GPUs Research Articles

Обеспечение функционирования графических GLX приложений на промышленном оборудовании с использованием API EGL

This paper discusses the operation of engineering software on systems with graphics adapter drivers that provide limited to no support for certain graphics subsystems of Linux OS. The paper demonstrates that software with GLX API-based visualization subsystem needs modification to be ported to the embedded systems using graphics accelerators with drivers that support only the EGL API. The purpose of the study is to develop approaches for enabling the operation of software on industrial equipment and utilizing the GLX API via the EGL API of the Linux-based graphics subsystem. The paper discussed the method for translating GLX API calls to the EGL API and its applicability. A new algorithm has been developed for organizing interaction between user applications and the Linux OS graphics subsystem, which allows running applications with the GLX API in the absence of DRI support by the graphics accelerator driver. The correctness of the algorithm’s operation has been tested and compared with the results of the tests of open drivers with DRI2 support. The findings will make it possible to reduce the resources needed to support individual graphics driver subsystems and make a smooth transition to using the EGL API in the embedded systems.

Novel Scheme for GPU‐Accelerated Finite‐Difference Time‐Domain Simulation of Electromagnetic Wave Interaction With Magnetic Plasma

AbstractBased on graphical processing unit acceleration, a new method of finite‐difference time‐domain scheme is proposed to simulate the interaction between electromagnetic waves and magnetized plasma in two‐dimensional conditions. In this study, transversely electric and transversely magnetic are computed in time to avoid matrix operations involving Lorentz equations of motion. Compared to Young's method, the new method reduces addition and multiplication by about 63% and 66%, respectively. The simulation results of ionospheric wave propagation show that the new method agrees well with Young's method and the calculation speed is improved significantly.

Speedy Vision-based Human Detection Using Lightweight Deep Learning Network

Person detection plays a role as the initial system of video surveillance analysis with various implementations, such as activity analysis, person re-id, behavior analysis, and tracking analysis. The demand for efficient models drives a deep learning architecture with a superficial structure that can operate in real-time. You look only once (YOLO) object detection has been presented as an accurate detector that can operate in real-time. The speed limitation, huge computation cost, and abundant parameters still leave vital issues to improve the efficiency of this architecture. Lightweight human detection is proposed by utilizing the YOLOv5n framework. Modifying layer depth promotes a detection system that can operate fast and without stuttering. As a result, the proposed detector has satisfactory performance and is competitive with existing models. It achieves a mAP of 45.2%, closely competing with other person detectors. Additionally, it can run fast without stumbling at 26 frames per second. The detector's speed offers the advantage of this work that it can be feasibly implemented on a cpu device without a graphics accelerator.

Conversational AI

Abstract: Conversational AI systems are becoming increasingly popular across many industries and are transforming the way people interact with technology. For a more authentic, human-like connection and a smooth user experience, these systems should combine text-based interactions with multimodal capabilities. The authors of this work suggest a new approach to improving conversational AI systems' usability by combining speech and visual analysis. By combining visual and auditory processing capabilities, AI systems can better understand human inquiries and instructions. Both visual data and speech can be better understood with the use of computer vision algorithms and natural language processing techniques, respectively. Conversational AI systems can provide more accurate and tailored replies by integrating many modalities to better grasp human intent and context. The development of multimodal conversational AI presents a significant difficulty in ensuring the smooth integration of voice and visual processing units. A strong architectural design and advanced algorithms are necessary for the simultaneous synchronization and comprehension of data from several modalities in real-time. The system needs to keep track of the conversation's context even when it switches between different forms of communication so it can keep providing fair and relevant responses all through the engagement. Customization is key to making multimodal conversational AI better for users. Based on user data and preferences, the system may tailor interactions to offer more relevant ideas and support. Users are more invested in the AI system over time, and they have a better experience overall because to customization. Ensuring the privacy and security of important audiovisual data is of the utmost importance while building multimodal conversational AI. Strong encryption, anonymization technologies, and compliance with data protection regulations are vital for user privacy and system confidence. Continuous improvement is key to the success of multimodal conversational AI systems. The feedback from users can help the developers improve the system and add new features. Thanks to this iterative technique, the AI system stays flexible and can adjust to changing consumer preferences. By combining voice and picture processing, conversational AI systems have a great deal of promise for improving the user experience. Through the integration of visual and auditory signals, these systems have the ability to comprehend user intent more accurately, provide customized experiences, and completely transform the way humans engage with technology.

