基本思想：sections在OpenMP里面属于分迭代任务的分担，总感觉这玩意像FPGA硬件逻辑代码中的分段模块

(1)分段执行代码逻辑sections

#pragma omp parallel sections
    {
#pragma omp section
        for (int i = 0; i < num/2; i++) {
          .....
        }
#pragma omp section
        for (int i = num/2; i < num; i++) {
          ........
        }
}

测试代码

#include <iostream>
#include <omp.h>
#include<chrono>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
       // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {



#pragma omp parallel sections
    {
#pragma omp section
        for (int i = 0; i < num/2; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
#pragma omp section
        for (int i = num/2; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
}

}

int main() {


    int num=omp_get_num_procs();
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

测试结果，可以看到 两个for 循环会被分配到各个线程里面独立处理

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
i=1 the current thread id: 0
i=2 the current thread id: 0
i=3 the current thread id: 0
i=4 the current thread id: 0
i=5 the current thread id: 0
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
sequentialProgram elapse time: 0.0373026 seconds
i=0 the current thread id: 1
i=1 the current thread id: 1
i=2 the current thread id: 1
i=3 the current thread id: 1
i=4 the current thread id: 1
i=5 the current thread id: 1
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
parallelProgram elapse time: 0.0211751 seconds

Process finished with exit code 0

(2)可以同时并行处理两个任务sections

#pragma omp parallel
    {
#pragma omp sections
        {

#pragma omp section
            for (int i = 0; i < num/4; i++) {
               .....
            }
#pragma omp section
            for (int i =  num/4; i < num/2; i++) {
               .....
            }
        }
#pragma omp sections
        {

#pragma omp section
            for (int i = num/2; i < 3*num/4; i++) {
                .....
            }
#pragma omp section
            for (int i =  3*num/4; i < num; i++) {
                .....
            }
        }

}

测试代码

#include <iostream>
#include <omp.h>
#include<chrono>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
       // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {



#pragma omp parallel
    {
#pragma omp sections
        {

#pragma omp section
            for (int i = 0; i < num/4; i++) {
                //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
                printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
            }
#pragma omp section
            for (int i =  num/4; i < num/2; i++) {
                //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
                printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
            }
        }
#pragma omp sections
        {

#pragma omp section
            for (int i = num/2; i < 3*num/4; i++) {
                //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
                printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
            }
#pragma omp section
            for (int i =  3*num/4; i < num; i++) {
                //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
                printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
            }
        }

}

}

int main() {


    int num=omp_get_num_procs();
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

测试结果，可以看到各个线程是独立运行的，这里发现 各个section还是按照代码的先后顺序执行的&同时还需要让各个section的任务量尽量均衡~

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
i=1 the current thread id: 0
i=2 the current thread id: 0
i=3 the current thread id: 0
i=4 the current thread id: 0
i=5 the current thread id: 0
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
sequentialProgram elapse time: 0.0212041 seconds
i=0 the current thread id: 2
i=1 the current thread id: 2
i=2 the current thread id: 2
i=3 the current thread id: 8
i=4 the current thread id: 8
i=5 the current thread id: 8
i=6 the current thread id: 11
i=7 the current thread id: 11
i=8 the current thread id: 11
i=9 the current thread id: 9
i=10 the current thread id: 9
i=11 the current thread id: 9
parallelProgram elapse time: 0.0202144 seconds

Process finished with exit code 0

(3)单线程运行制导指令 single

#pragma omp parallel
    {
#pragma omp single
      ....
        for (int i = 0; i < num; i++) {
            ....
        }

    };

   
#pragma omp parallel
    {
#pragma omp single
        {
          ....
        }
#pragma omp single nowait
        {
          ....
        }
        for (int i = 0; i < num; i++) {
              ....
        }

    };

其中single保证，其限制的任务为单线程执行

测试代码

#include <iostream>
#include <omp.h>
#include<chrono>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
       // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {



#pragma omp parallel
    {
#pragma omp single
    printf("i am students the current thread id: %d\n",  omp_get_thread_num());
        for (int i = 0; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }

    };

    printf("--------------------\n");
#pragma omp parallel
    {
#pragma omp single
        {
        printf("i am students the current thread id: %d\n", omp_get_thread_num());
        }
#pragma omp single nowait
        {
        printf("i am college the current thread id: %d\n",  omp_get_thread_num());
        }
        for (int i = 0; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }

    };

