OffTarget

Observed issues › harvest/cce-use-device-ptr-c-ptr-not-converted

Appears to behave incorrectlyWrong answer · Fortran · OpenMP · Open

use_device_ptr does not convert a C_PTR list item: the region sees the host address

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 24.3.0behaves incorrectlycorrectunknown
not run
correctcorrectn/an/a
CCE 19.0.0correctbehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 20.0.2correctbehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 21.0.3correctbehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
CCE 21.0.4correctbehaves incorrectlyunknown
not run
unknown
not run
unknown
not run
n/an/a
NVHPC 25.11n/an/an/an/an/afails another wayfails another way
NVHPC 26.9n/an/an/an/an/afails another wayfails another way
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
show
host   555555B61EC0
mapped 0
seen   555555B61EC0
loc(cp) 5555556200C0
RESULT ran_on_host=0 check=fail
behaves incorrectlyCCE 19.0.0, MI300A copy
show
host   48C040
mapped 155544D00000
seen   48C040
loc(cp) 408740
RESULT ran_on_host=0 check=fail
behaves incorrectlyCCE 20.0.2, MI300A copy
show
host   48C2C0
mapped 155544D00000
seen   48C2C0
loc(cp) 408780
RESULT ran_on_host=0 check=fail
behaves incorrectlyCCE 21.0.3, MI300A copy
CCE 21.0.4, MI300A copy
show
host   494480
mapped 155544200000
seen   494480
loc(cp) 4087C0
RESULT ran_on_host=0 check=fail
fails another wayNVHPC 25.11, A100 separate
show
Present table dump for device[1]: NVIDIA Tesla GPU 0, compute capability 8.0, threadid=1
Hint: specify 0x800 bit in NV_ACC_DEBUG for verbose info.
host:0x308d1120 device:0x792a792fa000 size:4096 presentcount:0+1 line:44 name:tab(:)
allocated block device:0x792a792fa000 size:4096 thread:1
FATAL ERROR: data in use_device clause was not found on device 1: host:0x413e88
fails another wayNVHPC 25.11, A100 managed
show
FATAL ERROR: data in use_device clause was not found on device 1: host:0x413e88
fails another wayNVHPC 26.9, A100 separate
show
Present table dump for device[1]: NVIDIA Tesla GPU 0, compute capability 8.0, threadid=1
Hint: specify 0x800 bit in NV_ACC_DEBUG for verbose info.
host:0xf262130 device:0x71e5ab2fa000 size:4096 presentcount:0+1 line:44 name:tab(:)
allocated block device:0x71e5ab2fa000 size:4096 thread:1
FATAL ERROR: data in use_device clause was not found on device 1: host:0x41a888
fails another wayNVHPC 26.9, A100 managed
show
FATAL ERROR: data in use_device clause was not found on device 1: host:0x41a888

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 copy

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=0 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 20.0.2, MI300A copy

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=0 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 21.0.3, MI300A copy

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=0 OMP_TARGET_OFFLOAD=mandatory ./repro

