Using OpenMP target offloading in llvm-8.0.0

Runtimes OpenMP
#1

Hello,

i came across this message on the list and found no better place to ask for help on a similar problem.
I am just experimenting offloading to x86_64 code from a x86_64 host system (ubuntu 18.04).
I have the packages:
- libomp-8-dev , containing among others :
   /usr/lib/llvm-8/lib/libomptarget.so
   /usr/lib/llvm-8/lib/libgomp.so
   /usr/lib/llvm-8/lib/libiomp5.so
   /usr/lib/llvm-8/lib/libomp.so
- clang-8

I compiled the same offload.c program as the one given in the referenced mail.

My first try was :
$ clang-8 -g -fopenmp -fopenmp-targets=x86_64-pc-linux-gnu offload.c

which did produce a binary a.out, but this one fails with the cryptic error message:
$ export LD_LIBRARY_PATH=/usr/lib/llvm-8/lib/
$ ./a.out
Libomptarget fatal error 1: default offloading policy must switched to mandatory or disabled

Second try, after reading some of the hints was
$ clang-8 -g -I//usr/lib/llvm-8/include/ -L/usr/lib/llvm-8/lib -Xopenmp-target -march=x86-64 -fopenmp -fopenmp-targets=x86_64-pc-linux-gnu offload.c

which resulted in the same error.
By the way, i have no clue why -Xopenmp-target -march=x86-64 would be required, which quite seem redundant with -fopenmp_targets.

I went to see the source code at
https://github.com/llvm-mirror/openmp/blob/master/libomptarget/src/interface.cpp line:56
but still can’t figure out what makes libomptarget unhappy.
In case this is related to ‘device’, i tested omp_get_default_device() which returns 1.

Any help would be appreciated.
Thanks.

#2

This looks like a bug in 8.0 that is fixed in trunk.
You have a `target teams distribute`, right?

Can you put an empty target region before that and see if it helps.

Cheers,
  Johannes

P.S.
This is the right place to ask these questions

#3

Thanks for the quick answer.

This looks like a bug in 8.0 that is fixed in trunk.
You have a `target teams distribute`, right?

Can you put an empty target region before that and see if it helps.

Yes, indeed the weird error message disappears after inserting a parallel target region,
that is, the following code works.

#include <omp.h>
#include <stdio.h>
int a[100];

int main() {

  printf("* number of devices: %d\n",omp_get_num_devices());
#pragma omp target
{
}

#pragma omp target teams distribute parallel for
  for(int i=0; i<100; i++)
        a[i]++;
  return 0;
}