CUDA Research Articles

GPU Programming Productivity in Different Abstraction Paradigms

Coprocessor architectures in High Performance Computing are prevalent in today’s scientific computing clusters and require specialized knowledge for proper utilization. Various alternative paradigms for parallel and offload computation exist, but little is known about the human factors impacts of using the different paradigms. With computer science student participants from the University of Nevada, Las Vegas with no previous exposure to Graphics Processing Unit programming, our study compared NVIDIA CUDA C/C++ as a control group and the Thrust library. The designers of Thrust claim their higher level of abstraction enhances programmer productivity. The trial was conducted on 91 participants and was administered through our computerized testing platform. Although the study was narrowly focused on the basic steps of an offloaded computation problem and was not intended to be a comprehensive evaluation of the superiority of one approach or the other, we found evidence that although Thrust was designed for ease of use, the abstractions tended to be confusing to students and in several cases diminished productivity. Specifically, abstractions in Thrust for (i) memory allocation through a C++ Standard Template Library-style vector library call, (ii) memory transfers between the host and Graphics Processing Unit coprocessor through an overloaded assignment operator, and (iii) execution of an offloaded routine through a generic transform library call instead of a CUDA kernel routine all performed either equal to or worse than CUDA.

Fast Reconstruction of 3D Point Cloud Model Using Visual SLAM on Embedded UAV Development Platform

In recent years, the rapid development of unmanned aerial vehicle (UAV) technologies has made data acquisition increasingly convenient, and three-dimensional (3D) reconstruction has emerged as a popular subject of research in this context. These 3D models have many advantages, such as the ability to represent realistic scenes and a large amount of information. However, traditional 3D reconstruction methods are expensive, and require long and complex processing. As a result, they cannot rapidly respond when used in time-sensitive applications, e.g., those for such natural disasters as earthquakes, debris flow, etc. Computer vision-based simultaneous localization and mapping (SLAM) along with hardware development based on embedded systems, can provide a solution to this problem. Based on an analysis of the principle and implementation of the visual SLAM algorithm, this study proposes a fast method to quickly reconstruct a dense 3D point cloud model on a UAV platform combined with an embedded graphics processing unit (GPU). The main contributions are as follows: (1) to resolve the contradiction between the resource limitations and the computational complexity of visual SLAM on UAV platforms, the technologies needed to compute resource allocation, communication between nodes, and data transmission and visualization in an embedded environment were investigated to achieve real-time data acquisition and processing. Visual monitoring to this end is also designed and implemented. (2) To solve the problem of time-consuming algorithmic processing, a corresponding parallel algorithm was designed and implemented based on the parallel programming framework of the compute unified device architecture (CUDA). (3) The visual odometer and methods of 3D “map” reconstruction were designed using under a monocular vision sensor to implement the prototype of the fast 3D reconstruction system. Based on preliminary results of the 3D modeling, the following was noted: (1) the proposed method was feasible. By combining UAV, SLAM, and parallel computing, a simple and efficient 3D reconstruction model of an unknown area was obtained for specific applications. (2) The parallel SLAM algorithm used in this method improved the efficiency of the SLAM algorithm. On the one hand, the SLAM algorithm required 1/6 of the time taken by the structure-from-motion algorithm. On the other hand, the speedup obtained using the parallel SLAM algorithm based on the embedded GPU on our test platform was 7.55 × that of the serial algorithm. (3) The depth map results show that the effective pixel with an error less than 15cm is close to 60%.

