111
|
1 /* OpenACC Runtime Library Definitions.
|
|
2
|
145
|
3 Copyright (C) 2013-2020 Free Software Foundation, Inc.
|
111
|
4
|
|
5 Contributed by Mentor Embedded.
|
|
6
|
|
7 This file is part of the GNU Offloading and Multi Processing Library
|
|
8 (libgomp).
|
|
9
|
|
10 Libgomp is free software; you can redistribute it and/or modify it
|
|
11 under the terms of the GNU General Public License as published by
|
|
12 the Free Software Foundation; either version 3, or (at your option)
|
|
13 any later version.
|
|
14
|
|
15 Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
|
|
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
|
|
17 FOR A PARTICULAR PURPOSE. See the GNU General Public License for
|
|
18 more details.
|
|
19
|
|
20 Under Section 7 of GPL version 3, you are granted additional
|
|
21 permissions described in the GCC Runtime Library Exception, version
|
|
22 3.1, as published by the Free Software Foundation.
|
|
23
|
|
24 You should have received a copy of the GNU General Public License and
|
|
25 a copy of the GCC Runtime Library Exception along with this program;
|
|
26 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
|
27 <http://www.gnu.org/licenses/>. */
|
|
28
|
|
29 #include <assert.h>
|
145
|
30 #include <string.h>
|
111
|
31 #include "openacc.h"
|
|
32 #include "libgomp.h"
|
|
33 #include "oacc-int.h"
|
|
34
|
145
|
35 static struct goacc_thread *
|
|
36 get_goacc_thread (void)
|
111
|
37 {
|
|
38 struct goacc_thread *thr = goacc_thread ();
|
|
39
|
|
40 if (!thr || !thr->dev)
|
|
41 gomp_fatal ("no device active");
|
|
42
|
145
|
43 return thr;
|
|
44 }
|
|
45
|
|
46 static int
|
|
47 validate_async_val (int async)
|
|
48 {
|
|
49 if (!async_valid_p (async))
|
|
50 gomp_fatal ("invalid async-argument: %d", async);
|
|
51
|
|
52 if (async == acc_async_sync)
|
|
53 return -1;
|
|
54
|
|
55 if (async == acc_async_noval)
|
|
56 return 0;
|
|
57
|
|
58 if (async >= 0)
|
|
59 /* TODO: we reserve 0 for acc_async_noval before we can clarify the
|
|
60 semantics of "default_async". */
|
|
61 return 1 + async;
|
|
62 else
|
|
63 __builtin_unreachable ();
|
|
64 }
|
|
65
|
|
66 /* Return the asyncqueue to be used for OpenACC async-argument ASYNC. This
|
|
67 might return NULL if no asyncqueue is to be used. Otherwise, if CREATE,
|
|
68 create the asyncqueue if it doesn't exist yet.
