|
111
|
1 /* OpenACC Runtime initialization routines
|
|
|
2
|
|
131
|
3 Copyright (C) 2013-2018 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 "openacc.h"
|
|
|
30 #include "config.h"
|
|
|
31 #include "libgomp.h"
|
|
|
32 #include "gomp-constants.h"
|
|
|
33 #include "oacc-int.h"
|
|
|
34 #include <stdint.h>
|
|
|
35 #include <string.h>
|
|
|
36 #include <assert.h>
|
|
|
37
|
|
|
38 /* Return block containing [H->S), or NULL if not contained. The device lock
|
|
|
39 for DEV must be locked on entry, and remains locked on exit. */
|
|
|
40
|
|
|
41 static splay_tree_key
|
|
|
42 lookup_host (struct gomp_device_descr *dev, void *h, size_t s)
|
|
|
43 {
|
|
|
44 struct splay_tree_key_s node;
|
|
|
45 splay_tree_key key;
|
|
|
46
|
|
|
47 node.host_start = (uintptr_t) h;
|
|
|
48 node.host_end = (uintptr_t) h + s;
|
|
|
49
|
|
|
50 key = splay_tree_lookup (&dev->mem_map, &node);
|
|
|
51
|
|
|
52 return key;
|
|
|
53 }
|
|
|
54
|
|
|
55 /* Return block containing [D->S), or NULL if not contained.
|
|
|
56 The list isn't ordered by device address, so we have to iterate
|
|
|
57 over the whole array. This is not expected to be a common
|
|
|
58 operation. The device lock associated with TGT must be locked on entry, and
|
|
|
59 remains locked on exit. */
|
|
|
60
|
|
|
61 static splay_tree_key
|
|
|
62 lookup_dev (struct target_mem_desc *tgt, void *d, size_t s)
|
|
|
63 {
|
|
|
64 int i;
|
|
|
65 struct target_mem_desc *t;
|
|
|
66
|
|
|
67 if (!tgt)
|
|
|
68 return NULL;
|
|
|
69
|
|
|
70 for (t = tgt; t != NULL; t = t->prev)
|
|
|
71 {
|
|
|
72 if (t->tgt_start <= (uintptr_t) d && t->tgt_end >= (uintptr_t) d + s)
|
|
|
73 break;
|
|
|
74 }
|
|
|
75
|
|
|
76 if (!t)
|
|
|
77 return NULL;
|
|
|
78
|
|
|
79 for (i = 0; i < t->list_count; i++)
|
|
|
80 {
|
|
|
81 void * offset;
|
|
|
82
|
|
|
83 splay_tree_key k = &t->array[i].key;
|
|
|
84 offset = d - t->tgt_start + k->tgt_offset;
|
|
|
85
|
|
|
86 if (k->host_start + offset <= (void *) k->host_end)
|
|
|
87 return k;
|
|
|
88 }
|
|
|
89
|
|
|
90 return NULL;
|
|
|
91 }
|
|
|
92
|
|
|
93 /* OpenACC is silent on how memory exhaustion is indicated. We return
|
|
|
94 NULL. */
|
|
|
95
|
|
|
96 void *
|
|
|
97 acc_malloc (size_t s)
|
|
|
98 {
|
|
|
99 if (!s)
|
|
|
100 return NULL;
|
|
|
101
|
|
|
102 goacc_lazy_initialize ();
|
|
|
103
|
|
|
104 struct goacc_thread *thr = goacc_thread ();
|
|
|
105
|
|
|
106 assert (thr->dev);
|
|
|
107
|
|
|
108 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
109 return malloc (s);
|
|
|
110
|
|
|
111 return thr->dev->alloc_func (thr->dev->target_id, s);
|
|
|
112 }
|
|
|
113
|
|
|
114 /* OpenACC 2.0a (3.2.16) doesn't specify what to do in the event
|
|
|
115 the device address is mapped. We choose to check if it mapped,
|
|
|
116 and if it is, to unmap it. */
|
|
|
117 void
|
|
|
118 acc_free (void *d)
|
|
|
119 {
|
|
|
120 splay_tree_key k;
|
|
|
121
|
|
|
122 if (!d)
|
|
|
123 return;
|
|
|
124
|
|
|
125 struct goacc_thread *thr = goacc_thread ();
|
|
|
126
|
|
|
127 assert (thr && thr->dev);
|
|
|
128
|
|
|
129 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
130
|
|
|
131 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
132 return free (d);
