本文共 16836 字，大约阅读时间需要 56 分钟。
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 #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
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.exei=0 the current thread id: 0i=1 the current thread id: 0i=2 the current thread id: 0i=3 the current thread id: 0i=4 the current thread id: 0i=5 the current thread id: 0i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0sequentialProgram elapse time: 0.0373026 secondsi=0 the current thread id: 1i=1 the current thread id: 1i=2 the current thread id: 1i=3 the current thread id: 1i=4 the current thread id: 1i=5 the current thread id: 1i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0parallelProgram elapse time: 0.0211751 secondsProcess 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 #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
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.exei=0 the current thread id: 0i=1 the current thread id: 0i=2 the current thread id: 0i=3 the current thread id: 0i=4 the current thread id: 0i=5 the current thread id: 0i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0sequentialProgram elapse time: 0.0212041 secondsi=0 the current thread id: 2i=1 the current thread id: 2i=2 the current thread id: 2i=3 the current thread id: 8i=4 the current thread id: 8i=5 the current thread id: 8i=6 the current thread id: 11i=7 the current thread id: 11i=8 the current thread id: 11i=9 the current thread id: 9i=10 the current thread id: 9i=11 the current thread id: 9parallelProgram elapse time: 0.0202144 secondsProcess 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 #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
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.exei=0 the current thread id: 0sequentialProgram elapse time: 0.0025314 secondsi am students the current thread id: 1i=0 the current thread id: 0i=0 the current thread id: 4i=0 the current thread id: 5i=0 the current thread id: 7i=0 the current thread id: 3i=0 the current thread id: 8i=0 the current thread id: 10i=0 the current thread id: 9i=0 the current thread id: 2i=0 the current thread id: 6i=0 the current thread id: 11i=0 the current thread id: 1--------------------i am students the current thread id: 5i am college the current thread id: 7i=0 the current thread id: 7i=0 the current thread id: 6i=0 the current thread id: 9i=0 the current thread id: 11i=0 the current thread id: 1i=0 the current thread id: 8i=0 the current thread id: 0i=0 the current thread id: 4i=0 the current thread id: 10i=0 the current thread id: 2i=0 the current thread id: 5i=0 the current thread id: 3parallelProgram elapse time: 0.0525391 secondsProcess finished with exit code 0
执行结果中这两个线程号（nowait）
i am college the current thread id: 7i=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 #include #include #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
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.exei=0 the current thread id: 0i=1 the current thread id: 0i=2 the current thread id: 0i=3 the current thread id: 0i=4 the current thread id: 0i=5 the current thread id: 0i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0sequentialProgram elapse time: 0.0194814 secondsA i=1 the current thread id: 1B i=7 the current thread id: 1A i=2 the current thread id: 2B i=8 the current thread id: 2A i=0 the current thread id: 0B i=6 the current thread id: 0A i=3 the current thread id: 3B i=9 the current thread id: 3A i=4 the current thread id: 4A i=5 the current thread id: 5B i=11 the current thread id: 5B i=10 the current thread id: 4parallelProgram elapse time: 0.0217952 secondsProcess 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 #include #include #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
测试结果
F:\OpenMP\cmake-build-debug\OpenMP.exei=0 the current thread id: 0i=1 the current thread id: 0i=2 the current thread id: 0i=3 the current thread id: 0i=4 the current thread id: 0i=5 the current thread id: 0i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0sequentialProgram elapse time: 0.0240967 secondsA i=1 the current thread id: 1A i=2 the current thread id: 2A i=4 the current thread id: 4A i=3 the current thread id: 3A i=5 the current thread id: 5A i=0 the current thread id: 0B i=7 the current thread id: 1B i=9 the current thread id: 3B i=11 the current thread id: 5B i=8 the current thread id: 2B i=10 the current thread id: 4B i=6 the current thread id: 0parallelProgram elapse time: 0.0256972 secondsProcess finished with exit code 0
（5）设置只有一个线程去执行任务
#pragma omp masterfor (int i = 0; i < num; i++) { ......} 测试代码
#include #include #include #include #include using namespace std;using namespace chrono;void sequentialProgram(int num){ for(int i=0;i (end_time-start_time).count()<<" seconds"< (end_time-start_time).count()<<" seconds"<
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 orderedfor (int i = 0; i < num; i++) { ....... } F:\OpenMP\cmake-build-debug\OpenMP.exei=0 the current thread id: 0i=1 the current thread id: 0i=2 the current thread id: 0i=3 the current thread id: 0i=4 the current thread id: 0i=5 the current thread id: 0i=6 the current thread id: 0i=7 the current thread id: 0i=8 the current thread id: 0i=9 the current thread id: 0i=10 the current thread id: 0i=11 the current thread id: 0sequentialProgram elapse time: 0.0232429 secondsA i=1 the current thread id: 1A i=2 the current thread id: 2A i=3 the current thread id: 3A i=5 the current thread id: 5A i=11 the current thread id: 11A i=0 the current thread id: 0A i=4 the current thread id: 4A i=8 the current thread id: 8A i=6 the current thread id: 6A i=7 the current thread id: 7A i=9 the current thread id: 9A i=10 the current thread id: 10B i=0 the current thread id: 1B i=1 the current thread id: 1B i=2 the current thread id: 1B i=3 the current thread id: 1B i=4 the current thread id: 1B i=5 the current thread id: 1B i=6 the current thread id: 1B i=7 the current thread id: 1B i=8 the current thread id: 1B i=9 the current thread id: 1B i=10 the current thread id: 1B i=11 the current thread id: 1C i=0 the current thread id: 0C i=1 the current thread id: 0C i=2 the current thread id: 0C i=3 the current thread id: 0C i=4 the current thread id: 0C i=5 the current thread id: 0C i=6 the current thread id: 0C i=7 the current thread id: 0C i=8 the current thread id: 0C i=9 the current thread id: 0C i=10 the current thread id: 0C i=11 the current thread id: 0D i=0 the current thread id: 0D i=8 the current thread id: 8D i=2 the current thread id: 2D i=1 the current thread id: 1D i=9 the current thread id: 9D i=3 the current thread id: 3D i=5 the current thread id: 5D i=4 the current thread id: 4D i=11 the current thread id: 11D i=7 the current thread id: 7D i=10 the current thread id: 10D i=6 the current thread id: 6parallelProgram elapse time: 0.112705 secondsProcess 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
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: "< (end_time-start_time).count()<<" seconds"< 测试时间
the current thread id: 0the current thread id: 0the current thread id: 0output elapse time: 0.278379 secondsyolov5s elapse time: 0.338553 seconds15 = 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: "< (end_time-start_time).count()<<" seconds"< 测试时间 好像大部分比原来的快 ~
F:\window10_yolo5_mingw32\cmake-build-debug\window10_yolo5_mingw32.exethe current thread id: 5the current thread id: 3the current thread id: 11output elapse time: 0.244162 secondsyolov5s elapse time: 0.303863 seconds15 = 0.54197 at 256.27 15.57 826.90 x 603.65Process 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: "< (end_time-start_time).count()<<" seconds"< 测试时间
F:\window10_yolo5_mingw32\cmake-build-debug\window10_yolo5_mingw32.exethe current thread id: 1the current thread id: 0the current thread id: 7output elapse time: 0.829006 secondsyolov5s elapse time: 0.895948 seconds15 = 0.54197 at 256.27 15.57 826.90 x 603.65
转载地址：http://dtyai.baihongyu.com/