/* OpenACC Runtime initialization routines
- Copyright (C) 2013-2019 Free Software Foundation, Inc.
+ Copyright (C) 2013-2024 Free Software Foundation, Inc.
Contributed by Mentor Embedded.
if (from)
gomp_copy_dev2host (thr->dev, aq, h, d, s);
else
- gomp_copy_host2dev (thr->dev, aq, d, h, s, /* TODO: cbuf? */ NULL);
+ gomp_copy_host2dev (thr->dev, aq, d, h, s, false, /* TODO: cbuf? */ NULL);
if (profiling_p)
{
void
acc_map_data (void *h, void *d, size_t s)
{
- struct target_mem_desc *tgt = NULL;
size_t mapnum = 1;
void *hostaddrs = h;
void *devaddrs = d;
gomp_mutex_unlock (&acc_dev->lock);
- tgt = gomp_map_vars (acc_dev, mapnum, &hostaddrs, &devaddrs, &sizes,
- &kinds, true, GOMP_MAP_VARS_OPENACC);
+ struct target_mem_desc *tgt
+ = goacc_map_vars (acc_dev, NULL, mapnum, &hostaddrs, &devaddrs, &sizes,
+ &kinds, true, GOMP_MAP_VARS_ENTER_DATA);
+ assert (tgt);
+ assert (tgt->list_count == 1);
splay_tree_key n = tgt->list[0].key;
+ assert (n);
assert (n->refcount == 1);
assert (n->dynamic_refcount == 0);
/* Special reference counting behavior. */
acc_api_info api_info;
bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
- size_t host_size;
-
gomp_mutex_lock (&acc_dev->lock);
splay_tree_key n = lookup_host (acc_dev, h, 1);
- struct target_mem_desc *t;
if (!n)
{
gomp_fatal ("%p is not a mapped block", (void *)h);
}
- host_size = n->host_end - n->host_start;
+ size_t host_size = n->host_end - n->host_start;
if (n->host_start != (uintptr_t) h)
{
(void *) h, (int) host_size);
}
- /* Mark for removal. */
- n->refcount = 1;
-
- t = n->tgt;
+ struct target_mem_desc *tgt = n->tgt;
- if (t->refcount == 2)
+ if (tgt->refcount == REFCOUNT_INFINITY)
{
- /* This is the last reference, so pull the descriptor off the
- chain. This avoids gomp_unmap_vars via gomp_unmap_tgt from
- freeing the device memory. */
- t->tgt_end = 0;
- t->to_free = 0;
+ gomp_mutex_unlock (&acc_dev->lock);
+ gomp_fatal ("cannot unmap target block");
}
- gomp_mutex_unlock (&acc_dev->lock);
+ /* Above, we've verified that the mapping must have been set up by
+ 'acc_map_data'. */
+ assert (tgt->refcount == 1);
- gomp_unmap_vars (t, true);
+ /* Nullifying these fields prevents 'gomp_unmap_tgt' via 'gomp_remove_var'
+ from freeing the target memory. */
+ tgt->tgt_end = 0;
+ tgt->to_free = NULL;
+
+ bool is_tgt_unmapped = gomp_remove_var (acc_dev, n);
+ assert (is_tgt_unmapped);
+
+ gomp_mutex_unlock (&acc_dev->lock);
if (profiling_p)
{
}
}
-#define FLAG_PRESENT (1 << 0)
-#define FLAG_CREATE (1 << 1)
-#define FLAG_COPY (1 << 2)
+
+/* Helper function to map a single dynamic data item, represented by a single
+ mapping. The acc_dev->lock should be held on entry, and remains locked on
+ exit. */
static void *
-present_create_copy (unsigned f, void *h, size_t s, int async)
+goacc_map_var_existing (struct gomp_device_descr *acc_dev, void *hostaddr,
+ size_t size, splay_tree_key n)
+{
+ assert (n);
+
+ /* Present. */
+ void *d = (void *) (n->tgt->tgt_start + n->tgt_offset + hostaddr
+ - n->host_start);
+
+ if (hostaddr + size > (void *) n->host_end)
+ {
+ gomp_mutex_unlock (&acc_dev->lock);
+ gomp_fatal ("[%p,+%d] not mapped", hostaddr, (int) size);
+ }
+
+ assert (n->refcount != REFCOUNT_LINK);
+ if (n->refcount != REFCOUNT_INFINITY)
+ n->refcount++;
+ n->dynamic_refcount++;
+
+ return d;
+}
+
+/* Enter dynamic mapping for a single datum. Return the device pointer. */
+
+static void *
+goacc_enter_datum (void **hostaddrs, size_t *sizes, void *kinds, int async)
{
void *d;
splay_tree_key n;
- if (!h || !s)
- gomp_fatal ("[%p,+%d] is a bad range", (void *)h, (int)s);
+ if (!hostaddrs[0] || !sizes[0])
+ gomp_fatal ("[%p,+%d] is a bad range", hostaddrs[0], (int) sizes[0]);
goacc_lazy_initialize ();
struct gomp_device_descr *acc_dev = thr->dev;
if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
- return h;
+ return hostaddrs[0];
acc_prof_info prof_info;
acc_api_info api_info;
gomp_mutex_lock (&acc_dev->lock);
- n = lookup_host (acc_dev, h, s);
+ n = lookup_host (acc_dev, hostaddrs[0], sizes[0]);
if (n)
{
- /* Present. */