}

int main() {


    int num=1;
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

其中nowait保证 该线程也是为独立的线程执行，但是并不用等待线程执行结束，别的线程向下执行就行

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
sequentialProgram elapse time: 0.0025314 seconds
i am students the current thread id: 1
i=0 the current thread id: 0
i=0 the current thread id: 4
i=0 the current thread id: 5
i=0 the current thread id: 7
i=0 the current thread id: 3
i=0 the current thread id: 8
i=0 the current thread id: 10
i=0 the current thread id: 9
i=0 the current thread id: 2
i=0 the current thread id: 6
i=0 the current thread id: 11
i=0 the current thread id: 1
--------------------
i am students the current thread id: 5
i am college the current thread id: 7
i=0 the current thread id: 7
i=0 the current thread id: 6
i=0 the current thread id: 9
i=0 the current thread id: 11
i=0 the current thread id: 1
i=0 the current thread id: 8
i=0 the current thread id: 0
i=0 the current thread id: 4
i=0 the current thread id: 10
i=0 the current thread id: 2
i=0 the current thread id: 5
i=0 the current thread id: 3
parallelProgram elapse time: 0.0525391 seconds

Process finished with exit code 0

执行结果中这两个线程号（nowait）

i am college the current thread id: 7
i=0 the current thread id: 7

另一种测试方法，在并行的for循环上、section段上都可以使用nowait，不别再等待同步点执行~

#include <iostream>
#include <omp.h>
#include<chrono>
#include<vector>
#include<thread>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
        // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {




#pragma omp parallel
    {
#pragma omp for nowait
        for (int i = 0; i < num/2; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
#pragma omp for
        for (int i = num/2; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
    };

}

int main() {


    int num=omp_get_num_procs();
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

测试结果中看出，两个并行for循环，将真正的不存在先后到关系，开始并行执行

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
i=1 the current thread id: 0
i=2 the current thread id: 0
i=3 the current thread id: 0
i=4 the current thread id: 0
i=5 the current thread id: 0
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
sequentialProgram elapse time: 0.0194814 seconds
A i=1 the current thread id: 1
B i=7 the current thread id: 1
A i=2 the current thread id: 2
B i=8 the current thread id: 2
A i=0 the current thread id: 0
B i=6 the current thread id: 0
A i=3 the current thread id: 3
B i=9 the current thread id: 3
A i=4 the current thread id: 4
A i=5 the current thread id: 5
B i=11 the current thread id: 5
B i=10 the current thread id: 4
parallelProgram elapse time: 0.0217952 seconds

Process finished with exit code 0

（4） 设置路障 等待前面的线程运行完成，才能往下运行 barrier

#pragma omp barrier
{
.....
}

测试代码

#include <iostream>
#include <omp.h>
#include<chrono>
#include<vector>
#include<thread>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
        // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {




#pragma omp parallel
    {
#pragma omp for nowait
        for (int i = 0; i < num/2; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
#pragma omp barrier
        {
#pragma omp for
        for (int i = num/2; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
        }
    };

}

int main() {


    int num=omp_get_num_procs();
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

测试结果

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
i=1 the current thread id: 0
i=2 the current thread id: 0
i=3 the current thread id: 0
i=4 the current thread id: 0
i=5 the current thread id: 0
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
sequentialProgram elapse time: 0.0240967 seconds
A i=1 the current thread id: 1
A i=2 the current thread id: 2
A i=4 the current thread id: 4
A i=3 the current thread id: 3
A i=5 the current thread id: 5
A i=0 the current thread id: 0
B i=7 the current thread id: 1
B i=9 the current thread id: 3
B i=11 the current thread id: 5
B i=8 the current thread id: 2
B i=10 the current thread id: 4
B i=6 the current thread id: 0
parallelProgram elapse time: 0.0256972 seconds

Process finished with exit code 0

（5）设置只有一个线程去执行任务

#pragma omp master
for (int i = 0; i < num; i++) {
 ......
}

测试代码

#include <iostream>
#include <omp.h>
#include<chrono>
#include<vector>
#include<thread>
using namespace std;
using namespace chrono;

void sequentialProgram(int num)
{

    for(int i=0;i<num;i++)
    {
        // std::cout<<"hello world"<<std::endl;
        printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
    }
}

void  parallelProgram(int num) {




#pragma omp parallel
    {
#pragma omp for
        for (int i = 0; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }

#pragma omp single
        for (int i = 0; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }

#pragma omp master
        for (int i = 0; i < num; i++) {
            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("C i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
#pragma omp for ordered
        for (int i = 0; i < num; i++) {