|
|
69
|
|
70 Unless CREATE, this will not generate any OpenACC Profiling Interface
|
|
71 events. */
|
|
72
|
|
73 attribute_hidden struct goacc_asyncqueue *
|
|
74 lookup_goacc_asyncqueue (struct goacc_thread *thr, bool create, int async)
|
|
75 {
|
|
76 async = validate_async_val (async);
|
|
77 if (async < 0)
|
|
78 return NULL;
|
|
79
|
|
80 struct goacc_asyncqueue *ret_aq = NULL;
|
|
81 struct gomp_device_descr *dev = thr->dev;
|
|
82
|
|
83 gomp_mutex_lock (&dev->openacc.async.lock);
|
|
84
|
|
85 if (!create
|
|
86 && (async >= dev->openacc.async.nasyncqueue
|
|
87 || !dev->openacc.async.asyncqueue[async]))
|
|
88 goto end;
|
|
89
|
|
90 if (async >= dev->openacc.async.nasyncqueue)
|
|
91 {
|
|
92 int diff = async + 1 - dev->openacc.async.nasyncqueue;
|
|
93 dev->openacc.async.asyncqueue
|
|
94 = gomp_realloc (dev->openacc.async.asyncqueue,
|
|
95 sizeof (goacc_aq) * (async + 1));
|
|
96 memset (dev->openacc.async.asyncqueue + dev->openacc.async.nasyncqueue,
|
|
97 0, sizeof (goacc_aq) * diff);
|
|
98 dev->openacc.async.nasyncqueue = async + 1;
|
|
99 }
|
|
100
|
|
101 if (!dev->openacc.async.asyncqueue[async])
|
|
102 {
|
|
103 dev->openacc.async.asyncqueue[async]
|
|
104 = dev->openacc.async.construct_func (dev->target_id);
|
|
105
|
|
106 if (!dev->openacc.async.asyncqueue[async])
|
|
107 {
|
|
108 gomp_mutex_unlock (&dev->openacc.async.lock);
|
|
109 gomp_fatal ("async %d creation failed", async);
|
|
110 }
|
|
111
|
|
112 /* Link new async queue into active list. */
|
|
113 goacc_aq_list n = gomp_malloc (sizeof (struct goacc_asyncqueue_list));
|
|
114 n->aq = dev->openacc.async.asyncqueue[async];
|
|
115 n->next = dev->openacc.async.active;
|
|
116 dev->openacc.async.active = n;
|
|
117 }
|
|
118
|
|
119 ret_aq = dev->openacc.async.asyncqueue[async];
|
|
120
|
|
121 end:
|
|
122 gomp_mutex_unlock (&dev->openacc.async.lock);
|
|
123 return ret_aq;
|
|
124 }
|
|
125
|
|
126 /* Return the asyncqueue to be used for OpenACC async-argument ASYNC. This
|
|
127 might return NULL if no asyncqueue is to be used. Otherwise, create the
|
|
128 asyncqueue if it doesn't exist yet. */
|
|
129
|
|
130 attribute_hidden struct goacc_asyncqueue *
|
|
131 get_goacc_asyncqueue (int async)
|
|
132 {
|
|
133 struct goacc_thread *thr = get_goacc_thread ();
|
|
134 return lookup_goacc_asyncqueue (thr, true, async);
|
|
135 }
|
|
136
|
|
137 int
|
|
138 acc_async_test (int async)
|
|
139 {
|
|
140 struct goacc_thread *thr = goacc_thread ();
|
|
141
|
|
142 if (!thr || !thr->dev)
|
|
143 gomp_fatal ("no device active");
|
|
144
|
|
145 goacc_aq aq = lookup_goacc_asyncqueue (thr, false, async);
|
|
146 if (!aq)
|
|
147 return 1;
|
|
148
|
|
149 acc_prof_info prof_info;
|
|
150 acc_api_info api_info;
|
|
151 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
152 if (profiling_p)
|
|
153 {
|
|
154 prof_info.async = async;
|
|
155 prof_info.async_queue = prof_info.async;
|
|
156 }
|
|
157
|
|
158 int res = thr->dev->openacc.async.test_func (aq);
|
|
159
|
|
