CUDA Mesh projector

Projects triangular meshes with (ASTRA) cone and cone vector geometries.

License

The software is released under the GNU General Public License v3.0 - See the LICENSE file for more details.

How to use

Dependencies

NVIDIA display driver >= 495, cuda 10.x or 11.x, NVIDIA OptiX 7.4 (for ray tracing projectors), g++-8 if cuda 10.x is used, g++-9 if cuda 11.x is used, python 3.x, numpy, cython, trimesh (optional, for reading STL models), mayavi (optional, used in examples for 3D plotting)

Linux build

Python package

This will build the cad-astra package without OptiX projectors and install it into your current conda environment:

git clone git@github.com:cad-astra/cad-astra.git
cd cad-astra
./build_python.sh install

If you wish to preserve your conda environment unchanged, then simply build the package without install option and make it accessible via PYTHONPATH environment variable:

git clone git@github.com:cad-astra/cad-astra.git
cd cad-astra
./build_python.sh
cd build/python/lib*
export PYTHONPATH=$(pwd)

To build python package with OptiX projector, you need to specify OptiX 7.x root directory via OPTIX_DIR environment variable, and build with environment variable WITH_OPTIX=TRUE:

cd <OptiX 7 root>
export OPTIX_DIR=$(pwd)
cd <cad-astra directory>
WITH_OPTIX=TRUE ./build_python.sh install

To check other build options and environment variables that control the build, run

./build_python.sh --help

Usage

Import mesh_projector:

import mesh_projector as mp

Create cuda arrays for vertices, normals, faces, projection angles (views) and sinogram from numpy arrays:

vert_id = mp.create_cuda_array('vertices', nd_array=vertices)
face_id = mp.create_cuda_array('faces', nd_array=faces)
norm_id = mp.create_cuda_array('normals', nd_array=normals)
ang_id = mp.create_cuda_array('angles', nd_array=angles)
det_rows, det_cols = 512, 512
sino_id = mp.create_cuda_array('sino', shape=(angles.shape[0], det_rows, det_cols))

Create projection geometry, e.g., 'cone' geometry:

det_size_x, det_size_y = 1, 1
proj_geom = mp.create_projection_geometry('cone', det_size_x, det_size_y, det_rows, det_cols, ang_id, 5000, 200)

Project:

mp.project(vert_id, face_id, norm_id, sino_id, proj_geom)

Copy sinogram to host memory and reshape it:

sino = mp.get_cuda_array(sino_id)
sino = sino.reshape((angles.shape[0], det_rows, det_cols))

See the examples directory for more examples.

Running CUDA profiler

nvprof requires root priviliges for its work. To run it, create a bash script that starts nvprof and run it with sudo. Example of nvprof invocation script:

#! /bin/bash -x

nvprof --track-memory-allocations on <script_that_runs_projector>.py

Running the projector as differentiable program

Whilst the main projectors are distributed with the copyleft license (GPLv3), the projectors implemented as differential programs are distributed as closed code. To obtain the code for academic or commercial use, please send an email to Jan Sijbers <Jan.Sijbers@uantwerpen.be> or Jan De Beenhouwer <jan.debeenhouwer@uantwerpen.be>.

How to cite CAD-ASTRA

If you use CAD-ASTRA in your research, please cite our paper describing the toolkit published in Optics Express.

• P. Paramonov, N. Francken, J. Renders, D. Iuso, T. Elberfeld, J. De Beenhouwer, and J. Sijbers, "CAD-ASTRA: a versatile and efficient mesh projector for X-ray tomography with the ASTRA-toolbox," Opt. Express 32, 3425-3439 (2024); DOI: https://doi.org/10.1364/OE.498194