devbox
AllinToyou
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

3、OpenMP的分段sections、single、nowait、barrier、master、order_pragma omp section

作者：AllinToyou | 2024-03-18 20:40:24

赞

踩

pragma omp section

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

  1. #pragma omp parallel sections
  2.     {
  3. #pragma omp section
  4.         for (int i = 0; i < num/2; i++) {
  5.           .....
  6.         }
  7. #pragma omp section
  8.         for (int i = num/2; i < num; i++) {
  9.           ........
  10.         }
  11. }

测试代码

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. using namespace std;
  5. using namespace chrono;
  6.  
  7. void sequentialProgram(int num)
  8. {
  9.  
  10.     for(int i=0;i<num;i++)
  11.     {
  12.        // std::cout<<"hello world"<<std::endl;
  13.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  14.     }
  15. }
  16.  
  17. void  parallelProgram(int num) {
  18.  
  19.  
  20.  
  21. #pragma omp parallel sections
  22.     {
  23. #pragma omp section
  24.         for (int i = 0; i < num/2; i++) {
  25.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  26.             printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  27.         }
  28. #pragma omp section
  29.         for (int i = num/2; i < num; i++) {
  30.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  31.             printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  32.         }
  33. }
  34.  
  35. }
  36.  
  37. int main() {
  38.  
  39.  
  40.     int num=omp_get_num_procs();
  41.     auto start_time=std::chrono::steady_clock::now();
  42.     sequentialProgram(num);
  43.     auto end_time=std::chrono::steady_clock::now();
  44.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  45.  
  46.     start_time=std::chrono::steady_clock::now();
  47.     parallelProgram(num);
  48.     end_time=std::chrono::steady_clock::now();
  49.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  50.     return 0;
  51. }

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

  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. i=1 the current thread id: 0
  4. i=2 the current thread id: 0
  5. i=3 the current thread id: 0
  6. i=4 the current thread id: 0
  7. i=5 the current thread id: 0
  8. i=6 the current thread id: 0
  9. i=7 the current thread id: 0
  10. i=8 the current thread id: 0
  11. i=9 the current thread id: 0
  12. i=10 the current thread id: 0
  13. i=11 the current thread id: 0
  14. sequentialProgram elapse time: 0.0373026 seconds
  15. i=0 the current thread id: 1
  16. i=1 the current thread id: 1
  17. i=2 the current thread id: 1
  18. i=3 the current thread id: 1
  19. i=4 the current thread id: 1
  20. i=5 the current thread id: 1
  21. i=6 the current thread id: 0
  22. i=7 the current thread id: 0
  23. i=8 the current thread id: 0
  24. i=9 the current thread id: 0
  25. i=10 the current thread id: 0
  26. i=11 the current thread id: 0
  27. parallelProgram elapse time: 0.0211751 seconds
  28.  
  29. Process finished with exit code 0

(2)可以同时并行处理两个任务sections

  1. #pragma omp parallel
  2.     {
  3. #pragma omp sections
  4.         {
  5.  
  6. #pragma omp section
  7.             for (int i = 0; i < num/4; i++) {
  8.                .....
  9.             }
  10. #pragma omp section
  11.             for (int i =  num/4; i < num/2; i++) {
  12.                .....
  13.             }
  14.         }
  15. #pragma omp sections
  16.         {
  17.  
  18. #pragma omp section
  19.             for (int i = num/2; i < 3*num/4; i++) {
  20.                 .....
  21.             }
  22. #pragma omp section
  23.             for (int i =  3*num/4; i < num; i++) {
  24.                 .....
  25.             }
  26.         }
  27.  
  28. }

测试代码

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. using namespace std;
  5. using namespace chrono;
  6.  
  7. void sequentialProgram(int num)
  8. {
  9.  
  10.     for(int i=0;i<num;i++)
  11.     {
  12.        // std::cout<<"hello world"<<std::endl;
  13.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  14.     }
  15. }
  16.  
  17. void  parallelProgram(int num) {
  18.  
  19.  
  20.  
  21. #pragma omp parallel
  22.     {
  23. #pragma omp sections
  24.         {
  25.  
  26. #pragma omp section
  27.             for (int i = 0; i < num/4; i++) {
  28.                 //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  29.                 printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  30.             }
  31. #pragma omp section
  32.             for (int i =  num/4; i < num/2; i++) {
  33.                 //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  34.                 printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  35.             }
  36.         }
  37. #pragma omp sections
  38.         {
  39.  
  40. #pragma omp section
  41.             for (int i = num/2; i < 3*num/4; i++) {
  42.                 //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  43.                 printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  44.             }
  45. #pragma omp section
  46.             for (int i =  3*num/4; i < num; i++) {
  47.                 //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  48.                 printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  49.             }
  50.         }
  51.  
  52. }
  53.  
  54. }
  55.  
  56. int main() {
  57.  
  58.  
  59.     int num=omp_get_num_procs();
  60.     auto start_time=std::chrono::steady_clock::now();
  61.     sequentialProgram(num);
  62.     auto end_time=std::chrono::steady_clock::now();
  63.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  64.  
  65.     start_time=std::chrono::steady_clock::now();
  66.     parallelProgram(num);
  67.     end_time=std::chrono::steady_clock::now();
  68.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  69.     return 0;
  70. }

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~

  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. i=1 the current thread id: 0
  4. i=2 the current thread id: 0
  5. i=3 the current thread id: 0
  6. i=4 the current thread id: 0
  7. i=5 the current thread id: 0
  8. i=6 the current thread id: 0
  9. i=7 the current thread id: 0
  10. i=8 the current thread id: 0
  11. i=9 the current thread id: 0
  12. i=10 the current thread id: 0
  13. i=11 the current thread id: 0
  14. sequentialProgram elapse time: 0.0212041 seconds
  15. i=0 the current thread id: 2
  16. i=1 the current thread id: 2
  17. i=2 the current thread id: 2
  18. i=3 the current thread id: 8
  19. i=4 the current thread id: 8
  20. i=5 the current thread id: 8
  21. i=6 the current thread id: 11
  22. i=7 the current thread id: 11
  23. i=8 the current thread id: 11
  24. i=9 the current thread id: 9
  25. i=10 the current thread id: 9
  26. i=11 the current thread id: 9
  27. parallelProgram elapse time: 0.0202144 seconds
  28.  
  29. Process finished with exit code 0

(3)单线程运行制导指令 single

  1. #pragma omp parallel
  2.     {
  3. #pragma omp single
  4.       ....
  5.         for (int i = 0; i < num; i++) {
  6.             ....
  7.         }
  8.  
  9.     };
  10.  
  11.    
  12. #pragma omp parallel
  13.     {
  14. #pragma omp single
  15.         {
  16.           ....
  17.         }
  18. #pragma omp single nowait
  19.         {
  20.           ....
  21.         }
  22.         for (int i = 0; i < num; i++) {
  23.               ....
  24.         }
  25.  
  26.     };

其中single保证，其限制的任务为单线程执行

测试代码

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. using namespace std;
  5. using namespace chrono;
  6.  
  7. void sequentialProgram(int num)
  8. {
  9.  
  10.     for(int i=0;i<num;i++)
  11.     {
  12.        // std::cout<<"hello world"<<std::endl;
  13.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  14.     }
  15. }
  16.  
  17. void  parallelProgram(int num) {
  18.  
  19.  
  20.  
  21. #pragma omp parallel
  22.     {
  23. #pragma omp single
  24.     printf("i am students the current thread id: %d\n",  omp_get_thread_num());
  25.         for (int i = 0; i < num; i++) {
  26.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  27.             printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  28.         }
  29.  
  30.     };
  31.  
  32.     printf("--------------------\n");
  33. #pragma omp parallel
  34.     {
  35. #pragma omp single
  36.         {
  37.         printf("i am students the current thread id: %d\n", omp_get_thread_num());
  38.         }
  39. #pragma omp single nowait
  40.         {
  41.         printf("i am college the current thread id: %d\n",  omp_get_thread_num());
  42.         }
  43.         for (int i = 0; i < num; i++) {
  44.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  45.             printf("i=%d the current thread id: %d\n", i, omp_get_thread_num());
  46.         }
  47.  
  48.     };
  49.  
  50. }
  51.  
  52. int main() {
  53.  
  54.  
  55.     int num=1;
  56.     auto start_time=std::chrono::steady_clock::now();
  57.     sequentialProgram(num);
  58.     auto end_time=std::chrono::steady_clock::now();
  59.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  60.  
  61.     start_time=std::chrono::steady_clock::now();
  62.     parallelProgram(num);
  63.     end_time=std::chrono::steady_clock::now();
  64.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  65.     return 0;
  66. }

