ZLUDA inomhanya pane AMD's ROCm stack
Nhau yakaburitswa iyoyo AMD yaita danho rekumisa mari yekuvandudza ZLUDA, zvinova kushandiswa kwakavhurika kweCUDA tekinoroji ye AMD GPUs, ichikubvumidza kuti umhanye CUDA zvikumbiro usingazvigadzirise nekuita padhuze neiyo yemaapplication ari kushanda pasina mamwe mataira.
Mukati memakore maviri apfuura, basa rakaitwa pakuvandudzwa kweiyi layer kufambisa kuitwa kweCUDA maapplication pane masisitimu ane AMD GPUs. Asi kusvika munguva pfupi yapfuura AMD yakaita sarudzo iyo mhanyisa CUDA zvikumbiro pa AMD GPUs yakanga isiri yezvido zvekutengeserana, izvo zvakabvumira mugadziri kuvhura budiriro yayo sezvakabvumiranwa muchibvumirano.
About ZLUDA
Iyo purojekiti yakatanga kugadzirwa kutsigira basa reCUDA paIntel GPUs uye shanduko yegwaro rekutsigira reGPU imhaka yekuti pakutanga mugadziri weZLUDA aive mushandi weIntel. Muna 2021, Intel yakafunga kuti yakanga isingaite zvekutengesa kupa kugona kumhanyisa CUDA zvikumbiro paIntel GPUs, izvo zvakakonzera kumisa kusimudzira kwekutanga.
Mukutanga kwa2022, mugadziri akasiya Intel uye akasaina chibvumirano ne AMD kuti aenderere mberi nebudiriro yeCUDA inoenderana layer. Panguva yebudiriro, AMD yakakumbira kuti kufarira kwekambani iyi chirongwa cheZLUDA kusasimudzirwe uye kuti isazvipira kuZLUDA inochengeterwa veruzhinji.
ZLUDA inopa rubatsiro rwebhinari yeCUDA zvikumbiro aripo akaunganidzwa neCUDA compiler yeNVIDIA GPUs, pasina chikonzero chekuchinja padanho rekodhi kodhi. Iko kuita kunoshanda pa AMD's ROCm stack uye iyo HIP yekumhanya nguva.
Mumamiriro ayo azvino, ZLUDA's shanduko yemhando yepamusoro inoonekwa sealpha vhezheni. Kunyangwe zvakakodzera kutaura kuti yave kutokwanisa kumhanyisa akawanda eCUDA zvikumbiro, pamusoro pekupa rutsigiro rushoma rwemaprimitives uye maraibhurari, ine zvakare prototype kuita kweOptiX chimiro.
ZLUDA Comparison Chart
Panguva yekutanga kuitiswa kweCUDA zvikumbiro pasi peZLUDA kutonga unoona kunonoka kunooneka nekuti ZLUDA inounganidza kodhi yeGPU. Zvisinei, pakumhanya kunotevera uku kunonoka kunopera, sezvo iyo kodhi yakaunganidzwa yakavharirwa, zvichikonzera kuita kwepedyo-kwenyika paunenge uchimhanyisa kodhi yakaunganidzwa. Semuenzaniso, paunenge uchimhanyisa Geekbench pa AMD Radeon 6800 XT GPU, iyo ZLUDA-yakavakirwa vhezheni yeCUDA benchmark suite yakaratidza kuita zvirinani zvakanyanya kupfuura OpenCL-based version.
Mukuwedzera, zvinonzi kutsigirwa kweCUDA mutyairi API uye reverse engineering yeCUDA API isina kunyorwa inoshandiswa muZLUDA nekutsiva mafoni ekushanda nemabasa akafanana anopiwa muHIP runtime, iyo yakafanana zvikuru neCUDA.
Semuenzaniso, iyo basa cuDeviceGetAttribute() inotsiviwa nehipDeviceGetAttribute(). ZLUDA zvakare inovimbisa kuenderana neNVIDIA maraibhurari akadai seNVML, cuBLAS uye cuSPARSE. Kune aya maraibhurari, ZLUDA inopa maraibhurari eshanduro ane zita rimwechete uye yakafanana ficha yakaiswa, yakavakirwa pamusoro peakafanana AMD raibhurari. Iyo GPU application kodhi, yakaunganidzwa muPTX (Parallel Thread Execution) inomiririra, inodudzirwa neakakosha compiler kune yepakati LLVM IR inomiririra, kubva panogadzirwa iyo binary kodhi ye AMD GPU.
Chekupedzisira, zvinofanirwa kutaurwa izvo Ramangwana rechirongwa rinoenderana nekufarira kwenharaunda uye zvingangoitwa zvikumbiro zvekushandira pamwe kubva kune mamwe makambani. Pasina rutsigiro rwekunze, chirongwa ichi chicharamba chichikura munzvimbo dzekufarira kwemunyori, senge DLSS.
For the kufarira kodhi yeprojekiti, iwe unofanirwa kuziva kuti yakanyorwa muRust uye yakagoverwa pasi peMIT neApache 2.0 marezinesi, uye chirongwa ichi chinoenderana neLinux neWindows. Kana iwe uchida kudzidza zvakawanda nezvazvo, unogona kubvunza nezve ruzivo Mune inotevera chinongedzo.