CUDA Technology Research Articles

A NEW APPROACH FOR TEXT RECOGNITION ON A VIDEO CARD

An important task is to develop a computer system that can automatically read text content from images or videos with a complex background. Due to a large number of calculations, it is quite difficult to apply them in real-time. Therefore, the use of parallel and distributed computing in the development of real-time or near real-time systems is relevant. The latter is especially relevant in such areas as automation of video recording of traffic violations, text recognition, machine vision, fingerprint recognition, speech, and more. The paper proposes a new approach to text recognition on a video card. A parallel algorithm for processing a group of images and a video sequence has been developed and tested. Parallelization on the video-core is provided by the OpenCL framework and CUDA technology. Without reducing the generality, the problem of processing images on which there are vehicles, which allowed to obtain text from the license plate. A system was developed that was tested for the processing speed of a group of images and videos while achieving an average processing speed of 207 frames per second. As for the execution time of the parallel algorithm, for 50 images and video in 63 frames, image preprocessing took 0.4 seconds, which is sufficient for real-time or near real-time systems. The maximum acceleration of image processing is obtained up to 8 times, and the video sequence – up to 12. The general tendency to increase the acceleration with increasing dimensionality of the processed image is preserved, which indicates the relevance of parallel calculations in solving the problem.

Solving the problem of one-dimensional thermal conductivity on graphics processors using CUDA technology

A mathematical model for solving the problem of one-dimensional thermal conductivity has been developed and implemented programmatically. The purpose of the simulation is to compare the performance of algorithms on the central and graphics processors. The task of parallelization is relevant, since back in 2015 the number of stream processors in the most powerful video card was 2816, and in 2021 there were video cards with 10 496 stream processors. Applications running on NVIDIA GPUs demonstrate greater performance per dollar of invested funds and per watt of energy consumed compared to implementations built on the basis of central processors alone. This is confirmed by the high demand for video cards from miners, which has led to a 1.5-2.5 times increase in the price of video cards at the moment. The requirements for the hardware and software components necessary for the start of modeling are presented. Three methods of finite difference approximation are implemented: explicit, implicit and Crank-Nicolson on the central and graphics processors. The programming languages chosen are C (CPU) and CUDA C (GPU). For a well-parallelized task, when each thread is executed separately and it does not need data from other threads, the acceleration of calculations on the video card increased up to 60 times (an entry-level video card was used). The CUDA C language appeared relatively recently in 2006 and has a number of features when implementing a parallel algorithm. For the selected schemes: explicit, implicit, Crank-Nicolson, at each iteration, it is necessary to access neighboring threads and synchronize the threads. Synchronization of threads occurs in such a way that all threads wait for the slowest of them at each iteration, so solving problems using finite-difference approximation will be performed slower. A fragment of code on a GPU for implementing the Crank-Nicolson scheme is presented. The implementation of the Crank-Nicolson scheme requires the use of fast shared memory for data exchange between threads. The amount of shared memory is limited and affects the number of cells in the grid. The use of graphics cards gave a significant increase in execution speed even on an entry-level card with a number of 384 stream processors. The article presents a comparative analysis of the computing speed for different grid sizes from 1024 to 4000, as well as for different amounts of computing volumes in one thread.

Мультиагентні метаевристичні методи рішення задачі управління запасами

Currently, the problem of insufficient efficiency of supply chain management is relevant. One of the problems solved within the limits of the specified problem is the optimization problem of inventory management. Optimization methods that find an approximate solution using a directed search have a high probability of reaching a local extremum. Optimization methods that find an exact solution have a high computational complexity. Random search methods do not guarantee convergence. In this connection, there is a problem of insufficient efficiency of optimization methods, which needs to be solved. The article considers the task of inventory management as a component of the task of effective supply chain management. To solve this problem, the existing multi-agent metaheuristic methods were investigated. To improve the quality of solving this problem, particle swarm optimization and artificial fish swarm algorithm were chosen, which are modified by introducing dynamic parameters and Cauchy and Gaussian distributions. Parallel algorithms based on CUDA technology are proposed for these methods. This made it possible to ensure high speed and accuracy of the decision. The proposed methods are designed for software implementation in the Matlab package using the Parallel Computing Toolbox, which speeds up the process of finding a solution. The software that implements the proposed methods was developed and researched based on the data of the logistics company "Ekol Ukraine". The conducted experiments confirmed the functionality of the developed software and allow us to recommend it for practical use in solving supply chain management problems. Prospects for further research are to test the proposed methods on a wider set of test databases.