- d = (void *) (n->tgt->tgt_start + n->tgt_offset + h - n->host_start);
-
- if (!(f & FLAG_PRESENT))
- {
- gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("[%p,+%d] already mapped to [%p,+%d]",
- (void *)h, (int)s, (void *)d, (int)s);
- }
- if ((h + s) > (void *)n->host_end)
- {
- gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
- }
-
- assert (n->refcount != REFCOUNT_LINK);
- if (n->refcount != REFCOUNT_INFINITY)
- n->refcount++;
- n->dynamic_refcount++;
-
- gomp_mutex_unlock (&acc_dev->lock);
- }
- else if (!(f & FLAG_CREATE))
- {
+ d = goacc_map_var_existing (acc_dev, hostaddrs[0], sizes[0], n);
gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("[%p,+%d] not mapped", (void *)h, (int)s);
}
else
{
- struct target_mem_desc *tgt;
- size_t mapnum = 1;
- unsigned short kinds;
- void *hostaddrs = h;
-
- if (f & FLAG_COPY)
- kinds = GOMP_MAP_TO;
- else
- kinds = GOMP_MAP_ALLOC;
+ const size_t mapnum = 1;
gomp_mutex_unlock (&acc_dev->lock);
goacc_aq aq = get_goacc_asyncqueue (async);
- tgt = gomp_map_vars_async (acc_dev, aq, mapnum, &hostaddrs, NULL, &s,
- &kinds, true, GOMP_MAP_VARS_OPENACC);
+ struct target_mem_desc *tgt
+ = goacc_map_vars (acc_dev, aq, mapnum, hostaddrs, NULL, sizes,
+ kinds, true, GOMP_MAP_VARS_ENTER_DATA);
+ assert (tgt);
+ assert (tgt->list_count == 1);
n = tgt->list[0].key;
+ assert (n);
assert (n->refcount == 1);
assert (n->dynamic_refcount == 0);
n->dynamic_refcount++;
- d = tgt->to_free;
+ d = (void *) tgt->tgt_start;
}
if (profiling_p)
void *
acc_create (void *h, size_t s)
{
- return present_create_copy (FLAG_PRESENT | FLAG_CREATE, h, s, acc_async_sync);
+ unsigned short kinds[1] = { GOMP_MAP_ALLOC };
+ return goacc_enter_datum (&h, &s, &kinds, acc_async_sync);
}
void
acc_create_async (void *h, size_t s, int async)
{
- present_create_copy (FLAG_PRESENT | FLAG_CREATE, h, s, async);
+ unsigned short kinds[1] = { GOMP_MAP_ALLOC };
+ goacc_enter_datum (&h, &s, &kinds, async);
}
/* acc_present_or_create used to be what acc_create is now. */
void *
acc_copyin (void *h, size_t s)
{
- return present_create_copy (FLAG_PRESENT | FLAG_CREATE | FLAG_COPY, h, s,
- acc_async_sync);
+ unsigned short kinds[1] = { GOMP_MAP_TO };
+ return goacc_enter_datum (&h, &s, &kinds, acc_async_sync);
}
void
acc_copyin_async (void *h, size_t s, int async)
{
- present_create_copy (FLAG_PRESENT | FLAG_CREATE | FLAG_COPY, h, s, async);
+ unsigned short kinds[1] = { GOMP_MAP_TO };
+ goacc_enter_datum (&h, &s, &kinds, async);
}
/* acc_present_or_copyin used to be what acc_copyin is now. */
}
#endif
-#define FLAG_COPYOUT (1 << 0)
-#define FLAG_FINALIZE (1 << 1)
+
+/* Helper function to unmap a single data item. Device lock should be held on
+ entry, and remains locked on exit. */
static void
-delete_copyout (unsigned f, void *h, size_t s, int async, const char *libfnname)
+goacc_exit_datum_1 (struct gomp_device_descr *acc_dev, void *h, size_t s,
+ unsigned short kind, splay_tree_key n, goacc_aq aq)
{
- splay_tree_key n;
- struct goacc_thread *thr = goacc_thread ();
- struct gomp_device_descr *acc_dev = thr->dev;
-
- if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
- return;
-
- acc_prof_info prof_info;
- acc_api_info api_info;
- bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
- if (profiling_p)
- {
- prof_info.async = async;
- prof_info.async_queue = prof_info.async;
- }
-
- gomp_mutex_lock (&acc_dev->lock);
-
- n = lookup_host (acc_dev, h, s);
-
- /* No need to call lazy open, as the data must already have been
- mapped. */
-
- if (!n)
- {
- gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("[%p,%d] is not mapped", (void *)h, (int)s);
- }
+ assert (kind != GOMP_MAP_DETACH
+ && kind != GOMP_MAP_FORCE_DETACH);
if ((uintptr_t) h < n->host_start || (uintptr_t) h + s > n->host_end)
{
(void *) h, (int) s, (void *) n->host_start, (int) host_size);
}
+ bool finalize = (kind == GOMP_MAP_FORCE_FROM
+ || kind == GOMP_MAP_DELETE);
+
assert (n->refcount != REFCOUNT_LINK);
if (n->refcount != REFCOUNT_INFINITY
&& n->refcount < n->dynamic_refcount)
gomp_fatal ("Dynamic reference counting assert fail\n");
}
- if (f & FLAG_FINALIZE)
