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
;
495 #define FLAG_PRESENT (1 << 0)
496 #define FLAG_CREATE (1 << 1)
497 #define FLAG_COPY (1 << 2)
500 present_create_copy (unsigned f
, void *h
, size_t s
, int async
)
506 gomp_fatal ("[%p,+%d] is a bad range", (void *)h
, (int)s
);
508 goacc_lazy_initialize ();
510 struct goacc_thread
*thr
= goacc_thread ();
511 struct gomp_device_descr
*acc_dev
= thr
->dev
;
513 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
516 acc_prof_info prof_info
;
517 acc_api_info api_info
;
518 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
521 prof_info
.async
= async
;
522 prof_info
.async_queue
= prof_info
.async
;
525 gomp_mutex_lock (&acc_dev
->lock
);
527 n
= lookup_host (acc_dev
, h
, s
);
531 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
+ h
- n
->host_start
);
533 if (!(f
& FLAG_PRESENT
))
535 gomp_mutex_unlock (&acc_dev
->lock
);
536 gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
537 (void *)h
, (int)s
, (void *)d
, (int)s
);
539 if ((h
+ s
) > (void *)n
->host_end
)
541 gomp_mutex_unlock (&acc_dev
->lock
);
542 gomp_fatal ("[%p,+%d] not mapped", (void *)h
, (int)s
);
545 assert (n
->refcount
!= REFCOUNT_LINK
);
546 if (n
->refcount
!= REFCOUNT_INFINITY
)
548 n
->dynamic_refcount
++;
550 gomp_mutex_unlock (&acc_dev
->lock
);
552 else if (!(f
& FLAG_CREATE
))
554 gomp_mutex_unlock (&acc_dev
->lock
);
555 gomp_fatal ("[%p,+%d] not mapped", (void *)h
, (int)s
);
559 struct target_mem_desc
*tgt
;
561 unsigned short kinds
;
567 kinds
= GOMP_MAP_ALLOC
;
569 gomp_mutex_unlock (&acc_dev
->lock
);
571 goacc_aq aq
= get_goacc_asyncqueue (async
);
573 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, &hostaddrs
, NULL
, &s
,
574 &kinds
, true, GOMP_MAP_VARS_ENTER_DATA
);
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
)
663 struct goacc_thread
*thr
= goacc_thread ();
664 struct gomp_device_descr
*acc_dev
= thr
->dev
;
666 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
669 acc_prof_info prof_info
;
670 acc_api_info api_info
;
671 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
674 prof_info
.async
= async
;
675 prof_info
.async_queue
= prof_info
.async
;
678 gomp_mutex_lock (&acc_dev
->lock
);
680 n
= lookup_host (acc_dev
, h
, s
);
682 /* No need to call lazy open, as the data must already have been
687 gomp_mutex_unlock (&acc_dev
->lock
);
688 gomp_fatal ("[%p,%d] is not mapped", (void *)h
, (int)s
);
691 if ((uintptr_t) h
< n
->host_start
|| (uintptr_t) h
+ s
> n
->host_end
)
693 size_t host_size
= n
->host_end
- n
->host_start
;
694 gomp_mutex_unlock (&acc_dev
->lock
);
695 gomp_fatal ("[%p,+%d] outside mapped block [%p,+%d]",
696 (void *) h
, (int) s
, (void *) n
->host_start
, (int) host_size
);
699 assert (n
->refcount
!= REFCOUNT_LINK
);
700 if (n
->refcount
!= REFCOUNT_INFINITY
701 && n
->refcount
< n
->dynamic_refcount
)
703 gomp_mutex_unlock (&acc_dev
->lock
);
704 gomp_fatal ("Dynamic reference counting assert fail\n");
707 if (f
& FLAG_FINALIZE
)
709 if (n
->refcount
!= REFCOUNT_INFINITY
)
710 n
->refcount
-= n
->dynamic_refcount
;
711 n
->dynamic_refcount
= 0;
713 else if (n
->dynamic_refcount
)
715 if (n
->refcount
!= REFCOUNT_INFINITY
)
717 n
->dynamic_refcount
--;
720 if (n
->refcount
== 0)
722 goacc_aq aq
= get_goacc_asyncqueue (async
);
724 if (f
& FLAG_COPYOUT
