GPGPU Research Articles

A comparison of numerical schemes for the GPU-accelerated simulation of variably-saturated groundwater flow

The optimal strategy to numerically solve the Richards equation is problem-dependent, especially on GPUs because an efficient numerical scheme on CPU might not scale well on GPU. In this work, four numerical schemes are coded to investigate their parallel performance on CPU and GPU. The results indicate that the scaling of Richards solvers on GPU is affected by the numerical scheme, the linear system solver, the soil constitutive model, the code structure, the problem size and the adaptive time stepping strategies. Compared with CPU, parallel simulations on GPU exhibit stronger variance in the scaling of different code sections. The poorly-scaled components could deteriorate the overall scaling. Under all circumstances, using a GPU significantly enhances computational speed, especially for large problems. Clearly, GPU computing have significant potential in accelerating large-scale hydrological simulations, but care must be taken on the design and implementation of the model structure.

Novel GPU-accelerated high-resolution solar potential estimation in urban areas by using a modified diffuse irradiance model

In the past years various methods have been developed to estimate high-resolution solar potential in urban areas, by simulating solar irradiance over surface models that originate from remote sensing data. In general, this requires discretisation of solar irradiance models that estimate direct, reflective, and diffuse irradiances. The latter is most accurately estimated by an anisotropic model, where the hemispherical sky dome from arbitrary surface’s viewpoint consists of the horizon, the circumsolar and sky regions. Such model can be modified to incorporate the effects of shadowing from obstruction with a view factor for each sky region. However, state-of-the-art using such models for estimating solar potential in urban areas, only considers the sky view factor, and not circumsolar view factor, due to high computational load. In this paper, a novel parallelisation of solar potential estimation is proposed by using General Purpose computing on Graphics Processing Units (GPGPU). Modified anisotropic Perez model is used by considering diffuse shadowing with all three sky view factors. Moreover, we provide validation based on sensitivity analysis of the method’s accuracy with independent meteorological measurements, by changing circumsolar sky region’s half-angle and resolution of the hemispherical sky dome. Finally, the presented method using GPPGU was compared to multithreaded Central Processing Unit (CPU) approach, where on average a 70x computational speedup was achieved. Finally, the proposed method was applied over a ∼21km2 urban area, obtained from Light Detection And Ranging (LiDAR) data, where the computation of solar potential was performed in a reasonable time.

Solution of nonlinear fractional-order models of nuclear reactor with parallel computing: Implementation on GPU platform

This paper present novel parallel computing algorithms for numerical solution of nonlinear fractional-order (FO) models of nuclear reactor. These FO models arise by virtue of considering neutron transport in nuclear reactor as subdiffusion. The numerical computation of these FO models is carried out by parallelizing the sequential steps in Adams–Bashforth–Moulton method. A comparison of parallel execution on MATLAB and CUDA platforms is also given. The parallel computation is executed by harnessing multicore architecture of the general purpose graphics processing unit (GPGPU). A detailed analysis of the proposed procedure is presented considering different values of fractional derivative order in the fractional differential equations (FDEs). It is shown with extensive computational exercise that the proposed methodology achieves substantial speed up for all the three FDEs models used in nuclear reactor.

Waterwave: A GPU Memory Flow Engine for Concurrent DNN Training

Training Deep Neural Networks (DNN) concurrently is becoming increasingly important for deep learning practitioners, e.g., <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">hyperparameter optimization (HPO)</i> and <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">neural architecture search (NAS)</i> . The GPU memory capacity is the impediment that prohibits multiple DNNs from being trained on the same GPU due to the large memory usage during training. In this paper, we propose Waterwave a GPU memory flow engine for concurrent deep learning training. Firstly, to address the memory explosion brought by the long time lag between memory allocation and deallocation time, we develop an allocator tailored for multi-streams. By making the allocator aware of the stream information, a <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">prioritized allocation</i> is conducted based on the chunk's <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">synchronization</i> attributes, allowing us to provide useable memory after scheduling rather than waiting it to be really released after GPU computation. Secondly, Waterwave partitions the compute graph to a set of continuous <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">node groups</i> and then performs finer-grained scheduling: <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">NodeGroup pipeline execution</i> , to guarantee a proper memory requests order. Waterwave can accomplish up to 96.8% of the maximum batch size of solo training. Additionally, in scenarios with high memory demand, Waterwave can outperform existing spatial sharing and temporal sharing by up to 12x and 1.49x, respectively.