+ if (finalize)
{
if (n->refcount != REFCOUNT_INFINITY)
n->refcount -= n->dynamic_refcount;
if (n->refcount == 0)
{
- goacc_aq aq = get_goacc_asyncqueue (async);
-
- if (f & FLAG_COPYOUT)
+ bool copyout = (kind == GOMP_MAP_FROM
+ || kind == GOMP_MAP_FORCE_FROM);
+ if (copyout)
{
void *d = (void *) (n->tgt->tgt_start + n->tgt_offset
+ (uintptr_t) h - n->host_start);
gomp_copy_dev2host (acc_dev, aq, h, d, s);
}
- gomp_remove_var_async (acc_dev, n, aq);
+
+ if (aq)
+ /* TODO We can't do the 'is_tgt_unmapped' checking -- see the
+ 'gomp_unref_tgt' comment in
+ <http://mid.mail-archive.com/878snl36eu.fsf@euler.schwinge.homeip.net>;
+ PR92881. */
+ gomp_remove_var_async (acc_dev, n, aq);
+ else
+ {
+ size_t num_mappings = 0;
+ /* If the target_mem_desc represents a single data mapping, we can
+ check that it is freed when this splay tree key's refcount reaches
+ zero. Otherwise (e.g. for a 'GOMP_MAP_STRUCT' mapping with
+ multiple members), fall back to skipping the test. */
+ for (size_t l_i = 0; l_i < n->tgt->list_count; ++l_i)
+ if (n->tgt->list[l_i].key
+ && !n->tgt->list[l_i].is_attach)
+ ++num_mappings;
+ bool is_tgt_unmapped = gomp_remove_var (acc_dev, n);
+ assert (is_tgt_unmapped || num_mappings > 1);
+ }
+ }
+}
+
+
+/* Exit a dynamic mapping for a single variable. */
+
+static void
+goacc_exit_datum (void *h, size_t s, unsigned short kind, int async)
+{
+ /* No need to call lazy open, as the data must already have been
+ mapped. */
+
+ kind &= 0xff;
+
+ struct goacc_thread *thr = goacc_thread ();
+ struct gomp_device_descr *acc_dev = thr->dev;
+
+ if (acc_dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
+ return;
+
+ acc_prof_info prof_info;
+ acc_api_info api_info;
+ bool profiling_p = GOACC_PROFILING_SETUP_P (thr, &prof_info, &api_info);
+ if (profiling_p)
+ {
+ prof_info.async = async;
+ prof_info.async_queue = prof_info.async;
+ }
+
+ gomp_mutex_lock (&acc_dev->lock);
+
+ splay_tree_key n = lookup_host (acc_dev, h, s);
+ /* Non-present data is a no-op: PR92726, RP92970, PR92984. */
+ if (n)
+ {
+ goacc_aq aq = get_goacc_asyncqueue (async);
+ goacc_exit_datum_1 (acc_dev, h, s, kind, n, aq);
}
gomp_mutex_unlock (&acc_dev->lock);
void
acc_delete (void *h , size_t s)
{
- delete_copyout (0, h, s, acc_async_sync, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_RELEASE, acc_async_sync);
}
void
acc_delete_async (void *h , size_t s, int async)
{
- delete_copyout (0, h, s, async, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_RELEASE, async);
}
void
acc_delete_finalize (void *h , size_t s)
{
- delete_copyout (FLAG_FINALIZE, h, s, acc_async_sync, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_DELETE, acc_async_sync);
}
void
acc_delete_finalize_async (void *h , size_t s, int async)
{
- delete_copyout (FLAG_FINALIZE, h, s, async, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_DELETE, async);
}
void
acc_copyout (void *h, size_t s)
{
- delete_copyout (FLAG_COPYOUT, h, s, acc_async_sync, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_FROM, acc_async_sync);
}
void
acc_copyout_async (void *h, size_t s, int async)
{
- delete_copyout (FLAG_COPYOUT, h, s, async, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_FROM, async);
}
void
acc_copyout_finalize (void *h, size_t s)
{
- delete_copyout (FLAG_COPYOUT | FLAG_FINALIZE, h, s, acc_async_sync,
- __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_FORCE_FROM, acc_async_sync);
}
void
acc_copyout_finalize_async (void *h, size_t s, int async)
{
- delete_copyout (FLAG_COPYOUT | FLAG_FINALIZE, h, s, async, __FUNCTION__);
+ goacc_exit_datum (h, s, GOMP_MAP_FORCE_FROM, async);
}
static void
goacc_aq aq = get_goacc_asyncqueue (async);
if (is_dev)
- gomp_copy_host2dev (acc_dev, aq, d, h, s, /* TODO: cbuf? */ NULL);
+ gomp_copy_host2dev (acc_dev, aq, d, h, s, false, /* TODO: cbuf? */ NULL);
else
gomp_copy_dev2host (acc_dev, aq, h, d, s);
update_dev_host (0, h, s, async);
}
-
-/* OpenACC 'enter data', 'exit data': 'GOACC_enter_exit_data' and its helper
- functions. */
-
-/* Special handling for 'GOMP_MAP_POINTER', 'GOMP_MAP_TO_PSET'.
