libgomp/taskloop.c

   1 /* Copyright (C) 2015-2018 Free Software Foundation, Inc.
   2    Contributed by Jakub Jelinek <jakub@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 taskloop construct.  It is included twice, once
  27    for the long and once for unsigned long long variant.  */
  28
  29 /* Called when encountering an explicit task directive.  If IF_CLAUSE is
  30    false, then we must not delay in executing the task.  If UNTIED is true,
  31    then the task may be executed by any member of the team.  */
  32
  33 void
  34 GOMP_taskloop (void (*fn) (void *), void *data, void (*cpyfn) (void *, void *),
  35                long arg_size, long arg_align, unsigned flags,
  36                unsigned long num_tasks, int priority,
  37                TYPE start, TYPE end, TYPE step)
  38 {
  39   struct gomp_thread *thr = gomp_thread ();
  40   struct gomp_team *team = thr->ts.team;
  41
  42 #ifdef HAVE_BROKEN_POSIX_SEMAPHORES
  43   /* If pthread_mutex_* is used for omp_*lock*, then each task must be
  44      tied to one thread all the time.  This means UNTIED tasks must be
  45      tied and if CPYFN is non-NULL IF(0) must be forced, as CPYFN
  46      might be running on different thread than FN.  */
  47   if (cpyfn)
  48     flags &= ~GOMP_TASK_FLAG_IF;
  49   flags &= ~GOMP_TASK_FLAG_UNTIED;
  50 #endif
  51
  52   /* If parallel or taskgroup has been cancelled, don't start new tasks.  */
  53   if (team && gomp_team_barrier_cancelled (&team->barrier))
  54     return;
  55
  56 #ifdef TYPE_is_long
  57   TYPE s = step;
  58   if (step > 0)
  59     {
  60       if (start >= end)
  61         return;
  62       s--;
  63     }
  64   else
  65     {
  66       if (start <= end)
  67         return;
  68       s++;
  69     }
  70   UTYPE n = (end - start + s) / step;
  71 #else
  72   UTYPE n;
  73   if (flags & GOMP_TASK_FLAG_UP)
  74     {
  75       if (start >= end)
  76         return;
  77       n = (end - start + step - 1) / step;
  78     }
  79   else
  80     {
  81       if (start <= end)
  82         return;
  83       n = (start - end - step - 1) / -step;
  84     }
  85 #endif
  86
  87   TYPE task_step = step;
  88   unsigned long nfirst = n;
  89   if (flags & GOMP_TASK_FLAG_GRAINSIZE)
  90     {
  91       unsigned long grainsize = num_tasks;
  92 #ifdef TYPE_is_long
  93       num_tasks = n / grainsize;
  94 #else
  95       UTYPE ndiv = n / grainsize;
  96       num_tasks = ndiv;
  97       if (num_tasks != ndiv)
  98         num_tasks = ~0UL;
  99 #endif
 100       if (num_tasks <= 1)
 101         {
 102           num_tasks = 1;
 103           task_step = end - start;
 104         }
 105       else if (num_tasks >= grainsize
 106 #ifndef TYPE_is_long
 107                && num_tasks != ~0UL
 108 #endif
 109               )
 110         {
 111           UTYPE mul = num_tasks * grainsize;
 112           task_step = (TYPE) grainsize * step;
 113           if (mul != n)
 114             {
 115               task_step += step;
 116               nfirst = n - mul - 1;
 117             }
 118         }
 119       else
 120         {
 121           UTYPE div = n / num_tasks;
 122           UTYPE mod = n % num_tasks;
 123           task_step = (TYPE) div * step;
 124           if (mod)
 125             {
 126               task_step += step;
 127               nfirst = mod - 1;
 128             }
 129         }
 130     }
 131   else
 132     {
 133       if (num_tasks == 0)
 134         num_tasks = team ? team->nthreads : 1;
 135       if (num_tasks >= n)
 136         num_tasks = n;
 137       else
 138         {
 139           UTYPE div = n / num_tasks;
 140           UTYPE mod = n % num_tasks;
 141           task_step = (TYPE) div * step;
 142           if (mod)
 143             {
 144               task_step += step;
 145               nfirst = mod - 1;
 146             }
 147         }
 148     }
 149
 150   if (flags & GOMP_TASK_FLAG_NOGROUP)
 151     {
 152       if (__builtin_expect (gomp_cancel_var, 0)
 153           && thr->task
 154           && thr->task->taskgroup)
 155         {
 156           if (thr->task->taskgroup->cancelled)
 157             return;
 158           if (thr->task->taskgroup->workshare
 159               && thr->task->taskgroup->prev
 160               && thr->task->taskgroup->prev->cancelled)
