Isifundo sokufakela iCUDA kwiWindows

  • Ukuhambelana: Iinguqulelo zeWindows ezixhaswayo, abaqhubi abafanelekileyo, kunye nemifuziselo eqwalaselwe ngokufanelekileyo yeWDDM/TCC.
  • Ufakelo oluguquguqukayo: I-Toolkit ye-Graphical okanye ethuleyo, i-Conda kunye ne-Wheels, kunye nokuqinisekiswa ngeesampuli.
  • Ubume be-Hybrid: Inkxaso ye-WSL ene-kernel encinci, i-Docker, kunye nesikhokelo se-AI ngaphakathi kweLinux kwiWindows.
Daniel Terrasa

Imizuzu ye8

CUDA kwiWindows

Ukufakela iCUDA kwiWindows Akunyanzelekanga ukuba ibe yintloko ebuhlungu ukuba ulandela indlela ecacileyo engqinisiswe zizikhokelo ezisemthethweni. Ngaso sonke esi sikhokelo sisebenzayo kunye nezobugcisa, siza kugubungela yonke into eyimfuneko ukulungisa kakuhle i-Toolkit, abaqhubi, kunye nezixhobo zokuqinisekisa, kwiiWindows zomthonyama nakwi-WSL. Injongo kukuba ugqibezele ngokufakela okusebenzayo nokuvavanywayo., ukulungele ukukhawulezisa ukuhamba kwakho kweGPU.

Ukongeza kumanyathelo eklasikhi, siya kugubungela ukuhambelana kweVisual Studio, ukufakwa okuthe cwaka, iipakethe zeConda kunye namavili eNVIDIA, ukuqinisekiswa ngemizekelo yeCUDA, izinto ezikhethekileyo zeWSL, kunye namanqaku angundoqo okukhululwa (njengokususwa komqhubi kwi-installer kwi-CUDA 13). Uya kubona kwakhona ukuba ufumanisa njani iingxaki eziqhelekileyo. (kubandakanywa nemeko yehlabathi langempela kunye ne-Lenovo Ideapad kunye nabaqhubi be-526.56) kunye neemfuno ezithile zeemeko ezithile ezifana ne-Model Builder kunye ne-CUDA 10.1 kunye ne-cuDNN 7.6.4.

Yintoni iCUDA?

I-CUDA liqonga lenkqubo elinxuseneyo le-NVIDIA kunye nemodeli evumela ukukhawuleziswa kwezicelo ezisebenza ngamandla kwi-GPU. NgeCUDA C/C++ kunye nezandiso zayo Unokugxila ekufaniseni i-algorithm, ngelixa ixesha le-CUDA lilawula ukubulawa kunye nememori phakathi kwe-CPU kunye ne-GPU. Ngokwahlula umsebenzi we-serial (CPU) kunye ne-parallel (GPU), Kuyenzeka ukulinganisa ukusebenza ngokuxhasa amawaka eentambo kunye nememori ekwabelwana ngayo kwi-chip.

Esi sikhokelo sigxile kwindlela yokulungisa kakuhle imeko-bume kwi-Windows (yemveli kunye ne-WSL), faka i-Toolkit, uqokelele imizekelo, kwaye uqinisekise ukuba inkqubo iyibhaqa ngokuchanekileyo i-GPU. Ukuqinisekiswa ngeesampulu ezifana neDeviceQuery kunye ne-bandwidthTest Liqhekeza elibalulekileyo elihlala lingahoywa kwaye esithetha ngalo ngokucacileyo apha.

CUDA

Iimfuno zenkqubo kunye neenguqulelo zeWindows ezixhaswayo

Ukusebenzisa i-CUDA udinga i-NVIDIA GPU ehambelanayo kunye nenkqubo yokusebenza yeWindows exhaswayo. Iinguqulelo ezixhaswayo zenkqubo yokusebenza zibandakanyaWindows 11 24H2, 23H2 kunye ne-22H2-SV2; Windows 10 22H2; I-Windows Server 2022 kunye ne-Windows Server 2025. Khangela kuMphathi weSixhobo ukuba ikhadi lakho livela phantsi kwe-Display adapters kwaye uqinisekise ukuhambelana kwayo kuluhlu olusemthethweni lwe-CUDA-ekwaziyo i-GPUs (iphepha le-NVIDIA).

