xtensor is a C++ template tensor algebra library supporting numpy-style broadcasting and universal functions, aiming at feature parity with numpy with a native feel. In this talk we will present the highlights of the expression system. Then we will show how xtensor can be used to create numpy-aware Python extension modules with the xtensor-python project and the xtensor-cookiecutter template.
In this presentation, we will focus on the authoring numpy-aware Python extensions with xtensor.
First, we present a general overview of the xtensor expression system and its main features:
Then we show how the expression-based system can be used to symbolically manipulate array expressions based on other data structures, such as a numpy arrays thanks to Python's buffer protocol, but also a database or the file system.
The special case of the bindings with numpy is handled by the xtensor-python project which makes it extremely easy to author Python extension operating inplace on numpy arrays, with a modern STL-compliant API. We conclude the talk with a demonstration of the xtensor-cookiecutter template project for a Python extension defining a NumPy ufunc, and exposing other C++ functionalities.
The cookiecutter produces a complete project including a documentation skeleton, unit testing and all the python packaging boilerplate.