线性系统求解中迭代算法的GPU加速方法

在求解线性系统时,迭代法是一种基本的方法,特别是在系数矩阵为大规模稀疏矩阵的情况下,高效地使用迭代法求解变得十分重要。本文通过分析迭代法的一般特点,提出了使用具有强大计算能力和存储带宽的GPU加速迭代法的一般方法。利用这些方法,在两种主流GPU平台上实现了一个经典的迭代法PQMRCGSTAB,并且针对不同的GPU平台特点提出了具体的优化方法。与AMD Opteron 2.4GHz 4核处理器相比,双精度版本的PQMRCGSTAB算法经NVIDIA Tesla S1070加速后性能提高31倍,经AMD Radeon HD 4870 X2加速后性能提高9倍。     

使用GPU加速BLAST算法初探

应用GPU通用高性能编程技术实现了一种加速BLAST算法的新方法。BLAST是目前最常用的用于生物序列查询比对的算法和软件包,其处理速度受到串行化执行和磁盘I/O等因素的影响。本文通过实验分析了BLAST软件包中的典型程序BLASTN的运行热点,并选定关键热点模块,应用CUDA编程技术对其进行并行化改造。对比实验结果表明,对于平均序列长度较大的序列库,应用GPGPU并行化可明显缩短该模块的运行时间,获得超过35倍的加速比。这说明,我们可以利用GPGPU对BLAST进行并行化加速,以满足高性能生物序列查询的需求。     

基于GPU的多分辨率体数据重构和渲染

基于小波变换的多分辨率压缩算法能够获得很高的压缩比,因而被广泛地用于压缩体数据.针对这种压缩策略,研究基于GPU的数据重构的方法,可以只从CPU向GPU传输少量的压缩数据,从而提高数据传输效率.因为好的数据结构是实现基于GPU的重构算法的关键,所以文中提出适合使用矩形纹理表示的数据结构--Nested Tileboard;然后给出基于该数据结构在GPU上实现多分辨率重构的方法,使用Nested Tileboard保存中间数据及重构结果;还提出了基于Nested Tileboard的多分辨率体绘制方法,直接对重构数据进行体绘制,从而实现数据重构和体绘制的无缝连接.     

大尺寸纹理的实时合成

提出一种纹理合成方法,可实时高质量地生成大纹理.它先基于纹理特征变化的周期性分析,得到合适的纹理块尺寸,以使所划分的纹理块能高效反映这种周期性变化,便于生成高质量的纹理;然后,它在目标纹理上均衡地分布纹理块,使得垂直方向和水平方向上相邻的纹理块之间都留有一个块尺寸大小的空白区域,再对空白区域进行填充,以完成目标纹理的生成.显然,布块操作和填充操作均可并行地进行.同时,为每个纹理块预先生成可与其邻接匹配的纹理块集合,以便在填充计算时可用简便的集合求交计算来进行邻域约束的搜索,并将这种求交计算放在CPU中进行,而将邻接纹理块在重叠区域的缝合计算放到GPU中进行,以综合利用CPU和GPU的优势.实验表明,新方法可在一般微机上以45帧/秒的速度高质量地实时合成1024*1024的大纹理,而这是已有技术难以达到的.     

基于GPU的无网格流体实时渲染

提出基于平滑粒子流体力学的自由界面流体模拟方法,采用了范德瓦尔斯方程与粒子间短距离排斥力和长距离吸引力作用的表面张力,设计出基于GPU的粒子泼溅算法。渲染算法完全消除了时间离散假象,具有交互式的高质量渲染效果。与传统拉格朗日算法相比,该方法具有简化的表面张力模型,快速的渲染方式,减小了运算的复杂性,有效提高了系统的运行速度。     

基于硬件加速的几何失真校正研究

分析研究一种新的图形硬件架构CUDA,基于其类C语言硬件开发环境和多线程的硬件执行机制,实现了基于GPU的几何失真校正的数学贴图映射,并根据像素分辨率的不同情况分别进行了实验验证,证明该方法比基于CPU的实现具有更高的场景交互速度,保证了视景系统的真实感和实时性.     