Canny Edge Detection Analysis Based on Parallel Algorithm, Constructed Complexity Scale and CUDA

Edge detection is especially important for computer vision and generally for image processing and visual recognition. On the other hand, digital image processing is widely used in multiple science fields such as medicine, X-ray analysis, magnetic resonance tomography, computed tomography, and cosmology, i.e. information collection from satellites, its transferring, and analysis. Any step of image processing, from obtaining the image to its segmentation and object recognition is followed by image noise. The processing speed is vital in popular fields that demand image analysis in real time. In this work, we have proposed an approach of parallel computing of the Canny algorithm using CUDA technology, the complexity of object recognition is analyzed according to the type of the image noise and the level of its density. The sequenced implementation on GPU and the parallel implementation on GPU was considered. The results were analyzed for efficiency and reliability. Also, parallel acceleration is calculated according to the size of the incoming image. The manipulations with the image showed the growth of processing speed of 68 times, whereas the manipulations with the size of the kernel showed the growth of processing speed of 26 times. Another contribution of this work is the analysis of the complexity of object recognition depending on the type of image noise and the level of its density. Furthermore, the increase of Gaussian noise density linearly increases the complexity of object recognition.

Influence of gravity effect to the recovery rate at uranium in-situ leaching

In-Situ Leaching is a method of extracting minerals by selectively dissolving it with a leaching solution directly in the place of occurrence of the mineral. In practice, during the development of deposits with the In-Situ Leaching method, situations arise when the solution tends to go down below the active thickness of the stratum. This may be due to geological heterogeneity of the rock or gravitational sedimentation of the solution in the rock due to the difference in the densities of the solution and groundwater. As a result of the deposition of the solution along the height, there is a decrease in the recovery of the metal located in the upper part of the geological layers. This article examines the effect of gravity on the flow regime during the filtration of the solution in the rock. The influence of the gravitational effect on the flow of solution in the rock is studied for different ratios of the densities of the solution and groundwater without taking into account the interaction of the solution with the rock. The CUDA technology is used to improve the performance of calculations. The results show that the use of CUDA technology allows to increase the performance of calculations by 40-80 times compared to calculations on a central processing unit (CPU) for different computational grids.

GPU-Accelerated Laplace Equation Model Development Based on CUDA Fortran

In this study, a CUDA Fortran-based GPU-accelerated Laplace equation model was developed and applied to several cases. The Laplace equation is one of the equations that can physically analyze the groundwater flows, and is an equation that can provide analytical solutions. Such a numerical model requires a large amount of data to physically regenerate the flow with high accuracy, and requires computational time. These numerical models require a large amount of data to physically reproduce the flow with high accuracy and require computational time. As a way to shorten the computation time by applying CUDA technology, large-scale parallel computations were performed on the GPU, and a program was written to reduce the number of data transfers between the CPU and GPU. A GPU consists of many ALUs specialized in graphic processing, and can perform more concurrent computations than a CPU using multiple ALUs. The computation results of the GPU-accelerated model were compared with the analytical solution of the Laplace equation to verify the accuracy. The computation results of the GPU-accelerated Laplace equation model were in good agreement with the analytical solution. As the number of grids increased, the computational time of the GPU-accelerated model gradually reduced compared to the computational time of the CPU-based Laplace equation model. As a result, the computational time of the GPU-accelerated Laplace equation model was reduced by up to about 50 times.

