Merge pull request #5775 from mvieth/normal3d_dynamic_omp

NormalEstimationOMP: use dynamic scheduling for faster computation
PointCloudLibrary · Aug 7, 2023 · 97ef1b7 · 97ef1b7
2 parents de6d715 + be05749
commit 97ef1b7
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 3 deletions.
diff --git a/features/include/pcl/features/impl/normal_3d_omp.hpp b/features/include/pcl/features/impl/normal_3d_omp.hpp
@@ -77,7 +77,8 @@ pcl::NormalEstimationOMP<PointInT, PointOutT>::computeFeature (PointCloudOut &ou
   default(none) \
   shared(output) \
   firstprivate(nn_indices, nn_dists) \
-  num_threads(threads_)
+  num_threads(threads_) \
+  schedule(dynamic, chunk_size_)
     // Iterating over the entire index vector
     for (std::ptrdiff_t idx = 0; idx < static_cast<std::ptrdiff_t> (indices_->size ()); ++idx)
     {
@@ -106,7 +107,8 @@ pcl::NormalEstimationOMP<PointInT, PointOutT>::computeFeature (PointCloudOut &ou
   default(none) \
   shared(output) \
   firstprivate(nn_indices, nn_dists) \
-  num_threads(threads_)
+  num_threads(threads_) \
+  schedule(dynamic, chunk_size_)
     // Iterating over the entire index vector
     for (std::ptrdiff_t idx = 0; idx < static_cast<std::ptrdiff_t> (indices_->size ()); ++idx)
     {

diff --git a/features/include/pcl/features/normal_3d_omp.h b/features/include/pcl/features/normal_3d_omp.h
@@ -72,8 +72,9 @@ namespace pcl
     public:
       /** \brief Initialize the scheduler and set the number of threads to use.
         * \param nr_threads the number of hardware threads to use (0 sets the value back to automatic)
+        * \param chunk_size PCL will use dynamic scheduling with this chunk size. Setting it too low will lead to more parallelization overhead. Setting it too high will lead to a worse balancing between the threads.
         */
-      NormalEstimationOMP (unsigned int nr_threads = 0)
+      NormalEstimationOMP (unsigned int nr_threads = 0, int chunk_size = 256): chunk_size_(chunk_size)
       {
         feature_name_ = "NormalEstimationOMP";
 
@@ -90,6 +91,8 @@ namespace pcl
       /** \brief The number of threads the scheduler should use. */
       unsigned int threads_;
 
+      /** \brief Chunk size for (dynamic) scheduling. */
+      int chunk_size_;
     private:
       /** \brief Estimate normals for all points given in <setInputCloud (), setIndices ()> using the surface in
         * setSearchSurface () and the spatial locator in setSearchMethod ()