷ஌͕ࣝඞཁͰ͋ͬͨΓɼ طଘͷϓϩάϥϛϯάख๏ͱ͸ҟͳΔख๏Λशಘ͢Δ ඞཁ͕͋ͬͨΓ͢Δɽ

ࠓ೔Ͱ͸MAC OS XͷAqua/Quartz΍Windows VistaͷAeroGlassΛ͸͡Ίͱͯ͠OS΍σεΫτο ϓ؀ڥͷϨϕϧͰ͋Δఔ౓ߴੑೳͳGPUΛཁٻ͢Δ έʔε͕૿Ճ͍ͯ͠Δ4)∼7)_{ɽߋʹಈը࠶ੜࢧԉػೳ} Λ࣋ͭGPU(ϏσΦΧʔυ)΋૿Ճ͍ͯ͠ΔͨΊɼί ϯγϡʔϚ޲͚PCͷଟ͘ʹGPU͕౥ࡌ͞Ε͍ͯΔɽ GPUͷ࣋ͭཧ࿦ԋࢉੑೳͷߴ͞Λߟྀ͢ΔͱɼGPU ͷ࣋ͭશͯͷੑೳΛ׆༻͢Δ͜ͱ͸Ͱ͖ͳͯ͘΋ɼ͋ Δఔ౓ͷੑೳ͕׆༻Ͱ͖Ε͹༷ʑͳΞϓϦέʔγϣϯ ͷߴ଎Խ͕ظ଴Ͱ͖Δɽ͔͠͠ͳ͕Βɼݱࡏը૾ॲཧ Ҏ֎ͷ෼໺ʹ͓͍ͯGPGPU͕׆༻͞Ε͍ͯΔͷ͸ Ұ෦ͷ਺஋ܭࢉ΍Պֶٕज़ܭࢉʹݶΒΕ͍ͯΔɽͦͷ ࠷ͨΔཧ༝ͱͯ͠ɼGPUΛ׆༻͢ΔϓϩάϥϜͷ࡞ ੒͕༰қͰͳ͍͜ͱ͕ڍ͛ΒΕΔɽGPGPUʹৄ͘͠ ͳ͍ଟ͘ͷΞϓϦέʔγϣϯϓϩάϥϚʹͱͬͯ͸ɼ GPGPU͕༰қʹར༻Ͱ͖Δ͜ͱ͕ॏཁͰ͋Δɽ͜Ε ͸΋ͪΖΜίϯγϡʔϚPCʹݶͬͨ࿩Ͱ͸ͳ͘ɼಛ ʹߴੑೳͳܭࢉ؀ڥΛٻΊΔHPC༻్Ͱ΋ڞ௨ͷ՝ ୊Ͱ͋Δɽ

3. ఏ Ҋ ಺ ༰

3.1 طଘͷฒྻԽख๏Λར༻ͨ͠GPGPUϓϩ άϥϛϯάͷఏҊ GPU͸CPUͱൺ΂ͯߴ͍ϋʔυ΢ΣΞϨϕϧͷฒ ྻੑΛඋ͍͑ͯΔͨΊɼGPUʹΑΔԋࢉੑೳͷ޲্ ͕ظ଴͞Ε͍ͯΔͷ͸΋ͬͺΒߴ͍ฒྻੑΛ࣋ͭΞϓ ϦέʔγϣϯͰ͋ΔɽҰํͰฒྻԽ͸SIMDɼSMPɼ

