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。

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。

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.

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.

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.

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.

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。

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.

晶硅分子动力学模拟的GPU加速算法优化

分子动力学(MD)模拟是研究硅纳米薄膜热力学性质的主要方法,但存在数据处理量大、计算密集、原子间作用模型复杂等问题,限制了MD模拟的深入应用。针对晶硅分子动力学模拟算法中数据访问不连续和大量分支判断造成并行资源浪费、线程等待等问题,结合Nvidia Tesla V100 GPU硬件体系结构特点,对晶硅MD模拟算法进行设计。通过全局内存的合并访存、循环展开、原子操作等优化方法,利用GPU强大并行计算和浮点运算能力,减少显存访问及算法执行过程中的分支冲突和判断指令,提升算法整体计算性能。测试结果表明,优化后的晶硅MD模拟算法的计算速度相比于优化前提升了1.69~1.97倍,相比于国际上主流的GPU加速MD模拟软件HOOMDblue和LAMMPS分别提升了3.20~3.47倍和17.40~38.04倍,具有较好的模拟加速效果。

Speeding up the MATLAB complex networks package using graphic processors

The availability of computers and communication networks allows us to gather and analyse data on a far larger scale than previously. At present, it is believed that statistics is a suitable method to analyse networks with millions, or more, of vertices. The MATLAB language, with its mass of statistical functions, is a good choice to rapidly realize an algorithm prototype of complex networks. The performance of the MATLAB codes can be further improved by using graphic processor units （GPU）. This paper presents the strategies and performance of the GPU implementation of a complex networks package, and the Jacket toolbox of MATLAB is used. Compared with some commercially available CPU implementations, GPU can achieve a speedup of, on average, 11.3x. The experimental result proves that the GPU platform combined with the MATLAB language is a good combination for complex network research.

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.

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.

Fast modeling of gravity gradients from topographic surface data using GPU parallel algorithm

The gravity gradient is a secondary derivative of gravity potential,containing more high-frequency information of Earth’s gravity field.Gravity gradient observation data require deducting its prior and intrinsic parts to obtain more variational information.A model generated from a topographic surface database is more appropriate to represent gradiometric effects derived from near-surface mass,as other kinds of data can hardly reach the spatial resolution requirement.The rectangle prism method,namely an analytic integration of Newtonian potential integrals,is a reliable and commonly used approach to modeling gravity gradient,whereas its computing efficiency is extremely low.A modified rectangle prism method and a graphical processing unit(GPU)parallel algorithm were proposed to speed up the modeling process.The modified method avoided massive redundant computations by deforming formulas according to the symmetries of prisms’integral regions,and the proposed algorithm parallelized this method’s computing process.The parallel algorithm was compared with a conventional serial algorithm using 100 elevation data in two topographic areas(rough and moderate terrain).Modeling differences between the two algorithms were less than 0.1 E,which is attributed to precision differences between single-precision and double-precision float numbers.The parallel algorithm showed computational efficiency approximately 200 times higher than the serial algorithm in experiments,demonstrating its effective speeding up in the modeling process.Further analysis indicates that both the modified method and computational parallelism through GPU contributed to the proposed algorithm’s performances in experiments.

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 re-sults 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.

GPU上的查询算子的设计与优化

选择、连接、投影和聚集等是传统关系型数据库中的基本操作。为了实现关系型数据库在GPU上的查询优化,必须使用相应的GPU算法实现对应的关系算子。借鉴GDB分而治之的分层设计思想将关系代数拆分成算子层和原语层。数据查询处理过程中存在着一些难点问题,如数据传输时延、过度使用共享内存、活跃线程数减少和线程之间数据通信产生的通信时延。针对这些问题,基于较新的Pascal架构实现了查询优化算法,在原有的连接、聚集和条件选择算法原理基础上,对相应的算法进行了设计与优化。提高了每个工作线程的工作负载,实现了内核计算与数据传输之间的延迟隐藏,解决了连接操作中的数据倾斜问题。

基于GPU的数字下变频累积误差控制方法

针对数字下变频系统中高速数据处理实时性需求,设计了基于中央处理器(central processing unit,CPU)+图形处理单元(graphic processing unit,GPU)的异构平台数字下变频处理算法,并完成了相应并行程序设计,进行了实际数据验证。鉴于GPU运算采用单精度浮点数格式进行数据处理,进行长时程数据处理过程中,会产生浮点数舍入误差累积,导致结果的不准确不可靠,设计了联合相位循环归零法和无误差变换补偿法,实现了浮点数舍入误差的控制与补偿。数据测试结果表明,所提方法可将相位误差控制在10-11rad以内,且不随时间累积,计算结果准确可靠。

基于GPU的固态晶体硅分子动力学算法优化

分子动力学模拟通常用于晶体硅热力学性质的研究,因原子间采用复杂的多体作用势,分子模拟通常面临较高的计算负载,导致计算的时间和空间尺度受限。图形处理器(GPU)采用并行多线程技术,用于计算密集型处理任务,在分子动力学模拟领域中显示巨大的应用潜力。因此,充分利用GPU硬件架构特性提升固态共价晶体硅分子动力学模拟的时空尺度对晶体硅导热机制的研究具有重要意义。基于固态共价晶体硅分子动力学模拟算法,提出面向GPU计算平台的固定邻居算法设计与优化。利用数据结构、分支结构优化等方法解决分子动力学模拟的固定邻居算法全局访存和分支结构的耗时问题,降低数据访存消耗和分支冲突,通过改变线程并行调度方式,在GPU计算平台上实现高性能并行计算,有效解决计算负载问题。实验结果表明,LAMMPS双精度固态晶体硅分子动力学模拟与双精度固定邻居算法的加速比为11.62,HOOMD-blue双精度固态晶体硅分子动力学模拟与双精度固定邻居算法和单精度固定邻居算法的加速比分别为9.39和12.18。

带状稀疏矩阵乘法及高效GPU实现

稀疏-稠密矩阵乘法(SpMM)广泛应用于科学计算和深度学习等领域,提高它的效率具有重要意义。针对具有带状特征的一类稀疏矩阵,提出一种新的存储格式BRCV(Banded Row Column Value)以及基于此格式的SpMM算法和高效图形处理单元(GPU)实现。由于每个稀疏带可以包含多个稀疏块,所提格式可看成块稀疏矩阵格式的推广。相较于常用的CSR(Compressed Sparse Row)格式,BRCV格式通过避免稀疏带中列下标的冗余存储显著降低存储复杂度;同时,基于BRCV格式的SpMM的GPU实现通过同时复用稀疏和稠密矩阵的行更高效地利用GPU的共享内存,提升SpMM算法的计算效率。在两种不同GPU平台上针对随机生成的带状稀疏矩阵的实验结果显示,BRCV的性能不仅优于cuBLAS(CUDA Basic Linear Algebra Subroutines),也优于基于CSR和块稀疏两种不同格式的cuSPARSE。其中,相较于基于CSR格式的cuSPARSE,BRCV的最高加速比分别为6.20和4.77。此外,将新的实现应用于图神经网络(GNN)中的SpMM算子的加速。在实际应用数据集上的测试结果表明,BRCV的性能优于cuBLAS和基于CSR格式的cuSPARSE,且在大多数情况下优于基于块稀疏格式的cuSPARSE。其中,相较于基于CSR格式的cuSPARSE,BRCV的最高加速比为4.47。以上结果表明BRCV可以有效提升SpMM的效率。