OpenCL运行测试

---

由于双目的分辨率会非常大，因此特来测试下OpenCL的运算性能（奈何手中无英伟达）
环境：MacOS x86_64 Darwin 17.7.0
CPU：Intel Core i5-5350U
GPU：Intel HD Graphics 6000
OpenCV：OpenCV 4.2
编译器：clang++

以下是选用一张4480x6720和一张640x480图像做测试的结果，原始数据输出：

测试结果（仅供参考）

---

1. 像素：4480x6720，循环次数：100次
   Size:4480x6720
   OpenCL time consume = 85.6954ms
   Size:4480x6720
   CPU time consume = 120.533ms
   Size:4480 x 6720
   OpenCL loops time consume = 1050.15 ms
   Size:4480x6720
   CPU loop time consume = 1058.4ms
   -------------Process Complete------------

2. 像素：4480x6720，循环次数：100次
   Size:4480x6720
   OpenCL time consume = 86.6929 ms
   Size:4480x6720
   CPU time consume = 118.128 ms
   Size:4480 x 6720
   OpenCL loops time consume = 1055.27 ms
   Size:4480x6720
   CPU loops time consume = 1065.48 ms
   -------------Process Complete------------

3. 像素4480x6720，循环次数：1000
   Size:4480x6720
   OpenCL time consume = 86.6043 ms
   Size:4480x6720
   CPU time consume = 119.366 ms
   Size:4480 x 6720
   OpenCL loops time consume = 97815.7 ms
   Size:4480x6720
   CPU loops time consume = 105973 ms
   -------------Process Complete------------

4. 像素：640x480，循环次数：100次
   Size:640x480
   OpenCL time consume = 3.01333 ms
   Size:640x480
   CPU time consume = 1.55183 ms
   Size:640 x 480
   OpenCL loops time consume = 109.044 ms
   Size:640x480
   CPU loops time consume = 84.6049 ms
   -------------Process Complete------------

5. 像素：640x480，循环次数：100次
   Size:640x480
   OpenCL time consume = 3.2602 ms
   Size:640x480
   CPU time consume = 1.29072 ms
   Size:640 x 480
   OpenCL loops time consume = 107.968 ms
   Size:640x480
   CPU loops time consume = 86.4707 ms
   -------------Process Complete------------

6. 像素：640x480，1000次
   Size:640x480
   OpenCL time consume = 3.23095 ms
   Size:640x480
   CPU time consume = 1.17594 ms
   Size:640 x 480
   OpenCL loops time consume = 1164.68 ms
   Size:640x480
   CPU loops time consume = 940.939 ms
   -------------Process Complete------------

7. 像素：640x480，10000次
   Size:640x480
   OpenCL time consume = 3.64011 ms
   Size:640x480
   CPU time consume = 1.2437 ms
   Size:640 x 480
   OpenCL loops time consume = 10745.1 ms
   Size:640x480
   CPU loops time consume = 8566.75 ms
   -------------Process Complete------------

表格总结

表1：

分辨率	CPU处理1次	OpenCL处理1次	CPU循环100次	OpenCL循环100次
640x480	1.55 ms	3.01 ms	84.60 ms	109.04 ms
640x480	1.29 ms	3.26 ms	86.47 ms	107.96 ms
4480x6720	120.53 ms	85.69 ms	1058.4 ms	1050.15 ms
4480x6720	118.12 ms	86.69 ms	1065.48 ms	1055.27 ms

表2：

分辨率	循环次数	CPU处理1次	OpenCL处理1次	CPU循环	OpenCL循环
640x480	100次	1.55 ms	3.01 ms	84.60 ms	109.04 ms
640x480	100次	1.29 ms	3.26 ms	86.47 ms	107.96 ms
640x480	1000次	1.17 ms	3.23 ms	940.93 ms	1164.68 ms
640x480	10000次	1.24 ms	3.64 ms	8566.75 ms	10745.10 ms
4480x6720	100次	120.53 ms	85.69 ms	1058.4 ms	1050.15 ms
4480x6720	100次	118.12 ms	86.69 ms	1065.48 ms	1055.27 ms
4480x6720	1000次	119.36 ms	86.60 ms	1059.73 s	978.15 s

