Scalar with the TARGET attribute is treated as firstprivate in a target region (device write lost)

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	behaves incorrectly	behaves incorrectly	unknown not run	behaves incorrectly	behaves incorrectly	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	AFAR 24.3.0, MI300A unified AFAR 24.3.0, MI300A copy AFAR 24.3.0, MI300X AFAR 24.3.0, MI350X NVHPC 25.11, A100 separate NVHPC 25.11, A100 managed NVHPC 26.9, A100 separate NVHPC 26.9, A100 managed	show rd= 101 s= 101 dev= 1

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

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

CCE 19.0.0, MI300A unified (runs correctly)

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

Program (35 lines)

! # Finding: NVHPC treats a TARGET scalar as firstprivate in a target region
!
! - compilers: NVHPC nvfortran 25.11 and 26.9, A100 (sm_80), -mp=gpu -gpu=cc80, separate
!   (-gpu=mem:separate) and managed (-gpu=mem:managed) memory, -O2
! - expected: rd= 101 s= 203 dev= 1. A scalar with the TARGET attribute referenced in a target region
!   "is treated as if it had appeared in a map clause with a map-type of tofrom" (5.2 §5.8.1 p.149
!   l.6-7; 5.0 §2.19.7 p.315 l.7-8; 6.0 §7.9.3 p.277 l.17-18), so the write s = 2*s+1 returns 203
! - actual: rd= 101 s= 101 dev= 1 on both releases and both memory modes (the region ran on the
!   device and read s; its write was not copied back: s behaves as firstprivate)
! - control (same compiler): the same program with map(tofrom: rd, dev, s)
!   (data/reports/fuzz-stress-audit/r1-target-scalar-control.f90): 203 on both releases and modes
! - other families: CCE 19/20/21 give 203 on the same program (MI300A HSA_XNACK 1 and 0; MI250X).
!   AFAR fails the same way: stress/afar-target-scalar-implicit-firstprivate.f90 (sibling entry)
! - legality: legal by reading: one initial thread in the region; rd and dev mapped tofrom, s
!   implicitly tofrom; no data race
! - evidence: results/stress/audit/runs-nvhpc.jsonl (tools/jobs/run_wingtip_fix1.sh, tools/campaigns/stress_audit.py
!   on data/reports/fuzz-stress-audit/r1-target-scalar*.f90). The confirmer gave needs_human (R22): it
!   does not draw a witness across machines; the audit counts the executed MI300A results
!   as the other-family witness for this deterministic single-construct program
program target_scalar_tofrom
  use omp_lib
  implicit none
  integer(8), target :: s
  integer(8) :: rd
  integer :: dev
  s = 101
  rd = -1
  dev = -1
  !$omp target map(tofrom: rd, dev)
    rd = s
    s = 2 * s + 1
    dev = merge(0, 1, omp_is_initial_device())
  !$omp end target
  print '(a,i0,a,i0,a,i0)', 'rd= ', rd, ' s= ', s, ' dev= ', dev
end program target_scalar_tofrom

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

rd= 101 s= 203 dev= 1

Observed where it behaves incorrectly

rd= 101 s= 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.149 l.6-7

is treated as if it had appeared in a map clause with a map-type of tofrom

OpenMP 5.0 §2.19.7 p.315 l.7-8

OpenMP 6.0 §7.9.3 p.277 l.17-18

Evidence

Same compiler, one change

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

Control program

program target_scalar_tofrom
  use omp_lib
  implicit none
  integer(8), target :: s
  integer(8) :: rd
  integer :: dev
  s = 101
  rd = -1
  dev = -1
  !$omp target map(tofrom: rd, dev, s)
    rd = s
    s = 2 * s + 1
    dev = merge(0, 1, omp_is_initial_device())
  !$omp end target
  print '(a,i0,a,i0,a,i0)', 'rd= ', rd, ' s= ', s, ' dev= ', dev
end program target_scalar_tofrom

Other compilers run it correctly

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