Headers in universal2 wheels are not architecture independent

[isuruf]

Since auto-generated file _numpyconfig.h has NPY_SIZEOF_LONGDOUBLE and NPY_SIZEOF_COMPLEX_LONGDOUBLE generated, it's wrong for x86_64 as the universal2 wheel distributes the arm64 version of it.

Easiest option would be to stop distributing universal2 wheels and only distribute thin wheels.

[isuruf]

We can ignore universal2 wheels also because pip prefers thin wheels over universal2 wheels if the thin wheel exist.

[tylerjereddy]

Ah ok, this would make sense as the source of the problems based on some back and forth with @charris as well

[rgommers]

Thanks for tracking that one down!

Easiest option would be to stop distributing universal2 wheels and only distribute thin wheels.

That's fine with me. Probably a good thing in the long run; universal2 is a bit pointless, just increases download sizes.

[isuruf]

Or we merge numpy/numpy#20270

[isuruf]

In the meantime we should either upload a thin wheel or yank numpy-1.21.3-cp310-cp310-macosx_10_9_universal2.whl

[charris]

@isuruf Do you know if rosetta supports extended precision long doubles? We ran into a problem with cpp where we needed to use typedef long double npy_longdouble, which I am thinking may have been a wrong workaround.

[isuruf]

Yes.

(base) idf2@M1MD ~ % cat asd.cpp         
#include<iostream>
  
int main()
{
    long double eps = 0.5;
    while ((1+eps) != 1)
    {
        eps /=2;
    }
    std::cout << eps*2 << std::endl;
}
(base) idf2@M1MD ~ % clang++ asd.cpp -arch arm64 -arch x86_64
(base) idf2@M1MD ~ % arch -x86_64 ./a.out                    
1.0842e-19
(base) idf2@M1MD ~ % arch -arm64 ./a.out                     
2.22045e-16

[charris]

@isuruf I'm happy to do both. Do the thin 3.10 builds work now?

[isuruf]

You'll have to use multi-build/multibuild#438 to get a thin wheel with 3.10

[charris]

I've merged #20270. I'll trigger a nightly build so it can be tested.

[charris]

@isuruf Is there a problem getting multi-build/multibuild#438 merged?

[isuruf]

Nope. Just needs a review

[EwoutH]

multi-build/multibuild#438 was merged yesterday!

[charris]

Both numpy and multibuild have been updated to deal with this.