20.500.12556/DKUM-77511 Algoritem za učinkovit izračun verige elementarnih morfoloških filtrov na centralni procesni enoti doktorska disertacija Algorithm for efficient computation of elementary morphological filter chain on central processing unit V doktorski disertaciji predstavimo nov algoritem za učinkovit izračun verige elementarnih filtrov na centralni procesni enoti. Verige filtrov so temeljni del tako imenovanih geodetskih operatorjev, ki uporabljajo postopek morfološke rekonstrukcije. Ta iterativno filtrira in omejuje vrednosti filtrirane slike. Dobljeno učinkovitost razvitega algoritma delimo na dva dela. Prvi del se zanaša na učinkovit enonitni izračun elementarnih morfoloških filtrov, kar dosežemo s procesiranjem na mestu, dekompozicijo strukturnega elementa in uporabo vektorskih registrov za pohitritev izračuna. Drugi del pa razvite filtre uporabi za učinkovit vzporeden in hkraten izračun več filtrov v podani verigi. Vzporedni izračun je osnovan na emulaciji podatkovno pretokovnega procesiranja, kjer z analizo topologije predpomnilnika CPE in primernim pripenjanjem niti zagotovimo, da prenos podatkov med jedri CPE poteka preko predpomnilnika. Pokazali smo, da dosežemo največjo prepustnost procesiranja ob uporabi več neodvisnih vzporednih procesnih cevovodov, medtem ko najhitrejše odzivne čase dosežemo z uporabo enega cevovoda. Primerjava z algoritmi stanja tehnike in odprtokodnimi knjižnicami je pokazala, da je predlagani algoritem vedno dosegel boljše računske čase pri obdelavi verig filtrov. V primerjavi z iterativnim izračunom verige filtrov na splošnonamenskih grafičnih procesnih enotah se je predlagani algoritem prav tako izkazal za bistveno hitrejšega, tudi ko smo uporabili CPE nižjega cenovnega ranga. In this doctoral dissertation we introduce a new algorithm for efficient calculation of filter chains on central processing unit. Filter chains are a fundamental part of the so-called geodesic operators, which use morphological reconstruction. The latter iteratively filters and bounds the values of the input image. The efficiency of the developed algorithm is compared from two points of view. The first relies on the efficient single-threaded calculation of elementary morphological filters, which is achieved by in-place processing, decomposition of the structuring element and the use of vector registers to speed up calculations. The second part uses the developed filters to efficiently process multiple filters from a given filter chain in parallel. The parallel calculation is based on emulating dataflow processing, where analysis of CPU's cache topology and appropriate thread pinning ensures optimal data transfers between threads via the hierarchical cache. We have shown that the highest processing throughput is achieved using several independent processing pipelines in parallel, while the lowest processing latency is obtained with a single pipeline. A comparison with state-of-the-art algorithms and open source libraries showed that the proposed algorithm always achieves better computational times when processing filter chains. A comparison of computation times of filter chains using general-purpose graphics processing units showed that the proposed algorithm also significantly outperforms them, even when using a lower priced CPU. morfološki filtri geodetski operatorji veriga filtrov obdelava slik vzporedno procesiranje podatkovno pretočna obdelava vektorsko procesiranje SIMD centralna procesna enota morphological filters geodesic operators filter chain image processing parallel processing dataflow processing vector processing SIMD central processing unit true true false D. Žlaus Slovenski jezik Angleški jezik Doktorsko delo/naloga 2020-09-02 09:47:47 2021-04-13 10:53:49 2023-01-31 10:38:51 2020-09-03 03:20:17 2021 0 Maribor 0 Maribor VI, 123 str. 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.925.8:519.254(043.3) 59591683 URN:SI:UM:DK:V4UXNCNU DOK_Zlaus_Danijel_2021.pdf DOK_Zlaus_Danijel_2021.pdf 1 B680368471BAE7CD0A570CF7716F6887 9803435791bb14e6e7feac735fa4973e2626de32e92b1965f5783440052e6d19 7457da57-9c2d-11eb-b78a-00155d0001ca 20.500.12556/dkum/e2769bd8-26aa-4d19-9e7d-f26b02b7f274 https://dk.um.si/Dokument.php?lang=slv&id=149746 DOK_Zlaus_Danijel_2021.zip DOK_Zlaus_Danijel_2021.zip 1 3C80D1FC2AA2E5C39E0B62FBB899F9A2 6b240698218208d57eee07136ed2ccdadebd77c11220e7ecd44b281027c6c644 243149c5-922f-11eb-bc2e-00155d0001ca 20.500.12556/dkum/c614d117-7613-46d6-b564-4803857cb489 https://dk.um.si/Dokument.php?lang=slv&id=149589 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0