 161             return;
 162         }
 163     }
 164   else
 165     {
 166       ialias_call (GOMP_taskgroup_start) ();
 167       if (flags & GOMP_TASK_FLAG_REDUCTION)
 168         {
 169           struct gomp_data_head { TYPE t1, t2; uintptr_t *ptr; };
 170           uintptr_t *ptr = ((struct gomp_data_head *) data)->ptr;
 171           ialias_call (GOMP_taskgroup_reduction_register) (ptr);
 172         }
 173     }
 174
 175   if (priority > gomp_max_task_priority_var)
 176     priority = gomp_max_task_priority_var;
 177
 178   if ((flags & GOMP_TASK_FLAG_IF) == 0 || team == NULL
 179       || (thr->task && thr->task->final_task)
 180       || team->task_count + num_tasks > 64 * team->nthreads)
 181     {
 182       unsigned long i;
 183       if (__builtin_expect (cpyfn != NULL, 0))
 184         {
 185           struct gomp_task task[num_tasks];
 186           struct gomp_task *parent = thr->task;
 187           arg_size = (arg_size + arg_align - 1) & ~(arg_align - 1);
 188           char buf[num_tasks * arg_size + arg_align - 1];
 189           char *arg = (char *) (((uintptr_t) buf + arg_align - 1)
 190                                 & ~(uintptr_t) (arg_align - 1));
 191           char *orig_arg = arg;
 192           for (i = 0; i < num_tasks; i++)
 193             {
 194               gomp_init_task (&task[i], parent, gomp_icv (false));
 195               task[i].priority = priority;
 196               task[i].kind = GOMP_TASK_UNDEFERRED;
 197               task[i].final_task = (thr->task && thr->task->final_task)
 198                                    || (flags & GOMP_TASK_FLAG_FINAL);
 199               if (thr->task)
 200                 {
 201                   task[i].in_tied_task = thr->task->in_tied_task;
 202                   task[i].taskgroup = thr->task->taskgroup;
 203                 }
 204               thr->task = &task[i];
 205               cpyfn (arg, data);
 206               arg += arg_size;
 207             }
 208           arg = orig_arg;
 209           for (i = 0; i < num_tasks; i++)
 210             {
 211               thr->task = &task[i];
 212               ((TYPE *)arg)[0] = start;
 213               start += task_step;
 214               ((TYPE *)arg)[1] = start;
 215               if (i == nfirst)
 216                 task_step -= step;
 217               fn (arg);
 218               arg += arg_size;
 219               if (!priority_queue_empty_p (&task[i].children_queue,
 220                                            MEMMODEL_RELAXED))
 221                 {
 222                   gomp_mutex_lock (&team->task_lock);
 223                   gomp_clear_parent (&task[i].children_queue);
 224                   gomp_mutex_unlock (&team->task_lock);
 225                 }
 226               gomp_end_task ();
 227             }
 228         }
 229       else
 230         for (i = 0; i < num_tasks; i++)
 231           {
 232             struct gomp_task task;
 233
 234             gomp_init_task (&task, thr->task, gomp_icv (false));
 235             task.priority = priority;
 236             task.kind = GOMP_TASK_UNDEFERRED;
 237             task.final_task = (thr->task && thr->task->final_task)
 238                               || (flags & GOMP_TASK_FLAG_FINAL);
 239             if (thr->task)
 240               {
 241                 task.in_tied_task = thr->task->in_tied_task;
 242                 task.taskgroup = thr->task->taskgroup;
 243               }
 244             thr->task = &task;
 245             ((TYPE *)data)[0] = start;
 246             start += task_step;
 247             ((TYPE *)data)[1] = start;
 248             if (i == nfirst)
 249               task_step -= step;
 250             fn (data);
 251             if (!priority_queue_empty_p (&task.children_queue,
 252                                          MEMMODEL_RELAXED))
 253               {
 254                 gomp_mutex_lock (&team->task_lock);
 255                 gomp_clear_parent (&task.children_queue);
 256                 gomp_mutex_unlock (&team->task_lock);
 257               }
 258             gomp_end_task ();
 259           }
 260     }
 261   else
 262     {
 263       struct gomp_task *tasks[num_tasks];
 264       struct gomp_task *parent = thr->task;
 265       struct gomp_taskgroup *taskgroup = parent->taskgroup;
 266       char *arg;
 267       int do_wake;
