GPGPU Research Articles

GPU-Accelerated Real-Time Stereo Estimation With Binary Neural Network

Depth estimation from stereo images is essential to many applications such as robotics and autonomous vehicles, most of which ask for the real-time response, high energy and storage efficiency. Recent work has shown deep neural networks (DNN) perform extremely well for stereo estimation. However, these state-of-the-art DNN based algorithms are challenging to be deployed into real-world applications due to the high computational complexities of DNNs. Most of them are too slow for real-time inference and require several seconds of GPU computation to process image frames. In this article, we address the problem of fast stereo estimation and propose an efficient and light-weighted stereo matching system, called StereoBit, to produce a disparity map in a real-time manner while achieving close to state-of-the-art accuracy. To achieve this goal, we propose a binary neural network to generate weighted Hamming distance for an efficient similarity join in stereo estimation. In addition, we propose a novel approximation approach to derive StereoBit network directly from the well-trained network with the cosine similarity. Our approximation strategies enable a significant speedup while maintaining almost the same accuracy compared to the network with the cosine similarity. Furthermore, we present an optimization framework for fully exploiting the computing power of StereoBit. The framework provides a significant speedup of stereo estimation routines, and at the same time, reduces the memory usage for storing parameters. The effectiveness of StereoBit is evaluated by comprehensive experiments. StereoBit can achieve 60 frames per second on an NVIDIA TITAN Xp GPU on KITTI 2012 benchmark while achieving 3-pixel non-occluded stereo error 3.56 percent.

Part-scale thermal process modeling for laser powder bed fusion with matrix-free method and GPU computing

This paper presents an efficient GPU-based part-scale thermal process simulator for laser powder bed fusion (L-PBF) additive manufacturing (AM). To take full advantage of modern GPU computing, a matrix-free preconditioned conjugate gradient (PCG) finite element algorithm with voxel mesh is proposed to solve the transient heat transfer problem involved in the L-PBF process. The accuracy of the developed simulator is demonstrated by comparing with a commercial software (ANSYS) using representative L-PBF process parameters and temperature-dependent thermal properties for Ti6Al4V. For efficiency, it is found that the process simulation has a significant speedup going from a single CPU to a single GPU implementation. A speedup is also observed with the matrix-free method compared to a linear solver using a sparse matrix, both on a single GPU. In addition, several schemes devised to gain higher efficiency are discussed in details, which include exclusion of inactive elements from the memory, adaptive meshing in the build direction, preconditioner, and layer lumping. Using these schemes, the adaptability and scalability of the developed simulator are demonstrated on a complex geometry. A calibration of the model is also performed in realistic conditions with a thermocouple measurement coming from experimental data.

GEVO

GPUs are a key enabler of the revolution in machine learning and high-performance computing, functioning as de facto co-processors to accelerate large-scale computation. As the programming stack and tool support have matured, GPUs have also become accessible to programmers, who may lack detailed knowledge of the underlying architecture and fail to fully leverage the GPU’s computation power. GEVO (Gpu optimization using EVOlutionary computation) is a tool for automatically discovering optimization opportunities and tuning the performance of GPU kernels in the LLVM representation. GEVO uses population-based search to find edits to GPU code compiled to LLVM-IR and improves performance on desired criteria while retaining required functionality. We demonstrate that GEVO improves the execution time of general-purpose GPU programs and machine learning (ML) models on NVIDIA Tesla P100. For the Rodinia benchmarks, GEVO improves GPU kernel runtime performance by an average of 49.48% and by as much as 412% over the fully compiler-optimized baseline. If kernel output accuracy is relaxed to tolerate up to 1% error, GEVO can find kernel variants that outperform the baseline by an average of 51.08%. For the ML workloads, GEVO achieves kernel performance improvement for SVM on the MNIST handwriting recognition (3.24×) and the a9a income prediction (2.93×) datasets with no loss of model accuracy. GEVO achieves 1.79× kernel performance improvement on image classification using ResNet18/CIFAR-10, with less than 1% model accuracy reduction.

COMPARISON OF GPU AND CPU EFFICIENCY WHILE SOLVING HEAT CONDUCTION PROBLEMS

Overview of GPU usage while solving different engineering problems, comparison between CPU and GPU computations and overview of the heat conduction problem are provided in this paper. The Jacobi iterative algorithm was implemented by using Python, TensorFlow GPU library and NVIDIA CUDA technology. Numerical experiments were conducted with 6 CPUs and 4 GPUs. The fastest used GPU completed the calculations 19 times faster than the slowest CPU. On average, GPU was from 9 to 11 times faster than CPU. Significant relative speed-up in GPU calculations starts when the matrix contains at least 4002 floating-point numbers.

