CUDA到异构众核架构的线程映射模型

统一计算设备架构(CUDA)程序移植到其他异构众核架构时的线程数不匹配。为此,提出一种层次化的线程映射模型。在第1个映射层次上,将CUDA主机端线程和设备端线程分别映射到目标平台的主核和从核阵列上,在第2个映射层次上,采用线程循环的方法消除协作线程阵列(CTA)中线程间同步操作,将整个CTA映射到从核阵列的一个从核上。实验结果表明,该模型能使CUDA程序在其他异构众核系统上得到有效运行。     

面向异构众核的CUDA程序二进制翻译

通过二进制翻译手段将CUDA程序移植到其他异构众核处理器平台特别是国产处理器平台,对扩展CUDA程序应用范围,发挥目标平台的众核优势以及支持民族产业都具有现实意义。设计了CUDA程序的二进制翻译框架,从CUDA程序可执行代码入手,采用“分而治之”的手段,将主机端代码和设备端代码分别翻译。重点介绍了移植过程中几个关键问题的解决,包括设备端代码的提取,计算模型的映射,存储模型的映射,栅栏同步和指令翻译问题。通过实验验证了系统功能的正确性。     

基于CUDA技术的卷积神经网络识别算法

针对具有高浮点运算能力的流处理器设备GPU对神经网络的适用性问题,提出卷积神经网络的并行化识别算法,采用计算统一设备架构(CUDA)技术,并定义其上的并行化数据结构,描述计算任务到CUDA的映射机制。实验结果证明,在GTX200硬件架构的GPU上实现的并行识别算法的平均浮点运算能力峰值较CPU上串行算法提高了近60倍,更适用于神经网络的相关应用。     

一种单GPU程序向多GPU移植的模板化技术

图形处理器(GPU)作为一种高度并行化的处理器架构,已得到越来越多的重视,目前已诞生了以NVIDIA CUDA为代表的各种GPU通用计算技术,同时多GPU并行计算也已有了实际的应用.多GPU并行计算涉及GPU与CPU两者之间的协调和交互,对程序员有着更高的要求.为此,提出一种基于模板的源代码生成技术,通过模板转化来支持单GPU程序的并行化移植.最后通过一个实例表明使用提出的CUDA源代码移植框架能够自动生成与手写程序等价的代码,可以显著降低多GPU下CUDA程序的开发代价,提高CUDA应用程序员的生产效率.     

基于CUDA的并行粒子群优化算法的设计与实现

针对处理大量数据和求解大规模复杂问题时粒子群优化(PSO)算法计算时间过长的问题, 进行了在显卡(GPU)上实现细粒度并行粒子群算法的研究。通过对传统PSO算法的分析, 结合目前被广泛使用的基于GPU的并行计算技术, 设计实现了一种并行PSO方法。本方法的执行基于统一计算架构（CUDA）, 使用大量的GPU线程并行处理各个粒子的搜索过程来加速整个粒子群的收敛速度。程序充分使用CUDA自带的各种数学计算库, 从而保证了程序的稳定性和易写性。通过对多个基准优化测试函数的求解证明, 相对于基于CPU的串行计算方法, 在求解收敛性一致的前提下, 基于CUDA架构的并行PSO求解方法可以取得高达90倍的计算加速比。     

多核CPU和GPU加速分子动力学模拟

在多核中央处理器(CPU)—图形处理器(GPU)异构并行体系结构上,采用OpenMP和计算统一设备架构(CUDA)编程实现了基于AMBER力场的蛋白质分子动力学模拟程序。通过合理地将程序划分为CPU单线程、CPU多线程和GPU多线程执行部分,高效地利用了计算机的处理能力。性能测试结果表明,相对于优化后的CPU串行计算,多核CPU-GPU异构并行计算模型有强大的性能优势,特别是将占整个程序执行时间90%的作用力的计算移植到GPU上执行,获得了最高可达12倍的计算加速比。     

面向异构多核架构的自适应编译框架

针对应用在移植到异构多核高性能计算机系统中所面临的可移植性差以及性能优化难度大的问题,文中提出一种面向异构多核架构的自适应编译框架.通过源到源编译解决传统并行编程模型应用向异构多核架构的映射问题;同时利用动态剖分信息,自适应地调整插桩并配置优化策略,形成迭代式的自动优化过程.文中自适应编译框架将软硬件映射机制与优化策略结合,有效地解决了同构并行应用向异构多核架构的移植问题并提高了应用的整体性能.实验结果表明,文中基于Cell架构实现的原型系统,很好地解决了异构多核架构下应用移植性等问题,同时应用性能有所提高.     

