[Numpy-discussion] Mapping of dtype to C types

[Bruce Southey]

On Wed, May 11, 2011 at 2:07 PM, Olivier Delalleau <shish at keba.be> wrote:
> 2011/5/11 Sturla Molden <sturla at molden.no>
>>
>> Den 09.05.2011 15:58, skrev Keith Goodman:
>> > On Mon, May 9, 2011 at 1:46 AM, Pauli Virtanen<pav at iki.fi>  wrote:
>> >> Sun, 08 May 2011 14:45:45 -0700, Keith Goodman wrote:
>> >>> I'm writing a function that accepts four possible dtypes: int32,
>> >>> int64,
>> >>> float32, float64. The function will call a C extension (wrapped in
>> >>> Cython). What are the equivalent C types? int, long, float, double,
>> >>> respectively? Will that work on all systems?
>> >> Long can be 32-bit or 64-bit, depending on the platform.
>> >>
>> >> The types available in Numpy are listed here, including
>> >> the information which of them are compatible with which C types:
>> >>
>> >> http://docs.scipy.org/doc/numpy/reference/arrays.scalars.html
>> >>
>> >> The C long seems not to be listed -- but it's the same as
>> >> "Python int", i.e., np.int_ will work.
>> >>
>> >> IIRC, the Numpy type codes, dtype.kind, map directly to C types.
>> > Does this mapping look right?
>> >
>>
>> No.
>>
>> C int is at least 16 bits, C long is at least 32 bits.
>>
>> The size of long and size of int depend on compiler and platform.
>>
>> I'd write a small Cython function and ask the C compiler for an
>> authorative answer.
>>
>> Alternatively you could use ctypes. This is 64-bit Python on Windows:
>>
>>  >>> import ctypes
>>  >>> ctypes.sizeof(ctypes.c_long)
>> 4
>>  >>> ctypes.sizeof(ctypes.c_int)
>> 4
>>
>
> I think Keith's approach should work, as long as there is one C type listed
> on the url you mention that corresponds to your four dtypes.
> Something like (not tested at all):
>
> map = {}
> for dtype in ('int32', 'int64', 'float32', 'float64'):
>   for ctype (N.byte, N.short, N.intc, N.long.long, ...): # List all
> compatible C types here
>     if N.dtype(ctype) == dtype:
>        map{dtype} = ctype
>        break
> assert len(map) == 4
>
> -=- Olivier
>
>
> _______________________________________________
> NumPy-Discussion mailing list
> NumPy-Discussion at scipy.org
> http://mail.scipy.org/mailman/listinfo/numpy-discussion
>
>

But that result is just due to how Microsoft implemented it's 64-bit
support so it will not work for 64-bit Linux and other similar OSes.

I thought that you would just use things like npy_intp (or intp) and
npy_float as used by mtrandom:
numpy/random/mtrand/numpy.pxi

These are defined in
numpy/core/include/numpy/ndarraytypes.h


Bruce