1 /* OpenACC Runtime initialization routines
3 Copyright (C) 2013-2019 Free Software Foundation, Inc.
5 Contributed by Mentor Embedded.
7 This file is part of the GNU Offloading and Multi Processing Library
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)
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
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.
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/>. */
31 #include "gomp-constants.h"
36 /* Return block containing [H->S), or NULL if not contained. The device lock
37 for DEV must be locked on entry, and remains locked on exit. */
40 lookup_host (struct gomp_device_descr
*dev
, void *h
, size_t s
)
42 struct splay_tree_key_s node
;
45 node
.host_start
= (uintptr_t) h
;
46 node
.host_end
= (uintptr_t) h
+ s
;
48 key
= splay_tree_lookup (&dev
->mem_map
, &node
);
53 /* Helper for lookup_dev. Iterate over splay tree. */
56 lookup_dev_1 (splay_tree_node node
, uintptr_t d
, size_t s
)
58 splay_tree_key key
= &node
->key
;
59 if (d
>= key
->tgt
->tgt_start
&& d
+ s
<= key
->tgt
->tgt_end
)
64 key
= lookup_dev_1 (node
->left
, d
, s
);
65 if (!key
&& node
->right
)
66 key
= lookup_dev_1 (node
->right
, d
, s
);
71 /* Return block containing [D->S), or NULL if not contained.
73 This iterates over the splay tree. This is not expected to be a common
76 The device lock associated with MEM_MAP must be locked on entry, and remains
80 lookup_dev (splay_tree mem_map
, void *d
, size_t s
)
82 if (!mem_map
|| !mem_map
->root
)
85 return lookup_dev_1 (mem_map
->root
, (uintptr_t) d
, s
);
89 /* OpenACC is silent on how memory exhaustion is indicated. We return
98 goacc_lazy_initialize ();
100 struct goacc_thread
*thr
= goacc_thread ();
104 if (thr
->dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
107 acc_prof_info prof_info
;
108 acc_api_info api_info
;
109 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
111 void *res
= thr
->dev
->alloc_func (thr
->dev
->target_id
, s
);
115 thr
->prof_info
= NULL
;
116 thr
->api_info
= NULL
;
130 struct goacc_thread
*thr
= goacc_thread ();
132 assert (thr
&& thr
->dev
);
134 struct gomp_device_descr
*acc_dev
= thr
->dev
;
136 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
139 acc_prof_info prof_info
;
140 acc_api_info api_info
;
141 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
143 gomp_mutex_lock (&acc_dev
->lock
);
145 /* We don't have to call lazy open here, as the ptr value must have
146 been returned by acc_malloc. It's not permitted to pass NULL in
147 (unless you got that null from acc_malloc). */
148 if ((k
= lookup_dev (&acc_dev
->mem_map
, d
, 1)))
150 void *offset
= d
- k
->tgt
->tgt_start
+ k
->tgt_offset
;
151 void *h
= k
->host_start
+ offset
;
152 size_t h_size
= k
->host_end
- k
->host_start
;
153 gomp_mutex_unlock (&acc_dev
->lock
);
154 /* PR92503 "[OpenACC] Behavior of 'acc_free' if the memory space is still
155 used in a mapping". */
156 gomp_fatal ("refusing to free device memory space at %p that is still"
157 " mapped at [%p,+%d]",
161 gomp_mutex_unlock (&acc_dev
->lock
);
163 if (!acc_dev
->free_func (acc_dev
->target_id
, d
))
164 gomp_fatal ("error in freeing device memory in %s", __FUNCTION__
);
168 thr
->prof_info
= NULL
;
169 thr
->api_info
= NULL
;
174 memcpy_tofrom_device (bool from
