|
0
|
1 /* Copyright (C) 2005, 2008, 2009 Free Software Foundation, Inc.
|
|
|
2 Contributed by Richard Henderson <rth@redhat.com>.
|
|
|
3
|
|
|
4 This file is part of the GNU OpenMP Library (libgomp).
|
|
|
5
|
|
|
6 Libgomp is free software; you can redistribute it and/or modify it
|
|
|
7 under the terms of the GNU General Public License as published by
|
|
|
8 the Free Software Foundation; either version 3, or (at your option)
|
|
|
9 any later version.
|
|
|
10
|
|
|
11 Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
|
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
|
|
13 FOR A PARTICULAR PURPOSE. See the GNU General Public License for
|
|
|
14 more details.
|
|
|
15
|
|
|
16 Under Section 7 of GPL version 3, you are granted additional
|
|
|
17 permissions described in the GCC Runtime Library Exception, version
|
|
|
18 3.1, as published by the Free Software Foundation.
|
|
|
19
|
|
|
20 You should have received a copy of the GNU General Public License and
|
|
|
21 a copy of the GCC Runtime Library Exception along with this program;
|
|
|
22 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
|
23 <http://www.gnu.org/licenses/>. */
|
|
|
24
|
|
|
25 /* This file contains routines to manage the work-share queue for a team
|
|
|
26 of threads. */
|
|
|
27
|
|
|
28 #include "libgomp.h"
|
|
|
29 #include <stddef.h>
|
|
|
30 #include <stdlib.h>
|
|
|
31 #include <string.h>
|
|
|
32
|
|
|
33
|
|
|
34 /* Allocate a new work share structure, preferably from current team's
|
|
|
35 free gomp_work_share cache. */
|
|
|
36
|
|
|
37 static struct gomp_work_share *
|
|
|
38 alloc_work_share (struct gomp_team *team)
|
|
|
39 {
|
|
|
40 struct gomp_work_share *ws;
|
|
|
41 unsigned int i;
|
|
|
42
|
|
|
43 /* This is called in a critical section. */
|
|
|
44 if (team->work_share_list_alloc != NULL)
|
|
|
45 {
|
|
|
46 ws = team->work_share_list_alloc;
|
|
|
47 team->work_share_list_alloc = ws->next_free;
|
|
|
48 return ws;
|
|
|
49 }
|
|
|
50
|
|
|
51 #ifdef HAVE_SYNC_BUILTINS
|
|
|
52 ws = team->work_share_list_free;
|
|
|
53 /* We need atomic read from work_share_list_free,
|
|
|
54 as free_work_share can be called concurrently. */
|
|
|
55 __asm ("" : "+r" (ws));
|
|
|
56
|
|
|
57 if (ws && ws->next_free)
|
|
|
58 {
|
|
|
59 struct gomp_work_share *next = ws->next_free;
|
|
|
60 ws->next_free = NULL;
|
|
|
61 team->work_share_list_alloc = next->next_free;
|
|
|
62 return next;
|
|
|
63 }
|
|
|
64 #else
|
|
|
65 gomp_mutex_lock (&team->work_share_list_free_lock);
|
|
|
66 ws = team->work_share_list_free;
|
|
|
67 if (ws)
|
|
|
68 {
|
|
|
69 team->work_share_list_alloc = ws->next_free;
|
|
|
70 team->work_share_list_free = NULL;
|
|
|
71 gomp_mutex_unlock (&team->work_share_list_free_lock);
|
|
|
72 return ws;
|
|
|
73 }
|
|
|
74 gomp_mutex_unlock (&team->work_share_list_free_lock);
|
|
|
75 #endif
|
|
|
76
|
|
|
77 team->work_share_chunk *= 2;
|
|
|
78 ws = gomp_malloc (team->work_share_chunk * sizeof (struct gomp_work_share));
|
|
|
79 ws->next_alloc = team->work_shares[0].next_alloc;
|
|
|
80 team->work_shares[0].next_alloc = ws;
|
|
|
81 team->work_share_list_alloc = &ws[1];
|
|
|
82 for (i = 1; i < team->work_share_chunk - 1; i++)
|
|
|
83 ws[i].next_free = &ws[i + 1];
|
|
|
84 ws[i].next_free = NULL;
|
|
|
85 return ws;
|
|
|
86 }
|
|
|
87
|
|
|
88 /* Initialize an already allocated struct gomp_work_share.
