setting nogil at cimport

[Björn Dahlgren]

Hi,

Here's an odd situation, I have a C++ function which takes a template typename argument (a C++ class).
Depending on what I specialize this function as, it may or may not touch any python objects (depending on whether that class
holds any python object references).

I guess there is no way to indicate at cimport time whether I consider it to be 'nogil' or not?
I can just create another copy of the pxd of course: myobj.pxd & myobj_nogil.pxd - perhaps that is easier here?

Regards,
Björn

[Robert Bradshaw]

No, this isn't possible. You're probably better off generating two copies of the pxd. (That is, assuming you actually take advantage of releasing the gil in the one cases where you can--if not then you don't even need to worry about it.)

--

---
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+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Björn Dahlgren]

On Friday, 23 September 2016 23:45:16 UTC+2, Robert Bradshaw wrote:

No, this isn't possible. You're probably better off generating two copies of the pxd. (That is, assuming you actually take advantage of releasing the gil in the one cases where you can--if not then you don't even need to worry about it.)

Thanks. Yes, I was hoping to use prange. Is it enough to just do:

with nogil:
    c_callback(&my_numpy_array[0])

or does c_callback need to be declared "nogil" in my pxd as well?

[Stefan Behnel]

Cython validates your code for GIL dependencies inside of nogil sections,
so no, this will require the declaration to be there.

I'd be careful with keeping Python object reference in a C++ container,
though. It prevents automatic reference cleanup on deallocation, so you're
very much on your own here.

Stefan