Improved numerical solution of anisotropic poroelastic wave equation in microseismicity: Graphic process unit acceleration and moment tensor implementation

AbstractThe accuracy and computational efficiency of full waveform forward modelling in poroelastic media are crucial for microseismic monitoring. It enables intuitive, precise and efficient simulation of subsurface responses, thereby improving the reliability of moment tensor inversion and seismic source mechanism interpretation. Additionally, it reflects the role of fluid effects in waveform evolution. In this paper, based on the Biot mechanism, we derived the first‐order velocity–stress equation of poroelastic media and discretized the model using a rotated staggered grid. The rotated staggered grid can well adapt to anisotropic media with high contrast parameters. We provide the finite difference formula based on graphic process unit–acceleration and moment tensor and also provide the graphic process unit workflow for forward modelling of anisotropic poroelastic media. First, two models with different grid sizes were run based on single graphic process unit, 1‐thread Central Processing Unit (CPU) and 16‐thread CPU. The results show that the speedup factors are approximately 14.3 and 3.7, respectively, compared with the running time of 1‐thread CPU and 16‐thread CPU. Then, we compare and evaluate the response of three basic source mechanisms (isotropic expansion, double couple and compensated linear vector dipole) in the model. The comparison of analytical and numerical results verifies the effectiveness of the method. Furthermore, wavefield snapshots of two typical anisotropic media (vertical transversely isotropic and horizontal transversely isotropic) are analysed to correspond to different moment tensor sources. The results showed that the source mechanism does not change the isotropic and anisotropic plane and the wave travel time, but it does change the polarization amplitude of the velocity component. The attenuation of slow qP‐wave increases along with the increase of the value of viscosity. The effect of permeability on wavefield appears with the opposite effect of viscosity. Finally, the seismic waveform differences between multi‐layer elastic media and poroelastic media are compared and analysed. The results showed that the seismic wavefield and waveform of poroelastic media are more complex, the propagation speed of seismic waves is faster, but the attenuation of seismic waves is stronger.

РАСПАРАЛЛЕЛИВАНИЕ ВЫЧИСЛЕНИЙ НА ГРАФИЧЕСКОМ ПРОЦЕССОРЕ ДЛЯ УСКОРЕНИЯ ГРАНИЧНО-ЭЛЕМЕНТНЫХ РАСЧЕТОВ В МЕХАНИКЕ

In solving problems of computer modeling using various methods, accuracy and computational efficiency questions always arise. This study explores the application of two modifications of the boundary element method to solve the problem of potential distribution within a closed two-dimensional domain with a uniform potential distribution on its boundary. The first modification involves using three nonlinear shape functions instead of one. The second modification applies the Galerkin method to the boundary element approach with three nonlinear shape functions. The essence of this modification lies in the fact that the system of equations is formulated in integral form over the entire boundary element, rather than at collocation points. In addition to this, the paper describes and investigates the advantages and disadvantages of the smoothing modification applied to these approaches. Since the influence matrix consists of independently computable elements, parallelization of calculations using NVIDIA CUDA technology has been proposed to enhance computational efficiency, significantly accelerating the calculation of interaction matrix. The choice of this technology is advantageous due to the prevalence of NVIDIA graphics accelerators in almost every personal computer or laptop, as well as it is easy to use. The study presents an approach to the application of this technology and demonstrates the results, showing the acceleration of parallelized calculations which show the dependence on the number of boundary elements. A comparison of the efficiency of the selected technology when applied to two methods, collocation and Galerkin, is also presented, indicating a significant speedup of up to 22 times by computing the influence matrix of the boundary elements.

Analyzing GPU Performance in Virtualized Environments: A Case Study

The graphics processing unit (GPU) plays a crucial role in boosting application performance and enhancing computational tasks. Thanks to its parallel architecture and energy efficiency, the GPU has become essential in many computing scenarios. On the other hand, the advent of GPU virtualization has been a significant breakthrough, as it provides scalable and adaptable GPU resources for virtual machines. However, this technology faces challenges in debugging and analyzing the performance of GPU-accelerated applications. Most current performance tools do not support virtual GPUs (vGPUs), highlighting the need for more advanced tools. Thus, this article introduces a novel performance analysis tool that is designed for systems using vGPUs. Our tool is compatible with the Intel GVT-g virtualization solution, although its underlying principles can apply to many vGPU-based systems. Our tool uses software tracing techniques to gather detailed runtime data and generate relevant performance metrics. It also offers many synchronized graphical views, which gives practitioners deep insights into GVT-g operations and helps them identify potential performance bottlenecks in vGPU-enabled virtual machines.