GPU computing of yield stress fluid flows in narrow gaps

GPU computing of yield stress fluid flows in narrow gaps

A GPU-accelerated viewer for HEALPix maps

HEALPix by Górskiet al. (2005) is a de-facto standard for Cosmic Microwave Background (CMB) data storage and analysis, and is widely used in current and upcoming CMB experiments. Almost all the datasets in Legacy Archive for Microwave Background Data Analysis (LAMBDA) use HEALPix as a format of choice. Visualizing the data plays important role in research, and several toolsets were developed to do that for HEALPix maps, most notably original Fortran facilities and Python integration with healpy. With the current state of GPU performance, it is now possible to visualize extremely large maps in real time on a laptop or a tablet. HEALPix Viewer described here is developed for macOS, and takes full advantage of GPU acceleration to handle extremely large datasets in real time. It compiles natively on Intel and Arm64 architectures, and uses Metal framework for high-performance GPU computations. The aim of this project is to reduce the effort required for interactive data exploration, as well as time overhead for producing publication-quality maps. Drag and drop integration with Keynote and Powerpoint makes creating presentations easy. The main codebase is written in Swift, a modern and efficient compiled language, with high-performance computing parts delegated entirely to GPU, and a few inserts in C interfacing to cfitsio library for I/O. Graphical user interface is written in SwiftUI, a new declarative UI framework based on Swift. Most common spherical projections and colormaps are supported out of the box, and the available source code makes it easy to customize the application and to add new features if desired. On a M1 Max laptop, an nside=8192 maps are processed in real time, with geometry effects rendered at 60fps in full resolution with no appreciable load to the machine. Main user-facing delays are limited to CPU-bound cfitsio load times, and sorting needed to construct Cumulative Distribution Function (CDF) estimators for statistical analysis (hidden in background queue). Overall performance improves on the current Python software stack by a factor of 3–180x depending on the task at hand.

Investigation of the dynamic tensile fracture process of rocks associated with spalling using 3-D FDEM

Understanding the tensile fracturing of rocks under high strain-rate conditions (>20 s−1) is challenging due to experimental constraints. The spalling test has been known as a potential method to measure the tensile dynamic strength in the high strain-rate regime. However to date, the different evaluation methods proposed to interpret the spalling test results yield inconsistent dynamic tensile strength and strain-rate. To understand this inconsistency, a 3-dimensional finite-discrete element method (3-D FDEM), accelerated by general-purpose computing on graphics processing units (GPGPU), is applied to simulate the dynamic tensile fracturing process in the spalling test. The 3-D FDEM simulation results agree with the experimental results in terms of fracture patterns and free surface particle velocity to evaluate the dynamic tensile strength and corresponding strain-rate under various loading rates. The 3-D FDEM simulations can capture the discrepancies in dynamic tensile strength and strain-rate obtained based on the different evaluation methods used in the spalling test. When integrated with experimental analyses, the 3-D FDEM could provide insights, and has potential application in estimating local strain-rate and fracture location, which are key factors for the reliable evaluation of dynamic tensile strength and strain-rate in spalling tests, but are challenging to evaluate by experiments only.

An Efficient and Accurate GPU-based Deep Learning Model for Multimedia Recommendation

This article proposes the use of deep learning in human-computer interaction and presents a new explainable hybrid framework for recommending relevant hashtags on a set of orpheline tweets, which are tweets with hashtags. The approach starts by determining the set of batches used in the convolution neural network based on frequent pattern mining solutions. The convolutional neural network is then applied to the set of batches of tweets to learn the hashtags of the tweets. An optimization strategy has been proposed to accurately perform the learning process by reducing the number of frequent patterns. Moreover, eXplainable AI is introduced for hashtag recommendations by analyzing the user preferences and understanding the different weights of the deep learning model used in the learning process. This is performed by learning the hyper-parameters of the deep architecture using the genetic algorithm. GPU computing is also investigated to achieve high speed and enable the execution of the overall framework in real time. Extensive experimental analysis has been performed to show that our methodology is useful on different collections of tweets. The experimental results clearly show the efficiency of our proposed approach compared to baseline approaches in terms of both runtime and accuracy. Thus, the proposed solution achieves an accuracy of 90% when analyzing complex Wikipedia data while the other algorithms did not achieve 85% when processing the same amount of data.