ਤ2 ඳըॲཧΛར༻ͯ͠ܭࢉॲཧΛߦ͏ϓϩάϥϜͷྫ

Fig. 2 Example of numerical culculation programming using graphics programming

ਤ3 GeForce8000 γϦʔζͷΞʔΩςΫνϟ

Fig. 3 Architecture of GeForce8000 series

PCΫϥελɼGridͳͲطʹ༷ʑͳݚڀ͕ਐΊΒΕͯ ͍ΔςʔϚͰ͋Γɼݱࡏ΋ϚϧνίΞCPUͷීٴͳ ͲʹΑΓ஫໨͞Ε͍ͯΔɽطଘͷฒྻԽख๏Λ༻͍ͯ GPGPUϓϩάϥϛϯάΛߦ͑ΔΑ͏ʹ͢Δ͜ͱ͕Ͱ ͖Ε͹ɼGPUΛطଘͷฒྻ؀ڥʹΑΓ͍ۙ΋ͷͱ͠ ͯѻ͑ΔΑ͏ʹͳΔͱߟ͑ΒΕΔɽGPU͕਎ۙͰ࢖ ͍΍͍͢΋ͷͱͳΔ͜ͱͰɼଟ͘ͷϢʔβ͕GPUͷ ࣋ͭߴ͍ܭࢉੑೳΛ༗ޮʹ׆༻Ͱ͖ΔΑ͏ʹͳΔ͜ͱ ͕ظ଴Ͱ͖Δɽ ͦ͜ͰɼGPGPUϓϩάϥϛϯάʹ͓͍ͯطଘͷฒ ྻԽख๏Λར༻͢Δ͜ͱΛఏҊ͢Δɽຊষͷ࢒Γͷ෦ ෼Ͱ͸ɼ࠷৽GPUͷϋʔυ΢ΣΞߏ੒Λ֬ೝͨ͠͏ ͑Ͱɼطଘͷฒྻϓϩάϥϛϯάʹ͓͍ͯ͸ฒྻԽର ৅ͷཻ౓ʹԠ༷ͯ͡ʑͳฒྻԽख๏͕༻͍ΒΕ͍ͯΔ ͜ͱΛߟྀ͠ɼGPUͷϋʔυ΢ΣΞߏ଄ͱฒྻԽख ๏ͱͷରԠ෇͚Λݕ౼͢Δɽ·ͨ࣍ষͰ͸ຊఏҊʹର ͢Δ࣮૷ͷྫͱͯ͠ɼCUDA޲͚ʹطଘͷฒྻԽख ๏Λ࣮૷͢Δ͜ͱΛݕ౼͢Δɽ 3.2 GPUʹର͢ΔطଘͷฒྻԽख๏ͷׂΓ౰ͯ ͷݕ౼ ࠷৽GPUͷϋʔυ΢ΣΞߏ଄Λ֓؍͠ɼطଘͷฒ ྻԽख๏ΛGPUʹద༻͢Δʹ͸Ͳ͏͢Ε͹Α͍͔ɼ طଘͷฒྻԽख๏Ͱ༻͍Δ࣮ߦϞσϧͱGPUϋʔ υ΢ΣΞΛͲͷΑ͏ʹؔ࿈෇͚Δ͔ʹ͍ͭͯͷݕ౼Λ ߦ͏ɽ ຊߘࣥච࣌఺ͷ࠷৽ੈ୅GPUͰ͋ΔGeForce8000 γϦʔζ͓ΑͼRadeonHD2000/3800γϦʔζͷ಺ ෦ߏ଄ͷ֓ཁΛਤ3͓Αͼਤ4ʹࣔ͢ɽ͜ΕΒͷGPU Ͱ͸ϓϩάϥϚϒϧॲཧϢχοτͷฒྻԽ͕ਐΜͰ͓ ΓɼGPUશମͰ࠷େ100Ҏ্ͷԋࢉΛฒྻ࣮ߦՄೳ

ਤ4 RadeonHD2000/3800 γϦʔζͷΞʔΩςΫνϟ

Fig. 4 Architecture of RadeonHD2000/3800 series

ͱͳ͍ͬͯΔɽͨͩ͠ɼGPU্ͷશͯͷԋࢉث͕ಉ ͡ػೳΛ࣋ͪۉ౳ʹ഑ஔ͞Ε͍ͯΔΘ͚Ͱ͸ͳ͍ɽ͍ ͔ͭ͘ͷԋࢉث͕ू·ͬͯॲཧϢχοτΛߏ੒͓ͯ͠ ΓɼߋʹॲཧϢχοτ͕ू·ͬͯGPUΛߏ੒͍ͯ͠ Δɽ·ͨϝϞϦʹ͍ͭͯ΋ɼGPUશମ͔Βಁաతʹ ΞΫηεՄೳͳ(άϩʔόϧͳ)ϝϞϦͷΈΛඋ͑ͯ ͍ΔΘ͚Ͱ͸ͳ͘ɼॲཧϢχοτ͝ͱʹϩʔΧϧͳϝ ϞϦ΋උ͍͑ͯΔɽ͢ͳΘͪݱࡏͷGPU͸ɼϋʔυ ΢ΣΞϨϕϧͰͷߴ͍ฒྻੑΛඋ͍͑ͯΔͷΈͳΒͣɼ ౥ࡌ͍ͯ͠ΔԋࢉثͱϝϞϦʹ͸֊૚ੑ͕උΘ͍ͬͯ Δͱଊ͑Δ͜ͱ͕Ͱ͖Δɽ طଘͷฒྻϓϩάϥϛϯάʹ͓͍ͯ͸ɼσʔλϨϕ ϧͷฒྻੑɼ໋ྩϨϕϧͷฒྻੑɼεϨουϨϕϧͷ ฒྻੑɼϓϩηεϨϕϧͷฒྻੑͱ͍ͬͨฒྻੑ͕ར ༻͞Ε͍ͯΔɽ·ͨ۩ମతͳฒྻϓϩάϥϛϯάͷख ஈͱͯ͠͸SIMD໋ྩɼOpenMP͓Αͼ֤छεϨο υϥΠϒϥϦɼMPIͳͲ͕༻͍ΒΕ͍ͯΔ(ਤ5)ɽͦ ͜ͰɼGPUͷ֊૚ੑͱطଘͷฒྻԽख๏ͷ֊૚ੑͱ ΛରԠ෇͚ɼGPU޲͚ʹ͜ΕΒͷؔ਺ɾϥΠϒϥϦ Λ࣮૷͢Δ͜ͱΛߟ͑Δɽ 3.2.1 GPU޲͚SIMD໋ྩͷݕ౼ SIMD໋ྩ͸̍౓ͷԋࢉͰෳ਺ͷσʔλʹରͯ͠ॲ ཧΛߦ͏͍ΘΏΔϕΫτϧԋࢉ໋ྩͰ͋Γɼσʔλฒ ྻ޲͚ͷࡉཻ౓ͳฒྻԽʹར༻͞Ε͍ͯΔɽGPUϓ ϩάϥϛϯάʹ͓͍ͯ͸ɼը૾ඳըͷࡍʹߦ͏௖఺΍ ৭ͷԋࢉʹ4࣍ݩϕΫτϧԋࢉ͕ద͍ͯ͠Δ͜ͱ͔Βɼ ϋʔυ΢ΣΞϨϕϧͰϕΫτϧԋࢉʹద͍ͯ͠ΔͷΈ ͳΒͣɼγΣʔμݴޠʹϕΫτϧԋࢉ޲͚ͷΠϯλʔ ਤ5 CPU ʹ͓͚Δ֊૚ܕฒྻϓϩάϥϛϯά