-
- Only the first mapping is considered in reference counting; the following
- ones implicitly follow suit. */
-
-static void
-goacc_insert_pointer (size_t mapnum, void **hostaddrs, size_t *sizes,
- void *kinds, int async)
+void
+acc_attach_async (void **hostaddr, int async)
{
- struct target_mem_desc *tgt;
struct goacc_thread *thr = goacc_thread ();
struct gomp_device_descr *acc_dev = thr->dev;
+ goacc_aq aq = get_goacc_asyncqueue (async);
+
+ struct splay_tree_key_s cur_node;
+ splay_tree_key n;
- if (*hostaddrs == NULL)
+ if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
return;
- if (acc_is_present (*hostaddrs, *sizes))
+ gomp_mutex_lock (&acc_dev->lock);
+
+ cur_node.host_start = (uintptr_t) hostaddr;
+ cur_node.host_end = cur_node.host_start + sizeof (void *);
+ n = splay_tree_lookup (&acc_dev->mem_map, &cur_node);
+
+ if (n == NULL)
{
- splay_tree_key n;
- gomp_mutex_lock (&acc_dev->lock);
- n = lookup_host (acc_dev, *hostaddrs, *sizes);
- assert (n->refcount != REFCOUNT_INFINITY
- && n->refcount != REFCOUNT_LINK);
gomp_mutex_unlock (&acc_dev->lock);
-
- tgt = n->tgt;
- for (size_t i = 0; i < tgt->list_count; i++)
- if (tgt->list[i].key == n)
- {
- for (size_t j = 0; j < mapnum; j++)
- if (i + j < tgt->list_count && tgt->list[i + j].key)
- {
- tgt->list[i + j].key->refcount++;
- tgt->list[i + j].key->dynamic_refcount++;
- }
- return;
- }
- /* Should not reach here. */
- gomp_fatal ("Dynamic refcount incrementing failed for pointer/pset");
+ gomp_fatal ("struct not mapped for acc_attach");
}
- gomp_debug (0, " %s: prepare mappings\n", __FUNCTION__);
- goacc_aq aq = get_goacc_asyncqueue (async);
- tgt = gomp_map_vars_async (acc_dev, aq, mapnum, hostaddrs,
- NULL, sizes, kinds, true, GOMP_MAP_VARS_OPENACC);
- splay_tree_key n = tgt->list[0].key;
- assert (n->refcount == 1);
- assert (n->dynamic_refcount == 0);
- n->dynamic_refcount++;
- gomp_debug (0, " %s: mappings prepared\n", __FUNCTION__);
+ gomp_attach_pointer (acc_dev, aq, &acc_dev->mem_map, n, (uintptr_t) hostaddr,
+ 0, NULL, false);
+
+ gomp_mutex_unlock (&acc_dev->lock);
+}
+
+void
+acc_attach (void **hostaddr)
+{
+ acc_attach_async (hostaddr, acc_async_sync);
}
static void
-goacc_remove_pointer (void *h, size_t s, bool force_copyfrom, int async,
- int finalize, int mapnum)
+goacc_detach_internal (void **hostaddr, int async, bool finalize)
{
struct goacc_thread *thr = goacc_thread ();
struct gomp_device_descr *acc_dev = thr->dev;
+ struct splay_tree_key_s cur_node;
splay_tree_key n;
- struct target_mem_desc *t;
- int minrefs = (mapnum == 1) ? 2 : 3;
+ struct goacc_asyncqueue *aq = get_goacc_asyncqueue (async);
- if (!acc_is_present (h, s))
+ if (thr->dev->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM)
return;
gomp_mutex_lock (&acc_dev->lock);
- n = lookup_host (acc_dev, h, 1);
+ cur_node.host_start = (uintptr_t) hostaddr;
+ cur_node.host_end = cur_node.host_start + sizeof (void *);
+ n = splay_tree_lookup (&acc_dev->mem_map, &cur_node);
- if (!n)
+ if (n == NULL)
{
gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("%p is not a mapped block", (void *)h);
+ gomp_fatal ("struct not mapped for acc_detach");
}
- gomp_debug (0, " %s: restore mappings\n", __FUNCTION__);
+ gomp_detach_pointer (acc_dev, aq, n, (uintptr_t) hostaddr, finalize, NULL);
- t = n->tgt;
+ gomp_mutex_unlock (&acc_dev->lock);
+}
- assert (n->refcount != REFCOUNT_INFINITY
- && n->refcount != REFCOUNT_LINK);
- if (n->refcount < n->dynamic_refcount)
- {
- gomp_mutex_unlock (&acc_dev->lock);
- gomp_fatal ("Dynamic reference counting assert fail\n");
- }
+void
+acc_detach (void **hostaddr)
+{
+ goacc_detach_internal (hostaddr, acc_async_sync, false);
+}
- if (finalize)
- {
- n->refcount -= n->dynamic_refcount;
- n->dynamic_refcount = 0;
- }
- else if (n->dynamic_refcount)
+void
+acc_detach_async (void **hostaddr, int async)
+{
+ goacc_detach_internal (hostaddr, async, false);
+}
+
+void
+acc_detach_finalize (void **hostaddr)
+{
+ goacc_detach_internal (hostaddr, acc_async_sync, true);
+}
+
+void
+acc_detach_finalize_async (void **hostaddr, int async)
+{
+ goacc_detach_internal (hostaddr, async, true);
+}
+
+/* Some types of (pointer) variables use several consecutive mappings, which
+ must be treated as a group for enter/exit data directives. This function
+ returns the last mapping in such a group (inclusive), or POS for singleton
+ mappings. */
+
+static int
+find_group_last (int pos, size_t mapnum, size_t *sizes, unsigned short *kinds)
+{
+ unsigned char kind0 = kinds[pos] & 0xff;
+ int first_pos = pos;
+
+ switch (kind0)
{
- n->refcount--;
- n->dynamic_refcount--;
+ case GOMP_MAP_TO_PSET:
+ if (pos + 1 < mapnum
+ && (kinds[pos + 1] & 0xff) == GOMP_MAP_ATTACH)
+ return pos + 1;
+
