OffTarget

Observed issues › cce-defaultmap-tofrom-explicit-map

Appears to behave incorrectlyWrong answer · Fortran · OpenMP · Open · regression

With a defaultmap clause on target, an explicitly mapped map(tofrom) scalar is not copied back, and with defaultmap(tofrom) an array in map(alloc) or map(to) is copied back

Legality: not independently confirmed. no independent review yet How this is checked

Same issue as cce21-defaultmap-aggregate-flag; this program is another reproducer of it.

Affected variants

CompilerMI300A gfx942
unified memory (XNACK 1)
MI300A gfx942
copy (XNACK 0)
MI250X gfx90a
copy (XNACK 0)
MI300X gfx942
copy (XNACK 0)
MI350X gfx950
copy (XNACK 0)
A100 sm_80
separate memory
A100 sm_80
managed memory
AFAR 24.3.0behaves incorrectlycorrectcorrectcorrectcorrectn/an/a
CCE 19.0.0behaves incorrectlybehaves incorrectlybehaves incorrectlyunknown
not run
unknown
not run
n/an/a
CCE 20.0.2behaves incorrectlycorrectcorrectunknown
not run
unknown
not run
n/an/a
CCE 21.0.3behaves incorrectlybehaves incorrectlybehaves incorrectlyunknown
not run
unknown
not run
n/an/a
CCE 21.0.4behaves incorrectlybehaves incorrectlybehaves incorrectlyunknown
not run
unknown
not run
n/an/a
NVHPC 25.11n/an/an/an/an/acorrectbehaves incorrectly
NVHPC 26.9n/an/an/an/an/acorrectbehaves incorrectly
GCC 16.2.0correctcorrectcorrectcorrectcorrectcorrectn/a

"behaves incorrectly" / "not supported": this issue, confirmed in up to four environments (another build directory, another node, a repeat). "fails another way": a different failure, not counted as this issue. "unknown": not run there, or the runs disagree.

Messages

ResultCompiler and GPUMessage
behaves incorrectlyAFAR 24.3.0, MI300A unified
CCE 19.0.0, MI300A unified
CCE 20.0.2, MI300A unified
NVHPC 25.11, A100 managed
NVHPC 26.9, A100 managed
show
region 1: host_flag=0 host a unchanged: 0
region 2: host_flag=0 host a unchanged: 0
region 3: host_flag=0 host a unchanged: 0
region 4: host_flag=0 host a unchanged: 0
behaves incorrectlyCCE 19.0.0, MI300A copy
CCE 19.0.0, MI250X
show
region 1: host_flag=0 host a unchanged: 64
region 2: host_flag=0 host a unchanged: 0
region 3: host_flag=0 host a unchanged: 64
region 4: host_flag=0 host a unchanged: 0
behaves incorrectlyCCE 21.0.3, MI300A unified
CCE 21.0.4, MI300A unified
show
region 1: host_flag=0 host a unchanged: 0
region 2: host_flag=1 host a unchanged: 0
region 3: host_flag=0 host a unchanged: 0
region 4: host_flag=1 host a unchanged: 0
behaves incorrectlyCCE 21.0.3, MI300A copy
CCE 21.0.3, MI250X
CCE 21.0.4, MI300A copy
CCE 21.0.4, MI250X
show
region 1: host_flag=0 host a unchanged: 64
region 2: host_flag=1 host a unchanged: 0
region 3: host_flag=0 host a unchanged: 64
region 4: host_flag=1 host a unchanged: 0

Reproduce

AFAR 24.3.0, MI300A unified

$AFAR/bin/amdflang -fopenmp --offload-arch=gfx942 -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 19.0.0, MI300A unified

module load cce/19.0.0 craype-accel-amd-gfx942 rocm/6.4.3
module unload cray-mpich
ftn -fopenmp -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 20.0.2, MI300A unified

module load cce/20.0.2 craype-accel-amd-gfx942 rocm/6.4.3
module unload cray-mpich
ftn -fopenmp -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 21.0.3, MI300A unified

module load cce/21.0.3 craype-accel-amd-gfx942 rocm/7.2.0
module unload cray-mpich
ftn -fopenmp -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 21.0.4, MI300A unified

module load cce/21.0.4 craype-accel-amd-gfx942 rocm/7.2.0
module unload cray-mpich
ftn -fopenmp -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro

NVHPC 25.11, A100 managed

export LD_LIBRARY_PATH=$NVHPC_25_11/compilers/lib
nvfortran -mp=gpu -gpu=cc80 -gpu=mem:managed -O2 repro.f90 -o repro
OMP_TARGET_OFFLOAD=mandatory ./repro

NVHPC 26.9, A100 managed

export LD_LIBRARY_PATH=$NVHPC_26_9/compilers/lib
nvfortran -mp=gpu -gpu=cc80 -gpu=mem:managed -O2 repro.f90 -o repro
OMP_TARGET_OFFLOAD=mandatory ./repro

GCC 16.2.0, MI300A unified (runs correctly)

export LD_LIBRARY_PATH=$GCC_16_2/lib64:/opt/rocm-7.2.1/lib
$GCC_16_2/bin/gfortran -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx942 -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro
Program (85 lines)
! CCE 19 applies defaultmap(tofrom) to an array the
! same target construct maps EXPLICITLY with map(alloc:) or map(to:): on
! exit it copies the device values back to the host.
! Spec: on exit from the region the original list item is updated only "if
! the map-type is from or tofrom" (5.2 §5.8.3 p.154, exit step 1a); alloc
! and to have no from transfer. "The defaultmap clause determines the
! implicit data-mapping or data-sharing attribute of certain variables"
! (5.2 §5.8.7 p.161), and "explicit data-environment attribute clauses on
! target directives determine the attributes" of the variables they name
! (5.2 §5.8.1 p.148), so defaultmap(tofrom) does not change a's map type.
! Premise: the original and corresponding items do not share storage. The
! spec allows shared storage (5.2 §5.8.3 p.153, "may share storage"), so
! the evidence holds only where they are separate: run at HSA_XNACK=0 (on
! MI300A copy mode, on MI250X a discrete GPU), and the control regions 1
! and 3 (the same maps without defaultmap) must leave the host a unchanged
! (64 of 64): separate storage in the same configuration without
! defaultmap; for regions 2 and 4 separate storage is the premise.
! Expected: every region prints "host a unchanged: 64".
! Observed (2026-10-22, -O2, OMP_TARGET_OFFLOAD=mandatory, HSA_XNACK=0, two
! runs each, the same every run; <host> MI250X gfx90a and <host>
! MI300A gfx942):
! - CCE 19.0.0: regions 1 and 3: 64; regions 2 and 4
!   (with defaultmap(tofrom)), ran_on_host=0: 0 -- the device's writes
!   reached the host a.
! - CCE 20.0.2 and AFAR 24.0.0git: 64 in every region (correct).
! - CCE 21.0.3 and 21.0.4 : regions 2 and 4 print
!   ran_on_host=1, but that is the map(tofrom) host_flag, which CCE 21 does
!   not copy back in a region with defaultmap (cce21-defaultmap-aggregate-
!   flag.f90): the label `ran_on_host=` names no independent measurement,
!   and the regions run on the device. At HSA_XNACK=0 CCE 21.0.3 and 21.0.4
!   copy a back in regions 2 and 4 (host a unchanged: 0): this defect, fixed
!   in 20.0.2, is back in 21 (data/vendor/hpe/cce21-defaultmap-aggregate-
!   flag.md, variants: confirmed in the four environments on both releases).
! Found by the gfx90a map-mutant audit (data/reports/map-mutant-audit-
! gfx90a.md): the weaken_map and swap_to_from mutants of the defaultmap-
! tofrom programs survive on CCE 19.
! CCE:  ftn -fopenmp -O2 (module load cce/<v> craype-accel-amd-<gfx>
!       rocm/<r>; module unload cray-mpich)
! AFAR: amdflang -fopenmp --offload-arch=<gfx> -O2
! Run:  OMP_TARGET_OFFLOAD=mandatory HSA_XNACK=0 ./a.out
program p
  use omp_lib, only: omp_is_initial_device
  implicit none
  integer, parameter :: n = 64
  integer, allocatable :: a(:)
  integer :: i, k, host_flag
  do k = 1, 4
    allocate(a(n))
    a = [(i, i = 1, n)]
    host_flag = 1
    select case (k)
    case (1)
      !$omp target map(alloc: a) map(tofrom: host_flag)
        host_flag = merge(1, 0, omp_is_initial_device())
        do i = 1, n
          a(i) = -i
        end do
      !$omp end target
    case (2)
      !$omp target map(alloc: a) map(tofrom: host_flag) defaultmap(tofrom)
        host_flag = merge(1, 0, omp_is_initial_device())
        do i = 1, n
          a(i) = -i
        end do
      !$omp end target
    case (3)
      !$omp target map(to: a) map(tofrom: host_flag)
        host_flag = merge(1, 0, omp_is_initial_device())
        do i = 1, n
          a(i) = -i
        end do
      !$omp end target
    case (4)
      !$omp target map(to: a) map(tofrom: host_flag) defaultmap(tofrom)
        host_flag = merge(1, 0, omp_is_initial_device())
        do i = 1, n
          a(i) = -i
        end do
      !$omp end target
    end select
    print '(a,i0,a,i0,a,i0)', 'region ', k, ': ran_on_host=', host_flag, &
      ' host a unchanged: ', count(a == [(i, i = 1, n)])
    deallocate(a)
  end do
end program p