Fig. 5 Hierarchical parallel programming on CPU

ϑΣΠε͕උΘ͍ͬͯΔɽϕΫτϧԋࢉʹΑΔߴ଎Խ ͕ߦ͑ΔΑ͏ʹϓϩάϥϜΛ૊Ή͜ͱ͸GPGPUϓ ϩάϥϛϯάͷجຊͰ͋ΔɽҰํCPUϓϩάϥϛϯ άʹ͓͍ͯ͸MMX΍SSEɼVMXͳͲ֤CPUͷ උ͑Δ໋ྩηοτΛར༻ͯ͠SIMD໋ྩΛར༻͢Δɽ ͦͷͨΊCPUͷରԠ͢Δ໋ྩηοτ͝ͱʹҟͳΔϓ ϩάϥϜهड़Λ͢Δඞཁ͕͋Δ্ʹɼΠϯϥΠϯΞη ϯϒϥΛهड़͢Δඞཁ͕͋Δ৔߹΋͋ΔͨΊɼGPU ޲͚ͷSIMD໋ྩ͸࡞੒͠ʹ͍͘ɽͦ͜Ͱɼத੢Β8)

ʹΑΔهड़ํࣜΛར༻͢Δ͜ͱΛߟ͑Δɽ͜ͷهड़ํ ࣜ͸Ϣʔβ(ΞϓϦέʔγϣϯϓϩάϥϚ)ʹରͯ͠ SIMD໋ྩηοτ͝ͱͷهड़ͷҧ͍ΛӅณ͠ɼ࣮ߦ؀ ڥΛ໰Θͣ౷Ұతͳهड़ʹΑͬͯSIMDԽΛߦ͏͜ ͱͷͰ͖ΔํࣜͰ͋Δɽ࣮૷తʹ͸ఏҊ͢Δهड़ํࣜ Λ༻͍ͯॻ͔ΕͨϓϩάϥϜʹର֤ͯ͠CPUͷରԠ ͢ΔSIMD໋ྩ΁ͷϓϩάϥϜม׵Λߦ͏΋ͷͰ͋ ΔͨΊɼهड़ํ͔ࣜΒGPUͷରԠ͢ΔϕΫτϧԋࢉ ΁ͷม׵ػೳΛ࣮૷͢Δ͜ͱͰGPU޲͚SIMD໋ ྩͷ࣮૷͕ՄೳͱͳΔɽ 3.2.2 GPU޲͚OpenMPͷݕ౼ OpenMP͓Αͼ֤छεϨουϥΠϒϥϦΛ༻͍ͨ ฒྻԽ͸ɼڞ༗ϝϞϦΛඋ͑ͨ؀ڥʹదͨ͠ࡉཻ౓ͷ ฒྻԽख๏Ͱ͋ΔɽOpenMP͸֤छεϨουϥΠϒϥ Ϧ(ओʹpthread)ͷ্Ґʹ࣮૷͞Ε͓ͯΓ͔ͭهड़͕ ༰қͰ࢖͍΍͍ͨ͢Ίɼ͜͜Ͱ͸GPU޲͚OpenMP ͷݕ౼ʹ࿩ΛߜΔɽOpenMPͰ͸ஞ࣍ϓϩάϥϜͷ ιʔείʔυʹϓϥάϚΛૠೖ͠ɼίϯύΠϧ࣌ʹϓ ϥάϚΛղऍͯ͠ڞ༗ϝϞϦΛ༻͍ͨεϨουฒྻϓ ϩάϥϜʹม׵͢Δɽϧʔϓม਺ͷࣗಈॻ͖׵͑ʹΑ ΔϧʔϓͷฒྻԽͱ͍ͬͨࡉཻ౓ͷฒྻԽʹ࢖ΘΕΔ ͜ͱ͕ଟ͘ɼஞ࣍ϓϩάϥϜͷஈ֊తͳฒྻԽʹ΋ద ͍ͯ͠ΔɽOpenMPΛར༻͢Δʹ͸OpenMPͷϓ ϥάϚΛղऍ͢Δ͜ͱ͕ՄೳͳίϯύΠϥ͕ඞཁͰ ͋ΔɽରԠίϯύΠϥͷ୅දతͳ΋ͷͱͯ͠͸Omni

