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_OPENACC
);
407 splay_tree_key n
= tgt
->list
[0].key
;
408 assert (n
->refcount
== 1);
409 assert (n
->dynamic_refcount
== 0);
410 /* Special reference counting behavior. */
411 n
->refcount
= REFCOUNT_INFINITY
;
415 thr
->prof_info
= NULL
;
416 thr
->api_info
= NULL
;
422 acc_unmap_data (void *h
)
424 struct goacc_thread
*thr
= goacc_thread ();
425 struct gomp_device_descr
*acc_dev
= thr
->dev
;
427 /* No need to call lazy open, as the address must have been mapped. */
429 /* This is a no-op on shared-memory targets. */
430 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
433 acc_prof_info prof_info
;
434 acc_api_info api_info
;
435 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
439 gomp_mutex_lock (&acc_dev
->lock
);
441 splay_tree_key n
= lookup_host (acc_dev
, h
, 1);
442 struct target_mem_desc
*t
;
446 gomp_mutex_unlock (&acc_dev
->lock
);
447 gomp_fatal ("%p is not a mapped block", (void *)h
);
450 host_size
= n
->host_end
- n
->host_start
;
452 if (n
->host_start
!= (uintptr_t) h
)
454 gomp_mutex_unlock (&acc_dev
->lock
);
455 gomp_fatal ("[%p,%d] surrounds %p",
456 (void *) n
->host_start
, (int) host_size
, (void *) h
);
458 /* TODO This currently doesn't catch 'REFCOUNT_INFINITY' usage different from
459 'acc_map_data'. Maybe 'dynamic_refcount' can be used for disambiguating
460 the different 'REFCOUNT_INFINITY' cases, or simply separate
461 'REFCOUNT_INFINITY' values per different usage ('REFCOUNT_ACC_MAP_DATA'
463 else if (n
->refcount
!= REFCOUNT_INFINITY
)
465 gomp_mutex_unlock (&acc_dev
->lock
);
466 gomp_fatal ("refusing to unmap block [%p,+%d] that has not been mapped"
467 " by 'acc_map_data'",
468 (void *) h
, (int) host_size
);
471 /* Mark for removal. */
476 if (t
->refcount
== 2)
478 /* This is the last reference, so pull the descriptor off the
479 chain. This avoids gomp_unmap_vars via gomp_unmap_tgt from
480 freeing the device memory. */
485 gomp_mutex_unlock (&acc_dev
->lock
);
487 gomp_unmap_vars (t
, true);
491 thr
->prof_info
= NULL
;
492 thr
->api_info
= NULL
;
496 #define FLAG_PRESENT (1 << 0)
497 #define FLAG_CREATE (1 << 1)
498 #define FLAG_COPY (1 << 2)
501 present_create_copy (unsigned f
, void *h
, size_t s
, 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 (!(f
& FLAG_PRESENT
))
536 gomp_mutex_unlock (&acc_dev
->lock
);
537 gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
538 (void *)h
, (int)s
, (void *)d
, (int)s
);
540 if ((h
+ s
) > (void *)n
->host_end
)
542 gomp_mutex_unlock (&acc_dev
->lock
);
543 gomp_fatal ("[%p,+%d] not mapped", (void *)h
, (int)s
);
546 assert (n
->refcount
!= REFCOUNT_LINK
);
547 if (n
->refcount
!= REFCOUNT_INFINITY
)
549 n
->dynamic_refcount
++;
551 gomp_mutex_unlock (&acc_dev
->lock
);
553 else if (!(f
& FLAG_CREATE
))
555 gomp_mutex_unlock (&acc_dev
->lock
);
556 gomp_fatal ("[%p,+%d] not mapped", (void *)h
, (int)s
);
560 struct target_mem_desc
*tgt
;
562 unsigned short kinds
;
568 kinds
= GOMP_MAP_ALLOC
;
570 gomp_mutex_unlock (&acc_dev
->lock
);
572 goacc_aq aq
= get_goacc_asyncqueue (async
);
574 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, &hostaddrs