A radix sorting parallel algorithm suitable for graphic processing unit computing

SummaryRadix sorting is an essential basic data processing operation in many computer fields. It has important practical significance to accelerate its performance through Graphic Processing Unit (GPU). The heterogeneous parallel computing technology attracts much attention and is widely applied for its effective computation efficiency and parallel real‐time data processing capability. Taking advantage of the parallelism of GPU in numerical computation processing, a parallelization design method of the Binary_Least Significant Digit (LSD) first Radix Sorting (B_LSD_RS) algorithm based on Open Computing Language (OpenCL) is proposed. The radix sorting algorithm is divided into multiple kernel tasks, and the kernels are sequentially controlled by the event information transfer. The parallel algorithm is implemented and verified on the GPU + CPU heterogeneous platform. The experimental results show that compared with the performance of the B_LSD_RS sequential algorithm based on AMD Ryzen5 1600X CPU, B_LSD_RS parallel algorithm based on Open Multi‐Processing (OpenMP) and B_LSD_RS parallel algorithm based on Compute Unified Device Architecture (CUDA), the B_LSD_RS parallel algorithm based on OpenCL obtained 28.86 times, 11.01 times and 2.14 times speedup in the NVIDIA GTX 1070 computing platform respectively, not only achieves high performance but also achieves performance portability among different GPU computing platforms.

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.

Gradient clustering algorithm based on deep learning aerial image detection

• The two change detection algorithms proposed in this paper use KFCM gradient clustering algorithm . • KFCM algorithm is a classic clustering algorithm in unsupervised clustering, which can effectively cluster most of the data. • We will study how to use the special features of the image to complete the clustering of the difference maps. In recent years, computer vision, especially deep learning, has been widely used in various fields. Through the deep learning aerial image detection gradient clustering algorithm automatic recognition, it can solve the limitations of manual shooting by humans, can shoot from a high altitude to a panoramic view of a specific area, and provide a more comprehensive solution. The traditional forest resource management and management work is mainly carried out by forestry personnel to carry out a large number of investigations and investigations on the forest. This method not only consumes a lot of manpower and material resources, but also does not have real-time nature. It is difficult to deal with all kinds of forest management. Problems, causing unnecessary losses. In this regard, this paper proposes an aerial image change detection algorithm based on H-KFCM, and designs related experiments to verify and demonstrate the performance of the algorithm. In this paper, we conduct a parallel study based on deep learning on the gradient clustering algorithm of deep learning in aerial image processing. By using CUDA (Compute Unified Device Architecture) to perform large-scale parallel processing of aerial data. Can greatly shorten the time to obtain results, improve the efficiency of relevant personnel. Experiment analysis. It can be seen from the results that the deep learning parallelization program implemented in this paper has a faster calculation speed and uses less time in high-resolution images, and has a good acceleration ratio compared to the CPU.

NormiRazor: tool applying GPU-accelerated computing for determination of internal references in microRNA transcription studies

BackgroundMulti-gene expression assays are an attractive tool in revealing complex regulatory mechanisms in living organisms. Normalization is an indispensable step of data analysis in all those studies, since it removes unwanted, non-biological variability from data. In targeted qPCR assays it is typically performed with respect to prespecified reference genes, but the lack of robust strategy of their selection is reported in literature, especially in studies concerning circulating microRNAs (miRNA). Unfortunately, this problem impedes translation of scientific discoveries on miRNA biomarkers into widely available laboratory assays. Previous studies concluded that averaged expressions of multi-miRNA combinations are more stable references than single genes. However, due to the number of such combinations the computational load is considerable and may be hindering for objective reference selection in large datasets. Existing implementations of normalization algorithms (geNorm, NormFinder and BestKeeper) have poor performance and may require days to compute stability values for all potential reference as the evaluation is performed sequentially.ResultsWe designed NormiRazor - an integrative tool which implements those methods in a parallel manner on a graphics processing unit (GPU) using CUDA platform. We tested our approach on publicly available miRNA expression datasets. As a result, the times of executions on 8 datasets containing from 50 to 400 miRNAs (subsets of GSE68314) decreased 18.7 ±0.6 (mean ±SD), 104.7 ±4.2 and 76.5 ±2.2 times for geNorm, BestKeeper and NormFinder with respect to previous Python implementation. To allow for easy access to normalization pipeline for biomedical researchers we implemented NormiRazor as an online platform where a user could normalize their datasets based on the automatically selected references. It is available at norm.btm.umed.pl, together with instruction manual and exemplary datasets.ConclusionsNormiRazor allows for an easy, informed choice of reference genes for qPCR transcriptomic studies. As such it can improve comparability and repeatability of experiments and in longer perspective help translate newly discovered biomarkers into readily available assays.

