Abathuthukisi bephrojekthi I-LLVM ikhishwe isanda kukhipha i-Heterogeneous Parallel Virtual Machine compiler (HPVM) 1.0, ehlose yenza lula izinhlelo zezinhlelo ezahlukahlukene futhi inikeze amathuluzi wokukhiqiza amakhodi we-CPU, i-GPU, i-FPGA, nama-accelerator wehadiwe i-domain ethize (ukusekelwa kwe-FGPA nama-accelerators akufakiwe kunguqulo 1.0).
Umqondo oyinhloko ngemuva kwe-HPVM ukuhlanganisa ukumelwa okuhlangene kwezinhlelo ezisebenzisekayo ngasikhathi sinye engasetshenziselwa ukusebenza kuzinhlobonhlobo zezingxenyekazi zekhompiyutha ezihambisanayo, kufaka phakathi ama-GPU, imiyalo ye-vector, ama-processor we-multi-core, ama-FPGA, kanye nama-chip accelerator ahlukahlukene akhethekile.
Ukuhlelela izinhlelo ezihambisanayo ezixakile kuyinkimbinkimbi ngokuba khona kohlelo lwezinto (ama-CPU cores, imiyalo ye-vector, ama-GPU, njll.) Asebenzisa amamodeli ahlukahlukene ukufeza ukufana, amasethi wokufundisa ahlukile kanye nezinhlu zememori ezihlukile, futhi uhlelo ngalunye lunenhlanganisela izingxenye ezinjalo ziyashintsha.
I-Illinois LLVM Compiler Research Group iyajabula ukumemezela ukukhishwa komthombo ovulekile we-HPVM (inguqulo 1.0). I-HPVM iyi-compiler yohlaka oluqondisa kabusha ama-CPU, ama-GPU, nama-accelerators (lokhu kukhishwa akubandakanyi ukuxhaswa kwe-accelerator) [1]. I-HPVM isebenzisa isihlanganisi esizimele esihlanganayo i-IR eselula i-LLVM 9.0.0 compiler IR ngesethulo sokugeleza kwedatha okucacile okuthatha imisebenzi, idatha, kanye nokufana kwepayipi.
Le nguqulo ingeze okubalulekile enguqulweni yethu yokuqala (inguqulo engu-0.5), engeza ukusekelwa kokusebenza kwe-algebra tensor, i-Pytorch ne-Keras interface, ukulinganiselwa kwabasebenza i-convolution, nohlaka olusebenzayo noluguquguqukayo lokulinganisa okulinganayo.
Kumele kuqashelwe ukuthi ukusetshenziswa kwe- I-HPVM ingafinyelela ukuzuza okuphawulekayo kokusebenza kusukela eUkusebenza kokukhipha kwabahumushi be-HPVM kuqhathaniswa nokwikhodi ye-OpenCL ibhalwe ngesandla kumadivayisi wekhompuyutha we-GPU kanye ne-vector. Ngokungafani nezinye izinhlelo, i-HPVM izame ukuhlanganisa amathuba amathathu okuhlela i-heterogeneous computing: ukumelwa kolimi- ne-hardware-ezimele okuphakathi, imiyalo ebonakalayo yokusetha izakhiwo (i-V-ISA), kanye nohlelo lwesikhathi sokusebenza.
Isethulo esiphakathi esizimele (IR) yohlelo oluqondiwe nolimi lohlelo olusetshenziswe ku-HPVM lususelwa ekumelelekeni okuphakathi kwemiyalo ye-LLVM 9.0 futhi ilinweba ngegrafu yokugeleza kwedatha okulandelanayo ukumboza ukufana emsebenzini, idatha, nasezingeni lamapayipi wokubala.
I-HPVM Lesisemkhatsini ibuye ifake imiyalo ye-vector nememori eyabiwe. Umgomo omkhulu wokusebenzisa ukumelwa okuphakathi ukwenziwa kwekhodi nokusebenza kahle kwezinhlelo ezihlukile.
I-Virtual Instruction Set Architecture (V-ISA) ifaka izici ezisezingeni eliphansi zezingxenyekazi zekhompiyutha futhi ihlanganisa izindlela ezahlukahlukene zokuqhathanisa nokwakhiwa kwememori kusetshenziswa kuphela imodeli eyisisekelo ye-concurrency, igrafu yokuhamba kwedatha. I-V-ISA ivumela ukufezekisa ukuthwala phakathi kwezinhlobo ezahlukahlukene zemishini yokusebenza ngekhompyutha futhi ivumela ukungalahlekelwa ukusebenza lapho kusetshenziswa izinto ezahlukahlukene zezinhlelo ezixakile. I-Virtual ISA nayo ingasetshenziselwa ukuhlinzeka ngekhodi yohlelo olusebenzisekayo jikelele engaqhutshwa kusetshenziswa i-CPU, i-GPU, i-FPGA, nama-accelerator ahlukahlukene.
Esigabeni samanje sentuthuko, I-HPVM inikeza ama-generator ekhodi akwazi ukuhumusha ama-node wohlelo lokusebenza achazwe yi-ISA I-Virtual ukusebenzisa i-NVIDIA GPUs (cuDNN ne-OpenCL), imiyalo ye-Intel AVX vector, nama-CPUs ama-x86 amaningi. }
Ngesikhathi sokwenziwa, i-HPVM isebenzisa izinqubomgomo zokuhlela eziguquguqukayo zenqubo yekhompiyutha, zenziwe zombili ngesisekelo solwazi mayelana nohlelo (isakhiwo sokuqhafaza) nangokuhlanganiswa kwamanothi wohlelo ngamanye wokwenza kunoma yimaphi amadivayisi wekhompyutha lapho eya khona atholakala ohlelweni.
Uma kuqhathaniswa nenguqulo yokuqala yokubuka kuqala, I-HPVM 1.0 ifaka ukusekelwa kwemisebenzi eqondile ye-algebra tensor, izixhumi esibonakalayo sePytorch neKeras nohlaka lokushuna lokulinganisa olukhetha ngokuzenzakalela ukulinganisa okulungile kwemisebenzi ethile ye-tensor futhi lukhethe ukulungiselelwa kokusebenza okulungile.
Ekugcineni, uma unentshisekelo yokwazi okwengeziwe ngalo mhlanganisi, Ungahlola imininingwane kusixhumanisi esilandelayo.