OpenMP Compiler9)ɼIntel C++ Compiler,GCCͳ Ͳ͕͋Γɼଞʹ΋঎༻ίϯύΠϥͷ͍͔͕ͭ͘ରԠ͠ ͍ͯΔɽOpenMPίϯύΠϥͷجຊతͳ࣮૷͸ɼϓϥ άϚ͕ૠೖ͞Εͨ෦෼ʹεϨουͷ੍ޚΛߦ͏هड़Λ ૠೖ͢Δ͜ͱͰ࣮ݱͰ͖ΔɽGPUʹ͸ڞ༗ϝϞϦ͕ ౥ࡌ͞Ε͍ͯΔͨΊɼOpenMPʹ͓͚ΔεϨουͷ ੍ޚʹ૬౰͢Δهड़ΛGPU্Ͱ໛฿Ͱ͖Ε͹GPU ޲͚OpenMPͷ࣮૷͕ՄೳͱͳΔɽ 3.2.3 GPU޲͚MPIͷݕ౼ MPI͸ϝοηʔδ௨৴Λར༻͠ɼෳ਺ϊʔυؒʹ͓ ͚Δσʔλ΍੍ޚͷ΍ΓऔΓΛߦ͏͜ͱͰฒྻϓϩά ϥϜΛܗ੒͢ΔͨΊͷن֨Ͱ͋Δɽϊʔυ͝ͱʹऔಘ ͨ͠rank৘ใʹج͍ͮͯॲཧΛৼΓ෼͚Δͱ͍ͬͨ ࢖͍ํʹద͓ͯ͠ΓɼOpenMPͱൺ΂ͯૄཻ౓ͷฒ ྻԽʹ࢖ΘΕΔ͜ͱ͕ଟ͍ɽMPI͸෼ࢄϝϞϦ؀ڥ ޲͚ͷฒྻԽख๏Ͱ͋Δ͕ɼڞ༗ϝϞϦ؀ڥͰ΋ಛʹ ໰୊ͳ͘ར༻͢Δ͜ͱ͕Ͱ͖ΔɽΉ͠Ζ௨৴ͷࡉ͔͍ ੍ޚ͕ՄೳͳͨΊɼڞ༗ϝϞϦ؀ڥͰ΋OpenMPΑ Γߴ͍ੑೳ͕ಘΒΕΔ͜ͱ͕͋ΔɽMPIͷ࣮૷͸ί ϯύΠϥ΍ෳࡶͳιʔείʔυม׵ػߏΛඞཁͱ͠ͳ ͍୅ΘΓʹɼ(ෳ਺ͷPCʹލͬͯ)ෳ਺ͷϓϩηεΛ ্ཱͪ͛ΔͨΊͷػߏ౳Λඋ͍͑ͯΔɽMPIͷ࣮૷ ͱͯ͠͸mpich͓Αͼmpich2,LAM10),11)ͳͲ͕ڍ ͛ΒΕΔɽGPU޲͚MPIͷ࣮૷ʹ͍ͭͯ͸ɼԋࢉϢ χοτؒʹ͓͚Δڞ༗ϝϞϦΛհͨ͠σʔλ΍੍ޚͷ ΍ΓऔΓʹMPIͷΠϯλʔϑΣΠεΛར༻Ͱ͖ΔΑ ͏ʹ͢Δͱ͍͏࣮૷͕ߟ͑ΒΕΔɽ·ͨɼCPU-GPU ؒͷ௨৴ʹMPIͷΠϯλʔϑΣΠεΛར༻Ͱ͖ΔΑ ͏ʹ͢Δ͜ͱͰɼCPU্ͷϝϞϦͱGPU্ͷϝϞ Ϧͱ͍͏෼ࢄϝϞϦͷ؅ཧΛطଘͷ෼ࢄϝϞϦʹ͍ۙ هड़ʹΑͬͯѻ͑ΔΑ͏ʹ͢Δ͜ͱ͕Ͱ͖Δɽ Ҏ্ͷΑ͏ʹɼطଘͷ༷ʑͳฒྻԽख๏ʹGPUΛ ରԠ෇͚Δ͜ͱͰɼGPU಺෦ͷԋࢉث΍ϝϞϦ͕࣋ ͭ֊૚ੑΛ׆༻ͨ͠GPU޲͚ฒྻϓϩάϥϜΛ༰ қʹ࡞੒ՄೳͱͳΔ͜ͱ͕ظ଴Ͱ͖Δɽ·ͨSMP΍ ϚϧνίΞCPUΛ౥ࡌͨ͠ϊʔυʹΑͬͯߏ੒͞Ε ΔPCΫϥελͳͲʹ͓͍ͯ͸ɼϊʔυ಺ͷฒྻԽΛ OpenMPɼϊʔυؒͷฒྻԽΛMPIʹΑͬͯߦ͏ϋ ΠϒϦουͳฒྻԽͳͲʹΑͬͯޮ཰తͳฒྻԽ͕ߦ ΘΕ͍ͯΔɽGPUΛ༻͍Δ৔߹ʹ͓͍ͯ΋ɼGPU಺ ͷ֊૚తͳฒྻੑΛར༻ͨ͠ϋΠϒϦουͳฒྻԽ΍ɼ GPUͱCPUͷฒྻॲཧͳͲ༷ʑͳϨϕϧͰͷฒྻ ॲཧ͕༰қʹߦ͑ΔΑ͏ʹͳΔ͜ͱ͕ظ଴Ͱ͖Δɽ

