[Numpy-discussion] __array_interface__ / __array_struct__

[Travis E. Oliphant]

Thomas Heller wrote:
> Travis E. Oliphant schrieb:
>   
>> Thomas Heller wrote:
>>     
>>> I am experimenting with implementing __array_interface__ and/or __array_struct__
>>> properties for ctypes instances, and have problems to create numpy arrays
>>> from them that share the memory.  Probably I'm doing something wrong;
>>> what is the correct function in numpy to create these shared objects?
>>>
>>> I am using numpy.core.multiarray.array(ctypes-object), is that correct?
>>>   
>>>       
>> Yes, this should work, as the array function goes through several checks 
>> including looking for the __array_struct__ and/or __array_interface__ 
>> attributes.    If you can point me to the code, I can probably help.
>>
>>     
>
> The pure-python code, using __array_interface__, is here:
>
> http://ctypes-stuff.googlecode.com/svn/trunk/numpy/
>
> Use a webbrowser to view or download it, or an svn client
> to checkout a copy.  I use it like this:
>
>   from ctypes_array mport as_ctypes, as_array
>   c_array = (c_double * 3)()
>   numpy_array = as_array(c_array)
>
> or
>
>   numpy_array = zeros(32)
>   c_array = as_array(numpy_array)
>
> In the former example the objects to bot share memory, in the
> latter example they do.
>
> I also have an extension in the works that uses __array_struct__,
> but this is not yet uploaded.
>   

This is all very good.  The only thing missing that causes the former to 
not share memory is you are missing a  copy=False argument to the 
multi_array function. 

Thus:

return multi_array(obj, copy=0)

is what you need to use.


Also, the address of the memory is also available as

arr.ctypes.data  if arr is a NumPy array (but this is less general than 
using the array interface for sure). 


-Travis O.