Hybrid MPI and CUDA Parallelization for CFD Applications on Multi-GPU HPC Clusters

Graphics processing units (GPUs) have a strong floating-point capability and a high memory bandwidth in data parallelism and have been widely used in high-performance computing (HPC). Compute unified device architecture (CUDA) is used as a parallel computing platform and programming model for the GPU to reduce the complexity of programming. The programmable GPUs are becoming popular in computational fluid dynamics (CFD) applications. In this work, we propose a hybrid parallel algorithm of the message passing interface and CUDA for CFD applications on multi-GPU HPC clusters. The AUSM + UP upwind scheme and the three-step Runge–Kutta method are used for spatial discretization and time discretization, respectively. The turbulent solution is solved by the K−ω SST two-equation model. The CPU only manages the execution of the GPU and communication, and the GPU is responsible for data processing. Parallel execution and memory access optimizations are used to optimize the GPU-based CFD codes. We propose a nonblocking communication method to fully overlap GPU computing, CPU_CPU communication, and CPU_GPU data transfer by creating two CUDA streams. Furthermore, the one-dimensional domain decomposition method is used to balance the workload among GPUs. Finally, we evaluate the hybrid parallel algorithm with the compressible turbulent flow over a flat plate. The performance of a single GPU implementation and the scalability of multi-GPU clusters are discussed. Performance measurements show that multi-GPU parallelization can achieve a speedup of more than 36 times with respect to CPU-based parallel computing, and the parallel algorithm has good scalability.

A Graphics Process Unit-Based Multiple-Relaxation-Time Lattice Boltzmann Simulation of Non-Newtonian Fluid Flows in a Backward Facing Step

A modified power-law (MPL) viscosity model of non-Newtonian fluid flow has been used for the multiple-relaxation-time (MRT) lattice Boltzmann methods (LBM) and then validated with the benchmark problems using the graphics process unit (GPU) parallel computing via Compute Unified Device Architecture (CUDA) C platform. The MPL model for characterizing the non-Newtonian behavior is an empirical correlation that considers the Newtonian behavior of a non-Newtonian fluid at a very low and high shear rate. A new time unit parameter (λ) governing the flow has been identified, and this parameter is the consequence of the induced length scale introduced by the power law. The MPL model is free from any singularities due to the very low or even zero shear-rate. The proposed MPL model was first validated for the benchmark study of the lid-driven cavity and channel flows. The model was then applied for shear-thinning and shear-thickening fluid flows through a backward-facing step with relatively low Reynolds numbers, Re = 100–400. In the case of shear-thinning fluids (n=0.5), laminar to transitional flow arises while Re≥300, and the large vortex breaks into several small vortices. The numerical results are presented regarding the velocity distribution, streamlines, and the lengths of the reattachment points.

End to end simulators: a flexible and scalable cloud-based architecture

Simulations of frames from existing and upcoming high-resolution spectrographs, targeted for high accuracy radial velocity measurements, are computationally demanding (both in time and space). We present in this paper an innovative approach based on both parallelization and distribution of the workload. By using NVIDIA CUDA custom-made kernels and state-of-the-art cloud-computing architectures in a Platform as a Service (PaaS) approach, we implemented a modular and scalable end-to-end simulator that is able to render synthetic frames with an accuracy of the order of few cm/sec, while keeping the computational time low. We applied our approach to two spectrographs. For VLT-ESPRESSO we give a sound comparison between the actual data and the simulations showing the obtained spectral formats and the recovered instrumental profile. We also simulate data for the upcoming HIRES at the ELT and investigate the overall performance in terms of computational time and scalability against the size of the problem. In addition we demonstrate the interface with data-reduction systems and we preliminary show that the data can be reduced successfully by existing methods.

