memory view from fortran-contigous pointer

[Soren Christensen]

Hi,

 I am having problems casting a short* array to a memory view (with the ultimate goal of exposing this memoryview to a numpy array).

I get an error when running cython on my .pyx file:

...

        cdef getData(self)

                 .....

        cdef short[ ::1,:,:] mya= <short[:128 , :128, :14]> self._thisptr.getBufferPtr()    #this is the short pointer

        z=np.asarray(mya)

        return z

...

Error:

 cdef short[ ::1,:,:] mya= <short[:128 , :128, :14]> self._thisptr.getBufferPtr()

                                 ^

DataImporter.pyx:72:34: Memoryview 'short[:, :, ::1]' not conformable to memoryview 'short[::contiguous, :, :]'.

If I declare the mya array to be c-contigous, there are no technical issues, but the numpy array is then nominally c-contiguous with an f-contiguous raw buffer and my 3D matrix is messed up (as expected).

1) Why am I getting this error?   Is there a different casting syntax to specify that the buffer is fortran contiguous?

2) If not, are there any good work arounds?  Casting to 1D buffer and reshaping perhaps? 

Thanks!

Soren

[Soren Christensen]

Sorry, did eventually find a previous post with this.

Should be:

cdef short[ ::1,:,:] mya= <short[:128:1 , :128, :14]> self._thisptr.getBufferPtr()
This works well.

Is this level of detail documented anywhere right now?

Thanks!

Soren

[Nathaniel Smith]

Probably there is some better solution, but a silly/simple hack would be to exploit the fact that if some memory buffer is an f-contiguous representation of an array A, then it is also a c-contiguous representation of the array A-transpose. So you could take a C contiguous view with your shape reversed, and then return resulting_array.T

(The .T attribute on numpy arrays never makes a copy, because it can exploit the above trick internally.)

-n

--

---
You received this message because you are subscribed to the Google Groups "cython-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cython-users...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Soren Christensen]

Thanks for the suggestion!

I think my follow up post may have crossed yours. Eventually the bold faced addition did the trick of making the pointer cast Fortran contigous.

cdef short[ ::1,:,:] mya= <short[:128:1 , :128, :14]> 
self._thisptr.getBufferPtr()

I think transposing would also have worked - but it seems to be limited to 2 dimensions?

Thanks in any case!

Soren

--

---
You received this message because you are subscribed to a topic in the Google Groups "cython-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/cython-users/A7tUKycLZfM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to cython-users...@googlegroups.com.

[Nathaniel Smith]

On Sat, Jan 2, 2016 at 6:02 PM, Soren Christensen <sor...@gmail.com> wrote:
> Thanks for the suggestion!
> I think my follow up post may have crossed yours. Eventually the bold faced
> addition did the trick of making the pointer cast Fortran contigous.
> cdef short[ ::1,:,:] mya= <short[:128:1 , :128, :14]>
> self._thisptr.getBufferPtr()
>
> I think transposing would also have worked - but it seems to be limited to 2
> dimensions?

Sounds like you solved your problem, but for the record, no, the numpy
transpose operator (np.transpose or array.T) applies to
higher-dimensional arrays by reversing the dimension order, which is
also what you want for flipping between C and F ordering.

-n

--
Nathaniel J. Smith -- http://vorpus.org

[Soren Christensen]

Thanks again!

I assumed it was for 2D matrices only, but I see now it is nd and would also work.

Soren