4. CUDA Λ༻͍ͨ GPU ޲͚طଘͷฒྻԽ

ख๏ͷ࣮૷ʹ޲͚ͨݕ౼

ຊষͰ͸ɼCUDAΛ༻͍ͯGPU޲͚ʹطଘͷฒ ྻԽख๏Λ࣮૷͢Δํ๏ʹ͍ͭͯΑΓ۩ମతͳݕ౼Λ ߦ͏ɽ ͜Ε·ͰGPUͷػೳΛར༻͢Δʹ͸άϥϑΟοΫ εAPIΛհ͢Δඞཁ͕͋ͬͨͨΊɼGPGPUʹ͓͍ͯ ͸શͯͷॲཧΛάϥϑΟοΫεϓϩάϥϛϯάͷํࣜ ʹ͋Θ࣮ͤͯ૷͢Δඞཁ͕͋ͬͨɽͦͷͨΊGPUϓ ϩάϥϛϯάΛߦ͏ͨΊʹ͸άϥϑΟοΫεϓϩάϥ ϛϯάΛशಘ͢Δඞཁ͕͋ͬͨɽ͜Εʹରͯ͠CUDA Ͱ͸ଟ͘ͷॲཧΛΑΓ௚ײతʹهड़͢Δ͜ͱ͕Ͱ͖Δ (ਤ6)ɽ·ͨ͜Ε·ͰGPUͷ಺෦৘ใʹ͍ͭͯ͸ά ϥϑΟοΫεϓϩάϥϛϯάʹඞཁͳఔ౓ͷ৘ใͷΈ ͕ެ։͞Ε͍͕ͯͨɼCUDAͱͱ΋ʹଟ͘ͷ৘ใ͕ఏ ڙ͞ΕΔΑ͏ʹͳͬͨɽ͜ΕʹΑΓGPUϓϩάϥϜ ͷσόοά΍࠷దԽʹ༗ӹͳ৘ใ͕ೖख͠΍͘͢ͳͬ ͨͱݴ͑Δɽ ਤ7ʹCUDAͷฒྻ࣮ߦϞσϧ͓ΑͼϝϞϦϞσ ϧΛࣔ͢ɽ͜ΕΛݩʹɼCUDAΛ༻͍ͯطଘͷฒྻ

Խख๏Ͱ͋ΔSIMD໋ྩɼOpenMP͓ΑͼMPIͷ

ਤ6 CUDA Λ༻͍ͨ GPGPU ϓϩάϥϛϯά

Fig. 6 GPGPU programming using CUDA

ਤ7 CUDA ͷฒྻ࣮ߦϞσϧͱϝϞϦϞσϧ

Fig. 7 Parallel execution model and memory model of CUDA

4.1 CUDA޲͚SIMD໋ྩͷ࣮૷

͸͡ΊʹCUDA޲͚SIMD໋ྩͷ࣮૷ʹ͍ͭͯߟ

͑ΔɽCUDAʹ͓͚Δ֤Block಺ͷThread͸ɼର

৅ͷσʔλ܈ʹରͯ͠ಉҰͷॲཧΛಉ࣌ʹ࣮ߦ͢Δ͜ ͱ͕ՄೳͰ͋Δɽ͜Ε͸యܕతͳSIMD໋ྩͷॲཧ ͱಉҰͰ͋ΔͨΊɼThreadΛ༻͍ͨԋࢉʹରͯ͠ط ଘͷSIMD໋ྩηοτ(ڞ௨هड़ํࣜ)ʹରԠͨ͠Π ϯλʔϑΣΠεΛ༻ҙ͢Δ͜ͱͰCUDA޲͚SIMD ໋ྩ͕࣮ݱͰ͖Δͱߟ͑ΒΕΔɽ 4.2 CUDA޲͚OpenMPͷ࣮૷ ଓ͍ͯ CUDA ޲͚OpenMP ʹ͍ͭͯߟ͑Δɽ OpenMPΛ༻͍ͨฒྻॲཧʹ͸ɼ֤ԋࢉث͕ࣗ༝ʹ ΞΫηε͢Δ͜ͱͷͰ͖Δڞ༗ϝϞϦ͕ඞཁͰ͋Δɽ CUDAʹ͸ڞ༗ϝϞϦͱͯ͠ར༻Մೳͳෳ਺छྨͷ

