This project serves as a wrapper around ASTRA-toolbox for GPU acccelerated and custom written parallel C++ reconstraction algorithms. In both cases, we recommend using the GNU GCC compiler, any version above 4.8 that supports OpenMP is suitable for compiling these scripts. To install the C++ accelerated scripts:
Compiling the C++ scripts is straightforward assuming GCC is avaiable and all the thirdparty dependencies are cloned. Simply move the Utility folder and compile the core script.
cd cpu_3D/Utils/
make ctvlib
On MacOS: the build command requires passing of the -undefined dynamic_lookup flag. If you're compiling the C++ for any other operating system remove that flag from line. M1 and any future architectures requires the passing of an additional -mcpu=apple-m1 flag. You can append these flags in the CXXFLAGS variable in make.inc.
Compiling ctvlib will create a shared object file in the following format: (ctvlib.cpython-3X-___.so). You can now import ctvlib as a python module and initialize it as a class. Refer to the demo jupyter-notebook for more details.
To compile multi - nodal scripts compile the mpi_ctvlib library instead (assuming MPI is available).
For GPU accelerated forward and back-projection operators, we first will need to be in the ASTRA-toolbox folder in thirdparty/astra-toolbox. Tomo_tv uses a couple custom written C++ scripts to link the two packages together.
Note: boost is a requirement for astra. Run sudo apt-get install libboost-all-dev to install boost or compile it from source (https://www.boost.org/doc/libs/1_43_0/more/getting_started/unix-variants.html).
Complete the following steps to build Astra from source:
cd thirdparty/astra-toolbox/build/linux
./autogen.sh
./configure --with-cuda=/path/to/cuda --with-python --with-install-type=prefix --prefix=/path/to/astra
make all
To get your cuda path run: whereis cuda. You'll also need to specify the path where you would like to build the Astra libary. I'd recommend spefcifying the path to where tomo_TV is located. Example: /path/to/tomo_TV/thirdparty/astra.
After the configuration compeletes, add --prefix=/path/to/astra to the end of Line 508 in the Makefile where /path/to/astra is the path specified as the prefix in the above (e.g. /path/to/tomo_TV/thirdparty/astra).
Finish installing the package.
make install
This will install the Python package to /path/to/astra (e.g. /path/to/tomo_TV/thirdparty/astra)
Once Astra is available, now let's compile tomo_TV.
cd tomo_TV/gpu_3D/Utils
make shared_library
Before compiling astra_ctvlib we'll need to specify the paths for ASTRA and CUDA. Open make.inc and edit the following lines
ASTRA_LIB = -L /path/to/astra/lib -lasatra
CUDA = -I /path/to/cuda/include -L /path/to/cuda/lib64 -lcudart -lz
Complete the process by running:
make astra_ctvlib
This will create a shared object file in the following format: (astra_ctvlib.cpython-3X-___.so). The final step requires linking the shared libraries and adding the necessary modules to the python path. Refer to setup_tomo_tv.sh for a sample shell script.
Multi-GPU parallel scripts requires parallel HDF5 as an additional dependency.
To run the scripts enter : mpirun -n XX python3 dynamicTomo.py where XX is the number of cores or GPUs available. \
tomo_TV uses the Eigen library for C++ linear algebra and sparse computations (available as a thirdparty submodule). vc