The use of functional programming library to parallelize on graphics accelerators with CUDA technology

Modern graphics accelerators (GPUs) can signi cantly speed up the execution of numerical tasks. However, porting programs to graphics accelerators is not an easy task, sometimes requiring their almost complete rewriting. CUDA graphics accelerators, thanks to technology developed by NVIDIA, allow you to have a single source code for both conventional processors (CPUs) and CUDA. However, in this single source code, you need to somehow tell the compiler which parts of this code to parallelize on shared memory. 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.

Shoe for Blind People

Eyes play important role in our day to day lives and are perhaps the most valuable gift we have. This world is visible to us because we are blessed with eyesight. But there are some people who lag this ability of visualizing these things. Due to this, they will undergo a lot of troubles o move comfortably in public places. Hence, wearable device should design for such visual impaired people. A smart shoe is wearable system design to provide directional information to visually impaired people. To provide smart and sensible navigation guidance to visually impaired people, the system has great potential especially when integrated with visual processing units

Study on the flooding characteristics of a deep-water submarine based on δ plus-smoothed particle hydrodynamics method and graphic processing units acceleration

The process of a damaged ship flooding is a complicated free surface flow problem. There is a complex coupling effect between the ship cabin and the flow inside and outside of the cabin. In this paper, a GPU (graphic processing unit)-δ+-SPH (smoothed particle hydrodynamics) numerical model for the cabin flooding in deep-water environments is developed based on GPU parallel acceleration technology and Nvidia's CUDA (compute unified device architecture). First, the computational accuracy and efficiency of this numerical model are verified by experiments results on the water flooding of a simple damaged cabin model. Furthermore, the flooding characteristics of a submarine cabin are analyzed, considering different numbers of damaged cabins, depths, and opening positions. Finally, the progressive flooding and the dynamic response characteristics of a full-scale submarine model are investigated. The results show that the process of progressive flooding in a submarine cabin is characterized by its rapidity and intensity. Different factors, for example, damaged cabin numbers, cabin depths, and opening positions, have great influences on the process of flooding and the motion of the submarine cabin. This study can offer valuable technical assistance in the post-damage remediation process.

Neuromorphic Analog Machine Vision Enabled by Nanoelectronic Memristive Devices

Arrays of memristive devices coupled with photosensors can be used for capturing and processing visual information, thereby realizing the concept of “in-sensor computing”. This is a promising concept associated with the development of compact and low-power machine vision devices, which is crucial important for bionic prostheses of eyes, on-board image recognition systems for unmanned vehicles, computer vision in robotics, etc. This concept can be applied for the creation of a memristor based neuromorphic analog machine vision systems, and here, we propose a new architecture for these systems in which captured visual data are fed to a spiking artificial neural network (SNN) based on memristive devices without analog-to-digital and digital-to-analog conversions. Such an approach opens up the opportunities of creating more compact, energy-efficient visual processing units for wearable, on-board, and embedded electronics for such areas as robotics, the Internet of Things, and neuroprosthetics, as well as other practical applications in the field of artificial intelligence.

BPMF: A Backprojection and Matched-Filtering Workflow for Automated Earthquake Detection and Location

Abstract We introduce BPMF (backprojection and matched filtering)—a complete and fully automated workflow designed for earthquake detection and location, and distributed in a Python package. This workflow enables the creation of comprehensive earthquake catalogs with low magnitudes of completeness using no or little prior knowledge of the study region. BPMF uses the seismic wavefield backprojection method to construct an initial earthquake catalog that is then densified with matched filtering. BPMF integrates recent machine learning tools to complement physics-based techniques, and improve the detection and location of earthquakes. In particular, BPMF offers a flexible framework in which machine learning detectors and backprojection can be harmoniously combined, effectively transforming single-station detectors into multistation detectors. The modularity of BPMF grants users the ability to control the contribution of machine learning tools within the workflow. The computation-intensive tasks (backprojection and matched filtering) are executed with C and CUDA-C routines wrapped in Python code. This leveraging of low-level, fast programming languages and graphic processing unit acceleration enables BPMF to efficiently handle large datasets. Here, we first summarize the methodology and describe the application programming interface. We then illustrate BPMF’s capabilities to characterize microseismicity with a 10 yr long application in the Ridgecrest, California area. Finally, we discuss the workflow’s runtime scaling with numerical resources and its versatility across various tectonic environments and different problems.

