JNumPy: writing high-performance C extensions for Python in minutes

Requirements:

• Python >= 3.7

You can install the Python package jnumpy with the following command:

pip install julia-numpy.

Note that JNumPy will install julia in JNUMPY_HOME for you, if there is no Julia installation available.

1. create a Python package example, write and export julia functions in the file example/src/example.jl

   module example
   
   using TyPython
   using TyPython.CPython
   
   @export_py function mat_mul(a::AbstractArray, b::AbstractArray)::Array
       return a * b
   end
   
   function init()
       @export_pymodule _example begin
           jl_mat_mul = Pyfunc(mat_mul)
       end
   end
   
   # the following code is optional,
   # but makes Python code loading much faster since the second time.
   precompile(init, ())
   
   end

2. create example/Project.toml as follows:

   name = "example"  # this is required to find the julia's entry module
   
   [deps]
   # specify your julia dependencies here

3. initialize and import the julia functions at example/__init__.py:

   import jnumpy as np
   # you may call np.set_julia_mirror(server) to set the julia package server,
   # or leave the argument server empty to automatically search the nearest mirror.
   np.init_jl()
   np.init_project(__file__)
   
   from _example import jl_mat_mul
   
   __all__ = ['jl_mat_mul']

4. enjoy your Python extension package:

   This is the structure of your Python extension package:

   > ls -R
   example/:
       __init__.py  Project.toml  src
   
   example/src:
       example.jl

   This is how you use it:

   from example import jl_mat_mul
   x = np.array([[1,2],[3,4]])
   y = np.array([[4,3],[2,1]])
   jl_mat_mul(x, y)
   # array([[ 8,  5],
   #       [20, 13]])

• JNUMPY_HOME:

  The home directory for JNumPy-specific settings. The default value is ~/.jnumpy. In case that you don't have a julia executable, JNumPy installs julia into $JNUMPY_HOME using jill.py.

• TYPY_JL_EXE:

  The path of the julia executable in use.

• TYPY_JL_OPTS:

  Command-line options when launching julia. JNumPy runs julia in a default environment (path/to/package/jnumpy/JNumPyEnv). If you want to use a custom environment, you could set --project=<dir>. TYPY_JL_OPTS is the same as those arguments passed to julia.

There are several examples presented in the demo directory. Those examples are standalone Python packages created using JNumPy, and can be imported if you have JNumPy installed.

• demo/basic: a tiny Python package to give an example of how to use JNumPy.

• demo/kmeans: a tiny Python package wrapping ParallelKMeans.jl. It produces a 10x performance gain against Scikit-Learn.

• demo/fft: a tiny Python package wrapping FFTW.jl, and indirectly the GPL-licenced FFTW library. It allows users to access FFT plans for accelerating FFTs.

Open-source contributions are kindly accepted and appreciated including bug reports, documentations, pull requests, and general suggestions.