1 /* Copyright (C) 2015-2017 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 (thr->task && thr->task->taskgroup && thr->task->taskgroup->cancelled)
153 	return;
154     }
155   else
156     ialias_call (GOMP_taskgroup_start) ();
157 
158   if (priority > gomp_max_task_priority_var)
159     priority = gomp_max_task_priority_var;
160 
161   if ((flags & GOMP_TASK_FLAG_IF) == 0 || team == NULL
162       || (thr->task && thr->task->final_task)
163       || team->task_count + num_tasks > 64 * team->nthreads)
164     {
165       unsigned long i;
166       if (__builtin_expect (cpyfn != NULL, 0))
167 	{
168 	  struct gomp_task task[num_tasks];
169 	  struct gomp_task *parent = thr->task;
170 	  arg_size = (arg_size + arg_align - 1) & ~(arg_align - 1);
171 	  char buf[num_tasks * arg_size + arg_align - 1];
172 	  char *arg = (char *) (((uintptr_t) buf + arg_align - 1)
173 				& ~(uintptr_t) (arg_align - 1));
174 	  char *orig_arg = arg;
175 	  for (i = 0; i < num_tasks; i++)
176 	    {
177 	      gomp_init_task (&task[i], parent, gomp_icv (false));
178 	      task[i].priority = priority;
179 	      task[i].kind = GOMP_TASK_UNDEFERRED;
180 	      task[i].final_task = (thr->task && thr->task->final_task)
181 				   || (flags & GOMP_TASK_FLAG_FINAL);
182 	      if (thr->task)
183 		{
184 		  task[i].in_tied_task = thr->task->in_tied_task;
185 		  task[i].taskgroup = thr->task->taskgroup;
186 		}
187 	      thr->task = &task[i];
188 	      cpyfn (arg, data);
189 	      arg += arg_size;
190 	    }
191 	  arg = orig_arg;
192 	  for (i = 0; i < num_tasks; i++)
193 	    {
194 	      thr->task = &task[i];
195 	      ((TYPE *)arg)[0] = start;
196 	      start += task_step;
197 	      ((TYPE *)arg)[1] = start;
198 	      if (i == nfirst)
199 		task_step -= step;
200 	      fn (arg);
201 	      arg += arg_size;
202 	      if (!priority_queue_empty_p (&task[i].children_queue,
203 					   MEMMODEL_RELAXED))
204 		{
205 		  gomp_mutex_lock (&team->task_lock);
206 		  gomp_clear_parent (&task[i].children_queue);
207 		  gomp_mutex_unlock (&team->task_lock);
208 		}
209 	      gomp_end_task ();
210 	    }
211 	}
212       else
213 	for (i = 0; i < num_tasks; i++)
214 	  {
215 	    struct gomp_task task;
216 
217 	    gomp_init_task (&task, thr->task, gomp_icv (false));
218 	    task.priority = priority;
219 	    task.kind = GOMP_TASK_UNDEFERRED;
220 	    task.final_task = (thr->task && thr->task->final_task)
221 			      || (flags & GOMP_TASK_FLAG_FINAL);
222 	    if (thr->task)
223 	      {
224 		task.in_tied_task = thr->task->in_tied_task;
225 		task.taskgroup = thr->task->taskgroup;
226 	      }
227 	    thr->task = &task;
228 	    ((TYPE *)data)[0] = start;
229 	    start += task_step;
230 	    ((TYPE *)data)[1] = start;
231 	    if (i == nfirst)
232 	      task_step -= step;
233 	    fn (data);
234 	    if (!priority_queue_empty_p (&task.children_queue,
235 					 MEMMODEL_RELAXED))
236 	      {
237 		gomp_mutex_lock (&team->task_lock);
238 		gomp_clear_parent (&task.children_queue);
239 		gomp_mutex_unlock (&team->task_lock);
240 	      }
241 	    gomp_end_task ();
242 	  }
243     }
244   else
245     {
246       struct gomp_task *tasks[num_tasks];
247       struct gomp_task *parent = thr->task;