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

  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. sequentialProgram elapse time: 0.0025314 seconds
  4. i am students the current thread id: 1
  5. i=0 the current thread id: 0
  6. i=0 the current thread id: 4
  7. i=0 the current thread id: 5
  8. i=0 the current thread id: 7
  9. i=0 the current thread id: 3
  10. i=0 the current thread id: 8
  11. i=0 the current thread id: 10
  12. i=0 the current thread id: 9
  13. i=0 the current thread id: 2
  14. i=0 the current thread id: 6
  15. i=0 the current thread id: 11
  16. i=0 the current thread id: 1
  17. --------------------
  18. i am students the current thread id: 5
  19. i am college the current thread id: 7
  20. i=0 the current thread id: 7
  21. i=0 the current thread id: 6
  22. i=0 the current thread id: 9
  23. i=0 the current thread id: 11
  24. i=0 the current thread id: 1
  25. i=0 the current thread id: 8
  26. i=0 the current thread id: 0
  27. i=0 the current thread id: 4
  28. i=0 the current thread id: 10
  29. i=0 the current thread id: 2
  30. i=0 the current thread id: 5
  31. i=0 the current thread id: 3
  32. parallelProgram elapse time: 0.0525391 seconds
  33.  
  34. Process finished with exit code 0

执行结果中这两个线程号（nowait）

  1. i am college the current thread id: 7
  2. 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~

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. #include<vector>
  5. #include<thread>
  6. using namespace std;
  7. using namespace chrono;
  8.  
  9. void sequentialProgram(int num)
  10. {
  11.  
  12.     for(int i=0;i<num;i++)
  13.     {
  14.         // std::cout<<"hello world"<<std::endl;
  15.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  16.     }
  17. }
  18.  
  19. void  parallelProgram(int num) {
  20.  
  21.  
  22.  
  23.  
  24. #pragma omp parallel
  25.     {
  26. #pragma omp for nowait
  27.         for (int i = 0; i < num/2; i++) {
  28.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  29.             printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
  30.         }
  31. #pragma omp for
  32.         for (int i = num/2; i < num; i++) {
  33.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  34.             printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
  35.         }
  36.     };
  37.  
  38. }
  39.  
  40. int main() {
  41.  
  42.  
  43.     int num=omp_get_num_procs();
  44.     auto start_time=std::chrono::steady_clock::now();
  45.     sequentialProgram(num);
  46.     auto end_time=std::chrono::steady_clock::now();
  47.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  48.  
  49.     start_time=std::chrono::steady_clock::now();
  50.     parallelProgram(num);
  51.     end_time=std::chrono::steady_clock::now();
  52.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  53.     return 0;
  54. }

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

  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. i=1 the current thread id: 0
  4. i=2 the current thread id: 0
  5. i=3 the current thread id: 0
  6. i=4 the current thread id: 0
  7. i=5 the current thread id: 0
  8. i=6 the current thread id: 0
  9. i=7 the current thread id: 0
  10. i=8 the current thread id: 0
  11. i=9 the current thread id: 0
  12. i=10 the current thread id: 0
  13. i=11 the current thread id: 0
  14. sequentialProgram elapse time: 0.0194814 seconds
  15. A i=1 the current thread id: 1
  16. B i=7 the current thread id: 1
  17. A i=2 the current thread id: 2
  18. B i=8 the current thread id: 2
  19. A i=0 the current thread id: 0
  20. B i=6 the current thread id: 0
  21. A i=3 the current thread id: 3
  22. B i=9 the current thread id: 3
  23. A i=4 the current thread id: 4
  24. A i=5 the current thread id: 5
  25. B i=11 the current thread id: 5
  26. B i=10 the current thread id: 4
  27. parallelProgram elapse time: 0.0217952 seconds
  28.  
  29. 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