+ while (pos + 1 < mapnum
+ && (kinds[pos + 1] & 0xff) == GOMP_MAP_POINTER)
+ pos++;
+ /* We expect at least one GOMP_MAP_POINTER (if not a single
+ GOMP_MAP_ATTACH) after a GOMP_MAP_TO_PSET. */
+ assert (pos > first_pos);
+ break;
+
+ case GOMP_MAP_STRUCT:
+ case GOMP_MAP_STRUCT_UNORD:
+ pos += sizes[pos];
+ break;
+
+ case GOMP_MAP_POINTER:
+ case GOMP_MAP_ALWAYS_POINTER:
+ /* These mappings are only expected after some other mapping. If we
+ see one by itself, something has gone wrong. */
+ gomp_fatal ("unexpected mapping");
+ break;
+
+ case GOMP_MAP_ATTACH:
+ break;
+
+ default:
+ /* GOMP_MAP_ALWAYS_POINTER can only appear directly after some other
+ mapping. */
+ if (pos + 1 < mapnum)
+ {
+ unsigned char kind1 = kinds[pos + 1] & 0xff;
+ if (kind1 == GOMP_MAP_ALWAYS_POINTER)
+ return pos + 1;
+ }
+
+ /* We can have a single GOMP_MAP_ATTACH mapping after a to/from
+ mapping. */
+ if (pos + 1 < mapnum
+ && (kinds[pos + 1] & 0xff) == GOMP_MAP_ATTACH)
+ return pos + 1;
+
+ /* We can have zero or more GOMP_MAP_POINTER mappings after a to/from
+ (etc.) mapping. */
+ while (pos + 1 < mapnum
+ && (kinds[pos + 1] & 0xff) == GOMP_MAP_POINTER)
+ pos++;
}
- gomp_mutex_unlock (&acc_dev->lock);
+ return pos;
+}
- if (n->refcount == 0)
+/* Map variables for OpenACC "enter data". We can't just call
+ goacc_map_vars once, because individual mapped variables might have
+ "exit data" called for them at different times. */
+
+static void
+goacc_enter_data_internal (struct gomp_device_descr *acc_dev, size_t mapnum,
+ void **hostaddrs, size_t *sizes,
+ unsigned short *kinds, goacc_aq aq)
+{
+ gomp_mutex_lock (&acc_dev->lock);
+
+ for (size_t i = 0; i < mapnum; i++)
{
- /* Set refcount to 1 to allow gomp_unmap_vars to unmap it. */
- n->refcount = 1;
- t->refcount = minrefs;
- for (size_t i = 0; i < t->list_count; i++)
- if (t->list[i].key == n)
+ splay_tree_key n;
+ size_t group_last = find_group_last (i, mapnum, sizes, kinds);
+ bool struct_p = false;
+ size_t size, groupnum = (group_last - i) + 1;
+
+ switch (kinds[i] & 0xff)
+ {
+ case GOMP_MAP_STRUCT:
+ case GOMP_MAP_STRUCT_UNORD:
{
- t->list[i].copy_from = force_copyfrom ? 1 : 0;
- break;
+ size = (uintptr_t) hostaddrs[group_last] + sizes[group_last]
+ - (uintptr_t) hostaddrs[i];
+ struct_p = true;
}
+ break;
- /* If running synchronously, unmap immediately. */
- if (async < acc_async_noval)
- gomp_unmap_vars (t, true);
- else
+ case GOMP_MAP_ATTACH:
+ size = sizeof (void *);
+ break;
+
+ default:
+ size = sizes[i];
+ }
+
+ n = lookup_host (acc_dev, hostaddrs[i], size);
+
+ if (n && struct_p)
{
- goacc_aq aq = get_goacc_asyncqueue (async);
- gomp_unmap_vars_async (t, true, aq);
+ for (size_t j = i + 1; j <= group_last; j++)
+ {
+ struct splay_tree_key_s cur_node;
+ cur_node.host_start = (uintptr_t) hostaddrs[j];
+ cur_node.host_end = cur_node.host_start + sizes[j];
+ splay_tree_key n2
+ = splay_tree_lookup (&acc_dev->mem_map, &cur_node);
+ if (!n2
+ || n2->tgt != n->tgt
+ || n2->host_start - n->host_start
+ != n2->tgt_offset - n->tgt_offset)
+ {
+ gomp_mutex_unlock (&acc_dev->lock);
+ gomp_fatal ("Trying to map into device [%p..%p) structure "
+ "element when other mapped elements from the "
+ "same structure weren't mapped together with "
+ "it", (void *) cur_node.host_start,
+ (void *) cur_node.host_end);
+ }
+ }
+ /* This is a special case because we must increment the refcount by
+ the number of mapped struct elements, rather than by one. */
+ if (n->refcount != REFCOUNT_INFINITY)
+ n->refcount += groupnum - 1;
+ n->dynamic_refcount += groupnum - 1;
}
- }
+ else if (n && groupnum == 1)
+ {
+ void *h = hostaddrs[i];
+ size_t s = sizes[i];
- gomp_mutex_unlock (&acc_dev->lock);
+ if ((kinds[i] & 0xff) == GOMP_MAP_ATTACH)
+ {
+ gomp_attach_pointer (acc_dev, aq, &acc_dev->mem_map, n,
+ (uintptr_t) h, s, NULL, false);
+ /* OpenACC 'attach'/'detach' doesn't affect structured/dynamic
+ reference counts ('n->refcount', 'n->dynamic_refcount'). */
+ }
+ else
+ goacc_map_var_existing (acc_dev, h, s, n);
+ }
+ else if (n && groupnum > 1)
+ {
+ assert (n->refcount != REFCOUNT_LINK);
- gomp_debug (0, " %s: mappings restored\n", __FUNCTION__);
-}
+ for (size_t j = i + 1; j <= group_last; j++)
+ if ((kinds[j] & 0xff) == GOMP_MAP_ATTACH)
+ {
+ splay_tree_key m
+ = lookup_host (acc_dev, hostaddrs[j], sizeof (void *));
+ gomp_attach_pointer (acc_dev, aq, &acc_dev->mem_map, m,
+ (uintptr_t) hostaddrs[j], sizes[j], NULL,
+ false);
+ }
-/* Return the number of mappings associated with 'GOMP_MAP_TO_PSET' or
- 'GOMP_MAP_POINTER'. */
+ bool processed = false;
-static int
-find_pointer (int pos, size_t mapnum, unsigned short *kinds)
-{
- if (pos + 1 >= mapnum)
- return 0;
+ struct target_mem_desc *tgt = n->tgt;
- unsigned char kind = kinds[pos+1] & 0xff;