, NULL
, &s
,
575 &kinds
, true, GOMP_MAP_VARS_OPENACC
);
576 n
= tgt
->list
[0].key
;
577 assert (n
->refcount
== 1);
578 assert (n
->dynamic_refcount
== 0);
579 n
->dynamic_refcount
++;
586 thr
->prof_info
= NULL
;
587 thr
->api_info
= NULL
;
594 acc_create (void *h
, size_t s
)
596 return present_create_copy (FLAG_PRESENT
| FLAG_CREATE
, h
, s
, acc_async_sync
);
600 acc_create_async (void *h
, size_t s
, int async
)
602 present_create_copy (FLAG_PRESENT
| FLAG_CREATE
, h
, s
, async
);
605 /* acc_present_or_create used to be what acc_create is now. */
606 /* acc_pcreate is acc_present_or_create by a different name. */
607 #ifdef HAVE_ATTRIBUTE_ALIAS
608 strong_alias (acc_create
, acc_present_or_create
)
609 strong_alias (acc_create
, acc_pcreate
)
612 acc_present_or_create (void *h
, size_t s
)
614 return acc_create (h
, s
);
618 acc_pcreate (void *h
, size_t s
)
620 return acc_create (h
, s
);
625 acc_copyin (void *h
, size_t s
)
627 return present_create_copy (FLAG_PRESENT
| FLAG_CREATE
| FLAG_COPY
, h
, s
,
632 acc_copyin_async (void *h
, size_t s
, int async
)
634 present_create_copy (FLAG_PRESENT
| FLAG_CREATE
| FLAG_COPY
, h
, s
, async
);
637 /* acc_present_or_copyin used to be what acc_copyin is now. */
638 /* acc_pcopyin is acc_present_or_copyin by a different name. */
639 #ifdef HAVE_ATTRIBUTE_ALIAS
640 strong_alias (acc_copyin
, acc_present_or_copyin
)
641 strong_alias (acc_copyin
, acc_pcopyin
)
644 acc_present_or_copyin (void *h
, size_t s
)
646 return acc_copyin (h
, s
);
650 acc_pcopyin (void *h
, size_t s
)
652 return acc_copyin (h
, s
);
656 #define FLAG_COPYOUT (1 << 0)
657 #define FLAG_FINALIZE (1 << 1)
660 delete_copyout (unsigned f
, void *h
, size_t s
, int async
, const char *libfnname
)
664 struct goacc_thread
*thr
= goacc_thread ();
665 struct gomp_device_descr
*acc_dev
= thr
->dev
;
667 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
670 acc_prof_info prof_info
;
671 acc_api_info api_info
;
672 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
675 prof_info
.async
= async
;
676 prof_info
.async_queue
= prof_info
.async
;
679 gomp_mutex_lock (&acc_dev
->lock
);
681 n
= lookup_host (acc_dev
, h
, s
);
683 /* No need to call lazy open, as the data must already have been
688 gomp_mutex_unlock (&acc_dev
->lock
);
689 gomp_fatal ("[%p,%d] is not mapped", (void *)h
, (int)s
);
692 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
693 + (uintptr_t) h
- n
->host_start
);
695 if ((uintptr_t) h
< n
->host_start
|| (uintptr_t) h
+ s
> n
->host_end
)
697 size_t host_size
= n
->host_end
- n
->host_start
;
698 gomp_mutex_unlock (&acc_dev
->lock
);
699 gomp_fatal ("[%p,+%d] outside mapped block [%p,+%d]",
700 (void *) h
, (int) s
, (void *) n
->host_start
, (int) host_size
);
703 assert (n
->refcount
!= REFCOUNT_LINK
);
704 if (n
->refcount
!= REFCOUNT_INFINITY
705 && n
->refcount
< n
->dynamic_refcount
)
707 gomp_mutex_unlock (&acc_dev
->lock
);
708 gomp_fatal ("Dynamic reference counting assert fail\n");
711 if (f
& FLAG_FINALIZE
)
713 if (n
->refcount
!= REFCOUNT_INFINITY
)
714 n
->refcount
-= n
->dynamic_refcount
;
715 n