|
|
|
133
|
|
|
134 gomp_mutex_lock (&acc_dev->lock);
|
|
|
135
|
|
|
136 /* We don't have to call lazy open here, as the ptr value must have
|
|
|
137 been returned by acc_malloc. It's not permitted to pass NULL in
|
|
|
138 (unless you got that null from acc_malloc). */
|
|
|
139 if ((k = lookup_dev (acc_dev->openacc.data_environ, d, 1)))
|
|
|
140 {
|
|
|
141 void *offset;
|
|
|
142
|
|
|
143 offset = d - k->tgt->tgt_start + k->tgt_offset;
|
|
|
144
|
|
|
145 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
146
|
|
|
147 acc_unmap_data ((void *)(k->host_start + offset));
|
|
|
148 }
|
|
|
149 else
|
|
|
150 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
151
|
|
|
152 if (!acc_dev->free_func (acc_dev->target_id, d))
|
|
|
153 gomp_fatal ("error in freeing device memory in %s", __FUNCTION__);
|
|
|
154 }
|
|
|
155
|
|
|
156 void
|
|
|
157 acc_memcpy_to_device (void *d, void *h, size_t s)
|
|
|
158 {
|
|
|
159 /* No need to call lazy open here, as the device pointer must have
|
|
|
160 been obtained from a routine that did that. */
|
|
|
161 struct goacc_thread *thr = goacc_thread ();
|
|
|
162
|
|
|
163 assert (thr && thr->dev);
|
|
|
164
|
|
|
165 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
166 {
|
|
|
167 memmove (d, h, s);
|
|
|
168 return;
|
|
|
169 }
|
|
|
170
|
|
|
171 if (!thr->dev->host2dev_func (thr->dev->target_id, d, h, s))
|
|
|
172 gomp_fatal ("error in %s", __FUNCTION__);
|
|
|
173 }
|
|
|
174
|
|
|
175 void
|
|
|
176 acc_memcpy_from_device (void *h, void *d, size_t s)
|
|
|
177 {
|
|
|
178 /* No need to call lazy open here, as the device pointer must have
|
|
|
179 been obtained from a routine that did that. */
|
|
|
180 struct goacc_thread *thr = goacc_thread ();
|
|
|
181
|
|
|
182 assert (thr && thr->dev);
|
|
|
183
|
|
|
184 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
185 {
|
|
|
186 memmove (h, d, s);
|
|
|
187 return;
|
|
|
188 }
|
|
|
189
|
|
|
190 if (!thr->dev->dev2host_func (thr->dev->target_id, h, d, s))
|
|
|
191 gomp_fatal ("error in %s", __FUNCTION__);
|
|
|
192 }
|
|
|
193
|
|
|
194 /* Return the device pointer that corresponds to host data H. Or NULL
|
|
|
195 if no mapping. */
|
|
|
196
|
|
|
197 void *
|
|
|
198 acc_deviceptr (void *h)
|
|
|
199 {
|
|
|
200 splay_tree_key n;
|
|
|
201 void *d;
|
|
|
202 void *offset;
|
|
|
203
|
|
|
204 goacc_lazy_initialize ();
|
|
|
205
|
|
|
206 struct goacc_thread *thr = goacc_thread ();
|
|
|
207 struct gomp_device_descr *dev = thr->dev;
|
|
|
208
|
|
|
209 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
210 return h;
|
|
|
211
|
|
|
212 gomp_mutex_lock (&dev->lock);
|
|
|
213
|
|
|
214 n = lookup_host (dev, h, 1);
|
|
|
215
|
|
|
216 if (!n)
|
|
|
217 {
|
|
|
218 gomp_mutex_unlock (&dev->lock);
|
|
|
219 return NULL;
|
|
|
220 }
|
|
|
221
|
|
|
222 offset = h - n->host_start;
|
|
|
223
|
|
|
224 d = n->tgt->tgt_start + n->tgt_offset + offset;
|
|
|
225
|
|
|
226 gomp_mutex_unlock (&dev->lock);
|
|
|
227
|
|
|
228 return d;
|
|
|
229 }
|
|
|
230
|
|
|
231 /* Return the host pointer that corresponds to device data D. Or NULL
|
|
|
232 if no mapping. */
|
|
|
233
|
|
|
234 void *
|
|
|
235 acc_hostptr (void *d)
|
|
|
236 {
|
|