图表分析

---

为此我专门做了如下图表：

处理一张图
循环处理

因为这个数据差距有些大，不好直观看出差距，那么我们不妨采用log10为底的坐标刻度：

对数1
对数2

初步分析

1. 在分辨率较小的情况下，使用CPU处理会快于GPU(OpenCL),分辨率较大则GPU占优如图表中显示。
2. 在分辨率大(>1080P)而且循环次数多（>1000）的情况下，使用GPU(OpenCL)明显会快出CPU,参见表2最后一组数据，1000次循环整整快出了81.58s。
   当然，由于平台不同，这个结论不一定具有普适性

结尾放上测试代码，简单的canny边缘检测

edge_test.cpp

#include <opencv2/opencv.hpp>
using namespace cv;
void opencl_process(std::string &filename); //处理一张图（OpenCL）
void cpu_process(std::string &filename);  //处理一张图（CPU）
void loops_opencl(std::string &filename,int &times); //循环处理（OpenCL）
void loops_cpu(std::string &filename,int &times); //循环处理（CPU）
int main(int argc, char** argv)
{
    std::string filename = 文件名;
    int times = 循环次数;
    opencl_process(filename);
    cpu_process(filename);
    loops_opencl(filename,times);
    loops_cpu(filename,times);
    std::cout <<"-------------Process Complete------------"<<std::endl;
    return 0;
}

void opencl_process(std::string &filename){
    double start = (double)getTickCount();
    UMat img, gray;
    // 复制，从Mat->UMat
    imread(filename, IMREAD_COLOR).copyTo(img);
    cvtColor(img, gray, COLOR_BGR2GRAY);
    GaussianBlur(gray, gray,Size(7, 7), 1.5);
    Canny(gray, gray, 0, 50);
    double time_consume = ((double)getTickCount() - start) / getTickFrequency();
    std::cout << "Size:" << gray.cols << "x" << gray.rows << std::endl;
    std::cout << "OpenCL time consume = " << time_consume * 100<< " ms" << std::endl;
}

void cpu_process(std::string &filename){
    double start = (double)getTickCount();
    Mat img, gray;
    imread(filename, IMREAD_COLOR).copyTo(img);
    cvtColor(img, gray, COLOR_BGR2GRAY);
    GaussianBlur(gray, gray,Size(7, 7), 1.5);
    Canny(gray, gray, 0, 50);
    double time_consume = ((double)getTickCount() - start) / getTickFrequency();
    std::cout << "Size:" << gray.cols << "x" << gray.rows << std::endl;
    std::cout << "CPU time consume = " << time_consume * 100<< " ms" << std::endl;
}

void loops_opencl(std::string &filename,int &times){
    double start = (double)getTickCount();
    UMat img, gray;
    for(int i=0;i<=times;i++){
        imread(filename, IMREAD_COLOR).copyTo(img);
        cvtColor(img, gray, COLOR_BGR2GRAY);
        GaussianBlur(gray, gray,Size(7, 7), 1.5);
        Canny(gray, gray, 0, 50);
    }
    double time_consume = ((double)getTickCount() - start) / getTickFrequency();
    std::cout << "Size:" << gray.cols << " x " << gray.rows << std::endl;
    std::cout << "OpenCL loops time consume = " << time_consume * 100<< " ms" << std::endl;

}

void loops_cpu(std::string &filename,int &times){
    double start = (double)getTickCount();
    Mat img, gray;
    for (int i =0; i < times; i++){
        imread(filename, IMREAD_COLOR).copyTo(img);
        cvtColor(img, gray, COLOR_BGR2GRAY);
        GaussianBlur(gray, gray,Size(7, 7), 1.5);
        Canny(gray, gray, 0, 50);
    }
    double time_consume = ((double)getTickCount() - start) / getTickFrequency();
    std::cout << "Size:" << gray.cols << "x" << gray.rows << std::endl;
    std::cout << "CPU loops time consume = " << time_consume * 100<< " ms" << std::endl;
}

更改文件名和循环次数即可。
编译命令参考：

clang++ -std=c++11 edge_test.cpp -o edge_test `pkg-config --cflags --libs opencv4`

注意：opencv 4，开启c++11编译选项。