Modeling of Flooding of Settlements During the Spring Flood

The problem of flooding of territories with flood waters for settlements of the Volgograd region is considered. A numerical model of flood water dynamics is constructed, taking into account the topography of the terrain. The simulation is based on two-dimensional shallow water equations. For computational experiments, a parallel implementation of the numerical scheme CSPH-TVD for NVIDIA graphics accelerators with CUDA technology is used. The digital model of the river-bed and floodplain relief is based on spatial data SRTM3 and SRTMGL, topographic maps of the area, longitudinal profiles of rivers. The water flow rates for the Buzuluk and Perevozinka rivers are determined and a hydrograph of the Volga hydroelectric power station is constructed for the given probabilities of exceeding the water level. Flood maps were obtained for the following localities: Novoannisky, Berezovka 1st, kh. Vyazovka. The maximum values of the depths for 1 %, 3 %, 5 %, 10 %, 25 % and 50 % of water security are presented. An analysis of the flood situation was carried out and appropriate engineering and protective measures for settlements were proposed. At maximum flood water levels, residential buildings of kh. Berezovka 1-I are not subject to flooding. In order to minimize the negative impact for the city of Novoanninsky, it is proposed to carry out timely clearing of the river-bed of the Perevozinka river from congestion. To protect from flooding ch. Vyazovka it is recommended to use diversion dyke.

METHOD OF NEURAL NETWORK DETECTION OF ANOMALIES IN DATA OF WASTE-FREE PRODUCTION AUDIT

The paper presents a method for the detection of anomalies in waste-free production audit data based on the neural network model of Gauss-Bernoulli of the forward only restricted Cauchy machine (FORCM). The purpose of the work is to increase the efficiency of audit data analysis of waste-free production on the basis of the neural network model of anomalies detection without the use of the marked data that simplifies audit.
 To achieve this goal, the following tasks have been set and solved: offered model of generalized multiple transformations of audit data in the form of a two-layer neural network. Cauchy offered neural network model of Gauss-Bernoulli of the forward only restricted Cauchy machine possesses a heteroassociative memory; works real data; has no restrictions for storage capacity; provide high accuracy of detection of anomalies; uses Cauchy's distribution that increases the speed of convergence of a method of parametrical identification. To increase the speed of Gauss-Bernoulli parametric identification of a forward only restricted Cauchy machine, a parametric identification algorithm was developed to be implemented on a GPU using CUDA technology. The offered algorithm allows increasing training speed by approximately proportional to the product of numbers of neurons in the hidden layer and power of a training set.
 The experiments confirmed the operability of the developed software and allow to recommend it for use in practice in a subsystem of the automated analysis of DSS of audit for detection of anomalies.

Evaluation of the performance of algorithms for synthesizing radar images using Cuda technology

The features of the implementation of the algorithm for the synthesis of detail radar images for an aperture synthesis radar using the built-in functions of the Cuda library are presented. The estimation of computational complexity from the standpoint of the organization of parallel computing on Nvidia GPUs is given. The estimation of the real performance of radar synthesis is given, taking into account the volume and features of the placement of primary radar information.

A multi-level parallel algorithm for seismic imaging based on one-way wave equation migration

This article presents a parallel algorithm for seismic imaging based on depth wavefield extrapolation (the solution of a one-way wave equation [OWE]) employing the pseudospectral method. The algorithm is essentially oriented on common-offset vector (COV) gathers seismic migration based on an OWE, includes several parallelism levels, and utilizes message passing interface (MPI), Nested OpenMP, and CUDA technologies. The uppermost level of parallelism involves data splitting to ensure that each dataset can be processed independently to construct parts of the COV images. The algorithm is embarrassingly parallel at this level. Next, each independent run processes several datasets to compute COV images. Each MPI process computes all COV images for a single dataset, then MPI processes exchange data to build up a single COV image for all datasets at a single node. Computation of the COV image at a node requires OpenMP parallelization so that one thread governs the GPU-based calculations and facilitates the construction of images within a thick slab. All other threads perform image interpolation within the slab. Finally, CUDA technology is used for the most computationally intense part of the algorithm—wavefield extrapolation. Such a complex algorithm structure makes it possible to process all COV images for full-azimuth seismic data employing OWE-based migration.

