libgomp/parallel.c

   1 /* Copyright (C) 2005-2016 Free Software Foundation, Inc.
   2    Contributed by Richard Henderson <rth@redhat.com>.
   3
   4    This file is part of the GNU Offloading and Multi Processing Library
   5    (libgomp).
   6
   7    Libgomp is free software; you can redistribute it and/or modify it
   8    under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3, or (at your option)
  10    any later version.
  11
  12    Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
  13    WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
  14    FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  15    more details.
  16
  17    Under Section 7 of GPL version 3, you are granted additional
  18    permissions described in the GCC Runtime Library Exception, version
  19    3.1, as published by the Free Software Foundation.
  20
  21    You should have received a copy of the GNU General Public License and
  22    a copy of the GCC Runtime Library Exception along with this program;
  23    see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  24    <http://www.gnu.org/licenses/>.  */
  25
  26 /* This file handles the (bare) PARALLEL construct.  */
  27
  28 #include "libgomp.h"
  29 #include <limits.h>
  30
  31
  32 /* Determine the number of threads to be launched for a PARALLEL construct.
  33    This algorithm is explicitly described in OpenMP 3.0 section 2.4.1.
  34    SPECIFIED is a combination of the NUM_THREADS clause and the IF clause.
  35    If the IF clause is false, SPECIFIED is forced to 1.  When NUM_THREADS
  36    is not present, SPECIFIED is 0.  */
  37
  38 unsigned
  39 gomp_resolve_num_threads (unsigned specified, unsigned count)
  40 {
  41   struct gomp_thread *thr = gomp_thread ();
  42   struct gomp_task_icv *icv;
  43   unsigned threads_requested, max_num_threads, num_threads;
  44   unsigned long busy;
  45   struct gomp_thread_pool *pool;
  46
  47   icv = gomp_icv (false);
  48
  49   if (specified == 1)
  50     return 1;
  51   else if (thr->ts.active_level >= 1 && !icv->nest_var)
  52     return 1;
  53   else if (thr->ts.active_level >= gomp_max_active_levels_var)
  54     return 1;
  55
  56   /* If NUM_THREADS not specified, use nthreads_var.  */
  57   if (specified == 0)
  58     threads_requested = icv->nthreads_var;
  59   else
  60     threads_requested = specified;
  61
  62   max_num_threads = threads_requested;
  63
  64   /* If dynamic threads are enabled, bound the number of threads
  65      that we launch.  */
  66   if (icv->dyn_var)
  67     {
  68       unsigned dyn = gomp_dynamic_max_threads ();
  69       if (dyn < max_num_threads)
  70         max_num_threads = dyn;
  71
  72       /* Optimization for parallel sections.  */
  73       if (count && count < max_num_threads)
  74         max_num_threads = count;
  75     }
  76
  77   /* UINT_MAX stands for infinity.  */
  78   if (__builtin_expect (icv->thread_limit_var == UINT_MAX, 1)
  79       || max_num_threads == 1)
  80     return max_num_threads;
  81
  82   /* The threads_busy counter lives in thread_pool, if there
  83      isn't a thread_pool yet, there must be just one thread
  84      in the contention group.  If thr->team is NULL, this isn't
  85      nested parallel, so there is just one thread in the
  86      contention group as well, no need to handle it atomically.  */
  87   pool = thr->thread_pool;
  88   if (thr->ts.team == NULL || pool == NULL)
  89     {
  90       num_threads = max_num_threads;
  91       if (num_threads > icv->thread_limit_var)
  92         num_threads = icv->thread_limit_var;
  93       if (pool)
  94         pool->threads_busy = num_threads;
  95       return num_threads;
  96     }
  97
  98 #ifdef HAVE_SYNC_BUILTINS
  99   do
 100     {
 101       busy = pool->threads_busy;
 102       num_threads = max_num_threads;
 103       if (icv->thread_limit_var - busy + 1 < num_threads)
 104         num_threads = icv->thread_limit_var - busy + 1;
 105     }
 106   while (__sync_val_compare_and_swap (&pool->threads_busy,
 107                                       busy, busy + num_threads - 1)
 108          != busy);
 109 #else
 110   gomp_mutex_lock (&gomp_managed_threads_lock);
 111   num_threads = max_num_threads;
 112   busy = pool->threads_busy;
 113   if (icv->thread_limit_var - busy + 1 < num_threads)
 114     num_threads = icv->thread_limit_var - busy + 1;
 115   pool->threads_busy += num_threads - 1;
 116   gomp_mutex_unlock (&gomp_managed_threads_lock);
 117 #endif
 118
 119   return num_threads;
 120 }
 121
 122 void
 123 GOMP_parallel_start (void (*fn) (void *), void *data, unsigned num_threads)
 124 {
 125   num_threads = gomp_resolve_num_threads (num_threads, 0);
 126   gomp_team_start (fn, data, num_threads, 0, gomp_new_team (num_threads));
 127 }
 128
 129 void
 130 GOMP_parallel_end (void)
 131 {
 132   struct gomp_task_icv *icv = gomp_icv (false);
 133   if (__builtin_expect (icv->thread_limit_var != UINT_MAX, 0))
 134     {
 135       struct gomp_thread *thr = gomp_thread ();
