Fix compiler error with MSVC 17 and CUDA 10.2

[SimeonEhrig]

Fixes the compiler error described in this issue: #2180

[atilag]

I can confirm that solved the problem (we were facing the same exact issue).
Would love to see this merged :)

[rwgk]

Looks good to me.
Could you please add a one-line comment to explain why the static_cast was inserted?
Maybe:
// static_cast works around compiler error with MSVC 17 and CUDA 10.2 (PR #2240).

[wjakob]

I'll merge this once you've had a chance to add an extra comment. (As we do in a few other places).

Side comment: compiling pybind11 code with nvcc sounds really scary (even if it's just for the "host" part).

[SimeonEhrig]

I added the comment.

Side comment: compiling pybind11 code with nvcc sounds really scary (even if it's just for the "host" part).

In our case, pybind11 is not the scary part to compile ;-) We use a highly templated C++ library which also supports gcc, clang and icc as host compiler for nvcc.

[rwgk]

Thanks!

[ax3l]

Thank you for digging into this and finding a work-around, Simeon!

I co-reported your reproducer as Nvidia Bug 3013788 so they have a chance to get their splitter/frontend into order for upcoming CUDA releases.

[wjakob]

Thanks!

[SimeonEhrig]

Thank you for digging into this and finding a work-around, Simeon!

I co-reported your reproducer as Nvidia Bug 3013788 so they have a chance to get their splitter/frontend into order for upcoming CUDA releases.

Thanks for reporting the bug.