Three-Dimensional Combined Finite-Discrete Element Modeling of Shear Fracture Process in Direct Shearing of Rough Concrete–Rock Joints

A three-dimensional combined finite-discrete element element method (FDEM), parallelized by a general-purpose graphic-processing-unit (GPGPU), was applied to identify the fracture process of rough concrete–rock joints under direct shearing. The development process of shear resistance under the complex interaction between the rough concrete–rock joint surfaces, i.e., asperity dilatation, sliding, and degradation, was numerically simulated in terms of various asperity roughness under constant normal confinement. It was found that joint roughness significantly affects the development of overall joint shear resistance. The main mechanism for the joint shear resistance was identified as asperity sliding in the case of smoother joint roughness and asperity degradation in the case of rougher joint asperity. Moreover, it was established that the bulk internal friction angle increased with asperity angle increments in the Mohr–Coulomb criterion, and these results follow Patton’s theoretical model. Finally, the friction coefficient in FDEM appears to be an important parameter for simulating the direct shear test because the friction coefficient affects the bulk shear strength as well as the bulk internal friction angle. In addition, the friction coefficient of the rock–concrete joints contributes to the variation of the internal friction angle at the smooth joint than the rough joint.

Development of regional simulation of seismic ground‐motion and induced liquefaction enhanced by GPU computing

AbstractMuch research has been conducted for physics‐based ground‐motion simulation to reproduce seismic response of soil and structures precisely and to mitigate damages caused by earthquakes. We aimed at enabling physics‐based ground‐motion simulations of complex three‐dimensional (3D) models with multiple materials, such as a digital twin (high‐fidelity 3D model of the physical world that is constructed in cyberspace). To perform one case of such simulation requires high computational cost and it is necessary to perform a number of simulations for the estimation of parameters or consideration of the uncertainty of underground soil structure data. To overcome this problem, we proposed a fast simulation method using graphics processing unit computing that enables a simulation with small computational resources. We developed a finite‐element‐based method for large‐scale 3D seismic response analysis with small programming effort and high maintainability by using OpenACC, a directive‐based parallel programming model. A lower precision variable format was introduced to achieve further speeding up of the simulation. For an example usage of the developed method, we applied the developed method to soil liquefaction analysis and conducted two sets of simulations that compared the effect of countermeasures against soil liquefaction: grid‐form ground improvement to strengthen the earthquake resistance of existing houses and replacement of liquefiable backfill soil of river wharves for seismic reinforcement of the wharf structure. The developed method accelerates the simulation and enables us to quantitatively estimate the effect of countermeasures using the high‐fidelity 3D soil‐structure models on a small cluster of computers.

RETRACTED ARTICLE: An effective secret image sharing using quantum logic and GPGPU based EDNN super-resolution

This paper presented an effective secret image sharing with super-resolution utilizing quantum logic and enthalpy based adaptive deep neural network. The proposed technique is processed as; at the sender side, initially secret input image is converted into a halftone image format by utilizing Error diffusion with varying thresholds (EDVT) method. Then in share generation phase, shares are produced with the basis matrix. Here, the basis matrix is created utilizing the quantum logic methodology. Then in embedding phase, discrete wavelet transform (DWT) is utilized for encoding shares. At the receiver side, encoded image is reconstructed using XOR operation and results the low-resolution image. Finally, enthalpy based adaptive deep neural network (EDNN) is designed with the General Purpose Graphic Processing Unit (GPGPU) to enhance the resolution of the reconstructed images and to lessen the time complexity of deep learning. Here, the EDNN is adapted with the enthalpy based normalization to mitigate the over fitting in layers of deep neural network. Furthermore, the proficiency of the proposed work improved in terms of normalized cross correlation, normalized absolute error, peak signal to noise ratio, mean square error and execution time by deploying images among CPU and GPGPU in an enhanced manner.

Arioc: High-concurrency short-read alignment on multiple GPUs.