|
237 splay_tree_key n;
|
|
|
238 void *h;
|
|
|
239 void *offset;
|
|
|
240
|
|
|
241 goacc_lazy_initialize ();
|
|
|
242
|
|
|
243 struct goacc_thread *thr = goacc_thread ();
|
|
|
244 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
245
|
|
|
246 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
247 return d;
|
|
|
248
|
|
|
249 gomp_mutex_lock (&acc_dev->lock);
|
|
|
250
|
|
|
251 n = lookup_dev (acc_dev->openacc.data_environ, d, 1);
|
|
|
252
|
|
|
253 if (!n)
|
|
|
254 {
|
|
|
255 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
256 return NULL;
|
|
|
257 }
|
|
|
258
|
|
|
259 offset = d - n->tgt->tgt_start + n->tgt_offset;
|
|
|
260
|
|
|
261 h = n->host_start + offset;
|
|
|
262
|
|
|
263 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
264
|
|
|
265 return h;
|
|
|
266 }
|
|
|
267
|
|
|
268 /* Return 1 if host data [H,+S] is present on the device. */
|
|
|
269
|
|
|
270 int
|
|
|
271 acc_is_present (void *h, size_t s)
|
|
|
272 {
|
|
|
273 splay_tree_key n;
|
|
|
274
|
|
|
275 if (!s || !h)
|
|
|
276 return 0;
|
|
|
277
|
|
|
278 goacc_lazy_initialize ();
|
|
|
279
|
|
|
280 struct goacc_thread *thr = goacc_thread ();
|
|
|
281 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
282
|
|
|
283 if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
284 return h != NULL;
|
|
|
285
|
|
|
286 gomp_mutex_lock (&acc_dev->lock);
|
|
|
287
|
|
|
288 n = lookup_host (acc_dev, h, s);
|
|
|
289
|
|
|
290 if (n && ((uintptr_t)h < n->host_start
|
|
|
291 || (uintptr_t)h + s > n->host_end
|
|
|
292 || s > n->host_end - n->host_start))
|
|
|
293 n = NULL;
|
|
|
294
|
|
|
295 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
296
|
|
|
297 return n != NULL;
|
|
|
298 }
|
|
|
299
|
|
|
300 /* Create a mapping for host [H,+S] -> device [D,+S] */
|
|
|
301
|
|
|
302 void
|
|
|
303 acc_map_data (void *h, void *d, size_t s)
|
|
|
304 {
|
|
|
305 struct target_mem_desc *tgt = NULL;
|
|
|
306 size_t mapnum = 1;
|
|
|
307 void *hostaddrs = h;
|
|
|
308 void *devaddrs = d;
|
|
|
309 size_t sizes = s;
|
|
|
310 unsigned short kinds = GOMP_MAP_ALLOC;
|
|
|
311
|
|
|
312 goacc_lazy_initialize ();
|
|
|
313
|
|
|
314 struct goacc_thread *thr = goacc_thread ();
|
|
|
315 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
316
|
|
|
317 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
318 {
|
|
|
319 if (d != h)
|
|
|
320 gomp_fatal ("cannot map data on shared-memory system");
|
|
|
321 }
|
|
|
322 else
|
|
|
323 {
|
|
|
324 struct goacc_thread *thr = goacc_thread ();
|
|
|
325
|
|
|
326 if (!d || !h || !s)
|
|
|
327 gomp_fatal ("[%p,+%d]->[%p,+%d] is a bad map",
|
|
|
328 (void *)h, (int)s, (void *)d, (int)s);
|
|
|
329
|
|
|
330 gomp_mutex_lock (&acc_dev->lock);
|
|
|
331
|
|
|
332 if (lookup_host (acc_dev, h, s))
|
|
|
333 {
|
|
|
334 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
335 gomp_fatal ("host address [%p, +%d] is already mapped", (void *)h,
|
|
|
336 (int)s);
|
|
|
337 }
|
|
|
338
|
|
|
339 if (lookup_dev (thr->dev->openacc.data_environ, d, s))
|
|
|
340 {
|
|
|
341 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
342 gomp_fatal ("device address [%p, +%d] is already mapped", (void *)d,
|
|
|
343 (int)s);
|