Visualization and Analysis of the Volume Dataset for Time‐Varing Electromagnetic Simulation

The three Dimensional (3D) visualization of time-varying Electromagnetic (EM) data in antenna devices can be used to evaluate various parameters, such as the quality of beams and the power performance of antennas, for optimizing the antenna design. Compared with the single volume data, the time-varying data has characteristics of strong continuity, large scale and multi-variable. It is challenging. We proposed a real-time dynamic rendering method using Compute unified device architecture (CUDA)-based volume rendering algorithm. We designed the opacity transfer function as multiple trapezoid, which represents each energy band, and provided a fusion strategy for a horn-like surface and a EM volume model that considers the contribution of both intersection points of surface with viewing ray and sample points of the volume. We also proposed a framework for time-varying 3D EM datasets. Two assessment tests were conducted by both experts and students. Results showed that our 3D visualization method were effective.

Optimizing non-coalesced memory access for irregular applications with GPU computing

General purpose graphics processing units (GPGPUs) can be used to improve computing performance considerably for regular applications. However, irregular memory access exists in many applications, and the benefits of graphics processing units (GPUs) are less substantial for irregular applications. In recent years, several studies have presented some solutions to remove static irregular memory access. However, eliminating dynamic irregular memory access with software remains a serious challenge. A pure software solution without hardware extensions or offline profiling is proposed to eliminate dynamic irregular memory access, especially for indirect memory access. Data reordering and index redirection are suggested to reduce the number of memory transactions, thereby improving the performance of GPU kernels. To improve the efficiency of data reordering, an operation to reorder data is offloaded to a GPU to reduce overhead and thus transfer data. Through concurrently executing the compute unified device architecture (CUDA) streams of data reordering and the data processing kernel, the overhead of data reordering can be reduced. After these optimizations, the volume of memory transactions can be reduced by 16.7%–50% compared with CUSPARSE-based benchmarks, and the performance of irregular kernels can be improved by 9.64%–34.9% using an NVIDIA Tesla P4 GPU.

A versatile smoothed particle hydrodynamics code for graphic cards

Abstract We present the second release of the now open source smoothed particle hydrodynamics code miluphcuda . The code is designed to run on Nvidia CUDA capable devices. It handles one to three dimensional problems and includes modules to solve the equations for viscid and inviscid hydrodynamical flows, the equations of continuum mechanics using SPH, and self-gravity with a Barnes–Hut tree. The covered material models include different porosity and plasticity models. Several equations of states, especially for impact physics, are implemented. The basic ideas of the numerical scheme are presented, the usage of the code is explained and its versatility is shown by means of different applications. The code is hereby publicly available.

An Overview of Hardware Implementation of Membrane Computing Models

The model of membrane computing, also known under the name of P systems, is a bio-inspired large-scale parallel computing paradigm having a good potential for the design of massively parallel algorithms. For its implementation it is very natural to choose hardware platforms that have important inherent parallelism, such as field-programmable gate arrays (FPGAs) or compute unified device architecture (CUDA)-enabled graphic processing units (GPUs). This article performs an overview of all existing approaches of hardware implementation in the area of P systems. The quantitative and qualitative attributes of FPGA-based implementations and CUDA-enabled GPU-based simulations are compared to evaluate the two methodologies.

Parallel implementation of the non-overlapping template matching test using CUDA

NIST (National Institute of Standards and Technology) statistical test recognized as the most authoritative is widely used in verifying the randomness of binary sequences. The Non-overlapping Template Matching Test as the 7th test of the NIST Test Suit is remarkably time consuming and the slow performance is one of the major hurdles in the testing process. In this paper, we present an efficient bit-parallel matching algorithm and segmented scan-based strategy for execution on Graphics Processing Unit (GPU) using NVIDIA Compute Unified Device Architecture (CUDA). Experimental results show the significant performance improvement of the parallelized Non-overlapping Template Matching Test, the running speed is 483 times faster than the original NIST implementation without attenuating the test result accuracy.