ϝϞϦ͕ଘࡏ͢ΔɽGlobal Memory, Constant

Mem-ory, Texture Memory͸શͯͷBlock͓Αͼશͯͷ

Thread͔Βڞ༗ϝϞϦͱͯ͠ར༻ՄೳͰ͋ΓɼShared

Memory͸ಉҰBlock಺ͷશͯͷThread͔Βڞ༗ ϝϞϦͱͯ͠ར༻Մೳͱͳ͍ͬͯΔɽ͜ΕΒͷ͏ͪɼ

Constant MemoryͱTexture Memory͸GPU͔Β ݟΔͱಡΈࠐΈઐ༻ͷϝϞϦͰ͋ΔͨΊɼࠓճ͸ར༻

ର৅͔Βআ֎͢Δɽ࢒ΔGlobal MemoryͱShared

Memoryʹ͍ͭͯ͸ɼGlobal MemoryΛར༻͢Ε͹

ThreadϨϕϧͱBlockϨϕϧ૒ํͰɼShared

Mem-oryΛར༻͢Ε͹BlockϨϕϧͰOpenMPͷฒྻॲ

ཧΛஔ͖׵͑ΒΕΔͱߟ͑ΒΕΔɽΑͬͯɼ͜ΕΒͷ ϝϞϦΛར༻ͯ͠OpenMPͷػೳΛ࣮ݱ͢Δ͜ͱͰɼ CUDA޲͚OpenMP͕࣮૷Ͱ͖Δͱߟ͑ΒΕΔɽ OpenMP ͸ࡉཻ౓ͳฒྻॲཧʹར༻͢ΔͨΊɼ BlockϨϕϧͷฒྻॲཧΑΓ΋ThreadϨϕϧͷฒ ྻॲཧʹద͍ͯ͠Δͱߟ͑Δ͜ͱ͕Ͱ͖ΔɽҰํͰ GPU಺෦ͷϝϞϦసૹ଎౓͸CPU-ϝΠϯϝϞϦؒ ΑΓ΋ߴ଎Ͱ͋Γɼ·ͨBlockϨϕϧͷฒྻॲཧ͸ ThreadϨϕϧͷॲཧͱൺ΂ͯॲཧͷࣗ༝౓͕ߴ͍ɽ

ਤ8 CUDA Λ༻͍ͨϋΠϒϦουฒྻϓϩάϥϜͷಈ࡞Πϝʔδ

Fig. 8 Executive image of hybrid parallel program using CUDA

ͦͷͨΊCUDA޲͚OpenMPͷ࣮૷ʹ͍ͭͯ͸ɼ ThreadϨϕϧͱBlockϨϕϧͷͦΕͧΕͰ࣮૷ͯ͠ ੑೳൺֱΛߦ͍ɼΑΓྑ͍࣮૷Λ୳Δ͜ͱʹ͢Δɽ 4.3 CUDA޲͚MPIͷ࣮૷ ࠷ޙʹCUDA޲͚MPIͷ࣮૷ʹ͍ͭͯߟ͑Δɽ CUDAͷฒྻ࣮ߦϞσϧͷ͏ͪɼThreadϨϕϧͷฒ ྻॲཧ͸SIMDʹۙ͘ɼҟछͷԋࢉΛฒྻ࣮ߦ͢Δ͜ ͱ͕Ͱ͖ͳ͍ɽ͜Ε͸ૄཻ౓ͷฒྻॲཧΛߦ͏MPI ʹ͓͍ͯக໋తͳ໰୊Ͱ͋ΔͨΊɼBlockϨϕϧͷฒ ྻॲཧΛར༻͢Δ͜ͱʹ͢ΔɽMPIΛ༻͍ͨฒྻॲ ཧʹ͸ϊʔυؒͷ௨৴͕ෆՄܽͰ͋Γɼ·ͨಉظ௨৴ ؔ਺Λ໛฿͢Δʹ͸Blockؒͷಉظػߏ͕ඞཁͰ͋ ΔɽCUDAʹ͸BlockؒͰ௚઀తͳ௨৴Λߦ͏ػߏ ͕ଘࡏ͠ͳ͍͕ɼGlobal MemoryΛհͯ͠σʔλΛ ΍ΓऔΓ͢Δ͜ͱͰBlockؒͷ௨৴Λ໛฿͢Δɽߋʹ CUDAʹ͸BlockؒͷಉظΛऔΔػߏ͕༻ҙ͞Εͯ ͍ͳ͍ͨΊɼGlobal MemoryΛར༻ͯ͠ಉظॲཧΛ ໛฿͢ΔɽҎ্ʹΑͬͯCUDA޲͚MPI͕࣮૷Ͱ ͖Δͱߟ͑ΒΕΔɽ ҰํɼCPU-GPUؒͷσʔλసૹΛߦ͏ؔ਺Λఆٛ ͠ɼCPUͱGPUͱ͍͏ҟͳΔϋʔυ΢ΣΞؒͷ෼ࢄ ϝϞϦΛMPIͷΠϯλʔϑΣΠεͰ؅ཧՄೳʹ͢Δ ͱ͍͏ద༻ํ๏΋ߟ͑ΒΕΔɽ͜ͷࡍͷ໰୊ͱͯ͠͸ɼ

CUDAͷ࣮ߦϞσϧͱͯ͠Grid(CPU͔ΒGPUʹ

ରͯ͠ԋࢉࢦࣔΛߦ͏ࡍͷ୯Ґ)ͷ్தͰCPU-GPU ؒͷσʔλૹड৴͕Ͱ͖ͳ͍͜ͱɼͦͷͨΊGridͷ ෼ׂͳͲ͕ඞཁͱͳΔՄೳੑ͕ڍ͛ΒΕΔɽ͔͠͠ͳ ͕ΒطଘͷΠϯλʔϑΣΠεͰ؅ཧͰ͖ΔϝϦοτ͸ େ͖͍ͨΊɼ͜ͷద༻ํ๏͸࣮૷͢ΔՁ஋͕͋Δɽ Ҏ্ͷΑ͏ʹThreadϨϕϧͷฒྻԽͱBlockϨϕ ϧͷฒྻԽΛద੾ʹ࢖͍෼͚Δ͜ͱͰɼCUDA޲͚