248       struct gomp_taskgroup *taskgroup = parent->taskgroup;
249       char *arg;
250       int do_wake;
251       unsigned long i;
252 
253       for (i = 0; i < num_tasks; i++)
254 	{
255 	  struct gomp_task *task
256 	    = gomp_malloc (sizeof (*task) + arg_size + arg_align - 1);
257 	  tasks[i] = task;
258 	  arg = (char *) (((uintptr_t) (task + 1) + arg_align - 1)
259 			  & ~(uintptr_t) (arg_align - 1));
260 	  gomp_init_task (task, parent, gomp_icv (false));
261 	  task->priority = priority;
262 	  task->kind = GOMP_TASK_UNDEFERRED;
263 	  task->in_tied_task = parent->in_tied_task;
264 	  task->taskgroup = taskgroup;
265 	  thr->task = task;
266 	  if (cpyfn)
267 	    {
268 	      cpyfn (arg, data);
269 	      task->copy_ctors_done = true;
270 	    }
271 	  else
272 	    memcpy (arg, data, arg_size);
273 	  ((TYPE *)arg)[0] = start;
274 	  start += task_step;
275 	  ((TYPE *)arg)[1] = start;
276 	  if (i == nfirst)
277 	    task_step -= step;
278 	  thr->task = parent;
279 	  task->kind = GOMP_TASK_WAITING;
280 	  task->fn = fn;
281 	  task->fn_data = arg;
282 	  task->final_task = (flags & GOMP_TASK_FLAG_FINAL) >> 1;
283 	}
284       gomp_mutex_lock (&team->task_lock);
285       /* If parallel or taskgroup has been cancelled, don't start new
286 	 tasks.  */
287       if (__builtin_expect ((gomp_team_barrier_cancelled (&team->barrier)
288 			     || (taskgroup && taskgroup->cancelled))
289 			    && cpyfn == NULL, 0))
290 	{
291 	  gomp_mutex_unlock (&team->task_lock);
292 	  for (i = 0; i < num_tasks; i++)
293 	    {
294 	      gomp_finish_task (tasks[i]);
295 	      free (tasks[i]);
296 	    }
297 	  if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
298 	    ialias_call (GOMP_taskgroup_end) ();
299 	  return;
300 	}
301       if (taskgroup)
302 	taskgroup->num_children += num_tasks;
303       for (i = 0; i < num_tasks; i++)
304 	{
305 	  struct gomp_task *task = tasks[i];
306 	  priority_queue_insert (PQ_CHILDREN, &parent->children_queue,
307 				 task, priority,
308 				 PRIORITY_INSERT_BEGIN,
309 				 /*last_parent_depends_on=*/false,
310 				 task->parent_depends_on);
311 	  if (taskgroup)
312 	    priority_queue_insert (PQ_TASKGROUP, &taskgroup->taskgroup_queue,
313 				   task, priority, PRIORITY_INSERT_BEGIN,
314 				   /*last_parent_depends_on=*/false,
315 				   task->parent_depends_on);
316 	  priority_queue_insert (PQ_TEAM, &team->task_queue, task, priority,
317 				 PRIORITY_INSERT_END,
318 				 /*last_parent_depends_on=*/false,
319 				 task->parent_depends_on);
320 	  ++team->task_count;
321 	  ++team->task_queued_count;
322 	}
323       gomp_team_barrier_set_task_pending (&team->barrier);
324       if (team->task_running_count + !parent->in_tied_task
325 	  < team->nthreads)
326 	{
327 	  do_wake = team->nthreads - team->task_running_count
328 		    - !parent->in_tied_task;
329 	  if ((unsigned long) do_wake > num_tasks)
330 	    do_wake = num_tasks;
331 	}
332       else
333 	do_wake = 0;
334       gomp_mutex_unlock (&team->task_lock);
335       if (do_wake)
336 	gomp_team_barrier_wake (&team->barrier, do_wake);
337     }
338   if ((flags & GOMP_TASK_FLAG_NOGROUP) == 0)
339     ialias_call (GOMP_taskgroup_end) ();
340 }
341