            //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
            printf("D i=%d the current thread id: %d\n", i, omp_get_thread_num());
        }
    };

}

int main() {


    int num=omp_get_num_procs();
    auto start_time=std::chrono::steady_clock::now();
    sequentialProgram(num);
    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

    start_time=std::chrono::steady_clock::now();
    parallelProgram(num);
    end_time=std::chrono::steady_clock::now();
    std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
    return 0;
}

测试结果  抛开single的功能点说明，single和master在某种意义很类似都是控制任务为单一线程去完成。

其中的order 可以保证线程依次顺序处理各个任务

#pragma omp for ordered
for (int i = 0; i < num; i++) {
 .......
 }

F:\OpenMP\cmake-build-debug\OpenMP.exe
i=0 the current thread id: 0
i=1 the current thread id: 0
i=2 the current thread id: 0
i=3 the current thread id: 0
i=4 the current thread id: 0
i=5 the current thread id: 0
i=6 the current thread id: 0
i=7 the current thread id: 0
i=8 the current thread id: 0
i=9 the current thread id: 0
i=10 the current thread id: 0
i=11 the current thread id: 0
sequentialProgram elapse time: 0.0232429 seconds
A i=1 the current thread id: 1
A i=2 the current thread id: 2
A i=3 the current thread id: 3
A i=5 the current thread id: 5
A i=11 the current thread id: 11
A i=0 the current thread id: 0
A i=4 the current thread id: 4
A i=8 the current thread id: 8
A i=6 the current thread id: 6
A i=7 the current thread id: 7
A i=9 the current thread id: 9
A i=10 the current thread id: 10
B i=0 the current thread id: 1
B i=1 the current thread id: 1
B i=2 the current thread id: 1
B i=3 the current thread id: 1
B i=4 the current thread id: 1
B i=5 the current thread id: 1
B i=6 the current thread id: 1
B i=7 the current thread id: 1
B i=8 the current thread id: 1
B i=9 the current thread id: 1
B i=10 the current thread id: 1
B i=11 the current thread id: 1
C i=0 the current thread id: 0
C i=1 the current thread id: 0
C i=2 the current thread id: 0
C i=3 the current thread id: 0
C i=4 the current thread id: 0
C i=5 the current thread id: 0
C i=6 the current thread id: 0
C i=7 the current thread id: 0
C i=8 the current thread id: 0
C i=9 the current thread id: 0
C i=10 the current thread id: 0
C i=11 the current thread id: 0
D i=0 the current thread id: 0
D i=8 the current thread id: 8
D i=2 the current thread id: 2
D i=1 the current thread id: 1
D i=9 the current thread id: 9
D i=3 the current thread id: 3
D i=5 the current thread id: 5
D i=4 the current thread id: 4
D i=11 the current thread id: 11
D i=7 the current thread id: 7
D i=10 the current thread id: 10
D i=6 the current thread id: 6
parallelProgram elapse time: 0.112705 seconds

Process finished with exit code 0

 测试ncnn 提供的yolov5.cpp源码 每次测试都不一样，好像大部分修改的快一点

https://github.com/Tencent/ncnn/blob/master/examples/yolov5.cpp

代码片段

// anchor setting from yolov5/models/yolov5s.yaml
    auto start_time=std::chrono::steady_clock::now();

//#pragma omp parallel sections firstprivate(ex)
    {
 //#pragma omp section
    // stride 8
    {
        ncnn::Mat out;
        ex.extract("output", out);

        ncnn::Mat anchors(6);

        anchors[0] = 10.f;
        anchors[1] = 13.f;
        anchors[2] = 16.f;
        anchors[3] = 30.f;
        anchors[4] = 33.f;
        anchors[5] = 23.f;


        generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects);
        proposals.insert(proposals.end(), objects.begin(), objects.end());
      printf("the current thread id: %d\n",omp_get_thread_num());

    }
 //#pragma omp  section
    // stride 16
    {
        ncnn::Mat out;
        ex.extract("781", out);

        ncnn::Mat anchors(6);
        anchors[0] = 30.f;
        anchors[1] = 61.f;
        anchors[2] = 62.f;
        anchors[3] = 45.f;
        anchors[4] = 59.f;
        anchors[5] = 119.f;


        generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects);
        proposals.insert(proposals.end(), objects.begin(), objects.end());
        printf("the current thread id: %d\n",omp_get_thread_num());

