How best to parallelize numba.stencil operations

In https://numba.pydata.org/numba-doc/dev/user/stencil.html#numba-stencil it says the following:

Thus, the stencil decorator allows clearer, more concise code and in conjunction with the parallel jit option enables higher performance through parallelization of the stencil execution.

However, my naive attempt to use the parallel=True keyword with numba.stencil has failed:

In [1]: import numba

In [2]: @numba.stencil(parallel=True)
   ...: def f(x):
   ...:     return (x[-1, -1] + x[-1, 0] + x[-1, 1] + x[0, -1] + x[0, 0] + x[0, 1] + x[1, -1] + x[1, 0] + x[1, 1]) 
   ...: / 9
   ...: 
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-2-9fe53b83f8a0> in <module>()
----> 1 @numba.stencil(parallel=True)
      2 def f(x):
      3     return (x[-1, -1] + x[-1, 0] + x[-1, 1] + x[0, -1] + x[0, 0] + x[0, 1] + x[1, -1] + x[1, 0] + x[1, 1]) / 9

~/Software/anaconda/lib/python3.6/site-packages/numba/stencil.py in stencil(func_or_mode, **options)
    722     for option in options:
    723         if option not in ["cval", "standard_indexing", "neighborhood"]:
--> 724             raise ValueError("Unknown stencil option " + option)
    725 
    726     wrapper = _stencil(mode, options)

ValueError: Unknown stencil option parallel

What is the right way to parallelize stencil operations? If I’ve missed documentation on this I apologize.

[ehsantn]

numba.stencil decorator is for creating stencil kernels. It doesn’t even jit the kernel; it remains as regular python. The njit functions that use the kernels can have the parallel=True annotation.

Looks like the documentation should be more clear on this.

[ehsantn]

Makes sense.