feat: improved data handling in the ogm. When using the full concaten…

…ated pointcloud the processing time decreases from 8ms to 4ms

Signed-off-by: Kenzo Lobos-Tsunekawa <[email protected]>

perception/autoware_probabilistic_occupancy_grid_map/include/autoware/probabilistic_occupancy_grid_map/costmap_2d/occupancy_grid_map_base.hpp

@@ -102,6 +102,8 @@ class OccupancyGridMapInterface : public nav2_costmap_2d::Costmap2D
 
   bool isCudaEnabled() const;
 
+  void setCudaStream(const cudaStream_t & stream);
+
   const autoware::cuda_utils::CudaUniquePtr<std::uint8_t[]> & getDeviceCostmap() const;
 
   void copyDeviceCostmapToHost() const;

perception/autoware_probabilistic_occupancy_grid_map/include/autoware/probabilistic_occupancy_grid_map/utils/cuda_pointcloud.hpp

@@ -24,31 +24,37 @@
 class CudaPointCloud2 : public sensor_msgs::msg::PointCloud2
 {
 public:
-  void fromROSMsgAsync(const sensor_msgs::msg::PointCloud2 & msg)
+  void fromROSMsgAsync(const sensor_msgs::msg::PointCloud2::ConstSharedPtr & msg_ptr)
   {
     // cSpell: ignore knzo25
     // NOTE(knzo25): replace this with the cuda blackboard later
-    header = msg.header;
-    fields = msg.fields;
-    height = msg.height;
-    width = msg.width;
-    is_bigendian = msg.is_bigendian;
-    point_step = msg.point_step;
-    row_step = msg.row_step;
-    is_dense = msg.is_dense;
-
-    if (!data || capacity_ < msg.data.size()) {
-      capacity_ = 1024 * (msg.data.size() / 1024 + 1);
+    cudaStreamSynchronize(stream);
+    internal_msg_ = msg_ptr;
+
+    header = msg_ptr->header;
+    fields = msg_ptr->fields;
+    height = msg_ptr->height;
+    width = msg_ptr->width;
+    is_bigendian = msg_ptr->is_bigendian;
+    point_step = msg_ptr->point_step;
+    row_step = msg_ptr->row_step;
+    is_dense = msg_ptr->is_dense;
+
+    if (!data || capacity_ < msg_ptr->data.size()) {
+      const int factor = 4096 * point_step;
+      capacity_ = factor * (msg_ptr->data.size() / factor + 1);
       data = autoware::cuda_utils::make_unique<std::uint8_t[]>(capacity_);
     }
 
-    cudaMemcpyAsync(data.get(), msg.data.data(), msg.data.size(), cudaMemcpyHostToDevice, stream);
+    cudaMemcpyAsync(
+      data.get(), msg_ptr->data.data(), msg_ptr->data.size(), cudaMemcpyHostToDevice, stream);
   }
 
   autoware::cuda_utils::CudaUniquePtr<std::uint8_t[]> data;
   cudaStream_t stream;
 
 private:
+  sensor_msgs::msg::PointCloud2::ConstSharedPtr internal_msg_;
   std::size_t capacity_{0};
 };
 

perception/autoware_probabilistic_occupancy_grid_map/lib/costmap_2d/occupancy_grid_map_base.cpp

@@ -90,7 +90,6 @@ OccupancyGridMapInterface::OccupancyGridMapInterface(
     const auto num_cells_x = this->getSizeInCellsX();
     const auto num_cells_y = this->getSizeInCellsY();
 
-    cudaStreamCreate(&stream_);
     device_costmap_ = autoware::cuda_utils::make_unique<std::uint8_t[]>(num_cells_x * num_cells_y);
     device_costmap_aux_ =
       autoware::cuda_utils::make_unique<std::uint8_t[]>(num_cells_x * num_cells_y);
@@ -136,8 +135,9 @@ void OccupancyGridMapInterface::updateOrigin(double new_origin_x, double new_ori
       device_costmap_.get(), lower_left_x, lower_left_y, size_x_, size_y_,
       device_costmap_aux_.get(), 0, 0, cell_size_x, cell_size_y, cell_size_x, cell_size_y, stream_);
 
-    cudaMemset(
-      device_costmap_.get(), cost_value::NO_INFORMATION, size_x_ * size_y_ * sizeof(std::uint8_t));
+    cudaMemsetAsync(
+      device_costmap_.get(), cost_value::NO_INFORMATION, size_x_ * size_y_ * sizeof(std::uint8_t),
+      stream_);
   } else {
     local_map = new unsigned char[cell_size_x * cell_size_y];
 