A hybrid MPI-CUDA approach for nonequispaced discrete Fourier transformation

Nonequispaced discrete Fourier transformation (NDFT) is widely applied in all aspects of computational science and engineering. The computational efficiency and accuracy of NDFT has always been a critical issue in hindering its comprehensive applications both in intensive and in extensive aspects of scientific computing. In our previous work Yang et al. (2018), a CUNFFT method was proposed and it shown outstanding performance in handling NDFT at intermediate scale based on CUDA (Compute Unified Device Architecture) technology. In the current work, we further improved the computational efficiency of the CUNTTF method using an efficient MPI-CUDA hybrid parallelization (HP) scheme of NFFT to achieve a cutting-edge treatment of NDFT at super extended scale. Within this HP-NFFT method, the spatial domain of NDFT is decomposed into several parts according to the accumulative feature of NDFT and the detailed number of CPU and GPU nodes. These decomposed NDFT subcells are independently calculated on different CPU nodes using a MPI process-level parallelization mode, and on different GPU nodes using a CUDA thread-level parallelization mode and CUNFFT algorithm. A massive benchmarking of the HP-NFFT method indicates that this method exhibits a dramatic improvement in computational efficiency for handling NDFT at super extended scale without loss of computational precision. Furthermore, the HP-NFFT method is validated via the calculation of Madelung constant of fluorite crystal structure, and thereafter verified that this method is robust for the calculation of electrostatic interactions between charged ions in molecular dynamics simulation systems. Program summaryProgram title: HP-NFFTCPC Library link to program files:http://dx.doi.org/10.17632/ys2y92jkwy.1Licensing provisions: GNU General Public License 3Programming language: MPI, C, and CUDA CSupplementary material: The program is designed for effective computation of large-scale nonequispaced discrete Fourier transformation (NDFT), which runs on particular computers equipped with NVIDIA GPUs. It has been tested on (a) one single computer node with Intel(R) Core(TM) i7-3770 @ 3.40 GHz (CPU) and GTX 980 Ti (GPU), and (b) MPI parallel computer nodes with the same configurations.Nature of problem: For NDFT, the computation is extremely time-consuming in many domains of computational physics due to the failure to utilize FFT directly, which often affects the computational efficiency of whole system seriously. Although the parallel method CUNFFT based on GPU has achieved a qualitative leap compared with previous methods in NDFT computation, the computation capability is limited to the throughput capacity of GPU when the size of NDFT system is big enough.Solution method: We constructed a hybrid parallel architecture, in which CPU and GPU are combined to accelerate the NDFT computation effectively. Firstly, the NDFT system is divided into several subcells via domain-decomposition method. Then MPI (Message Passing Interface) is used to implement the CPU-parallel computation with each computer node corresponding to a particular subcell, and furthermore each subcell in one computer node will be executed in GPU in parallel efficiently. In this hybrid parallel method, the most critical technical problem is how to parallelize a CUNFFT in the parallel strategy, which is conquered effectively by deep-seated research of basic principles and some algorithm skills.Restrictions: The HP-NFFT is mainly oriented to, and has shown significant computational efficiency to deal with large-scale NDFT computations. However, for a small NDFT system containing less than 10^6 particles, the mode of multiple computer nodes has no apparent efficiency advantage or even lower efficiency compared with the mode of a single computer node because of the serious network delay.

A parallel hybrid implementation of the 2D acoustic wave equation

Abstract In this paper, we propose a hybrid parallel programming approach for a numerical solution of a two-dimensional acoustic wave equation using an implicit difference scheme for a single computer. The calculations are carried out in an implicit finite difference scheme. First, we transform the differential equation into an implicit finite-difference equation and then using the alternating direction implicit (ADI) method, we split the equation into two sub-equations. Using the cyclic reduction algorithm, we calculate an approximate solution. Finally, we change this algorithm to parallelize on graphics processing unit (GPU), GPU + Open Multi-Processing (OpenMP), and Hybrid (GPU + OpenMP + message passing interface (MPI)) computing platforms. The special focus is on improving the performance of the parallel algorithms to calculate the acceleration based on the execution time. We show that the code that runs on the hybrid approach gives the expected results by comparing our results to those obtained by running the same simulation on a classical processor core, Compute Unified Device Architecture (CUDA), and CUDA + OpenMP implementations.