+ /* Minimal OpenACC variant corresponding to PR96668
+ "[OpenMP] Re-mapping allocated but previously unallocated
+ allocatable does not work" 'libgomp/target.c' changes, so that
+ OpenACC 'declare' code à la PR106643
+ "[gfortran + OpenACC] Allocate in module causes refcount error"
+ has a chance to work. */
+ if ((kinds[i] & 0xff) == GOMP_MAP_TO_PSET
+ && tgt->list_count == 0)
+ {
+ /* 'declare target'. */
+ assert (n->refcount == REFCOUNT_INFINITY);
- if (kind == GOMP_MAP_TO_PSET)
- return 3;
- else if (kind == GOMP_MAP_POINTER)
- return 2;
+ for (size_t k = 1; k < groupnum; k++)
+ {
+ /* The only thing we expect to see here. */
+ assert ((kinds[i + k] & 0xff) == GOMP_MAP_POINTER);
+ }
- return 0;
-}
+ /* Let 'goacc_map_vars' -> 'gomp_map_vars_internal' handle
+ this. */
+ gomp_mutex_unlock (&acc_dev->lock);
+ struct target_mem_desc *tgt_
+ = goacc_map_vars (acc_dev, aq, groupnum, &hostaddrs[i], NULL,
+ &sizes[i], &kinds[i], true,
+ GOMP_MAP_VARS_ENTER_DATA);
+ assert (tgt_ == NULL);
+ gomp_mutex_lock (&acc_dev->lock);
+
+ /* Given that 'goacc_exit_data_internal'/'goacc_exit_datum_1'
+ will always see 'n->refcount == REFCOUNT_INFINITY',
+ there's no need to adjust 'n->dynamic_refcount' here. */
+
+ processed = true;
+ }
+ else
+ assert (n->refcount != REFCOUNT_INFINITY);
-void
-GOACC_enter_exit_data (int flags_m, size_t mapnum, void **hostaddrs,
- size_t *sizes, unsigned short *kinds, int async,
- int num_waits, ...)
-{
- int flags = GOACC_FLAGS_UNMARSHAL (flags_m);
+ for (size_t j = 0; j < tgt->list_count; j++)
+ if (tgt->list[j].key == n)
+ {
+ /* We are processing a group of mappings (e.g.
+ [GOMP_MAP_TO, GOMP_MAP_TO_PSET, GOMP_MAP_POINTER]).
+ Find the right group in the target_mem_desc's variable
+ list, and increment the refcounts for each item in that
+ group. */
+ for (size_t k = 0; k < groupnum; k++)
+ if (j + k < tgt->list_count
+ && tgt->list[j + k].key
+ && !tgt->list[j + k].is_attach)
+ {
+ tgt->list[j + k].key->refcount++;
+ tgt->list[j + k].key->dynamic_refcount++;
+ }
+ processed = true;
+ break;
+ }
- struct goacc_thread *thr;
- struct gomp_device_descr *acc_dev;
- bool data_enter = false;
- size_t i;
+ if (!processed)
+ {
+ gomp_mutex_unlock (&acc_dev->lock);
+ gomp_fatal ("dynamic refcount incrementing failed for "
+ "pointer/pset");
+ }
+ }
+ else if (hostaddrs[i])
+ {
+ /* The data is not mapped already. Map it now, unless the first
+ member in the group has a NULL pointer (e.g. a non-present
+ optional parameter). */
+ gomp_mutex_unlock (&acc_dev->lock);
- goacc_lazy_initialize ();
+ struct target_mem_desc *tgt
+ = goacc_map_vars (acc_dev, aq, groupnum, &hostaddrs[i], NULL,
+ &sizes[i], &kinds[i], true,
+ GOMP_MAP_VARS_ENTER_DATA);
+ assert (tgt);
- thr = goacc_thread ();
- acc_dev = thr->dev;
+ gomp_mutex_lock (&acc_dev->lock);
- /* Determine whether "finalize" semantics apply to all mappings of this
- OpenACC directive. */
- bool finalize = false;
- if (mapnum > 0)
- {
- unsigned char kind = kinds[0] & 0xff;
- if (kind == GOMP_MAP_DELETE
- || kind == GOMP_MAP_FORCE_FROM)
- finalize = true;
+ for (size_t j = 0; j < tgt->list_count; j++)
+ {
+ n = tgt->list[j].key;
+ if (n && !tgt->list[j].is_attach)
+ n->dynamic_refcount++;
+ }
+ }
+
+ i = group_last;
}
- /* Determine if this is an "acc enter data". */
- for (i = 0; i < mapnum; ++i)
+ gomp_mutex_unlock (&acc_dev->lock);
+}
+
+/* Unmap variables for OpenACC "exit data". */
+
+static void
+goacc_exit_data_internal (struct gomp_device_descr *acc_dev, size_t mapnum,
+ void **hostaddrs, size_t *sizes,
+ unsigned short *kinds, goacc_aq aq)
+{
+ gomp_mutex_lock (&acc_dev->lock);
+
+ /* Handle "detach" before copyback/deletion of mapped data. */
+ for (size_t i = 0; i < mapnum; ++i)
{
unsigned char kind = kinds[i] & 0xff;
+ bool finalize = false;
+ switch (kind)
+ {
+ case GOMP_MAP_FORCE_DETACH:
+ finalize = true;
+ /* Fallthrough. */
- if (kind == GOMP_MAP_POINTER || kind == GOMP_MAP_TO_PSET)
- continue;
+ case GOMP_MAP_DETACH:
+ {
+ struct splay_tree_key_s cur_node;
+ uintptr_t hostaddr = (uintptr_t) hostaddrs[i];
+ cur_node.host_start = hostaddr;
+ cur_node.host_end = cur_node.host_start + sizeof (void *);
+ splay_tree_key n
+ = splay_tree_lookup (&acc_dev->mem_map, &cur_node);
+
+ if (n == NULL)
+ {
+ gomp_mutex_unlock (&acc_dev->lock);
+ gomp_fatal ("struct not mapped for detach operation");
+ }
- if (kind == GOMP_MAP_FORCE_ALLOC
- || kind == GOMP_MAP_FORCE_PRESENT
- || kind == GOMP_MAP_FORCE_TO
- || kind == GOMP_MAP_TO
- || kind == GOMP_MAP_ALLOC)
- {
- data_enter = true;
+ gomp_detach_pointer (acc_dev, aq, n, hostaddr, finalize, NULL);
+ }
break;
+ default:
+ ;
}
+ }
- if (kind == GOMP_MAP_RELEASE
- || kind == GOMP_MAP_DELETE
- || kind == GOMP_MAP_FROM