->dynamic_refcount
= 0;
717 else if (n
->dynamic_refcount
)
719 if (n
->refcount
!= REFCOUNT_INFINITY
)
721 n
->dynamic_refcount
--;
724 if (n
->refcount
== 0)
726 if (f
& FLAG_COPYOUT
)
728 goacc_aq aq
= get_goacc_asyncqueue (async
);
729 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
731 gomp_remove_var (acc_dev
, n
);
734 gomp_mutex_unlock (&acc_dev
->lock
);
738 thr
->prof_info
= NULL
;
739 thr
->api_info
= NULL
;
744 acc_delete (void *h
, size_t s
)
746 delete_copyout (0, h
, s
, acc_async_sync
, __FUNCTION__
);
750 acc_delete_async (void *h
, size_t s
, int async
)
752 delete_copyout (0, h
, s
, async
, __FUNCTION__
);
756 acc_delete_finalize (void *h
, size_t s
)
758 delete_copyout (FLAG_FINALIZE
, h
, s
, acc_async_sync
, __FUNCTION__
);
762 acc_delete_finalize_async (void *h
, size_t s
, int async
)
764 delete_copyout (FLAG_FINALIZE
, h
, s
, async
, __FUNCTION__
);
768 acc_copyout (void *h
, size_t s
)
770 delete_copyout (FLAG_COPYOUT
, h
, s
, acc_async_sync
, __FUNCTION__
);
774 acc_copyout_async (void *h
, size_t s
, int async
)
776 delete_copyout (FLAG_COPYOUT
, h
, s
, async
, __FUNCTION__
);
780 acc_copyout_finalize (void *h
, size_t s
)
782 delete_copyout (FLAG_COPYOUT
| FLAG_FINALIZE
, h
, s
, acc_async_sync
,
787 acc_copyout_finalize_async (void *h
, size_t s
, int async
)
789 delete_copyout (FLAG_COPYOUT
| FLAG_FINALIZE
, h
, s
, async
, __FUNCTION__
);
793 update_dev_host (int is_dev
, void *h
, size_t s
, int async
)
798 goacc_lazy_initialize ();
800 struct goacc_thread
*thr
= goacc_thread ();
801 struct gomp_device_descr
*acc_dev
= thr
->dev
;
803 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
806 /* Fortran optional arguments that are non-present result in a
807 NULL host address here. This can safely be ignored as it is
808 not possible to 'update' a non-present optional argument. */
812 acc_prof_info prof_info
;
813 acc_api_info api_info
;
814 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
817 prof_info
.async
= async
;
818 prof_info
.async_queue
= prof_info
.async
;
821 gomp_mutex_lock (&acc_dev
->lock
);
823 n
= lookup_host (acc_dev
, h
, s
);
827 gomp_mutex_unlock (&acc_dev
->lock
);
828 gomp_fatal ("[%p,%d] is not mapped", h
, (int)s
);
831 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
832 + (uintptr_t) h
- n
->host_start
);
834 goacc_aq aq
= get_goacc_asyncqueue (async
);
837 gomp_copy_host2dev (acc_dev
, aq
, d
, h
, s
, /* TODO: cbuf? */ NULL
);
839 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
841 gomp_mutex_unlock (&acc_dev
->lock
);
845 thr
->prof_info
= NULL
;
846 thr
->api_info
= NULL
;
851 acc_update_device (void *h
, size_t s
)
853 update_dev_host (1, h
, s
, acc_async_sync
);
857 acc_update_device_async (void *h
, size_t s
, int async
)
859 update_dev_host (1, h
, s
, async
);
863 acc_update_self (void *h
, size_t s
)
865 update_dev_host (0, h
, s
, acc_async_sync
);
869 acc_update_self_async (void *h
, size_t s
, int async
)
871 update_dev_host (0, h
, s
, async
);
875 /* OpenACC 'enter data', 'exit data': 'GOACC_enter_exit_data' and its helper
878 /* Special handling for 'GOMP_MAP_POINTER', 'GOMP_MAP_TO_PSET'.
