OffTarget

Observed issues › harvest/cce21-declare-target-link-module-variable-write-lost

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

ftn 21 loses a target region's write to a declare target link module variable mapped tofrom

Legality: independently confirmed. 2 independent readings find the program legal and the result required. How this is checked

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 24.3.0correctcorrectunknown
not run
correctcorrectn/an/a
CCE 19.0.0correctcorrectunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 20.0.2correctcorrectunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 21.0.3behaves incorrectlybehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 21.0.4behaves incorrectlybehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
NVHPC 25.11n/an/an/an/an/acorrectcorrect
NVHPC 26.9n/an/an/an/an/acorrectcorrect
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 incorrectlyCCE 21.0.3, MI300A unified
CCE 21.0.3, MI300A copy
CCE 21.0.4, MI300A unified
CCE 21.0.4, MI300A copy
show
v= 0 dev= T
RESULT ran_on_host=0 check=fail

Reproduce

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

AFAR 24.3.0, MI300A unified (runs correctly)

$AFAR/bin/amdflang -fopenmp --offload-arch=gfx942 -O2 repro.f90 -o repro
HSA_XNACK=1 OMP_TARGET_OFFLOAD=mandatory ./repro
Program (60 lines)
! # Finding: CCE 21 ftn loses the target region's write to a module
! # variable in a declare target link clause that the target construct maps
! # tofrom (CCE 19 and 20 correct)
!
! - compilers: CCE ftn 21.0.3 and 21.0.4, -O2: this file on MI300A gfx942
!   (HSA_XNACK 0 and 1); the two AOMP tests it comes from also on MI250X
!   gfx90a with 21.0.3.
! - expected: v= 1 dev= T
! - actual: v= 0 dev= T (the region runs on the device; the original v keeps
!   its value), every run, both memory modes.
! - rule: the link clause "does not map the list items. Instead, they are
!   mapped according to the data-mapping rules described in Section 5.8"
!   (5.2 §5.8.5 p.159 l.24-27), so map(tofrom: v) on the target construct
!   maps v with a reference count of one and on exit "the original list item
!   is updated as if the list item appeared in a from clause on a target
!   update directive" (5.2 §5.8.3 p.154 l.3-5). OpenMP 4.5 already has the
!   link clause.
! - regression: CCE 19.0.0 and 20.0.2 print v= 1 dev= T (gfx942 both modes;
!   gfx90a).
! - same-compiler controls (CCE 21.0.3 and 21.0.4 correct, v= 1 dev= T):
!   the same map on an enclosing target data construct (target data
!   map(tofrom: v), then target without a map of v); v without the declare
!   target directive. (With enter(v) instead of link(v) v= 0 is the expected
!   result: the variable is present for the whole program and the exit update
!   happens only at a reference count of one, 5.2 §5.8.3 p.154 l.3-5; AFAR and
!   CCE 21 print v= 0, CCE 19 and 20 reject enter.)
! - other families: AFAR amdflang prints v= 1 dev= T (gfx942 X0 and X1,
!   gfx90a).
! - harvest: dev programs h2-aomp-smoke-fort-declare-target-link-2 and
!   h2-aomp-fortran-map-tests-even-basicer-example (data/bench/harvest2; AOMP
!   test/smoke-fort/declare-target-link-2 and
!   test/fortran-map-tests/even-basicer-example.f95), wrong_answer on CCE 21
!   only, gate confirmed on gfx90a.
! - legality: offload.engine.prerun.check accepts it (the rules with the race
!   rule, and the AFAR flang front end).
! CCE:  ftn -fopenmp -O2 (module load cce/<v> craype-accel-amd-gfx942 rocm/<r>;
!       module unload cray-mpich)
! Run:  OMP_TARGET_OFFLOAD=mandatory ./a.out
module m
  implicit none
  integer :: v = 0
  !$omp declare target link(v)
end module m
program p
  use m
  use omp_lib
  implicit none
  logical :: dev
  dev = .false.
  !$omp target map(tofrom: v) map(from: dev)
  v = 1
  dev = .not. omp_is_initial_device()
  !$omp end target
  print '(a,i0,a,l1)', 'v= ', v, ' dev= ', dev
  if (v == 1) then
    print '(a,i0,a)', 'RESULT ran_on_host=', merge(0, 1, dev), ' check=pass'
  else
    print '(a,i0,a)', 'RESULT ran_on_host=', merge(0, 1, dev), ' check=fail'
  end if
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

Expected

v= 1 dev= T

Observed where it behaves incorrectly

v= 0 dev= T
RESULT ran_on_host=0 check=fail

What the specification requires

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

OpenMP 5.2 §5.8.5 p.159 l.24-27

does not map the list items. Instead, they are mapped according to the data-mapping rules described in Section 5.8

OpenMP 5.2 §5.8.3 p.154 l.3-5

the original list item is updated as if the list item appeared in a from clause on a target update directive

OpenMP 5.2 §5.8.3 p.154 l.3-5

Evidence

Other compilers run it correctly

AFAR 24.3.0, NVHPC 25.11, NVHPC 26.9, GCC 16.2.0