[Cython] new FFI library for Python

[Dag Sverre Seljebotn]

Saravanan Shanmugham <sarvilive at gmail.com> wrote:

>I have been using CFFI for wrapping a whole bunch of C libraries and I
>can
>tell you its way easier than Cython.
>
>And a couple of clarifications of misconceptions mentioned on this
>list.
>1. It does not "require" GCC at run time. The extension can be compiled
>at package/install time using the GCC and can run without need GCC at
>run
>time.
> 2. Its way easier to wrap C libraries than Cython. I would know, since
>I started with Cython. And after many email back and forth with the
>cython-users mailing list, I was referred to CFFI which turned out
>great.
>The trouble was wrapping C structures and datastructures with Python
>wrappers.
>3. I had to just copy over the C header files as is and had a replace a
>few #define values with "..." to ask it to automatically get these
>values
>during compile time from the compiler.  This makes it even more
>simpler.
>
>
>Now, I am still a Cython lover when it comes optimizing portions of my
>Python code and having it compiled to C.
>
>The question I had in mind is, If I had say a shared math library
>libmath.so and wrapped it with a CFFI based python wrapper, which
>includes
>a portion of python wrapper code that calls into the generated Python/C
>wrapper shared library.
>And then wrote a piece of cython code that invoked that CFFI based
>python
>wrapper, how will my performance be compared to
>writing the python wrapper for libmath.so with  cython cdef or cpdef
>wrapper function definitions.

It will be worse, it will be like calling Python functions rather than C.

Fixing this can be done but would require support in both cffi and Cython.

We were discussing a wrapper-language-neutral protocol for fixing this problem (also so that Cython could call numpy.sin and friends without any overhead). CEP1000 is an early draft of that idea; it has developed a lot since then but the details are in mailing list threads and my and Robert's heads.

If anybody would like to work on that then I'll find the time to write down all the ideas me and Robert worked out. (but unless anybody stands ready I'd rather wait another month)

Dag Sverre


>
>Thanks,
>Sarvi
>_______________________________________________
>cython-devel mailing list
>cython-devel at python.org
>http://mail.python.org/mailman/listinfo/cython-devel

-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity.