Parallel Image Processing: Taking Grayscale Conversion Using OpenMP as an Example

In recent years, the widespread adoption of parallel computing, especially in multi-core processors and high-performance computing environments, ushered in a new era of efficiency and speed. This trend was particularly noteworthy in the field of image processing, which witnessed significant advancements. This parallel computing project explored the field of parallel image processing, with a focus on the grayscale conversion of colorful images. Our approach involved integrating OpenMP into our framework for parallelization to execute a critical image processing task: grayscale conversion. By using OpenMP, we strategically enhanced the overall performance of the conversion process by distributing the workload across multiple threads. The primary objectives of our project revolved around optimizing computation time and improving overall efficiency, particularly in the task of grayscale conversion of colorful images. Utilizing OpenMP for concurrent processing across multiple cores significantly reduced execution times through the effective distribution of tasks among these cores. The speedup values for various image sizes highlighted the efficacy of parallel processing, especially for large images. However, a detailed examination revealed a potential decline in parallelization efficiency with an increasing number of cores. This underscored the importance of a carefully optimized parallelization strategy, considering factors like load balancing and minimizing communication overhead. Despite challenges, the overall scalability and efficiency achieved with parallel image processing underscored OpenMP’s effectiveness in accelerating image manipulation tasks.

The inversion of density structure by graphic processing unit(GPU) and identification of igneous rocks in Xisha area

Organic reefs, the targets of deep-water petro- leum exploration, developed widely in Xisha area. However, there are concealed igneous rocks undersea, to which organic rocks have nearly equal wave impedance. So the igneous rocks have become interference for future explo- ration by having similar seismic reflection characteristics. Yet, the density and magnetism of organic reefs are very different from igneous rocks. It has obvious advantages to identify organic reefs and igneous rocks by gravity and magnetic data. At first, frequency decomposition was applied to the free-air gravity anomaly in Xisha area to obtain the 2D subdivision of the gravity anomaly and magnetic anomaly in the vertical direction. Thus, the dis- tribution of igneous rocks in the horizontal direction can be acquired according to high-frequency field, low-frequency field, and its physical properties. Then, 3D forward model- ing of gravitational field was carried out to establish the density model of this area by reference to physical properties of rocks based on former researches. Furthermore, 3D inversion of gravity anomaly by genetic algorithm method of the graphic processing unit （GPU） parallel processing in Xisha target area was applied, and 3D density structure of this area was obtained. By this way, we can confine the igneous rocks to the certain depth according to the density of the igneous rocks. The frequency decomposition and 3D inversion of gravity anomaly by genetic algorithm method of the GPU parallel processing proved to be a useful method for recognizing igneous rocks to its 3D geological position. So organic reefs and igneous rocks can be identified, which provide a prescient information for further exploration.

TEB:GPU上矩阵分解重构的高效SpMV存储格式

稀疏矩阵向量乘法(SpMV)是科学与工程领域中一个至关重要的计算过程,CSR(compressed sparse row)格式是最常用的稀疏矩阵存储格式之一,在图形处理器(GPU)平台上实现并行SpMV的过程中,其只存储稀疏矩阵的非零元,避免零元素填充所带来的计算冗余,节约存储空间,但存在着负载不均衡的问题,浪费了计算资源。针对上述问题,对近年来效果良好的存储格式进行了研究,提出了一种逐行分解重组存储格式——TEB(threshold-exchangeorder block)格式。该格式采用启发式阈值选择算法确定合适分割阈值,并结合基于重排序的行归并算法,对稀疏矩阵进行重构分解,使得块与块之间非零元个数尽可能得相近,其次结合CUDA(computer unified device architecture)线程技术,提出了基于TEB存储格式的子块间并行SpMV算法,能够合理分配计算资源,解决负载不均衡问题,从而提高SpMV并行计算效率。为了验证TEB存储格式的有效性,在NVIDIA Tesla V100平台上进行实验,结果表明TEB相较于PBC(partition-block-CSR)、AMF-CSR(adaptive multi-row folding of CSR)、CSR-Scalar(compressed sparse row-scalar)和CSR5(compressed sparse row 5)存储格式,在SpMV的时间性能方面平均可提升3.23、5.83、2.33和2.21倍;在浮点计算性能方面,平均可提高3.36、5.95、2.29和2.13倍。