Source in the repository

Something to add? Dispute its legality · Report it fixed in a newer release · Report a mistake on this page

Expected and observed

Observed where it behaves incorrectly

region 1: host_flag=0 host a unchanged: 0
region 2: host_flag=0 host a unchanged: 0
region 3: host_flag=0 host a unchanged: 0
region 4: host_flag=0 host a unchanged: 0

What the specification requires

The specification text this program relies on and the compiler is expected to support.

OpenMP 5.2 §5.8.3 p.154

if the map-type is from or tofrom

OpenMP 5.2 §5.8.7 p.161

The defaultmap clause determines the implicit data-mapping or data-sharing attribute of certain variables

OpenMP 5.2 §5.8.1 p.148

OpenMP 5.2 §5.8.3 p.153

Evidence

Same compiler, one change

  • defaultmap(tofrom) removed from regions 2 and 4: AFAR 24.3.0: behaves incorrectly; AFAR 24.3.0: correct; CCE 19.0.0: behaves incorrectly; CCE 19.0.0: correct; CCE 20.0.2: behaves incorrectly; CCE 20.0.2: correct; CCE 21.0.3: behaves incorrectly; CCE 21.0.3: correct; CCE 21.0.4: behaves incorrectly; CCE 21.0.4: correct
    Control program
    program p
      use omp_lib, only: omp_is_initial_device
      implicit none
      integer, parameter :: n = 64
      integer, allocatable :: a(:)
      integer :: i, k, host_flag
      do k = 1, 4
        allocate(a(n))
        a = [(i, i = 1, n)]
        host_flag = 1
        select case (k)
        case (1)
          !$omp target map(alloc: a) map(tofrom: host_flag)
            host_flag = merge(1, 0, omp_is_initial_device())
            do i = 1, n
              a(i) = -i
            end do
          !$omp end target
        case (2)
          !$omp target map(alloc: a) map(tofrom: host_flag)
            host_flag = merge(1, 0, omp_is_initial_device())
            do i = 1, n
              a(i) = -i
            end do
          !$omp end target
        case (3)
          !$omp target map(to: a) map(tofrom: host_flag)
            host_flag = merge(1, 0, omp_is_initial_device())
            do i = 1, n
              a(i) = -i
            end do
          !$omp end target
        case (4)
          !$omp target map(to: a) map(tofrom: host_flag)
            host_flag = merge(1, 0, omp_is_initial_device())
            do i = 1, n
              a(i) = -i
            end do
          !$omp end target
        end select
        print '(a,i0,a,i0,a,i0)', 'region ', k, ': host_flag=', host_flag, &
          ' host a unchanged: ', count(a == [(i, i = 1, n)])
        deallocate(a)
      end do
    end program p
    

Other compilers run it correctly

GCC 16.2.0