  1. #pragma omp barrier
  2. {
  3. .....
  4. }

测试代码

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. #include<vector>
  5. #include<thread>
  6. using namespace std;
  7. using namespace chrono;
  8.  
  9. void sequentialProgram(int num)
  10. {
  11.  
  12.     for(int i=0;i<num;i++)
  13.     {
  14.         // std::cout<<"hello world"<<std::endl;
  15.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  16.     }
  17. }
  18.  
  19. void  parallelProgram(int num) {
  20.  
  21.  
  22.  
  23.  
  24. #pragma omp parallel
  25.     {
  26. #pragma omp for nowait
  27.         for (int i = 0; i < num/2; i++) {
  28.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  29.             printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
  30.         }
  31. #pragma omp barrier
  32.         {
  33. #pragma omp for
  34.         for (int i = num/2; i < num; i++) {
  35.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  36.             printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
  37.         }
  38.         }
  39.     };
  40.  
  41. }
  42.  
  43. int main() {
  44.  
  45.  
  46.     int num=omp_get_num_procs();
  47.     auto start_time=std::chrono::steady_clock::now();
  48.     sequentialProgram(num);
  49.     auto end_time=std::chrono::steady_clock::now();
  50.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  51.  
  52.     start_time=std::chrono::steady_clock::now();
  53.     parallelProgram(num);
  54.     end_time=std::chrono::steady_clock::now();
  55.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  56.     return 0;
  57. }

测试结果

  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. i=1 the current thread id: 0
  4. i=2 the current thread id: 0
  5. i=3 the current thread id: 0
  6. i=4 the current thread id: 0
  7. i=5 the current thread id: 0
  8. i=6 the current thread id: 0
  9. i=7 the current thread id: 0
  10. i=8 the current thread id: 0
  11. i=9 the current thread id: 0
  12. i=10 the current thread id: 0
  13. i=11 the current thread id: 0
  14. sequentialProgram elapse time: 0.0240967 seconds
  15. A i=1 the current thread id: 1
  16. A i=2 the current thread id: 2
  17. A i=4 the current thread id: 4
  18. A i=3 the current thread id: 3
  19. A i=5 the current thread id: 5
  20. A i=0 the current thread id: 0
  21. B i=7 the current thread id: 1
  22. B i=9 the current thread id: 3
  23. B i=11 the current thread id: 5
  24. B i=8 the current thread id: 2
  25. B i=10 the current thread id: 4
  26. B i=6 the current thread id: 0
  27. parallelProgram elapse time: 0.0256972 seconds
  28.  
  29. Process finished with exit code 0

（5）设置只有一个线程去执行任务

  1. #pragma omp master
  2. for (int i = 0; i < num; i++) {
  3.  ......
  4. }

测试代码

  1. #include <iostream>
  2. #include <omp.h>
  3. #include<chrono>
  4. #include<vector>
  5. #include<thread>
  6. using namespace std;
  7. using namespace chrono;
  8.  
  9. void sequentialProgram(int num)
  10. {
  11.  
  12.     for(int i=0;i<num;i++)
  13.     {
  14.         // std::cout<<"hello world"<<std::endl;
  15.         printf("i=%d the current thread id: %d\n",i,omp_get_thread_num());
  16.     }
  17. }
  18.  
  19. void  parallelProgram(int num) {
  20.  
  21.  
  22.  
  23.  
  24. #pragma omp parallel
  25.     {
  26. #pragma omp for
  27.         for (int i = 0; i < num; i++) {
  28.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  29.             printf("A i=%d the current thread id: %d\n", i, omp_get_thread_num());
  30.         }
  31.  
  32. #pragma omp single
  33.         for (int i = 0; i < num; i++) {
  34.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  35.             printf("B i=%d the current thread id: %d\n", i, omp_get_thread_num());
  36.         }
  37.  
  38. #pragma omp master
  39.         for (int i = 0; i < num; i++) {
  40.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  41.             printf("C i=%d the current thread id: %d\n", i, omp_get_thread_num());
  42.         }
  43. #pragma omp for ordered
  44.         for (int i = 0; i < num; i++) {
  45.  
  46.             //std::cout<<"hello world"<<"the current thread id: "<<omp_get_thread_num()<<std::endl;
  47.             printf("D i=%d the current thread id: %d\n", i, omp_get_thread_num());
  48.         }
  49.     };
  50.  
  51. }
  52.  
  53. int main() {
  54.  
  55.  
  56.     int num=omp_get_num_procs();
  57.     auto start_time=std::chrono::steady_clock::now();
  58.     sequentialProgram(num);
  59.     auto end_time=std::chrono::steady_clock::now();
  60.     std::cout<<"sequentialProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  61.  
  62.     start_time=std::chrono::steady_clock::now();
  63.     parallelProgram(num);
  64.     end_time=std::chrono::steady_clock::now();
  65.     std::cout<<"parallelProgram elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  66.     return 0;
  67. }