, void *d
, void *h
, size_t s
, int async
,
175 const char *libfnname
)
177 /* No need to call lazy open here, as the device pointer must have
178 been obtained from a routine that did that. */
179 struct goacc_thread
*thr
= goacc_thread ();
181 assert (thr
&& thr
->dev
);
183 if (thr
->dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
192 acc_prof_info prof_info
;
193 acc_api_info api_info
;
194 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
197 prof_info
.async
= async
;
198 prof_info
.async_queue
= prof_info
.async
;
201 goacc_aq aq
= get_goacc_asyncqueue (async
);
203 gomp_copy_dev2host (thr
->dev
, aq
, h
, d
, s
);
205 gomp_copy_host2dev (thr
->dev
, aq
, d
, h
, s
, /* TODO: cbuf? */ NULL
);
209 thr
->prof_info
= NULL
;
210 thr
->api_info
= NULL
;
215 acc_memcpy_to_device (void *d
, void *h
, size_t s
)
217 memcpy_tofrom_device (false, d
, h
, s
, acc_async_sync
, __FUNCTION__
);
221 acc_memcpy_to_device_async (void *d
, void *h
, size_t s
, int async
)
223 memcpy_tofrom_device (false, d
, h
, s
, async
, __FUNCTION__
);
227 acc_memcpy_from_device (void *h
, void *d
, size_t s
)
229 memcpy_tofrom_device (true, d
, h
, s
, acc_async_sync
, __FUNCTION__
);
233 acc_memcpy_from_device_async (void *h
, void *d
, size_t s
, int async
)
235 memcpy_tofrom_device (true, d
, h
, s
, async
, __FUNCTION__
);
238 /* Return the device pointer that corresponds to host data H. Or NULL
242 acc_deviceptr (void *h
)
248 goacc_lazy_initialize ();
250 struct goacc_thread
*thr
= goacc_thread ();
251 struct gomp_device_descr
*dev
= thr
->dev
;
253 if (thr
->dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
256 /* In the following, no OpenACC Profiling Interface events can possibly be
259 gomp_mutex_lock (&dev
->lock
);
261 n
= lookup_host (dev
, h
, 1);
265 gomp_mutex_unlock (&dev
->lock
);
269 offset
= h
- n
->host_start
;
271 d
= n
->tgt
->tgt_start
+ n
->tgt_offset
+ offset
;
273 gomp_mutex_unlock (&dev
->lock
);
278 /* Return the host pointer that corresponds to device data D. Or NULL
282 acc_hostptr (void *d
)
288 goacc_lazy_initialize ();
290 struct goacc_thread
*thr
= goacc_thread ();
291 struct gomp_device_descr
*acc_dev
= thr
->dev
;
293 if (thr
->dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
296 /* In the following, no OpenACC Profiling Interface events can possibly be
299 gomp_mutex_lock (&acc_dev
->lock
);
301 n
= lookup_dev (&acc_dev
->mem_map
, d
, 1);
305 gomp_mutex_unlock (&acc_dev
->lock
);
309 offset
= d
- n
->tgt
->tgt_start
+ n
->tgt_offset
;
311 h
= n
->host_start
+ offset
;
313 gomp_mutex_unlock (&acc_dev
->lock
);
318 /* Return 1 if host data [H,+S] is present on the device. */
321 acc_is_present (void *h
, size_t s
)
328 goacc_lazy_initialize ();
330 struct goacc_thread
*thr
= goacc_thread ();
331 struct gomp_device_descr
*acc_dev
= thr
->dev
;
333 if (thr
->dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
336 /* In the following, no OpenACC Profiling Interface events can possibly be
339 gomp_mutex_lock (&acc_dev
->lock
);
341 n
= lookup_host (acc_dev
, h
, s
);
343 if (n
&& ((uintptr_t)h
< n
->host_start
344 || (uintptr_t)h
+ s
> n
->host_end
345 || s
> n
->host_end
- n
->host_start
))
348 gomp_mutex_unlock (&acc_dev
->lock
);
353 /* Create a mapping for host [H,+S] -> device [D,+S] */