)
726 void *d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
727 + (uintptr_t) h
- n
->host_start
);
728 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
732 /* TODO We can't do the 'is_tgt_unmapped' checking -- see the
733 'gomp_unref_tgt' comment in
734 <http://mid.mail-archive.com/878snl36eu.fsf@euler.schwinge.homeip.net>;
736 gomp_remove_var_async (acc_dev
, n
, aq
);
739 bool is_tgt_unmapped
= gomp_remove_var (acc_dev
, n
);
740 assert (is_tgt_unmapped
);
744 gomp_mutex_unlock (&acc_dev
->lock
);
748 thr
->prof_info
= NULL
;
749 thr
->api_info
= NULL
;
754 acc_delete (void *h
, size_t s
)
756 delete_copyout (0, h
, s
, acc_async_sync
, __FUNCTION__
);
760 acc_delete_async (void *h
, size_t s
, int async
)
762 delete_copyout (0, h
, s
, async
, __FUNCTION__
);
766 acc_delete_finalize (void *h
, size_t s
)
768 delete_copyout (FLAG_FINALIZE
, h
, s
, acc_async_sync
, __FUNCTION__
);
772 acc_delete_finalize_async (void *h
, size_t s
, int async
)
774 delete_copyout (FLAG_FINALIZE
, h
, s
, async
, __FUNCTION__
);
778 acc_copyout (void *h
, size_t s
)
780 delete_copyout (FLAG_COPYOUT
, h
, s
, acc_async_sync
, __FUNCTION__
);
784 acc_copyout_async (void *h
, size_t s
, int async
)
786 delete_copyout (FLAG_COPYOUT
, h
, s
, async
, __FUNCTION__
);
790 acc_copyout_finalize (void *h
, size_t s
)
792 delete_copyout (FLAG_COPYOUT
| FLAG_FINALIZE
, h
, s
, acc_async_sync
,
797 acc_copyout_finalize_async (void *h
, size_t s
, int async
)
799 delete_copyout (FLAG_COPYOUT
| FLAG_FINALIZE
, h
, s
, async
, __FUNCTION__
);
803 update_dev_host (int is_dev
, void *h
, size_t s
, int async
)
808 goacc_lazy_initialize ();
810 struct goacc_thread
*thr
= goacc_thread ();
811 struct gomp_device_descr
*acc_dev
= thr
->dev
;
813 if (acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
816 /* Fortran optional arguments that are non-present result in a
817 NULL host address here. This can safely be ignored as it is
818 not possible to 'update' a non-present optional argument. */
822 acc_prof_info prof_info
;
823 acc_api_info api_info
;
824 bool profiling_p
= GOACC_PROFILING_SETUP_P (thr
, &prof_info
, &api_info
);
827 prof_info
.async
= async
;
828 prof_info
.async_queue
= prof_info
.async
;
831 gomp_mutex_lock (&acc_dev
->lock
);
833 n
= lookup_host (acc_dev
, h
, s
);
837 gomp_mutex_unlock (&acc_dev
->lock
);
838 gomp_fatal ("[%p,%d] is not mapped", h
, (int)s
);
841 d
= (void *) (n
->tgt
->tgt_start
+ n
->tgt_offset
842 + (uintptr_t) h
- n
->host_start
);
844 goacc_aq aq
= get_goacc_asyncqueue (async
);
847 gomp_copy_host2dev (acc_dev
, aq
, d
, h
, s
, /* TODO: cbuf? */ NULL
);
849 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
851 gomp_mutex_unlock (&acc_dev
->lock
);
855 thr
->prof_info
= NULL
;
856 thr
->api_info
= NULL
;
861 acc_update_device (void *h
, size_t s
)
863 update_dev_host (1, h
, s
, acc_async_sync
);
867 acc_update_device_async (void *h
, size_t s
, int async
)
869 update_dev_host (1, h
, s
, async
);
873 acc_update_self (void *h
, size_t s
)
875 update_dev_host (0, h
, s
, acc_async_sync
);
879 acc_update_self_async (void *h
, size_t s
, int async
)
881 update_dev_host (0, h
, s
, async
);
885 /* OpenACC 'enter data', 'exit data': 'GOACC_enter_exit_data' and its helper
888 /* Special handling for 'GOMP_MAP_POINTER', 'GOMP_MAP_TO_PSET'.
