nvfortran -mp=gpu: write to a declare target module scalar in a target region is lost (implicit map)

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

Affected variants

Compiler	MI300A 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 24.3.0	correct	correct	unknown not run	correct	correct	n/a	n/a
CCE 19.0.0	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 20.0.2	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 21.0.3	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
CCE 21.0.4	correct	correct	unknown not run	unknown not run	unknown not run	n/a	n/a
NVHPC 25.11	n/a	n/a	n/a	n/a	n/a	behaves incorrectly	behaves incorrectly
NVHPC 26.9	n/a	n/a	n/a	n/a	n/a	behaves incorrectly	behaves incorrectly
GCC 16.2.0	correct	correct	correct	correct	correct	correct	n/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

Result	Compiler and GPU	Message
behaves incorrectly	NVHPC 25.11, A100 separate NVHPC 25.11, A100 managed NVHPC 26.9, A100 separate NVHPC 26.9, A100 managed	show g= 101 dev= 1

Reproduce

NVHPC 25.11, A100 separate

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

NVHPC 26.9, A100 separate

export LD_LIBRARY_PATH=$NVHPC_26_9/compilers/lib
nvfortran -mp=gpu -gpu=cc80 -gpu=mem:separate -O2 repro.f90 -o repro
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 (38 lines)

! # Finding: NVHPC loses the device write to a declare target scalar
!
! - compilers: NVHPC nvfortran 25.11 and 26.9, A100 (sm_80), -mp=gpu -gpu=cc80, separate and
!   managed memory, -O2
! - expected: g= 203 dev= 1. g appears in an enter clause (the 5.0 declare target(list) form). A
!   variable in an enter clause referenced in a target construct is treated as map(tofrom) (5.2
!   §5.8.1 p.148 l.9-11); its corresponding list item is never removed, "as if its reference count
!   was initialized to positive infinity" (§5.8.4 p.159 l.10-14), so on entry and exit there is no
!   transfer (§5.8.3 p.153 l.1-17, p.153 l.31-35, p.154 l.1-11): the region reads the device copy set by
!   "target update to" (101), writes 203 to it, and "target update from" returns 203
! - actual: g= 101 dev= 1 on both releases and both memory modes
! - control (same compiler): the same program with an explicit map(tofrom: dev, g)
!   (data/reports/fuzz-stress-audit/r4-declare-target-scalar-control.f90): 203 on both releases and
!   modes. (CCE 21 fails that control: stress/cce21-declare-target-scalar-map-write-lost.f90.)
! - other families: AFAR and CCE 19/20/21 give 203 on this program (MI300A HSA_XNACK 1 and 0)
! - legality: legal by reading: a module variable, no parallelism in the region
! - evidence: results/stress/audit/runs-nvhpc.jsonl (tools/jobs/run_wingtip_fix1.sh). The confirmer gave
!   needs_human (R22, no cross-machine witness); the audit counts the MI300A results
module dt_scalar_m
  implicit none
  integer(8) :: g
  !$omp declare target(g)
end module dt_scalar_m
program declare_target_scalar
  use omp_lib
  use dt_scalar_m
  implicit none
  integer :: dev
  g = 101
  dev = -1
  !$omp target update to(g)
  !$omp target map(tofrom: dev)
    g = 2 * g + 1
    dev = merge(0, 1, omp_is_initial_device())
  !$omp end target
  !$omp target update from(g)
  print '(a,i0,a,i0)', 'g= ', g, ' dev= ', dev
end program declare_target_scalar

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

g= 203 dev= 1

Observed where it behaves incorrectly

g= 101 dev= 1

What the specification requires

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

OpenMP 5.2 §5.8.1 p.148 l.9-11

OpenMP 5.2 §5.8.4 p.159 l.10-14

as if its reference count was initialized to positive infinity

OpenMP 5.2 §5.8.3 p.153 l.1-17

Evidence

Same compiler, one change

g mapped explicitly: map(tofrom: dev, g): AFAR 24.3.0: correct; CCE 19.0.0: correct; CCE 20.0.2: correct; CCE 21.0.3: behaves incorrectly; CCE 21.0.4: behaves incorrectly; NVHPC 25.11: correct; NVHPC 26.9: correct

Control program

module dt_scalar_m
  implicit none
  integer(8) :: g
  !$omp declare target(g)
end module dt_scalar_m
program declare_target_scalar
  use omp_lib
  use dt_scalar_m
  implicit none
  integer :: dev
  g = 101
  dev = -1
  !$omp target update to(g)
  !$omp target map(tofrom: dev, g)
    g = 2 * g + 1
    dev = merge(0, 1, omp_is_initial_device())
  !$omp end target
  !$omp target update from(g)
  print '(a,i0,a,i0)', 'g= ', g, ' dev= ', dev
end program declare_target_scalar

Other compilers run it correctly

AFAR 24.3.0, CCE 19.0.0, CCE 20.0.2, CCE 21.0.3, CCE 21.0.4, GCC 16.2.0