Optimization of GPU parallel scheme for simulating ultrafast magnetization dynamics model

In this paper, we propose an optimized parallel scheme based on graphics processor unit (GPU) for simulating thermal induced ultrafast magnetization dynamics model using finite difference method. By optimizing the solution algorithm, optimizing memory accessing patterns and utilizing the characteristics of the compute unified device architecture (CUDA), the scheme of simulating ultrafast magnetization dynamics model is optimized. The simulation of the ultrafast magnetization dynamics for GdFe alloy filled with 32*32*256 cells on the NIVDIA Titan V GPU shows that our optimized GPU parallel scheme has about 208 times and 9 times acceleration respectively compared to serial CPU and parallel GPU schemes. The results of the simulation are in qualitative agreement with the reported simulations and experiments, making a great significance in the study of laser thermal induced ultrafast magnetization dynamics.

Stabilized variational formulation of an oldroyd-B fluid flow equations on a Graphic Processing Unit (GPU) architecture

The governing equations of the flow of an oldroyd-B fluid are discretized using the finite element method. To overcome the convective nature of the momentum equation, the Galerkin/Least-Squares Finite Element Method (GLS/FEM) is used while the Discrete Elastic–Viscous Stress-Splitting (DEVSS) method is used to overcome the instability due to the absence of diffusion in the constitutive equations. The discretized equations are implemented on a hybrid system between the Graphics Processing Unit (GPU) architecture using Compute-Unified-Device-Architecture (CUDA) and a multi-core CPU. The implementation is applied successfully to simulate the blood flow in abdominal aortic aneurysm. To accelerate application performance on the GPU several optimized approaches are adopted. The most significant approach is the coloring technique that is used to assemble the global matrix. Numerical experiments show that the hybrid CPU-GPU implementation has a 26 time speedup over the multi-core CPU implementations.

Performance gains with Compute Unified Device Architecture-enabled eddy current correction for diffusion MRI.

Correcting for eddy currents, movement-induced distortion and gradient inhomogeneities is imperative when processing diffusion MRI (dMRI) data, but is highly computing resource-intensive. Recently, Compute Unified Device Architecture (CUDA) was implemented for the widely-used eddy-correction software, 'eddy', which reduces processing time and allows more comprehensive correction. We investigated processing speed, performance and compatibility of CUDA-enabled eddy-current correction processing compared to commonly-used non-CUDA implementations. Four representative dMRI datasets from the Human Connectome Project, Alzheimer's Disease Neuroimaging Initiative and Chronic Diseases Connectome Project were processed on high-specification and regular workstations through three different configurations of 'eddy'. Processing times and graphics processing unit (GPU) resources used were monitored and compared. Using CUDA reduced the 'eddy' processing time by a factor of up to five. The CUDA slice-to-volume correction method was also faster than non-CUDA eddy except when datasets were large. We make a series of recommendations for eddy configuration and hardware. We suggest that users of eddy-correction software for dMRI processing utilise CUDA and take advantage of the slice-to-volume correction option. We recommend that users run eddy on computers with at least 32GB motherboard random access memory (RAM), and a graphics card with at least 4.5GB RAM and 3750 cores to optimise processing time.

Direct numerical simulations of turbulent channel flows with mesh-refinement lattice Boltzmann methods on GPU cluster

Wall-bounded turbulent channel flows up to Reτ=640 are simulated with multi-relaxation-time (MRT) lattice Boltzmann models. In order to perform direct numerical simulations and to resolve the wall boundary layer, a successive refinement procedure is developed to capture the rapid variation of the flow properties in the near-wall region. In the present implementation, three levels of grids are adopted with grid spacings in wall unit being Δ+∼1, 2 and 4, respectively. Due to the explicit nature of the numerical scheme, the numerical procedure is built on the CUDA framework, and the simulations are conducted on the GPU cluster. Strong scaling shows satisfactory scalability of the present numerical procedure. Overall, for the present channel flows, simulated results (Reτ=180, 395, and 640) with D3Q19 and D3Q27 models show marginal differences. With the grid refinement, both models produce good mean and turbulent quantities compared with the benchmark solutions.