]> gcc.gnu.org Git - gcc.git/commit
openmp: Honor OpenMP 5.1 num_teams lower bound
authorJakub Jelinek <jakub@redhat.com>
Fri, 12 Nov 2021 11:41:22 +0000 (12:41 +0100)
committerJakub Jelinek <jakub@redhat.com>
Fri, 12 Nov 2021 11:41:22 +0000 (12:41 +0100)
commit7d6da11fce054b25b50d0dec7f8d49cf22852680
treea52c163e2c2d9eb14ec2cc43f59029673863cc12
parent5f516a6a5d7ecce48a86d01fed1aeb4fc4ccc534
openmp: Honor OpenMP 5.1 num_teams lower bound

The following patch implements what I've been talking about earlier,
honor that for explicit num_teams clause we create at least the
lower-bound (if not specified, upper-bound) teams in the league.
For host fallback, it still means we only have one thread doing all the
teams, sequentially one after another.
For PTX and GCN, I think the new teams-2.c test and maybe teams-4.c too
will or might fail.
For these offloads, I think it is ok to remove symbols no longer used
from libgomp.a.
If num_teams_lower is bigger than the provided num_blocks or num_workgroups,
we should arrange for gomp_num_teams_var to be num_teams_lower - 1,
stop using the %ctaid.x or __builtin_gcn_dim_pos (0) for omp_get_team_num ()
and instead use for it some .shared var that GOMP_teams4 initializes to
%ctaid.x or __builtin_gcn_dim_pos (0) when first and for !first
increment that by num_blocks or num_workgroups each time and only
return false when we are above num_teams_lower.
Any help with actually implementing this for the 2 architectures highly
appreciated.

2021-11-12  Jakub Jelinek  <jakub@redhat.com>

gcc/
* omp-builtins.def (BUILT_IN_GOMP_TEAMS): Remove.
(BUILT_IN_GOMP_TEAMS4): New.
* builtin-types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
* omp-low.c (lower_omp_teams): Use GOMP_teams4 instead of
GOMP_teams, pass to it also num_teams lower-bound expression
or a dup of upper-bound if it is missing and a flag whether
it is the first call or not.
gcc/fortran/
* types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
libgomp/
* libgomp_g.h (GOMP_teams4): Declare.
* libgomp.map (GOMP_5.1): Export GOMP_teams4.
* target.c (GOMP_teams4): New function.
* config/nvptx/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* config/gcn/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* testsuite/libgomp.c/teams-4.c (main): Expect exactly 2
teams instead of <= 2.
* testsuite/libgomp.c-c++-common/teams-2.c: New test.
gcc/builtin-types.def
gcc/fortran/types.def
gcc/omp-builtins.def
gcc/omp-low.c
libgomp/config/gcn/target.c
libgomp/config/nvptx/target.c
libgomp/libgomp.map
libgomp/libgomp_g.h
libgomp/target.c
libgomp/testsuite/libgomp.c-c++-common/teams-2.c [new file with mode: 0644]
libgomp/testsuite/libgomp.c/teams-4.c
This page took 0.065227 seconds and 5 git commands to generate.