@@ -207,6 +207,13 @@ bool OccupancyGridMapInterface::isCudaEnabled() const
   return use_cuda_;
 }
 
+void OccupancyGridMapInterface::setCudaStream(const cudaStream_t & stream)
+{
+  if (isCudaEnabled()) {
+    stream_ = stream;
+  }
+}
+
 const autoware::cuda_utils::CudaUniquePtr<std::uint8_t[]> &
 OccupancyGridMapInterface::getDeviceCostmap() const
 {
@@ -215,9 +222,10 @@ OccupancyGridMapInterface::getDeviceCostmap() const
 
 void OccupancyGridMapInterface::copyDeviceCostmapToHost() const
 {
-  cudaMemcpy(
+  cudaMemcpyAsync(
     costmap_, device_costmap_.get(), getSizeInCellsX() * getSizeInCellsY() * sizeof(std::uint8_t),
-    cudaMemcpyDeviceToHost);
+    cudaMemcpyDeviceToHost, stream_);
+  cudaStreamSynchronize(stream_);
 }
 
 }  // namespace costmap_2d

perception/autoware_probabilistic_occupancy_grid_map/lib/costmap_2d/occupancy_grid_map_fixed.cpp

@@ -125,14 +125,12 @@ void OccupancyGridMapFixedBlindSpot::updateWithPointCloud(
   const std::size_t num_raw_points = raw_pointcloud.width * raw_pointcloud.height;
   float range_resolution_inv = 1.0 / map_res;
 
-  cudaStreamSynchronize(stream_);
-
   map_fixed::prepareTensorLaunch(
     reinterpret_cast<const float *>(raw_pointcloud.data.get()), num_raw_points,
     raw_pointcloud.point_step / sizeof(float), angle_bin_size, range_bin_size, min_height_,
     max_height_, min_angle_, angle_increment_inv_, range_resolution_inv, device_rotation_map_.get(),
     device_translation_map_.get(), device_rotation_scan_.get(), device_translation_scan_.get(),
-    raw_points_tensor_.get(), raw_pointcloud.stream);
+    raw_points_tensor_.get(), stream_);
 
   const std::size_t num_obstacle_points = obstacle_pointcloud.width * obstacle_pointcloud.height;
 
@@ -141,27 +139,23 @@ void OccupancyGridMapFixedBlindSpot::updateWithPointCloud(
     obstacle_pointcloud.point_step / sizeof(float), angle_bin_size, range_bin_size, min_height_,
     max_height_, min_angle_, angle_increment_inv_, range_resolution_inv, device_rotation_map_.get(),
     device_translation_map_.get(), device_rotation_scan_.get(), device_translation_scan_.get(),
-    obstacle_points_tensor_.get(), obstacle_pointcloud.stream);
+    obstacle_points_tensor_.get(), stream_);
 
   map_fixed::fillEmptySpaceLaunch(
     raw_points_tensor_.get(), angle_bin_size, range_bin_size, range_resolution_inv,
     scan_origin.position.x, scan_origin.position.y, origin_x_, origin_y_, num_cells_x, num_cells_y,
-    cost_value::FREE_SPACE, device_costmap_.get(), raw_pointcloud.stream);
-
-  cudaStreamSynchronize(obstacle_pointcloud.stream);
+    cost_value::FREE_SPACE, device_costmap_.get(), stream_);
 
   map_fixed::fillUnknownSpaceLaunch(
     raw_points_tensor_.get(), obstacle_points_tensor_.get(), distance_margin_, angle_bin_size,
     range_bin_size, range_resolution_inv, scan_origin.position.x, scan_origin.position.y, origin_x_,
     origin_y_, num_cells_x, num_cells_y, cost_value::FREE_SPACE, cost_value::NO_INFORMATION,
-    device_costmap_.get(), raw_pointcloud.stream);
+    device_costmap_.get(), stream_);
 
   map_fixed::fillObstaclesLaunch(
     obstacle_points_tensor_.get(), distance_margin_, angle_bin_size, range_bin_size,
     range_resolution_inv, origin_x_, origin_y_, num_cells_x, num_cells_y,
-    cost_value::LETHAL_OBSTACLE, device_costmap_.get(), raw_pointcloud.stream);
-
-  cudaStreamSynchronize(raw_pointcloud.stream);
+    cost_value::LETHAL_OBSTACLE, device_costmap_.get(), stream_);
 }
 
 void OccupancyGridMapFixedBlindSpot::initRosParam(rclcpp::Node & node)

perception/autoware_probabilistic_occupancy_grid_map/lib/updater/binary_bayes_filter_updater.cpp

