MPI_Send and MPI_Recv not found in "omp target is_device_ptr"-region

Accelerated Computing GPU-Accelerated Libraries
1

I’m trying to send and receive an array that is only allocated on the device:


#include  <mpi.h>
#include  <omp.h>

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

const size_t N = 16;

MPI_Init(NULL, NULL);
int rank, nranks;
MPI_Comm_rank (MPI_COMM_WORLD, &rank);
MPI_Comm_size (MPI_COMM_WORLD, &nranks);

int device_id = omp_get_default_device();

int *x = (int*)omp_target_alloc(N * sizeof(int), device_id);

#pragma omp target is_device_ptr(x)
{
   if (rank == 0) {
      for (int i = 0; i < N; i++) 
      {
         x[i] = i;
      }
      MPI_Send (x, N, MPI_INT, 1, 0, MPI_COMM_WORLD);
   } else {
      MPI_Recv (x, N, MPI_INT, 0, 0, MPI_COMM_WORLD, NULL);
   }
}

omp_target_free(x, device_id);}

I compile with the nvhpc-hpcx module from HPC SDK 25.7, using mpicc -mp=gpu. This results in


nvlink error   : Undefined reference to ‘MPI_Recv’ in ‘/tmp/nvcVHThpnC_h-wdA.o’
nvlink error   : Undefined reference to ‘MPI_Send’ in ‘/tmp/nvcVHThpnC_h-wdA.o’
pgacclnk: child process exit status 2: /home/cweiss/local/hpc_sdk/Linux_x86_64/25.1/compilers/bin/tools/nvdd

I have confirmed with the -v option that -lmpi is linked. Also, when I leave out the pragma, the test code compiles fine.

2

I recommend asking questions like this on the HPC SDK forum.