一种基于Hadoop+CUDA实现相关器的方法

根据21CMA相关器的算法特点,在对比基于CPU并行的MPI集群、MPI+CUDA异构并行集群和Hadoop+CUDA异构并行集群的架构特点的基础上,提出了一种基于Hadoop+CUDA平台实现软相关器的方法。本方法利用GPU在计算FFT、向量乘和向量加等密集型计算模型的优势,设计相关器的并行模型,使其性能较前期在CPU并行的MPI集群实现的相关器有了大幅提升。同时,本文选择广泛应用于大数据处理平台的Hadoop软件架构,利用Hadoop Streaming工具实现非Java编写的程序在分布式系统中并行执行,非常便捷地获得了集群系统的线性加速比。Hadoop HDFS并行文件系统管理结果数据和过程日志更加灵活可靠,为后续的大数据分析提供了支撑环境。     

CUDA软硬件环境简介

CUDA是用于GPU计算的开发环境,它是一个全新的软硬件架构,可以将GPU视为一个并行数据计算的设备,对所进行的计算进行分配和管理。在CUDA的架构中,这些计算不再像过去所谓的GPGPU架构那样必须将计算映射到图形API（OpenGL和Direct3D）中,因此对于开发者来说,CUDA的开发门槛大大降低了。CUDA的GPU编程语言基于标准的C语言,因此任何有C语言基础的用户都很容易地开发CUDA的应用程序。     

基于CUDA的快速图像压缩

为了进一步提高JPEG编码效率,对JPEG压缩算法进行研究,分析得出JPEG核心步骤可以并行化处理.因此,实现平台宜采用以并行计算为优势的GPU,而不是以串行计算为主的CPU.NVIDIA新推出的CUDA(计算统一设备架构)为此实现提供了软硬件环境.CUDA是基于GPU进行通用计算的开发平台,非常适合大规模的并行数据计算.在GPU流处理器架构下用CUDA技术实现编码并行化,并针对流处理器架构特点进行内存读写等方面的优化,提高了JPEG编码的速度.实验结果表明了CUDA技术在并行处理方面的优越性,JPEG编码效率得到了极大提高.     

Parallel Computing Experiences with CUDA

The CUDA programming model provides a straightforward means of describing inherently parallel computations, and NVIDIA's Tesla GPU architecture delivers high computational throughput on massively parallel problems. This article surveys experiences gained in applying CUDA to a diverse set of problems and the parallel speedups over sequential codes running on traditional CPU architectures attained by executing key computations on the GPU.     

Optimizations of a GPU accelerated heat conduction equation by a programming of CUDA Fortran from an analysis of a PTX file

The Fortran language has been commonly used for many kinds of scientific computation. In this paper, we focus on the solution of an unsteady heat conduction equation, which is one of the simplest problems for thermal dynamics. Recently, a GPU (graphics processing unit) has been enhanced with a Fortran programming language capability employing CUDA (compute unified device architecture), known as CUDA Fortran. We find that the speed performance of a system using an ordinary program coding of CUDA Fortran is lower than that of systems using a program coding of CUDA C. We also find that intermediate assembly files PTX (parallel thread execution) of the two languages are not coincident. Therefore, by comparing the PTX files from the two coding programs we could detect the bottleneck that causes the speed reduction. We propose three optimization techniques that can enable the calculated speeds using CUDA Fortran and CUDA C to be coincident. The optimizations can be performed by the Fortran language when improved by an analyzed PTX file. It is thus possible to improve the performance of CUDA Fortran by adding a correction to it, which happens to be at a programming language level.     

A compound OpenMP/MPI program development toolkit for hybrid CPU/GPU clusters

In this paper, we propose a program development toolkit called OMPICUDA for hybrid CPU/GPU clusters. With the support of this toolkit, users can make use of a familiar programming model, i.e., compound OpenMP and MPI instead of mixed CUDA and MPI or SDSM to develop their applications on a hybrid CPU/GPU cluster. In addition, they can adapt the types of resources used for executing different parallel regions in the same program by means of an extended device directive according to the property of each parallel region. On the other hand, this programming toolkit supports a set of data-partition interfaces for users to achieve load balance at the application level no matter what type of resources are used for the execution of their programs.     

基于CUDA的汇流分析并行算法的研究与实现*

针对基于数字高程模型（DEM）生成流域等流时线的快速运算问题,提出了一种基于统一设备计算架构（CUDA）平台同时可发挥图形处理器（GPU）并行运算特性的汇流分析的快速并行算法。采用改进后的归并排序算法进行数据排序及新的内存分配策略和改进的并行算法进行汇流分析。用该并行算法和CPU上的串行算法, 对生成基于DEM的等流时线运算时间和矩阵乘法运算时间进行分析验证。实验结果表明,基于CUDA的汇流分析并行算法能提高系统的计算效率,具有较好的效果。     

