Wannan aikin An buɗe VeriGPU 'yan kwanakin da suka gabata ya fara aiki don haɓaka GPU mai buɗewa, wanda ke nufin haɓakawa ƙarƙashin yaren ƙirar ƙira da bayanin tsarin lantarki Verilog.
Ga waɗanda ba su da masaniya da VeriGPU, ya kamata ku sani cewa an sanya shi azaman Ƙaƙwalwar Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararren Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙwararrun Ƙirar Ƙididdiga ta Ƙarfafa Ƙididdiga masu dangantaka da tsarin ilmantarwa na inji.
hanyar farawa, Ana haɓaka aikin ta amfani da na'urar kwaikwayo ta Verilog, amma bayan an shirya, ana iya amfani dashi don samar da kwakwalwan kwamfuta na gaske.
Tsare-tsare sun haɗa da goyan baya ga tsarin ilmantarwa mai zurfi na PyTorch da ikon haɓaka aikace-aikace don VeriGPU ta amfani da HIP (Hanyoyin Ƙididdigar Ƙididdigar Ƙididdigar Ƙididdiga) API. A nan gaba, ba a cire haɗin haɗin gwiwa na wasu APIs, kamar SYCL da NVIDIA CUDA.
Yana da mahimmanci a faɗi hakan ci gaban wannan GPU yana nufin horar da na'ura kai tsaye. Don haka yakamata ya dace da goyan bayan tsarin koyon injin na yanzu kamar PyTorch da Tensorflow, wannan yana nufin tabbas tabbas yana buƙatar tallafawa NVIDIA CUDA ko AMD HIP.
Ko da yake kuma an ambaci hakan Na san zan iya aiwatar da hanyar sadarwa ta OpenCL ko SYCL, kodayake goyan bayan manyan tsare-tsare a halin yanzu yana iyakance. Akwai keɓaɓɓen tsarin ilmantarwa mai zurfi na OpenCL a cikin DeepCL, amma yana da ƙayyadaddun ƙayyadaddun saitin hanyoyin sadarwa na jijiyoyi da yuwuwar hanyoyin sadarwa, idan aka kwatanta da PyTorch da Tensorflow.
A halin yanzu, ba mu da niyyar aiwatar da aiwatar da kisa ba tare da izini ba, watau fara umarni ɗaya kafin farkon wanda ya gabata, saboda wannan yana da wayo a yanayin yanayin koyarwa guda ɗaya na multithreading (SIMT), kuma saboda wannan yana amfani da yanki mai mutuƙar mutu. , sabili da haka kadan ne a lamba (ko mafi tsada).
A gefe guda kuma, za mu aiwatar da aiwatar da aiwatar da umarni guda ɗaya, inda za mu fara koyarwa yayin da koyarwar da ta gabata tana ci gaba da aiwatarwa. Wannan daidaitaccen nauyi ne kuma mara nauyi, baya ɗaukar yanki mai mutuƙar yawa.
Ba a aiwatar da caching kowane nau'i a halin yanzu (babu matakin 1, babu matakin 2, babu matakin 3, har ma da caching na umarni: P). Tun da na yi niyyar yin GPU, wanda ke da tsarin cache daban-daban fiye da CPU, zan yi tunanin wannan da zarar ya fara kama da GPU.
GPU yana tasowa bisa tsarin umarni na RISC-V, amma sakamakon tsarin tsarin koyarwa na GPU na ciki bai dace da RISC-V ISA ba, tun da yake a cikin yanayin da tsarin GPU bai dace da ma'anar RISC-V ba, ban da saita aikin kiyaye daidaituwa tare da RISC-V. .
Tunda cigaba yana mai da hankali kan iyawar da ake buƙata don tsarin koyon injin, don haka, don rage girman da rikitarwa na guntu array, kawai tsarin BF16 floating-point ana amfani da shi, kuma kawai ayyuka masu iyo kawai waɗanda ke buƙatar koyon inji, irin su exp, log, tanh, da sqrt. samuwa.
Daga cikin abubuwan da aka riga aka samu, alal misali, direban GPU, APU (Accelerated Processing Unit) don ayyukan haɗin gwiwa («+»,»-«,»/»,»*»), naúrar don ayyukan iyo ayyuka («+», »*») da sashin reshe, da kuma tallafi ga mai tara HIP.
Don ƙirƙirar apps Ana ba da mai tarawa da goyan baya don haɗa lambar C++ ta hanyar LLVM. Daga cikin abubuwan da ake sa ran, aiwatar da umarni guda ɗaya, caching bayanai da ƙwaƙwalwar koyarwa, ayyukan SIMT (Umar guda ɗaya da Multiple Thread) ta bambanta.
A ƙarshe, idan kuna sha'awar ƙarin sani game da shi game da ci gaban wannan buɗaɗɗen tushen GPU, ya kamata ku san cewa ana rarraba abubuwan ci gaban aikin a ƙarƙashin lasisin MIT kuma kuna iya tuntuɓar lambar, da kuma ci gaban aikin daga. mahada mai zuwa.