880 Only the first mapping is considered in reference counting; the following
881 ones implicitly follow suit. */
884 goacc_insert_pointer (size_t mapnum
, void **hostaddrs
, size_t *sizes
,
885 void *kinds
, int async
)
887 struct target_mem_desc
*tgt
;
888 struct goacc_thread
*thr
= goacc_thread ();
889 struct gomp_device_descr
*acc_dev
= thr
->dev
;
891 if (*hostaddrs
== NULL
)
894 if (acc_is_present (*hostaddrs
, *sizes
))
897 gomp_mutex_lock (&acc_dev
->lock
);
898 n
= lookup_host (acc_dev
, *hostaddrs
, *sizes
);
899 assert (n
->refcount
!= REFCOUNT_INFINITY
900 && n
->refcount
!= REFCOUNT_LINK
);
901 gomp_mutex_unlock (&acc_dev
->lock
);
904 for (size_t i
= 0; i
< tgt
->list_count
; i
++)
905 if (tgt
->list
[i
].key
== n
)
907 for (size_t j
= 0; j
< mapnum
; j
++)
908 if (i
+ j
< tgt
->list_count
&& tgt
->list
[i
+ j
].key
)
910 tgt
->list
[i
+ j
].key
->refcount
++;
911 tgt
->list
[i
+ j
].key
->dynamic_refcount
++;
915 /* Should not reach here. */
916 gomp_fatal ("Dynamic refcount incrementing failed for pointer/pset");
919 gomp_debug (0, " %s: prepare mappings\n", __FUNCTION__
);
920 goacc_aq aq
= get_goacc_asyncqueue (async
);
921 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, hostaddrs
,
922 NULL
, sizes
, kinds
, true, GOMP_MAP_VARS_OPENACC
);
923 splay_tree_key n
= tgt
->list
[0].key
;
924 assert (n
->refcount
== 1);
925 assert (n
->dynamic_refcount
== 0);
926 n
->dynamic_refcount
++;
927 gomp_debug (0, " %s: mappings prepared\n", __FUNCTION__
);
931 goacc_remove_pointer (void *h
, size_t s
, bool force_copyfrom
, int async
,
932 int finalize
, int mapnum
)
934 struct goacc_thread
*thr
= goacc_thread ();
935 struct gomp_device_descr
*acc_dev
= thr
->dev
;
937 struct target_mem_desc
*t
;
938 int minrefs
= (mapnum
== 1) ? 2 : 3;
940 if (!acc_is_present (h
, s
))
943 gomp_mutex_lock (&acc_dev
->lock
);
945 n
= lookup_host (acc_dev
, h
, 1);
949 gomp_mutex_unlock (&acc_dev
->lock
);
950 gomp_fatal ("%p is not a mapped block", (void *)h
);
953 gomp_debug (0, " %s: restore mappings\n", __FUNCTION__
);
957 assert (n
->refcount
!= REFCOUNT_INFINITY
958 && n
->refcount
!= REFCOUNT_LINK
);
959 if (n
->refcount
< n
->dynamic_refcount
)
961 gomp_mutex_unlock (&acc_dev
->lock
);
962 gomp_fatal ("Dynamic reference counting assert fail\n");
967 n
->refcount
-= n
->dynamic_refcount
;
968 n
->dynamic_refcount
= 0;
970 else if (n
->dynamic_refcount
)
973 n
->dynamic_refcount
--;
976 gomp_mutex_unlock (&acc_dev
->lock
);
978 if (n
->refcount
== 0)
980 /* Set refcount to 1 to allow gomp_unmap_vars to unmap it. */
982 t
->refcount
= minrefs
;
983 for (size_t i
= 0; i
< t
->list_count
; i
++)
984 if (t
->list
[i
].key
== n
)
986 t
->list
[i
].copy_from
= force_copyfrom
? 1 : 0;
990 /* If running synchronously, unmap immediately. */
991 if (async
< acc_async_noval
)
992 gomp_unmap_vars (t
, true);
995 goacc_aq aq
= get_goacc_asyncqueue (async
);
996 gomp_unmap_vars_async (t