|
|
344 }
|
|
|
345
|
|
|
346 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
347
|
|
|
348 tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, &devaddrs, &sizes,
|
|
|
349 &kinds, true, GOMP_MAP_VARS_OPENACC);
|
|
131
|
350 tgt->list[0].key->refcount = REFCOUNT_INFINITY;
|
|
111
|
351 }
|
|
|
352
|
|
|
353 gomp_mutex_lock (&acc_dev->lock);
|
|
|
354 tgt->prev = acc_dev->openacc.data_environ;
|
|
|
355 acc_dev->openacc.data_environ = tgt;
|
|
|
356 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
357 }
|
|
|
358
|
|
|
359 void
|
|
|
360 acc_unmap_data (void *h)
|
|
|
361 {
|
|
|
362 struct goacc_thread *thr = goacc_thread ();
|
|
|
363 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
364
|
|
|
365 /* No need to call lazy open, as the address must have been mapped. */
|
|
|
366
|
|
|
367 /* This is a no-op on shared-memory targets. */
|
|
|
368 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
369 return;
|
|
|
370
|
|
|
371 size_t host_size;
|
|
|
372
|
|
|
373 gomp_mutex_lock (&acc_dev->lock);
|
|
|
374
|
|
|
375 splay_tree_key n = lookup_host (acc_dev, h, 1);
|
|
|
376 struct target_mem_desc *t;
|
|
|
377
|
|
|
378 if (!n)
|
|
|
379 {
|
|
|
380 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
381 gomp_fatal ("%p is not a mapped block", (void *)h);
|
|
|
382 }
|
|
|
383
|
|
|
384 host_size = n->host_end - n->host_start;
|
|
|
385
|
|
|
386 if (n->host_start != (uintptr_t) h)
|
|
|
387 {
|
|
|
388 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
389 gomp_fatal ("[%p,%d] surrounds %p",
|
|
|
390 (void *) n->host_start, (int) host_size, (void *) h);
|
|
|
391 }
|
|
|
392
|
|
131
|
393 /* Mark for removal. */
|
|
|
394 n->refcount = 1;
|
|
|
395
|
|
111
|
396 t = n->tgt;
|
|
|
397
|
|
|
398 if (t->refcount == 2)
|
|
|
399 {
|
|
|
400 struct target_mem_desc *tp;
|
|
|
401
|
|
|
402 /* This is the last reference, so pull the descriptor off the
|
|
|
403 chain. This avoids gomp_unmap_vars via gomp_unmap_tgt from
|
|
|
404 freeing the device memory. */
|
|
|
405 t->tgt_end = 0;
|
|
|
406 t->to_free = 0;
|
|
|
407
|
|
|
408 for (tp = NULL, t = acc_dev->openacc.data_environ; t != NULL;
|
|
|
409 tp = t, t = t->prev)
|
|
|
410 if (n->tgt == t)
|
|
|
411 {
|
|
|
412 if (tp)
|
|
|
413 tp->prev = t->prev;
|
|
|
414 else
|
|
|
415 acc_dev->openacc.data_environ = t->prev;
|
|
|
416
|
|
|
417 break;
|
|
|
418 }
|
|
|
419 }
|
|
|
420
|
|
|
421 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
422
|
|
|
423 gomp_unmap_vars (t, true);
|
|
|
424 }
|
|
|
425
|
|
|
426 #define FLAG_PRESENT (1 << 0)
|
|
|
427 #define FLAG_CREATE (1 << 1)
|
|
|
428 #define FLAG_COPY (1 << 2)
|
|
|
429
|
|
|
430 static void *
|
|
|
431 present_create_copy (unsigned f, void *h, size_t s)
|
|
|
432 {
|
|
|
433 void *d;
|
|
|
434 splay_tree_key n;
|
|
|
435
|
|
|
436 if (!h || !s)
|
|
|
437 gomp_fatal ("[%p,+%d] is a bad range", (void *)h, (int)s);
|
|
|
438
|
|
|
439 goacc_lazy_initialize ();
|
|
|
440
|
|
|
441 struct goacc_thread *thr = goacc_thread ();
|
|
|
442 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
443
|
|
|
444 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
445 return h;
|
|
|
446
|
|
|
447 gomp_mutex_lock (&acc_dev->lock);