ͷSIMD໋ྩɼOpenMPɼMPIΛ࣮૷͢Δ͜ͱ͕Ͱ

͖Δͱߟ͑ΒΕΔɽߋʹɼطଘͷCPU޲͚ฒྻϓϩ άϥϛϯάʹ͓͚Δ֊૚ܕͷ(ϋΠϒϦουͳ)ฒྻ ϓϩάϥϜΛ໛฿͢Δ͜ͱ΋ՄೳͱͳΔɽਤ8ʹզʑ ͕ݕ౼ΛਐΊ͍ͯΔCUDAΛ༻͍ͨϋΠϒϦουฒ ྻϓϩάϥϜͷಈ࡞ΠϝʔδΛࣔ͢ɽ͜ΕΒΛ༻͍Δ ͜ͱͰɼGPUͷੑೳΛΑΓ༰қʹ͔ͭޮ཰ྑ͘ൃش Ͱ͖ΔΑ͏ʹͳΔ͜ͱ͕ظ଴Ͱ͖Δɽ

5. ؔ ࿈ ݚ ڀ

ݱࡏɼޮ཰ྑ͘GPGPUϓϩάϥϛϯάΛߦ͏ͨ Ίͷݴޠ΍ϥΠϒϥϦͷݚڀ͕͍͔ͭ͘ߦΘΕ͍ͯ Δɽதʹ͸BrookGPU12)΍RapidMind13)ͷΑ͏

ʹɼGPUͷΈͰ͸ͳ͘Cell Broadband Engine΍

ϚϧνίΞCPUͳͲ΁ͷద༻Λࢹ໺ʹೖΕͨ΋ͷ΋

ଘࡏ͢Δɽ͜ΕΒ͸GPGPUϓϩάϥϛϯάΛ༰қ

ʹ͍ͨ͠ͱ͍͏ཁٻ͸ຊݚڀͱಉ༷Ͱ͋Δɽ͔͠͠ͳ ͕ΒɼC/C++ʹର͢Δ֦ுͱ͍͏ܗΛͱΓطଘͷϓ ϩάϥϛϯάݴޠʹ͍ۙهड़͕Ͱ͖ΔΑ͏ʹ͍ͯ͠Δ

ͱ͸͍͑ɼ৽͍͠ݴޠ΍ϥΠϒϥϦΛ࡞੒͍ͯ͠Δͨ ΊطଘͷΞϓϦέʔγϣϯϓϩάϥϚʹର͢Δशಘί ετͷେ͖͞͸൱Ίͳ͍ɽҰํɼର৅ΞʔΩςΫνϟ ʹదͨ͠ݴޠΛ࡞੒͢Δ͜ͱͰطଘͷݴޠΛ༻͍ΔΑ Γߴ͍ੑೳΛಘ΍͍͢Մೳੑ΋͋ΔͨΊɼهड़΍शಘ ͷ͠΍͢͞ͱੑೳʹ͍ͭͯͷٞ࿦ɾධՁΛߦ͏Ձ஋͕ ͋Δɽ ຊݚڀͰ࣮૷ʹར༻͍ͯ͠ΔCUDA͸ɼGeForce8000 γϦʔζͷΞʔΩςΫνϟʹڧ͘ґଘͨ͠ϥΠϒϥϦ Ͱ͋ΔɽCUDA͸ࠓޙNVIDIAͷϦϦʔε͢Δ৽͠ ͍GPUͰ΋ར༻Ͱ͖Δͱ͞Ε͍ͯΔ͕ɼࠓͷͱ͜Ζ ଞࣾͷGPUͰ༻͍Δ͜ͱ͸Ͱ͖ͳ͍ͨΊɼGPUϓϩ άϥϜΛ͢΂ͯCUDAͰهड़͢Δͱ͍͏ͷ͸ݱ࣮త Ͱ͸ͳ͍ɽ͜Εʹର͠ɼNVIDIAͱGPUͷγΣΞΛ ೋ෼͍ͯ͠ΔAMD΋Close-to-the-Metal(CTM)14) ͷఏڙΛද໌͍ͯ͠ΔɽGPUϓϩάϥϜͷهड़͠΍ ͢͞΍ੑೳͷൃش͠΍͢͞ͷ໘͔Βɼࠓޙ΋GPUϕ ϯμʔ͕GPGPU޲͚ʹ։ൃ؀ڥ΍ࢿྉΛఏڙ͠ଓ ͚ΔՄೳੑ͸ߴ͍ɽҰํͰɼCUDA΍CTM͕ݴޠ ࢓༷ͷมߋͳ͘ܧଓͯ͠ఏڙ͞ΕΔ͔͸ෆ໌Ͱ͋Γɼ ·ͨGPUͷछྨ͝ͱʹݸผͷϓϩάϥϜΛ࡞੒͢Δ ඞཁ͕͋Δ͜ͱ͸ɼGPGPUϓϩάϥϛϯάΛ༰қ ʹ͢Δͱ͍͏؍఺͔Β͸େ͖ͳো֐Ͱ͋ΔɽຊఏҊ͸ GPUͷҧ͍Λٵऩ͠ڞ௨ʹར༻Ͱ͖Δ؀ڥΛఏڙ͢ Δ΋ͷͰ͋Γɼ͜͏ͨ͠໰୊ΛղܾͰ͖ΔՄೳੑ͕͋ Δͱݴ͑Δɽ

