openMP 的多线程崩溃

Question

我的功能是对 17000 个点的点云（大约可以波动）进行排序，以提取其中的相关点并将它们存储在一个向量中。一切正常，但速度很慢。所以我正在尝试使用 openMp 来并行化任务，但我遇到了崩溃。

这是不工作的多线程版本：

  void IntervalMapEstimator::extract_relevant_points_multithread(std::vector<Point3D>& relevant_points, std::vector<Point3D>& pointcloud, doubleIE cell_min_angle_sensor_rot, doubleIE cell_max_angle_sensor_rot)
{

#pragma omp parallel for shared( pointcloud, cell_min_angle_sensor_rot, cell_max_angle_sensor_rot)
        for(int i = 0; i < pointcloud.size(); i++) {
            //int numThread = omp_get_thread_num();
            //std::cout << "numThread = " << numThread << std::endl;

            // Check whether the cell is between the 2nd and 3rd quadrant (--> e.g. -170 to 170°)
            if ( cell_min_angle_sensor_rot < 0 && cell_max_angle_sensor_rot >= 0 && abs(cell_min_angle_sensor_rot) > M_PI/2 && abs(cell_max_angle_sensor_rot) > M_PI/2) {
                // Point must be smaller than the minimum angle and bigger than the max angle (e.g. min-angle: -1.5 max-angle: 1.5 point angle bigger than 1.5 or smaller than -1.5)
                if ( pointcloud[i].pol_sensor_rot.phi <= cell_min_angle_sensor_rot || pointcloud[i].pol_sensor_rot.phi  >= cell_max_angle_sensor_rot ) {
                    relevant_points.push_back(pointcloud[i]);
                }

            } else {
                 if (pointcloud[i].pol_sensor_rot.phi  >= cell_min_angle_sensor_rot && pointcloud[i].pol_sensor_rot.phi  <= cell_max_angle_sensor_rot ) {
                     relevant_points.push_back(pointcloud[i]);
                 }
            }
        }

}

这里是我在输出中得到的响应：

7fcc93737000-7fcc9374b000 r-xp 00012000 103:05 7078283                   /lib/x86_64-linux-gnu/ld-2.23.so
7fcc938a3000-7fcc938f7000 rw-p 00000000 00:00 0 
7fcc9391e000-7fcc9392c000 rw-p 00000000 00:00 0 
7fcc93937000-7fcc93939000 rw-p 00000000 00:00 0 
7fcc93947000-7fcc9394a000 rw-p 00000000 00:00 0 
7fcc9394a000-7fcc9394b000 r--p 00025000 103:05 7078283                   /lib/x86_64-linux-gnu/ld-2.23.so
7fcc9394b000-7fcc9394c000 rw-p 00026000 103:05 7078283                   /lib/x86_64-linux-gnu/ld-2.23.so
7fcc9394c000-7fcc9394d000 rw-p 00000000 00:00 0 
7fff20b58000-7fff20b7a000 rw-p 00000000 00:00 0                          [stack]
7fff20bb8000-7fff20bbb000 r--p 00000000 00:00 0                          [vvar]
7fff20bbb000-7fff20bbd000 r-xp 00000000 00:00 0                          [vdso]
ffffffffff600000-ffffffffff601000 r-xp 00000000 00:00 0                  [vsyscall]
[interval_map-1] process has died [pid 12700, exit code -6, cmd /home/catkin_ws/SWC_INTERVAL_MAP/devel/lib/interval_map_test/interval_map_test __name:=interval_map __log:=/home/.ros/log/615acdf0-3714-11e8-bc07-9cebe84f847e/interval_map-1.log].
log file: /home/.ros/log/615acdf0-3714-11e8-bc07-9cebe84f847e/interval_map-1*.log
all processes on machine have died, roslaunch will exit
shutting down processing monitor...
... shutting down processing monitor complete
done

目前，我最好的猜测是，有一个等待列表用于查看点云向量内的值或相关点向量上的 push_back，它们越来越大，最终导致堆栈爆炸。 有人对这个问题有任何想法吗？

Answer 1

所以，问题是 openMp 的私有子句不适用于参考向量。 这不起作用（调用未定义的引用类型错误）：

#pragma omp parallel for private (relevant_points) 

我的解决方法是使用 boost 中的 lock_guard。

       void IntervalMapEstimator::extract_relevant_points_multithread(std::vector<Point3D>&  relevant_points ,std::vector<Point3D>& pointcloud, doubleIE cell_min_angle_sensor_rot, doubleIE cell_max_angle_sensor_rot)
    {

        boost::mutex mutex_relevant_points;
        omp_set_num_threads (5);

#pragma omp parallel for shared (relevant_points, pointcloud, cell_min_angle_sensor_rot, cell_max_angle_sensor_rot)
        for(int i = 0; i < pointcloud.size(); i++) {
            //int numThread = omp_get_thread_num();
            //std::cout << "numThread = " << numThread << std::endl;