- || kind == GOMP_MAP_FORCE_FROM)
- break;
+ for (size_t i = 0; i < mapnum; ++i)
+ {
+ unsigned char kind = kinds[i] & 0xff;
- gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
- kind);
+ switch (kind)
+ {
+ case GOMP_MAP_FROM:
+ case GOMP_MAP_FORCE_FROM:
+ case GOMP_MAP_TO_PSET:
+ case GOMP_MAP_POINTER:
+ case GOMP_MAP_DELETE:
+ case GOMP_MAP_RELEASE:
+ {
+ struct splay_tree_key_s cur_node;
+ size_t size;
+ if (kind == GOMP_MAP_POINTER)
+ size = sizeof (void *);
+ else
+ size = sizes[i];
+ cur_node.host_start = (uintptr_t) hostaddrs[i];
+ cur_node.host_end = cur_node.host_start + size;
+ splay_tree_key n
+ = splay_tree_lookup (&acc_dev->mem_map, &cur_node);
+
+ if (n == NULL)
+ continue;
+
+ goacc_exit_datum_1 (acc_dev, hostaddrs[i], size, kind, n, aq);
+ }
+ break;
+
+ case GOMP_MAP_STRUCT:
+ case GOMP_MAP_STRUCT_UNORD:
+ /* Skip the 'GOMP_MAP_STRUCT' itself, and use the regular processing
+ for all its entries. This special handling exists for GCC 10.1
+ compatibility; afterwards, we're not generating these no-op
+ 'GOMP_MAP_STRUCT's anymore. */
+ break;
+
+ case GOMP_MAP_DETACH:
+ case GOMP_MAP_FORCE_DETACH:
+ /* OpenACC 'attach'/'detach' doesn't affect structured/dynamic
+ reference counts ('n->refcount', 'n->dynamic_refcount'). */
+ break;
+
+ default:
+ gomp_fatal (">>>> goacc_exit_data_internal UNHANDLED kind 0x%.2x",
+ kind);
+ }
}
+ gomp_mutex_unlock (&acc_dev->lock);
+}
+
+static void
+goacc_enter_exit_data_internal (int flags_m, size_t mapnum, void **hostaddrs,
+ size_t *sizes, unsigned short *kinds,
+ bool data_enter, int async, int num_waits,
+ va_list *ap)
+{
+ int flags = GOACC_FLAGS_UNMARSHAL (flags_m);
+
+ struct goacc_thread *thr;
+ struct gomp_device_descr *acc_dev;
+
+ goacc_lazy_initialize ();
+
+ thr = goacc_thread ();
+ acc_dev = thr->dev;
+
bool profiling_p = GOACC_PROFILING_DISPATCH_P (true);
acc_prof_info prof_info;
}
if (num_waits)
- {
- va_list ap;
-
- va_start (ap, num_waits);
- goacc_wait (async, num_waits, &ap);
- va_end (ap);
- }
+ goacc_wait (async, num_waits, ap);
- /* In c, non-pointers and arrays are represented by a single data clause.
- Dynamically allocated arrays and subarrays are represented by a data
- clause followed by an internal GOMP_MAP_POINTER.
-
- In fortran, scalars and not allocated arrays are represented by a
- single data clause. Allocated arrays and subarrays have three mappings:
- 1) the original data clause, 2) a PSET 3) a pointer to the array data.
- */
+ goacc_aq aq = get_goacc_asyncqueue (async);
if (data_enter)
- {
- for (i = 0; i < mapnum; i++)
- {
- unsigned char kind = kinds[i] & 0xff;
-
- /* Scan for pointers and PSETs. */
- int pointer = find_pointer (i, mapnum, kinds);
-
- if (!pointer)
- {
- switch (kind)
- {
- case GOMP_MAP_ALLOC:
- case GOMP_MAP_FORCE_ALLOC:
- acc_create_async (hostaddrs[i], sizes[i], async);
- break;
- case GOMP_MAP_TO:
- case GOMP_MAP_FORCE_TO:
- acc_copyin_async (hostaddrs[i], sizes[i], async);
- break;
- default:
- gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
- kind);
- break;
- }
- }
- else
- {
- goacc_insert_pointer (pointer, &hostaddrs[i], &sizes[i], &kinds[i],
- async);
- /* Increment 'i' by two because OpenACC requires fortran
- arrays to be contiguous, so each PSET is associated with
- one of MAP_FORCE_ALLOC/MAP_FORCE_PRESET/MAP_FORCE_TO, and
- one MAP_POINTER. */
- i += pointer - 1;
- }
- }
- }
+ goacc_enter_data_internal (acc_dev, mapnum, hostaddrs, sizes, kinds, aq);
else
- for (i = 0; i < mapnum; ++i)
- {
- unsigned char kind = kinds[i] & 0xff;
-
- int pointer = find_pointer (i, mapnum, kinds);
-
- if (!pointer)
- {
- switch (kind)
- {
- case GOMP_MAP_RELEASE:
- case GOMP_MAP_DELETE:
- if (acc_is_present (hostaddrs[i], sizes[i]))
- {
- if (finalize)
- acc_delete_finalize_async (hostaddrs[i], sizes[i], async);
- else
- acc_delete_async (hostaddrs[i], sizes[i], async);
- }
- break;
- case GOMP_MAP_FROM:
- case GOMP_MAP_FORCE_FROM:
- if (finalize)
- acc_copyout_finalize_async (hostaddrs[i], sizes[i], async);
- else
- acc_copyout_async (hostaddrs[i], sizes[i], async);
- break;
- default:
- gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
- kind);
- break;
- }
- }
- else
- {
- bool copyfrom = (kind == GOMP_MAP_FORCE_FROM
- || kind == GOMP_MAP_FROM);
- goacc_remove_pointer (hostaddrs[i], sizes[i], copyfrom, async,
- finalize, pointer);
- /* See the above comment. */
- i += pointer - 1;
- }
- }
+ goacc_exit_data_internal (acc_dev, mapnum, hostaddrs, sizes, kinds, aq);
out_prof:
if (profiling_p)
thr->api_info = NULL;
}
}
+
+/* Legacy entry point (GCC 11 and earlier). */
+
+void
+GOACC_enter_exit_data (int flags_m, size_t mapnum, void **hostaddrs,
+ size_t *sizes, unsigned short *kinds, int async,
+ int num_waits, ...)