@@ -56,9 +56,9 @@ void OccupancyGridMapBBFUpdater::initRosParam(rclcpp::Node & node)
     }
   }
 
-  cudaMemcpy(
+  cudaMemcpyAsync(
     device_probability_matrix_.get(), probability_matrix_vector.data(),
-    sizeof(float) * Index::NUM_STATES * Index::NUM_STATES, cudaMemcpyHostToDevice);
+    sizeof(float) * Index::NUM_STATES * Index::NUM_STATES, cudaMemcpyHostToDevice, stream_);
 }
 
 inline unsigned char OccupancyGridMapBBFUpdater::applyBBF(
@@ -103,12 +103,6 @@ bool OccupancyGridMapBBFUpdater::update(
       Index::NUM_STATES, Index::FREE, Index::OCCUPIED, cost_value::FREE_SPACE,
       cost_value::LETHAL_OBSTACLE, cost_value::NO_INFORMATION, v_ratio_,
       getSizeInCellsX() * getSizeInCellsY(), device_costmap_.get(), stream_);
-
-    cudaMemcpy(
-      costmap_, device_costmap_.get(), getSizeInCellsX() * getSizeInCellsY() * sizeof(std::uint8_t),
-      cudaMemcpyDeviceToHost);
-
-    cudaStreamSynchronize(stream_);
   } else {
     for (unsigned int x = 0; x < getSizeInCellsX(); x++) {
       for (unsigned int y = 0; y < getSizeInCellsY(); y++) {

perception/autoware_probabilistic_occupancy_grid_map/lib/updater/log_odds_bayes_filter_updater.cpp

@@ -73,12 +73,6 @@ bool OccupancyGridMapLOBFUpdater::update(
     applyLOBFLaunch(
       single_frame_occupancy_grid_map.getDeviceCostmap().get(), cost_value::NO_INFORMATION,
       getSizeInCellsX() * getSizeInCellsY(), device_costmap_.get(), stream_);
-
-    cudaMemcpy(
-      costmap_, device_costmap_.get(), getSizeInCellsX() * getSizeInCellsY() * sizeof(std::uint8_t),
-      cudaMemcpyDeviceToHost);
-
-    cudaStreamSynchronize(stream_);
   } else {
     for (unsigned int x = 0; x < getSizeInCellsX(); x++) {
       for (unsigned int y = 0; y < getSizeInCellsY(); y++) {

perception/autoware_probabilistic_occupancy_grid_map/src/pointcloud_based_occupancy_grid_map/pointcloud_based_occupancy_grid_map_node.cpp

@@ -71,15 +71,13 @@ PointcloudBasedOccupancyGridMapNode::PointcloudBasedOccupancyGridMapNode(
   const double map_resolution = this->declare_parameter<double>("map_resolution");
 
   /* Subscriber and publisher */
-  obstacle_pointcloud_sub_.subscribe(
-    this, "~/input/obstacle_pointcloud",
-    rclcpp::SensorDataQoS{}.keep_last(1).get_rmw_qos_profile());
-  raw_pointcloud_sub_.subscribe(
-    this, "~/input/raw_pointcloud", rclcpp::SensorDataQoS{}.keep_last(1).get_rmw_qos_profile());
-  sync_ptr_ = std::make_shared<Sync>(SyncPolicy(5), obstacle_pointcloud_sub_, raw_pointcloud_sub_);
-
-  sync_ptr_->registerCallback(
-    std::bind(&PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw, this, _1, _2));
+  obstacle_pointcloud_sub_ptr_ = this->create_subscription<PointCloud2>(
+    "~/input/obstacle_pointcloud", rclcpp::SensorDataQoS{}.keep_last(1),
+    std::bind(&PointcloudBasedOccupancyGridMapNode::obstaclePointcloudCallback, this, _1));
+  raw_pointcloud_sub_ptr_ = this->create_subscription<PointCloud2>(
+    "~/input/raw_pointcloud", rclcpp::SensorDataQoS{}.keep_last(1),
+    std::bind(&PointcloudBasedOccupancyGridMapNode::rawPointcloudCallback, this, _1));
+
   occupancy_grid_map_pub_ = create_publisher<OccupancyGrid>("~/output/occupancy_grid_map", 1);
 
   const std::string updater_type = this->declare_parameter<std::string>("updater_type");
@@ -94,7 +92,6 @@ PointcloudBasedOccupancyGridMapNode::PointcloudBasedOccupancyGridMapNode(
     occupancy_grid_map_updater_ptr_ = std::make_unique<OccupancyGridMapBBFUpdater>(
       true, map_length / map_resolution, map_length / map_resolution, map_resolution);
   }
-  occupancy_grid_map_updater_ptr_->initRosParam(*this);
 
   const std::string grid_map_type = this->declare_parameter<std::string>("grid_map_type");
 