Parallel Cloth Simulation Using OpenGL Shading Language

The primary goal of cloth simulation is to express object behavior in a realistic manner and achieve real-time performance by following the fundamental concept of physic. In general, the mass–spring system is applied to real-time cloth simulation with three types of springs. However, hard spring cloth simulation using the mass–spring system requires a small integration time-step in order to use a large stiffness coefficient. Furthermore, to obtain stable behavior, constraint enforcement is used instead of maintenance of the force of each spring. Constraint force computation involves a large sparse linear solving operation. Due to the large computation, we implement a cloth simulation using adaptive constraint activation and deactivation techniques that involve the mass–spring system and constraint enforcement method to prevent excessive elongation of cloth. At the same time, when the length of the spring is stretched or compressed over a defined threshold, adaptive constraint activation and deactivation method deactivates the spring and generate the implicit constraint. Traditional method that uses a serial process of the Central Processing Unit (CPU) to solve the system in every frame cannot handle the complex structure of clothmodel in real-time. Our simulation utilizes the Graphic Processing Unit (GPU) parallel processing with compute shader in OpenGL Shading Language (GLSL) to solve the system effectively. In this paper, we design and implement parallel method for cloth simulation, and experiment on the performance and behavior comparison of the mass–spring system, constraint enforcement, and adaptive constraint activation and deactivation techniques the using GPU-based parallel method.     

Multi-GPU accelerated multi-spin Monte Carlo simulations of the 2D Ising model

A Modern Graphics Processing unit (GPU) is able to perform massively parallel scientific computations at low cost. We extend our implementation of the checkerboard algorithm for the two-dimensional Ising model [T. Preis et al., Journal of Chemical Physics 228 (2009) 4468-4477] in order to overcome the memory limitations of a single GPU which enables us to simulate significantly larger systems. Using multi-spin coding techniques, we are able to accelerate simulations on a single GPU by factors up to 35 compared to an optimized single Central Processor Unit (CPU) core implementation which employs multi-spin coding. By combining the Compute Unified Device Architecture (CUDA) with the Message Parsing Interface (MPI) on the CPU level, a single Ising lattice can be updated by a cluster of GPUs in parallel. For large systems, the computation time scales nearly linearly with the number of GPUs used. As proof of concept we reproduce the critical temperature of the 2D Ising model using finite size scaling techniques.     

On optimization of finite-difference time-domain (FDTD) computation on heterogeneous and GPU clusters

A model for the computational cost of the finite-difference time-domain (FDTD) method irrespective of implementation details or the application domain is given. The model is used to formalize the problem of optimal distribution of computational load to an arbitrary set of resources across a heterogeneous cluster. We show that the problem can be formulated as a minimax optimization problem and derive analytic lower bounds for the computational cost. The work provides insight into optimal design of FDTD parallel software. Our formulation of the load distribution problem takes simultaneously into account the computational and communication costs. We demonstrate that significant performance gains, as much as 75%, can be achieved by proper load distribution.     

Interactive Fibre Structure Visualization of the Heart

The heart consists of densely packed muscle fibres. The orientation of these fibres can be acquired by using Diffusion Tensor Imaging (DTI) ex vivo. A good way to visualize the fibre structure in a cross section of the heart is by showing short line segments originating from the cross section and aligned with the local direction of the fibres. If the line segments are placed dense enough, one can see how the fibre orientations change. However, generation of the line segments takes time and thus the user has to wait for new geometry to be generated when the plane defining the cross section is changed. We present a new direct rendering method for the visualization of the 3D vector field in a 2D user‐definable cross section of a heart. On the intersection of the plane with the vector field, the full 3D vectors are rendered as 3D line segments with a local ray casting approach. No preprocessing of the data is needed and no geometry is generated. This technique allows a fast inspection of the data to identify interesting areas where further analysis is necessary (e.g. quantification or generation of streamlines). We also show how the technique is generalized to other glyph shapes than line segments by implementing ellipsoids.