|
|
|
448
|
|
|
449 n = lookup_host (acc_dev, h, s);
|
|
|
450 if (n)
|
|
|
451 {
|
|
|
452 /* Present. */
|
|
|
453 d = (void *) (n->tgt->tgt_start + n->tgt_offset);
|
|
|
454
|
|
|
455 if (!(f & FLAG_PRESENT))
|
|
|
456 {
|
|
|
457 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
458 gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
|
|
|
459 (void *)h, (int)s, (void *)d, (int)s);
|
|
|
460 }
|
|
|
461 if ((h + s) > (void *)n->host_end)
|
|
|
462 {
|
|
|
463 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
464 gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
|
|
|
465 }
|
|
|
466
|
|
131
|
467 if (n->refcount != REFCOUNT_INFINITY)
|
|
|
468 {
|
|
|
469 n->refcount++;
|
|
|
470 n->dynamic_refcount++;
|
|
|
471 }
|
|
111
|
472 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
473 }
|
|
|
474 else if (!(f & FLAG_CREATE))
|
|
|
475 {
|
|
|
476 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
477 gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
|
|
|
478 }
|
|
|
479 else
|
|
|
480 {
|
|
|
481 struct target_mem_desc *tgt;
|
|
|
482 size_t mapnum = 1;
|
|
|
483 unsigned short kinds;
|
|
|
484 void *hostaddrs = h;
|
|
|
485
|
|
|
486 if (f & FLAG_COPY)
|
|
|
487 kinds = GOMP_MAP_TO;
|
|
|
488 else
|
|
|
489 kinds = GOMP_MAP_ALLOC;
|
|
|
490
|
|
|
491 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
492
|
|
|
493 tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, NULL, &s, &kinds, true,
|
|
|
494 GOMP_MAP_VARS_OPENACC);
|
|
131
|
495 /* Initialize dynamic refcount. */
|
|
|
496 tgt->list[0].key->dynamic_refcount = 1;
|
|
111
|
497
|
|
|
498 gomp_mutex_lock (&acc_dev->lock);
|
|
|
499
|
|
|
500 d = tgt->to_free;
|
|
|
501 tgt->prev = acc_dev->openacc.data_environ;
|
|
|
502 acc_dev->openacc.data_environ = tgt;
|
|
|
503
|
|
|
504 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
505 }
|
|
|
506
|
|
|
507 return d;
|
|
|
508 }
|
|
|
509
|
|
|
510 void *
|
|
|
511 acc_create (void *h, size_t s)
|
|
|
512 {
|
|
131
|
513 return present_create_copy (FLAG_PRESENT | FLAG_CREATE, h, s);
|
|
111
|
514 }
|
|
|
515
|
|
|
516 void *
|
|
|
517 acc_copyin (void *h, size_t s)
|
|
|
518 {
|
|
131
|
519 return present_create_copy (FLAG_PRESENT | FLAG_CREATE | FLAG_COPY, h, s);
|
|
111
|
520 }
|
|
|
521
|
|
|
522 void *
|
|
|
523 acc_present_or_create (void *h, size_t s)
|
|
|
524 {
|
|
|
525 return present_create_copy (FLAG_PRESENT | FLAG_CREATE, h, s);
|
|
|
526 }
|
|
|
527
|
|
|
528 /* acc_pcreate is acc_present_or_create by a different name. */
|
|
|
529 #ifdef HAVE_ATTRIBUTE_ALIAS
|
|
|
530 strong_alias (acc_present_or_create, acc_pcreate)
|
|
|
531 #else
|
|
|
532 void *
|
|
|
533 acc_pcreate (void *h, size_t s)
|
|
|
534 {
|
|
|
535 return acc_present_or_create (h, s);
|
|
|
536 }
|
|
|
537 #endif
|
|
|
538
|
|
|
539 void *
|
|
|
540 acc_present_or_copyin (void *h, size_t s)
|
|
|
541 {
|
|
|
542 return present_create_copy (FLAG_PRESENT | FLAG_CREATE | FLAG_COPY, h, s);
|
|
|
543 }
|
|
|
544
|
|
|
545 /* acc_pcopyin is acc_present_or_copyin by a different name. */
|
|
|
546 #ifdef HAVE_ATTRIBUTE_ALIAS
|
|
|
547 strong_alias (acc_present_or_copyin, acc_pcopyin)
|
|
|
548 #else
|
|
|
549 void *
|
|
|
