Error compiling with GPU accelerated OpenMP and Debug flag

[jeffrey.garretson]

We’re trying to run some fortran code using OpenMP and device offloading supported by nvfortran. Most of our code compiles and runs just fine, but a few of the files generate odd errors that won’t go away unless I completely remove the OpenMP directives. I’ve generated a simple test program that replicates the error:

module testmod

    use, intrinsic :: iso_fortran_env

    IMPLICIT NONE

    INTEGER, PARAMETER :: RPREC = REAL64
    INTEGER, PARAMETER :: IPREC = INT32

    INTEGER, PARAMETER :: ISIZE = 100
    INTEGER, PARAMETER :: JSIZE = 100
    INTEGER, PARAMETER :: KSIZE = 100
    INTEGER, PARAMETER :: ESIZE = 200

    REAL (rprec) :: bigArray1(isize, jsize, ksize, esize)
    REAL (rprec) :: storage(isize, jsize)

    CONTAINS

      SUBROUTINE testSubroutine ( )
      IMPLICIT NONE
      INTEGER (iprec) :: i, j, k
      REAL (rprec) :: offset

      storage (:,:) = 0.0_rprec

      do k=1,ksize

        !$OMP PARALLEL DO &
        !$OMP schedule(dynamic) &
        !$OMP DEFAULT (NONE) &
        !$OMP PRIVATE(i,j,offset) &
        !$OMP SHARED(k,jsize,isize,storage)
        DO  j = 1,jsize
          DO i = 1,isize
            offset = i+j
            storage (i, j) = storage (i, j) + offset
          ENDDO !i
        ENDDO !j

      enddo


      END SUBROUTINE testSubroutine

end module testmod

If I compile that code with this command:

nvfortran -mp=gpu -g -c testmod.F90

Where the code is in the file named testmod.F90, then it generates this output:

testsubroutine:
     29, !$omp parallel
<path>/bin/llc: error: <path>/bin/llc: <path>/nvfortranIWujMoeNmkBA.ll:338:75: error: expected unsigned integer
!75 = !DIExpression(DW_OP_plus_uconst, 0, DW_OP_deref, DW_OP_plus_uconst, -1094967168)

If I try to compile with “-mp” instead of “-mp=gpu”, or without the “-g” flag, then it compiles successfully. Also if I reduce the size of the arrays in the code, for example by reducing the “ESIZE” variable to 100, it will also compile successfully.

In our actual code we can’t reduce the size of the arrays, and we are using device offloading in the same file as code that isn’t offloaded. So we would like to be able to compile the non-offloaded code with debug info if there are any work-arounds possible.

Thanks for any assistance!

Jeff

[MatColgrove]

Hi Jeff,

Apologies for the late response, I somehow missed your post.

I’m able to recreate the error here and have filed a problem report, TPR #32377. My best guess is that it’s a problem with the DWARF (debug) code being generated with some type of overflow in the expression. No idea why it only occurs with -mp=gpu given you’re not using target directives and hence no GPU code will be created, but we’ll have a compiler engineer investigate.

Thanks for the report,
Mat

[jeffrey.garretson]

Mat,

No problem at all, I appreciate the response! Thanks for making the TPR, I’ll keep an eye on this post in the hope it gets patched in a future version.

Thanks,
Jeff

[MatColgrove]

Hi Jeff,

TPR #32377 should be resolved in the 22.9 release:

% nvfortran -mp=gpu -Minfo=mp test.f90 -c -g -V22.7
NVFORTRAN-S-0087-Non-constant expression where constant expression required (test.f90: 10)
  0 inform,   0 warnings,   1 severes, 0 fatal for p
% nvfortran -mp=gpu -Minfo=mp test.f90 -c -g -V22.9
%

-Mat