nvc++ errors with message "float128_type is undefined"

[davetaflin]

nVidia's HPC toolkit (https://developer.nvidia.com/hpc-sdk-downloads) supplies compiler nvc++, which defines macro __PGI. Using it to compiling a C++ file that #includes <boost/geometry/index/rtree.hpp> issues the following error message (this was with Boost 1.83.0, but the error began occurring with 1.79.0):

.../boost_1_83_0/boost/multiprecision/detail/number_base.hpp", line 205: error: identifier "float128_type" is undefined
struct bits_of<float128_type>

Tested operating system: Ubuntu 20.04 LTS
Compiler: nvc++ 23.7-0 64-bit target on x86-64 Linux -tp haswell

[DJDavies2]

Is there a solution for this? I am seeing this issue although the versions are slightly different: boost 1.86, nvidia 22.7.

[mborland]

Do you have a minimal reproducer? If multiprecision is being run in the CUDA Kernel there's no hope of it working because this library does not support being run in device code. The lines in question are extracted below:

#if defined(_GLIBCXX_USE_FLOAT128) && defined(BOOST_GCC) && !defined(__STRICT_ANSI__)
#define BOOST_MP_BITS_OF_FLOAT128_DEFINED
template <>
struct bits_of<float128_type>
{
   static constexpr unsigned value = 113;
};
#endif

If you're using CMake with GCC as the host compiler all of those conditions are being met.

[DJDavies2]

The code doesn't get as far as doing anything, the compilation fails just with the #include, e.g.:

`
#include <boost/multiprecision/float128.hpp>

int main()
{
}
`

[mborland]

This gets things compiling locally: #641. Boost.Config says that the NVC++/PGI compiler should not be providing any kind of Float128 support. The NVC++ documentation only mentions _Float128 in a bug fix so I assume this is still correct.

[DJDavies2]

Thanks, that fix seemed to work for me.

[mborland]

Thanks for the confirmation.