Please document if it is possible to build extensions interfacing with fortran libraries using real128

[callegar]

Because there are fortran libraries that use real128, it would be interesting to know if via numpy-quadtype numpy it can be possible to take advantage of them. Being able to use cython to build python extensions acting as bridges to them would be a nice option.

[ngoldbaum]

I think that would require numpy_quaddtype to expose a C API, which introduces a lot of headaches for a Python package. NumPy does it but it's a big pain.

Ideally it would all work using Cython's support for typed memoryviews, but that can only work with types supported by the buffer protocol, which is limited to a set of enumerated types that doesn't include float128.

I have long-term plans to improve the buffer protocol and enable support for custom data types, but that's something that needs a PEP and is not a short-term fix.

All that to say: yes this is possible and desirable, but it will require a lot of work and some design decisions. If numpy-quaddtype starts exposing a C API that adds a lot of complexity since we need to be careful to maintain both forward and backward compatibility. I'm not sure we want to take that on.

[carlkl]

You may take a look here: Add functionality to replace libm and libquadmath functions or aks the sleef maintainer directly

[carlkl]

Mayby TLFloat from the same author is the better libquadmath replacement. To my knowledege neither sleef nor tlfloat supports complex quadmath functions.

Reading you issue correctly you need to support fortran libraries with quadmath functions from python/numpy with the quadttype extension?

[callegar]

My desire is actually not specifically to replace libquadmath, rather it regards the many fortran libraries that are already available and that also provide quadruple precision operations (presumably using libquadmath on hardware that does not provide native quadruple precision floats — x86-64 mainly). Some of these libraries are directly supported in scipy or anyway have wrapping packages that let them be used in NumPy-based Python projects, but with a limitation: only the functions operating on doubles are made available in Python, since NumPy does not provide a quad data type. What I wonder is if there is documentation letting one code a Python extension package that lets the functions operating on quads be made available in Python using the quad type provided by numpy-quadtype.

Just an example: this quadpack implementation exports a single (real32), double (real64) and quadruple (real128) precision interface. Would it be possible to use numpy-quadtype to use the real128 interface?

[carlkl]

It should be possible to call a fortran90 wrapper function with ctypes, that gets quad scalars arguments by value and quad array per reference to avoid copies / marshalling. This is not as elegant as with f2py of course.

[ngoldbaum]

Not that I'm saying we should do this, but I think it might require numpy_quaddtype to have a PyCapsule-exported C API like the NumPy C API.

You also can't easily use cython memoryviews because that requires support from the buffer protocol. I have medium-term plans to work on improvements for the buffer protocol that would enable this but it's not quite there yet.