Development of vector algorithm using CUDA technology for three-dimensional retinal laser coagulation process modeling

For diabetic retinopathy treatment, laser coagulation is used in modern practice. During the laser surgery process, the parameters of laser exposure are selected manually by a doctor, which requires the doctor to have sufficient experience and knowledge to achieve a therapeutic effect. On the basis of mathematical modeling of the laser coagulation process, it is possible to estimate the crucial parameters without performing an operation. However, the retina has a rather complex structure, and when even low-cost numerical methods are used for modeling, it takes a long time to obtain a result. In this regard, the development of time-efficient algorithms for three-dimensional modeling is an urgent task, since the use of such algorithms will provide a compre-hensive study within a limited time. In this paper, we study the execution time of algorithms that implement various variations in the application of the splitting method and the finite difference method, adapted to the set problem of heat conduction. The study reveals the most efficient algorithm, which is then vectorized and implemented using the CUDA technology. The study was carried out using Intel Core i7-10875H and Nvidia RTX 2080 MAX Q and showed that an analog of the vector algorithm, focused on solving a multidimensional heat conduction problem, provides an acceleration of no more than 1.5 times compared to the sequential version. The developed vector-based algorithm, focused on the application of the sweep method in all directions of the three-dimensional problem, significantly reduces the time spent on copying into the memory of the video card and provides a 40-fold acceleration in comparison with the sequential three-dimensional modeling algorithm. On the basis of the same approach, a parallel algorithm of mathematical modeling was developed, which provided a 20-fold acceleration at full processor load.

REAL-TIME DEEP NEURAL NETWORKS FOR MULTIPLE OBJECT TRACKING AND SEGMENTATION ON MONOCULAR VIDEO

Abstract. The paper is devoted to the task of multiple objects tracking and segmentation on monocular video, which was obtained by the camera of unmanned ground vehicle. The authors investigate various architectures of deep neural networks for this task solution. Special attention is paid to deep models providing inference in real time. The authors proposed an approach based on combining the modern SOLOv2 instance segmentation model, a neural network model for embedding generation for each found object, and a modified Hungarian tracking algorithm. The Hungarian algorithm was modified taking into account the geometric constraints on the positions of the found objects on the sequence of images. The investigated solution is a development and improvement of the state-of-the-art PointTrack method. The effectiveness of the proposed approach is demonstrated quantitatively and qualitatively on the popular KITTI MOTS dataset collected using the cameras of a driverless car. The software implementation of the approach was carried out. The acceleration of the procedure for the formation of a two-dimensional point cloud in the found image segment was done using the NVidia CUDA technology. At the same time, the proposed instance segmentation module provides a mean processing time of one image of 68 ms, the embedding and tracking module of 24 ms using the NVidia Tesla V100 GPU. This indicates that the proposed solution is promising for on-board computer vision systems for both unmanned vehicles and various robotic platforms.

Parallel, multi-purpose Monte Carlo code for simulation of light propagation in segmented tissues

We have developed a massive-parallel, multi-purpose Monte Carlo (MC) code for simulation of light propagation in complex structures modelled with e.g. magnetic resonance images. The code is designed to execute parallel threads on a Linux-based cluster of computers equipped with multiple graphical processing units (GPU) utilizing NVidia CUDA technology. We show steps one can take to implements such code itself. Furthermore, we provide methodology of building a MRI-based head model and populating it with realistic optical properties at excitation and fluorescence emission wavelengths during inflow of fluorescence agent (indocyanine green – ICG). The proposed code provides following original features: (i) Simulation of fluorescence light propagation in media with spatial distribution of multiple different fluorophores characterized by concentration, quantum yield and fluorescence lifetime. (ii) The fluorescence light tracking does not need extra photon tracking and works in parallel with tracking photons at the excitation wavelength, introducing execution time overhead of 0.02% only. (iii) Calculation of high-resolution spatial distributions of sensitivity factors mapping voxels absorption change to parameters measured on a model surface: statistical parameters (moments) of the distribution of time of flight of photons. (iv) Random number generators states are preserved between runs to greatly improve calculation time of the sensitivity factors. (v) Simulation of fluorescence lifetime wide field imaging in diffusively scattering media. The code is designed to handle big models, divided into e.g. 200 million voxels or more, which other methods struggle to handle. Simulations on a human head model with 0.3 mm voxel size require 2 GB of a GPU memory only. This is supported by developed non-standard floating-point data storage format. The proposed code is cross-validated with field-leading MC and finite element methods on the same hardware environment for varying model sizes and temporal resolutions. Tests revealed competitive execution time and high temporal resolution of boundary data for high spatial resolution of MRI-based head model.