CCE 21.0.4, MI300A copy

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=0 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 (61 lines)
! # Finding (known-issues harvest): CCE does not convert a C_PTR list item of
!   use_device_ptr: the region sees the host address
!
! - source: the NVIDIA forum topic 290317 (nvfortran dropping use_device_addr on an allocatable,
!   TPR 36867) gave data/bench/harvest-known hk-040, which compares the use_device_addr address of an
!   allocatable with the use_device_ptr address of a C_PTR to it; this file is the use_device_ptr
!   half (data/reports/harvest-known/v040-ptr-print.f90)
! - compilers: CCE 19.0.0, 20.0.2, 21.0.3 (ftn -fopenmp), copy mode: gfx90a (Tioga, HSA_XNACK=0)
!   and gfx942 at HSA_XNACK=0 (MI300A); at HSA_XNACK=1 on MI300A the mapped address is the host
!   address, so the missing conversion is not visible (check=pass)
! - expected: seen equals mapped (the device address of tab), RESULT ... check=pass
! - actual: seen equals host: e.g. CCE 19 "host 48BEC0 mapped 155544200000 seen 48BEC0", the same
!   on 20 and 21; RESULT ... check=fail. A device kernel given this pointer would read host memory.
!   The observation does not rest on omp_get_mapped_ptr (a routine of the compiler under test):
!   in copy mode device memory is separate, so seen = the host address is wrong whatever it returns
! - control (same compiler): use_device_addr(tab) on the allocatable itself gives the mapped
!   address (data/reports/harvest-known/v040-addr.f90, check=pass on CCE 19/20/21)
! - other families: AFAR converts it (seen = mapped); NVHPC 25.11 and 26.9 abort at the construct
!   (harvest/nvhpc-use-device-ptr-c-ptr-not-found.f90)
!   This file uses omp_get_mapped_ptr as the reference: AFAR at HSA_XNACK=1 (MI300A) returns a
!   null pointer for it, the known defect of afar-xnack1-target-is-present.f90 (nothing is
!   recorded present in unified memory), so that cell fails for another reason
! - legality: "If a list item that appears in a use_device_ptr clause is of type C_PTR and
!   points to a data entity that is mapped to the device data environment, references to the list
!   item in the structured block ... are converted into references to a device pointer that is
!   local to the structured block and that refers to the device address of the corresponding
!   entity" (OpenMP 5.2 §5.4.8 p.122 l.1-5); tab is mapped by target enter data and cp =
!   c_loc(tab); the C_PTR form is the one 5.2 keeps (the non-C_PTR form is deprecated, l.8-10).
!   omp_get_mapped_ptr is OpenMP 5.1 (§3.8.11); the pre-run check accepts the program
! - evidence: results/harvest-known/main.tioga (hk-040, gate: confirmed); the MI250X probe runs' output
!   (results/harvest-known/logs/tioga/probe.f656gDkdAd9R.log and probe.f656k2mjQdoD.log; their rows file
!   probes.tioga.jsonl was not kept); results/harvest-known/probes.tuolumne.jsonl (MI300A, v040-ptr.f90:
!   fail at HSA_XNACK=0, pass at 1)
program v040pp
  use iso_c_binding
  use omp_lib
  implicit none
  integer, parameter :: n = 1024
  integer, allocatable, target :: tab(:)
  type(c_ptr), target :: cp
  type(c_ptr) :: p_ptr, p_map
  integer :: hk_host
  allocate(tab(n))
  tab = 7
  !$omp target enter data map(to: tab)
  p_map = omp_get_mapped_ptr(c_loc(tab), omp_get_default_device())
  cp = c_loc(tab)
  !$omp target data use_device_ptr(cp)
  p_ptr = cp
  !$omp end target data
  !$omp target exit data map(delete: tab)
  print '(a,z0)', 'host   ', transfer(c_loc(tab), 0_c_intptr_t)
  print '(a,z0)', 'mapped ', transfer(p_map, 0_c_intptr_t)
  print '(a,z0)', 'seen   ', transfer(p_ptr, 0_c_intptr_t)
  print '(a,z0)', 'loc(cp) ', transfer(c_loc(cp), 0_c_intptr_t)
  hk_host = 1
  !$omp target map(tofrom: hk_host)
  if (.not. omp_is_initial_device()) hk_host = 0
  !$omp end target
  print '(a,i0,a,a)', 'RESULT ran_on_host=', hk_host, ' check=', merge('pass', 'fail', c_associated(p_ptr, p_map))
end program v040pp

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

check=pass

Observed where it behaves incorrectly

host   555555B61EC0
mapped 0
seen   555555B61EC0
loc(cp) 5555556200C0
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 §5.4.8 p.122 l.1-5

If a list item that appears in a use_device_ptr clause is of type C_PTR and points to a data entity that is mapped to the device data environment, references to the list item in the structured block ... are converted into references to a device pointer that is local to the structured block and that refers to the device address of the corresponding entity

OpenMP 5.1 §3.8.11

Evidence

Other compilers run it correctly

GCC 16.2.0