[High Performance Computing] openMP

学习来源

链接

预备知识

parallel foris a combination of two instructions,This is generally used rather than simply usedfor指令

案例1

#include <iostream>
#include "omp.h"

int main()
{
    
    int tid,mcpu;
    tid=omp_get_thread_num(); //返回线程号
    mcpu=omp_get_num_threads(); //Returns the number of threads in the current parallel region
    printf("---before parallel---\n");
    printf("thread %d in %d cpus\n",tid,mcpu);

    printf("---during parallel---\n");
    #pragma omp parallel num_threads(3) private(tid,mcpu) //开启三个线程 And specify the two variables Each thread has its own private copy
    {
    
        tid=omp_get_thread_num();
        mcpu=omp_get_num_threads();
        printf("thread %d in %d cpus\n",tid,mcpu);
    }
    printf("---after parallel---\n");
    printf("thread %d in %d cpus\n",tid,mcpu);
    return 0;
}

OpenMPUse in parallel construction

常用指令

除此之外,还有一些要求,如：
forThe instruction is followed by the body of the loop;
只能是for循环不能是while循环;
for循环中的iMust be an integer,And the increment cannot be changed;
单入口单出口;
The variables inside the loop are public,但是i是私有的;
Only if there is no data correlationforcycle to usefor指令并行优化;

Parallelism starts from the outermost where possible,可以节约资源

The above is an example of data correlation

如果不使用collapse 我们只能对iThis layer loops in parallel,内部的j、k是串行的,The loop length becomes M;使用collapse(2),can be achieved fori,ktwo layers merged,The loop length becomes MxP

Declaration of private variables

这个变量的默认值是0,Does not inherit the value of external variables of the same name

如果是for循环,The value of the child thread's private variable of the last iteration is passed,如果是section的话,同理.

规约

规约：Often do some kind of operation on a variable,The result is also stored in the current variable



注意：
After each child thread ends,The value of the original reduction variable is indeterminate,It will be determined after all threads are executed.
The order in which threads are executed is indeterminate