, true, aq
);
1000 gomp_mutex_unlock (&acc_dev
->lock
);
1002 gomp_debug (0, " %s: mappings restored\n", __FUNCTION__
);
1005 /* Return the number of mappings associated with 'GOMP_MAP_TO_PSET' or
1006 'GOMP_MAP_POINTER'. */
1009 find_pointer (int pos
, size_t mapnum
, unsigned short *kinds
)
1011 if (pos
+ 1 >= mapnum
)
1014 unsigned char kind
= kinds
[pos
+1] & 0xff;
1016 if (kind
== GOMP_MAP_TO_PSET
)
1018 else if (kind
== GOMP_MAP_POINTER
)
1025 GOACC_enter_exit_data (int flags_m
, size_t mapnum
, void **hostaddrs
,
1026 size_t *sizes
, unsigned short *kinds
, int async
,
1029 int flags
= GOACC_FLAGS_UNMARSHAL (flags_m
);
1031 struct goacc_thread
*thr
;
1032 struct gomp_device_descr
*acc_dev
;
1033 bool data_enter
= false;
1036 goacc_lazy_initialize ();
1038 thr
= goacc_thread ();
1041 /* Determine whether "finalize" semantics apply to all mappings of this
1042 OpenACC directive. */
1043 bool finalize
= false;
1046 unsigned char kind
= kinds
[0] & 0xff;
1047 if (kind
== GOMP_MAP_DELETE
1048 || kind
== GOMP_MAP_FORCE_FROM
)
1052 /* Determine if this is an "acc enter data". */
1053 for (i
= 0; i
< mapnum
; ++i
)
1055 unsigned char kind
= kinds
[i
] & 0xff;
1057 if (kind
== GOMP_MAP_POINTER
|| kind
== GOMP_MAP_TO_PSET
)
1060 if (kind
== GOMP_MAP_FORCE_ALLOC
1061 || kind
== GOMP_MAP_FORCE_PRESENT
1062 || kind
== GOMP_MAP_FORCE_TO
1063 || kind
== GOMP_MAP_TO
1064 || kind
== GOMP_MAP_ALLOC
)
1070 if (kind
== GOMP_MAP_RELEASE
1071 || kind
== GOMP_MAP_DELETE
1072 || kind
== GOMP_MAP_FROM
1073 || kind
== GOMP_MAP_FORCE_FROM
)
1076 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1080 bool profiling_p
= GOACC_PROFILING_DISPATCH_P (true);
1082 acc_prof_info prof_info
;
1085 thr
->prof_info
= &prof_info
;
1087 prof_info
.event_type
1088 = data_enter
? acc_ev_enter_data_start
: acc_ev_exit_data_start
;
1089 prof_info
.valid_bytes
= _ACC_PROF_INFO_VALID_BYTES
;
1090 prof_info
.version
= _ACC_PROF_INFO_VERSION
;
1091 prof_info
.device_type
= acc_device_type (acc_dev
->type
);
1092 prof_info
.device_number
= acc_dev
->target_id
;
1093 prof_info
.thread_id
= -1;
1094 prof_info
.async
= async
;
1095 prof_info
.async_queue
= prof_info
.async
;
1096 prof_info
.src_file
= NULL
;
1097 prof_info
.func_name
= NULL
;
1098 prof_info
.line_no
= -1;
1099 prof_info
.end_line_no
= -1;
1100 prof_info
.func_line_no
= -1;
1101 prof_info
.func_end_line_no
= -1;
1103 acc_event_info enter_exit_data_event_info
;
1106 enter_exit_data_event_info
.other_event
.event_type
1107 = prof_info
.event_type
;
1108 enter_exit_data_event_info
.other_event
.valid_bytes
1109 = _ACC_OTHER_EVENT_INFO_VALID_BYTES
;
1110 enter_exit_data_event_info
.other_event
.parent_construct
1111 = data_enter
? acc_construct_enter_data
: acc_construct_exit_data
;
1112 enter_exit_data_event_info
.other_event
.implicit
= 0;
1113 enter_exit_data_event_info
.other_event
.tool_info
= NULL
;
1115 acc_api_info api_info
