Can't build simple OpenMP offloading programs (needs -no-pie)

[Mehdi Chinoune]

$cat prod.f90
program main
  implicit none
  integer, parameter :: sp = selected_real_kind(6,37)
  integer, parameter :: n = 4096
  real(sp), allocatable :: a(:), b(:), c(:), c2(:)
  integer :: i

  allocate( a(n), b(n), c(n) )

  c = 0._sp
  call random_number(a)
  call random_number(b)

  !$omp target teams map(to: a, b) map(from: c)
  !$omp distribute parallel do
  do i = 1, n
    c(i) = a(i)*b(i)
  end do
  !$omp end target teams

  c2 = a*b

  print*, maxval( abs(c2-c) )

end program

$ gfortran -O3 -flto -fopenmp -foffload=nvptx-none prod.f90 -o test.x
/usr/bin/ld: /tmp/ccyvMF16.crtoffloadtable.o: warning: relocation against `__offload_vars_end' in read-only section `.rodata'
/usr/bin/ld: warning: creating DT_TEXTREL in a PIE

Ubuntu 20.10 x86_64
Xeon E3-1505m-v6
Nvidia Quadro M2200

That doesn't happen on Ubuntu 20.04

[Mehdi Chinoune (mchinoune)]

Building with "-no-pie" resolves the problem.

[Mehdi Chinoune (mchinoune)]

It's getting worse on Ubuntu 21.04

$ gfortran -fopenmp matmul.f90 -o test.x
lto-wrapper: fatal error: could not find accel/amdgcn-amdhsa/mkoffload in /usr/lib/gcc/x86_64-linux-gnu/10/:/usr/lib/gcc/x86_64-linux-gnu/10/:/usr/lib/gcc/x86_64-linux-gnu/:/usr/lib/gcc/x86_64-linux-gnu/10/:/usr/lib/gcc/x86_64-linux-gnu/ (consider using ‘-B’)
compilation terminated.
/usr/bin/ld: error: lto-wrapper failed
collect2: error: ld returned 1 exit status

I reported that before but someone has closed it.

$ gfortran -fopenmp -foffload=nvptx-none matmul.f90 -o test.x
/usr/bin/ld: /tmp/ccvGeRgU.crtoffloadtable.o: warning: relocation against `__offload_vars_end' in read-only section `.rodata'
/usr/bin/ld: warning: creating DT_TEXTREL in a PIE

$ gfortran -fopenmp -foffload=nvptx-none -no-pie matmul.f90 -o test.x
$ ./test.x 1024
 dim1 = dim2 = 1024
 MATMUL Time = 6.80000037E-02

libgomp: while loading libgomp-plugin-gcn.so.1: libgomp-plugin-gcn.so.1: cannot open shared object file: No such file or directory
 Offloading Time = 11.8520002
 Max(|C2-C1|) = 5.79833984E-04
$ ./test.x 1024
 dim1 = dim2 = 1024
 MATMUL Time = 6.30000010E-02

libgomp: while loading libgomp-plugin-gcn.so.1: libgomp-plugin-gcn.so.1: cannot open shared object file: No such file or directory

libgomp: cuCtxSynchronize error: the launch timed out and was terminated

libgomp: cuMemFree_v2 error: the launch timed out and was terminated

libgomp: device finalization failed

[Mehdi Chinoune (mchinoune)]

No need to pass -no-pie with gcc-11 on Ubuntu 21.10, but still throwing some warnings:

/usr/bin/ld: /tmp/ccpIxsYV.crtoffloadtable.o: warning: relocation against `__offload_vars_end' in read-only section `.rodata'
/usr/bin/ld: warning: creating DT_TEXTREL in a PIE

Still stop running with big size matrices:

 ./test.x 1024
 dim1 = dim2 = 1024
 MATMUL Time = 0.12400000000000000

libgomp: cuCtxSynchronize error: the launch timed out and was terminated

libgomp: cuMemFree_v2 error: the launch timed out and was terminated

libgomp: device finalization failed

[Mehdi Chinoune (mchinoune)]

No will to fix it.
Close

[Mehdi Chinoune (mchinoune)]

No will to fix it
Close