Deep-learning accelerated super-resolution radial fluctuations (SRRF) enables real-time live cell imaging

The Super-resolution radial fluctuations (SRRF) algorithm analyzes radial and temporal fluorescence intensity fluctuations in an image sequence, which typically includes tens or hundreds of raw images to generate one super-resolution image. At present, most of the SRRF applications rely on large amounts of raw images or tedious post-processing and computation, thus are not feasible for real-time live cell super-resolution imaging. Here, we developed a novel deep learning accelerated SRRF method, which significantly reduced the requirement of raw images for super-resolution reconstruction. Our results showed that by using only 5 low signal-to-noise ratio (SNR) images, we were able to achieve super-resolution SRRF reconstruction to a comparable resolution as the traditional method. We demonstrated that by integration of GPU computing and the sliding window reconstruction method, the dynamic contraction of microtubules and the interactions between microtubules and clathrin-coated pits (CCPs) can be visualized in real-time with super-resolution. In summary, we established the deep learning accelerated SRRF method, which permits real-time, long-term and multi-color live cell super-resolution imaging, and we anticipate it will have vast biomedical applications.

Collective dynamics of a dense streamer front

We explore the dynamics of dense streamer channel fronts. We introduce a novel, fully three-dimensional, adaptive mesh refinement streamer simulation code, which leverages the power of general-purpose graphical processing units to accelerate computations. Our code enables the simulation of systems comprising several parallel-propagating streamers, using appropriate boundary conditions to emulate an infinitely extended front of positive streamers in ambient air. Our findings reveal that denser streamer packings result in slower front propagation and increased electric field screening within the streamers. To interpret these results and progress towards developing a coarse-grained corona model, we present a streamlined model that effectively approximates the behavior of the comprehensive microscopic system.

Memory-friendly fixed-point iteration method for nonlinear surface mode oscillations of acoustically driven bubbles: from the perspective of high-performance GPU programming

A fixed-point iteration technique is presented to handle the implicit nature of the governing equations of nonlinear surface mode oscillations of acoustically excited microbubbles. The model is adopted from the theoretical work of Shaw [1], where the dynamics of the mean bubble radius and the surface modes are bi-directionally coupled via nonlinear terms. The model comprises a set of second-order ordinary differential equations. It extends the classic Keller–Miksis equation and the linearized dynamical equations for each surface mode. Only the implicit parts (containing the second derivatives) are reevaluated during the iteration process. The performance of the technique is tested at various parameter combinations. The majority of the test cases needs only a single reevaluation to achieve 10-9 error. Although the arithmetic operation count is higher than the Gauss elimination, due to its memory-friendly matrix-free nature, it is a viable alternative for high-performance GPU computations of massive parameter studies.

Gaiwan: A size-polymorphic typesystem for GPU programs

General-purpose computing on graphics processing units (GPGPU) is increasingly used for number crunching tasks such as analyzing time series data. GPUs are a good fit for these tasks as they can execute many computations in parallel. To leverage this parallelism, the programmer is forced to carefully divide their input data into data blocks that are then distributed over the many GPU cores. The optimal block sizes are unrelated to the programmers goals, instead, they are based on characteristics of the used GPU and the input data. GPGPU programmers must additionally be wary of introducing race conditions in their programs. We believe that GPGPU programmers should be able to express GPU transformations without worrying about splitting data or race conditions. For this, we created Gaiwan, a GPGPU programming language with a size-polymorphic type system that only features data race free operations. Programmers can declare the effects of program steps on the sizes of buffers by using affine functions (eg. ). From a step sequence, Gaiwan derives a set of constraints on the size and shape of valid inputs. Gaiwan guarantees that the program will run for any input satisfying these constraints. This means that one program may analyze both a hundred data points and millions of data points, as long as the input satisfies the constraints. We prove that our system is sound and show it works with two usage examples. Our benchmarks show that our initial OpenCL-based implementation of Gaiwan scales to handling large programs.