356 acc_map_data (void *h
, void *d
, size_t s
)
358 struct target_mem_desc
*tgt
= NULL
;
363 unsigned short kinds
= GOMP_MAP_ALLOC
;
365 goacc_lazy_initialize ();
367 struct goacc_thread
*thr
= goacc_thread ();
368 struct gomp_device_descr
*acc_dev
= thr
->dev
;
370 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
373 gomp_fatal ("cannot map data on shared-memory system");
377 struct goacc_thread
*thr
= goacc_thread ();
380 gomp_fatal ("[%p,+%d]->[%p,+%d] is a bad map",
381 (void *)h
, (int)s
, (void *)d
, (int)s
);
383 acc_prof_info prof_info
;
384 acc_api_info api_info
;
385 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
387 gomp_mutex_lock (&acc_dev
->lock
);
389 if (lookup_host (acc_dev
, h
, s
))
391 gomp_mutex_unlock (&acc_dev
->lock
);
392 gomp_fatal ("host address [%p, +%d] is already mapped", (void *)h
,
396 if (lookup_dev (&thr
->dev
->mem_map
, d
, s
))
398 gomp_mutex_unlock (&acc_dev
->lock
);
399 gomp_fatal ("device address [%p, +%d] is already mapped", (void *)d
,
403 gomp_mutex_unlock (&acc_dev
->lock
);
405 tgt
= gomp_map_vars (acc_dev
, mapnum
, &hostaddrs
, &devaddrs
, &sizes
,
406 &kinds
, true, GOMP_MAP_VARS_ENTER_DATA
);
408 splay_tree_key n
= tgt
->list
[0].key
;
409 assert (n
->refcount
== 1);
410 assert (n
->dynamic_refcount
== 0);
411 /* Special reference counting behavior. */
412 n
->refcount
= REFCOUNT_INFINITY
;
416 thr
->prof_info
= NULL
;
417 thr
->api_info
= NULL
;
423 acc_unmap_data (void *h
)
425 struct goacc_thread
*thr
= goacc_thread ();
426 struct gomp_device_descr
*acc_dev
= thr
->dev
;
428 /* No need to call lazy open, as the address must have been mapped. */
430 /* This is a no-op on shared-memory targets. */
431 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
434 acc_prof_info prof_info
;
435 acc_api_info api_info
;
436 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
440 gomp_mutex_lock (&acc_dev
->lock
);
442 splay_tree_key n
= lookup_host (acc_dev
, h
, 1);
443 struct target_mem_desc
*t
;
447 gomp_mutex_unlock (&acc_dev
->lock
);
448 gomp_fatal ("%p is not a mapped block", (void *)h
);
451 host_size
= n
->host_end
- n
->host_start
;
453 if (n
->host_start
!= (uintptr_t) h
)
455 gomp_mutex_unlock (&acc_dev
->lock
);
456 gomp_fatal ("[%p,%d] surrounds %p",
457 (void *) n
->host_start
, (int) host_size
, (void *) h
);
459 /* TODO This currently doesn't catch 'REFCOUNT_INFINITY' usage different from
460 'acc_map_data'. Maybe 'dynamic_refcount' can be used for disambiguating
461 the different 'REFCOUNT_INFINITY' cases, or simply separate
462 'REFCOUNT_INFINITY' values per different usage ('REFCOUNT_ACC_MAP_DATA'
464 else if (n
->refcount
!= REFCOUNT_INFINITY
)
466 gomp_mutex_unlock (&acc_dev
->lock
);
467 gomp_fatal ("refusing to unmap block [%p,+%d] that has not been mapped"
468 " by 'acc_map_data'",
469 (void *) h
, (int) host_size
);
474 if (t
->refcount
== 1)
476 /* This is the last reference, so pull the descriptor off the
477 chain. This prevents 'gomp_unmap_tgt' via 'gomp_remove_var' from
478 freeing the device memory. */
483 bool is_tgt_unmapped
= gomp_remove_var (acc_dev
, n
);
484 assert (is_tgt_unmapped
);
486 gomp_mutex_unlock (&acc_dev
->lock
);
490 thr
->prof_info
= NULL
;
491 thr
->api_info
= NULL
;
496 /* Enter a dynamic mapping.
