Parallelization of Minimal Problem Solver Generator
Paralelizace generátoru minimálních solverů
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učení technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date
Abstract
Navrhli jsme paralelizaci minimalního solveru pro RANSAC. Optimální využití
GPU je dosaženo řešením několika problémů zároveň. Hlavní pozornost věnujeme
výpočtu vlastních čísel, neboť je to časově nejnáročnější část solveru.
Zkoumáním RANSACU jsme zjistili, že verifikace představuje potenciální
zdroj zrychlení paralelizací.
Porovnali jsme implementaci v CUDA C/C++ se seriovym řešením. Pracovali
jsme s pěti bodovým problémem relativní polohy.
Minimální solver je až dvakrát rychlejší na hybridním systému (GPU + jedno
jádro CPU) než na jednom jádru CPU. Proces verifikace je pro zvolená data až
devadesátkrát rychlejší na GPU než na jednom jádru CPU.
We proposed a parallelization of the minimal solvers intended for the RANSAC scheme. To utilize the GPU fully, we are solving several instances at once. We op- timize the computation of the eigenvectors because it is the most time-consuming part of the solver. We also examined the other parts of RANSAC and found that the verification process has much greater potential for parallelization. We implemented both improvements in CUDA C/C++ and compared the results with serial implementation. The selected minimal problem was the five- point relative pose problem. The minimal solver is often more than two times faster in a hybrid system (GPU + single-core CPU) than on a single-core CPU. The verification process is about 90 times faster on GPU than on a single-core CPU for the selected data set.
We proposed a parallelization of the minimal solvers intended for the RANSAC scheme. To utilize the GPU fully, we are solving several instances at once. We op- timize the computation of the eigenvectors because it is the most time-consuming part of the solver. We also examined the other parts of RANSAC and found that the verification process has much greater potential for parallelization. We implemented both improvements in CUDA C/C++ and compared the results with serial implementation. The selected minimal problem was the five- point relative pose problem. The minimal solver is often more than two times faster in a hybrid system (GPU + single-core CPU) than on a single-core CPU. The verification process is about 90 times faster on GPU than on a single-core CPU for the selected data set.