[SciPy-dev] numpy long vs. int

[Damian Eads]

> On Fri, Jul 25, 2008 at 17:36, Damian Eads <eads at soe.ucsc.edu> wrote:
>> Hi there,
>>
>> I had some problems porting some internal extension code I wrote
>> recently.
>> The code worked fine on 32-bit but did not work on 64-bit. When np.int_
>> is
>> used as the dtype argument to np.zeros or np.asarray, an array results
>> that has typecode NPY_INT on 32-bit and NPY_LONG on 64-bit. This
>> inconsistency is problematic when type checking arrays in C-space prior
>> to
>> passing them to a C function expecting a specific type, like int.
>> Changing
>> dtype=np.int_ to dtype='i' seems to consistently result in an array with
>> typecode NPY_INT on both architectures, which is desired.
>>
>> I didn't think to write about this until I encountered the very same
>> problem today when trying to compile the ANN scikit on 64-bit. All the
>> tests failed because the kd-tree array passed was of type long.
>>
>> ======================================================================
>> ERROR:
>> test_ann.TestANNWrapper.test_knn_returns_nearest_neighbor(array([[
>> 0.53209373,  0.2149725 ],
>> ----------------------------------------------------------------------
>> Traceback (most recent call last):
>>  File
>> "/mirror/ssrc/eads/sci-tools/prefix-x86-64/lib/python2.5/site-packages/nose-0.10.3-py2.5.egg/nose/case.py",
>> line 182, in runTest
>>    self.test(*self.arg)
>>  File
>> "/mirror/ssrc/eads/sci-tools-source/x86-64/ann/scikits/ann/tests/test_ann.py",
>> line 49, in checkReturnNN
>>    nn,nn_distances = tree.knn(pt,1)
>>  File "scikits/ann/__init__.py", line 113, in knn
>>    self._knn2(pts, idx, d2, eps)
>>  File "scikits/ann/ANN.py", line 45, in _knn2
>>    def _knn2(*args): return _ANN._kdtree__knn2(*args)
>> TypeError: Array of type 'int' required.  Array of type 'long' given
>> =================================
>>
>> Changing line 110 from dtype=np.int_ to dtype='i' fixed the problem.
>> Some
>> people seem insistent on using a type object (e.g. np.int_ or np.float_)
>> instead of a string. In fact, when I checked in my hierarchical
>> clustering
>> code, I noticed someone eventually changed all the dtype's in my code to
>> use the type objects. I had no qualms with this until now. Are there
>> type
>> objects that can be passed to dtype to guarantee consistency in
>> translation to NPY_XXX type codes? We should probably write a caveat in
>> the Numpy C extensions help document explaining this inconsistency.
>
> numpy.intc gives NPY_INT. numpy.int_ gives NPY_LONG (since Python ints
> are C longs).

Any reason why there aren't np.floatc or np.doublec equivalents?

Just curious,

Damian