498 Return the device pointer. */
501 goacc_enter_data (void *h
, size_t s
, unsigned short kind
, int async
)
507 gomp_fatal ("[%p,+%d] is a bad range", (void *)h
, (int)s
);
509 goacc_lazy_initialize ();
511 struct goacc_thread
*thr
= goacc_thread ();
512 struct gomp_device_descr
*acc_dev
= thr
->dev
;
514 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
517 acc_prof_info prof_info
;
518 acc_api_info api_info
;
519 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
522 prof_info
.async
= async
;
523 prof_info
.async_queue
= prof_info
.async
;
526 gomp_mutex_lock (&acc_dev
->lock
);
528 n
= lookup_host (acc_dev
, h
, s
);
532 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
+ h
- n
->host_start
);
534 if ((h
+ s
) > (void *)n
->host_end
)
536 gomp_mutex_unlock (&acc_dev
->lock
);
537 gomp_fatal ("[%p,+%d] not mapped", (void *)h
, (int)s
);
540 assert (n
->refcount
!= REFCOUNT_LINK
);
541 if (n
->refcount
!= REFCOUNT_INFINITY
)
543 n
->dynamic_refcount
++;
545 gomp_mutex_unlock (&acc_dev
->lock
);
549 struct target_mem_desc
*tgt
;
553 gomp_mutex_unlock (&acc_dev
->lock
);
555 goacc_aq aq
= get_goacc_asyncqueue (async
);
557 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, &hostaddrs
, NULL
, &s
,
558 &kind
, true, GOMP_MAP_VARS_ENTER_DATA
);
560 n
= tgt
->list
[0].key
;
561 assert (n
->refcount
== 1);
562 assert (n
->dynamic_refcount
== 0);
563 n
->dynamic_refcount
++;
570 thr
->prof_info
= NULL
;
571 thr
->api_info
= NULL
;
578 acc_create (void *h
, size_t s
)
580 return goacc_enter_data (h
, s
, GOMP_MAP_ALLOC
, acc_async_sync
);
584 acc_create_async (void *h
, size_t s
, int async
)
586 goacc_enter_data (h
, s
, GOMP_MAP_ALLOC
, async
);
589 /* acc_present_or_create used to be what acc_create is now. */
590 /* acc_pcreate is acc_present_or_create by a different name. */
591 #ifdef HAVE_ATTRIBUTE_ALIAS
592 strong_alias (acc_create
, acc_present_or_create
)
593 strong_alias (acc_create
, acc_pcreate
)
596 acc_present_or_create (void *h
, size_t s
)
598 return acc_create (h
, s
);
602 acc_pcreate (void *h
, size_t s
)
604 return acc_create (h
, s
);
609 acc_copyin (void *h
, size_t s
)
611 return goacc_enter_data (h
, s
, GOMP_MAP_TO
, acc_async_sync
);
615 acc_copyin_async (void *h
, size_t s
, int async
)
617 goacc_enter_data (h
, s
, GOMP_MAP_TO
, async
);
620 /* acc_present_or_copyin used to be what acc_copyin is now. */
621 /* acc_pcopyin is acc_present_or_copyin by a different name. */
622 #ifdef HAVE_ATTRIBUTE_ALIAS
623 strong_alias (acc_copyin
, acc_present_or_copyin
)
624 strong_alias (acc_copyin
, acc_pcopyin
)
627 acc_present_or_copyin (void *h
, size_t s
)
629 return acc_copyin (h
, s
);
633 acc_pcopyin (void *h
, size_t s
)
635 return acc_copyin (h
, s
);
640 /* Exit a dynamic mapping. */
643 goacc_exit_data (void *h
, size_t s
, unsigned short kind
, int async
)
645 /* No need to call lazy open, as the data must already have been
650 struct goacc_thread
*thr
= goacc_thread ();
651 struct gomp_device_descr
*acc_dev
= thr
->dev
;
653 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
656 acc_prof_info prof_info
;
657 acc_api_info api_info
;
658 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
661 prof_info
.async
= async
;
662 prof_info
.async_queue
= prof_info
.async
;
665 gomp_mutex_lock (&acc_dev
->lock
);
667 splay_tree_key n
= lookup_host (acc_dev
, h
, s
);
669 /* PR92726, RP92970, PR92984: no-op. */
672 if ((uintptr_t) h
< n
->host_start
|| (uintptr_t) h
+ s
> n
->host_end
)
674 size_t host_size
= n
->host_end
- n
->host_start
;
675 gomp_mutex_unlock (&acc_dev
->lock
);
676 gomp_fatal ("[%p,+%d] outside mapped block [%p,+%d]",
677 (void *) h
, (int) s
, (void *) n
->host_start
, (int) host_size
);
680 assert (n
->refcount
!= REFCOUNT_LINK
);
681 if (n
->refcount
!= REFCOUNT_INFINITY
682 && n