;
1118 thr
->api_info
= &api_info
;
1120 api_info
.device_api
= acc_device_api_none
;
1121 api_info
.valid_bytes
= _ACC_API_INFO_VALID_BYTES
;
1122 api_info
.device_type
= prof_info
.device_type
;
1123 api_info
.vendor
= -1;
1124 api_info
.device_handle
= NULL
;
1125 api_info
.context_handle
= NULL
;
1126 api_info
.async_handle
= NULL
;
1130 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1133 if ((acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
1134 || (flags
& GOACC_FLAG_HOST_FALLBACK
))
1136 prof_info
.device_type
= acc_device_host
;
1137 api_info
.device_type
= prof_info
.device_type
;
1146 va_start (ap
, num_waits
);
1147 goacc_wait (async
, num_waits
, &ap
);
1151 /* In c, non-pointers and arrays are represented by a single data clause.
1152 Dynamically allocated arrays and subarrays are represented by a data
1153 clause followed by an internal GOMP_MAP_POINTER.
1155 In fortran, scalars and not allocated arrays are represented by a
1156 single data clause. Allocated arrays and subarrays have three mappings:
1157 1) the original data clause, 2) a PSET 3) a pointer to the array data.
1162 for (i
= 0; i
< mapnum
; i
++)
1164 unsigned char kind
= kinds
[i
] & 0xff;
1166 /* Scan for pointers and PSETs. */
1167 int pointer
= find_pointer (i
, mapnum
, kinds
);
1173 case GOMP_MAP_ALLOC
:
1174 case GOMP_MAP_FORCE_ALLOC
:
1175 acc_create_async (hostaddrs
[i
], sizes
[i
], async
);
1178 case GOMP_MAP_FORCE_TO
:
1179 acc_copyin_async (hostaddrs
[i
], sizes
[i
], async
);
1182 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1189 goacc_insert_pointer (pointer
, &hostaddrs
[i
], &sizes
[i
], &kinds
[i
],
1191 /* Increment 'i' by two because OpenACC requires fortran
1192 arrays to be contiguous, so each PSET is associated with
1193 one of MAP_FORCE_ALLOC/MAP_FORCE_PRESET/MAP_FORCE_TO, and
1200 for (i
= 0; i
< mapnum
; ++i
)
1202 unsigned char kind
= kinds
[i
] & 0xff;
1204 int pointer
= find_pointer (i
, mapnum
, kinds
);
1210 case GOMP_MAP_RELEASE
:
1211 case GOMP_MAP_DELETE
:
1212 if (acc_is_present (hostaddrs
[i
], sizes
[i
]))
1215 acc_delete_finalize_async (hostaddrs
[i
], sizes
[i
], async
);
1217 acc_delete_async (hostaddrs
[i
], sizes
[i
], async
);
1221 case GOMP_MAP_FORCE_FROM
:
1223 acc_copyout_finalize_async (hostaddrs
[i
], sizes
[i
], async
);
1225 acc_copyout_async (hostaddrs
[i
], sizes
[i
], async
);
1228 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1235 bool copyfrom
= (kind
== GOMP_MAP_FORCE_FROM
1236 || kind
== GOMP_MAP_FROM
);
1237 goacc_remove_pointer (hostaddrs
[i
], sizes
[i
], copyfrom
, async
,
1239 /* See the above comment. */
1247 prof_info
.event_type
1248 = data_enter
? acc_ev_enter_data_end
: acc_ev_exit_data_end
;
1249 enter_exit_data_event_info
.other_event
.event_type
= prof_info
.event_type
;
1250 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1253 thr
->prof_info
= NULL
;
1254 thr
->api_info
= NULL
;