A WebGPU-based acoustic wave simulator for ultrasound NDT

Ultrasound NDT is a standard industrial technique for health monitoring and flaw identification in several structures, such as subsea oil pipelines. Its primary purpose is to form a representative image of the internal structure of the inspected object, which provides, in many cases, early warning of structure discontinuities, making it possible to reduce repair costs and mitigate risks. One of the imaging methods that has been receiving increasing attention in the ultrasound and seismology communities is the Full Waveform Inversion (FWI), which estimates the velocity model for the inspection area by confronting the acquired information obtained by the NDT system with theoretical simulated data. Since FWI algorithms rely on successive ultrasonic wave simulations, it is essential to ensure that each simulation is optimized and contributes little to the total running time. Acoustic wave simulations can be optimized using a graphical processing unit (GPU), which provides hardware acceleration through parallelism, where each point of the simulated area can be computed simultaneously. Among all available GPU APIs, WebGPU is the next-generation standard graphics Web API that exposes modern computer graphics capabilities by giving the user low-level, general-purpose access to the GPUs. This API is designed to efficiently map to native GPU APIs, which enables applications to run on different GPUs, making implementations more scalable and interchangeable. Motivated by this flexibility, we explore the acceleration capabilities of the WebGPU API by implementing a WebGPU-based acoustic wave simulator for ultrasound NDT.

Program Analysis and Machine Learning–based Approach to Predict Power Consumption of CUDA Kernel

The General Purpose Graphics Processing Unit has secured a prominent position in the High-Performance Computing world due to its performance gain and programmability. Understanding the relationship between Graphics Processing Unit (GPU) power consumption and program features can aid developers in building energy-efficient sustainable applications. In this work, we propose a static analysis-based power model built using machine learning techniques. We have investigated six machine learning models across three NVIDIA GPU architectures: Kepler, Maxwell, and Volta with Random Forest, Extra Trees, Gradient Boosting, CatBoost, and XGBoost reporting favorable results. We observed that the XGBoost technique-based prediction model is the most efficient technique with an R 2 value of 0.9646 on Volta Architecture. The dataset used for these techniques includes kernels from different benchmarks suits, sizes, nature (e.g., compute-bound, memory-bound), and complexity (e.g., control divergence, memory access patterns). Experimental results suggest that the proposed solution can help developers precisely predict GPU applications power consumption using program analysis across GPU architectures. Developers can use this approach to refactor their code to build energy-efficient GPU applications.

Efficient computation of discrete games: Estimating the effect of Apple on market structure

Discrete games provide the means to analyze market dynamics with limited data. However, computing such games with many players—especially in a complete information setting—is computationally infeasible because the strategy space increases exponentially with the number of players. This study presents a novel and practical method to compute and estimate discrete games. To do so, the study introduces two methodological innovations. First, we develop an efficient simulator that requires fewer random draws to evaluate the likelihood of discrete games with multiple equilibria. The augmented simulator avoids random draws that are not compatible with the observed equilibrium outcome and, thus, efficiently uses all draws to evaluate the likelihood. Second, we utilize general‐purpose computing on graphics‐processing unit (GPGPU), using multiple processing cores in a graphics‐processing unit, to increase computational speed. The two features allow us to estimate the model significantly faster compared to traditional methods. The study's empirical application examines the effect of Apple's company‐owned stores on the retail market structure. The results show that agglomeration effects exist between Apple and upscale firms. The presence of an Apple store attracts high‐income customers, promoting the entry of upscale firms and the exit of discount firms.

Motion Vector Extrapolation for Video Object Detection.

Despite the continued successes of computationally efficient deep neural network architectures for video object detection, performance continually arrives at the great trilemma of speed versus accuracy versus computational resources (pick two). Current attempts to exploit temporal information in video data to overcome this trilemma are bottlenecked by the state of the art in object detection models. This work presents motion vector extrapolation (MOVEX), a technique which performs video object detection through the use of off-the-shelf object detectors alongside existing optical flow-based motion estimation techniques in parallel. This work demonstrates that this approach significantly reduces the baseline latency of any given object detector without sacrificing accuracy performance. Further latency reductions up to 24 times lower than the original latency can be achieved with minimal accuracy loss. MOVEX enables low-latency video object detection on common CPU-based systems, thus allowing for high-performance video object detection beyond the domain of GPU computing.