->refcount
< n
->dynamic_refcount
)
684 gomp_mutex_unlock (&acc_dev
->lock
);
685 gomp_fatal ("Dynamic reference counting assert fail\n");
688 bool finalize
= (kind
== GOMP_MAP_DELETE
689 || kind
== GOMP_MAP_FORCE_FROM
);
692 if (n
->refcount
!= REFCOUNT_INFINITY
)
693 n
->refcount
-= n
->dynamic_refcount
;
694 n
->dynamic_refcount
= 0;
696 else if (n
->dynamic_refcount
)
698 if (n
->refcount
!= REFCOUNT_INFINITY
)
700 n
->dynamic_refcount
--;
703 if (n
->refcount
== 0)
705 goacc_aq aq
= get_goacc_asyncqueue (async
);
707 bool copyout
= (kind
== GOMP_MAP_FROM
708 || kind
== GOMP_MAP_FORCE_FROM
);
711 void *d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
712 + (uintptr_t) h
- n
->host_start
);
713 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
717 /* TODO We can't do the 'is_tgt_unmapped' checking -- see the
718 'gomp_unref_tgt' comment in
719 <http://mid.mail-archive.com/878snl36eu.fsf@euler.schwinge.homeip.net>;
721 gomp_remove_var_async (acc_dev
, n
, aq
);
724 bool is_tgt_unmapped
= gomp_remove_var (acc_dev
, n
);
725 assert (is_tgt_unmapped
);
730 gomp_mutex_unlock (&acc_dev
->lock
);
734 thr
->prof_info
= NULL
;
735 thr
->api_info
= NULL
;
740 acc_delete (void *h
, size_t s
)
742 goacc_exit_data (h
, s
, GOMP_MAP_RELEASE
, acc_async_sync
);
746 acc_delete_async (void *h
, size_t s
, int async
)
748 goacc_exit_data (h
, s
, GOMP_MAP_RELEASE
, async
);
752 acc_delete_finalize (void *h
, size_t s
)
754 goacc_exit_data (h
, s
, GOMP_MAP_DELETE
, acc_async_sync
);
758 acc_delete_finalize_async (void *h
, size_t s
, int async
)
760 goacc_exit_data (h
, s
, GOMP_MAP_DELETE
, async
);
764 acc_copyout (void *h
, size_t s
)
766 goacc_exit_data (h
, s
, GOMP_MAP_FROM
, acc_async_sync
);
770 acc_copyout_async (void *h
, size_t s
, int async
)
772 goacc_exit_data (h
, s
, GOMP_MAP_FROM
, async
);
776 acc_copyout_finalize (void *h
, size_t s
)
778 goacc_exit_data (h
, s
, GOMP_MAP_FORCE_FROM
, acc_async_sync
);
782 acc_copyout_finalize_async (void *h
, size_t s
, int async
)
784 goacc_exit_data (h
, s
, GOMP_MAP_FORCE_FROM
, async
);
788 update_dev_host (int is_dev
, void *h
, size_t s
, int async
)
793 goacc_lazy_initialize ();
795 struct goacc_thread
*thr
= goacc_thread ();
796 struct gomp_device_descr
*acc_dev
= thr
->dev
;
798 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
801 /* Fortran optional arguments that are non-present result in a
802 NULL host address here. This can safely be ignored as it is
803 not possible to 'update' a non-present optional argument. */
807 acc_prof_info prof_info
;
808 acc_api_info api_info
;
809 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
812 prof_info
.async
= async
;
813 prof_info
.async_queue
= prof_info
.async
;
816 gomp_mutex_lock (&acc_dev
->lock
);
818 n
= lookup_host (acc_dev
, h
, s
);
822 gomp_mutex_unlock (&acc_dev
->lock
);
823 gomp_fatal ("[%p,%d] is not mapped", h
, (int)s
);
826 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
827 + (uintptr_t) h
- n
->host_start
);
829 goacc_aq aq
= get_goacc_asyncqueue (async
);
832 gomp_copy_host2dev (acc_dev
, aq
, d
, h
, s
, /* TODO: cbuf? */ NULL
);
834 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
836 gomp_mutex_unlock (&acc_dev
->lock
);
840 thr
->prof_info
= NULL
;
841 thr
->api_info
= NULL
;
846 acc_update_device (void *h
, size_t s
)
848 update_dev_host (1, h
, s
, acc_async_sync
);
852 acc_update_device_async (void *h
, size_t s
, int async
)
854 update_dev_host (1, h
, s
, async
);
858 acc_update_self (void *h
, size_t s
)
860 update_dev_host (0, h
, s
, acc_async_sync
);
864 acc_update_self_async (void *h
, size_t s
, int async
)
866 update_dev_host (0, h
, s
, async
);
870 /* OpenACC 'enter data', 'exit data': 'GOACC_enter_exit_data' and its helper
873 /* Special handling for 'GOMP_MAP_POINTER', 'GOMP_MAP_TO_PSET'.