 268       unsigned long i;
 269
 270       for (i = 0; i < num_tasks; i++)
 271         {
 272           struct gomp_task *task
 273             = gomp_malloc (sizeof (*task) + arg_size + arg_align - 1);
 274           tasks[i] = task;
 275           arg = (char *) (((uintptr_t) (task + 1) + arg_align - 1)
 276                           & ~(uintptr_t) (arg_align - 1));
 277           gomp_init_task (task, parent, gomp_icv (false));
 278           task->priority = priority;
 279           task->kind = GOMP_TASK_UNDEFERRED;
 280           task->in_tied_task = parent->in_tied_task;
 281           task->taskgroup = taskgroup;
 282           thr->task = task;
 283           if (cpyfn)
 284             {
 285               cpyfn (arg, data);
 286               task->copy_ctors_done = true;
 287             }
 288           else
 289             memcpy (arg, data, arg_size);
 290           ((TYPE *)arg)[0] = start;
 291           start += task_step;
 292           ((TYPE *)arg)[1] = start;
 293           if (i == nfirst)
 294             task_step -= step;
 295           thr->task = parent;
 296           task->kind = GOMP_TASK_WAITING;
 297           task->fn = fn;
 298           task->fn_data = arg;
 299           task->final_task = (flags & GOMP_TASK_FLAG_FINAL) >> 1;
 300         }
 301       gomp_mutex_lock (&team->task_lock);
 302       /* If parallel or taskgroup has been cancelled, don't start new
 303          tasks.  */
 304       if (__builtin_expect (gomp_cancel_var, 0)
 305           && cpyfn == NULL)
 306         {
 307           if (gomp_team_barrier_cancelled (&team->barrier))
 308             {
 309             do_cancel:
 310               gomp_mutex_unlock (&team->task_lock);
 311               for (i = 0; i < num_tasks; i++)
 312                 {
 313                   gomp_finish_task (tasks[i]);
 314                   free (tasks[i]);
 315                 }
 316               if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
 317                 ialias_call (GOMP_taskgroup_end) ();
 318               return;
 319             }
 320           if (taskgroup)
 321             {
 322               if (taskgroup->cancelled)
 323                 goto do_cancel;
 324               if (taskgroup->workshare
 325                   && taskgroup->prev
 326                   && taskgroup->prev->cancelled)
 327                 goto do_cancel;
 328             }
 329         }
 330       if (taskgroup)
 331         taskgroup->num_children += num_tasks;
 332       for (i = 0; i < num_tasks; i++)
 333         {
 334           struct gomp_task *task = tasks[i];
 335           priority_queue_insert (PQ_CHILDREN, &parent->children_queue,
 336                                  task, priority,
 337                                  PRIORITY_INSERT_BEGIN,
 338                                  /*last_parent_depends_on=*/false,
 339                                  task->parent_depends_on);
 340           if (taskgroup)
 341             priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
 342                                    task, priority, PRIORITY_INSERT_BEGIN,
 343                                    /*last_parent_depends_on=*/false,
 344                                    task->parent_depends_on);
 345           priority_queue_insert (PQ_TEAM, &team->task_queue, task, priority,
 346                                  PRIORITY_INSERT_END,
 347                                  /*last_parent_depends_on=*/false,
 348                                  task->parent_depends_on);
 349           ++team->task_count;
 350           ++team->task_queued_count;
 351         }
 352       gomp_team_barrier_set_task_pending (&team->barrier);
 353       if (team->task_running_count + !parent->in_tied_task
 354           < team->nthreads)
 355         {
 356           do_wake = team->nthreads - team->task_running_count
 357                     - !parent->in_tied_task;
 358           if ((unsigned long) do_wake > num_tasks)
 359             do_wake = num_tasks;
 360         }
 361       else
 362         do_wake = 0;
 363       gomp_mutex_unlock (&team->task_lock);
 364       if (do_wake)
 365         gomp_team_barrier_wake (&team->barrier, do_wake);
 366     }
 367   if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
 368     ialias_call (GOMP_taskgroup_end) ();
 369 }