Simulating fluid behavior has proven to be a demanding challenge which requires complex computational models and highly efficient data structures. Smoothed Particle Hydrodynamics (SPH) is a particle based computational model used to simulate fluid behavior that has been found capable of producing convincing results. However, the SPH algorithm is computational heavy which makes it cumbersome to work with. This master thesis describes how the SPH algorithm can be accelerated by utilizing the GPU’s computational resources. It describes a model for how to distribute the work load on the GPU and presents a suitable data structure. In addition, it proposes a method to represent and handle moving objects in the fluids surroundings. Finally, the performance gain due to the GPU is evaluated by comparing processing times with an identical implementation running solely on the CPU.

Forecasting the stock market is well-known to be a very complex and difficult task, and even by many considered to be impossible. The new model, emph{Historical Consistent Neural Networks} (HCNN), has recently been successfully applied for prediction and risk estimation on the energy markets. HCNN is developed by Dr. Hans Georg Zimmermann, Siemens AG, Corporate Technology Dpt., Munich, and implemented in the SENN (Simulation Environment for Neural Network) package, distributed by Siemens. The evalution is made by tests on a large database of historical price data for global indicies, currencies, commodities and interest rates. Tests have been done, using the Linux version of the SENN package, provided by Dr. Zimmermann and his research team. This thesis takes on the task given by Eturn Fonder AB, to develop a sound basis for evaluating and using HCNN, in a fast and easy manner. An important part of our work has been to develop a rapid and improved implementation of HCNN, as an interactive software package. Our approach has been to take advantage of the parallelization capabilities of the graphics card, using the CUDA library together with an intuitive and flexible interface for HCNN built in MATLAB. We can show that the computational power of our CUDA implementation (using a cheap graphics device), compared to SENN, is about 33 times faster. With our new optimized implementation of HCNN, we have been able to test the model on large data sets, consisting of multidimensional financial time series. We present the results with respect to some common statistical measures, evaluates the prediction qualities and performance of HCNN, and give our analysis of how to move forward and do further testing.

Image completion is a process of removing an area from a photograph and replacing it with suitable data. Earlier methods either search for this relevant data within the image itself, or extends the search to some form of additional data, usually some form of database. Methods that search for suitable data within the image itself has problems when no suitable data can be found in the image. Methods that extend their search has in earlier work either used some form of database with labeled images or a massive database with photos from the Internet. For the labels in a database to be useful they typically needs to be entered manually, which is a very time consuming process. Methods that uses databases with millions of images from the Internet has issues with copyrighted images, storage of the photographs and computation time. This work shows that a small database of the user’s own private, or professional, photos can be used to improve the quality of image completions. A photographer today typically take many similar photographs on similar scenes during a photo session. Therefore a smaller number of images are needed to find images that are visually and structurally similar, than when random images downloaded from the internet are used. Thus, this approach gains most of the advantages of using additional data for the image completions, while at the same time minimizing the disadvantages. It gains a better ability to find suitable data without having to process millions of irrelevant photos.

Thesis focuses on analysis of contemporary machine learning methods used for text classification based on emotion and testing several deep neural nework architectures. Outcome of this thesis is a neural network architecture, which is tuned for using with text data and which had the best result of 79,94 percent. Proposed method is language independent and it doesn’t require as precisely classified training datasets as current methods. Training and testing datasets were consisted of short amateur movie reviews in Czech and in English. Thesis contains also overview of theoretical basics for convolutional neural networks and history of neural networks and language processing Scripts were written in Python, neural networks were simulated using Keras library and Theano framework. We used CUDA for better performance.