+{
+ /* Determine if this is an OpenACC "enter data". */
+ bool data_enter = false;
+ for (size_t i = 0; i < mapnum; ++i)
+ {
+ unsigned char kind = kinds[i] & 0xff;
+
+ if (kind == GOMP_MAP_POINTER
+ || kind == GOMP_MAP_TO_PSET
+ || kind == GOMP_MAP_STRUCT
+ || kind == GOMP_MAP_STRUCT_UNORD)
+ continue;
+
+ if (kind == GOMP_MAP_FORCE_ALLOC
+ || kind == GOMP_MAP_FORCE_PRESENT
+ || kind == GOMP_MAP_ATTACH
+ || kind == GOMP_MAP_FORCE_TO
+ || kind == GOMP_MAP_TO
+ || kind == GOMP_MAP_ALLOC)
+ {
+ data_enter = true;
+ break;
+ }
+
+ if (kind == GOMP_MAP_RELEASE
+ || kind == GOMP_MAP_DELETE
+ || kind == GOMP_MAP_DETACH
+ || kind == GOMP_MAP_FORCE_DETACH
+ || kind == GOMP_MAP_FROM
+ || kind == GOMP_MAP_FORCE_FROM)
+ break;
+
+ gomp_fatal (">>>> GOACC_enter_exit_data UNHANDLED kind 0x%.2x",
+ kind);
+ }
+
+ va_list ap;
+ va_start (ap, num_waits);
+ goacc_enter_exit_data_internal (flags_m, mapnum, hostaddrs, sizes, kinds,
+ data_enter, async, num_waits, &ap);
+ va_end (ap);
+}
+
+void
+GOACC_enter_data (int flags_m, size_t mapnum, void **hostaddrs,
+ size_t *sizes, unsigned short *kinds, int async,
+ int num_waits, ...)
+{
+ va_list ap;
+ va_start (ap, num_waits);
+ goacc_enter_exit_data_internal (flags_m, mapnum, hostaddrs, sizes, kinds,
+ true, async, num_waits, &ap);
+ va_end (ap);
+}
+
+void
+GOACC_exit_data (int flags_m, size_t mapnum, void **hostaddrs,
+ size_t *sizes, unsigned short *kinds, int async,
+ int num_waits, ...)
+{
+ va_list ap;
+ va_start (ap, num_waits);
+ goacc_enter_exit_data_internal (flags_m, mapnum, hostaddrs, sizes, kinds,
+ false, async, num_waits, &ap);
+ va_end (ap);
+}
+
+void
+GOACC_declare (int flags_m, size_t mapnum,
+ void **hostaddrs, size_t *sizes, unsigned short *kinds)
+{
+ for (size_t i = 0; i < mapnum; i++)
+ {
+ unsigned char kind = kinds[i] & 0xff;
+
+ if (kind == GOMP_MAP_POINTER || kind == GOMP_MAP_TO_PSET)
+ continue;
+
+ switch (kind)
+ {
+ case GOMP_MAP_ALLOC:
+ if (acc_is_present (hostaddrs[i], sizes[i]))
+ continue;
+ /* FALLTHRU */
+ case GOMP_MAP_FORCE_ALLOC:
+ case GOMP_MAP_TO:
+ case GOMP_MAP_FORCE_TO:
+ goacc_enter_exit_data_internal (flags_m, 1, &hostaddrs[i], &sizes[i],
+ &kinds[i], true, GOMP_ASYNC_SYNC, 0, NULL);
+ break;
+
+ case GOMP_MAP_FROM:
+ case GOMP_MAP_FORCE_FROM:
+ case GOMP_MAP_RELEASE:
+ case GOMP_MAP_DELETE:
+ goacc_enter_exit_data_internal (flags_m, 1, &hostaddrs[i], &sizes[i],
+ &kinds[i], false, GOMP_ASYNC_SYNC, 0, NULL);
+ break;
+
+ case GOMP_MAP_FORCE_DEVICEPTR:
+ break;
+
+ case GOMP_MAP_FORCE_PRESENT:
+ if (!acc_is_present (hostaddrs[i], sizes[i]))
+ gomp_fatal ("[%p,%ld] is not mapped", hostaddrs[i],
+ (unsigned long) sizes[i]);
+ break;
+
+ default:
+ assert (0);
+ break;
+ }
+ }
+}