550 acc_pcopyin (void *h, size_t s)
|
|
|
551 {
|
|
|
552 return acc_present_or_copyin (h, s);
|
|
|
553 }
|
|
|
554 #endif
|
|
|
555
|
|
131
|
556 #define FLAG_COPYOUT (1 << 0)
|
|
|
557 #define FLAG_FINALIZE (1 << 1)
|
|
111
|
558
|
|
|
559 static void
|
|
|
560 delete_copyout (unsigned f, void *h, size_t s, const char *libfnname)
|
|
|
561 {
|
|
|
562 size_t host_size;
|
|
|
563 splay_tree_key n;
|
|
|
564 void *d;
|
|
|
565 struct goacc_thread *thr = goacc_thread ();
|
|
|
566 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
567
|
|
|
568 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
569 return;
|
|
|
570
|
|
|
571 gomp_mutex_lock (&acc_dev->lock);
|
|
|
572
|
|
|
573 n = lookup_host (acc_dev, h, s);
|
|
|
574
|
|
|
575 /* No need to call lazy open, as the data must already have been
|
|
|
576 mapped. */
|
|
|
577
|
|
|
578 if (!n)
|
|
|
579 {
|
|
|
580 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
581 gomp_fatal ("[%p,%d] is not mapped", (void *)h, (int)s);
|
|
|
582 }
|
|
|
583
|
|
|
584 d = (void *) (n->tgt->tgt_start + n->tgt_offset
|
|
|
585 + (uintptr_t) h - n->host_start);
|
|
|
586
|
|
|
587 host_size = n->host_end - n->host_start;
|
|
|
588
|
|
|
589 if (n->host_start != (uintptr_t) h || host_size != s)
|
|
|
590 {
|
|
|
591 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
592 gomp_fatal ("[%p,%d] surrounds2 [%p,+%d]",
|
|
|
593 (void *) n->host_start, (int) host_size, (void *) h, (int) s);
|
|
|
594 }
|
|
|
595
|
|
131
|
596 if (n->refcount == REFCOUNT_INFINITY)
|
|
|
597 {
|
|
|
598 n->refcount = 0;
|
|
|
599 n->dynamic_refcount = 0;
|
|
|
600 }
|
|
|
601 if (n->refcount < n->dynamic_refcount)
|
|
|
602 {
|
|
|
603 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
604 gomp_fatal ("Dynamic reference counting assert fail\n");
|
|
|
605 }
|
|
111
|
606
|
|
131
|
607 if (f & FLAG_FINALIZE)
|
|
|
608 {
|
|
|
609 n->refcount -= n->dynamic_refcount;
|
|
|
610 n->dynamic_refcount = 0;
|
|
|
611 }
|
|
|
612 else if (n->dynamic_refcount)
|
|
|
613 {
|
|
|
614 n->dynamic_refcount--;
|
|
|
615 n->refcount--;
|
|
|
616 }
|
|
111
|
617
|
|
131
|
618 if (n->refcount == 0)
|
|
|
619 {
|
|
|
620 if (n->tgt->refcount == 2)
|
|
|
621 {
|
|
|
622 struct target_mem_desc *tp, *t;
|
|
|
623 for (tp = NULL, t = acc_dev->openacc.data_environ; t != NULL;
|
|
|
624 tp = t, t = t->prev)
|
|
|
625 if (n->tgt == t)
|
|
|
626 {
|
|
|
627 if (tp)
|
|
|
628 tp->prev = t->prev;
|
|
|
629 else
|
|
|
630 acc_dev->openacc.data_environ = t->prev;
|
|
|
631 break;
|
|
|
632 }
|
|
|
633 }
|
|
111
|
634
|
|
131
|
635 if (f & FLAG_COPYOUT)
|
|
|
636 acc_dev->dev2host_func (acc_dev->target_id, h, d, s);
|
|
|
637
|
|
|
638 gomp_remove_var (acc_dev, n);
|
|
|
639 }
|
|
|
640
|
|
|
641 gomp_mutex_unlock (&acc_dev->lock);
|
|
111
|
642 }
|
|
|
643
|
|
|
644 void
|
|
|
645 acc_delete (void *h , size_t s)
|
|
|
646 {
|
|
|
647 delete_copyout (0, h, s, __FUNCTION__);
|
|
|
648 }
|
|
|
649
|
|
|
650 void
|
|
131
|
651 acc_delete_finalize (void *h , size_t s)
|
|
|
652 {
|
|
|
653 delete_copyout (FLAG_FINALIZE, h, s, __FUNCTION__);
|
|
|
654 }
|
|
|
655
|
|
|
656 void
|
|
|
657 acc_delete_finalize_async (void *h , size_t s, int async)
|
|
|
658 {
|
