compilation failed with the new cython version (0.28)

[skoudoro]

Description

As you can see on this build, the new Cython version (0.28) does not allow to use memoryviews as struct members.

In this discussion, it is hard to say if it will be something permanent or not.

BTW, We have to find a workaround.

[MarcCote]

If I remember correctly, this comment (cython/cython#1453 (comment)) is exactly what we are doing, i.e. manually nulling the memview at creation and freeing it in dealloc.

[Garyfallidis]

Check the travis error the message is clear. Memory views not allowed as struct members.

[Garyfallidis]

And even if you nullify it the possibility for memory leak is still there.

[Garyfallidis]

Basically, they are not allowing the code to compile when we add memviews inside structs. We can suggest to reconsider not changing this option.

[Garyfallidis]

cdef struct Centroid:
    Data2D features # float* features
    int size
    float[6] aabb

[Garyfallidis]

Perhaps changing from Data2D to float * will do the job? Let us know if you have time to hangout on this. We need to coordinate.

[Garyfallidis]

Otherwise perhaps we can change from struct to class for structures with memviews?

[Garyfallidis]

Okay @MarcCote the class idea was not good.
But I think we can make something like this

cdef struct Centroid2:
    float * features #using pointer rather than memview
    int size

and then have functions to move from pointer to memview and memview to pointer

def memview_to_pointer(float [:, ::1] mv):
    cdef Centroid2 c
    c.features = <float *> &mv[0, 0]

def pointer_to_memview(size_t N, size_t M):

    cdef Centroid2 c
    c.features = <float *>PyMem_Malloc(N * M * sizeof(float))  
    cdef float[:, ::1] mv = <float[:N, :M]> c.features
    return mv
    
def free_memview(float [:, ::1] mv): #not working
    if mv._data is not NULL:
        PyMem_Free(mv._data) 

[MarcCote]

I like the Centroid2 struct idea. Do we really need the conversion functions, though?

[Garyfallidis]

The data are coming in as memviews or numpy arrays. And you use 2D memviews internally too. So... probably yes. But if you think otherwise we would love to hear an alternative.

[MarcCote]

I don't have a precise idea on how to do it. I wasn't able to install cython 0.28 on my machine. Is 0.28 officially released?

[skoudoro]

This is a future release @MarcCote. Create a new environment and you can install it via :

pip install --find-links=https://5cf40426d9f06eb7461d-6fe47d9331aba7cd62fc36c7196769e4.ssl.cf2.rackcdn.com --find-links=https://7933911d6844c6c53a7d-47bd50c35cd79bd838daf386af554a83.ssl.cf2.rackcdn.com --pre cython numpy scipy h5py nibabel

[skoudoro]

fix by #1434, closing !