Work-Around: NVCC 11.4.0 - 11.8.0

[ax3l]

Adds a targeted NVCC work around for limited number of CUDA releases. Fixed in NVCC development.

Fix#4193

Description

Suggested changelog entry:

* work-around for Nvidia's CUDA nvcc compiler in versions 11.4.0 - 11.8.0

[henryiii]

Could we also bump the docker image used to build the cuda tests?

[ax3l]

Good point. Bumped.

[henryiii]

Thanks! I'd like to target a 2.10.1 release in the next (max) 2.5 weeks, hopefully sooner. (That's before 3.11, if you are curious to the selection of the max date ;) ).

[rwgk]

I think this PR is giving me trouble when updating the smart_holder branch (#4227). See below.

@ax3l did you try to add enable_if to the workaround? It looks like the current version is too generic. Now that I look at it more carefully, I'm actually surprised that this didn't bite on master.

https://github.com/pybind/pybind11/actions/runs/3207229735/jobs/5241897361

/__w/pybind11/pybind11/tests/test_class_sh_factory_constructors.cpp(94): error: more than one instance of overloaded function "pybind11::class_<type_, options...>::def [with type_=pybind11_tests::class_sh_factory_constructors::atyp_valu, options=<pybindit::memory::smart_holder>]" matches the argument list:
function template "pybind11::class_<pybind11_tests::class_sh_factory_constructors::atyp_valu, pybindit::memory::smart_holder> &pybind11::class_<type_, options...>::def(const char *, Func &&, const Extra &...) [with type_=pybind11_tests::class_sh_factory_constructors::atyp_valu, options=<pybindit::memory::smart_holder>]"
/__w/pybind11/pybind11/include/pybind11/pybind11.h(1787): here
function template "pybind11::class_<pybind11_tests::class_sh_factory_constructors::atyp_valu, pybindit::memory::smart_holder> &pybind11::class_<type_, options...>::def(const T &, const Extra &...) [with type_=pybind11_tests::class_sh_factory_constructors::atyp_valu, options=<pybindit::memory::smart_holder>]"
/__w/pybind11/pybind11/include/pybind11/pybind11.h(1816): here
argument types are: (const char [9], <unknown-type>)
object type is: pybind11::class_<pybind11_tests::class_sh_factory_constructors::atyp_valu, pybindit::memory::smart_holder>

• test_class_sh_factory_constructors.cpp(94):

pybind11/tests/test_class_sh_factory_constructors.cpp

Line 94 in 59ef530

.def("get_mtxt", get_mtxt<atyp_valu>);

• pybind11.h(1787):

 template <typename Func, typename... Extra>
class_ &def(const char *name_, Func &&f, const Extra &...extra) { <<<<<<<< 1787
cpp_function cf(method_adaptor<type>(std::forward<Func>(f)),
name(name_),
is_method(*this),
sibling(getattr(*this, name_, none())),
extra...);
add_class_method(*this, name_, cf);
return *this;
}

• pybind11.h(1816):

// Nvidia's NVCC is broken between 11.4.0 and 11.8.0
// https://github.com/pybind/pybind11/issues/4193
#if defined(__CUDACC__) && (__CUDACC_VER_MAJOR__ == 11) && (__CUDACC_VER_MINOR__ >= 4) \
&& (__CUDACC_VER_MINOR__ <= 8)
template <typename T, typename... Extra>
class_ &def(const T &op, const Extra &...extra) { <<<<<<<< 1816
op.execute(*this, extra...);
return *this;
}

[rwgk]

I made a simple reproducer for master: PR #4228

Evidently this workaround only trades one error for another.

I recommend rolling back this PR as soon as possible. (I'll do that on the smart_holder branch now, to unblock the update.)

My test is pretty small but can probably be reduced further.

[ax3l]

Continued in #4228

[henryiii]

Thanks! I'd like to target a 2.10.1 release in the next (max) 2.5 weeks, hopefully sooner. (That's before 3.11, if you are curious to the selection of the max date ;) ).

Missed by a week, but it's out now.