openmp 和 thread 性能实测

#include <stdio.h>
#include <iostream>
#include <thread>
#include<omp.h>
//#include <opencv2/core.hpp>
//#include <cv/cv_tools.h>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/core/version.hpp>
using namespace std;
using namespace cv;

#define TEST_BY_OPENMP

void childFunc(Mat& pic, Mat &dst)
{
    int w = pic.cols;
    int h = pic.rows;
    //do{
    int sz = 3 * w*h;
    for (int i = 0; i != sz; ++i) {
        //int src_data = pic.data[i];
        //float ss= 1.234556;
        //ss*=1.2;
        //src_data+=ss;
        //dst.data[i] = ((src_data >>5)<<5);
        dst.data[i] = ((pic.data[i] >> 5) << 5);
    }
    //}while(0);
}


vector<Mat> split(Mat& _src, int n) 
{
    int sz = _src.cols*_src.rows*3;
    vector<Mat> pics(n);
    for (int i = 0; i != n; ++i) {
        pics[i] = Mat(_src.rows/n,_src.cols,CV_8UC3, _src.data + i*sz/n );
    }
    return pics;
}
int main()
{
    const int LOOPS = 1e10;
    Mat src = imread("bg.jpg");
    resize(src,src,Size(3840,2160));
    Mat dst(src.size(),src.type());
    vector<Mat> child_pics   = split(src, 4);
    vector<Mat> child_pics_d = split(dst, 4);



    long long t0 = getTickCount();

    for(int k=0;k!=LOOPS;++k){
#ifdef TEST_BY_OPENMP
    #pragma omp parallel num_threads(4)
    {
        int index = omp_get_thread_num();
        childFunc(child_pics[index], child_pics_d[index]);
    }//7.3 ms   cpu=60%

    for (int i = 0; i != 4; ++i) {
        childFunc(child_pics[i], child_pics_d[i]);
    }//6.97ms  cpu=45%
#else
        thread t[4];
        for (int i = 0; i != 4; ++i) {
            t[i] = thread(childFunc, child_pics[i], child_pics_d[i]);
        }
        for (int i = 0; i != 4; ++i) {
            t[i].join();
        }//21.32ms  cpu=55%
#endif

    };

    long long t1 = getTickCount();
    double time_waste = double(t1 - t0) / getTickFrequency();
    time_waste /= LOOPS;
    printf("time waste=%.2f ms\n", time_waste * 1000);
    imshow("test", dst);
    waitKey(0);
}