Is there a straightforward way to pass numpy.float128 arguments to C++?

[mreineck]

I'd like to pass an object of type numpy.float128 from Python to a C++ function wrapped by pybind11. My first attempt was to simply use "long double" as argument type on the C++ side, but the values I'm passing are truncated to about "double" accuracy - I expect this happens because the value is cast to a standard Python "float" somewhere on the way.

What would be the proper way to achieve this?

[eacousineau]

TTBOMK, there may not be a straightforward "scalar" caster for NumPy types that are not directly tied to Python types (e.g. np.uint*, np.float16, etc.).

As a first step, here's showing float128 works as an array (but avoiding the scalar conversion):
eacousineau/repro@b04db11#diff-b4dfa50924d08e44f5f4cd7861a2fa26R56

Next step would be to just lie to pybind11 about the scalar, and return a scalar np.ndarray.
Will briefly try that out.

[eacousineau]

Hm... Trying out a type_caster<float128> has proven to be a bit persnickety...
eacousineau/repro@2c30668#diff-b4dfa50924d08e44f5f4cd7861a2fa26R44

Also, through the current API, I'm not sure how to enforce a scalar ndarray vs an empty ndarray:
https://docs.scipy.org/doc/numpy/reference/arrays.scalars.html

I'm not sure if that's what's causing the circular 'base' error...

[eacousineau]

Ah, was just not releasing references correctly.

Working code:
https://github.com/eacousineau/repro/blob/5e6acf6/python/pybind11/custom_tests/test_numpy_issue1785.cc#L103-L105

[mreineck]

Thanks a lot for looking into this!

I have tried to compile your code, but got:

martin@debian:~/codes/pypocketfft$ python3 setup.py install --user -f
running install
running bdist_egg
running egg_info
writing pypocketfft.egg-info/PKG-INFO
writing dependency_links to pypocketfft.egg-info/dependency_links.txt
writing requirements to pypocketfft.egg-info/requires.txt
writing top-level names to pypocketfft.egg-info/top_level.txt
reading manifest file 'pypocketfft.egg-info/SOURCES.txt'
writing manifest file 'pypocketfft.egg-info/SOURCES.txt'
installing library code to build/bdist.linux-x86_64/egg
running install_lib
running build_ext
building 'pypocketfft' extension
x86_64-linux-gnu-gcc -pthread -DNDEBUG -g -fwrapv -O2 -Wall -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 -fPIC -I./ -I/usr/include/eigen3/ -I/usr/local/include/python3.7 -I/home/martin/.local/include/python3.7m -I/usr/include/python3.7m -c pypocketfft.cc -o build/temp.linux-x86_64-3.7/pypocketfft.o -march=native -O3 -Wfatal-errors -Wno-ignored-attributes --std=c++14
In file included from /home/martin/.local/include/python3.7m/pybind11/embed.h:12,
                 from pypocketfft.cc:4:
/home/martin/.local/include/python3.7m/pybind11/pybind11.h: In instantiation of ���void pybind11::cpp_function::initialize(Func&&, Return (*)(Args ...), const Extra& ...) [with Func = init_module(pybind11::module)::<lambda(Vector<long double, -1>)>; Return = long double; Args = {Eigen::Matrix<long double, -1, 1, 0, -1, 1>}; Extra = {pybind11::name, pybind11::scope, pybind11::sibling}]���:
/home/martin/.local/include/python3.7m/pybind11/pybind11.h:65:9:   required from ���pybind11::cpp_function::cpp_function(Func&&, const Extra& ...) [with Func = init_module(pybind11::module)::<lambda(Vector<long double, -1>)>; Extra = {pybind11::name, pybind11::scope, pybind11::sibling}; <template-parameter-1-3> = void]���
/home/martin/.local/include/python3.7m/pybind11/pybind11.h:813:22:   required from ���pybind11::module& pybind11::module::def(const char*, Func&&, const Extra& ...) [with Func = init_module(pybind11::module)::<lambda(Vector<long double, -1>)>; Extra = {}]���
pypocketfft.cc:62:4:   required from here
/home/martin/.local/include/python3.7m/pybind11/pybind11.h:167:95: error: no match for call to ���(const pybind11::detail::descr<8, 0>) ()���
         PYBIND11_DESCR signature = _("(") + cast_in::arg_names() + _(") -> ") + cast_out::name();
                                                                                 ~~~~~~~~~~~~~~^~
compilation terminated due to -Wfatal-errors.
error: command 'x86_64-linux-gnu-gcc' failed with exit status 1

(Don't worry about the file names, I just replaced my C++ code with your file.)
Should I be using the current git HEAD of pybind11 instead of 2.2.4?

[eacousineau]

Oops! Sorry I didn't see this earlier!

Huh... TBH, I dunno... I've hacked around with doing custom dtypes, and I didn't have to make name a function...

However, @a-sevin also ran into that same thing here:
#1776 (comment)

Are you right at v2.2.4?

[mreineck]

Are you right at v2.2.4?

Yes, to the best of my knowledge.

I probably won't be able to require Eigen on the target machines, so I guess I'll have to wait anyway until "native" support materializes in pybind11. Luckily this is not a terribly important feature for me, more of a "nice-to-have" :)