Installation

PyNX supports python version 3.7 and above

You should follow Full installation in a python virtualenv, which can be used on any system which supports a python virtual environment. Installation has been tested on linux (e.g. debian >=8) systems as well as macOS computers. See below the notes regarding installation on windows computers.

If you are using an nVidia card and already have CUDA development tools, you can also use the quick install which only uses pip.

Generic Instructions

PyNX is focused on using Graphical Processing Units (GPU) for faster calculations, so you will need:

  • a GPU (which can be an integrated GPU)

  • an OpenCL installation (drivers and libraries)

  • and/or CUDA (which gives better performance), which requires CUDA drivers and development tools (nvcc)

PyNX should still work on a CPU only, but without any optimisation, and will therefore be very slow, especially for 3D CDI and Ptychography algorithms. Also, not all features are available in the CPU operators, so the CPU version is mostly intended for basic testing and educational purposes.

Quick installation

This allows to install PyNX quickly, assuming that you already have:

  • a python (>=3.7) installation

  • pip for python package installation

  • already-installed CUDA development tools (nvcc)

It is advised to use a python virtual environment (or a conda one)

You can install PyNX and the dependencies using the following commands:

pip install --upgrade pip
pip install setuptools wheel --upgrade
pip install numpy cython scipy matplotlib ipython notebook scikit-image ipywidgets ipympl
pip install h5py hdf5plugin h5glance silx fabio
curl -O http://ftp.esrf.fr/pub/scisoft/PyNX/pynx-latest.tar.bz2
pip install pynx-latest.tar.bz2[cuda]

Note you can remove the [cuda] if you do not want cuda but only OpenCL support, and you can use instead [cuda,gui] if you want also gui components, or even [cuda,gui,mpi] if you want to also have MPI support (if you have >=2 GPU).

The [cuda] option does not change anything to the PyNX installation, but only adds specific dependencies. So you can also install pycuda afterwards.

Once this has been run, you can test pynx.

Full installation in a python virtualenv

The following script should work on any POSIX (Linux, MacOS X) system, and requires:

  • a python (>=3.7) installation

  • pip for python package installation

  • GPU dependencies, for CUDA and/or OpenCL (at least one should be present, both can be used):

    • OpenCL libraries (out-of-the box on MacOS X, using nvidia/AMD drivers on Linux)

    • CUDA libraries and development tools

  • git, cmake and standard development tools (for C/C++/python, depending on the operating system)

The script can be found in the source code as ‘install_scripts/install-pynx-venv.sh’, or can be downloaded from http://ftp.esrf.fr/pub/scisoft/PyNX/install-scripts/install-pynx-venv.sh

Once this has been run, you can test pynx

Full installation in a conda environment

You can also install PyNX in a conda environment in the same way as a python virtual environment. You should preferably install all packages available through conda or conda-forge, before using pip to install PyNX.

You can use the installation script found in the source code as ‘install_scripts/install-pynx-conda.sh’,

Windows installation on a PC with an nVidia card

Windows is not officially supported, as it can be complex to get the correct environment tools to make PyNX work. You are therefore encouraged to use PyNX on a linux workstation or cluster node, or on a macOS machine.

Nevertheless the following was found to work on a PC with Windows 10 and an nVidia GTX 1080:

With all this PyNX runs correctly with CUDA in a notebook or ipython console.

Notes:

  • command-line scripts can also be used but without the ``.py`` extension, e.g. to test PynX you should run pynx-info and pynx-test from the anaconda console.

  • on-the-fly compilation of GPU kernels is significantly slower on windows than on linux and macOS, so when running a script or notebook, the first time the kernels are compiled (which happens transparently), the process may seem to hang. Once this has been done the optimisation will run at the same speed as on other platforms, depending on the GPU.

Development version

If you want to live on the wild side, you can install the (public) development version (updated nightly) using:

curl -O http://ftp.esrf.fr/pub/scisoft/PyNX/pynx-latest.tar.bz2
pip install pynx-devel-nightly.tar.bz2[cuda]

Testing the installation

Once installed, you can use:

pynx-info.py

which will tell you the version of pynx installed, and if you have support for cuda and/or opencl.

You can then test pynx from the console by using:

pynx-test

To also test live-plotting, you can run:

pynx-test live_plot
You can also run more specific tests using command-line keywords (combinations are possible):

  • pynx-test processing_unit : only run basic OpenCL and CUDA tests

  • pynx-test cdi : only run CDI tests

  • pynx-test cdi_runner : only run CDI runner tests

  • pynx-test ptycho : only run ptychography tests

  • pynx-test ptycho_runner : only run ptychography runner tests

  • pynx-test cuda : only run CUDA tests

  • pynx-test opencl : only run opencl tests

Dependencies

Requirements:

  • git, cmake and standard development tools (compilers, headers…)

  • Python packages (all installable using pip):

  • numpy, scipy, matplotlib

  • cython

  • scikit-image, scikit-learn

  • h5py hdf5plugin

  • silx fabio

  • psutil

  • pyvkfft (fft using cuda and opencl)

  • Recommended:

  • ipython, notebook

  • pyqt5, pyopengl (for silx viewer)

  • sphinx and nbsphinx to generate documentation

  • For OpenCL

  • pyopencl, mako

  • For CUDA:

  • CUDA development tools (nvcc)

  • pycuda

  • Optionally:

  • the cctbx library, if you want to use the pynx.scattering.gid module for grazing incidence scattering. This is a bit complex to install, so it should probably be installed first, before all other packages.

  • pandoc for sphinx documentation generation

  • scikit-cuda or gpyfft if you want to use those libraries rather than pyvkfft