Dispersion and Radiation Modelling in ESTE System Using Urban LPM

In cases of accidental or deliberate incidents involving a harmful agent in urban areas, a detailed modelling approach is required to include the building shapes and spatial locations. Simultaneously, when applied to crisis management, a simulation tool must meet strict time constraints. This work presents a Lagrangian particle model (LPM) for computing atmospheric dispersion. The model is implemented in the nuclear decision support system ESTE CBRN, a software tool developed to calculate the atmospheric dispersion of airborne hazardous materials and radiological impacts in the built-up area. The implemented LPM is based on Thomson’s solution for the nonstationary, three-dimensional Langevin equation model for turbulent diffusion. The simulation results are successfully analyzed by testing compatibility with Briggs sigma functions in the case of continuous release. The implemented LPM is compared with the Joint Urban 2003 Street Canyon Experiment for instantaneous puff releases. We compare the maximum concentrations and peak times measured during two intensive operational periods. The modeled peak times are mostly 10–20% smaller than the measured. Except for a few detector locations, the maximum concentrations are reproduced consistently. In the end, we demonstrate via calculation on single computers utilizing general-purpose computing on graphics processing units (GPGPU) that the implementation is well suited for an actual emergency response since the computational times (including dispersion and dose calculation) for an acceptable level of result accuracy are similar to the modeled event duration itself.

A LIGHTWEIGHT TARGET DETECTION ALGORITHM BASED ON MOBILENET CONVOLUTION

Target detection algorithm based on deep learning needs high computer GPU configuration, even need to use high performance deep learning workstation, this not only makes the cost increase, also greatly limits the realizability of the ground, this paper introduces a kind of lightweight algorithm for target detection under the condition of the balance accuracy and computational efficiency, MobileNet as Backbone performs parameter The processing speed is 30fps on the RTX2060 card for images with the CNN separator layer. The processing speed is 30fps on the RTX2060 card for images with a resolution of 320×320.

Performance Investigation of the Conjunction Filter Methods and Enhancement of Computation Speed on Conjunction Assessment Analysis with CUDA Techniques

The growing number of space objects leads to increases in the potential risks of damage to satellites and generates space debris after colliding. Conjunction assessment analysis is the one of keys to evaluating the collision risk of satellites and satellite operators require the analyzed results as fast as possible to decide and execute collision maneuver planning. However, the computation time to analyze the potential risk of all satellites is proportional to the number of space objects. The conjunction filters and parallel computing techniques can shorten the computation cost of conjunction analysis to provide the analyzed results. Therefore, this paper shows the investigation of the conjunction filter performances (accuracy and computation speed): Smart Sieve, CSieve and CAOS-D (combination of both Smart Sieve and CSieve) in both a single satellite (one vs. all) and all space objects (all vs. all) cases. Then, all the screening filters are developed to implement an algorithm that executes General-purpose computing on graphics processing units (GPGPU) by using NVIDIAs Compute Unified Device Architecture (CUDA). The analyzed results show the comparison results of the accuracy of conjunction screening analysis and computation times of each filter when implemented with the parallel computation techniques.

Speed-up Measures and Associated Problems for Large-scale DEM-Models

This contribution discusses the possibilities to increase the efficiency of large Discrete Element Method (DEM) simulations. Simulations were conducted to test particle upscaling, decreasing shear modulus and using GPU instead of CPU for the computation. The conducted simulations modelled a simple ore extraction process from a defined outlet of a stope in a cave mine. The analysis is based on the influence of the approaches on the computation speed-up and the accuracy of the result. It was found that the real shear modulus could be decreased in the simulation by a factor of 103 without interfering strongly with the result, provided that the decreased shear modulus is not smaller than 108 Pa. However, a reduction of 102 was found to bring the highest speed-up to the present application.For the particle upscaling in this contribution without parameter calibration, the upscaling factor should not exceed 3. Higher upscaling factors affect the results significantly. However, the flow dynamics was already differing for an upscaling factor of 2, even though the flow zones were comparable. Using a GPU instead of a CPU is only recommended if the simulation contains a high number of particles. The speed-up for a simulation with almost 250,000 particles was over 8, but the advantage diminished for less particles. Furthermore, differences in the results between CPU and GPU computation could be found, which could be a starting point for future work. In summary, this research significantly aids in the development of more efficient DEM simulations for large-scale applications, such as cave mining.