Kanye ne-processor core entsha I-ARM yethule iphrosesa yezithombe eyenzelwe isizukulwane esilandelayo samasistimu e-single-chip. I-Mali-G77, okungafanele iphanjaniswe neprosesa entsha yokubonisa , iphawula inguquko isuka ekwakhiweni kwe-ARM Bifrost iye e-Valhall.
I-ARM imemezela ukwenyuka okuphawulekayo kokusebenza kwezithombe ze-Mali-G77 - ngo-40% uma kuqhathaniswa nesizukulwane samanje se-Mali-G76. Lokhu kuzuzwe kokubili ngenqubo yobuchwepheshe kanye nokuthuthukiswa kwezakhiwo. I-Mali-G77 ingaba nama-cores angu-7 kuya kwangu-16 (ukukala kusuka ku-1 kuya ku-32 kungenzeka esikhathini esizayo), futhi ngayinye yazo icishe ifane ne-G76. Ngakho-ke, ama-smartphones asezingeni eliphezulu cishe azofaka inani elifanayo lama-GPU cores.
Emidlalweni, ungalindela ukuthuthukiswa kokusebenza okuphakathi kuka-20 no-40%, kuye ngohlobo lomsebenzi wezithombe. Uma sibheka imiphumela yohlolo oludumile lwe-Manhattan GFXBench, ukuphakama okuphawulekayo kwe-GPU entsha kunesizukulwane samanje kuzophoqa imbangi i-Qualcomm ukuthi ikhathazeke ngokuthuthuka okuphawulekayo ekusebenzeni kwezithombe ze-Adreno.
Ngokwaso, i-architecture entsha ye-Mali-G77 iletha ukuthuthuka okungamaphesenti angama-30 ekusebenzeni kahle kwamandla noma ukusebenza, kusho i-ARM. Isizukulwane sesibili se-ARM Valhall scalar architecture sivumela i-GPU ukuthi yenze imiyalelo engu-16 ngomjikelezo ngamunye ngokuhambisana ku-CU, uma kuqhathaniswa nesishiyagalombili ku-Bifrost (Mali-G76). Okunye okuqanjwe kabusha kuhlanganisa ukuhlela iziyalezo eziqhutshwa yihadiwe ngokugcwele kanye nesethi yemiyalo entsha ngokuphelele kuyilapho kugcinwa ukuhambisana okusemuva ne-Bifrost. Ukusekelwa kwefomethi yokuminyanisa ye-ARM AFBC1.3 nokunye okusha (okuqondiwe kwe-FP16, ukunikezwa okunezingqimba kanye nokuphumayo kwe-vertex shader) nakho kungeziwe.
I-Bifrost CU iqukethe izinjini zokubulala ezingu-3, ngayinye ehlanganisa inqolobane yemiyalelo, irejista, kanye neyunithi yokulawula ye-Warp. Ukusabalalisa kuzo zonke lezi zinjini kuvumele imiyalelo engu-24 ye-FMA ukuthi isetshenziswe ku-32-bit floating point precision (FP32). E-Valhall, i-CU ngayinye inenjini eyodwa kuphela yokukhipha, ehlukaniswa phakathi kwamayunithi amabili ekhompyutha akwazi ukucubungula imiyalelo ye-Warp eyi-16 ngewashi ngalinye, okuholela ekuphumeni okuphelele kweziyalezo ezingu-32 ze-FMA FP32 nge-CU ngayinye. Ngenxa yalezi zinguquko zezakhiwo, i-Mali-G77 ingakwazi ukwenza izibalo zezibalo ezingaphezulu kwengxenye eyodwa kwezintathu ngezibalo ezifanayo uma kuqhathaniswa ne-Mali-G76.
Ukwengeza, ngayinye yalawa ma-CU iqukethe amabhulokhi amabili amasha okusebenza kwezibalo. Injini entsha yokuguqula (i-CVT) iphethe inombolo eyisisekelo, enengqondo, yegatsha, nemiyalo yokuguqulwa. I-Special Function Unit (SFU) isheshisa ukuphindaphindeka okuphelele, ukwahlukanisa, impande eyisikwele, ama-logarithm, neminye imisebenzi eyinkimbinkimbi ephelele.
Ibhulokhi ye-FMA evamile inezilungiselelo ezimbalwa ezisekela imiyalelo engu-16 FP32 ngomjikelezo ngamunye, 32 ye-FP16, noma 64 yomkhiqizo we-INT8 Dot. Lokhu kulungiselelwa kunganikeza ukuthuthukiswa kokusebenza okungafika kokungu-60% ezinhlelweni zokufunda zomshini.
Olunye ushintsho olubalulekile ku-Mali-G77 ukuphindwa kabili kokusebenza kwenjini yokuthungwa, manje esecubungula ama-texels ama-bilinear ama-4 iwashi ngalinye uma kuqhathaniswa namabili adlule, ama-texel angu-trilinear angu-2 iwashi ngalinye, okuvumela ukuhlunga kwe-FP16 ne-FP32 ngokushesha.
I-ARM yenze ezinye izinguquko eziningi, i-Mali-G77 ne-Valhall zithembisa ukuthuthukiswa okubalulekile kokusebenza kwemidlalo nokufunda kwemishini. Okubalulekile, ukusetshenziswa kwamandla nendawo ye-chip kugcinwa kumazinga e-Bifrost, okuthembisa amadivayisi eselula anokusebenza okuphezulu okuphezulu ngaphandle kokwenyuka kokusetshenziswa kwamandla, ukunqanyulwa kokushisa kanye nezidingo zosayizi.
Source: 3dnews.ru