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

  1. #pragma omp for ordered
  2. for (int i = 0; i < num; i++) {
  3.  .......
  4.  }
  1. F:\OpenMP\cmake-build-debug\OpenMP.exe
  2. i=0 the current thread id: 0
  3. i=1 the current thread id: 0
  4. i=2 the current thread id: 0
  5. i=3 the current thread id: 0
  6. i=4 the current thread id: 0
  7. i=5 the current thread id: 0
  8. i=6 the current thread id: 0
  9. i=7 the current thread id: 0
  10. i=8 the current thread id: 0
  11. i=9 the current thread id: 0
  12. i=10 the current thread id: 0
  13. i=11 the current thread id: 0
  14. sequentialProgram elapse time: 0.0232429 seconds
  15. A i=1 the current thread id: 1
  16. A i=2 the current thread id: 2
  17. A i=3 the current thread id: 3
  18. A i=5 the current thread id: 5
  19. A i=11 the current thread id: 11
  20. A i=0 the current thread id: 0
  21. A i=4 the current thread id: 4
  22. A i=8 the current thread id: 8
  23. A i=6 the current thread id: 6
  24. A i=7 the current thread id: 7
  25. A i=9 the current thread id: 9
  26. A i=10 the current thread id: 10
  27. B i=0 the current thread id: 1
  28. B i=1 the current thread id: 1
  29. B i=2 the current thread id: 1
  30. B i=3 the current thread id: 1
  31. B i=4 the current thread id: 1
  32. B i=5 the current thread id: 1
  33. B i=6 the current thread id: 1
  34. B i=7 the current thread id: 1
  35. B i=8 the current thread id: 1
  36. B i=9 the current thread id: 1
  37. B i=10 the current thread id: 1
  38. B i=11 the current thread id: 1
  39. C i=0 the current thread id: 0
  40. C i=1 the current thread id: 0
  41. C i=2 the current thread id: 0
  42. C i=3 the current thread id: 0
  43. C i=4 the current thread id: 0
  44. C i=5 the current thread id: 0
  45. C i=6 the current thread id: 0
  46. C i=7 the current thread id: 0
  47. C i=8 the current thread id: 0
  48. C i=9 the current thread id: 0
  49. C i=10 the current thread id: 0
  50. C i=11 the current thread id: 0
  51. D i=0 the current thread id: 0
  52. D i=8 the current thread id: 8
  53. D i=2 the current thread id: 2
  54. D i=1 the current thread id: 1
  55. D i=9 the current thread id: 9
  56. D i=3 the current thread id: 3
  57. D i=5 the current thread id: 5
  58. D i=4 the current thread id: 4
  59. D i=11 the current thread id: 11
  60. D i=7 the current thread id: 7
  61. D i=10 the current thread id: 10
  62. D i=6 the current thread id: 6
  63. parallelProgram elapse time: 0.112705 seconds
  64.  
  65. 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