@@ -118,7 +115,15 @@ PointcloudBasedOccupancyGridMapNode::PointcloudBasedOccupancyGridMapNode(
       occupancy_grid_map_updater_ptr_->getSizeInCellsY(),
       occupancy_grid_map_updater_ptr_->getResolution());
   }
+
+  cudaStreamCreateWithFlags(&stream_, cudaStreamNonBlocking);
+  raw_pointcloud_.stream = stream_;
+  obstacle_pointcloud_.stream = stream_;
+  occupancy_grid_map_ptr_->setCudaStream(stream_);
+  occupancy_grid_map_updater_ptr_->setCudaStream(stream_);
+
   occupancy_grid_map_ptr_->initRosParam(*this);
+  occupancy_grid_map_updater_ptr_->initRosParam(*this);
 
   // initialize debug tool
   {
@@ -140,14 +145,29 @@ PointcloudBasedOccupancyGridMapNode::PointcloudBasedOccupancyGridMapNode(
       time_keeper_ = std::make_shared<autoware::universe_utils::TimeKeeper>(time_keeper);
     }
   }
+}
+
+void PointcloudBasedOccupancyGridMapNode::obstaclePointcloudCallback(
+  const PointCloud2::ConstSharedPtr & input_obstacle_msg)
+{
+  obstacle_pointcloud_.fromROSMsgAsync(input_obstacle_msg);
 
-  cudaStreamCreate(&raw_pointcloud_.stream);
-  cudaStreamCreate(&obstacle_pointcloud_.stream);
+  if (obstacle_pointcloud_.header.stamp == raw_pointcloud_.header.stamp) {
+    onPointcloudWithObstacleAndRaw();
+  }
 }
 
-void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
-  const PointCloud2::ConstSharedPtr & input_obstacle_msg,
+void PointcloudBasedOccupancyGridMapNode::rawPointcloudCallback(
   const PointCloud2::ConstSharedPtr & input_raw_msg)
+{
+  raw_pointcloud_.fromROSMsgAsync(input_raw_msg);
+
+  if (obstacle_pointcloud_.header.stamp == raw_pointcloud_.header.stamp) {
+    onPointcloudWithObstacleAndRaw();
+  }
+}
+
+void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw()
 {
   std::unique_ptr<ScopedTimeTrack> st_ptr;
   if (time_keeper_) st_ptr = std::make_unique<ScopedTimeTrack>(__func__, *time_keeper_);
@@ -156,9 +176,6 @@ void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
     stop_watch_ptr_->toc("processing_time", true);
   }
 
-  raw_pointcloud_.fromROSMsgAsync(*input_raw_msg);
-  obstacle_pointcloud_.fromROSMsgAsync(*input_obstacle_msg);
-
   // if scan_origin_frame_ is "", replace it with raw_pointcloud_.header.frame_id
   if (scan_origin_frame_.empty()) {
     scan_origin_frame_ = raw_pointcloud_.header.frame_id;
@@ -172,7 +189,7 @@ void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
         return;
       }
     }
-    if (input_obstacle_msg->header.frame_id != base_link_frame_) {
+    if (obstacle_pointcloud_.header.frame_id != base_link_frame_) {
       if (!utils::transformPointcloudAsync(obstacle_pointcloud_, *tf2_, base_link_frame_)) {
         return;
       }
@@ -221,7 +238,7 @@ void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
 
     // publish
     occupancy_grid_map_pub_->publish(OccupancyGridMapToMsgPtr(
-      map_frame_, input_raw_msg->header.stamp, robot_pose.position.z,
+      map_frame_, raw_pointcloud_.header.stamp, robot_pose.position.z,
       *occupancy_grid_map_ptr_));  // (todo) robot_pose may be altered with gridmap_origin
   } else {
     std::unique_ptr<ScopedTimeTrack> inner_st_ptr;
@@ -231,10 +248,11 @@ void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
 