890 Only the first mapping is considered in reference counting; the following
891 ones implicitly follow suit. Similarly, 'copyout' ('force_copyfrom') is
892 done only for the first mapping. */
895 goacc_insert_pointer (size_t mapnum
, void **hostaddrs
, size_t *sizes
,
896 void *kinds
, int async
)
898 struct target_mem_desc
*tgt
;
899 struct goacc_thread
*thr
= goacc_thread ();
900 struct gomp_device_descr
*acc_dev
= thr
->dev
;
902 if (*hostaddrs
== NULL
)
905 if (acc_is_present (*hostaddrs
, *sizes
))
908 gomp_mutex_lock (&acc_dev
->lock
);
909 n
= lookup_host (acc_dev
, *hostaddrs
, *sizes
);
910 assert (n
->refcount
!= REFCOUNT_INFINITY
911 && n
->refcount
!= REFCOUNT_LINK
);
912 gomp_mutex_unlock (&acc_dev
->lock
);
915 for (size_t i
= 0; i
< tgt
->list_count
; i
++)
916 if (tgt
->list
[i
].key
== n
)
918 for (size_t j
= 0; j
< mapnum
; j
++)
919 if (i
+ j
< tgt
->list_count
&& tgt
->list
[i
+ j
].key
)
921 tgt
->list
[i
+ j
].key
->refcount
++;
922 tgt
->list
[i
+ j
].key
->dynamic_refcount
++;
926 /* Should not reach here. */
927 gomp_fatal ("Dynamic refcount incrementing failed for pointer/pset");
930 gomp_debug (0, " %s: prepare mappings\n", __FUNCTION__
);
931 goacc_aq aq
= get_goacc_asyncqueue (async
);
932 tgt
= gomp_map_vars_async (acc_dev
, aq
, mapnum
, hostaddrs
,
933 NULL
, sizes
, kinds
, true, GOMP_MAP_VARS_ENTER_DATA
);
935 splay_tree_key n
= tgt
->list
[0].key
;
936 assert (n
->refcount
== 1);
937 assert (n
->dynamic_refcount
== 0);
938 n
->dynamic_refcount
++;
939 gomp_debug (0, " %s: mappings prepared\n", __FUNCTION__
);
943 goacc_remove_pointer (void *h
, size_t s
, bool force_copyfrom
, int async
,
946 struct goacc_thread
*thr
= goacc_thread ();
947 struct gomp_device_descr
*acc_dev
= thr
->dev
;
949 struct target_mem_desc
*t
;
951 if (!acc_is_present (h
, s
))
954 gomp_mutex_lock (&acc_dev
->lock
);
956 n
= lookup_host (acc_dev
, h
, 1);
960 gomp_mutex_unlock (&acc_dev
->lock
);
961 gomp_fatal ("%p is not a mapped block", (void *)h
);
964 gomp_debug (0, " %s: restore mappings\n", __FUNCTION__
);
968 assert (n
->refcount
!= REFCOUNT_INFINITY
969 && n
->refcount
!= REFCOUNT_LINK
);
970 if (n
->refcount
< n
->dynamic_refcount
)
972 gomp_mutex_unlock (&acc_dev
->lock
);
973 gomp_fatal ("Dynamic reference counting assert fail\n");
978 n
->refcount
-= n
->dynamic_refcount
;
979 n
->dynamic_refcount
= 0;
981 else if (n
->dynamic_refcount
)
984 n
->dynamic_refcount
--;
987 if (n
->refcount
== 0)
989 goacc_aq aq
= get_goacc_asyncqueue (async
);
993 void *d
= (void *) (t
->tgt_start
+ n
->tgt_offset
994 + (uintptr_t) h
- n
->host_start
);
996 gomp_copy_dev2host (acc_dev
, aq
, h
, d
, s
);
1001 /* TODO The way the following code is currently implemented, we need
1002 the 'is_tgt_unmapped' return value from 'gomp_remove_var', so
1003 can't use 'gomp_remove_var_async' here -- see the 'gomp_unref_tgt'
1005 <http://mid.mail-archive.com/878snl36eu.fsf@euler.schwinge.homeip.net>;
1006 PR92881 -- so have to synchronize here. */
1007 if (!acc_dev
->openacc
.async
.synchronize_func (aq
))