In large DNA sequence repositories, archival data storage is often coupled with computers that provide 40 or more CPU threads and multiple GPU (general-purpose graphics processing unit) devices. This presents an opportunity for DNA sequence alignment software to exploit high-concurrency hardware to generate short-read alignments at high speed. Arioc, a GPU-accelerated short-read aligner, can compute WGS (whole-genome sequencing) alignments ten times faster than comparable CPU-only alignment software. When two or more GPUs are available, Arioc's speed increases proportionately because the software executes concurrently on each available GPU device. We have adapted Arioc to recent multi-GPU hardware architectures that support high-bandwidth peer-to-peer memory accesses among multiple GPUs. By modifying Arioc's implementation to exploit this GPU memory architecture we obtained a further 1.8x-2.9x increase in overall alignment speeds. With this additional acceleration, Arioc computes two million short-read alignments per second in a four-GPU system; it can align the reads from a human WGS sequencer run-over 500 million 150nt paired-end reads-in less than 15 minutes. As WGS data accumulates exponentially and high-concurrency computational resources become widespread, Arioc addresses a growing need for timely computation in the short-read data analysis toolchain.

Supercomputing Pipelines Search for Therapeutics Against COVID-19

The urgent search for drugs to combat SARS-CoV-2 has included the use of supercomputers. The use of general-purpose graphical processing units (GPUs), massive parallelism, and new software for high-performance computing (HPC) has allowed researchers to search the vast chemical space of potential drugs faster than ever before. We developed a new drug discovery pipeline using the Summit supercomputer at Oak Ridge National Laboratory to help pioneer this effort, with new platforms that incorporate GPU-accelerated simulation and allow for the virtual screening of billions of potential drug compounds in days compared to weeks or months for their ability to inhibit SARS-COV-2 proteins. This effort will accelerate the process of developing drugs to combat the current COVID-19 pandemic and other diseases.

Prediction-Based Error Correction for GPU Reliability with Low Overhead

Scientific and simulation applications are continuously gaining importance in many fields of research and industries. These applications require massive amounts of memory and substantial arithmetic computation. Therefore, general-purpose computing on graphics processing units (GPGPU), which combines the computing power of graphics processing units (GPUs) and general CPUs, have been used for computationally intensive scientific and big data processing applications. Because current GPU architectures lack hardware support for error detection in computation logic, GPGPU has low reliability. Unlike graphics applications, errors in GPGPU can lead to serious problems in general-purpose computing applications. These applications are often intertwined with human life, meaning that errors can be life threatening. Therefore, this paper proposes a novel prediction-based error correction method called Prediction-based Error Correction (PRECOR) for GPU reliability, which detects and corrects errors in GPGPU platforms with a focus on errors in computational elements. The implementation of the proposed architecture needs a small number of checkpoint buffers in order to fix errors in computational logic. The PRECOR architecture has prediction buffers and controller units for predicting erroneous outputs before performing rollback. Following a rollback, the architecture confirms the accuracy of its predictions. The proposed method effectively reduces the hardware and time overheads required to correct errors. Experimental results confirm that PRECOR efficiently fixes errors with low hardware and time overheads.

High Performance Simulation of Spiking Neural Network on GPGPUs

Spiking neural network (SNN) is the most commonly used computational model for neuroscience and neuromorphic computing communities. It provides more biological reality and possesses the potential to achieve high computational power and energy efficiency. Because existing SNN simulation frameworks on general-purpose graphics processing units (GPGPUs) do not fully consider the biological oriented properties of SNNs, like spike-driven, activity sparsity, etc., they suffer from insufficient parallelism exploration, irregular memory access, and load imbalance. In this article, we propose specific optimization methods to speed up the SNN simulation on GPGPU. First, we propose a fine-grained network representation as a flexible and compact intermediate representation (IR) for SNNs. Second, we propose the cross-population/-projection parallelism exploration to make full use of GPGPU resources. Third, sparsity aware load balance is proposed to deal with the activity sparsity. Finally, we further provide dedicated optimization to support multiple GPGPUs. Accordingly, BSim, a code generation framework for high-performance simulation of SNN on GPGPUs is also proposed. Tests show that, compared to a state-of-the-art GPU-based SNN simulator GeNN, BSim achieves 1.41× - 9.33× speedup for SNNs with different configurations; it outperforms other simulators much more.

Exhaustive similarity search on a many-core architecture for finger-vein massive identification

In massive biometric identification systems, response times mainly depends on the database searching algorithms. Thus, in large databases, an increment in the simultaneous queries traffic becomes a critical factor. This paper proposes an algorithm based on the use of a graphic processing unit to solve the exhaustive similarity search for the mass identification of finger veins, using the binary pattern descriptor of the local vertical line and the Hamming distance. The proposed approach reduces the computation time of the searching process over high query traffic by solving each query with a different processing block. The proposed method allows the identification of individuals in a database of 1 million elements, which is the largest database used for finger-vein identification. Experimental results show that our proposed method resolves up to 28 queries simultaneously (over a database of one million individuals) within a time lower than 3 seconds and achieving a speed-up of 283x. To our knowledge, our work is the first implementation of finger-vein recognition on a general-purpose graphics processing unit, which is the main contribution of this document.

