openmp: Add support for thread_limit clause on target
Commit Message
Hi!
OpenMP 5.1 says that thread_limit clause can also appear on target,
and similarly to teams should affect the thread-limit-var ICV.
On combined target teams, the clause goes to both.
We actually passed thread_limit internally on target already before,
but only used it for gcn/ptx offloading to hint how many threads should be
created and for ptx didn't set thread_limit_var in that case.
Similarly for host fallback.
Also, I found that we weren't copying the args array that contains encoded
thread_limit and num_teams clause for target (etc.) for async target.
Will commit to trunk once testing finishes.
2021-11-15 Jakub Jelinek <jakub@redhat.com>
gcc/
* gimplify.c (optimize_target_teams): Only add OMP_CLAUSE_THREAD_LIMIT
to OMP_TARGET_CLAUSES if it isn't there already.
gcc/c-family/
* c-omp.c (c_omp_split_clauses) <case OMP_CLAUSE_THREAD_LIMIT>:
Duplicate to both OMP_TARGET and OMP_TEAMS.
gcc/c/
* c-parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
gcc/cp/
* parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
libgomp/
* task.c (gomp_create_target_task): Copy args array as well.
* target.c (gomp_target_fallback): Add args argument.
Set gomp_icv (true)->thread_limit_var if thread_limit is present.
(GOMP_target): Adjust gomp_target_fallback caller.
(GOMP_target_ext): Likewise.
(gomp_target_task_fn): Likewise.
* config/nvptx/team.c (gomp_nvptx_main): Set
gomp_global_icv.thread_limit_var.
* testsuite/libgomp.c-c++-common/thread-limit-1.c: New test.
Jakub
Comments
Hi,
On 15.11.21 13:05, Jakub Jelinek wrote:
> OpenMP 5.1 says that thread_limit clause can also appear on target,
> and similarly to teams should affect the thread-limit-var ICV.
> On combined target teams, the clause goes to both.
This patch does this also for Fortran.
OK, once the post-bootstap testing finished successfully?
Tobias
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
On Mon, Nov 15, 2021 at 02:00:42PM +0100, Tobias Burnus wrote:
> Hi,
>
> On 15.11.21 13:05, Jakub Jelinek wrote:
> > OpenMP 5.1 says that thread_limit clause can also appear on target,
> > and similarly to teams should affect the thread-limit-var ICV.
> > On combined target teams, the clause goes to both.
>
> This patch does this also for Fortran.
>
> OK, once the post-bootstap testing finished successfully?
Ok, thanks.
> gcc/fortran/ChangeLog:
>
> * openmp.c (OMP_TARGET_CLAUSES): Add thread_limit.
> * trans-openmp.c (gfc_split_omp_clauses): Add thread_limit also to
> teams.
>
> libgomp/ChangeLog:
>
> * testsuite/libgomp.fortran/thread-limit-1.f90: New test.
Jakub
@@ -13637,10 +13637,13 @@ optimize_target_teams (tree target, gimp
if (!DECL_P (expr) && TREE_CODE (expr) != TARGET_EXPR)
OMP_CLAUSE_OPERAND (c, 0) = *p;
}
- c = build_omp_clause (thread_limit_loc, OMP_CLAUSE_THREAD_LIMIT);
- OMP_CLAUSE_THREAD_LIMIT_EXPR (c) = thread_limit;
- OMP_CLAUSE_CHAIN (c) = OMP_TARGET_CLAUSES (target);
- OMP_TARGET_CLAUSES (target) = c;
+ if (!omp_find_clause (OMP_TARGET_CLAUSES (target), OMP_CLAUSE_THREAD_LIMIT))
+ {
+ c = build_omp_clause (thread_limit_loc, OMP_CLAUSE_THREAD_LIMIT);
