2018-08-05 Cython, Pythran, nuitka, numba¶
I discovered than Pythran was now a possible backend for Cython: Pythran as a Numpy backend. On one benchmark, it increases the speed by 2. I recommend the reading of the following article: Optimizing your code with NumPy, Cython, pythran and numba which instigates the performance brought by the four following tool and gives some hints on how to write efficient code with these tools:
Pythran: it is C compiler for Python code as Cython is but optimized for numerical computation.
nuitka: another compiler
The first three tools converts Python into C code and then compiles it. The main issue comes from type inference as C is a strongly typed language and Python is not. The inference is not always possible and that explains why the generated code is faster if the user specifies the type of each variable used in the optimized section. numba relies on llvmlite. The first execution is slower because it includes a compilation based on types discovered along the first execution. If the function is called a second time with the same type, the program uses the optimized version of it.