Iphrojekthi ye-ZLUDA ilungiselele ukuqaliswa okuvulekile kobuchwepheshe be-CUDA be-AMD GPUs, okuvumela ukuthi usebenzise izinhlelo zokusebenza ze-CUDA ezingashintshiwe ngokusebenza eduze nokusebenza kwezinhlelo zokusebenza ezisebenza ngaphandle kwezendlalelo. Ikhithi yamathuluzi eshicilelwe inikeza ukuhambisana kanambambili nezinhlelo zokusebenza ezikhona ze-CUDA ezihlanganiswe kusetshenziswa i-CUDA compiler ye-NVIDIA GPUs. Ukuqaliswa kusebenza phezu kwesitaki se-ROCm kanye ne-Runtime HIP (I-Heterogeneous-computing Interface for Portability) eyakhiwe yi-AMD. Ikhodi yephrojekthi ibhalwe nge-Rust futhi isatshalaliswa ngaphansi kwamalayisensi e-MIT kanye ne-Apache 2.0. Isekela umsebenzi ku-Linux ne-Windows.
Isendlalelo sokuhlela umsebenzi we-CUDA kumasistimu anama-AMD GPU sakhiwe eminyakeni emibili edlule, kodwa iphrojekthi inomlando omude futhi yadalelwa ekuqaleni ukuze kuqinisekiswe ukuthi i-CUDA isebenza kuma-Intel GPU. Ushintsho kunqubomgomo yokwesekwa kwe-GPU luchazwa ukuthi ekuqaleni umthuthukisi we-ZLUDA wayeyisisebenzi se-Intel, kodwa ngo-2021 le nkampani yacabanga ukunikeza amandla okusebenzisa izinhlelo ze-CUDA kuma-Intel GPU ukuthi angabi nantshisekelo yebhizinisi futhi ayizange isheshise. ukuthuthukiswa kwesinyathelo.
Ekuqaleni kuka-2022, umthuthukisi washiya i-Intel futhi wanikezwa inkontileka yi-AMD ukuze athuthukise isendlalelo sokuhambisana kwe-CUDA. Ngesikhathi sokuthuthukiswa, i-AMD yacela ukuthi ingakhangisi intshisekelo ye-AMD kuphrojekthi ye-ZLUDA futhi ingazibophezeli endaweni yokugcina i-ZLUDA yomphakathi. Eminyakeni emibili kamuva, i-AMD yanquma ukuthi ukusebenzisa izinhlelo ze-CUDA kuma-GPU we-AMD kwakungeyona isithakazelo ebhizinisini, okuyinto, ngokwemibandela yenkontileka, yavumela umthuthukisi ukuba athole umsebenzi wakhe. Njengoba abakhiqizi be-GPU beyekile ukuxhasa le phrojekthi, isiphetho sayo manje sincike entshisakalweni yomphakathi kanye nokwamukela iziphakamiso zokubambisana kwezinye izinkampani. Ngaphandle kokusekelwa kwangaphandle, iphrojekthi izokwazi kuphela ukuthuthuka ngezindlela ezithokozisa umbhali, njenge-DLSS (Deep Learning Super Sampling).
Ngendlela yayo yamanje, izinga lekhwalithi lokusetshenziswa lihlolwa njengenguqulo ye-alpha. Nokho, i-ZLUDA isivele isetshenziselwe ukusebenzisa izinhlelo zokusebenza eziningi ze-CUDA, okuhlanganisa i-Geekbench, i-3DF Zephyr, i-Blender, i-Reality Capture, i-LAMMPS, i-NAMD, i-waifu2x, i-OpenFOAM ne-Arnold. Inikeza ukusekelwa okuncane kwe-cuDNN, i-cuBLAS, i-cusSPARSE, i-cuFFT, i-NCCL ne-NVML yokuqala nemitapo yolwazi.
Ukwethulwa kokuqala kwezinhlelo zokusebenza ze-CUDA ezisebenzisa i-ZLUDA kwenzeka ngokubambezeleka okubonakalayo ngenxa yokuthi i-ZLUDA ihlanganisa ikhodi ye-GPU. Akukho ukubambezeleka okunjalo ekugijimeni okulandelayo, njengoba ikhodi ehlanganisiwe igcinwa kunqolobane. Uma usebenzisa ikhodi ehlanganisiwe, ukusebenza kuseduze nokomdabu. Lapho usebenzisa i-Geekbench ku-AMD Radeon 6800 XT GPU, inguqulo ye-ZLUDA ye-CUDA benchmark suite yenze kangcono kakhulu kunenguqulo ye-OpenCL.
Ukusekelwa kwe-CUDA Driver API esemthethweni kanye nengxenye ehlanekezelwe ye-CUDA API engabhalisiwe isetshenziswa ku-ZLUDA ngokushintsha izingcingo zomsebenzi ngemisebenzi efanayo ehlinzekwe esikhathini sokusebenza se-HIP, esifana ngezindlela eziningi ku-CUDA. Isibonelo, umsebenzi we-cuDeviceGetAttribute() uthathelwa indawo yi-hipDeviceGetAttribute(). Ukuhambisana nemitapo yolwazi ye-NVIDIA efana ne-NVML, cuBLAS kanye ne-cuSPARSE kuqinisekiswa ngendlela efanayo - kumalabhulali anjalo, i-ZLUDA inikeza imitapo yolwazi yokuhumusha enegama elifanayo nesethi efanayo yemisebenzi, eyakhelwe njengezengezo phezu kwemitapo yolwazi ye-AMD efanayo.
Ikhodi yohlelo lokusebenza ye-GPU ehlanganiswe isethulo se-PTX (Parallel Thread Execution) iqala ukuhunyushwa umdidiyeli okhethekile ekumeleleni okuphakathi kwe-LLVM IR, ngesisekelo lapho kukhiqizwa khona ikhodi kanambambili ye-AMD GPUs.
Source: opennet.ru