Parallel Surface Reconstuction

Your Environment

• Operating System and version: Windows 10
• Compiler: Microsoft (R) C/C++ Optimizing Compiler Version 19.00.23506
• PCL Version: 1.8.1

Expected Behavior

Running Poisson surface reconstruction on separate point clouds each in separate threads should works without crashing

Current Behavior

An Access Violation Exceptions is thrown.

Possible Solution

Don't use static memory:

set MEMORY_ALLOCATOR_BLOCK_SIZE to zero in the following files:

https://github.com/PointCloudLibrary/pcl/blob/abf77fffd5b05140ebeb4cb71c204ad7458ae7c5/surface/include/pcl/surface/impl/poisson.hpp#L55

&

https://github.com/PointCloudLibrary/pcl/blob/abf77fffd5b05140ebeb4cb71c204ad7458ae7c5/surface/include/pcl/surface/3rdparty/poisson4/multi_grid_octree_data.hpp#L41

To test the effect on performance I did a really basic test. Using a single point cloud, default Poisson Recon settings and running on a single thread, I measured how long it takes to perform the reconstruction. The results:

Debug build, static memory, duration: 42 483 037 μs
Debug build, dynamic memory, duration: 40 880 744 μs
Release build, static memory, duration: 4 066 210 μs
Release build, dynamic memory, duration: 4 177 797 μs

At first glace it seems there is minimal performance effect, so what is the rational of using static memory?

Code to Reproduce

#define PCL_NO_PRECOMPILE

#include <thread>
#include <vector>

#include <pcl/point_types.h>
#include <pcl/surface/poisson.h>
#include <pcl/io/ply_io.h>

pcl::PolygonMesh mesh_1;
pcl::PolygonMesh mesh_2;
pcl::PointCloud<pcl::PointNormal>::Ptr cloud_in_1(new pcl::PointCloud<pcl::PointNormal>());
pcl::PointCloud<pcl::PointNormal>::Ptr cloud_in_2(new pcl::PointCloud<pcl::PointNormal>());

void thread1Process()
{
	pcl::Poisson<pcl::PointNormal> surf_con;

	surf_con.setInputCloud(cloud_in_1);
	surf_con.performReconstruction(mesh_1);
}

int main(int argc, char* argv[])
{
	//------------------ LOAD FILES ------------------//
	std::string path_dir = "C:\\Projects\\Snugg\\2.Development\\4.Repo\\temp\\MeasurementFinder\\build\\000Yon\\temp_m_f.ply";

	pcl::PLYReader reader;
	reader.read(path_dir, *cloud_in_1);
	reader.read(path_dir, *cloud_in_2);

	std::thread t1(thread1Process);

	pcl::Poisson<pcl::PointNormal> surf_con;

	surf_con.setInputCloud(cloud_in_2);
	surf_con.performReconstruction(mesh_2);
	
	t1.join();

	pcl::io::savePLYFile("mesh_1.ply", mesh_1);
	pcl::io::savePLYFile("mesh_2.ply", mesh_2);
}

Start with the allocator definitions at surface/include/pcl/surface/impl/poisson.hpp:55 and surface/include/pcl/surface/3rdparty/poisson4/multi_grid_octree_data.hpp:41. Build and run the supplied two-thread reproduction on the stated PCL and Windows environment, then trace the access violation during concurrent Poisson reconstruction. Done means separate point clouds can be reconstructed concurrently without crashing, with relevant regression coverage if the repository provides a suitable test location.