160 if (profiling_p)
|
|
161 {
|
|
162 thr->prof_info = NULL;
|
|
163 thr->api_info = NULL;
|
|
164 }
|
|
165
|
|
166 return res;
|
|
167 }
|
|
168
|
|
169 int
|
|
170 acc_async_test_all (void)
|
|
171 {
|
|
172 struct goacc_thread *thr = get_goacc_thread ();
|
|
173
|
|
174 acc_prof_info prof_info;
|
|
175 acc_api_info api_info;
|
|
176 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
177
|
|
178 int ret = 1;
|
|
179 gomp_mutex_lock (&thr->dev->openacc.async.lock);
|
|
180 for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
|
|
181 if (!thr->dev->openacc.async.test_func (l->aq))
|
|
182 {
|
|
183 ret = 0;
|
|
184 break;
|
|
185 }
|
|
186 gomp_mutex_unlock (&thr->dev->openacc.async.lock);
|
|
187
|
|
188 if (profiling_p)
|
|
189 {
|
|
190 thr->prof_info = NULL;
|
|
191 thr->api_info = NULL;
|
|
192 }
|
|
193
|
|
194 return ret;
|
111
|
195 }
|
|
196
|
|
197 void
|
|
198 acc_wait (int async)
|
|
199 {
|
145
|
200 struct goacc_thread *thr = get_goacc_thread ();
|
111
|
201
|
145
|
202 goacc_aq aq = lookup_goacc_asyncqueue (thr, false, async);
|
|
203 if (!aq)
|
|
204 return;
|
111
|
205
|
145
|
206 acc_prof_info prof_info;
|
|
207 acc_api_info api_info;
|
|
208 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
209 if (profiling_p)
|
|
210 {
|
|
211 prof_info.async = async;
|
|
212 prof_info.async_queue = prof_info.async;
|
|
213 }
|
111
|
214
|
145
|
215 if (!thr->dev->openacc.async.synchronize_func (aq))
|
|
216 gomp_fatal ("wait on %d failed", async);
|
|
217
|
|
218 if (profiling_p)
|
|
219 {
|
|
220 thr->prof_info = NULL;
|
|
221 thr->api_info = NULL;
|
|
222 }
|
111
|
223 }
|
|
224
|
|
225 /* acc_async_wait is an OpenACC 1.0 compatibility name for acc_wait. */
|
|
226 #ifdef HAVE_ATTRIBUTE_ALIAS
|
|
227 strong_alias (acc_wait, acc_async_wait)
|
|
228 #else
|
|
229 void
|
|
230 acc_async_wait (int async)
|
|
231 {
|
|
232 acc_wait (async);
|
|
233 }
|
|
234 #endif
|
|
235
|
|
236 void
|
|
237 acc_wait_async (int async1, int async2)
|
|
238 {
|
145
|
239 struct goacc_thread *thr = get_goacc_thread ();
|
|
240
|
|
241 goacc_aq aq1 = lookup_goacc_asyncqueue (thr, false, async1);
|
|
242 /* TODO: Is this also correct for acc_async_sync, assuming that in this case,
|
|
243 we'll always be synchronous anyways? */
|
|
244 if (!aq1)
|
|
245 return;
|
|
246
|
|
247 acc_prof_info prof_info;
|
|
248 acc_api_info api_info;
|
|
249 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
250 if (profiling_p)
|
|
251 {
|
|
252 prof_info.async = async2;
|
|
253 prof_info.async_queue = prof_info.async;
|
|
254 }
|
111
|
255
|
145
|
256 goacc_aq aq2 = lookup_goacc_asyncqueue (thr, true, async2);
|
|
257 /* An async queue is always synchronized with itself. */
|
|
258 if (aq1 == aq2)
|
|
259 goto out_prof;
|
111
|
260
|
145
|
261 if (aq2)
|
|
262 {
|
|
263 if (!thr->dev->openacc.async.serialize_func (aq1, aq2))
|
|
264 gomp_fatal ("ordering of async ids %d and %d failed", async1, async2);
|
|
265 }
|
|
266 else
|
|
267 {
|
|
268 /* TODO: Local thread synchronization.
