Long story short, LowLevelCallable.from_cython() needs to be pointed at a compiled, C-style function to work. You can do that with C, Cython, or Numba. Of these, the Cython method seems the easiest, but is strangely the least documented.
Calling Cython from SciPy
=========================
Perhaps you have noticed that SciPy offers [low-level callback
subdirectory of a repo by Ashwin Vishnu was useful.
-`setup.py` is the usual distutils thing, but it knows it's
building a Cython module because of the call to `cythonize()`.
-`run.py` uses the `integrand()` function from `test.pyx` to do
some numerical integrals.
Here are the files I have sent:
To try this out you will need: python 3, cython
- test.pyx is a Cython file, which contains a C-style function called integrand(). We want to turn this into C, compile it, and then later on use it to instantiate a scipy.LowLevelCallable instance so scipy.integrate can use than rather than working with interpreted python (which is slow).
1. Run:
- test.pxd is a Cython "header" file (I don't know what the correct name is). When (I think) cythonize() in setup.py detects this, it knows it's building a Cython file we want to call from C and it should behave itself accordingly, exporting symbols that a C linker can use if need be. Otherwise, Cython does not let c-style functions (those defined with cdef) be called from outside the Cython file, e.g. from other C code.
$ python3 setup.py build_ext --inplace
The corresponding python module then has the attribute __pyx_capi__, which is also needed for LowLevelCallable.from_cython() to work. I think this means that it has a 'parallel' C interface, as it were. There's another relevant example here: https://github.com/ashwinvis/cython_capi/tree/master/using_pxd
This builds the module from test.pyx (with symbols from
test.pxd). Although running setup.py like this does the
compilation for you, you can by the way see the C code generated
in test.c, the function call starts something like:
- setup.py is the usual distutils thing, but it knows it's building a Cython module because of the call to cythonize()
static double __pyx_f_4test_integrand(CYTHON_UNUSED int __pyx_v_n, double *__pyx_v_args)
- run.py uses the integrand() function from test.pyx to do some numerical integrals.
[`CYTHON_UNUSED` _I think_ because the argument n is not used in `integrand()`, but the standard rubric for `LowLevelCallable.from_cython()` needs it]
But note that it really churns out `double`s, not some wacky
higher-level type. This is really C at the end of the day.
to evaluate three numerical integrals. Note that I've tried to
make it clear how to pass numerical coefficients from the python
to the C code, so things don't need to be hardcoded.
1. Run:
$ python3 setup.py build_ext --inplace
Outstanding issues
------------------
This builds the module from test.pyx (with symbols from test.pxd). Although running setup.py like this does the compilation for you, you can by the way see the C code generated in test.c, the function call starts something like:
I didn't quite figure out how to make this work with NumPy data
types. But maybe that doesn't matter.
static double __pyx_f_4test_integrand(CYTHON_UNUSED int __pyx_v_n, double *__pyx_v_args)
[CYTHON_UNUSED because the argument n is not used in integrand(), but the standard rubric for LowLevelCallable.from_cython() needs it]
Further reading
---------------
But note that it really churns out doubles. This is really C at the end of the day.
*[Developing a Cython
library](http://pdebuyl.be/blog/2017/cython-module.html) by Pierre
de Buyl helped me figure out that I needed a `.pxd` file.
2. To test:
*[Experiments with Cython C API and
PyCapsules](https://github.com/ashwinvis/cython_capi) by Ashwin
Vishnu is also worth glancing at, but also has some valuable
references to other similar examples of interfacing Cython, Python
by Juan Nunez-Iglesias is written from the point of view of Numba
Should be 21.333...: 21.3333
and image processing (`ndimage`), but is useful too.
Uses scipy.integrate.quad() to evaluate three numerical integrals. Note that I've tried to make it clear how to pass numerical coefficients from the python to the C code, so things don't need to be hardcoded.