Efficient Compilation of High Level Python Numerical Programs with Pythran

The Python language (5) has a rich ecosystem that now provides a full toolkit to carry out scientific experiments, from core scientific routines with the Numpy package(3, 4), to scientific packages with Scipy, plotting facilities with the Matplotlib package, enhanced terminal and notebooks with IPython. As a consequence, there has been a move from historical languages like Fortran to Python, as showcased by the success of the Scipy conference. As Python based scientific tools get widely used, the question of High performance Computing naturally arises, and it is the focus of many recent research. Indeed, although there is a great gain in productivity when using these tools, there is also a performance gap that needs to be filled. This extended abstract focuses on compilation techniques that are relevant for the optimization of high-level numerical kernels written in Python using the Numpy package, illus- trated on a simple kernel.