并行计算简介和多核CPU编程Demo

2006年是双核的普及年，双核处理器出货量开始超过单核处理器出货量；2006年的11月份Intel开始供货4核；AMD今年也将发布4核，并计划今年下半年发布8核；
按照Intel一个文档所说:"假定22纳米处理时帧上有一枚13毫米大小的处理器，其上有40亿个晶体管、48MB高速缓存，功耗为100W。利用如此数量的晶体管，我们可设计拥有12个较大内核、48个（多核）中型内核、或144个小型内核（许多个内核）的处理器。"
而且Intel已经开发完成了一款80核心处理器原型，速度达到每秒一万亿次浮点运算。

随着个人多核CPU的普及，充分利用多核CPU的性能优势摆在了众多开发人员的面前；
以前的CPU升级，很多时候软件性能都能够自动地获得相应提升，而面对多核CPU，免费的午餐没有了，开发人员必须手工的完成软件的并行化，以从爆炸性增长的CPU性能中获益；
(ps:我想，以后的CPU很可能会集成一些专门用途的核(很可能设计成比较通用的模式)，比如GPU的核、图象处理的核、向量运算的核、加解密编解码的核、FFT计算的核、物理计算的核、神经网络计算的核等等:D )

先来看一下单个CPU上的并行计算:
单CPU上常见的并行计算：多级流水线(提高CPU频率的利器)、超标量执行(多条流水线并同时发送多条指令)、乱序执行(指令重排)、单指令流多数据流SIMD、超长指令字处理器(依赖于编译器分析)等

并行计算简介
并行平台的通信模型: 共享数据(POSIX、windows线程、OpenMP)、消息交换(MPI、PVM)
并行算法模型: 数据并行模型、任务依赖图模型、工作池模型、管理者-工作者模型、消费者模型对于并行计算一个任务可能涉及到的问题： 任务分解、任务依赖关系、任务粒度分配、并发度、任务交互并行算法性能的常见度量值： 并行开销、加速比、效率(加速比/CPU数)、成本(并行运行时间*CPU数)

A:一个简单的计算Demo
演示中主要完成的工作在Sum0函数(工作本身没有什么意义，主要是消耗一些时间来代表需要做的工作:),然后分别用OpenMP工具(vc和icc编译器支持)和一个自己手工写的线程工具来并行化该函数，来看看多核优化后的效果； 我测试用的编译器是vc2005;CPU是双核的AMD64x2 4200+(2.37G)；内存2G双通道DDR2 677MHz；

原始代码如下:

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <math.h>

//一个简单的耗时任务
double Sum0(double* data,long data_count);

int main()
{
long data_count=200000;
double* data=new double[data_count];
long i;

//初始化测试数据
for (i=0;i<data_count;++i)
data[i]=(double)(rand()*(1.0/RAND_MAX));

const long test_count=200*2;//为了能够测量出代码执行的时间，让函数执行多次
double sumresult=0;
double runtime=(double)clock();
for( i=0; i<test_count; ++i )
{
sumresult+=Sum0(data,data_count);
}
runtime=((double)clock()-runtime)/CLOCKS_PER_SEC;
printf ("< Sum0 > ");
printf (" 最后结果 = %10.4f ",sumresult);
printf (" 执行时间(秒) = %f ",runtime);

delete [] data;
return 0;
}


double Sum0(double* data,long data_count)
{
double result=0;
for (long i=0;i<data_count;++i)
{
data[i]=(double)sin(cos(data[i]));
result+=data[i];
}
return result;
}

在我的电脑上运行输出如下:

< Sum0 >
最后结果 = 55590743.4039
执行时间(秒) = 6.156000

B:使用OpenMP来优化(并行化)Sum0函数

OpenMP是基于编译器命令的并行编程标准，使用的共享数据模型，现在可以用在C/C++、Fortan中；OpenMP命令提供了对并发、同步、数据读写的支持；

(需要在项目属性中打开多线程和OpenMP支持,并要在多核CPU上执行才可以看到多CPU并行的优势)
OpenMP的实现如下:

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <math.h>

//需要在项目属性中打开多线程和OpenMP支持
#include <omp.h>

//用OpenMP实现
double Sum_OpenMP(double* data,long data_count);

int main()
{
long data_count=200000;
double* data=new double[data_count];
long i;