    }
 //#pragma omp section
    // stride 32
    {
        ncnn::Mat out;

        ex.extract("801", out);

        ncnn::Mat anchors(6);
        anchors[0] = 116.f;
        anchors[1] = 90.f;
        anchors[2] = 156.f;
        anchors[3] = 198.f;
        anchors[4] = 373.f;
        anchors[5] = 326.f;



        generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects);
        proposals.insert(proposals.end(), objects.begin(), objects.end());
        printf("the current thread id: %d\n",omp_get_thread_num());

    }
    }
    // sort all proposals by score from highest to lowest


    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"output elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

测试时间

the current thread id: 0
the current thread id: 0
the current thread id: 0
output elapse time: 0.278379 seconds
yolov5s elapse time: 0.338553 seconds
15 = 0.54197 at 256.27 15.57 826.90 x 603.65

修改代码

auto start_time=std::chrono::steady_clock::now();

    ncnn::Mat out0;
    ex.extract("output", out0);

    ncnn::Mat out1;
    ex.extract("781", out1);

    ncnn::Mat out2;

    ex.extract("801", out2);

#pragma omp parallel sections 
    {
#pragma omp section
        // stride 8
        {


            ncnn::Mat anchors(6);

            anchors[0] = 10.f;
            anchors[1] = 13.f;
            anchors[2] = 16.f;
            anchors[3] = 30.f;
            anchors[4] = 33.f;
            anchors[5] = 23.f;


            generate_proposals(anchors, 8, in_pad, out0, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
#pragma omp  section
        // stride 16
        {
            ncnn::Mat anchors(6);
            anchors[0] = 30.f;
            anchors[1] = 61.f;
            anchors[2] = 62.f;
            anchors[3] = 45.f;
            anchors[4] = 59.f;
            anchors[5] = 119.f;


            generate_proposals(anchors, 16, in_pad, out1, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
#pragma omp section
        // stride 32
        {


            ncnn::Mat anchors(6);
            anchors[0] = 116.f;
            anchors[1] = 90.f;
            anchors[2] = 156.f;
            anchors[3] = 198.f;
            anchors[4] = 373.f;
            anchors[5] = 326.f;



            generate_proposals(anchors, 32, in_pad, out2, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
    }
    // sort all proposals by score from highest to lowest


    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"output elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

测试时间 好像大部分比原来的快 ~

F:\window10_yolo5_mingw32\cmake-build-debug\window10_yolo5_mingw32.exe
the current thread id: 5
the current thread id: 3
the current thread id: 11
output elapse time: 0.244162 seconds
yolov5s elapse time: 0.303863 seconds
15 = 0.54197 at 256.27 15.57 826.90 x 603.65

Process finished with exit cod

这样修改 耗时比较长

 auto start_time=std::chrono::steady_clock::now();




#pragma omp parallel sections firstprivate(ex)
    {
#pragma omp section
        // stride 8
        {
            ncnn::Mat out;
            ex.extract("output", out);

            ncnn::Mat anchors(6);

            anchors[0] = 10.f;
            anchors[1] = 13.f;
            anchors[2] = 16.f;
            anchors[3] = 30.f;
            anchors[4] = 33.f;
            anchors[5] = 23.f;


            generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
#pragma omp  section
        // stride 16
        {
            ncnn::Mat out;
            ex.extract("781", out);

            ncnn::Mat anchors(6);
            anchors[0] = 30.f;
            anchors[1] = 61.f;
            anchors[2] = 62.f;
            anchors[3] = 45.f;
            anchors[4] = 59.f;
            anchors[5] = 119.f;


            generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
#pragma omp section
        // stride 32
        {

            ncnn::Mat out;
            ex.extract("801", out);

            ncnn::Mat anchors(6);
            anchors[0] = 116.f;
            anchors[1] = 90.f;
            anchors[2] = 156.f;
            anchors[3] = 198.f;
            anchors[4] = 373.f;
            anchors[5] = 326.f;



            generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects);
            proposals.insert(proposals.end(), objects.begin(), objects.end());
            printf("the current thread id: %d\n",omp_get_thread_num());

        }
    }
    // sort all proposals by score from highest to lowest


    auto end_time=std::chrono::steady_clock::now();
    std::cout<<"output elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

测试时间

F:\window10_yolo5_mingw32\cmake-build-debug\window10_yolo5_mingw32.exe
the current thread id: 1
the current thread id: 0
the current thread id: 7
output elapse time: 0.829006 seconds
yolov5s elapse time: 0.895948 seconds
15 = 0.54197 at 256.27 15.57 826.90 x 603.65