自己做的部分习题解答,因为时间关系,有些马虎,也不全面,欢迎探讨或指出错误

5.1 Consider the matrixaddition in Exercise 3.1. Can one use shared memory to reduce theglobal memory bandwidth consumption?

Hint: analyze the elementsaccessed by each thread and see if there is any commonality betweenthreads.

Answer:I think there is no need to use shared memory in Exercise3.1, becauseall threads only use their variables once and no variables need to beshared between threads.

5.2 Draw the equivalent ofFigure 5.6 for a 8*8 matrix multiplication with 2*2 tiling and 4*4tiling. Verify that the reduction in global memory bandwidth isindeed proportional to the dimension size of the tiles.

Answer:

1.A 8*8matrix multiplication with 2*2tiling

Block0,0

Phase1

Phase2

thread0,0

M0,0

Mds0,0

N0,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

M0,2

Mds0,0

N2,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

thread0,1

M0,1

Mds0,1

N0,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

M0,3

Mds0,1

N2,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

thread1,0

M1,0

Mds1,0

N1,0

Nds1,0

Pvalue1,0+=

Mds1,0*Nds0,0

+Mds1,1*Nds1,0

M1,2

Mds1,0

N3,0

Nds1,0

Pvalue1,0+=

Mds1,0*Nds0,0

+Mds1,1*Nds1,0

thread1,1

M1,1

Mds1,1

N1,1

Nds1,1

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

M1,3

Mds1,1

N3,1

Nds1,1

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

Phase3

Phase4

thread0,0

M0,4

Mds0,0

N4,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

M0,6

Mds0,0

N6,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

thread0,1

M0,5

Mds0,1

N4,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

M0,7

Mds0,1

N6,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

thread1,0

M1,4

Mds1,0

N5,0

Nds1,0

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

M1,6

Mds1,0

N7,0

Nds1,0

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

thread1,1

M1,5

Mds1,1

N5,1

Nds1,1

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

M1,7

Mds1,1

N7,1

Nds1,1

Pvalue1,1+=

Mds1,0*Nds0,1

+Mds1,1*Nds1,1

2.A 8*8matrix multiplication with 4*4tiling

Block0,0

Phase1

Phase2

thread0,0

M0,0

Mds0,0

N0,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

+Mds0,2*Nds2,0

+Mds0,3*Nds3,0

M0,4

Mds0,0

N4,0

Nds0,0

Pvalue0,0+=

Mds0,0*Nds0,0

+Mds0,1*Nds1,0

+Mds0,2*Nds2,0

+Mds0,3*Nds3,0

thread0,1

M0,1

Mds0,1

N0,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

+Mds0,2*Nds2,1

+Mds0,3*Nds3,1

M0,5

Mds0,1

N4,1

Nds0,1

Pvalue0,1+=

Mds0,0*Nds0,1

+Mds0,1*Nds1,1

+Mds0,2*Nds2,1

+Mds0,3*Nds3,1

thread0,2

M0,2

Mds0,2

N0,2

Nds0,2

Pvalue0,2+=

Mds0,0*Nds0,2

+Mds0,1*Nds1,2

+Mds0,2*Nds2,2

+Mds0,3*Nds3,2

M0,6

Mds0,2

N4,2

Nds0,2

Pvalue0,2+=

Mds0,0*Nds0,2

+Mds0,1*Nds1,2

+Mds0,2*Nds2,2

+Mds0,3*Nds3,2

thread0,3

M0,3

Mds0,3

N0,3

Nds0,3

Pvalue0,3+=

Mds0,0*Nds0,3

+Mds0,1*Nds1,3

+Mds0,2*Nds2,3

+Mds0,3*Nds3,3

M0,7

Mds0,3

N4,3

Nds0,3

Pvalue0,3+=

Mds0,0*Nds0,3

+Mds0,1*Nds1,3

+Mds0,2*Nds2,3

+Mds0,3*Nds3,3

Thread1.x-thread3.xellipsis

As shown in the tables,the reduction in global memory bandwidth is indeed proportional tothe dimension size of the tiles, cause the if the tile is bigger, thethread used is proportional bigger, the phase of read data fromglobal memory is proportional smaller, so the reduction in globalmemory bandwidth is proportional to the dimension size of the tiles.

5.3 What type of incorrectexecution behavior can happen if one forgot to use syncthreads() inthe kernel of Figure 5.12?

Answer: The barrier__syncthreads() in line 11 ensures that all threads have finishedloading the tiles of d_M and d_N into Mds and Nds before any of themcan move forward. The barrier __syncthread() in line 14 ensures thatall threads have finished using the d_M and d_N elements in theshared memory before any of them move on to the next iteration andload the elements in the next tiles. Without synthreads() in thekernel, the threads would load the elements too early and corrupt theinput values for other threads.

5.4 Assuming capacity was notan issue for register or shared memory, give one case that it wouldbe valuable to use shared memory instead of registers to hold valuesfetched from global memory?

Explain your answer?

Answer: Without concerningthe capacity of register or shared memory. The biggest differencebetween them is that a register is made for a single thread, butshared memory can be shared by all threads in one block.

So the matrixmultiplication maybe a good example because the data read by onethread may be useful to other threads.