Warehouse Small Cargo-carrying UAV Design and Environmental T265 Camera Placement Angle Study

The Intel RealSense Tracking Camera T265 is a tracking camera that uses its own Vi-Slam algorithm to output horizontal and vertical coordinates, which has a wide range of applications in drones, unmanned boats, and unmanned vehicles. In this paper, we utilize the T265 and the Pixhawk4 (PX4) flight controller to build a Robotics Operating System (ROS)-based, fixed-point cruise-capable, self-taking-off and landing unmanned aerial vehicle (UAV) for cargo loading. The T265 is a fisheye black and white camera with powerful visual simultaneous localization and mapping (Slam) localization and a large visual range. The T265 features the Movidius Myriad 2 visual processing unit. Slam construction based on the comparison of feature points to output coordinates. Nevertheless, when the environment has a single color and insufficient feature contrast, it is easy to cause a SLAM error, leading to drifting of the drone's coordinates, which is very dangerous. It is easy to injure drone operators and other pedestrians accidentally, and at the same time, it is often accompanied by drone crashes that cause economic losses. We experimented with placing the T265 camera at multiple angles to test the best placement angle for more accurate drone positioning. The article also describes the mechanical mechanism design, hardware and software design of the cargo UAV, the optimal placement angle was verified in the environment described in the paper.

Автоматизированное определение дисперсности воздушной фазы в мороженом с применением методов машинного обучения

Ice cream is a popular cold dessert. Its air phase consists of tiny bubbles with an average diameter of 15–60 µm. New ice cream formulations depend on the way the composition and production factors affect the air phase. As a result, ice cream producers need new time-saving and reliable methods to determine dispersion. The research objective was to create a computer program for marking the position of centers and diameter of air bubbles on microscopic images of a bounding circle type.
 The review part included 20 years of Russian and English publications on microscopic research methods in ice cream production indexed in Web of Science and Russian Research Citation Index. Microscopic images of ice cream air phase were obtained using an Olympus CX41RF microscope with a magnification of ×100. The automatic markup program employed the Python programming language, the Keras machine learning library, and the TensorFlow framework. The models were trained using the NVIDIA GTX video accelerator.
 The review showed that the dispersion of ice cream air phase depends on its composition and the freezing parameters whereas bubble formation is usually described in line with the existing foaming theories. A training data set was obtained by manual labeling of microscopic images. The optimal number channels in the convolutional layers of a neural network with LeNet-type architecture was determined, which made it possible to classify images as spheres or non-spheres with an accuracy of ≥ 0.995. The sliding window method helped to determine the limits of the neural network triggering for the sliding window method were determined, which reached 7.5% of the diameter with lateral displacement and 12.5% with scaling. The developed algorithm automatically marked bubbles on microscopic images. The error in determining the average diameter was below 1.8%.
 The new method for automated calculation of the number and diameter of air bubbles in ice cream proved to be user-friendly. It can be found in public domain, and researchers are free to adapt it to solve various computer vision issues.

Volume integral model for algebraic image reconstruction and computed tomography

Industrial computed tomography (CT) has seen widespread adoption as an inspection technique due to its ability to resolve small defects and perform high-resolution measurements on complex structures. The reconstruction of CT data is usually performed using filtered back-projection (FBP) methods, such as the Feldkamp-Davis-Kress (FDK) method, and are selected as they offer a good compromise between reconstruction time and quality. More recently, iterative reconstruction algorithms have seen a resurgence in research interest as computing power has increased. Iterative reconstruction algorithms, such as the algebraic reconstruction technique (ART), use a reconstruction approach based on linear algebra to determine voxel attenuation coefficients based on the measured attenuation of the sample at the detector and calculation of the ray paths traversing the voxel grid. This offers a more precise model for CT reconstruction but at the cost of computational complexity and reconstruction time. Existing ART implementations are based on the 2D weighting models of the binary integral method (BIM), line integral method (LIM) and area integral method (AIM). For full 3D reconstruction, BIM and LIM only offer approximations leading to numerical inaccuracies. AIM for 2D reconstruction is mathematically exact but considers the divergent nature of a fan beam for 2D only. For a full 3D volumetric reconstruction, the X-ray cone beam is divergent in all directions and therefore AIM cannot be applied in its current form. A novel voxel weighting method for 3D volumetric image reconstruction using ART and providing a mathematically exact fractional volume weighting is introduced in this paper and referred to as the volume integral method (VIM). A set of algorithms is provided based on computer graphics techniques to determine ray/voxel intersections with volume reconstruction computed based on the divergence theorem. A set of experimental configurations is developed to provide a comparison against existing methods and conclusions are provided. Optimisation is achieved through graphic acceleration.

