Private Variables in Offloaded Fortran Parallel Loop

Question

I am offloading code to a GPU using OpenMP 4.5. So far everything is working on the GPU, except when I try to make parallel sections with private variables that are allocated before I offload.

I am using gcc 7.2.0 and cuda 9.2.88. I am running on CentOS 7 and am compiling it with
gfortran ./testCode.F90 -fopenmp -o ./test

Here is a sample code:

#define LENGTH_X 4
#define LENGTH_Y 4
#define PRINT

program main
    use omp_lib
    implicit none
    real, allocatable :: testVar(:,:)
    real :: error = 0
    logical :: onCPU
    integer :: i, j,k

    allocate(testVar(LENGTH_X,LENGTH_Y))
    do i = 1, LENGTH_X
        testVar(i,:) = i
#ifdef PRINT
        print *, testVar(i,:)
#endif
    end do

    onCPU = omp_is_initial_device()

!$omp target map(tofrom:testVar, onCPU,error)
 !$OMP TEAMS DISTRIBUTE PARALLEL DO  private(testVar) reduction(max:error) 

    do i = 2, LENGTH_X-1
        do j = 2, LENGTH_Y-1
        testVar(i,j) = 0.25
        end do
    end do

 !$OMP END TEAMS DISTRIBUTE PARALLEL DO
    onCPU = omp_is_initial_device()
    !$omp end target

    print *, "Ran on CPU", onCPU
    print *, "New vars"
    do i = 1, LENGTH_X
#ifdef PRINT
        print *, testVar(i,:)
#endif
    end do
end program main

This fails to compile with

unresolved symbol _gfortran_os_error
collect2: error: ld returned 1 exit status
mkoffload: fatal error: x86_64-pc-linux-gnu-accel-nvptx-none-gcc returned 1 exit status
compilation terminated.
lto-wrapper: fatal error: /opt/software/GCC/7.2.0-cuda-9.2.88-offload/libexec/gcc/x86_64-pc-linux-gnu/7.2.0//accel/nvptx-none/mkoffload returned 1 exit status
compilation terminated.
/opt/software/binutils/2.28-GCCcore-6.4.0/bin/ld: error: lto-wrapper failed
collect2: error: ld returned 1 exit status

If I change the private to shared it works fine. I am not new to fortran but know how to program in C/C++ and python. Any advice would be appreciated!

It's not a compiler error, it's an error at link time. Can you tell us the compiler you are using, what version of the compiler, what operating system and what commands you are using to compile and link. — Ian Bush, Aug 08 '19 at 18:03

score 0 · Answer 1 · answered Oct 20 '20 at 12:25

Fortran allocatables may imply dynamic memory allocation (whether or not that is then actually done on the offloading device), and that is implemented via support routines in libgfortran. I suppose _gfortran_os_error would be called in case of a memory allocation error. Per https://gcc.gnu.org/PR90386 "Offloading: libgfortran, libm dependencies" you currently have to manually specify -foffload=-lgfortran to resolve such errors.

Private Variables in Offloaded Fortran Parallel Loop

1 Answers1