融合GPU的拟单层覆盖近似集计算方法

拟单层覆盖粗糙集是一种匹配集值信息系统且有高质量和高效率的粗糙集模型。拟单层覆盖近似集的计算过程中存在大量计算密集且逻辑简单的运算,为此,提出拟单层覆盖近似集的矩阵化表示方法,以利用图形处理器(GPU)强大的计算性能加速计算过程。为了实现这一目标,使用布尔矩阵表示拟单层覆盖近似空间中的元素,引入与集合运算对应的布尔矩阵算子,提出拟单层覆盖粗糙近似集(DE、DA、DE0与DA0)的矩阵表示,并设计矩阵化拟单层覆盖近似集算法(M_SMC)。同时,相应的定理证明了拟单层覆盖近似集的矩阵表示形式与原始定义的等价性。然而,M_SMC运行过程中出现了矩阵存储和计算步骤的内存消耗过多问题。为了将算法部署到显存有限的GPU上,优化矩阵存储和计算步骤,提出分批处理的矩阵化拟单层覆盖近似集算法(BM_SMC)。在10个数据集上的实验结果表明,融合GPU的BM_SMC算法与单纯使用中央处理器(CPU)的BM_SMC算法相比计算效率提高2.16~11.3倍,BM_SMC算法可以在有限的存储空间条件下充分利用GPU,能够有效地提高拟单层覆盖近似集的计算效率。

GNNSched:面向GPU的图神经网络推理任务调度框架

由于频繁的显存访问,图神经网络GNN在GPU上运行时往往资源利用率较低。现有的推理框架由于没有考虑GNN输入的不规则性,直接适用到GNN进行推理任务共置时可能会超出显存容量导致任务失败。对于GNN推理任务,需要根据其输入特点预先分析并发任务的显存占用情况,以确保并发任务在GPU上的成功共置。此外,多租户场景提交的推理任务亟需灵活的调度策略,以满足并发推理任务的服务质量要求。为了解决上述问题,提出了GNNSched,其在GPU上高效管理GNN推理任务的共置运行。具体来说,GNNSched将并发推理任务组织为队列,并在算子粒度上根据成本函数估算每个任务的显存占用情况。GNNSched实现了多种调度策略来生成任务组,这些任务组被迭代地提交到GPU并发执行。实验结果表明,GNNSched能够满足并发GNN推理任务的服务质量并降低推理任务的响应时延。

隐私计算环境下深度学习的GPU加速技术综述

随着深度学习技术的不断发展,神经网络模型的训练时间越来越长,使用GPU计算对神经网络训练进行加速便成为一项关键技术.此外,数据隐私的重要性也推动了隐私计算技术的发展.首先介绍了深度学习、GPU计算的概念以及安全多方计算、同态加密2种隐私计算技术,而后探讨了明文环境与隐私计算环境下深度学习的GPU加速技术.在明文环境下,介绍了数据并行和模型并行2种基本的深度学习并行训练模式,分析了重计算和显存交换2种不同的内存优化技术,并介绍了分布式神经网络训练过程中的梯度压缩技术.介绍了在隐私计算环境下安全多方计算和同态加密2种不同隐私计算场景下的深度学习GPU加速技术.简要分析了2种环境下GPU加速深度学习方法的异同.