|
|
659 delete_copyout (FLAG_FINALIZE, h, s, __FUNCTION__);
|
|
|
660 }
|
|
|
661
|
|
|
662 void
|
|
111
|
663 acc_copyout (void *h, size_t s)
|
|
|
664 {
|
|
|
665 delete_copyout (FLAG_COPYOUT, h, s, __FUNCTION__);
|
|
|
666 }
|
|
|
667
|
|
131
|
668 void
|
|
|
669 acc_copyout_finalize (void *h, size_t s)
|
|
|
670 {
|
|
|
671 delete_copyout (FLAG_COPYOUT | FLAG_FINALIZE, h, s, __FUNCTION__);
|
|
|
672 }
|
|
|
673
|
|
|
674 void
|
|
|
675 acc_copyout_finalize_async (void *h, size_t s, int async)
|
|
|
676 {
|
|
|
677 delete_copyout (FLAG_COPYOUT | FLAG_FINALIZE, h, s, __FUNCTION__);
|
|
|
678 }
|
|
|
679
|
|
111
|
680 static void
|
|
|
681 update_dev_host (int is_dev, void *h, size_t s)
|
|
|
682 {
|
|
|
683 splay_tree_key n;
|
|
|
684 void *d;
|
|
|
685
|
|
|
686 goacc_lazy_initialize ();
|
|
|
687
|
|
|
688 struct goacc_thread *thr = goacc_thread ();
|
|
|
689 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
690
|
|
|
691 if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
|
|
|
692 return;
|
|
|
693
|
|
|
694 gomp_mutex_lock (&acc_dev->lock);
|
|
|
695
|
|
|
696 n = lookup_host (acc_dev, h, s);
|
|
|
697
|
|
|
698 if (!n)
|
|
|
699 {
|
|
|
700 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
701 gomp_fatal ("[%p,%d] is not mapped", h, (int)s);
|
|
|
702 }
|
|
|
703
|
|
|
704 d = (void *) (n->tgt->tgt_start + n->tgt_offset
|
|
|
705 + (uintptr_t) h - n->host_start);
|
|
|
706
|
|
|
707 if (is_dev)
|
|
|
708 acc_dev->host2dev_func (acc_dev->target_id, d, h, s);
|
|
|
709 else
|
|
|
710 acc_dev->dev2host_func (acc_dev->target_id, h, d, s);
|
|
|
711
|
|
|
712 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
713 }
|
|
|
714
|
|
|
715 void
|
|
|
716 acc_update_device (void *h, size_t s)
|
|
|
717 {
|
|
|
718 update_dev_host (1, h, s);
|
|
|
719 }
|
|
|
720
|
|
|
721 void
|
|
|
722 acc_update_self (void *h, size_t s)
|
|
|
723 {
|
|
|
724 update_dev_host (0, h, s);
|
|
|
725 }
|
|
|
726
|
|
|
727 void
|
|
|
728 gomp_acc_insert_pointer (size_t mapnum, void **hostaddrs, size_t *sizes,
|
|
|
729 void *kinds)
|
|
|
730 {
|
|
|
731 struct target_mem_desc *tgt;
|
|
|
732 struct goacc_thread *thr = goacc_thread ();
|
|
|
733 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
734
|
|
131
|
735 if (acc_is_present (*hostaddrs, *sizes))
|
|
|
736 {
|
|
|
737 splay_tree_key n;
|
|
|
738 gomp_mutex_lock (&acc_dev->lock);
|
|
|
739 n = lookup_host (acc_dev, *hostaddrs, *sizes);
|
|
|
740 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
741
|
|
|
742 tgt = n->tgt;
|
|
|
743 for (size_t i = 0; i < tgt->list_count; i++)
|
|
|
744 if (tgt->list[i].key == n)
|
|
|
745 {
|
|
|
746 for (size_t j = 0; j < mapnum; j++)
|
|
|
747 if (i + j < tgt->list_count && tgt->list[i + j].key)
|
|
|
748 {
|
|
|
749 tgt->list[i + j].key->refcount++;
|
|
|
750 tgt->list[i + j].key->dynamic_refcount++;
|
|
|
751 }
|
|
|
752 return;
|
|
|
753 }
|
|
|
754 /* Should not reach here. */
|
|
|
755 gomp_fatal ("Dynamic refcount incrementing failed for pointer/pset");
|
|
|
756 }
|
|
|
757
|
|
111
|
758 gomp_debug (0, " %s: prepare mappings\n", __FUNCTION__);
|
|
|
759 tgt = gomp_map_vars (acc_dev, mapnum, hostaddrs,
|
|
|
760 NULL, sizes, kinds, true, GOMP_MAP_VARS_OPENACC);