PARALLEL CUDA IMPLEMENTATION OF THE ALGORITHM FOR SOLVING THE NAVIER-STOKES EQUATIONS USING THE FICTITIOUS DOMAIN METHOD

An important direction in the development of numerical simulation methods is the study of approximate methods for solving problems of mathematical physics in complex multidimensional fields. To solve many applied problems in irregular areas, the method of fictitious areas is widely used, which is characterized by a high degree of automation of programming. The main idea of the method of fictitious domains is that the problem is solved not in the original complex domain, but in some other, simpler domain. This allows you to create software immediately for a fairly wide class of problems with arbitrary computational domains. The possibilities of applying the method of fictitious domains to the problems of hydrodynamics in the variables “stream function, velocity vortex” have been considered in many works. In this paper, we study a numerical method for solving the Navier-Stokes equations in doubly connected domains. To solve the two-dimensional Navier-Stokes equations in irregular domains, an approximate method based on the method of fictitious domains is proposed. A computational finite difference algorithm for solving an auxiliary problem of the method of fictitious domains has been developed. The results of numerical simulation of two-dimensional Navier-Stokes equations by the method of fictitious domains with continuation by the lowest coefficient are presented. For this problem, a parallel algorithm was developed using CUDA technologies, which was tested on various mesh dimensions

Benchmark of Similar Blocks Search under Noisy Conditions

A similarity search in images has become a typical operation in many applications. A presence of noise in images greatly affects the correctness of detection of similar image blocks, resulting in a reduction of efficiency of image processing methods, e.g., non-local denoising. In this paper, we study noise immunity of various distance measures (similarity metrics). Taking into account a wide variety of information content in real life images and variations of noise type and intensity. We propose a set of test data and obtain preliminary results for several typical cases of image and noise properties. The recommendations for metrics' and threshold selection are given. Fast implementation of the proposed benchmark is realized using CUDA technology.

Using the functional programming library for solving numerical problems on graphics accelerators with CUDA technology

Modern graphics accelerators (GPUs) can significantly speed up the execution of numerical tasks. However, porting programs to graphics accelerators is not an easy task. Sometimes the transfer of programs to such accelerators is carried out by almost completely rewriting them (for example, when using the OpenCL technology). This raises the daunting task of maintaining two independent source codes. However, CUDA graphics accelerators, thanks to technology developed by NVIDIA, allow you to have a single source code for both conventional processors (CPUs) and CUDA. The machine code generated when compiling this single text depends on which compiler it is compiled with (the usual one, such as gcc, icc and msvc, or the compiler for CUDA, nvcc). However, in this single source code, you need to somehow tell the compiler which parts of this code to parallelize on shared memory. For the CPU, this is usually done using OpenMP and special pragmas to the compiler. For CUDA, parallelization is done in a completely different way. The use of the functional programming library developed by the authors allows you to hide the use of one or another parallelization mechanism on shared memory within the library and make the user source code completely independent of the computing device used (CPU or CUDA). This article shows how this can be done.

CHOICE OF CONDITIONS FOR MHD SIMULATIONS ABOVE THE ACTIVE REGION, ALLOWING THE STUDY OF THE SOLAR FLARE MECHANISM

Primordial release of solar flare energy high in corona (at altitudes 1/40 - 1/20 of the solar radius) is explained by release of the magnetic energy of the current sheet. The observed manifestations of the flare are explained by the electrodynamical model of a solar flare proposed by I. M. Podgorny. To study the flare mechanism is necessary to perform MHD simulations above a real active region (AR). MHD simulation in the solar corona in the real scale of time can only be carried out thanks to parallel calculations using CUDA technology. Methods have been developed for stabilizing numerical instabilities that arise near the boundary of the computational domain. Methods are applicable for low viscosities in the main part of the domain, for which the flare energy is effectively accumulated near the singularities of the magnetic field. Singular lines of the magnetic field, near which the field can have a rather complex configuration, coincide or are located near the observed positions of the flare.