Autoferry Gemini: a real-time simulation platform for electromagnetic radiation sensors on autonomous ships

Testing that ships are compliant to specified safety requirements have traditionally relied on real world data, which is not scalable and limited to testable scenarios due to financial and ethical reasons. Low fidelity simulations have been used to counteract some of these problems, which is sufficient for emulating simpler systems such as radar detectors, but not for testing complex systems as found in computer vision. In the automotive industry the use of game engines have shown to be a great testing platform due to their customizability, and combination of real-time physics with computer graphics to create large volumes of high fidelity images. In the work presented here, development of an open-source maritime platform named Autoferry Gemini based on the Unity game engine is used to simulate sensors in real-time. Utilizing simulated optics and general purpose GPU programs, the render pipeline is capable of modeling lidar, radar, visible-light and infrared camera sensors simultaneously. Results from visible-light cameras and lidar have already proven to satisfy other research activities on sensor fusion for autonomous ship technology. Infrared cameras motivates further research in gathering empirical data, while GPU algorithms have made it possible to simulate 3D radar models and multiple lidar types in real-time.

Run‐time neuro‐fuzzy type‐2 controller for power optimisation of GP‐GPU architecture

The increasing demand for high-performance computing has emphasised the invocation of sophisticated multi/many-core computing architecture. Graphical Processing Unit (GPU) is considered to be an essential innovation in this regard as GPU offers a significant amount of parallelism in the execution of complex computing applications. The performance of GPUs in reducing the computational time of such applications is worth mentioning. Although GPUs appear to be a problem-solving solution for complex applications yet high power consumption has been a challenging problem, associated with this many-core computer architecture. Efficient resource management is a emerging and promising solution to this challenge; however, reducing the resources would degrade the system's overall performance. On the other hand, reducing the resources based on the analysis of workload can save significant power without degrading the system's overall performance. Therefore, a smart controller to optimise the resources of general purpose-GPU (GP-GPU) architecture is required. AFBRMC-2, a neuro-fuzzy type-2 based controller, is presented for GP-GPU architecture and, based on a feedback mechanism, keeps analysing the stats of processor and manages resources using dynamic voltage frequency scaling and core gating techniques. The proposed controller achieved up to 55% reduction in power consumption against various benchmarks on the NVIDIA TK1 GPU kit.

Convergence versus Divergence Behaviors of Asynchronous Iterations, and Their Applications in Concrete Situations

Asynchronous iterations have long been used in distributed computing algorithms to produce calculation methods that are potentially faster than a serial or parallel approach, but whose convergence is more difficult to demonstrate. Conversely, over the past decade, the study of the complex dynamics of asynchronous iterations has been initiated and deepened, as well as their use in computer security and bioinformatics. The first work of these studies focused on chaotic discrete dynamical systems, and links were established between these dynamics on the one hand, and between random or complex behaviours in the sense of the theory of the same name. Computer security applications have focused on pseudo-random number generation, hash functions, hidden information, and various security aspects of wireless sensor networks. At the bioinformatics level, this study of complex systems has allowed an original approach to understanding the evolution of genomes and protein folding. These various contributions are detailed in this review article, which is an extension of the paper “An update on the topological properties of asynchronous iterations” presented during the Sixth International Conference on Parallel, Distributed, GPU and Cloud Computing (Pareng 2019).

Simulation of variational Gaussian process NARX models with GPGPU

Gaussian processes (GP) regression is a powerful probabilistic tool for modeling nonlinear dynamical systems. The downside of the method is its cubic computational complexity with respect to the training data that can be partially reduced using pseudo-inputs. The dynamics can be represented with an autoregressive model, which simplifies the training to that of the static case. When simulating an autoregressive model, the uncertainty is propagated through a nonlinear function and the simulation cannot be evaluated in closed-form. This paper combines the variational methods of GP approximations with a nonlinear autoregressive model with exogenous inputs (NARX) to form variational GP (VGP-NARX) models. We show how VGP-NARX models, on average, better approximate a full GP-NARX model than more commonly used GP-NARX (FITC) model on 10 chaotic time-series. The modeling capabilities of VGP-NARX models are compared with the existing approaches on two benchmarks for modeling nonlinear dynamical systems. The advantage of general-purpose computing on graphics processing units (GPGPU) for Monte Carlo simulation on large validation data sets is addressed.