1009 gomp_mutex_unlock (&acc_dev
->lock
);
1010 gomp_fatal ("synchronize failed");
1013 bool is_tgt_unmapped
= false;
1014 for (size_t i
= 0; i
< t
->list_count
; i
++)
1016 is_tgt_unmapped
= gomp_remove_var (acc_dev
, t
->list
[i
].key
);
1017 if (is_tgt_unmapped
)
1020 assert (is_tgt_unmapped
);
1023 gomp_mutex_unlock (&acc_dev
->lock
);
1025 gomp_debug (0, " %s: mappings restored\n", __FUNCTION__
);
1028 /* Return the number of mappings associated with 'GOMP_MAP_TO_PSET' or
1029 'GOMP_MAP_POINTER'. */
1032 find_pointer (int pos
, size_t mapnum
, unsigned short *kinds
)
1034 if (pos
+ 1 >= mapnum
)
1037 unsigned char kind
= kinds
[pos
+1] & 0xff;
1039 if (kind
== GOMP_MAP_TO_PSET
)
1041 else if (kind
== GOMP_MAP_POINTER
)
1048 GOACC_enter_exit_data (int flags_m
, size_t mapnum
, void **hostaddrs
,
1049 size_t *sizes
, unsigned short *kinds
, int async
,
1052 int flags
= GOACC_FLAGS_UNMARSHAL (flags_m
);
1054 struct goacc_thread
*thr
;
1055 struct gomp_device_descr
*acc_dev
;
1056 bool data_enter
= false;
1059 goacc_lazy_initialize ();
1061 thr
= goacc_thread ();
1064 /* Determine whether "finalize" semantics apply to all mappings of this
1065 OpenACC directive. */
1066 bool finalize
= false;
1069 unsigned char kind
= kinds
[0] & 0xff;
1070 if (kind
== GOMP_MAP_DELETE
1071 || kind
== GOMP_MAP_FORCE_FROM
)
1075 /* Determine if this is an "acc enter data". */
1076 for (i
= 0; i
< mapnum
; ++i
)
1078 unsigned char kind
= kinds
[i
] & 0xff;
1080 if (kind
== GOMP_MAP_POINTER
|| kind
== GOMP_MAP_TO_PSET
)
1083 if (kind
== GOMP_MAP_FORCE_ALLOC
1084 || kind
== GOMP_MAP_FORCE_PRESENT
1085 || kind
== GOMP_MAP_FORCE_TO
1086 || kind
== GOMP_MAP_TO
1087 || kind
== GOMP_MAP_ALLOC
)
1093 if (kind
== GOMP_MAP_RELEASE
1094 || kind
== GOMP_MAP_DELETE
1095 || kind
== GOMP_MAP_FROM
1096 || kind
== GOMP_MAP_FORCE_FROM
)
1099 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1103 bool profiling_p
= GOACC_PROFILING_DISPATCH_P (true);
1105 acc_prof_info prof_info
;
1108 thr
->prof_info
= &prof_info
;
1110 prof_info
.event_type
1111 = data_enter
? acc_ev_enter_data_start
: acc_ev_exit_data_start
;
1112 prof_info
.valid_bytes
= _ACC_PROF_INFO_VALID_BYTES
;
1113 prof_info
.version
= _ACC_PROF_INFO_VERSION
;
1114 prof_info
.device_type
= acc_device_type (acc_dev
->type
);
1115 prof_info
.device_number
= acc_dev
->target_id
;
1116 prof_info
.thread_id
= -1;
1117 prof_info
.async
= async
;
1118 prof_info
.async_queue
= prof_info
.async
;
1119 prof_info
.src_file
= NULL
;
1120 prof_info
.func_name
= NULL
;
1121 prof_info
.line_no
= -1;
1122 prof_info
.end_line_no
= -1;
1123 prof_info
.func_line_no
= -1;
1124 prof_info
.func_end_line_no
= -1;
1126 acc_event_info enter_exit_data_event_info
;
1129 enter_exit_data_event_info
.other_event
.event_type
1130 = prof_info
.event_type
;
1131 enter_exit_data_event_info
.other_event
.valid_bytes
1132 = _ACC_OTHER_EVENT_INFO_VALID_BYTES
;
1133 enter_exit_data_event_info
.other_event
.parent_construct
1134 = data_enter
? acc_construct_enter_data
: acc_construct_exit_data
;
1135 enter_exit_data_event_info
