CODEWITHSUNDEEP
pythoncython
stochastic
stochastic

Reputation: 3413

using numpy in cython: defining ndarray datatype/ndims

I'm trying to write some cython code to do computations with numpy arrays. Cython seems to not like the [] used in all the examples I've seen to define the datatype and number of dimensions.

For example, I have a file test.pyx:

cimport numpy as np
import numpy as np

ctypedef np.ndarray[np.float64_t, ndim=2] mymatrix

cpdef mymatrix hat (mymatrix x):
    a = np.zeros((3,3));
    a[0,1] =  x[2,0];
    a[0,2] = -x[1,0];
    a[1,2] =  x[0,0];
    a[1,0] = -x[2,0];
    a[2,0] =  x[1,0];
    a[2,1] = -x[0,0];
    return a;

I compile this using a setup.py (see end of post), which I run with "python setup.py build_ext --inplace"

I get the following output:

running build_ext
cythoning test.pyx to test.c

Error converting Pyrex file to C:
------------------------------------------------------------
...
cimport numpy as np
import numpy as np

ctypedef np.ndarray[np.float64_t, ndim=2] mymatrix
                                         ^
------------------------------------------------------------

test.pyx:4:42: Syntax error in ctypedef statement

<snip, irrelevant>

whereas if I remove the "[np.float64_t, ndim=2]" part, it works fine.

Does anyone have any ideas?

As to my system setup: OS: Windows XP

full, complete pythonxy installation, version 2.6.5.1 (latest at this point)

pythonxy supposedly comes with cython, but I ended up installing cython version 0.12.1 for Python 2.6 from this site: http://www.lfd.uci.edu/~gohlke/pythonlibs/#cython

I suspect that I somehow am missing a path or something: I solved some problems by explicitly adding the numpy header file directory to the include path used by mingw (see the setup.py file below)

here is that setup.py file I mentioned:

from distutils.core import setup
from distutils.extension import Extension
from distutils.sysconfig import get_python_inc
from Cython.Distutils import build_ext
import os.path

inc_base = get_python_inc( plat_specific=1 );
incdir = os.path.join( get_python_inc( plat_specific=1 ), );

#libraries=['math'],
ext_modules = [Extension("test", 
 ["test.pyx"], 
 include_dirs = [
  os.path.join(inc_base,'..\\Lib\\site-packages\\numpy\\core\\include\\numpy'),
  ]
 )
 ]

setup(
  name = 'test',
  cmdclass = {'build_ext': build_ext},
  ext_modules = ext_modules
)

Upvotes: 8

Views: 2588

Answers (2)

Fabian Pedregosa
Fabian Pedregosa

Reputation: 6539

Put the type information in the declaration of the function, as in:

def hat (ndarray[np.float64_t, ndim=2] x):
    a = np.zeros((3,3));
    a[0,1] =  x[2,0];
    etc.

Upvotes: 3

Louis
Louis

Reputation: 1432

I think you cannot do it directly: you have to check the shape and type in the function

assert x.shape[0] == 2
assert x.dtype == np.float64

and only cdeftype np.ndarray mymatrix in the header

BUT you lose the typing of matrix values you thus have to assign each value you process to float64_t: but what should be the efficency?

Louis

Upvotes: 0

Related Questions