875 Only the first mapping is considered in reference counting; the following
876 ones implicitly follow suit. Similarly, 'copyout' is done only for the
880 goacc_insert_pointer (size_t mapnum
, void **hostaddrs
, size_t *sizes
,
881 void *kinds
, int async
)
883 struct target_mem_desc
*tgt
;
884 struct goacc_thread
*thr
= goacc_thread ();
885 struct gomp_device_descr
*acc_dev
= thr
->dev
;
887 if (*hostaddrs
== NULL
)
890 if (acc_is_present (*hostaddrs
, *sizes
))
893 gomp_mutex_lock (&acc_dev
->lock
);
894 n
= lookup_host (acc_dev
, *hostaddrs
, *sizes
);
895 assert (n
->refcount
!= REFCOUNT_INFINITY
896 && n
->refcount
!= REFCOUNT_LINK
);
897 gomp_mutex_unlock (&acc_dev
->lock
);
900 for (size_t i
= 0; i
< tgt
->list_count
; i
++)
901 if (tgt
->list
[i
].key
== n
)
903 for (size_t j
= 0; j
< mapnum
; j
++)
904 if (i
+ j
< tgt
->list_count
&& tgt
->list
[i
+ j
].key
)
906 tgt
->list
[i
+ j
].key
->refcount
++;
907 tgt
->list
[i
+ j
].key
->dynamic_refcount
++;
911 /* Should not reach here. */
912 gomp_fatal ("Dynamic refcount incrementing failed for pointer/pset");
915 gomp_debug (0, " %s: prepare mappings\n", __FUNCTION__
);
916 goacc_aq aq
= get_goacc_asyncqueue (async
);
917 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, hostaddrs
,
918 NULL
, sizes
, kinds
, true, GOMP_MAP_VARS_ENTER_DATA
);
920 splay_tree_key n
= tgt
->list
[0].key
;
921 assert (n
->refcount
== 1);
922 assert (n
->dynamic_refcount
== 0);
923 n
->dynamic_refcount
++;
924 gomp_debug (0, " %s: mappings prepared\n", __FUNCTION__
);
928 goacc_remove_pointer (void *h
, size_t s
, unsigned short kind
, int async
)
932 struct goacc_thread
*thr
= goacc_thread ();
933 struct gomp_device_descr
*acc_dev
= thr
->dev
;
935 struct target_mem_desc
*t
;
937 if (!acc_is_present (h
, s
))
940 gomp_mutex_lock (&acc_dev
->lock
);
942 n
= lookup_host (acc_dev
, h
, 1);
946 gomp_mutex_unlock (&acc_dev
->lock
);
947 gomp_fatal ("%p is not a mapped block", (void *)h
);
950 gomp_debug (0, " %s: restore mappings\n", __FUNCTION__
);
954 assert (n
->refcount
!= REFCOUNT_INFINITY
955 && n
->refcount
!= REFCOUNT_LINK
);
956 if (n
->refcount
< n
->dynamic_refcount
)
958 gomp_mutex_unlock (&acc_dev
->lock
);
959 gomp_fatal ("Dynamic reference counting assert fail\n");
962 bool finalize
= (kind
== GOMP_MAP_DELETE
963 || kind
== GOMP_MAP_FORCE_FROM
);
966 n
->refcount
-= n
->dynamic_refcount
;
967 n
->dynamic_refcount
= 0;
969 else if (n
->dynamic_refcount
)
972 n
->dynamic_refcount
--;
975 if (n
->refcount
== 0)
977 goacc_aq aq
= get_goacc_asyncqueue (async
);
979 bool copyout
= (kind
== GOMP_MAP_FROM
980 || kind
== GOMP_MAP_FORCE_FROM
);
983 void *d
= (void *) (t
->tgt_start
+ n
->tgt_offset
984 + (uintptr_t) h
- n
->host_start
);
985 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
990 /* TODO The way the following code is currently implemented, we need