     // Update with bayes filter
     occupancy_grid_map_updater_ptr_->update(*occupancy_grid_map_ptr_);
+    occupancy_grid_map_updater_ptr_->copyDeviceCostmapToHost();
 
     // publish
     occupancy_grid_map_pub_->publish(OccupancyGridMapToMsgPtr(
-      map_frame_, input_raw_msg->header.stamp, robot_pose.position.z,
+      map_frame_, raw_pointcloud_.header.stamp, robot_pose.position.z,
       *occupancy_grid_map_updater_ptr_));
   }
 
@@ -244,7 +262,7 @@ void PointcloudBasedOccupancyGridMapNode::onPointcloudWithObstacleAndRaw(
     const double pipeline_latency_ms =
       std::chrono::duration<double, std::milli>(
         std::chrono::nanoseconds(
-          (this->get_clock()->now() - input_raw_msg->header.stamp).nanoseconds()))
+          (this->get_clock()->now() - raw_pointcloud_.header.stamp).nanoseconds()))
         .count();
     debug_publisher_ptr_->publish<autoware_internal_debug_msgs::msg::Float64Stamped>(
       "debug/cyclic_time_ms", cyclic_time_ms);

perception/autoware_probabilistic_occupancy_grid_map/src/pointcloud_based_occupancy_grid_map/pointcloud_based_occupancy_grid_map_node.hpp

@@ -30,10 +30,7 @@
 #include <sensor_msgs/msg/laser_scan.hpp>
 #include <sensor_msgs/point_cloud2_iterator.hpp>
 
-#include <message_filters/pass_through.h>
-#include <message_filters/subscriber.h>
-#include <message_filters/sync_policies/exact_time.h>
-#include <message_filters/synchronizer.h>
+#include <cuda_runtime.h>
 #include <tf2_ros/buffer.h>
 #include <tf2_ros/transform_listener.h>
 
@@ -59,30 +56,28 @@ class PointcloudBasedOccupancyGridMapNode : public rclcpp::Node
   explicit PointcloudBasedOccupancyGridMapNode(const rclcpp::NodeOptions & node_options);
 
 private:
-  void onPointcloudWithObstacleAndRaw(
-    const PointCloud2::ConstSharedPtr & input_obstacle_msg,
-    const PointCloud2::ConstSharedPtr & input_raw_msg);
+  void obstaclePointcloudCallback(const PointCloud2::ConstSharedPtr & input_obstacle_msg);
+  void rawPointcloudCallback(const PointCloud2::ConstSharedPtr & input_raw_msg);
+  void onPointcloudWithObstacleAndRaw();
+
   OccupancyGrid::UniquePtr OccupancyGridMapToMsgPtr(
     const std::string & frame_id, const Time & stamp, const float & robot_pose_z,
     const Costmap2D & occupancy_grid_map);
 
 private:
   rclcpp::Publisher<OccupancyGrid>::SharedPtr occupancy_grid_map_pub_;
-  message_filters::Subscriber<PointCloud2> obstacle_pointcloud_sub_;
-  message_filters::Subscriber<PointCloud2> raw_pointcloud_sub_;
+  rclcpp::Subscription<PointCloud2>::SharedPtr obstacle_pointcloud_sub_ptr_;
+  rclcpp::Subscription<PointCloud2>::SharedPtr raw_pointcloud_sub_ptr_;
   std::unique_ptr<autoware::universe_utils::StopWatch<std::chrono::milliseconds>> stop_watch_ptr_{};
   std::unique_ptr<autoware::universe_utils::DebugPublisher> debug_publisher_ptr_{};
 
   std::shared_ptr<Buffer> tf2_{std::make_shared<Buffer>(get_clock())};
   std::shared_ptr<TransformListener> tf2_listener_{std::make_shared<TransformListener>(*tf2_)};
 
-  using SyncPolicy = message_filters::sync_policies::ExactTime<PointCloud2, PointCloud2>;
-  using Sync = message_filters::Synchronizer<SyncPolicy>;
-  std::shared_ptr<Sync> sync_ptr_;
-
   std::unique_ptr<OccupancyGridMapInterface> occupancy_grid_map_ptr_;
   std::unique_ptr<OccupancyGridMapUpdaterInterface> occupancy_grid_map_updater_ptr_;
 
+  cudaStream_t stream_;
   CudaPointCloud2 raw_pointcloud_;
   CudaPointCloud2 obstacle_pointcloud_;