_LIBCUDACXX_CUDACC_VER major value needs to be computed as the seventh digit

[robertmaynard]

The introduction of an extra digit to handle minor CUDACC versions greater than 100 forgot to also increment the major value. Since this meant that all values of _LIBCUDACXX_CUDACC_VER stayed as 6 digits, but we compared against 7 digit values no CUDA versions reported decimal 128 support.

[jrhemstad]

Yeesh, math is hard. I think this is the 4th time we've had to correct this :)

Makes me feel like we should abandon trying to have a single representative number for the CUDA version. Is there an easier way to write <11.5 by just looking at the major/minor components?

#if defined(_LIBCUDACXX_COMPILER_MSVC) || (defined(_LIBCUDACXX_CUDACC_VER) && ((_LIBCUDACXX_CUDACC_VER_MAJOR < 11) ||  (_LIBCUDACXX_CUDACC_VER_MAJOR == 11 && _LIBCUDACXX_CUDACC_VER_MINOR < 5)) )
#  define _LIBCUDACXX_HAS_NO_INT128
#endif

[robertmaynard]

Yeesh, math is hard. I think this is the 4th time we've had to correct this :)

Makes me feel like we should abandon trying to have a single representative number for the CUDA version. Is there an easier way to write <11.5 by just looking at the major/minor components?

#if defined(_LIBCUDACXX_COMPILER_MSVC) || (defined(_LIBCUDACXX_CUDACC_VER) && ((_LIBCUDACXX_CUDACC_VER_MAJOR < 11) ||  (_LIBCUDACXX_CUDACC_VER_MAJOR == 11 && _LIBCUDACXX_CUDACC_VER_MINOR < 5)) )
#  define _LIBCUDACXX_HAS_NO_INT128
#endif

That is a possibility. I don't know what libcudacxx back-compat policy is, but we might need to keep _LIBCUDACXX_CUDACC_VER as a 6 digit number and use your check

[wmaxey]

Nobody should be using this value downstream. It's certainly not documented. It's purely an internal computation.

Visually I'd prefer the above solution, but I noticed a change I made elsewhere that has a competing idea.

from <cuda/std/chrono>:

#if defined(_LIBCUDACXX_COMPILER_NVCC)
#  if (CUDART_VERSION >= 11050)
#    pragma nv_diag_suppress cuda_demote_unsupported_floating_point
#  else
#    pragma diag_suppress cuda_demote_unsupported_floating_point
#  endif
#endif

[robertmaynard]

Nobody should be using this value downstream. It's certainly not documented. It's purely an internal computation.

Visually I'd prefer the above solution, but I noticed a change I made elsewhere that has a competing idea.

from <cuda/std/chrono>:

#if defined(_LIBCUDACXX_COMPILER_NVCC)
#  if (CUDART_VERSION >= 11050)
#    pragma nv_diag_suppress cuda_demote_unsupported_floating_point
#  else
#    pragma diag_suppress cuda_demote_unsupported_floating_point
#  endif
#endif

The current usage of CUDACC_VER_MAJOR / CUDACC_VER_MINOR captures both clang and nvcc, while moving over to _LIBCUDACXX_COMPILER_NVCC would drop clang. Now that actually might be desired as I don't know if clang + cuda 11.5 is sufficient for decimal 128 support.

[wmaxey]

Right, I would be wary of presuming it for other compilers. I'm happy to stick with the current solution.