            // Check whether the cell is between the 2nd and 3rd quadrant (--> e.g. -170 to 170°)
            if ( cell_min_angle_sensor_rot < 0 && cell_max_angle_sensor_rot >= 0 && abs(cell_min_angle_sensor_rot) > M_PI/2 && abs(cell_max_angle_sensor_rot) > M_PI/2) {
                // Point must be smaller than the minimum angle and bigger than the max angle (e.g. min-angle: -1.5 max-angle: 1.5 point angle bigger than 1.5 or smaller than -1.5)
                if ( pointcloud[i].pol_sensor_rot.phi <= cell_min_angle_sensor_rot || pointcloud[i].pol_sensor_rot.phi  >= cell_max_angle_sensor_rot ) {                   
                    boost::lock_guard<boost::mutex> lock(mutex_relevant_points);
                    relevant_points.push_back(pointcloud[i]);
                }
            } else {
                 if (pointcloud[i].pol_sensor_rot.phi  >= cell_min_angle_sensor_rot && pointcloud[i].pol_sensor_rot.phi  <= cell_max_angle_sensor_rot ) {
                    boost::lock_guard<boost::mutex> lock(mutex_relevant_points);
                    relevant_points.push_back(pointcloud[i]);
                 }
            }
        }
}

您还可以看到我设置了线程数，否则处理点的程序实际上占用了所有线程，而发送所有点的程序没有任何线程可以使用。 我希望它会对某人有所帮助。

Answer 2

实际上，您在这里尝试实现的目标与使用向量作为归约变量的归约非常相似。

如果我想并行化它，我将如何解决它：

void IntervalMapEstimator::extract_relevant_points_multithread( std::vector<Point3D>& relevant_points,
                                                                std::vector<Point3D>& pointcloud,
                                                                doubleIE cell_min_angle_sensor_rot,
                                                                doubleIE cell_max_angle_sensor_rot ) {
    #pragma omp parallel num_threads( 5 )
    {
        std::vector<Point3D> local_relevant_points;
        #pragma omp for
        for ( int i = 0; i < pointcloud.size(); i++ ) {
            // Check whether the cell is between the 2nd and 3rd quadrant (--> e.g. -170 to 170°)
            if ( cell_min_angle_sensor_rot < 0 && cell_max_angle_sensor_rot >= 0
               && abs( cell_min_angle_sensor_rot ) > M_PI/2 && abs( cell_max_angle_sensor_rot ) > M_PI/2 ) {
                // Point must be smaller than the minimum angle and bigger than the max angle
                // (e.g. min-angle: -1.5 max-angle: 1.5 point angle bigger than 1.5 or smaller than -1.5)
                if ( pointcloud[i].pol_sensor_rot.phi <= cell_min_angle_sensor_rot
                   || pointcloud[i].pol_sensor_rot.phi  >= cell_max_angle_sensor_rot ) {
                    local_relevant_points.push_back( pointcloud[i] );
                }
            }
            else {
                if ( pointcloud[i].pol_sensor_rot.phi  >= cell_min_angle_sensor_rot
                   && pointcloud[i].pol_sensor_rot.phi  <= cell_max_angle_sensor_rot ) {
                    local_relevant_points.push_back( pointcloud[i] );
                }
            }
        }
        auto Size = relevant_points.size();
        #pragma omp critical
        Size += local_relevant_points.size();
        #pragma omp barrier
        #pragma single
        relevant_points.reserve( Size );
        #pragma omp critical
        for ( int i = 0; i < local_relevant_points.size(); i++ ) {
            relevant_points.push_back( local_relevant_points[i] );
        }
    }
}

想法是并行进行局部处理，然后将局部结果累加到全局输出中。 我通过计算结果的预期大小添加了一个额外的改进，以便提前保留正确的大小以节省额外冗长的内存分配。

现在，有两个警告：

1. 我还没有编译它，所以那里可能有一些错字；
2. 我不确定并行化这会带来多大好处，因为整个过程可能太短且内存受限......但无论如何你都去了。

Answer 3

这是我的函数现在的样子：

void IntervalMapEstimator::extract_relevant_points_multithread(std::vector<Point3D>&  relevant_points ,std::vector<Point3D>& pointcloud, doubleIE cell_min_angle_sensor_rot, doubleIE cell_max_angle_sensor_rot)
{
    relevant_points.reserve (pointcloud.size ());

#pragma omp parallel for shared (relevant_points, pointcloud, cell_min_angle_sensor_rot, cell_max_angle_sensor_rot) num_threads(5)
        for(int i = 0; i < pointcloud.size(); i++) {
            //int numThread = omp_get_thread_num();
            //std::cout << "numThread = " << numThread << std::endl;

            // Check whether the cell is between the 2nd and 3rd quadrant (--> e.g. -170 to 170°)
            if ( cell_min_angle_sensor_rot < 0 && cell_max_angle_sensor_rot >= 0 && abs(cell_min_angle_sensor_rot) > M_PI/2 && abs(cell_max_angle_sensor_rot) > M_PI/2) {
                // Point must be smaller than the minimum angle and bigger than the max angle (e.g. min-angle: -1.5 max-angle: 1.5 point angle bigger than 1.5 or smaller than -1.5)
                if ( pointcloud[i].pol_sensor_rot.phi <= cell_min_angle_sensor_rot || pointcloud[i].pol_sensor_rot.phi  >= cell_max_angle_sensor_rot ) {                   
#pragma omp critical(push_in_relevant_points)
                    relevant_points.push_back(pointcloud[i]);
                }
            } else {
                 if (pointcloud[i].pol_sensor_rot.phi  >= cell_min_angle_sensor_rot && pointcloud[i].pol_sensor_rot.phi  <= cell_max_angle_sensor_rot ) {
#pragma omp critical(push_in_relevant_points)
                     relevant_points.push_back(pointcloud[i]);
                 }
            }
        }
}

我使用 mop critical 而不是 guard_lock 以获得更高的一致性，并且我在开始时做了一个保留。 因为在我的例子中，related_points 向量永远不会比点云大，而且大多数时候他只包含 50 到 100 个元素。而且我不再使用向量的副本，而是参考。