5.5 For our tiledmatrix-matrix multiplication kernel, if we use a 32*32 tile, what isthe reduction of memory bandwidth usage for input matrices M andN?
a. 1/8 of the original usage

b. 1/16 of the originalusage

c. 1/32 of the originalusage

d. 1/64 of the originalusage

Answer: c

5.6 Assume that a kernel islaunched with 1000 tread blocks each of which has 512 threads. If avariable is declared as a local variable in the kernel, how manyversions of the variable will be created through the life time of theexecution of the kernel?

a.1

b.1000

c.512

d.512000

Answer: d

5.7 In the previous question,if a variable is declared as a shared memory variable, how manyversions of the variable will be created through the life time of theexecution of the kernel?

a.1

b.1000

c.512

d.51200

Answer: b

5.9 Consider performing amatrix multiplication of two input matrices with dimensions N*N. Howmany times is each element in the input matrices request form globalmemory when:

a. There is no tiling?

b. Tiles of size T*T areused?

Answer: a. N

b. N/T

《Programming Massively Parallel Processors》Chapter5 习题解答的更多相关文章

  1. Coursera公开课Functional Programming Principles in Scala习题解答:Week 2

    引言 OK.时间非常快又过去了一周.第一周有五一假期所以感觉时间绰绰有余,这周中间没有假期仅仅能靠晚上加周末的时间来消化,事实上还是有点紧张呢! 后来发现每堂课的视频还有相应的课件(Slide).字幕 ...

  2. Massively parallel supercomputer

    A novel massively parallel supercomputer of hundreds of teraOPS-scale includes node architectures ba ...

  3. (搬运)《算法导论》习题解答 Chapter 22.1-1(入度和出度)

    (搬运)<算法导论>习题解答 Chapter 22.1-1(入度和出度) 思路:遍历邻接列表即可; 伪代码: for u 属于 Vertex for v属于 Adj[u] outdegre ...

  4. DirectX 11游戏编程学习笔记之8: 第6章Drawing in Direct3D(在Direct3D中绘制)(习题解答)

            本文由哈利_蜘蛛侠原创,转载请注明出处.有问题欢迎联系2024958085@qq.com         注:我给的电子版是700多页,而实体书是800多页,所以我在提到相关概念的时候 ...

  5. 现代控制理论习题解答与Matlab程序示例

    现代控制理论习题解答与Matlab程序示例 现代控制理论 第三版 课后习题参考解答: http://download.csdn.net/detail/zhangrelay/9544934 下面给出部分 ...

  6. 【AI】Exponential Stochastic Cellular Automata for Massively Parallel Inference - 大规模并行推理的指数随机元胞自动机

    [论文标题]Exponential Stochastic Cellular Automata for Massively Parallel Inference     (19th-ICAIS,PMLR ...

  7. P4: Programming Protocol-Independent Packet Processors

    P4: Programming Protocol-Independent Packet Processors 摘要 P4是一门高级语言,用于编程与协议无关的数据包处理器.P4与SDN控制协议相关联,类 ...

  8. 機器學習基石(Machine Learning Foundations) 机器学习基石 作业三 课后习题解答

    今天和大家分享coursera-NTU-機器學習基石(Machine Learning Foundations)-作业三的习题解答.笔者在做这些题目时遇到非常多困难,当我在网上寻找答案时却找不到,而林 ...

  9. 《C++编程思想》第四章 初始化与清除(原书代码+习题+解答)

    相关代码: 1. #include <stdio.h> class tree { int height; public: tree(int initialHeight); ~tree(); ...

随机推荐

  1. codevs 2451 互不侵犯(状丫dp)

    /* 好神奇好神奇...表示自己要学的还很多 注意到n<=9 不是搜索就是状丫 搜索+剪枝 70分 枚举放或者不放 这里用状丫 f[i][j][k] 表示前i行 放了j个国王 i行的状态是k的方 ...

  2. 小学生之浅谈Struts2与struts1的运行机制

    Struts1工作原理图: 1.初始化:struts框架的总控制器ActionServlet是一个Servlet,它在web.xml中配置成自动启动的Servlet,在启动时总控制器会读取配置文件(s ...

  3. Highcharts 设置折线图的显示与隐藏

    1.初始化隐藏某条曲线 series 配置如: [ {name:"发帖", postCountData}, {name:"删帖帖", deleteCountDa ...

  4. CDN的全称是Content Delivery Network,即内容分发网络

    CDN的全称是Content Delivery Network,即内容分发网络 http://baike.baidu.com/link?url=Wd-IGGgslfJemdpuT3Y0BUi88RPQ ...

  5. SqlServer数据库设计,纠结的问题,有胆你就来!

    最近,小菜遇到了数据库设计上的一个问题,希望各位大侠再次,直抒己见,帮忙分析一下.也可有更好的设计,请直接提出来. 要求:数据库(SQL Server 2005)存储每个学生,单个客观题答案.分数,主 ...

  6. SQL Server 2008 R2 的版本和组件

    SQL Server 2008 R2 的版本和组件 SQL Server 2008 R2   其他版本 SQL Server 2008 SQL Server 2005 SQL Server 2012 ...

  7. OC之字符串 NSString与NSMutableString

    一.NSString 不可变字符串的操作1)将字符串常量对象直接赋值给字符串引用 NSString *str1=@"hello"; 字符串对象的输出格式:NSLog(@" ...

  8. Struts2的创建Web项目入门

    之前学过javaEE开发,学了之后便去做前端了,再也没碰过javaEE.现在重新学习和部署,入门级,使用MyEclipse10版本. 简单说一下过程. struts步骤: 1. web project ...

  9. Date与Calendar

    Date date=new Date();//现在时间 Date date1=new Date(1000);//格林威治时间1997/01/01开始算,后面的是毫秒 Calendar calendar ...

  10. CentOS 7 之Cisco Anyconnect Secure Mobility Client

    公司使用的是Cisco VPN, 于是准备使用一下.先登录公司的vpn页面,意料之中的失败,所以下载了vpnsetup.sh这个来手动安装. 手动是要用root的,不过由于我是个人学习使用机器,一直用 ...