基于GPU和角正交投影视图的多视角投影全息图

针对多视角投影全息图生成速度慢的问题,提出一种基于计算机图形处理单元(graphics processing unit,GPU)的多视角投影计算全息图合成方法.获取多个角正交投影视图,充分利用GPU强大的并行计算能力,同时计算多幅投影视图对全息图的作用,即在计算过程中同时将沿着投影方向移位后的一系列角正交投影视图乘以其相应的常数相位因子.其中,每个投影图像的投影角决定了其移位的距离和常数相位因子.将所有并行计算结果累加,可以得到一个包含物体三维信息的二维复矩阵,即菲涅尔全息图.相较于使用计算机中央处理器(central processing unit,CPU)进行计算,本方法显著提升了计算速度,将计算效率提高了30~40倍,为多视角投影全息图的高效生成提供一种可行途径.

Optimization of a precise integration method for seismic modeling based on graphic processing unit

General purpose graphic processing unit （GPU） calculation technology is gradually widely used in various fields. Its mode of single instruction, multiple threads is capable of seismic numerical simulation which has a huge quantity of data and calculation steps. In this study, we introduce a GPU-based parallel calculation method of a precise integration method （PIM） for seismic forward modeling. Compared with CPU single-core calculation, GPU parallel calculating perfectly keeps the features of PIM, which has small bandwidth, high accuracy and capability of modeling complex substructures, and GPU calculation brings high computational efficiency, which means that high-performing GPU parallel calculation can make seismic forward modeling closer to real seismic records.

Volumetric lattice Boltzmann method for pore-scale mass diffusionadvection process in geopolymer porous structures

Porous materials present significant advantages for absorbing radioactive isotopes in nuclear waste streams.To improve absorption efficiency in nuclear waste treatment,a thorough understanding of the diffusion-advection process within porous structures is essential for material design.In this study,we present advancements in the volumetric lattice Boltzmann method(VLBM)for modeling and simulating pore-scale diffusion-advection of radioactive isotopes within geopolymer porous structures.These structures are created using the phase field method(PFM)to precisely control pore architectures.In our VLBM approach,we introduce a concentration field of an isotope seamlessly coupled with the velocity field and solve it by the time evolution of its particle population function.To address the computational intensity inherent in the coupled lattice Boltzmann equations for velocity and concentration fields,we implement graphics processing unit(GPU)parallelization.Validation of the developed model involves examining the flow and diffusion fields in porous structures.Remarkably,good agreement is observed for both the velocity field from VLBM and multiphysics object-oriented simulation environment(MOOSE),and the concentration field from VLBM and the finite difference method(FDM).Furthermore,we investigate the effects of background flow,species diffusivity,and porosity on the diffusion-advection behavior by varying the background flow velocity,diffusion coefficient,and pore volume fraction,respectively.Notably,all three parameters exert an influence on the diffusion-advection process.Increased background flow and diffusivity markedly accelerate the process due to increased advection intensity and enhanced diffusion capability,respectively.Conversely,increasing the porosity has a less significant effect,causing a slight slowdown of the diffusion-advection process due to the expanded pore volume.This comprehensive parametric study provides valuable insights into the kinetics of isotope uptake in porous structures,facilitating the development of porous materials for nuclear waste treatment applications.

GPU加速下的三维快速分解后向投影SAS成像算法

后向投影(back projection,BP)算法是一种精确的时域成像算法,但BP算法的计算复杂度高,难以实现实时性成像,特别是在考虑三维成像时,BP算法的计算复杂度会进一步增加。提出一种应用在合成孔径声纳(synthetic aperture sonar,SAS)上的三维快速分解BP(fast factorized BP,FFBP)成像算法,并利用图形处理器(graphics processing unit,GPU)加速三维FFBP算法。经过对点目标的测试,计算时间从原本的263 s降低到了2.3 s,解决了SAS中的三维成像实时性问题。同时,验证了所提算法在非理想航迹下的成像效果。结果表明,在添加幅度不超过0.1 m(一个波长以内)的正弦扰动时,所提算法对点目标仍有良好的聚焦效果。

