On Wed, Jul 7, 2010 at 10:13 PM, Christoph Gohlke <cgohlke@uci.edu> wrote:

Dear NumPy developers,

I am trying to solve some scipy.sparse TypeError failures reported in
[1] and reduced them to the following example:

>>> import numpy
>>> a = numpy.array([[1]])

>>> numpy.dot(a.astype('single'), a.astype('longdouble'))
array([[1.0]], dtype=float64)

>>> numpy.dot(a.astype('double'), a.astype('longdouble'))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: array cannot be safely cast to required type

Is this exception expected?

I think not. On some platforms longdouble is the same as double, on others it is extended precision or quad precision. On your platform this looks like a bug, on my platform it would be correct except there is a fallback version of dot that works with extended precision. Is there a mix of compilers here, or is it msvc all the way down.

In [5]: a = array([[1]])

In [6]: dot(a.astype('single'), a.astype('longdouble'))
Out[6]: array([[1.0]], dtype=float128)

Also I noticed this:

>>> numpy.array([1]).astype('longdouble').dtype.num
13
>>> numpy.array([1.0]).astype('longdouble').dtype.num
12

Yeah, that is probably correct in a strange sort of way since the two types are identical under the hood. On ubuntu I get

In [1]: array([1]).astype('longdouble').dtype.num
Out[1]: 13

In [2]: array([1.]).astype('longdouble').dtype.num
Out[2]: 13

Type numbers aren't a good way to determine precision in a platform independent way.

I am using Python 2.6.5 for Windows and numpy 1.4.1 compiled with msvc9,
where sizeof(longdouble) == sizeof(double).

Chuck