6. ͓ Θ Γ ʹ

ຊߘͰ͸GPGPUΛ༰қʹར༻͢ΔͨΊͷ৽͍͠ ख๏ͱͯ͠ɼطଘͷฒྻԽख๏Λ༻͍ͨGPGPUϓϩ άϥϛϯάͷఏҊΛߦͬͨɽ·ͨCUDAΛ༻͍࣮ͨ૷ ྫʹ͍ͭͯ΋ݕ౼ΛߦͬͨɽຊఏҊʹΑͬͯGPGPU ϓϩάϥϛϯάʹطଘͷฒྻԽख๏͕ར༻Ͱ͖ΔΑ ͏ʹͳΕ͹ɼ༷ʑͳΞϓϦέʔγϣϯʹରͯ͠༰қʹ GPUΛ׆༻Ͱ͖ΔΑ͏ʹͳΔɽߋʹGPUͱ͍͏ಛघ ͳϋʔυ΢ΣΞ޲͚ͷϓϩάϥϛϯάΛطଘͷCPU ޲͚ϓϩάϥϛϯάख๏Ͱߦ͏ྫͱͯ͠ߟ͑Δ͜ͱͰɼ ฒྻԽख๏ʹͱͲ·Β༷ͣʑͳCPU޲͚ͷϓϩάϥ ϛϯάख๏ΛGPU޲͚ʹ׆༻Ͱ͖ΔՄೳੑΛࣔͯ͠ ͍ΔɽຊఏҊ͸ɼࠓޙ·͢·͢ͷੑೳ޲্͕ظ଴͞Ε ΔGPUΛ༰қʹ׆༻͢Δ৽͍͠ՄೳੑΛࣔ͢΋ͷͰ ͋Δɽ ݱࡏ͸ఏҊ಺༰ͷ࣮૷ΛਐΊ͍ͯΔɽࠓޙ΋࣮૷Λ ਐΊɼ֤छΞϓϦέʔγϣϯʹద༻ͯ͠ੑೳධՁΛߦ ͏ɽ·ͨ͜ΕΒΛ௨ͯ͠GPGPUϓϩάϥϛϯάʹ طଘͷฒྻԽख๏Λ༻͍Δ͜ͱࣗମʹ͍ͭͯ΋ධՁΛ ߦ͏͜ͱͰɼΑΓྑ͍GPGPUϓϩάϥϛϯάख๏ ʹ͍ͭͯͷٞ࿦͕ਂ·Δ͜ͱ͕ظ଴Ͱ͖Δɽ

ࢀ

ߟ จ

ݙ

1) gpgpu.org: SIGGRAPH 2007 GPGPU COURSE, http://www.gpgpu.org/s2007/.

2) gpgpu.org: General-Purpose computation on GPUs(GPGPU), http://gpgpu.org/.

3) NVIDIA: CUDA Programming Guide 1.0 (CUDA NVIDIA Homepage),

http://developer.nvidia.com/cuda/. 4) Apple: Apple Human Interface Guidelines,

http://developer.apple.com/documentation/ UserExperience/Conceptual/OSXHIGuidelines/. 5) Apple: Graphics & Imaging - Quartz,

http://developer.apple.com/ graphicsimaging/quartz/.

6) Microsoft: Experience Windows Vista: Win-dows Aero, http://www.microsoft.com/ windows/products/windowsvista/features/ experiences/aero.mspx.

7) RenderingProject/aiglx: RenderingProject/aiglx - Fredora Project Wiki, http://fedoraproject. org/wiki/RenderingProject/aiglx.

8) த੢༔,౉ᬑܒਖ਼,ฏᖒকҰ,ຊଟ߂थ:ίʔυͷ

ੑೳՄൖੑΛఏڙ͢ΔSIMD޲͚ڞ௨هड़ํࣜ,

৘ใॲཧֶձ࿦จࢽ ίϯϐϡʔςΟϯάγες Ϝ, Vol. 48, pp. 95–105 (2007).

9) M.Sato, S.Satoh, K.Kusano and Y.Tanaka: Design of OpenMP Compiler for an SMP Clus-ter, EWOMP ’99 , pp. 32–39 (1999).

10) Burns, G., Daoud, R. and Vaigl, J.: LAM: An Open Cluster Environment for MPI,

Proceed-ings of Supercomputing Symposium, pp. 379–

386 (1994).

11) Squyres, J. M. and Lumsdaine, A.: A Com-ponent Architecture for LAM/MPI,

Proceed-ings, 10th European PVM/MPI Users’ Group Meeting, Lecture Notes in Computer Science,

No. 2840, Venice, Italy, Springer-Verlag, pp. 379–387 (2003).

12) BrookGPU: BrookGPU, http://graphics. stanford.edu/projects/brookgpu/.

13) RapidMind Inc.: RapidMind, http://www.rapidmind.net/.

14) ATI: ATI CTM Guide, http://ati.amd.com/ companyinfo/researcher/documents.html.