https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109837
Bug ID: 109837
Summary: [OpenMP] despite 'requires unified_address' there is
segfault when 'is_device_ptr' is not used
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: jakub at gcc dot gnu.org
Target Milestone: ---
Cf. https://github.com/SOLLVE/sollve_vv/pull/734 for a testcase.
For nvptx, we currently have:
GOMP_OFFLOAD_get_num_devices (unsigned int omp_requires_mask)
...
&& ((omp_requires_mask
& ~(GOMP_REQUIRES_UNIFIED_ADDRESS
| GOMP_REQUIRES_REVERSE_OFFLOAD)) != 0))
That is: we accept for nvptx
omp requires unfied_address
However, while the address space is the same, the following is not handled:
> the is_device_ptr clause is not necessary to obtain device addresses from
> device pointers for use inside target regions.
Expected:
(A) Mapping related: is_device_ptr can be left out.
(B) omp_target_is_accessible - will properly work for such pointers.
For nvptx, the check can be done via cudaPointerGetAttributes
if I understand
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__UNIFIED.html
correctly.
* * *
NOTE: Something similar is needed for GCN, except that its
GOMP_OFFLOAD_get_num_devices currently returns -1 when
GOMP_REQUIRES_UNIFIED_ADDRESS has been requested.
It seems as if hsa_amd_pointer_info is the function to be used,
cf.
https://github.com/RadeonOpenCompute/ROCR-Runtime/blob/master/src/inc/hsa_ext_amd.h