TIME-DOMAIN INTERPOLATION ON GRAPHICS PROCESSING UNIT

The signal processing speed of spectral domain optical coherence tomography(SD-OCT)has become a bottleneck in a lot of medical applications.Recently,a time-domain interpolation method was proposed.This method can get better signal-to-noise ratio(SNR)but much-reduced signal processing time in SD-OCT data processing as compared with the commonly used zeropadding interpolation method.Additionally,the resampled data can be obtained by a few data and coefficients in the cutoff window.Thus,a lot of interpolations can be performed simultaneously.So,this interpolation method is suitable for parallel computing.By using graphics processing unit(GPU)and the compute unified device architecture(CUDA)program model,time-domain interpolation can be accelerated significantly.The computing capability can be achieved more than 250,000 A-lines,200,000 A-lines,and 160,000 A-lines in a second for 2,048 pixel OCT when the cutoff length is L=11,L=21,and L=31,respectively.A frame SD-OCT data(400A-lines×2,048 pixel per line)is acquired and processed on GPU in real time.The results show that signal processing time of SD-OCT can befinished in 6.223 ms when the cutoff length L=21,which is much faster than that on central processing unit(CPU).Real-time signal processing of acquired data can be realized.

Compute Unified Device Architecture Implementation of Euler/Navier-Stokes Solver on Graphics Processing Unit Desktop Platform for 2-D Compressible Flows

Personal desktop platform with teraflops peak performance of thousands of cores is realized at the price of conventional workstations using the programmable graphics processing units(GPUs).A GPU-based parallel Euler/Navier-Stokes solver is developed for 2-D compressible flows by using NVIDIA′s Compute Unified Device Architecture(CUDA)programming model in CUDA Fortran programming language.The techniques of implementation of CUDA kernels,double-layered thread hierarchy and variety memory hierarchy are presented to form the GPU-based algorithm of Euler/Navier-Stokes equations.The resulting parallel solver is validated by a set of typical test flow cases.The numerical results show that dozens of times speedup relative to a serial CPU implementation can be achieved using a single GPU desktop platform,which demonstrates that a GPU desktop can serve as a costeffective parallel computing platform to accelerate computational fluid dynamics(CFD)simulations substantially.

Multi-relaxation-time lattice Boltzmann simulations of lid driven flows using graphics processing unit

Large eddy simulation （LES） using the Smagorinsky eddy viscosity model is added to the two-dimensional nine velocity components （D2Q9） lattice Boltzmann equation （LBE） with multi-relaxation-time （MRT） to simulate incompressible turbulent cavity flows with the Reynolds numbers up to 1 × 10^7. To improve the computation efficiency of LBM on the numerical simulations of turbulent flows, the massively parallel computing power from a graphic processing unit （GPU） with a computing unified device architecture （CUDA） is introduced into the MRT-LBE-LES model. The model performs well, compared with the results from others, with an increase of 76 times in computation efficiency. It appears that the higher the Reynolds numbers is, the smaller the Smagorinsky constant should be, if the lattice number is fixed. Also, for a selected high Reynolds number and a selected proper Smagorinsky constant, there is a minimum requirement for the lattice number so that the Smagorinsky eddy viscosity will not be excessively large.

Graphic Processing Unit-Accelerated Neural Network Model for Biological Species Recognition

A graphic processing unit (GPU)-accelerated biological species recognition method using partially connected neural evolutionary network model is introduced in this paper. The partial connected neural evolutionary network adopted in the paper can overcome the disadvantage of traditional neural network with small inputs. The whole image is considered as the input of the neural network, so the maximal features can be kept for recognition. To speed up the recognition process of the neural network, a fast implementation of the partially connected neural network was conducted on NVIDIA Tesla C1060 using the NVIDIA compute unified device architecture (CUDA) framework. Image sets of eight biological species were obtained to test the GPU implementation and counterpart serial CPU implementation, and experiment results showed GPU implementation works effectively on both recognition rate and speed, and gained 343 speedup over its counterpart CPU implementation. Comparing to feature-based recognition method on the same recognition task, the method also achieved an acceptable correct rate of 84.6% when testing on eight biological species.