//初始化测试数据
for (i=0;i<data_count;++i)
data[i]=(double)(rand()*(1.0/RAND_MAX));

const long test_count=200*2;//为了能够测量出代码执行的时间，让函数执行多次
double sumresult=0;
double runtime=(double)clock();
for( i=0; i<test_count; ++i )
{
sumresult+=Sum_OpenMP(data,data_count);
}
runtime=((double)clock()-runtime)/CLOCKS_PER_SEC;
printf ("< Sum_OpenMP > ");
printf (" 最后结果 = %10.4f ",sumresult);
printf (" 执行时间(秒) = %f ",runtime);

delete [] data;
return 0;
}

double Sum_OpenMP(double* data,long data_count)
{
double result=0;
#pragma omp parallel for schedule(static) reduction(+: result)
for (long i=0;i<data_count;++i)
{
data[i]=(double)sin(cos(data[i]));
result+=data[i];
}
return result;
}

Sum_OpenMP函数相对于Sum0函数只是增加了一句"#pragma omp parallel for schedule(static) reduction(+: result)" ; 它告诉编译器并行化下面的for循环，并将多个result变量值用+合并；(更多的OpenMP语法请参阅相关资料)；

程序运行输出如下:

< Sum_OpenMP >
最后结果 = 55590743.4039
执行时间(秒) = 3.078000

在我的双核电脑上，OpenMP优化的并行代码使程序速度提高了约100%！

C:利用多线程来并行化Sum0函数(使用了我的CWorkThreadPool多线程工具类,完整源代码在后面)

需要在项目属性中打开多线程支持； 多线程并行实现如下:

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <math.h>
#include <vector>
#include "WorkThreadPool.h" //使用CWorkThreadPool类

double Sum_WorkThreadPool(double* data,long data_count);

int main()
{
long data_count=200000;
double* data=new double[data_count];
long i;

//初始化测试数据
for (i=0;i<data_count;++i)
data[i]=(double)(rand()*(1.0/RAND_MAX));

const long test_count=200*2;//为了能够测量出代码执行的时间，让函数执行多次
double sumresult=0;
double runtime=(double)clock();
for( i=0; i<test_count; ++i )
{
sumresult+=Sum_WorkThreadPool(data,data_count);
}
runtime=((double)clock()-runtime)/CLOCKS_PER_SEC;
printf ("< Sum_WorkThreadPool > ");
printf (" 最后结果 = %10.4f ",sumresult);
printf (" 执行时间(秒) = %f ",runtime);

delete [] data;
return 0;
}


double Sum0(double* data,long data_count)
{
double result=0;
for (long i=0;i<data_count;++i)
{
data[i]=(double)sin(cos(data[i]));
result+=data[i];
}
return result;
}

struct TWorkData
{
double* part_data;
long part_data_count;
double result;
};

void sum_callback(TWorkData* wd)
{
wd->result=Sum0(wd->part_data,wd->part_data_count);
}

double Sum_WorkThreadPool(double* data,long data_count)
{
long work_count=CWorkThreadPool::best_work_count();
std::vector<TWorkData> work_list(work_count);
std::vector<TWorkData*> pwork_list(work_count);
long i;

//给线程分配任务
long part_data_count=data_count/work_count;
for (i=0;i<work_count;++i)
{
work_list[i].part_data=&data[part_data_count*i];
work_list[i].part_data_count=part_data_count;
}
work_list[work_count-1].part_data_count=data_count-part_data_count*(work_count-1);
for (i=0;i<work_count;++i)
pwork_list[i]=&work_list[i];

//利用多个线程执行任务 阻塞方式的调用
CWorkThreadPool::work_execute((TThreadCallBack)sum_callback,(void**)&pwork_list[0],pwork_list.size());

double result=0;
for (i=0;i<work_count;++i)
result+=work_list[i].result;

return result;
}

用多线程来把代码并行化从而利用多个CPU核的计算能力,这种方式具有比OpenMP更好的灵活性；但容易看出这种方式没有OpenMP的实现简便； Sum_WorkThreadPool函数更多的代码在处理将计算任务分解成多个独立任务，然后将这些任务交给CWorkThreadPool执行； 程序执行输出如下:

< Sum_WorkThreadPool >
最后结果 = 55590743.4039
执行时间(秒) = 3.063000

在我的双核电脑上，多线程优化的并行代码使程序速度提高了约101%！

D: 附录: CWorkThreadPool类的完整源代码

(欢迎改进CWorkThreadPool类的代码，使它满足各种各样的并行需求)

//CWorkThreadPool的声明文件 WorkThreadPool.h

//WorkThreadPool.h
/////////////////////////////////////////////////////////////
//工作线程池 CWorkThreadPool
//用于把一个任务拆分成多个线程任务,从而可以使用多个CPU
//[e-mail]
////////////////////////////
//todo:改成任务领取模式
//要求：1.任务分割时分割的任务量比较接近
// 2.任务也不要太小，否则线程的开销可能会大于并行的收益
// 3.任务数最好是CPU数的倍数