Lattice Boltzmann method for viscoplastic fluid flow based on regularization of ghost moments

• A viscoplastic fluid flow solver is developed with lattice Boltzmann method (LBM). • “Infinite viscosity” is computed by setting the relaxation frequency to zero. • Regularization of ghost moments keeps the simulations numerically stable. • The numerical scheme is second-order accurate and good yield surfaces are obtained. • With GPU computation typical execution time for 3D simulations is 90 min. In the Lattice Boltzmann Method (LBM) the viscosity is inversely proportional to the relaxation frequency. Hence, it should be possible to represent the singularity of some viscoplastic models by setting the relaxation frequency to zero. In the present paper we take full advantage of the LBM capabilities to propose an efficient and stable numerical scheme for viscoplastic fluid flow simulations invoking the exact Bingham constitutive equation. This scheme is expected to suit the need for more accurate viscoplastic simulations because it does compute the “infinite viscosity”, therefore dismissing the need for any viscosity regularization. Although allowing for a wide range of relaxation frequencies varying in space and time, we demonstrate that the present implementation does not degrade the standard LBM's error order. Numerical stability was promoted by of regularization of ghost moments for the lattice Boltzmann equation with force term. Since the locality of the LBM is preserved, so is the scheme's ability to be highly scalable in large computer clusters. We outline the method and the theory behind it through a detailed Chapman–Enskog expansion. Three laminar test cases are analyzed: parallel channel Poiseuille flow, square duct Poiseuille flow and lid-driven cavity flow. We show conformity with the standard LBM's error order for different wall boundary conditions and yield stress levels. Comparisons are made with other numerical studies employing augmented Lagrangian and viscosity regularization methods.

GPGPU-based randomized visual secret sharing (GRVSS) for grayscale and colour images

Visual Secret Sharing (VSS) is a technique used for sharing secret images between users. The existing VSS schemes reconstruct the original secret image as a halftone image with only a 50% contrast. The Randomized Visual Secret Sharing (RVSS) scheme overcomes the disadvantages of existing VSS schemes. Although RVSS extracts the secret image with better contrast, it is computationally expensive. This paper proposes a General Purpose Graphics Processing Unit (GPGPU)-based Randomized Visual Secret Sharing (GRVSS) technique that leverages data parallelism in the RVSS pipeline. The performance of the GRVSS is compared with the RVSS in a generic and PARAM Shavak supercomputer architecture. The GRVSS outperforms the RVSS in both architectures.

An Accurate Detection System Based on the Convolutional Neural Network

The purpose of this project is trying to detect tumors by computer based on deep learning techniques when a picture of a tumor is shown. In this research, a fast and accurate colon cancer detection is proposed, which means this research can dramatically increase the speed of diagnosis and can also improve the accuracy of confirming a diagnosis. During the experiment, a Convolutional Neural Network (CNN) structure akin to that of VGG Net and ResNet was built. A GPU computer with two 2080 Ti GPUs was used for training. The result of training produced 94% accuracy with a loss lower than 10%. Respectively, this result improved over 10% of accuracy compared to the detection by human eyes. Lastly, this program can be used by any computer to predict the tumor, which allows it transits to a practical tool in the future.

Implicit discrete ordinates discontinuous Galerkin method for radiation problems on shared-memory multicore CPU/many-core GPU computation architecture

The implicit discrete ordinates (SN ) discontinuous Galerkin (DG)/generalized minimal residual (GMRES) method has been implemented on shared-memory, multicore CPU/many-core GPU heterogeneous computation architecture via standard application programming interfaces (APIs) of open multiprocessing (OpenMP) and computer unified device architecture (CUDA). The compiler derivative based OpenMP parallelization has been performed to port the sequential SN DG/GMRES implementation to run in parallel on multiple CPU cores that share global memory. Benchmarked on a workstation comprising 12 Intel(R) Xeon(R) CPU cores, a parallel efficiency of ∼90% has been achieved for all reference problems. The ILUT preconditioning has been performed to improve algebraic conditions of associated linear systems, and it has been demonstrated that preconditioning has accelerated simulations of all reference problems by a factor of 10. Finally, it has been showcased that a K20m GPU has boosted the computation capacity of an Intel(R) Xeon(R) CPU core by a modest factor of 2–4 for all the reference problems.