Compilation with OmpSs-2 and OpenMP

[Z10Frank]

Hello,
I was exploring the compatibility with OpenMP with a simple test. I know that OpenMP compatibility is not fully supported, nor it is guaranteed in any case, but I am wondering if I am making some mistakes in the compilation.

My test:

#include <stdio.h>
#include <iostream>
#include <omp.h>
using namespace std;

int main(int argc, char *argv[])
{
    #pragma omp for 
    for (int i=0; i<10; i++)
    {
        cout<<"Hello omp, iteration "<<i<<"\n"<<endl;
    }

    for (int i=0; i<10; i++)
    {
        #pragma oss task firstprivate(i)
        {
            #pragma oss critical
            {
                cout<<"Hello OmpSs-2/, iteration "<<i<<endl;
            }
        }
    }
    return 0;
}

My compilation commands:

export CXX=icpc
export CC=icc
export MERCURIUM=/gpfs/users/massimof/mcxx/install/bin/
export PATH=$MERCURIUM:$PATH
 
I_MPI_CXX=mcxx MPICH_CXX=mcxx OMPI_CXX=mcxx mpiicpc --ompss-2 -O3 -std=c++11 -o test_OpenMP_OmpSs-2 --openmp-compatibility -fopenmp test_OpenMP_OmpSs-2.cpp -lz

In correspondence with the OpenMP pragma, an error occurs: error: this construct is not supported by Nanos6. Using a #pragma omp parallel for yields warning: explicit parallel regions do not have any effect in OmpSs.

Mercurium should have been compiled with OpenMP support:

export MERCURIUM=/gpfs/users/massimof/mcxx/install
export BISON=/gpfs/users/massimof/bison-2.6.5/install/bin/bison
export FLEX=/gpfs/users/massimof/flex-2.6.4/install/bin/flex
export GPERF=/gpfs/users/massimof/gperf-3.1/install/bin/gperf
export NANOSSIX=/gpfs/users/massimof/nanos6/install
export GIT=/gpfs/softs/spack/opt/spack/linux-centos7-cascadelake/gcc-9.2.0/git-2.25.0-lojbmyepxnaay2yotcnyzwwty36rnjyq/bin/git
export NANOSS=/gpfs/users/massimof/ompss-19.06/nanox-0.15/install

./configure BISON=$BISON FLEX=$FLEX GPERF=$GPERF GIT=$GIT --prefix=$MERCURIUM --enable-ompss-2 --with-nanos6=$NANOSSIX --enable-openmp --with-nanox=$NANOSS
make -j 10
make install

In the log file for the compilation of Mercurium it can be read:

checking if enabled OmpSs support in the compiler for Nanos++... no
checking if enabled OpenMP support in the compiler for Nanos++... yes
checking if enabled OmpSs-2 support in the compiler for Nanos6... yes
checking nanos.h usability... yes
checking nanos.h presence... yes
checking for nanos.h... yes
checking whether used installation of Nanos++ has instrumentation support... yes
...
checking nanos6.h usability... yes
checking nanos6.h presence... yes
checking for nanos6.h... yes
checking for nanos6_in_final in -lnanos6... yes

If I try to compile only with OmpSs-2 pragmas, everything works correctly. So, is there something missing in my compilation?

[rofirrim]

Hi @Z10Frank,

I think in this case you can use #pragma oss task for as described here

https://pm.bsc.es/ftp/ompss-2/doc/spec/directives/index.html#task-for-clause

Let me know if this is useful.

Kind regards,

[Z10Frank]

Hello,
thank you for your answer, I know the #pragma oss task for construct, but my purpose was rather to understand the compatibility with OpenMP itself.

The objective is to progressively "taskify" a load imbalanced part of an existing code that already uses OpenMP, in particular with parallel regions, omp for, omp simd. So a preliminary question is if the use of OmpSs-2 in this code would imply the substitution of all existing OpenMP pragmas with "equivalents" in OmpSs-2. Or if there is a way to switch between OpenMP and OmpSs-2 pragmas with an if condition.

[rofirrim]

The objective is to progressively "taskify" a load imbalanced part of an existing code that already uses OpenMP, in particular with parallel regions, omp for, omp simd. So a preliminary question is if the use of OmpSs-2 in this code would imply the substitution of all existing OpenMP pragmas with "equivalents" in OmpSs-2.

The OmpSs-2 compatibility with OpenMP is limited. It may happen that some constructs will require you to substitute/temporarily disable the unsupported OpenMP constructs.

Or if there is a way to switch between OpenMP and OmpSs-2 pragmas with an if condition.

Not sure what you mean by "if" condition here, one option is using the macro _OMPSS_2 so you can enable OpenMP or OmpSs-2 progressively in your code. This macro is enabled when you use --ompss-2 with Mercurium.

#if _OMPSS_2
#pragma oss task ...
#else 
#pragma omp ...
#endif

Hope this helps.

[Z10Frank]

Thank you!
To keep it in the archives, my test after the modifications:

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

using namespace std;

int main(int argc, char *argv[])
{
    #ifndef _OMPSS_2
    #pragma omp parallel
    #endif
    {

    #ifndef _OMPSS_2
    //with OpenMP
    #pragma omp for
    for (int i=0; i<10; i++)
    {
        #pragma omp critical
        cout<<"Hello omp for from thread "<<omp_get_thread_num()<<" iteration "<<i<<"\n"<<endl;
    }

    #else
    // with OmpSs-2
    for (int i=0; i<10; i++)
    {
        #pragma oss task firstprivate(i)
        {
        #pragma oss critical
        {
            cout<<"Hello OmpSs-2/, iteration "<<i<<endl;
        }
        }
    }
    #endif

    } // end parallel region if OpenMP is used
    return 0;

}

To compile with OpenMP:

mpiicpc -o test_OpenMP -fopenmp test_OpenMP_OmpSs-2.cpp

To compile with OmpSs-2:

I_MPI_CXX=mcxx MPICH_CXX=mcxx OMPI_CXX=mcxx mpiicpc -o test_OmpSs-2 --ompss-2 test_OpenMP_OmpSs-2.cpp