|
|
|
761 gomp_debug (0, " %s: mappings prepared\n", __FUNCTION__);
|
|
|
762
|
|
131
|
763 /* Initialize dynamic refcount. */
|
|
|
764 tgt->list[0].key->dynamic_refcount = 1;
|
|
|
765
|
|
111
|
766 gomp_mutex_lock (&acc_dev->lock);
|
|
|
767 tgt->prev = acc_dev->openacc.data_environ;
|
|
|
768 acc_dev->openacc.data_environ = tgt;
|
|
|
769 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
770 }
|
|
|
771
|
|
|
772 void
|
|
131
|
773 gomp_acc_remove_pointer (void *h, size_t s, bool force_copyfrom, int async,
|
|
|
774 int finalize, int mapnum)
|
|
111
|
775 {
|
|
|
776 struct goacc_thread *thr = goacc_thread ();
|
|
|
777 struct gomp_device_descr *acc_dev = thr->dev;
|
|
|
778 splay_tree_key n;
|
|
|
779 struct target_mem_desc *t;
|
|
|
780 int minrefs = (mapnum == 1) ? 2 : 3;
|
|
|
781
|
|
131
|
782 if (!acc_is_present (h, s))
|
|
|
783 return;
|
|
|
784
|
|
111
|
785 gomp_mutex_lock (&acc_dev->lock);
|
|
|
786
|
|
|
787 n = lookup_host (acc_dev, h, 1);
|
|
|
788
|
|
|
789 if (!n)
|
|
|
790 {
|
|
|
791 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
792 gomp_fatal ("%p is not a mapped block", (void *)h);
|
|
|
793 }
|
|
|
794
|
|
|
795 gomp_debug (0, " %s: restore mappings\n", __FUNCTION__);
|
|
|
796
|
|
|
797 t = n->tgt;
|
|
|
798
|
|
131
|
799 if (n->refcount < n->dynamic_refcount)
|
|
111
|
800 {
|
|
131
|
801 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
802 gomp_fatal ("Dynamic reference counting assert fail\n");
|
|
111
|
803 }
|
|
|
804
|
|
131
|
805 if (finalize)
|
|
|
806 {
|
|
|
807 n->refcount -= n->dynamic_refcount;
|
|
|
808 n->dynamic_refcount = 0;
|
|
|
809 }
|
|
|
810 else if (n->dynamic_refcount)
|
|
|
811 {
|
|
|
812 n->dynamic_refcount--;
|
|
|
813 n->refcount--;
|
|
|
814 }
|
|
111
|
815
|
|
|
816 gomp_mutex_unlock (&acc_dev->lock);
|
|
|
817
|
|
131
|
818 if (n->refcount == 0)
|
|
|
819 {
|
|
|
820 if (t->refcount == minrefs)
|
|
|
821 {
|
|
|
822 /* This is the last reference, so pull the descriptor off the
|
|
|
823 chain. This prevents gomp_unmap_vars via gomp_unmap_tgt from
|
|
|
824 freeing the device memory. */
|
|
|
825 struct target_mem_desc *tp;
|
|
|
826 for (tp = NULL, t = acc_dev->openacc.data_environ; t != NULL;
|
|
|
827 tp = t, t = t->prev)
|
|
|
828 {
|
|
|
829 if (n->tgt == t)
|
|
|
830 {
|
|
|
831 if (tp)
|
|
|
832 tp->prev = t->prev;
|
|
|
833 else
|
|
|
834 acc_dev->openacc.data_environ = t->prev;
|
|
|
835 break;
|
|
|
836 }
|
|
|
837 }
|
|
|
838 }
|
|
|
839
|
|
|
840 /* Set refcount to 1 to allow gomp_unmap_vars to unmap it. */
|
|
|
841 n->refcount = 1;
|
|
|
842 t->refcount = minrefs;
|
|
|
843 for (size_t i = 0; i < t->list_count; i++)
|
|
|
844 if (t->list[i].key == n)
|
|
|
845 {
|
|
|
846 t->list[i].copy_from = force_copyfrom ? 1 : 0;
|
|
|
847 break;
|
|
|
848 }
|
|
|
849
|
|
|
850 /* If running synchronously, unmap immediately. */
|
|
|
851 if (async < acc_async_noval)
|
|
|
852 gomp_unmap_vars (t, true);
|
|
|
853 else
|
|
|
854 t->device_descr->openacc.register_async_cleanup_func (t, async);
|
|
|
855 }
|
|
|
856
|
|
|
857 gomp_mutex_unlock (&acc_dev->lock);
|
|
111
|
858
|
|
|
859 gomp_debug (0, " %s: mappings restored\n", __FUNCTION__);
|
|
|
860 }
|
