CFFI variable array limitation

[Adam G]

I encountered a CFFI limitation while trying to pass in a double (*)[3] to add_array:

void add_array(int x, int y, double (*a)[y], double (*b)[y], double (*result)[y])

https://github.com/nszceta/cffi-numpy-demo/blob/e559296055c3b169d42359f5ba0e793ef0273e73/test.c#L11

To use this syntax, I need to define test.h as follows:

void add_array(int x, int y, double (*a)[NUM_COLS], double (*b)[NUM_COLS], double (*result)[NUM_COLS]);

https://github.com/nszceta/cffi-numpy-demo/blob/e559296055c3b169d42359f5ba0e793ef0273e73/test.h#L2

And then parse NUM_COLS as follows:

ffi.cdef(my_header.read().replace('NUM_COLS', str(num_cols)))  # IMPORTANT! cffi limitation

https://github.com/nszceta/cffi-numpy-demo/blob/e559296055c3b169d42359f5ba0e793ef0273e73/test.py#L35

Is it possible for CFFI to automatically infer the shape of the array?

[Adam G]

As a workaround, I ended up passing in a double * and performing the pointer arithmetic manually:

https://github.com/nszceta/cffi-numpy-demo/blob/0cdd02074dcfa9ed4e79efffcd87d887ce5d28ab/test.c

[Ryan Gonzalez]

I think this is a misunderstanding?

https://lkml.org/lkml/2015/9/3/428

--
Sent from my Nexus 5 with K-9 Mail. Please excuse my brevity.

[Armin Rigo]

Hi Adam,

On Fri, Nov 6, 2015 at 2:55 AM, Ryan Gonzalez <rym...@gmail.com> wrote:
> I think this is a misunderstanding?

>>> To use this syntax, I need to define test.h as follows:
>>>
>>> void add_array(int x, int y, double (*a)[NUM_COLS], double
>>> (*b)[NUM_COLS], double (*result)[NUM_COLS]);

Yes, indeed. This is half-nonsensical C code. Don't do that and pass
plain "double *a" and "double *b" arguments and do the arithmetic
yourself---you need to anyway if the number of columns is not a
constant. That's how C works.


A bientôt,

Armin.