+ OMP_CLAUSE_THREAD_LIMIT_EXPR (c) = thread_limit;
+ OMP_CLAUSE_CHAIN (c) = OMP_TARGET_CLAUSES (target);
+ OMP_TARGET_CLAUSES (target) = c;
+ }
c = build_omp_clause (num_teams_loc, OMP_CLAUSE_NUM_TEAMS);
OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR (c) = num_teams_upper;
OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR (c) = num_teams_lower;
@@ -1867,7 +1867,6 @@ c_omp_split_clauses (location_t loc, enu
s = C_OMP_CLAUSE_SPLIT_TARGET;
break;
case OMP_CLAUSE_NUM_TEAMS:
- case OMP_CLAUSE_THREAD_LIMIT:
s = C_OMP_CLAUSE_SPLIT_TEAMS;
break;
case OMP_CLAUSE_DIST_SCHEDULE:
@@ -2531,6 +2530,30 @@ c_omp_split_clauses (location_t loc, enu
else
s = C_OMP_CLAUSE_SPLIT_FOR;
break;
+ /* thread_limit is allowed on target and teams. Distribute it
+ to all. */
+ case OMP_CLAUSE_THREAD_LIMIT:
+ if ((mask & (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_MAP))
+ != 0)
+ {
+ if ((mask & (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_NUM_TEAMS))
+ != 0)
+ {
+ c = build_omp_clause (OMP_CLAUSE_LOCATION (clauses),
+ OMP_CLAUSE_THREAD_LIMIT);
+ OMP_CLAUSE_THREAD_LIMIT_EXPR (c)
+ = OMP_CLAUSE_THREAD_LIMIT_EXPR (clauses);
+ OMP_CLAUSE_CHAIN (c) = cclauses[C_OMP_CLAUSE_SPLIT_TARGET];
+ cclauses[C_OMP_CLAUSE_SPLIT_TARGET] = c;
+ }
+ else
+ {
+ s = C_OMP_CLAUSE_SPLIT_TARGET;
+ break;
+ }
+ }
+ s = C_OMP_CLAUSE_SPLIT_TEAMS;
+ break;
/* Allocate clause is allowed on target, teams, distribute, parallel,
for, sections and taskloop. Distribute it to all. */
case OMP_CLAUSE_ALLOCATE:
@@ -20963,6 +20963,7 @@ c_parser_omp_target_exit_data (location_
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_ALLOCATE) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_DEFAULTMAP) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_IN_REDUCTION) \
+ | (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_THREAD_LIMIT) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_IS_DEVICE_PTR))
static bool
@@ -44015,6 +44015,7 @@ cp_parser_omp_target_update (cp_parser *
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_DEFAULTMAP) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_ALLOCATE) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_IN_REDUCTION) \
+ | (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_THREAD_LIMIT) \
| (OMP_CLAUSE_MASK_1 << PRAGMA_OMP_CLAUSE_IS_DEVICE_PTR))
static bool
@@ -745,6 +745,7 @@ gomp_create_target_task (struct gomp_dev
size_t depend_size = 0;
uintptr_t depend_cnt = 0;
size_t tgt_align = 0, tgt_size = 0;
+ uintptr_t args_cnt = 0;
if (depend != NULL)
{
@@ -769,10 +770,22 @@ gomp_create_target_task (struct gomp_dev
tgt_size += tgt_align - 1;
else
tgt_size = 0;
+ if (args)
+ {
+ void **cargs = args;
+ while (*cargs)
+ {
+ intptr_t id = (intptr_t) *cargs++;
+ if (id & GOMP_TARGET_ARG_SUBSEQUENT_PARAM)
+ cargs++;
+ }
+ args_cnt = cargs + 1 - args;
+ }
}
task = gomp_malloc (sizeof (*task) + depend_size
+ sizeof (*ttask)
+ + args_cnt * sizeof (void *)
+ mapnum * (sizeof (void *) + sizeof (size_t)
+ sizeof (unsigned short))
+ tgt_size);
@@ -785,9 +798,18 @@ gomp_create_target_task (struct gomp_dev
ttask->devicep = devicep;
ttask->fn = fn;
