https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110364
Bug ID: 110364
Summary: [OpenMP] Environment vars - doc and device-specific
cleanup; make GOMP_DEBUG more useful; janitorial takes
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Keywords: documentation, openmp
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: ---
OpenMP TR11 added the _ALL, _DEV, and _DEV_<number> suffix.
After the patch work stated, the following addition was done to the spec:
• Device-specific environment variables must not specify the initial device.
TODO: Implement this.
OMP_STACKSIZE_DEV etc. → Make this applicable to the non-host devices,
→ GCN_STACK_SIZE and nvptx's nvptx_stacks_size() (which is per warp!).
Improve the documentation (and implementation) with regards to the
device-specific ICVs - namely: which are handled, which aren't + possibly
implement some if needed.
Cross ref: PR 109620 (OMP_CANCELLATION), PR109875 (OMP_NUM_TEAMS)
(First step for the documentation bit was commit
r14-2032-g2cd0689a79498dcaaadc8cc5c1c4d0a452a4fb09)
GOMP_DEBUG: Currently, it is a 0 or all flag, it would be much more useful it
the numeric values could be used - either as (n) less output than (n+1) or
using some bitmap flags like 1<<0, 1<<1, 1<<2 turning on different outputs
(possibly with supporting some a comma-separated list of names for them.
In particular, plugin-nvptx outputs too much - and GCN has its own GCN_DEBUG.
* * *
BTW: While all OMP_* env vars are documented, the "OpenMP Runtime Library
Routines" lacks many routines there were added.
I additionally wonder whether the named-constants and kinds documentation
should be moved from Fortran's ../gcc/fortran/intrinsic.texi to libgomp and
extended to cover also the C/C++.