|
|
|
89 This shouldn't touch the next_alloc field. */
|
|
|
90
|
|
|
91 void
|
|
|
92 gomp_init_work_share (struct gomp_work_share *ws, bool ordered,
|
|
|
93 unsigned nthreads)
|
|
|
94 {
|
|
|
95 gomp_mutex_init (&ws->lock);
|
|
|
96 if (__builtin_expect (ordered, 0))
|
|
|
97 {
|
|
|
98 #define INLINE_ORDERED_TEAM_IDS_CNT \
|
|
|
99 ((sizeof (struct gomp_work_share) \
|
|
|
100 - offsetof (struct gomp_work_share, inline_ordered_team_ids)) \
|
|
|
101 / sizeof (((struct gomp_work_share *) 0)->inline_ordered_team_ids[0]))
|
|
|
102
|
|
|
103 if (nthreads > INLINE_ORDERED_TEAM_IDS_CNT)
|
|
|
104 ws->ordered_team_ids
|
|
|
105 = gomp_malloc (nthreads * sizeof (*ws->ordered_team_ids));
|
|
|
106 else
|
|
|
107 ws->ordered_team_ids = ws->inline_ordered_team_ids;
|
|
|
108 memset (ws->ordered_team_ids, '\0',
|
|
|
109 nthreads * sizeof (*ws->ordered_team_ids));
|
|
|
110 ws->ordered_num_used = 0;
|
|
|
111 ws->ordered_owner = -1;
|
|
|
112 ws->ordered_cur = 0;
|
|
|
113 }
|
|
|
114 else
|
|
|
115 ws->ordered_team_ids = NULL;
|
|
|
116 gomp_ptrlock_init (&ws->next_ws, NULL);
|
|
|
117 ws->threads_completed = 0;
|
|
|
118 }
|
|
|
119
|
|
|
120 /* Do any needed destruction of gomp_work_share fields before it
|
|
|
121 is put back into free gomp_work_share cache or freed. */
|
|
|
122
|
|
|
123 void
|
|
|
124 gomp_fini_work_share (struct gomp_work_share *ws)
|
|
|
125 {
|
|
|
126 gomp_mutex_destroy (&ws->lock);
|
|
|
127 if (ws->ordered_team_ids != ws->inline_ordered_team_ids)
|
|
|
128 free (ws->ordered_team_ids);
|
|
|
129 gomp_ptrlock_destroy (&ws->next_ws);
|
|
|
130 }
|
|
|
131
|
|
|
132 /* Free a work share struct, if not orphaned, put it into current
|
|
|
133 team's free gomp_work_share cache. */
|
|
|
134
|
|
|
135 static inline void
|
|
|
136 free_work_share (struct gomp_team *team, struct gomp_work_share *ws)
|
|
|
137 {
|
|
|
138 gomp_fini_work_share (ws);
|
|
|
139 if (__builtin_expect (team == NULL, 0))
|
|
|
140 free (ws);
|
|
|
141 else
|
|
|
142 {
|
|
|
143 struct gomp_work_share *next_ws;
|
|
|
144 #ifdef HAVE_SYNC_BUILTINS
|
|
|
145 do
|
|
|
146 {
|
|
|
147 next_ws = team->work_share_list_free;
|
|
|
148 ws->next_free = next_ws;
|
|
|
149 }
|
|
|
150 while (!__sync_bool_compare_and_swap (&team->work_share_list_free,
|
|
|
151 next_ws, ws));
|
|
|
152 #else
|
|
|
153 gomp_mutex_lock (&team->work_share_list_free_lock);
|
|
|
154 next_ws = team->work_share_list_free;
|
|
|
155 ws->next_free = next_ws;
|
|
|
156 team->work_share_list_free = ws;
|
|
|
157 gomp_mutex_unlock (&team->work_share_list_free_lock);
|
|
|
158 #endif
|
|
|
159 }
|
|
|
160 }
|
|
|
161
|
|
|
162 /* The current thread is ready to begin the next work sharing construct.