Kwinqanaba lokuhlanganisa, i-Toolkit isekela i-MSVC 193x kwi-Visual Studio 2022 17.x (C ++ 11/14/17/20) kunye ne-MSVC 192x kwi-Visual Studio 2019 16.x (C ++ 11/14/17). Inkxaso ye-Visual Studio 2015 yachithwa kwi-11.1Kwi-VS 2017, yachithwa kwi-version 12.5 kwaye yasuswa kwi-13.0. Qaphela ukuba i-32-bit compilation isusiwe kwaqala nge-CUDA 12.0 (indalo kunye ne-cross-compilation); ukuba ufuna i-32-bit, kuya kufuneka usebenzise izixhobo ezindala.

Nangona umqhubi we-NVIDIA uya kuqhubeka nokuvumela ukuqhuba iibhinari ezingama-32 kwi-GeForce GPUs ukuya kuyilo lwe-Ada, UAda uya kuba ngowokugqibela ngaloo nkxasoIHopper ayisaxhasi usetyenziso lwe-32-bit. Oku kubalulekile ukuba ugcina isoftware yelifa kwimveliso kwaye kufuneka ucwangcisele ukuphuculwa.

In Windows 10 kwaye kamva, umqhubi we-NVIDIA unokusebenza phantsi kweemodeli ezimbini: I-WDDM (yezixhobo zokubonisa) kunye ne-TCC (i-Tesla Compute Cluster) yee-GPU ngaphandle kwemveliso yevidiyo, njengeTesla okanye imodeli ethile ye-Titan. Nge nvidia-smi Ungajonga imo yangoku kwaye, xa uxhaswa, tshintsha phakathi kwabo. KwiiPC zedesktop ezineGPU zokudlala, iWDDM iya kuba yinto eqhelekileyo; kwiindawo zokusebenza ezinamakhadi ekhompyutha, I-TCC idla ngokuvulwa ngokungagqibekanga..

Ukufakela i-Toolkit kwi-Windows: inyathelo ngenyathelo

  1. Qinisekisa ukuba unayo i-CUDA-GPU kunye nenguqulelo yeWindows exhaswayo.Ukusuka kuMphathi weSixhobo (Bonisa iiadaptha) ungabona umenzi kunye nemodeli.
  2. Khuphela i-NVIDIA CUDA Toolkit kwiphepha elisemthethweni lokukhuphela. Unefomathi ezimbini: I-Network Installer (ukhuphelo kuphela oko kukhethiweyo ngexesha lofakelo) kunye ne-Full Installer (ibandakanya onke amacandelo, aluncedo ngaphandle kwe-intanethi okanye ukusasazwa kweshishini).
  3. Qalisa isifakeli kwaye ulandele iwizadi yegraphical. ukufaka i-Toolkit, imizekelo kunye nokudibanisa kunye ne-Visual Studio ukuba kuyimfuneko.
  4. Kufakelo oluzenzekelayo ungasebenzisa imo yokuthula nge -skwaye yongeza iiparameters ukukhetha iipakethe ezithile. Iflegi -n Kuphephe ukuqalisa kwakhona ngokuzenzekelayo ukuba kuyafuneka.
  5. Gcwalisa kwaye uqalise kwakhona ukuba iceliwe.
  6. Vula i-Command Prompt kwaye uqhube nvcc -V ukubona inguqulelo yomqokeleli weCUDA efakiweyo.

Faka iCUDA kwiWindows

Ufakelo oluthe cwaka kunye neepakethe ezincinci ze-Toolkit

I-installer ikuvumela ukuba usebenzise amacandelo athile e-Toolkit, eluncedo kwiindawo ezilawulwayo okanye ze-CI. Ezinye iipakethi ezingaphantsi zabameli kwiCUDA 13.0 (indlela emiselweyo C:\iifayile zeNkqubo\NVIDIA GPU Computing Toolkit\CUDA\v13.0) zezi:

  • cublas_13.0 / cublas_dev_13.0: amathala eencwadi e-cuBLAS.
  • crt_13.0: umqokeleli wezicelo zeCUDA.
  • I-ctadvisor_13.0Uhlalutyo lwe-NVCC/NVRTC lokulandelela ukunciphisa amaxesha okuqokelela.
  • cuda_profiler_api_13.0 y cuadrt_13.0: Iprofayili yeCUDA kunye ne-runtime API.
  • cufft_13.0 / cufft_dev_13.0: cuFF.
  • cuobjdump_13.0 y nvdisasm_13.0: izinto eziluncedo zokuhlola iikhubhini.
  • cupti_13.0: ujongano lwezixhobo zokujonga.
  • cuand_13.0 / cuand_dev_13.0: IKURANDI.
  • cusolver_13.0 / cusolver_dev_13.0 y cusparse_13.0 / cusparse_dev_13.0: amathala eencwadi ealgebra.
  • cuxxfilt_13.0: isihluzi se-demangler cu++.
  • uxwebhu_13.0: I-HTML kunye nezikhokelo zePDF (inkqubo, eyona ndlela ibalaseleyo, njl.njl.).
  • i-sight_compute_13.0, Sight_systems_13.0 y umbono_vse_13.0Izixhobo zokujonga.
  • npp_13.0 / npp_dev_13.0: NPP.
  • nvcc_13.0Umqokeleli weCUDA.
  • nvfatbin_13.0, nvjitlink_13.0, nvtx_13.0, nvvm_13.0, nvptxcompiler_13.0: izixhobo zokusebenza/ amathala eencwadi kunye nemikhondo.
  • nvjpeg_13.0 / nvjpeg_dev_13.0: nvJPEG.
  • nvml_dev_13.0Iiheader zeNVML kunye namathala eencwadi.
  • nvprune_13.0: ukuthenwa kwezinto/i-statics ngokujoliswe kuko kwesixhobo.
  • indawo_yokubala_13.0: i-spreadsheet yokuhlala.
  • opencl_13.0: Ithala leencwadi le-OpenCL.
  • sanitizer_13.0Compute Sanitizer API.
  • thrust_13.0 y visual_studio_integration_13.0I-Thrast kunye nokudibanisa VS.

Ukuba ufuna ukuhlola iziqulatho ngaphambi kokufaka, unokukhupha uMfaki oPheleleyo ngesixhobo esihambelana ne-LZMA (umzekelo, i-7-Zip). Iifayile ze Toolkit zibekwe kwi CUDAToolkit ulawulo Emva kokukhupha iifayile, uya kufumana iifayile ze.dll kunye .nvi ngaphakathi, ongayihoya ukuba ufuna ukuphonononga kuphela umthi wofakelo. Ukukhutshwa kwenziwa ngePhaneli yoLawulo> Iinkqubo kunye neempawu.

Ufakelo ngepip (NVIDIA Wheels)

Kwiindawo zePython-centric, i-NVIDIA ibonelela ngamavili afaka iCUDA ixesha lokubaleka kunye namathala eencwadi ngepip. Ezi phakheji zenzelwe ukusetyenziswa ngexesha lokuqhuba. kwaye ungaquki izixhobo zophuhliso (ungazifaka ngokwahlukeneyo xa zifuneka). Qaphela: indawo ye-CUDA ilawulwa ngaphakathi kwendawo yombhobho, ngoko kuya kufuneka ulungelelanise indawo yokusingatha ukuba ujonge ukusebenzisa i-CUDA ngaphandle kwayo.

Okokuqala, faka nvidia-pyindex ukuze ipip ikwazi ukusombulula iimodyuli ukusuka kwindawo yokugcina ye-NVIDIA NGC. Ukuba kuyafuneka, ukuhlaziya umbhobho kunye ne-setuptools ukuphepha iimpazamo ezilandelayo. Ungongeza kwakhona umgca we nvidia-pyindex kuwe requirements.txt ukuba usebenza naloo ndibano.

Qinisekisa ukufakela: iisampuli kunye neemvavanyo

Jonga uguqulelo lomqokeleli nge nvcc -V ukusuka kwiCMD. Cola iisampulu zeCUDA ukusuka github.com/nvidia/cuda-samplesZiqokelele kwaye uziqhube njengoko kubonisiwe kwindawo yokugcina. Kucetyiswa kakhulu ukuba kwakhiwe kwaye kuqhutywe deviceQuery ukuqinisekisa ukuba i-GPU ichongiwe kwaye uqwalaselo luchanekile.

Uvavanyo bandwidthTest iqinisekisa ukusebenza kwesixhobo somkhosi kunye nonxibelelwano. Kuzo zombini iimeko, into ebalulekileyo kukuba isixhobo sibonakala.Ihambelana ne-hardware yakho kwaye iphumelele iimvavanyo. Ukuba deviceQuery I-GPU ayifunyaniswanga, jonga ukufakwa komqhubi kunye nokuba isixhobo siyasebenza kwinkqubo.

