Compilation error of Python bindings on linux-64

[adriendelsalle]

Description

I got some compilation errors when compiling the Python bindings on linux-64:

• the auto deduction of return type in multiple overloads causes pybind11::overload_cast compilation error on MeridionalGrid<T>::operator().

  • after investigating a bit, gcc=11 should help on that (see the comment I left on related pybind11 issue) but it looks like it's not for our case. I would need to spend more time on that to understand what's going on
  • I propose to wrap the call in a lambda or to use a template parameter for return type to solve this compilation error

• the reducer used in gbs-mesh/gbs::msh_curves_set_sizes causes a compilation error using gcc=10

  • gcc>=11.2 doesn't complain. I found a bug in std::reduce overload using a binary functor in gcc=10, tracked here
  • for compatibility with current reference conda toolchain using gcc=10 it would be better to use std::accumulate instead of std::reduce in upstream project gbs, which is not buggy

[ssg-aero]

Did you try with gcc>=11.2?

I think it's a minimum requirement:

• gcc parallel execution relies on tbb and newer versions of tbb requires gcc 11: see here
• while building the conda package version 12 of gcc is used by default.
• this solve the issue with std::reduce on Linux, I'd rather stick to reduce, whit is more efficient than std::accumulate see here and let opened perspectives for both CPU/GPU parallelization. : "Additional execution policies may be provided by a standard library implementation (possible future additions may include std::parallel::cuda and std::parallel::opencl)"

[adriendelsalle]

Yes I understand the perspective to get better performance using std::reduce. Maybe it would be a good thing to start a developer documentation to mention minimum version of gcc.

The binding of MeridionalGrid<T>::operator() still needs to be wrapped in a lambda. We could investigate later how to remove that limitation (returning a template parameter already works).

[ssg-aero]

Yes didn't find a workaround either for MeridionalGrid::operator(), so I've merged the PR