CUDA8.X + Visual C++ 2015 does not require "BOOST_NO_CXX11_NOEXCEPT"

[boostorg]

CUDA8.X + Visual C++ 2015 does not require "BOOST_NO_CXX11_NOEXCEPT"

[jzmaddock]

Oh :(

Unfortunately the original bug report that led to that define being set was reproduced using msvc as the host compiler - see https://svn.boost.org/trac/boost/ticket/13049.

[jzmaddock]

I don't have CUDA 8 installed at present, and have no great desire to "downgrade", I wonder can you:

• Reproduce the original bug report linked above?
• Does using #if (BOOST_CUDA_VERSION >= 8000000) && (BOOST_CUDA_VERSION < 8010000) && defined( __CUDA_ARCH__) fix both the original issue and your test case? In other words is the workaround required for device code only?

[Chachay]

• I could reproduce the original bug under my environment. Just I had to handle this issue at cmake option

• My case is related to boost::system::error_category::std_category::name which overrides std::error_category::name, and "BOOST_NO_CXX11_NOEXCEPT" is directly harm on this overrides. Making CUDA 8.0 pretend as 8.2 have a good result. What I suggested was only about compiler code (msvc or GCC...). But I understood that this issue is not so easy...

I didn't get the link between the original bug and BOOST_NO_CXX11_NOEXCEPT. Give me a time to digest..

---

At least, CLANG has no trouble with the original bug??
ComputationalRadiationPhysics/picongpu#2048 (comment)

[jzmaddock]

At least, CLANG has no trouble with the original bug??

Apparently so: actually I had no idea that clang was capable of compiling CUDA device code (as opposed to just working as a host compiler). I'm fairly sure that some more fixes will be needed to support that.

I've been able to reproduce the issue, and am stuck as we now have two different bug reports requiring completely opposite actions.

My suggestion to restrict the original fix to device code didn't work either - apparently the issue occurs inside the EDG compiler front end while splitting device and host code apart - which is evidently more than just a preprocessing phase.

CUDA-9 is fine BTW.

[Chachay]

For my side, no news.. and I stucked at the similar place (yet I do not have enough knowledge about EDG compiler)

Just switching to CUDA9 would be the easiest solution, but, because Windows CUDA 9's setup.exe does not work stably, if I think of setting up a CI environment on Appveyor, I prefer CUDA8... :S

[Chachay]

This issue is out of date