Merge pull request #24 from Guidebolt/multi-threading

Multi-threading

CMakeLists.txt

@@ -4,14 +4,16 @@ project(robot_body_filter)
 set (CMAKE_MODULE_PATH "${CMAKE_MODULE_PATH};${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 include(DetectOptional)
 
+add_compile_options(-fopenmp)
+
 if(${ROBOT_BODY_FILTER_HAVE_CXX_OPTIONAL})
   # add_compile_definitions would be nicer, but isn't available in Stretch
   add_compile_options(-DROBOT_BODY_FILTER_USE_CXX_OPTIONAL=1)
 else()
   add_compile_options(-DROBOT_BODY_FILTER_USE_CXX_OPTIONAL=0)
 endif()
 
-set(THIS_PACKAGE_DEPS backward_ros urdf urdf_parser_plugin filters geometric_shapes laser_geometry moveit_core moveit_ros_perception rclcpp sensor_msgs std_srvs pcl_conversions tf2 tf2_ros tf2_eigen urdf visualization_msgs GLUT )
+set(THIS_PACKAGE_DEPS backward_ros urdf urdf_parser_plugin filters geometric_shapes laser_geometry moveit_core moveit_ros_perception rclcpp sensor_msgs std_srvs pcl_conversions tf2 tf2_ros tf2_eigen urdf visualization_msgs GLUT OpenMP)
 set(MESSAGE_DEPS geometry_msgs std_msgs)
 
 find_package(ament_cmake REQUIRED)

src/RayCastingShapeMask.cpp

@@ -1,5 +1,6 @@
 /* HACK HACK HACK */
 /* We want to subclass ShapeMask and use its private members. */
+#include <boost/asio.hpp>
 #include <rclcpp/logger.hpp>
 #include <rclcpp/logging.hpp>
 #include <sstream>  // has to be there, otherwise we encounter build problems
@@ -14,6 +15,8 @@
 
 #include <geometric_shapes/body_operations.h>
 
+#include <omp.h>
+
 // #include <ros/console.h>
 
 namespace robot_body_filter
@@ -240,7 +243,7 @@ void RayCastingShapeMask::maskContainmentAndShadows(
 
   // Cloud iterators are not incremented in the for loop, because of the pragma
   // Comment out below parallelization as it can result in very high CPU consumption
-  //#pragma omp parallel for schedule(dynamic)
+  #pragma #pragma omp parallel for schedule(dynamic) num_threads(8)
   for (size_t i = 0; i < np; ++i)
   {
     const Eigen::Vector3d pt(static_cast<double>(*(iter_x + i)),

src/utils/tf2_sensor_msgs.cpp

@@ -9,6 +9,8 @@
 
 #include <unordered_set>
 
+#include <boost/asio.hpp>
+
 namespace robot_body_filter {
 
 const static std::unordered_map<std::string, CloudChannelType> XYZ_CHANNELS({
@@ -48,18 +50,22 @@ void transformChannel(const sensor_msgs::msg::PointCloud2& cloudIn, sensor_msgs:
   CloudIter z_out(cloudOut, channelPrefix + "z");
 
   Eigen::Vector3f point;
+  size_t np = num_points(cloudIn);
+
   // the switch has to be outside the for loop for performance reasons
   switch (type)
   {
-    case CloudChannelType::POINT:
-      for (; x_in != x_in.end(); ++x_in, ++y_in, ++z_in, ++x_out, ++y_out, ++z_out)
-      {
-        point = t * Eigen::Vector3f(*x_in, *y_in, *z_in);  // apply the whole transform
-        *x_out = point.x();
-        *y_out = point.y();
-        *z_out = point.z();
+    case CloudChannelType::POINT: {
+      // High CPU usage comment out below line for single threaded operation
+      #pragma #pragma omp parallel for schedule(dynamic) num_threads(8)
+      for (size_t i = 0; i < np; ++i) {
+        point = t * Eigen::Vector3f(*(x_in + i), *(y_in + i), *(z_in + i));  // apply the whole transform
+        *(x_out + i) = point.x();
+        *(y_out + i) = point.y();
+        *(z_out + i) = point.z();
       }
       break;
+    }
     case CloudChannelType::DIRECTION:
       for (; x_out != x_out.end(); ++x_in, ++y_in, ++z_in, ++x_out, ++y_out, ++z_out)
       {
@@ -68,7 +74,6 @@ void transformChannel(const sensor_msgs::msg::PointCloud2& cloudIn, sensor_msgs:
         *y_out = point.y();
         *z_out = point.z();
       }
-      RCLCPP_ERROR(rclcpp::get_logger("robot_body_filter"), "break");
       break;
     case CloudChannelType::SCALAR:
     //TODO: ADD WARNING FOR NOT SUPPORTED