NTRU格基密钥封装方案GPU高性能实现

随着量子计算技术的发展,传统加密算法受到的威胁日益严重.为应对量子计算时代的挑战,各国正积极加强后量子密码算法的实现和迁移部署工作.由于NTRU密码方案具有结构简洁、计算效率高、尺寸较小、无专利风险等优点,因此NTRU格基密钥封装算法对于后量子时代的密码技术储备和应用具有重要意义.同时,图形处理器(Graphics Processing Unit,GPU)以其强大的并行计算能力、高吞吐量、低能耗等特性,已成为当前高并发密码工程实现的重要平台.本文给出后量子密码算法CTRU/CNTR的首个GPU高性能实现方案.对GPU主要资源占用进行分析,我们综合考虑并行计算、内存访问、数据布局和算法优化等多个方面,采用一系列计算和内存优化技术,旨在并行加速计算、优化访存、合理占用GPU资源以及减少I/O时延,从而提高本方案的计算能力和性能.本文的主要贡献在于以下几个方面:首先,针对模约减操作,使用NVIDIA并行指令集实现,有效减少所需指令条数;其次,针对耗时的多项式乘法模块,采用混合基NTT,并采用层融合、循环展开和延迟约减等方法,加快计算速度;此外,针对内存重复访问和冲突访问等问题,通过合并访存、核函数融合等优化技术,实现内存的高效访问;最后,为实现高并行的算法,设计恰当的线程块大小和数量,采用内存池机制,实现多任务的快速访存和高效处理.基于NVIDIA RTX4090平台,本方案CTRU768实现中密钥生成、封装和解封装的吞吐量分别为每秒1170.9万次、926.7万次和315.4万次.与参考实现相比,密钥生成、封装和解封装的吞吐量分别提高了336倍、174倍和128倍.本方案CNTR768实现中密钥生成、封装和解封装的吞吐量分别为每秒1117.3万次、971.8万次和322.2万次.与参考实现相比,密钥生成、封装和解封装的吞吐量分别提高了329倍、175倍和134倍;与开源Kyber实现相比,密钥生成、密钥封装和密钥解封装的吞吐量分别提升10.84~11.36倍、9.49~9.95倍和5.11~5.22倍.高性能的密钥封装实现在大规模任务处理场景下具有较大的应用潜力,对保障后量子时代的信息和数据安全具有重要意义.

Falcon后量子算法的密钥树生成部件GPU并行优化设计与实现

近年来,后量子密码算法因其具有抗量子攻击的特性成为安全领域的研究热点。基于格的Falcon数字签名算法是美国国家标准与技术研究所(NIST)公布的首批4个后量子密码标准算法之一。密钥树生成是Falcon算法的核心部件,在实际运算中占用较多的时间和消耗较多的资源。为此,提出一种基于图形处理器(GPU)的Falcon密钥树并行生成方案。该方案使用奇偶线程联合控制的单指令多线程(SIMT)并行模式和无中间变量的直接计算模式,达到了提升速度和减少资源占用的目的。基于Python的CUDA平台进行了实验,验证结果的正确性。实验结果表明,Falcon密钥树生成在RTX 3060 Laptop的延迟为6 ms,吞吐量为167次/s,在计算单个Falcon密钥树生成部件时相对于CPU实现了1.17倍的加速比,在同时并行1024个Falcon密钥树生成部件时,GPU相对于CPU的加速比达到了约56倍,在嵌入式Jetson Xavier NX平台上的吞吐量为32次/s。

Graphic Processing Unit-Accelerated Mutual Information-Based 3D Image Rigid Registration

Mutual information （MI）-based image registration is effective in registering medical images, but it is computationally expensive. This paper accelerates MI-based image registration by dividing computation of mutual information into spatial transformation and histogram-based calculation, and performing 3D spatial transformation and trilinear interpolation on graphic processing unit （GPU）. The 3D floating image is downloaded to GPU as flat 3D texture, and then fetched and interpolated for each new voxel location in fragment shader. The transformed resuits are rendered to textures by using frame buffer object （FBO） extension, and then read to the main memory used for the remaining computation on CPU. Experimental results show that GPU-accelerated method can achieve speedup about an order of magnitude with better registration result compared with the software implementation on a single-core CPU.

