CS217_Project

Point cloud registration is the process of finding a transformation to align two geometric point sets. A classic algorithm to achieve this goal is called Iterative Closest Point (ICP). Since the ICP algorithm can be sped up with parallel programming, in this project we will implement ICP with CUDA.

The program code is split into two parts, the ICP algorithm module and the visualization module. In the runtime, we first generate a random rotation and translation to transform the original point clouds, then shuffle the points and save the result to a csv file. Then the ICP module reads the files containing the original point cloud and the transform one. Then ICP does the calculation and saves the estimated transformation and recovered point cloud to two csv files. Finally, the visualization module reads files containing the original point cloud, the transformed one, and the recovered one, and visualizes all the three point clouds. If there is no GUI on the machine, the visualization module can be skipped.

Keyword: ICP, CUDA

1. Setup

Conda is recommended to config this repo, more details about Miniconda can be found here

1.1 Setup python evironment

conda create --name gpu_proj python=3.6
conda activate gpu_proj
pip install -r requirements.txt

1.2 Setup C++ evironment

To build this project in command line, cmake is required, https://cmake.org/download/, after downloading Cmake, don't forget to add its path to PATH variable.

mkdir build && cd build
cmake ..
make -j4

For Mac Users, CLion is recommended for debugging CPU part.

2. Demo

If you have a GUI, simply run

python demo.py

If you don't have a GUI, first we run the C++ module to get ICP result (Not recommended)

./build/registration

Then we copy the files and run the python code to visualize the result

3. Configure Workspace

Nsight Eclipse is the default and best IDE to debug CUDA applications as far as I know. It also works perfectly with nvprof to show the porfiling results. The following steps show how to import this project into Nsight Eclipse

3.1 Prepare Directory

cd .. && mkdir icp_build && cd icp_build
cmake -G"Eclipse CDT4 - Unix Makefiles" -D CMAKE_BUILD_TYPE=RelWithDebInfo -D USE_VDEC=0 ../icp/

Now start up Eclipse and do the following to import the project.

File→Import...
General→Existing Projects into Workspace
For the root directory enter the build root directory, not the source root Leave other options unchecked and click Finish

3.2 Compilation

First, compile the programs in icp_build

    make

Then you have to tell Eclipse what binary to run and what project to compile.

Go to Run → Run Configurations...
Double click on C/C++ Application to create a new configuration.
Browse to the NG executable (build/src/NightshadeNG) for the Application field.
Browse to select the NG project for the Project field.

3.3 Debugging

Set up a debugging configuration in Eclipse.
Be sure to manually add the actual source directory in the source tab as a local filesystem path.
Add source paths to shared library code also so you can step through those.

Then, enjoy the IDE.

4. Nvidia Profiling

It seems that by default, non-root user don't have permision to profile GPU. If you can't profile, please try the followings:

4.1 Unload Old modules

systemctl isolate multi-user # Stop the window manager.
sudo modprobe -r nvidia_uvm nvidia_drm nvidia
# modprobe -r nvidia_uvm nvidia_drm nvidia_modeset nvidia-vgpu-vfio nvidia # Unload dependent modules

4.2 Allow all users to profile

sudo modprobe nvidia NVreg_RestrictProfilingToAdminUsers=0

4.3 Restart Windows Manager(GUI)

systemctl isolate graphical

lynklody/CS217_Project