 136       struct gomp_team *team = thr->ts.team;
 137       unsigned int nthreads = team ? team->nthreads : 1;
 138       gomp_team_end ();
 139       if (nthreads > 1)
 140         {
 141           /* If not nested, there is just one thread in the
 142              contention group left, no need for atomicity.  */
 143           if (thr->ts.team == NULL)
 144             thr->thread_pool->threads_busy = 1;
 145           else
 146             {
 147 #ifdef HAVE_SYNC_BUILTINS
 148               __sync_fetch_and_add (&thr->thread_pool->threads_busy,
 149                                     1UL - nthreads);
 150 #else
 151               gomp_mutex_lock (&gomp_managed_threads_lock);
 152               thr->thread_pool->threads_busy -= nthreads - 1;
 153               gomp_mutex_unlock (&gomp_managed_threads_lock);
 154 #endif
 155             }
 156         }
 157     }
 158   else
 159     gomp_team_end ();
 160 }
 161 ialias (GOMP_parallel_end)
 162
 163 void
 164 GOMP_parallel (void (*fn) (void *), void *data, unsigned num_threads, unsigned int flags)
 165 {
 166   num_threads = gomp_resolve_num_threads (num_threads, 0);
 167   gomp_team_start (fn, data, num_threads, flags, gomp_new_team (num_threads));
 168   fn (data);
 169   ialias_call (GOMP_parallel_end) ();
 170 }
 171
 172 bool
 173 GOMP_cancellation_point (int which)
 174 {
 175   if (!gomp_cancel_var)
 176     return false;
 177
 178   struct gomp_thread *thr = gomp_thread ();
 179   struct gomp_team *team = thr->ts.team;
 180   if (which & (GOMP_CANCEL_LOOP | GOMP_CANCEL_SECTIONS))
 181     {
 182       if (team == NULL)
 183         return false;
 184       return team->work_share_cancelled != 0;
 185     }
 186   else if (which & GOMP_CANCEL_TASKGROUP)
 187     {
 188       if (thr->task->taskgroup && thr->task->taskgroup->cancelled)
 189         return true;
 190       /* FALLTHRU into the GOMP_CANCEL_PARALLEL case,
 191          as #pragma omp cancel parallel also cancels all explicit
 192          tasks.  */
 193     }
 194   if (team)
 195     return gomp_team_barrier_cancelled (&team->barrier);
 196   return false;
 197 }
 198 ialias (GOMP_cancellation_point)
 199
 200 bool
 201 GOMP_cancel (int which, bool do_cancel)
 202 {
 203   if (!gomp_cancel_var)
 204     return false;
 205
 206   if (!do_cancel)
 207     return ialias_call (GOMP_cancellation_point) (which);
 208
 209   struct gomp_thread *thr = gomp_thread ();
 210   struct gomp_team *team = thr->ts.team;
 211   if (which & (GOMP_CANCEL_LOOP | GOMP_CANCEL_SECTIONS))
 212     {
 213       /* In orphaned worksharing region, all we want to cancel
 214          is current thread.  */
 215       if (team != NULL)
 216         team->work_share_cancelled = 1;
 217       return true;
 218     }
 219   else if (which & GOMP_CANCEL_TASKGROUP)
 220     {
 221       if (thr->task->taskgroup && !thr->task->taskgroup->cancelled)
 222         {
 223           gomp_mutex_lock (&team->task_lock);
 224           thr->task->taskgroup->cancelled = true;
 225           gomp_mutex_unlock (&team->task_lock);
 226         }
 227       return true;
 228     }
 229   team->team_cancelled = 1;
 230   gomp_team_barrier_cancel (team);
 231   return true;
 232 }
 233 \f
 234 /* The public OpenMP API for thread and team related inquiries.  */
 235
 236 int
 237 omp_get_num_threads (void)
 238 {
 239   struct gomp_team *team = gomp_thread ()->ts.team;
 240   return team ? team->nthreads : 1;
 241 }
 242
 243 int
 244 omp_get_thread_num (void)
 245 {
 246   return gomp_thread ()->ts.team_id;
 247 }
 248
 249 /* This wasn't right for OpenMP 2.5.  Active region used to be non-zero
 250    when the IF clause doesn't evaluate to false, starting with OpenMP 3.0
 251    it is non-zero with more than one thread in the team.  */
 252
 253 int
 254 omp_in_parallel (void)
 255 {
 256   return gomp_thread ()->ts.active_level > 0;
 257 }
 258
 259 int
 260 omp_get_level (void)
 261 {
 262   return gomp_thread ()->ts.level;
 263 }
 264
 265 int
 266 omp_get_ancestor_thread_num (int level)
 267 {
 268   struct gomp_team_state *ts = &gomp_thread ()->ts;
 269   if (level < 0 || level > ts->level)
 270     return -1;
 271   for (level = ts->level - level; level > 0; --level)
 272     ts = &ts->team->prev_ts;
 273   return ts->team_id;
 274 }
 275
 276 int
 277 omp_get_team_size (int level)
 278 {
 279   struct gomp_team_state *ts = &gomp_thread ()->ts;
 280   if (level < 0 || level > ts->level)
 281     return -1;
 282   for (level = ts->level - level; level > 0; --level)
 283     ts = &ts->team->prev_ts;
 284   if (ts->team == NULL)
 285     return 1;
 286   else
 287     return ts->team->nthreads;
 288 }
 289
 290 int
 291 omp_get_active_level (void)
 292 {
 293   return gomp_thread ()->ts.active_level;
 294 }
 295
 296 ialias (omp_get_num_threads)
 297 ialias (omp_get_thread_num)
 298 ialias (omp_in_parallel)
 299 ialias (omp_get_level)
 300 ialias (omp_get_ancestor_thread_num)
 301 ialias (omp_get_team_size)
 302 ialias (omp_get_active_level)