|
|
269 Necessary for the "async2 == acc_async_sync" case, or can just skip? */
|
|
270 if (!thr->dev->openacc.async.synchronize_func (aq1))
|
|
271 gomp_fatal ("wait on %d failed", async1);
|
|
272 }
|
|
273
|
|
274 out_prof:
|
|
275 if (profiling_p)
|
|
276 {
|
|
277 thr->prof_info = NULL;
|
|
278 thr->api_info = NULL;
|
|
279 }
|
111
|
280 }
|
|
281
|
|
282 void
|
|
283 acc_wait_all (void)
|
|
284 {
|
|
285 struct goacc_thread *thr = goacc_thread ();
|
|
286
|
145
|
287 acc_prof_info prof_info;
|
|
288 acc_api_info api_info;
|
|
289 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
111
|
290
|
145
|
291 bool ret = true;
|
|
292 gomp_mutex_lock (&thr->dev->openacc.async.lock);
|
|
293 for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
|
|
294 ret &= thr->dev->openacc.async.synchronize_func (l->aq);
|
|
295 gomp_mutex_unlock (&thr->dev->openacc.async.lock);
|
|
296
|
|
297 if (profiling_p)
|
|
298 {
|
|
299 thr->prof_info = NULL;
|
|
300 thr->api_info = NULL;
|
|
301 }
|
|
302
|
|
303 if (!ret)
|
|
304 gomp_fatal ("wait all failed");
|
111
|
305 }
|
|
306
|
|
307 /* acc_async_wait_all is an OpenACC 1.0 compatibility name for acc_wait_all. */
|
|
308 #ifdef HAVE_ATTRIBUTE_ALIAS
|
|
309 strong_alias (acc_wait_all, acc_async_wait_all)
|
|
310 #else
|
|
311 void
|
|
312 acc_async_wait_all (void)
|
|
313 {
|
|
314 acc_wait_all ();
|
|
315 }
|
|
316 #endif
|
|
317
|
|
318 void
|
|
319 acc_wait_all_async (int async)
|
|
320 {
|
145
|
321 struct goacc_thread *thr = get_goacc_thread ();
|
|
322
|
|
323 acc_prof_info prof_info;
|
|
324 acc_api_info api_info;
|
|
325 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
326 if (profiling_p)
|
|
327 {
|
|
328 prof_info.async = async;
|
|
329 prof_info.async_queue = prof_info.async;
|
|
330 }
|
|
331
|
|
332 goacc_aq waiting_queue = lookup_goacc_asyncqueue (thr, true, async);
|
|
333
|
|
334 bool ret = true;
|
|
335 gomp_mutex_lock (&thr->dev->openacc.async.lock);
|
|
336 for (goacc_aq_list l = thr->dev->openacc.async.active; l; l = l->next)
|
|
337 {
|
|
338 if (waiting_queue)
|
|
339 ret &= thr->dev->openacc.async.serialize_func (l->aq, waiting_queue);
|
|
340 else
|
|
341 /* TODO: Local thread synchronization.
|
|
342 Necessary for the "async2 == acc_async_sync" case, or can just skip? */
|
|
343 ret &= thr->dev->openacc.async.synchronize_func (l->aq);
|
|
344 }
|
|
345 gomp_mutex_unlock (&thr->dev->openacc.async.lock);
|
|
346
|
|
347 if (profiling_p)
|
|
348 {
|
|
349 thr->prof_info = NULL;
|
|
350 thr->api_info = NULL;
|
|
351 }
|
|
352
|
|
353 if (!ret)
|
|
354 gomp_fatal ("wait all async(%d) failed", async);
