OpenMP dependency not found with Nvidia HPC SDK

[nordmoen]

Describe the bug
When using openmp = dependency('openmp') with Nvidia HPC SDK and Meson 0.56.0 OpenMP is not detected.

To Reproduce
Have openmp = dependency('openmp') and run CC=nvc meson setup builddir

Expected behavior
Expected OpenMP to be detected since it is supported by Nvidia HPC SDK according to their documentation.

I expect the issue to be that Nvidia (previously PGI) still does not define _OPENMP. In the code there is a special handling of pgi and that should probably be extended to nvidia_hpc. As per the link above the new Nvidia HPC SDK compilers should support version 4.5 (except for the offloading capabilities).

system parameters

• Meson 0.56.0
• Nvidia HPC SDK 20.7 and 20.9

[dcbaker]

@scivision I think you're the only one with access to the Nvidia compilers.

[nordmoen]

The Nvidia HPC SDK should be accessible to everyone as long as you accept Nvidia's EULA.

[scivision]

I don't have the Nvidia compiler anymore

[dcbaker]

I think then we're at "patches welcome". I'd be happy to review something though.

[nordmoen]

It seems like my initial assessment was wrong, the Nvidia HPC SDK does indeed support the _OPENMP macro, however, meson was unable to detect this due to incompatible flags used with nvc when preprocessing to detect _OPENMP. One solution is to add nvc flags to PGICompiler or create a new Nvidia HPC SDK mixin with the required flags.

The following preprocessor flags (the order is important) will result in nvc handling the _OPENMP macro and printing the expected result.

def get_preprocess_only_args(self) -> T.List[str]:
    return ['-Mcpp', '-E']

[nordmoen]

@dcbaker Could you or anyone else comment on the above proposed changes?

[dcbaker]

Does that still work with the PGI compiler, if it does just flipping it makes sense, otherwise some kind of inheritance based override make sense.

[nordmoen]

Does that still work with the PGI compiler, if it does just flipping it makes sense, otherwise some kind of inheritance based override make sense.

Yes, I tested with both PGI/19.9 and PGI/20.4 and they both output the expected OpenMP support when compiled with pgcc -Mcpp -E -mp (note that using pgcc -E -Mcpp -mp results in no output at all)

[nordmoen]

One challenge I find when testing a custom get_preprocess_only_args is that the command used to compile is not strictly nvc -Mcpp -E ..., but rather nvc -Mcpp -E -Mcpp ... which seems to be down to how compilers.py:compile() is building the command. Before compile() is called the extra_args parameter is built up to contain the flags we are after (in the correct order), but on line compilers.py:757 the command is extended with the arguments for the desired mode (preprocessor in our case) before the extra_args list is added. The result being that the preprocessor flags are attempted added twice to the command. If I comment out line compilers.py:757 the compile works as expected (however, I don't know what else this may lead to down the road).

Could anyone advice on what is the correct course so that the flags appear in the correct order?

[e-kwsm]

Let’s consider https://github.com/dftd4/dftd4, which requires OpenMP, as an example.

Meson fails to detect OpenMP with nvc and nvfortran:

$ CC=nvc FC=nvfortran meson setup --default-library static NVIDIA
…
Fortran compiler for the host machine: nvfortran (nvidia_hpc 22.1-0)
Fortran linker for the host machine: nvfortran pgi 22.1-0
Host machine cpu family: x86_64
Host machine cpu: x86_64
C compiler for the host machine: nvc (nvidia_hpc 22.1-0)
C linker for the host machine: nvc pgi 22.1-0
Run-time dependency OpenMP found: NO (tried system)

config/meson.build:42:2: ERROR: Dependency "openmp" not found, tried system

Then what happens if pgcc and pgfortran are used instead, which are symlinks to nvc/nvfortran? Meson works as expected:

$ CC=pgcc FC=pgfortran meson setup --default-library static PGI
…
Fortran compiler for the host machine: pgfortran (pgi 22.1-0)
Fortran linker for the host machine: pgfortran pgi 22.1-0
Host machine cpu family: x86_64
Host machine cpu: x86_64
C compiler for the host machine: pgcc (pgi 22.1-0)
C linker for the host machine: pgcc pgi 22.1-0
Run-time dependency OpenMP found: YES 3.1
…

So a workaround is to use old PGI names, but pgi staffs should be replaced with nvidia_hpcas @nordmoen pointed out.

---

• nvc 22.1-0
• meson 0.61.1

[RudiFeiman]

I met a similar problem.
Describe the bug
I'm trying buid a c++ project with nvidia HPC compiler with openmp . This project compiled with meson.
When I using default compiler (gcc c++) with openmp, it's OK.
But nvidia hpc compiler is failed.

To Reproduce
meson.build file:

project('nvidiaHPCtest', 'cpp')
ompdep = dependency('openmp')
executable('demo', 'main.cpp', dependencies : ompdep)

main.cpp file:

#include <iostream>
#include <omp.h>

int main(int argc, char **argv) {

#if defined(_OPENMP)
  std::cout << "openmp is defined\n";
#endif

  #pragma omp parallel
  {
    std::cout << _OPENMP <<"Hello there.\n";
  }
  
  return 0;
}

Expected behavior
gcc meson compile command:

1. meson setup builddir  
2. cd builddir
3. ninja

nvidia hpc compile command:
1. CC=/mnt/f/work/nvhpc/Linux_x86_64/23.5/compilers/bin/nvc CXX=/mnt/f/work/nvhpc/Linux_x86_64/23.5/compilers/bin/nvc++ meson setup builddir

It will failed with ERROR: Dependency “openmp” not found, tried system.

note: my nvidia hpc compiler installed in /mnt/f/work/nvhpc/Linux_x86_64/23.5/compilers/bin

system parameters

• This is a Windows wsl2 Ubuntu subsystem compile
• Ubuntu 22.04
• Python 3.10.6
• meson 0.61.2
• ninja 1.10.1

I have also post this issue on nvidia forums for some comments with nvidia hpc compiler.
https://forums.developer.nvidia.com/t/nvhpc-22-2-error-use-of-undefined-value-l-lb26-8163/205286/19