Memoryview 'DTYPE_t[:, :]' not conformable to memoryview 'DTYPE_t[:, :]'

[rens holmer]

I am relatively new to cython, but I have managed to make some small functions work.

Currently I am trying to build a modular cython codebase where important functions are declared with 'cpdef' so I can write unit tests in python, but use the optimized c-version internally. As far as my understanding of cython goes, this is what cpdef is supposed to do: expose a function to both python and c.

My test codebase is as follows:

test.pyx:

cimport cython

import numpy as np

cimport numpy as np

from fit cimport fit #<--- CHANGING THIS FIXES THE ERROR

ctypedef np.float64_t DTYPE_t

cpdef DTYPE_t test(DTYPE_t[:,:] dat, DTYPE_t[:,:] regs):

cdef DTYPE_t score = fit(dat, regs)

return score

fit.pyx

cimport cython

import numpy as np

cimport numpy as np

cpdef DTYPE_t fit(DTYPE_t[:,:] A, DTYPE_t[:,:] B):

cdef DTYPE_t score

return score

fit.pxd

cimport cython

import numpy as np

cimport numpy as np

ctypedef np.float64_t DTYPE_t

cpdef DTYPE_t fit(DTYPE_t[:,:] dat, DTYPE_t[:,:] regs)

setup.py

from setuptools import setup, find_packages

from setuptools.extension import Extension

from Cython.Build import cythonize

import numpy

extensions = [

Extension(

'*',

include_dirs=[numpy.get_include()],

language = 'c++',

extra_compile_args = ['-ffast-math','-O3','-Wno-unused-function'],

sources = ['*.pyx']

)

]

setup(

name = 'test',

version = '0.0.1',

packages = find_packages(),

ext_modules = cythonize(extensions)

)

In summary: test takes two 2D numpy arrays as memoryviews and passes those to fit, which will do the actual calculations and return a score.

Ideally I would write unit tests in python that import these functions.

The above codebase results in the following error:

Memoryview 'DTYPE_t[:, :]' not conformable to memoryview 'DTYPE_t[:, :]'

Which seems odd, given that the error is reporting two times the same memoryview type, but still declares it 'not conformable'?

If I change the line 'from fit cimport fit'  to   'from fit import fit'  (i.e. use regular python import), the error no longer occurs.

What am I missing here?

[Stefan Behnel]

You are redefining the DTYPE_t both in fit.pxd and in test.pyx. While this
should normally work, it fails in your case for some reason. Probably a bug
in Cython somewhere, and probably also related to the fact that this is
using external NumPy types here...

I suggest using "fit.DTYPE_t" in test.pyx instead of redeclaring it. That's
definitely the clean way of doing it.

Stefan

[rens holmer]

Awesome, that fixed the error, thanks!

For the record, in test.pyx I changed

from fit cimport fit 

ctypedef np.float64_t DTYPE_t

To:

from fit cimport fit, DTYPE_t