Analiza performanci referentnih OpenCL implementacija

undergraduate thesis

• Title (english): Performance Analysis of OpenCL Benchmark Implementations
• Author: Ela Grga
• Mentor: Vlado Sruk (mentor)
• Committee member: Vlado Sruk (predsjednik povjerenstva)
• Committee member: Nikolina Frid (član povjerenstva)
• Committee member: Hrvoje Nuić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-09-12, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Primarna svrha grafičkih procesorskih jedinica je obrada računalne grafike, odnosno slika i videa. Grafička procesorska jedinica sastoji se od mnogo procesorskih elemenata koji imaju mogućnost raznih računalnih operacija. Iako centralna procesorska jedinica može brže izvršiti jednu operaciju zbog veće brzine rada, GPU je taj koji množe istodobno izvoditi mnogo operacija.
Mogućnost paralelnog izvođenja zadataka i izračuna čini GPU korisnim izvan računalne grafike. Jedan od načina iskorištavanja mogućnosti grafičke procesorske jedinice je OpenCL.
OpenCL radni je okvir koji se koristi za paralelizaciju koda na uređajima poput GPU, FPGA pa čak i CPU. Moguće je koristiti OpenCL za paralelizaciju zadataka i paralelizaciju podataka odnosno paraleliziranu obradu skupine podataka.
OpenCL kod napisan je u dva dijela, host kod i kod jezgrenih funkcija. Jezgrena funkcija predstavlja dio koda koji se želi paralelizirati. OpenCL jezgrene funkcije napisane u OpenCL C programskog jezika koji se razlikuje od C programskog jezika u ugradbenim funkcijama i dodatnim tipovima podataka. Ostatak koda može biti napisan u C ili C++ jezicima.
Host kod izvršavat će se na CPU, a kod jezgrene funkcije na odabranom akceleratoru.
OpenCL neće se koristiti za paralelizaciju jednostavnih zadataka koji se mogu vrlo jednostavno i brzo sami po sebi izvršiti ako su napisani u standardnom C-u. OpenCL koristit će se za zadatke znatnih skupova podataka nad kojima se treba izvršiti jedan ili više setova instrukcija. Na primjer računanje vrijednosti elemenata matrice dimenzija 1600x1600.
Parametri GPU ne utječu vidljivo na izvedbu programa, ali je usporedbom vremena izvršavanja jezgrene funkcije pokazalo se da je najpovoljnije prilagodljivo napajanje, uključen shader cache i isključena niska latentnost.

Abstract (english)

Primary purpose of graphics processing unit is computer graphics, that is processing images and videos. Graphics processing unit is made of many processing elements which have the ability to do diverse computations. Although the central processing unit is faster in executing one computational operation due to it's higher clock speed, the GPU is the one that can do many operations simultaneously.
The ability of parallel execution of tasks and computation makes the GPU useful outside of the area of computer graphics. One of the ways to use the graphics processing unit is OpenCL.
OpenCL is a framework used for writing code which can be executed on multiple devices like the GPU, FPGA and even CPU. OpenCL can be used for both task and data parallelism.
OpenCL code is written in two parts, the host and kernel code. The kernel is part of the code which can be parallelized. OpenCL kernels are written in OpenCL C programming language which differs from standard C due to some built in functions and extra data types. The rest of the code can be written either in C or C++ languages.
The host code is executed on the CPU, and the kernel is executed on the chosen accelerator.
OpenCL shouldn't be used for simple tasks or tasks which don't need a whole lot of data. Most of those tasks can be easily written and executed in standard C. OpenCL should be used in tasks that are data heavy, where one set of instructions needs to be done on a large set of data. For example calculating the value of elements of a matrixs with dimensions of 1600x1600.
The parameters of the GPU seemingly do not affect the performance of the program, but after comparing the time needed to run the kernel on the GPU it is shown that the adaptive power management, low latency and the use of the shader cache are favorable.

• Keywords: Grafička procesorska jedinica; paralelizam koda; OpenCL
• Keywords (english): Graphics processing unit; parallelism; OpenCL
• Language: croatian
• URN:NBN: urn:nbn:hr:168:022084
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Repository: FER Repository
• Created on: 2023-01-14 08:52:00