ttask->mapnum = mapnum;
- ttask->args = args;
memcpy (ttask->hostaddrs, hostaddrs, mapnum * sizeof (void *));
- ttask->sizes = (size_t *) &ttask->hostaddrs[mapnum];
+ if (args_cnt)
+ {
+ ttask->args = (void **) &ttask->hostaddrs[mapnum];
+ memcpy (ttask->args, args, args_cnt * sizeof (void *));
+ ttask->sizes = (size_t *) &ttask->args[args_cnt];
+ }
+ else
+ {
+ ttask->args = args;
+ ttask->sizes = (size_t *) &ttask->hostaddrs[mapnum];
+ }
memcpy (ttask->sizes, sizes, mapnum * sizeof (size_t));
ttask->kinds = (unsigned short *) &ttask->sizes[mapnum];
memcpy (ttask->kinds, kinds, mapnum * sizeof (unsigned short));
@@ -2362,7 +2362,7 @@ gomp_unload_device (struct gomp_device_d
static void
gomp_target_fallback (void (*fn) (void *), void **hostaddrs,
- struct gomp_device_descr *devicep)
+ struct gomp_device_descr *devicep, void **args)
{
struct gomp_thread old_thr, *thr = gomp_thread ();
@@ -2378,6 +2378,25 @@ gomp_target_fallback (void (*fn) (void *
thr->place = old_thr.place;
thr->ts.place_partition_len = gomp_places_list_len;
}
+ if (args)
+ while (*args)
+ {
+ intptr_t id = (intptr_t) *args++, val;
+ if (id & GOMP_TARGET_ARG_SUBSEQUENT_PARAM)
+ val = (intptr_t) *args++;
+ else
+ val = id >> GOMP_TARGET_ARG_VALUE_SHIFT;
+ if ((id & GOMP_TARGET_ARG_DEVICE_MASK) != GOMP_TARGET_ARG_DEVICE_ALL)
+ continue;
+ id &= GOMP_TARGET_ARG_ID_MASK;
+ if (id != GOMP_TARGET_ARG_THREAD_LIMIT)
+ continue;
+ val = val > INT_MAX ? INT_MAX : val;
+ if (val)
+ gomp_icv (true)->thread_limit_var = val;
+ break;
+ }
+
fn (hostaddrs);
gomp_free_thread (thr);
*thr = old_thr;
@@ -2478,7 +2497,7 @@ GOMP_target (int device, void (*fn) (voi
/* All shared memory devices should use the GOMP_target_ext function. */
|| devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM
|| !(fn_addr = gomp_get_target_fn_addr (devicep, fn)))
- return gomp_target_fallback (fn, hostaddrs, devicep);
+ return gomp_target_fallback (fn, hostaddrs, devicep, NULL);
htab_t refcount_set = htab_create (mapnum);
struct target_mem_desc *tgt_vars
@@ -2617,7 +2636,7 @@ GOMP_target_ext (int device, void (*fn)
tgt_align, tgt_size);
}
}
- gomp_target_fallback (fn, hostaddrs, devicep);
+ gomp_target_fallback (fn, hostaddrs, devicep, args);
return;
}
@@ -3052,7 +3071,8 @@ gomp_target_task_fn (void *data)
|| (devicep->can_run_func && !devicep->can_run_func (fn_addr)))
{
ttask->state = GOMP_TARGET_TASK_FALLBACK;
- gomp_target_fallback (ttask->fn, ttask->hostaddrs, devicep);
+ gomp_target_fallback (ttask->fn, ttask->hostaddrs, devicep,
+ ttask->args);
return false;
}
@@ -55,6 +55,7 @@ gomp_nvptx_main (void (*fn) (void *), vo
if (tid == 0)
{
gomp_global_icv.nthreads_var = ntids;
+ gomp_global_icv.thread_limit_var = ntids;
/* Starting additional threads is not supported. */
gomp_global_icv.dyn_var = true;
@@ -0,0 +1,23 @@
+#include <omp.h>
+#include <stdlib.h>
+
+void
+foo ()
+{
+ {
+ #pragma omp target parallel nowait thread_limit (4) num_threads (1)
+ if (omp_get_thread_limit () > 4)
+ abort ();
+ }
+ #pragma omp taskwait
+}
+
+int
+main ()
+{
+ #pragma omp target thread_limit (6)
+ if (omp_get_thread_limit () > 6)
+ abort ();
+ foo ();
+ return 0;
+}