https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88939
Bug ID: 88939
Summary: [nvptx, openacc, libgomp] Assertion `!s->map->active'
failed for synchronous parallel with abort
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: vries at gcc dot gnu.org
Target Milestone: ---
Atm, abort-1.c passes like this:
...
CheCKpOInT
libgomp: cuStreamSynchronize error: an illegal instruction was encountered
abort-1.exe: /libgomp/plugin/plugin-nvptx.c:271: map_fini: \
Assertion `!s->map->active' failed.
PASS: libgomp.oacc-c/../libgomp.oacc-c-c++-common/abort-1.c \
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0 -O0 execution test
...
The usual scenario in nvptx_exec for synchronous kernel launches is this
sequence of calls:
- map_push, allocating a device memory pointer for the kernel arguments, and
allocating a struct cuda_map (to keep track of the device pointer), setting
the active field of struct cuda_map to true
- cuLaunchKernel, launching a kernel on a CUDA stream
- cuStreamSynchronize, waiting till the stream task (i.e, the kernel) completes
- map_pop, freeing the device memory and the corresponding struct cuda_map
In the case of this test-case, the abort causes cuStreamSynchronize to return
CUDA_ERROR_ILLEGAL_INSTRUCTION, which is then handled using a GOMP_PLUGIN_fatal
call, which calls abort. This means map_pop is not called, so no memory is
freed.
Then during fini_streams_for_device, map_fini is called, the remaining map is
found, and the !s->map->active assert triggers.