991 the 'is_tgt_unmapped' return value from 'gomp_remove_var', so
992 can't use 'gomp_remove_var_async' here -- see the 'gomp_unref_tgt'
994 <http://mid.mail-archive.com/878snl36eu.fsf@euler.schwinge.homeip.net>;
995 PR92881 -- so have to synchronize here. */
996 if (!acc_dev
->openacc
.async
.synchronize_func (aq
))
998 gomp_mutex_unlock (&acc_dev
->lock
);
999 gomp_fatal ("synchronize failed");
1002 bool is_tgt_unmapped
= false;
1003 for (size_t i
= 0; i
< t
->list_count
; i
++)
1005 is_tgt_unmapped
= gomp_remove_var (acc_dev
, t
->list
[i
].key
);
1006 if (is_tgt_unmapped
)
1009 assert (is_tgt_unmapped
);
1012 gomp_mutex_unlock (&acc_dev
->lock
);
1014 gomp_debug (0, " %s: mappings restored\n", __FUNCTION__
);
1017 /* Return the number of mappings associated with 'GOMP_MAP_TO_PSET' or
1018 'GOMP_MAP_POINTER'. */
1021 find_pointer (int pos
, size_t mapnum
, unsigned short *kinds
)
1023 if (pos
+ 1 >= mapnum
)
1026 unsigned char kind
= kinds
[pos
+1] & 0xff;
1028 if (kind
== GOMP_MAP_TO_PSET
)
1030 else if (kind
== GOMP_MAP_POINTER
)
1037 GOACC_enter_exit_data (int flags_m
, size_t mapnum
, void **hostaddrs
,
1038 size_t *sizes
, unsigned short *kinds
, int async
,
1041 int flags
= GOACC_FLAGS_UNMARSHAL (flags_m
);
1043 struct goacc_thread
*thr
;
1044 struct gomp_device_descr
*acc_dev
;
1045 bool data_enter
= false;
1048 goacc_lazy_initialize ();
1050 thr
= goacc_thread ();
1053 /* Determine if this is an "acc enter data". */
1054 for (i
= 0; i
< mapnum
; ++i
)
1056 unsigned char kind
= kinds
[i
] & 0xff;
1058 if (kind
== GOMP_MAP_POINTER
|| kind
== GOMP_MAP_TO_PSET
)
1061 if (kind
== GOMP_MAP_FORCE_ALLOC
1062 || kind
== GOMP_MAP_FORCE_PRESENT
1063 || kind
== GOMP_MAP_FORCE_TO
1064 || kind
== GOMP_MAP_TO
1065 || kind
== GOMP_MAP_ALLOC
)
1071 if (kind
== GOMP_MAP_RELEASE
1072 || kind
== GOMP_MAP_DELETE
1073 || kind
== GOMP_MAP_FROM
1074 || kind
== GOMP_MAP_FORCE_FROM
)
1077 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1081 bool profiling_p
= GOACC_PROFILING_DISPATCH_P (true);
1083 acc_prof_info prof_info
;
1086 thr
->prof_info
= &prof_info
;
1088 prof_info
.event_type
1089 = data_enter
? acc_ev_enter_data_start
: acc_ev_exit_data_start
;
1090 prof_info
.valid_bytes
= _ACC_PROF_INFO_VALID_BYTES
;
1091 prof_info
.version
= _ACC_PROF_INFO_VERSION
;
1092 prof_info
.device_type
= acc_device_type (acc_dev
->type
);
1093 prof_info
.device_number
= acc_dev
->target_id
;
1094 prof_info
.thread_id
= -1;
1095 prof_info
.async
= async
;
1096 prof_info
.async_queue
= prof_info
.async
;
1097 prof_info
.src_file
= NULL
;
1098 prof_info
.func_name
= NULL
;
1099 prof_info
.line_no
= -1;
1100 prof_info
.end_line_no
= -1;
1101 prof_info
.func_line_no
= -1;
1102 prof_info
.func_end_line_no
= -1;
1104 acc_event_info enter_exit_data_event_info
;
1107 enter_exit_data_event_info
.other_event
.event_type
1108 = prof_info
.event_type
;
1109 enter_exit_data_event_info
.other_event