#ifndef _WorkThreadPool_H_
#define _WorkThreadPool_H_

typedef void (*TThreadCallBack)(void * pData);

class CWorkThreadPool
{
public:
static long best_work_count(); //返回最佳工作分割数,现在的实现为返回CPU个数
static void work_execute(const TThreadCallBack work_proc,void** word_data_list,int work_count); //并行执行工作，并等待所有工作完成
static void work_execute_multi(const TThreadCallBack* work_proc_list,void** word_data_list,int work_count); //同上，但不同的work调用不同的函数
static void work_execute_single_thread(const TThreadCallBack work_proc,void** word_data_list,int work_count) //单线程执行工作，并等待所有工作完成;用于调试等
{
for (long i=0;i<work_count;++i)
work_proc(word_data_list[i]);
}
static void work_execute_single_thread_multi(const TThreadCallBack* work_proc_list,void** word_data_list,int work_count) //单线程执行工作，并等待所有工作完成;用于调试等
{
for (long i=0;i<work_count;++i)
work_proc_list[i](word_data_list[i]);
}
};

#endif //_WorkThreadPool_H_

//CWorkThreadPool的实现文件 WorkThreadPool.cpp

/////////////////////////////////////////////////////////////
//工作线程池 TWorkThreadPool

#include <process.h>
#include <vector>
#include "windows.h"
#include "WorkThreadPool.h"

//#define _IS_SetThreadAffinity_
//定义该标志则执行不同的线程绑定到不同的CPU，减少线程切换开销； 不鼓励


class TCriticalSection
{
private:
RTL_CRITICAL_SECTION m_data;
public:
TCriticalSection() { InitializeCriticalSection(&m_data); }
~TCriticalSection() { DeleteCriticalSection(&m_data); }
inline void Enter() { EnterCriticalSection(&m_data); }
inline void Leave() { LeaveCriticalSection(&m_data); }
};

class TWorkThreadPool;

//线程状态
enum TThreadState{ thrStartup=0, thrReady, thrBusy, thrTerminate, thrDeath };

class TWorkThread
{
public:
volatile HANDLE thread_handle;
volatile enum TThreadState state;
volatile TThreadCallBack func;
volatile void * pdata; //work data
TCriticalSection* CriticalSection;
TCriticalSection* CriticalSection_back;
TWorkThreadPool* pool;
volatile DWORD thread_ThreadAffinityMask;

TWorkThread() { memset(this,0,sizeof(TWorkThread)); }
};

void do_work_end(TWorkThread* thread_data);


void __cdecl thread_dowork(TWorkThread* thread_data) //void __stdcall thread_dowork(TWorkThread* thread_data)
{
volatile TThreadState& state=thread_data->state;
#ifdef _IS_SetThreadAffinity_
SetThreadAffinityMask(GetCurrentThread(),thread_data->thread_ThreadAffinityMask);
#endif
state = thrStartup;

while(true)
{
thread_data->CriticalSection->Enter();
thread_data->CriticalSection->Leave();
if(state == thrTerminate)
break;

state = thrBusy;
volatile TThreadCallBack& func=thread_data->func;
if (func!=0)
func((void *)thread_data->pdata);
do_work_end(thread_data);
}
state = thrDeath;
_endthread();
//ExitThread(0);
}

class TWorkThreadPool
{
private:
std::vector<TCriticalSection*> CriticalSections;
std::vector<TCriticalSection*> CriticalSections_back;
std::vector<TWorkThread> work_threads;
mutable long cpu_count;
inline long get_cpu_count() const {
if (cpu_count>0) return cpu_count;

SYSTEM_INFO SystemInfo;
GetSystemInfo(&SystemInfo);
cpu_count=SystemInfo.dwNumberOfProcessors;
return cpu_count;
}
inline long passel_count() const { return (long)work_threads.size()+1; }
void inti_threads()
{
long best_count =get_cpu_count();