dOCAL: high-level distributed programming with OpenCL and CUDA

In the state-of-the-art parallel programming approaches OpenCL and CUDA, so-called host code is required for program’s execution. Efficiently implementing host code is often a cumbersome task, especially when executing OpenCL and CUDA programs on systems with multiple nodes, each comprising different devices, e.g., multi-core CPU and graphics processing units; the programmer is responsible for explicitly managing node’s and device’s memory, synchronizing computations with data transfers between devices of potentially different nodes and for optimizing data transfers between devices’ memories and nodes’ main memories, e.g., by using pinned main memory for accelerating data transfers and overlapping the transfers with computations. We develop distributed OpenCL/CUDA abstraction layer (dOCAL)—a novel high-level C++ library that simplifies the development of host code. dOCAL combines major advantages over the state-of-the-art high-level approaches: (1) it simplifies implementing both OpenCL and CUDA host code by providing a simple-to-use, high-level abstraction API; (2) it supports executing arbitrary OpenCL and CUDA programs; (3) it allows conveniently targeting the devices of different nodes by automatically managing node-to-node communications; (4) it simplifies implementing data transfer optimizations by providing different, specially allocated memory regions, e.g., pinned main memory for overlapping data transfers with computations; (5) it optimizes memory management by automatically avoiding unnecessary data transfers; (6) it enables interoperability between OpenCL and CUDA host code for systems with devices from different vendors. Our experiments show that dOCAL significantly simplifies the development of host code for heterogeneous and distributed systems, with a low runtime overhead.

     

PPModel: a modeling tool for source code maintenance and optimization of parallel programs

As the computation power in desktops advances, parallel programming has emerged as one of the essential skills needed by next generation software engineers. However, programs written in popular parallel programming paradigms have a substantial amount of sequential code mixed with the parallel code. Several such versions supporting different platforms are necessary to find the optimum version of the program for the available resources and problem size. As revealed by our study on benchmark programs, sequential code is often duplicated in these versions. This can affect code comprehensibility and re-usability of the software. In this paper, we discuss a framework named PPModel, which is designed and implemented to free programmers from these scenarios. Using PPModel, a programmer can separate parallel blocks in a program, map these blocks to various platforms, and re-execute the entire program. We provide a graphical modeling tool (PPModel) intended for Eclipse users and a Domain-Specific Language (tPPModel) for non-Eclipse users to facilitate the separation, the mapping, and the re-execution. This is illustrated with a case study from a benchmark program, which involves re-targeting a parallel block to CUDA and another parallel block to OpenMP. The modified program gave almost 5× performance gain compared to the sequential counterpart, and 1.5× gain compared to the existing OpenMP version.     

CUDA并行技术与数字图像几何变换

CUDA是GPU通过并发执行多个线程以实现大规模快速并行计算能力的技术,它能使对GPU编程变得更容易。介绍了CUDA基本特性及主要编程模型,在此基础上,提出并实现了基于NVIDIA CUDA技术的图像快速几何变换。采用位置偏移增量代替原变换算法中大量乘法运算,并把CUDA技术的快速并行计算能力应用到数字图像几何变换中,解决了基于CPU的传统图像几何变换运算效率低下的问题。实验结果证明使用CUDA技术,随着处理图像尺寸的增加,对数字图像几何变换处理效率最高能够提高到近100倍。     

Code compaction for parallel architectures

There are two principal methods used to exploit the parallelism available on a parallel machine: the program to be executed can be optimized by hand, or the program can be automatically converted to parallel machine code by a compiler. The first method usually derives parallelism at the procedure level; a parallel program is written in a high-level language and typically has various modules executing in parallel. By contrast, the compiler methodically transforms the program into parallel code using various transformations, such as code movement. The automatic conversion of a program to parallel code is called compaction or parallelization. This paper describes the evolution of a new compaction program and presents a new algorithm for determining legal code movements. A simulator of the target architecture was used to estimate the execution times of a sample suite of programs before and after compaction. The results verify that substantial advantages arise from applying this compaction technique.     

The Hydra Parallel Programming System

The Hydra Parallel Programming System, a new parallel language extension to Java, and its supporting software are described. It is a fairly simple yet powerful language designed to address a number of areas that have not received much attention. One of these areas is the recompilation of parallel programs at runtime to allow a parallel program to adapt to the architecture it is executing on. The first version of this software system focuses on smaller Symmetric Multiprocessing and compatible architectures which are becoming more common. This particular class of machines has a great need for more options in the area of parallel programming among the vastly popular Java language programmers. Hydra programs will run as sequential Java on machines that do not have the parallel support or do not have an implemented Hydra runtime system without requirement of any modifications to the program. This paper describes the language, compares it with other languages (specifically with JOMP, an OpenMP implementation for Java), presents a brief discussion on compiling and executing Hydra programs, presents some sample benchmarks and their performance on three platforms, and concludes with a discussion of issues and future directions for Hydra. Copyright © 2007 John Wiley & Sons, Ltd.