  1. // anchor setting from yolov5/models/yolov5s.yaml
  2.     auto start_time=std::chrono::steady_clock::now();
  3.  
  4. //#pragma omp parallel sections firstprivate(ex)
  5.     {
  6.  //#pragma omp section
  7.     // stride 8
  8.     {
  9.         ncnn::Mat out;
  10.         ex.extract("output", out);
  11.  
  12.         ncnn::Mat anchors(6);
  13.  
  14.         anchors[0] = 10.f;
  15.         anchors[1] = 13.f;
  16.         anchors[2] = 16.f;
  17.         anchors[3] = 30.f;
  18.         anchors[4] = 33.f;
  19.         anchors[5] = 23.f;
  20.  
  21.  
  22.         generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects);
  23.         proposals.insert(proposals.end(), objects.begin(), objects.end());
  24.       printf("the current thread id: %d\n",omp_get_thread_num());
  25.  
  26.     }
  27.  //#pragma omp  section
  28.     // stride 16
  29.     {
  30.         ncnn::Mat out;
  31.         ex.extract("781", out);
  32.  
  33.         ncnn::Mat anchors(6);
  34.         anchors[0] = 30.f;
  35.         anchors[1] = 61.f;
  36.         anchors[2] = 62.f;
  37.         anchors[3] = 45.f;
  38.         anchors[4] = 59.f;
  39.         anchors[5] = 119.f;
  40.  
  41.  
  42.         generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects);
  43.         proposals.insert(proposals.end(), objects.begin(), objects.end());
  44.         printf("the current thread id: %d\n",omp_get_thread_num());
  45.  
  46.     }
  47.  //#pragma omp section
  48.     // stride 32
  49.     {
  50.         ncnn::Mat out;
  51.  
  52.         ex.extract("801", out);
  53.  
  54.         ncnn::Mat anchors(6);
  55.         anchors[0] = 116.f;
  56.         anchors[1] = 90.f;
  57.         anchors[2] = 156.f;
  58.         anchors[3] = 198.f;
  59.         anchors[4] = 373.f;
  60.         anchors[5] = 326.f;
  61.  
  62.  
  63.  
  64.         generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects);
  65.         proposals.insert(proposals.end(), objects.begin(), objects.end());
  66.         printf("the current thread id: %d\n",omp_get_thread_num());
  67.  
  68.     }
  69.     }
  70.     // sort all proposals by score from highest to lowest
  71.  
  72.  
  73.     auto end_time=std::chrono::steady_clock::now();
  74.     std::cout<<"output elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;

测试时间

  1. the current thread id: 0
  2. the current thread id: 0
  3. the current thread id: 0
  4. output elapse time: 0.278379 seconds
  5. yolov5s elapse time: 0.338553 seconds
  6. 15 = 0.54197 at 256.27 15.57 826.90 x 603.65

修改代码

  1. auto start_time=std::chrono::steady_clock::now();
  2.  
  3.     ncnn::Mat out0;
  4.     ex.extract("output", out0);
  5.  
  6.     ncnn::Mat out1;
  7.     ex.extract("781", out1);
  8.  
  9.     ncnn::Mat out2;
  10.  
  11.     ex.extract("801", out2);
  12.  
  13. #pragma omp parallel sections 
  14.     {
  15. #pragma omp section
  16.         // stride 8
  17.         {
  18.  
  19.  
  20.             ncnn::Mat anchors(6);
  21.  
  22.             anchors[0] = 10.f;
  23.             anchors[1] = 13.f;
  24.             anchors[2] = 16.f;
  25.             anchors[3] = 30.f;
  26.             anchors[4] = 33.f;
  27.             anchors[5] = 23.f;
  28.  
  29.  
  30.             generate_proposals(anchors, 8, in_pad, out0, prob_threshold, objects);
  31.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  32.             printf("the current thread id: %d\n",omp_get_thread_num());
  33.  
  34.         }
  35. #pragma omp  section
  36.         // stride 16
  37.         {
  38.             ncnn::Mat anchors(6);
  39.             anchors[0] = 30.f;
  40.             anchors[1] = 61.f;
  41.             anchors[2] = 62.f;
  42.             anchors[3] = 45.f;
  43.             anchors[4] = 59.f;
  44.             anchors[5] = 119.f;
  45.  
  46.  
  47.             generate_proposals(anchors, 16, in_pad, out1, prob_threshold, objects);
  48.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  49.             printf("the current thread id: %d\n",omp_get_thread_num());
  50.  
  51.         }
  52. #pragma omp section
  53.         // stride 32
  54.         {
  55.  
  56.  
  57.             ncnn::Mat anchors(6);
  58.             anchors[0] = 116.f;
  59.             anchors[1] = 90.f;
  60.             anchors[2] = 156.f;
  61.             anchors[3] = 198.f;
  62.             anchors[4] = 373.f;
  63.             anchors[5] = 326.f;
  64.  
  65.  
  66.  
  67.             generate_proposals(anchors, 32, in_pad, out2, prob_threshold, objects);
  68.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  69.             printf("the current thread id: %d\n",omp_get_thread_num());
  70.  
  71.         }
  72.     }
  73.     // sort all proposals by score from highest to lowest
  74.  
  75.  
  76.     auto end_time=std::chrono::steady_clock::now();
  77.     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 ~