long newthrcount=best_count - 1;
work_threads.resize(newthrcount);
CriticalSections.resize(newthrcount);
CriticalSections_back.resize(newthrcount);
long i;
for( i= 0; i < newthrcount; ++i)
{
CriticalSections[i]=new TCriticalSection();
CriticalSections_back[i]=new TCriticalSection();
work_threads[i].CriticalSection=CriticalSections[i];
work_threads[i].CriticalSection_back=CriticalSections_back[i];
CriticalSections[i]->Enter();
CriticalSections_back[i]->Enter();
work_threads[i].state = thrTerminate;
work_threads[i].pool=this;
work_threads[i].thread_ThreadAffinityMask=1<<(i+1);
work_threads[i].thread_handle =(HANDLE)_beginthread((void (__cdecl *)(void *))thread_dowork, 0, (void*)&work_threads[i]);
//CreateThread(0, 0, (LPTHREAD_START_ROUTINE)thread_dowork,(void*) &work_threads[i], 0, &thr_id);
//todo: _beginthread 的错误处理
}
#ifdef _IS_SetThreadAffinity_
SetThreadAffinityMask(GetCurrentThread(),0x01);
#endif
for(i = 0; i < newthrcount; ++i)
{
while(true) {
if (work_threads[i].state == thrStartup) break;
else Sleep(0);
}
work_threads[i].state = thrReady;
}
}
void free_threads(void)
{
long thr_count=(long)work_threads.size();
long i;
for(i = 0; i <thr_count; ++i)
{
while(true) {
if (work_threads[i].state == thrReady) break;
else Sleep(0);
}
work_threads[i].state=thrTerminate;
}
for (i=0;i<thr_count;++i)
{
CriticalSections[i]->Leave();
CriticalSections_back[i]->Leave();
}
for(i = 0; i <thr_count; ++i)
{
while(true) {
if (work_threads[i].state == thrDeath) break;
else Sleep(0);
}
}
work_threads.clear();
for (i=0;i<thr_count;++i)
{
delete CriticalSections[i];
delete CriticalSections_back[i];
}
CriticalSections.clear();
CriticalSections_back.clear();
}
void passel_work(const TThreadCallBack* work_proc,int work_proc_inc,void** word_data_list,int work_count) {
if (work_count==1)
(*work_proc)(word_data_list[0]);
else
{
const TThreadCallBack* pthwork_proc=work_proc;
pthwork_proc+=work_proc_inc;

long i;
long thr_count=(long)work_threads.size();
for(i = 0; i < work_count-1; ++i)
{
work_threads[i].func = *pthwork_proc;
work_threads[i].pdata =word_data_list[i+1];
work_threads[i].state = thrBusy;
pthwork_proc+=work_proc_inc;
}
for(i = work_count-1; i < thr_count; ++i)
{
work_threads[i].func = 0;
work_threads[i].pdata =0;
work_threads[i].state = thrBusy;
}
for (i=0;i<thr_count;++i)
CriticalSections[i]->Leave();

//current thread do a work
(*work_proc)(word_data_list[0]);

//wait for work finish
for(i = 0; i <thr_count; ++i)
{
while(true) {
if (work_threads[i].state == thrReady) break;
else Sleep(0);
}
}
CriticalSections.swap(CriticalSections_back);
for (i=0;i<thr_count;++i)
CriticalSections_back[i]->Enter();
}
}
void private_work_execute(TThreadCallBack* pwork_proc,int work_proc_inc,void** word_data_list,int work_count) {
while (work_count>0)
{
long passel_work_count;
if (work_count>=passel_count())
passel_work_count=passel_count();
else
passel_work_count=work_count;

passel_work(pwork_proc,work_proc_inc,word_data_list,passel_work_count);

pwork_proc+=(work_proc_inc*passel_work_count);
word_data_list=&word_data_list[passel_work_count];
work_count-=passel_work_count;
}
}
public:
explicit TWorkThreadPool():work_threads(),cpu_count(0) { inti_threads(); }
~TWorkThreadPool() { free_threads(); }
inline long best_work_count() const { return passel_count(); }
inline void DoWorkEnd(TWorkThread* thread_data){
thread_data->func=0;
thread_data->state = thrReady;
std::swap(thread_data->CriticalSection,thread_data->CriticalSection_back);
}

inline void work_execute_multi(TThreadCallBack* pwork_proc,void** word_data_list,int work_count) {
private_work_execute(pwork_proc,1,word_data_list,work_count);
}
inline void work_execute(TThreadCallBack work_proc,void** word_data_list,int work_count) {
private_work_execute(&work_proc,0,word_data_list,work_count);
}
};
void do_work_end(TWorkThread* thread_data)
{
thread_data->pool->DoWorkEnd(thread_data);
}

//TWorkThreadPool end;
////////////////////////////////////////

TWorkThreadPool g_work_thread_pool;//工作线程池

long CWorkThreadPool::best_work_count() { return g_work_thread_pool.best_work_count(); }

void CWorkThreadPool::work_execute(const TThreadCallBack work_proc,void** word_data_list,int work_count)
{
g_work_thread_pool.work_execute(work_proc,word_data_list,work_count);
}

void CWorkThreadPool::work_execute_multi(const TThreadCallBack* work_proc_list,void** word_data_list,int work_count)
{
g_work_thread_pool.work_execute_multi((TThreadCallBack*)work_proc_list,word_data_list,work_count);
}