.other_event
.implicit
= 0;
1136 enter_exit_data_event_info
.other_event
.tool_info
= NULL
;
1138 acc_api_info api_info
;
1141 thr
->api_info
= &api_info
;
1143 api_info
.device_api
= acc_device_api_none
;
1144 api_info
.valid_bytes
= _ACC_API_INFO_VALID_BYTES
;
1145 api_info
.device_type
= prof_info
.device_type
;
1146 api_info
.vendor
= -1;
1147 api_info
.device_handle
= NULL
;
1148 api_info
.context_handle
= NULL
;
1149 api_info
.async_handle
= NULL
;
1153 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1156 if ((acc_dev
->capabilities
& GOMP_OFFLOAD_CAP_SHARED_MEM
)
1157 || (flags
& GOACC_FLAG_HOST_FALLBACK
))
1159 prof_info
.device_type
= acc_device_host
;
1160 api_info
.device_type
= prof_info
.device_type
;
1169 va_start (ap
, num_waits
);
1170 goacc_wait (async
, num_waits
, &ap
);
1174 /* In c, non-pointers and arrays are represented by a single data clause.
1175 Dynamically allocated arrays and subarrays are represented by a data
1176 clause followed by an internal GOMP_MAP_POINTER.
1178 In fortran, scalars and not allocated arrays are represented by a
1179 single data clause. Allocated arrays and subarrays have three mappings:
1180 1) the original data clause, 2) a PSET 3) a pointer to the array data.
1185 for (i
= 0; i
< mapnum
; i
++)
1187 unsigned char kind
= kinds
[i
] & 0xff;
1189 /* Scan for pointers and PSETs. */
1190 int pointer
= find_pointer (i
, mapnum
, kinds
);
1196 case GOMP_MAP_ALLOC
:
1197 case GOMP_MAP_FORCE_ALLOC
:
1198 acc_create_async (hostaddrs
[i
], sizes
[i
], async
);
1201 case GOMP_MAP_FORCE_TO
:
1202 acc_copyin_async (hostaddrs
[i
], sizes
[i
], async
);
1205 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1212 goacc_insert_pointer (pointer
, &hostaddrs
[i
], &sizes
[i
], &kinds
[i
],
1214 /* Increment 'i' by two because OpenACC requires fortran
1215 arrays to be contiguous, so each PSET is associated with
1216 one of MAP_FORCE_ALLOC/MAP_FORCE_PRESET/MAP_FORCE_TO, and
1223 for (i
= 0; i
< mapnum
; ++i
)
1225 unsigned char kind
= kinds
[i
] & 0xff;
1227 int pointer
= find_pointer (i
, mapnum
, kinds
);
1233 case GOMP_MAP_RELEASE
:
1234 case GOMP_MAP_DELETE
:
1235 if (acc_is_present (hostaddrs
[i
], sizes
[i
]))
1238 acc_delete_finalize_async (hostaddrs
[i
], sizes
[i
], async
);
1240 acc_delete_async (hostaddrs
[i
], sizes
[i
], async
);
1244 case GOMP_MAP_FORCE_FROM
:
1246 acc_copyout_finalize_async (hostaddrs
[i
], sizes
[i
], async
);
1248 acc_copyout_async (hostaddrs
[i
], sizes
[i
], async
);
1251 gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
1258 bool copyfrom
= (kind
== GOMP_MAP_FORCE_FROM
1259 || kind
== GOMP_MAP_FROM
);
1260 goacc_remove_pointer (hostaddrs
[i
], sizes
[i
], copyfrom
, async
,
1262 /* See the above comment. */
1270 prof_info
.event_type
1271 = data_enter
? acc_ev_enter_data_end
: acc_ev_exit_data_end
;
1272 enter_exit_data_event_info
.other_event
.event_type
= prof_info
.event_type
;
1273 goacc_profiling_dispatch (&prof_info
, &enter_exit_data_event_info
,
1276 thr
->prof_info
= NULL
;
1277 thr
->api_info
= NULL
;