|
|
355 }
|
|
356
|
|
357 void
|
|
358 GOACC_wait (int async, int num_waits, ...)
|
|
359 {
|
|
360 goacc_lazy_initialize ();
|
111
|
361
|
|
362 struct goacc_thread *thr = goacc_thread ();
|
|
363
|
145
|
364 /* No nesting. */
|
|
365 assert (thr->prof_info == NULL);
|
|
366 assert (thr->api_info == NULL);
|
|
367 acc_prof_info prof_info;
|
|
368 acc_api_info api_info;
|
|
369 bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
|
|
370 if (profiling_p)
|
|
371 {
|
|
372 prof_info.async = async;
|
|
373 prof_info.async_queue = prof_info.async;
|
|
374 }
|
|
375
|
|
376 if (num_waits)
|
|
377 {
|
|
378 va_list ap;
|
|
379
|
|
380 va_start (ap, num_waits);
|
|
381 goacc_wait (async, num_waits, &ap);
|
|
382 va_end (ap);
|
|
383 }
|
|
384 else if (async == acc_async_sync)
|
|
385 acc_wait_all ();
|
|
386 else
|
|
387 acc_wait_all_async (async);
|
|
388
|
|
389 if (profiling_p)
|
|
390 {
|
|
391 thr->prof_info = NULL;
|
|
392 thr->api_info = NULL;
|
|
393 }
|
|
394 }
|
|
395
|
|
396 attribute_hidden void
|
|
397 goacc_wait (int async, int num_waits, va_list *ap)
|
|
398 {
|
|
399 while (num_waits--)
|
|
400 {
|
|
401 int qid = va_arg (*ap, int);
|
|
402
|
|
403 /* Waiting on ACC_ASYNC_NOVAL maps to 'wait all'. */
|
|
404 if (qid == acc_async_noval)
|
|
405 {
|
|
406 if (async == acc_async_sync)
|
|
407 acc_wait_all ();
|
|
408 else
|
|
409 acc_wait_all_async (async);
|
|
410 break;
|
|
411 }
|
|
412
|
|
413 if (acc_async_test (qid))
|
|
414 continue;
|
111
|
415
|
145
|
416 if (async == acc_async_sync)
|
|
417 acc_wait (qid);
|
|
418 else if (qid == async)
|
|
419 /* If we're waiting on the same asynchronous queue as we're
|
|
420 launching on, the queue itself will order work as
|
|
421 required, so there's no need to wait explicitly. */
|
|
422 ;
|
|
423 else
|
|
424 acc_wait_async (qid, async);
|
|
425 }
|
|
426 }
|
|
427
|
|
428 attribute_hidden void
|
|
429 goacc_async_free (struct gomp_device_descr *devicep,
|
|
430 struct goacc_asyncqueue *aq, void *ptr)
|
|
431 {
|
|
432 if (!aq)
|
|
433 free (ptr);
|
|
434 else
|
|
435 devicep->openacc.async.queue_callback_func (aq, free, ptr);
|
111
|
436 }
|
145
|
437
|
|
438 /* This function initializes the asyncqueues for the device specified by
|
|
439 DEVICEP. TODO DEVICEP must be locked on entry, and remains locked on
|
|
440 return. */
|
|
441
|
|
442 attribute_hidden void
|
|
443 goacc_init_asyncqueues (struct gomp_device_descr *devicep)
|
|
444 {
|
|
445 devicep->openacc.async.nasyncqueue = 0;
|
|
446 devicep->openacc.async.asyncqueue = NULL;
|
|
447 devicep->openacc.async.active = NULL;
|
|
448 gomp_mutex_init (&devicep->openacc.async.lock);
|
|
449 }
|
|
450
|
|
451 /* This function finalizes the asyncqueues for the device specified by DEVICEP.
|
|
452 TODO DEVICEP must be locked on entry, and remains locked on return. */
|
|
453
|
|
454 attribute_hidden bool
|
|
455 goacc_fini_asyncqueues (struct gomp_device_descr *devicep)
|
|
456 {
|
|
457 bool ret = true;
|
|
458 gomp_mutex_lock (&devicep->openacc.async.lock);
|
|
459 if (devicep->openacc.async.nasyncqueue > 0)
|
|
460 {
|
|
461 goacc_aq_list next;
|
|
462 for (goacc_aq_list l = devicep->openacc.async.active; l; l = next)
|
|
463 {
|
|
464 ret &= devicep->openacc.async.destruct_func (l->aq);
|
|
465 next = l->next;
|
|
466 free (l);
|
|
467 }
|
|
468 free (devicep->openacc.async.asyncqueue);
|
|
469 devicep->openacc.async.nasyncqueue = 0;
|
|
470 devicep->openacc.async.asyncqueue = NULL;
|
|
471 devicep->openacc.async.active = NULL;
|
|
472 }
|
|
473 gomp_mutex_unlock (&devicep->openacc.async.lock);
|
|
474 gomp_mutex_destroy (&devicep->openacc.async.lock);
|
|
475 return ret;
|
|
476 }
|