基于GPU的LBM迁移模块算法优化

格子玻尔兹曼方法(LBM)是一种基于介观模拟尺度的计算流体力学方法,其在计算时设置大量的离散格点,具有适合并行的特性。图形处理器(GPU)中有大量的算术逻辑单元,适合大规模的并行计算。基于GPU设计LBM的并行算法,能够提高计算效率。但是LBM算法迁移模块中每个格点的计算都需要与其他格点进行通信,存在较强的数据依赖。提出一种基于GPU的LBM迁移模块算法优化策略。首先分析迁移部分的实现逻辑,通过模型降维,将三维模型按照速度分量离散为多个二维模型,降低模型的复杂度;然后分析迁移模块计算前后格点中的数据差异,通过数据定位找到迁移模块的通信规律,并对格点之间的数据交换方式进行分类;最后使用分类的交换方式对离散的二维模型进行区域划分,设计新的数据通信方式,由此消除数据依赖的影响,将迁移模块完全并行化。对并行算法进行测试,结果显示:该算法在1.3×10^(8)规模网格下能达到1.92的加速比,表明算法具有良好的并行效果;同时对比未将迁移模块并行化的算法,所提优化策略能提升算法30%的并行计算效率。

基于GPU对角稀疏矩阵向量乘法的动态划分算法

在图形处理器(GPU)上实现对角稀疏矩阵向量乘法(SpMV)可以充分利用GPU的并行计算能力,并加速矩阵向量乘法;然而,相关主流算法存在零元填充数据多、计算效率低的问题。针对上述问题,提出一种对角SpMV算法DIA-Dynamic(DIAgonal-Dynamic)。首先,设计一种全新的动态划分策略,根据矩阵的不同特征进行分块,在保证GPU高计算效率的同时大幅减少零元填充,去除冗余计算量;其次,提出一种对角稀疏矩阵存储格式BDIA(Block DIAgonal)存储分块数据,并调整数据布局,提高GPU上的访存性能;最后,基于GPU的底层进行条件分支优化,以减少分支判断,并使用动态共享内存解决向量的不规则访问问题。DIA-Dynamic与前沿Tile SpMV算法相比,平均加速比达到了1.88;与前沿BRCSD(Diagonal Compressed Storage based on Row-Blocks)-Ⅱ算法相比,平均零元填充减少了43%,平均加速比达到了1.70。实验结果表明,DIA-Dynamic能够有效提高GPU上对角SpMV的计算效率,缩短计算时间,提升程序性能。

基于GPU加速的全源对最短路径并行算法

针对最短路径算法处理大规模数据集低效的问题,提出了基于图形处理器(Graphics Processing Unit,GPU)加速的全源对最短路径并行算法.首先通过优化矩阵乘法算法实现了在工作组内和组间进行并行运算数据,然后减少了非规则行造成的工作项分支,最后降低了工作项对邻接矩阵计算条带存储资源的访问延时.实验结果表明,与基于AMD Ryzen5 1600X CPU的串行算法、基于开放多处理(Open Multi-Processing, OpenMP)并行算法和基于统一计算设备架构(Compute Unified Device Architecture, CUDA)并行算法相比,最短路径并行算法在开放式计算语言(Open Computing Language, OpenCL)架构下NVIDIA GeForce GTX 1 070计算平台上分别获得了196.35、36.76和2.25倍的加速比,验证了提出的并行优化方法的有效性和性能可移植性.