I-CUDA kunye ne-WSL ngaphakathi Windows 11 kunye ne-Windows 10 21H2+

Windows 11 kunye neenguqulelo zamva nje ze Windows 10 (i-21H2 nasemva koko) ixhasa i-GPU-isantya esiphezulu se-ML izixhobo usebenzisa i-WSL. Uya kuba nakho ukusebenzisa i-TensorFlow, i-PyTorch, i-Docker, kunye ne-NVIDIA Container Toolkit ngaphakathi kwe-glibc-based Linux distro (Ubuntu, Debian, njl.).

Amanyathelo aphambili: Faka umqhubi we-NVIDIA CUDA-enabled kwi-WSL, uvule i-WSL, kwaye wongeze ukuhanjiswa kohlobo lwe-Ubuntu/Debian. Qinisekisa ukuba unoguqulelo lwamva nje lwe-WSL kernel. (ubuncinci 5.10.43.3). Ungaqinisekisa oku kwi-PowerShell nge: wsl cat /proc/versionOkulandelayo, landela i-CUDA ye-NVIDIA kwisikhokelo somsebenzisi we-WSL ukuze usebenze ne-NVIDIA Docker okanye ufake i-PyTorch/TensorFlow ngaphakathi kwe-distro.

Iitshekhi kunye noxilongo kwi-Windows

Ukufumanisa ukuba yeyiphi i-GPU onayo: kwiiSetingi> Inkqubo> Umboniso> Useto oluphezulu, uya kufumana uphawu kunye nemodeli phantsi kwe “Display information”. KuMphathi woMsebenzi, ithebhu yokuSebenzaKhetha i-GPU ukujonga ukusetyenziswa, imodeli, kunye nememori. Ukuba ayibonakali, khangela uMphathi weDivaysi> Bonisa iiadaptha kwaye ufake umqhubi ofanelekileyo wekhadi.

Ukubona inguqulo yeCUDA "exelwe" ngumqhubi wakho, sebenzisa nvidia-smi. Ukufumana uguqulelo lwe-Toolkit efakiweyo compiler, isebenzisa nvcc --version. Gcina engqondweni oko nvidia-smi Ibonisa ubuninzi be-CUDA "API" uguqulelo oluxhaswa ngumqhubi, hayi olusuka kwi-Toolkit onayo kwidiski.

Ukusetyenziswa kwamafu: umzekelo wemizekelo ngeGPU

Ukuba ufuna amandla afunekayo, ababoneleli ngamafu babonelela ngeemeko ngee-GPU ezifana ne-NVIDIA A100, RTX 4090, okanye i-A6000. Le ndlela ivumela ukuhlawula-njengoko-uhamba, ukuthunyelwa ngokukhawuleza, kunye neetemplates esele zenziwe. yezakhelo ezidumileyo (PyTorch, TensorFlow). Ngomthwalo omkhulu wemisebenzi okanye ii-spikes zexeshana, yindlela esebenzayo ngaphandle kokutyala imali kwi-hardware yobunikazi kunye nenkxaso yeenguqulelo zeCUDA zamva nje.

Izaziso ezisemthethweni kunye neempawu zorhwebo

Amaxwebhu eNVIDIA kunye nesoftware zinikezelwe "njengoko zinjalo," kunye nawo onke amalungelo agciniwe kwaye ngaphandle kweziqinisekiso ezichaziweyo zokurhweba okanye ukulungela injongo ethile. I-NVIDIA inokuguqula iinkcukacha kunye namaxwebhu ngaphandle kwesaziso sangaphambiliSoloko uqinisekisa ulwazi lwamva nje kwaye uhloniphe intengiso yomntu wesithathu kunye nemigaqo yelayisenisi. I-OpenCL luphawu lwentengiso lwe-Apple Inc. olunelayisensi eya e-Khronos; I-NVIDIA kunye nelogo yayo ziimpawu zorhwebo ezibhalisiweyo eU.S. nakwamanye amazwe.

Ngayo yonke le nto ingasentla, ngoku unamaqhekeza okufaka, ukuqwalasela, kunye nokuqinisekisa iCUDA kwiWindows (kwaye nakwiWSL okanye kwiLinux xa uyifuna). Ukususela kwiimfuno kunye nabaqhubi ukudibanisa kunye ne-Visual Studio kunye nokuqinisekiswa kunye neesampuliInto ephambili kukulungelelanisa inguqulo yomqhubi, i-toolkit, kunye nokusingqongileyo, kunye nokuxhomekeka kwizixhobo ezifana nvidia-smi, nvcc kunye neesampuli zokuqinisekisa ukuba lonke ikhonkco lisebenza njengoko kulindelekile.