.valid_bytes
1110 = _ACC_OTHER_EVENT_INFO_VALID_BYTES
;
1111 enter_exit_data_event_info
.other_event
.parent_construct
1112 = data_enter
? acc_construct_enter_data
: acc_construct_exit_data
;
1113 enter_exit_data_event_info
.other_event
.implicit
= 0;
1114 enter_exit_data_event_info
.other_event
.tool_info
= NULL
;
1116 acc_api_info api_info
;
1119 thr
->api_info
= &api_info
;
1121 api_info
.device_api
= acc_device_api_none
;
1122 api_info
.valid_bytes
= _ACC_API_INFO_VALID_BYTES
;
1123 api_info
.device_type
= prof_info
.device_type
;
1124 api_info
.vendor
= -1;
1125 api_info
.device_handle
= NULL
;
1126 api_info
.context_handle
= NULL
;
1127 api_info
.async_handle
= NULL
;
1131 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1134 if ((acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
1135 || (flags
& GOACC_FLAG_HOST_FALLBACK
))
1137 prof_info
.device_type
= acc_device_host
;
1138 api_info
.device_type
= prof_info
.device_type
;
1147 va_start (ap
, num_waits
);
1148 goacc_wait (async
, num_waits
, &ap
);
1152 /* In c, non-pointers and arrays are represented by a single data clause.
1153 Dynamically allocated arrays and subarrays are represented by a data
1154 clause followed by an internal GOMP_MAP_POINTER.
1156 In fortran, scalars and not allocated arrays are represented by a
1157 single data clause. Allocated arrays and subarrays have three mappings:
1158 1) the original data clause, 2) a PSET 3) a pointer to the array data.
1163 for (i
= 0; i
< mapnum
; i
++)
1165 /* Scan for pointers and PSETs. */
1166 int pointer
= find_pointer (i
, mapnum
, kinds
);
1170 unsigned char kind
= kinds
[i
] & 0xff;
1173 case GOMP_MAP_ALLOC
:
1174 case GOMP_MAP_FORCE_ALLOC
:
1176 case GOMP_MAP_FORCE_TO
:
1179 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1184 goacc_enter_data (hostaddrs
[i
], sizes
[i
], kinds
[i
], async
);
1188 goacc_insert_pointer (pointer
, &hostaddrs
[i
], &sizes
[i
], &kinds
[i
],
1190 /* Increment 'i' by two because OpenACC requires fortran
1191 arrays to be contiguous, so each PSET is associated with
1192 one of MAP_FORCE_ALLOC/MAP_FORCE_PRESET/MAP_FORCE_TO, and
1199 for (i
= 0; i
< mapnum
; ++i
)
1201 int pointer
= find_pointer (i
, mapnum
, kinds
);
1205 unsigned char kind
= kinds
[i
] & 0xff;
1208 case GOMP_MAP_RELEASE
:
1209 case GOMP_MAP_DELETE
:
1211 case GOMP_MAP_FORCE_FROM
:
1214 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1219 goacc_exit_data (hostaddrs
[i
], sizes
[i
], kinds
[i
], async
);
1223 goacc_remove_pointer (hostaddrs
[i
], sizes
[i
], kinds
[i
], async
);
1224 /* See the above comment. */
1232 prof_info
.event_type
1233 = data_enter
? acc_ev_enter_data_end
: acc_ev_exit_data_end
;
1234 enter_exit_data_event_info
.other_event
.event_type
= prof_info
.event_type
;
1235 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1238 thr
->prof_info
= NULL
;
1239 thr
->api_info
= NULL
;