Fast Barnes–Hut-based algorithm in 2D vortex method of computational hydrodynamics

In the present paper, the efficient modification of the Barnes–Hut algorithm is proposed based on ideas of multipole and local expansions. Such an approach allows for forming a hierarchy of fast method schemes with different accuracy and computational complexity depending on the number of expansion terms. All the necessary formulae for computational are derived. The parallel implementation is developed using OpenMP and CUDA technologies, which allows performing computations on multicore processors and graphics accelerators.

Sparse matrix‐vector and matrix‐multivector products for the truncated SVD on graphics processors

SummaryMany practical algorithms for numerical rank computations implement an iterative procedure that involves repeated multiplications of a vector, or a collection of vectors, with both a sparse matrix and its transpose. Unfortunately, the realization of these sparse products on current high performance libraries often deliver much lower arithmetic throughput when the matrix involved in the product is transposed. In this work, we propose a hybrid sparse matrix layout, named CSRC, that combines the flexibility of some well‐known sparse formats to offer a number of appealing properties: (1) CSRC can be obtained at low cost from the popular CSR (compressed sparse row) format; (2) CSRC has similar storage requirements as CSR; and especially, (3) the implementation of the sparse product kernels delivers high performance for both the direct product and its transposed variant on modern graphics accelerators thanks to a significant reduction of atomic operations compared to a conventional implementation based on CSR. This solution thus renders considerably higher performance when integrated into an iterative algorithm for the truncated singular value decomposition (SVD), such as the randomized SVD or, as demonstrated in the experimental results, the block Golub–Kahan–Lanczos algorithm.

Grid Free Euler Flow Solver With Cuda Computing

Graphics Accelerators are increasingly used for general purpose high performance computing applications as they provide a low cost solution to high performance computing requirements. Intel also came out with a performance accelerator that offers a similar solution. However, the existing application software needs to be restructured to suit to the accelerator paradigm. Master-slave software architecture has been employed to enable two-dimensional and threedimensional grid-free Euler flow solvers in GPGPU computing platforms. Results showing significant improvement in the performance are presented in this paper. Convergence histories and aerodynamic forces obtained from GPGPU computing are compared with that of sequential computing results.

MPI-CUDA Implementation of Implicit Euler Flow Solver in Grid-Free Framework

Graphics accelerators are increasingly used for general purpose high performance computing applications as they provide a low cost solution to high performance computing requirements. However, the existing application software needs to be restructured to suit to the accelerator paradigm. Explicit methods are inherently suitable for parallelization whereas implicit methods are not suitable as the nodes need to be processed in specific order. However, the nodes can be grouped into clusters such that nodes within the cluster are independent and can be processed concurrently. CUDA kernel can be launched separately for each cluster of nodes to process nodes in parallel leading to same computation results as sequential program. One such successful attempt has been made and the speed up obtained along with computation results is presented in this paper.

Using a lightweight Siamese neural network for generating a feature vector in a vascular authentication system

The article analyzes the possibility of using a Siamese convolutional neural network to solve the problem of vascular authentication on an embedded hardware platform with limited computing resources (Orange Pi One). The authors give a brief review of modern methods for calculating image feature vectors used in the tasks of classifying, comparing or searching for images by content: based on variational series (histograms), local descriptors, singular point descriptors, descriptors based on hash functions, neural network descriptors. They suggest using the architecture of a biometric authentication system (BAS) based on images of palms in the visible and near-IR spectra based on a Siamese convolutional neural network. The developed software solution allows using the Siamese neural network in the "full network" (both symmetrical channels of the neural network are used) and "half of the neural network" (only one channel is used) modes to reduce the time for comparing biometric data vectors - images of the palms of registered BAS users. The authors demonstrate advantages of the neural network features: universality, scalability and competitiveness, including on embedded hardware and software solutions with limited computing resources without graphics accelerators. The studies have shown that using the Siamese neural network, the "overall accuracy" of palm image classification can be improved from 0.929 to 0.968 when compared with the image vectorization method based on a perceptual hash, while showing a comparable authentication time for individuals registered in BAS. In the experiments, the authors use a database of 2,000 images for 400 people.