The presented work highlights the remarkable potential of physical discretization – lattice model FyDiK in three-dimensional modelling of non-linear problems in structural mechanics. To achieve the objectives a software application, that implements the model FyDiK along with the 3D graphical user interface has been developed and thus is able to assemble a spring network model. Such a model was used for modelling the formation of cracks and fracture in the concrete specimens and also to model a plastic behaviour of steel I-beam. The calculations were performed by a massive parallelization on CUDA platform. In the first part the basic principles on which the work is based are introduced. Subsequently, a detailed description of individual parts of the model and the issue of parallelization by graphics cards are presented. In the next part the creation of the required software and improving of the model properties of mentioned materials are described. That is followed by evaluation of the achieved results with the comparison of other modelling software. The conclusion summarizes the achievements and suggestions for the further development possibilities of the presented method of modelling.

This diploma thesis deals with design of embedded computer vision system and transfer of existing computer vision application for 3D object detection from Windows OS to designed embedded system with Linux OS. Thesis focuses on design of communication interface for system control and camera video transfer through local network with video compression. Then, detection algorithm is enhanced by transferring computationally expensive functions to GPU using CUDA technology. Finally, a user application with graphical interface is designed for system control on Windows platform.

This thesis deals with Burrows-Wheeler transform (BWT) and possibilities of acceleration of this transform on graphics processing unit (GPU). Methods of compression based on BWT are introduced, as well as software libraries CUDA and OpenCL for writing programs for GPU. Parallel variants of BWT are implemented, as well as following steps necessary for compression, using CUDA library. Amount of compression of used approaches are tested and parallel versions are compared to their sequential counterparts.

This thesis deals with parallelization of complex tasks in reconstruction of dynamic magnetic resonance. It describes the basic principle of magnetic resonance and its relation to Fourier transform. It deals with the difference between static and dynamic magnetic resonance image reconstruction. It analyzes SVD algorithm and its use in magnetic resonance image reconstruction. It presents the principles and the importance of parallel computing in magnetic resonance imaging and describes CUDA technology. The thesis also contains a description and execution of the implementation of the reconstruction model in MATLAB and Java programming language which were optimized by JCuda library for Java implementation and gpuArray function in case of MATLAB implementation.

This thesis deals with parallelization of complex tasks in reconstruction of dynamic magnetic resonance. It describes the basic principle of magnetic resonance and its relation to Fourier transform. It deals with the difference between static and dynamic magnetic resonance image reconstruction. It analyzes SVD algorithm and its use in magnetic resonance image reconstruction. It presents the principles and the importance of parallel computing in magnetic resonance imaging and describes CUDA technology. The thesis also contains a description and execution of the implementation of the reconstruction model in MATLAB and Java programming language which were optimized by JCuda library for Java implementation and gpuArray function in case of MATLAB implementation.

Hlavním cílem této práce je navrhnout novu metodu rekonstrukce obrazu z dat fotoakustického snímkování. Fotoakustické snímkování je velmi populární neinvazivní metoda snímkování založená na detekování ultrazvukových vln vyvolaných laserovým paprskem. Proces snímkování generuje velké množství dat, a kvůli tomu je proces rekonstrukce obrazu velmi časově náročný. Táto práce demonstruje proces rekonstrukce obrazu pomocí zpětné projekce, algoritmu který je dostatečně jednoduchý na přizpůsobení moderním architekturám procesorů umožňující různé způsoby optimalizovaného výpočtu. Dvě různé variantu algoritmu byly navrženy: z pohledu pixelu a z pohledu senzoru, který detekuje ultrazvukové vlny. Obě varianty byly implementovány třemi různými způsoby: pomocí vektorového paralelismu, vláknového paralelismu a paralelismu na grafické karetě (GPU). Všechny 3 implementace obou variant algoritmu byly testovány a výsledky byly srovnány s výsledkem rekonstrukce algoritmu reverzního času, přesnějšího ale mnohokrát pomalejšího algoritmu. Výsledky ukázaly, že GPU paralelismus nabízí nejrychlejší výpočet, cca. 200 krát rychlejší než u algoritmu reverzního času, a proto se dá použit i v aplikacích pracující v reálném čase.