Support NVidia HPC SDK (PGI/NVCC) on Ubuntu/Debian Systems

[nachovizzo]

As far as I know, you've introduced support for the PGI compilers in #2475.

But I fail to build pybind11 with the PGI compilers. gcc and clang works in my case.

I also tried to reproduce the steps on the CI/CD scripts, but this looks to be running CentOS and not a Debian based distribution.

The compilers I have on my system are
pgi:

pgcc (aka nvc) 20.7-0 LLVM 64-bit target on x86-64 Linux -tp skylake 
PGI Compilers and Tools
Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.

pgc++:

pgc++ (aka nvc++) 20.7-0 LLVM 64-bit target on x86-64 Linux -tp skylake 
PGI Compilers and Tools
Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.

After downloading eigen from source -DDOWNLOAD_EIGEN=ON and following applying this patch I managed to "build more" stuff, but not everything:

[  4%] Built target cross_module_gil_utils
[  8%] Built target external_module
[ 14%] Built target test_embed
[ 18%] Built target pybind11_cross_module_tests
[ 20%] Building CXX object tests/CMakeFiles/pybind11_tests.dir/test_eigen.cpp.o
At end of source: warning: routine is both "inline" and "noinline"

NVC++-F-0000-Internal compiler error. msz_dtype, bad value    2554  (/home/ivizzo/dev/libs/pybind11/tests/test_eigen.cpp: 464)
NVC++/x86-64 Linux 20.7-0: compilation aborted
make[2]: *** [tests/CMakeFiles/pybind11_tests.dir/build.make:219: tests/CMakeFiles/pybind11_tests.dir/test_eigen.cpp.o] Error 2
make[1]: *** [CMakeFiles/Makefile2:291: tests/CMakeFiles/pybind11_tests.dir/all] Error 2
make: *** [Makefile:130: all] Error 2

And the build line is:

cd /home/ivizzo/dev/libs/pybind11/build/tests && /opt/nvidia/hpc_sdk/Linux_x86_64/2020/compilers/bin/pgc++  -DPYBIND11_TEST_BOOST -DPYBIND11_TEST_EIGEN -Dpybind11_tests_EXPORTS -I/home/ivizzo/dev/libs/pybind11/include -I/usr/include/python3.8 -I/home/ivizzo/dev/libs/pybind11/build/_deps/eigen-src  -O2 -s -DNDEBUG -fPIC   --c++14 -A -o CMakeFiles/pybind11_tests.dir/test_eigen.cpp.o -c /home/ivizzo/dev/libs/pybind11/tests/test_eigen.cpp
At end of source: warning: routine is both "inline" and "noinline"

NVC++-F-0000-Internal compiler error. msz_dtype, bad value    2554  (/home/ivizzo/dev/libs/pybind11/tests/test_eigen.cpp: 464)
NVC++/x86-64 Linux 20.7-0: compilation aborted
make[2]: *** [tests/CMakeFiles/pybind11_tests.dir/build.make:219: tests/CMakeFiles/pybind11_tests.dir/test_eigen.cpp.o] Error 2
make[2]: Leaving directory '/home/ivizzo/dev/libs/pybind11/build'
make[1]: *** [CMakeFiles/Makefile2:291: tests/CMakeFiles/pybind11_tests.dir/all] Error 2
make[1]: Leaving directory '/home/ivizzo/dev/libs/pybind11/build'
make: *** [Makefile:130: all] Error 2

I've also tried skipping the pybind11_tests and installing the header files, but when I try to build a simple application I get these type errors on my build application:

"/usr/local/include/pybind11/cast.h", line 2196: error: no instance of
          overloaded function "pybind11::detail::collect_arguments" matches the
          argument list
            argument types are: (const char *const)
      return detail::collect_arguments<policy>(std::forward<Args>(args)...).call(derived().ptr());
             ^
          detected during:
            instantiation of "pybind11::object
                      pybind11::detail::object_api<Derived>::operator()(Args
                      &&...) const [with
                      Derived=pybind11::detail::accessor<pybind11::detail::acce
                      ssor_policies::str_attr>,
                      policy=pybind11::return_value_policy::automatic_reference
                      , Args=<const char *const &>]" at line 1573 of
                      "/usr/local/include/pybind11/pytypes.h"
            instantiation of "bool
                      pybind11::detail::object_api<Derived>::contains(T &&)
                      const [with Derived=pybind11::handle, T=const char *const
                      &]" at line 260 of
                      "/usr/local/include/pybind11/detail/internals.h"

Any idea on how can I make this work? Thanks in advance

[YannickJadoul]

Pinging @henryiii & @andriish, who probably know best.

[andriish]

Hi all ,
the error looks quite cryptic to me. Cannot say much more. But if that helps, here is some extra information

• one can try to use the older PGI compilers and/or RedHat system
• not that the actuall combination of stl headers + PGI version matters.
• 20.7 from HPC SDK has problems on some systems with the setup ... The developers working now to fix it.
• If the error is actually a compiler problem, that should be reported to PGI.
• not sure if that patch for Eigen is needed and I'm also not sure if the PGI identification is __PGI and not __PGI__.
  Hard to say much more about the error w/o the actual code.

Best regards

Andrii

[nachovizzo]

Hi, I'm glad that I'm not the only one who thinks this is cryptic.

I managed to build my application by only using the 2014 C++ standard(aka --c++14 in pgi compilers). I also tried 11 and 17 but it didn't work out in my case.

Thanks for the hints and for replying! I'm still not sure if PGI compilers can build much more stuff than just the NVIDIA examples. Let's see how this improves in the future!

Thanks again!