  1. F:\window10_yolo5_mingw32\cmake-build-debug\window10_yolo5_mingw32.exe
  2. the current thread id: 5
  3. the current thread id: 3
  4. the current thread id: 11
  5. output elapse time: 0.244162 seconds
  6. yolov5s elapse time: 0.303863 seconds
  7. 15 = 0.54197 at 256.27 15.57 826.90 x 603.65
  8.  
  9. Process finished with exit cod

这样修改 耗时比较长

  1.  auto start_time=std::chrono::steady_clock::now();
  2.  
  3.  
  4.  
  5.  
  6. #pragma omp parallel sections firstprivate(ex)
  7.     {
  8. #pragma omp section
  9.         // stride 8
  10.         {
  11.             ncnn::Mat out;
  12.             ex.extract("output", out);
  13.  
  14.             ncnn::Mat anchors(6);
  15.  
  16.             anchors[0] = 10.f;
  17.             anchors[1] = 13.f;
  18.             anchors[2] = 16.f;
  19.             anchors[3] = 30.f;
  20.             anchors[4] = 33.f;
  21.             anchors[5] = 23.f;
  22.  
  23.  
  24.             generate_proposals(anchors, 8, in_pad, out, prob_threshold, objects);
  25.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  26.             printf("the current thread id: %d\n",omp_get_thread_num());
  27.  
  28.         }
  29. #pragma omp  section
  30.         // stride 16
  31.         {
  32.             ncnn::Mat out;
  33.             ex.extract("781", out);
  34.  
  35.             ncnn::Mat anchors(6);
  36.             anchors[0] = 30.f;
  37.             anchors[1] = 61.f;
  38.             anchors[2] = 62.f;
  39.             anchors[3] = 45.f;
  40.             anchors[4] = 59.f;
  41.             anchors[5] = 119.f;
  42.  
  43.  
  44.             generate_proposals(anchors, 16, in_pad, out, prob_threshold, objects);
  45.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  46.             printf("the current thread id: %d\n",omp_get_thread_num());
  47.  
  48.         }
  49. #pragma omp section
  50.         // stride 32
  51.         {
  52.  
  53.             ncnn::Mat out;
  54.             ex.extract("801", out);
  55.  
  56.             ncnn::Mat anchors(6);
  57.             anchors[0] = 116.f;
  58.             anchors[1] = 90.f;
  59.             anchors[2] = 156.f;
  60.             anchors[3] = 198.f;
  61.             anchors[4] = 373.f;
  62.             anchors[5] = 326.f;
  63.  
  64.  
  65.  
  66.             generate_proposals(anchors, 32, in_pad, out, prob_threshold, objects);
  67.             proposals.insert(proposals.end(), objects.begin(), objects.end());
  68.             printf("the current thread id: %d\n",omp_get_thread_num());
  69.  
  70.         }
  71.     }
  72.     // sort all proposals by score from highest to lowest
  73.  
  74.  
  75.     auto end_time=std::chrono::steady_clock::now();
  76.     std::cout<<"output elapse time: "<<std::chrono::duration<double>(end_time-start_time).count()<<" seconds"<<std::endl;
  77.

测试时间

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

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/AllinToyou/article/detail/265594
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号