https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111214
Bug ID: 111214
Summary: omp_get_num_procs: Improve documentation - especially
for devices
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Keywords: documentation
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: ---
Current wording:
https://gcc.gnu.org/onlinedocs/libgomp/omp_005fget_005fnum_005fprocs.html
"Returns the number of processors online on that device."
(A) For the host, I wonder whether it should mention the affinity bits, which
we have in Linux:
if (gomp_places_list == NULL)
... && pthread_getaffinity_np (pthread_self (), gomp_get_cpuset_size,
gomp_cpusetp) == 0)
...
(B) We are completely silent for devices.
Seems as if the number of independent hardware threads is what is the sentiment
during today's OpenMP accel talk, i.e. #warps (nvptx) and #wavefronts (amdgcn)
in hardware (possibly: minus those removed via explicit num_threads settings).
We currently have for accelerators:
return gomp_icv (false)->nthreads_var
with gomp_icv(false) is:
struct gomp_task *task = gomp_thread ()->task;
if (task)
return &task->icv; /*... */
else
return &gomp_global_icv;
And set on GCN gomp_global_icv.nthreads_var = 16 and for nvptx:
gomp_global_icv.nthreads_var = ntids.
Possibly, it should be omp_get_num_teams() * nthreads_var, or not.
In any case, it needs to be documented.