PARALLELIZATION OF RSA CRYPTOGRAPHIC ALGORITHM BASED ON CUDA TECHNOLOGIES

The paper examines the efficiency of the application of CUDA technologies for the parallelization of the cryptographic algorithm with the public key. The speed of execution of several implementations of the algorithm is compared: sequential implementation on the CPU and two parallel implementations – on the CPU and GPU. A description of the public key algorithm is presented, as well as properties that allow it to be parallelized. The advantages and disadvantages of parallel implementations are analyzed. It is shown that each of them can be suitable for different scenarios. The software was developed and several numerical experiments were performed. The reliability of the obtained results of encryption and decryption is confirmed. To eliminate the influence of external factors at the time of execution the algorithm was tested ten times in a row and the average value was calculated. Acceleration coefficients for message encryption and decryption algorithms were estimated based on OpenMP and CUDA technology. The proposed approach focuses on the possibility of further optimization through the prospects of developing a multi-core architecture of computer systems and graphic processors.

Use of graphics processors to cardinally acceleration the calculation procedure of large subaperture active phased antenna arrays radiation pattern

For traditional APAA with a rectangular and triangular grid of nodes, in which partial emitters are installed, simple analytical formulas are known that express the dependence of the array radiation pattern (ARP) on the angular coordinates. If the grid of nodes is irregular, or in a regular grid there are failures of individual digits in the control elements (phase shifters, attenuators, delay lines) or in power amplifiers, then these formulas are not applicable. Therefore, it is necessary to address to general relations of additive summation of the specific contribution to the ARP of the field generated by each emitter separately. In large arrays, their number can range from several to many tens or even hundreds of thousands. Further, if it is necessary to synthesize a given type of array RP, it is necessary to form special types of the amplitude-phase distribution of the signal on the emitters. The procedures for synthesizing such distributions require hundreds or even thousands of repeated calculations of the array RP, which may require several million angular points to analyze in the upper hemisphere of radiation. The article shows that using parallel computing with the help of programming technology on graphics processors (GPU), produced by NVIDIA and CUDA technology in a system designed to quickly create high-performance codes in the extended C language, C++ is able to speed up the calculations of the large antenna ARP by 2…3 orders of magnitude, even on standard office computers. When using modern CPUs (AMD Ryzen Threadripper PRO 3995 WX) and GPU (NVIDIA GeForce 3090), the acceleration in calculations can additionally increase by hundreds of times. At such computing speeds, even the problems of antenna synthesis and statistical analysis of the characteristics of the array RP will be solved almost in real time. An example of a statistical estimation of the ARP parameters of a large subaperture array with 1280 emitters, used in the COSMOSkyMed system, when random failures of the inverse type occur in the phase shifter digits. The use of graphics processors allows us to solve problems such as the synthesis of an array RP having a prescribed shape, and the statistical analysis of the characteristics of the array in the event of various failure options in its elements, as well as mechanical distortions of the aperture shape, in almost real time.

Distributions of Two Atoms Collisions over the Surface of the Condensed Phase

The processes of heat and mass transfer are closely related to the evaporation of a substance from the surface of the condensed phase. The interaction of outgoing molecules from the surface of the condensed phase with condensed phase molecules plays a fundamental role. A simpler case of evaporation is the departure of atoms from the surface of the condensed phase, i.e. the atoms overcome the potential barrier on the surface of the condensed phase. Depending on the evaporation rate, a Knudsen layer appears above the surface of the condensed phase. In this paper, based on the model of rigid spheres, the density distributions of the collision distances and the average values of the collision distances of two atoms emitted simultaneously from the surface of the condensed phase above the surface are analyzed. Distributions of the collision distance depending on the surface temperature, the size of the potential barrier, and the size of the evaporation area are obtained. Computer experiments were performed using the Monte Carlo method. To obtain the results of numerical simulation, a parallel algorithm adapted to calculations on graphics processors with CUDA technology was developed.