|
|
|
163 In all cases, thr->ts.work_share is updated to point to the new
|
|
|
164 structure. In all cases the work_share lock is locked. Return true
|
|
|
165 if this was the first thread to reach this point. */
|
|
|
166
|
|
|
167 bool
|
|
|
168 gomp_work_share_start (bool ordered)
|
|
|
169 {
|
|
|
170 struct gomp_thread *thr = gomp_thread ();
|
|
|
171 struct gomp_team *team = thr->ts.team;
|
|
|
172 struct gomp_work_share *ws;
|
|
|
173
|
|
|
174 /* Work sharing constructs can be orphaned. */
|
|
|
175 if (team == NULL)
|
|
|
176 {
|
|
|
177 ws = gomp_malloc (sizeof (*ws));
|
|
|
178 gomp_init_work_share (ws, ordered, 1);
|
|
|
179 thr->ts.work_share = ws;
|
|
|
180 return ws;
|
|
|
181 }
|
|
|
182
|
|
|
183 ws = thr->ts.work_share;
|
|
|
184 thr->ts.last_work_share = ws;
|
|
|
185 ws = gomp_ptrlock_get (&ws->next_ws);
|
|
|
186 if (ws == NULL)
|
|
|
187 {
|
|
|
188 /* This thread encountered a new ws first. */
|
|
|
189 struct gomp_work_share *ws = alloc_work_share (team);
|
|
|
190 gomp_init_work_share (ws, ordered, team->nthreads);
|
|
|
191 thr->ts.work_share = ws;
|
|
|
192 return true;
|
|
|
193 }
|
|
|
194 else
|
|
|
195 {
|
|
|
196 thr->ts.work_share = ws;
|
|
|
197 return false;
|
|
|
198 }
|
|
|
199 }
|
|
|
200
|
|
|
201 /* The current thread is done with its current work sharing construct.
|
|
|
202 This version does imply a barrier at the end of the work-share. */
|
|
|
203
|
|
|
204 void
|
|
|
205 gomp_work_share_end (void)
|
|
|
206 {
|
|
|
207 struct gomp_thread *thr = gomp_thread ();
|
|
|
208 struct gomp_team *team = thr->ts.team;
|
|
|
209 gomp_barrier_state_t bstate;
|
|
|
210
|
|
|
211 /* Work sharing constructs can be orphaned. */
|
|
|
212 if (team == NULL)
|
|
|
213 {
|
|
|
214 free_work_share (NULL, thr->ts.work_share);
|
|
|
215 thr->ts.work_share = NULL;
|
|
|
216 return;
|
|
|
217 }
|
|
|
218
|
|
|
219 bstate = gomp_barrier_wait_start (&team->barrier);
|
|
|
220
|
|
|
221 if (gomp_barrier_last_thread (bstate))
|
|
|
222 {
|
|
|
223 if (__builtin_expect (thr->ts.last_work_share != NULL, 1))
|
|
|
224 free_work_share (team, thr->ts.last_work_share);
|
|
|
225 }
|
|
|
226
|
|
|
227 gomp_team_barrier_wait_end (&team->barrier, bstate);
|
|
|
228 thr->ts.last_work_share = NULL;
|
|
|
229 }
|
|
|
230
|
|
|
231 /* The current thread is done with its current work sharing construct.
|
|
|
232 This version does NOT imply a barrier at the end of the work-share. */
|
|
|
233
|
|
|
234 void
|
|
|
235 gomp_work_share_end_nowait (void)
|
|
|
236 {
|
|
|
237 struct gomp_thread *thr = gomp_thread ();
|
|
|
238 struct gomp_team *team = thr->ts.team;
|
|
|
239 struct gomp_work_share *ws = thr->ts.work_share;
|
|
|
240 unsigned completed;
|
|
|
241
|
|
|
242 /* Work sharing constructs can be orphaned. */
|
|
|
243 if (team == NULL)
|
|
|
244 {
|
|
|
245 free_work_share (NULL, ws);
|
|
|
246 thr->ts.work_share = NULL;
|
|
|
247 return;
|
|
|
248 }
|
|
|
249
|
|
|
250 if (__builtin_expect (thr->ts.last_work_share == NULL, 0))
|
|
|
251 return;
|
|
|
252
|
|
|
253 #ifdef HAVE_SYNC_BUILTINS
|
|
|
254 completed = __sync_add_and_fetch (&ws->threads_completed, 1);
|
|
|
255 #else
|
|
|
256 gomp_mutex_lock (&ws->lock);
|
|
|
257 completed = ++ws->threads_completed;
|
|
|
258 gomp_mutex_unlock (&ws->lock);
|
|
|
259 #endif
|
|
|
260
|
|
|
261 if (completed == team->nthreads)
|
|
|
262 free_work_share (team, thr->ts.last_work_share);
|
|
|
263 thr->ts.last_work_share = NULL;
|
|
|
264 }
|
