{
if (aq->queue_n == ASYNC_QUEUE_SIZE)
{
- pthread_mutex_lock (&aq->mutex);
-
/* Queue is full. Wait for it to not be full. */
while (aq->queue_n == ASYNC_QUEUE_SIZE)
pthread_cond_wait (&aq->queue_cond_out, &aq->mutex);
-
- pthread_mutex_unlock (&aq->mutex);
}
}
{
assert (aq->agent == kernel->agent);
- wait_for_queue_nonfull (aq);
-
pthread_mutex_lock (&aq->mutex);
+ wait_for_queue_nonfull (aq);
+
int queue_last = ((aq->queue_first + aq->queue_n)
% ASYNC_QUEUE_SIZE);
if (DEBUG_QUEUES)
queue_push_callback (struct goacc_asyncqueue *aq, void (*fn)(void *),
void *data)
{
- wait_for_queue_nonfull (aq);
-
pthread_mutex_lock (&aq->mutex);
+ wait_for_queue_nonfull (aq);
+
int queue_last = ((aq->queue_first + aq->queue_n)
% ASYNC_QUEUE_SIZE);
if (DEBUG_QUEUES)
queue_push_asyncwait (struct goacc_asyncqueue *aq,
struct placeholder *placeholderp)
{
- wait_for_queue_nonfull (aq);
-
pthread_mutex_lock (&aq->mutex);
+ wait_for_queue_nonfull (aq);
+
int queue_last = ((aq->queue_first + aq->queue_n) % ASYNC_QUEUE_SIZE);
if (DEBUG_QUEUES)
HSA_DEBUG ("queue_push_asyncwait %d:%d: at %i\n", aq->agent->device_id,
{
struct placeholder *placeholderp;
- wait_for_queue_nonfull (aq);
-
pthread_mutex_lock (&aq->mutex);
+ wait_for_queue_nonfull (aq);
+
int queue_last = ((aq->queue_first + aq->queue_n) % ASYNC_QUEUE_SIZE);
if (DEBUG_QUEUES)
HSA_DEBUG ("queue_push_placeholder %d:%d: at %i\n", aq->agent->device_id,
bool
GOMP_OFFLOAD_openacc_async_host2dev (int device, void *dst, const void *src,
- size_t n, struct goacc_asyncqueue *aq)
+ size_t n, bool ephemeral,
+ struct goacc_asyncqueue *aq)
{
struct agent_info *agent = get_agent_info (device);
assert (agent == aq->agent);
- /* The source data does not necessarily remain live until the deferred
- copy happens. Taking a snapshot of the data here avoids reading
- uninitialised data later, but means that (a) data is copied twice and
- (b) modifications to the copied data between the "spawning" point of
- the asynchronous kernel and when it is executed will not be seen.
- But, that is probably correct. */
- void *src_copy = GOMP_PLUGIN_malloc (n);
- memcpy (src_copy, src, n);
- queue_push_copy (aq, dst, src_copy, n, true);
+
+ if (ephemeral)
+ {
+ /* The source data is on the stack or otherwise may be deallocated
+ before the asynchronous copy takes place. Take a copy of the source
+ data. */
+ void *src_copy = GOMP_PLUGIN_malloc (n);
+ memcpy (src_copy, src, n);
+ src = src_copy;
+ }
+ queue_push_copy (aq, dst, src, n, ephemeral);
return true;
}
}
}
-static inline void
-goacc_device_copy_async (struct gomp_device_descr *devicep,
- bool (*copy_func) (int, void *, const void *, size_t,
- struct goacc_asyncqueue *),
- const char *dst, void *dstaddr,
- const char *src, const void *srcaddr,
- size_t size, struct goacc_asyncqueue *aq)
-{
- if (!copy_func (devicep->target_id, dstaddr, srcaddr, size, aq))
- {
- gomp_mutex_unlock (&devicep->lock);
- gomp_fatal ("Copying of %s object [%p..%p) to %s object [%p..%p) failed",
- src, srcaddr, srcaddr + size, dst, dstaddr, dstaddr + size);
- }
-}
-
/* Infrastructure for coalescing adjacent or nearly adjacent (in device
addresses) host to device memory transfers. */
}
/* Copy host memory to an offload device. In asynchronous mode (if AQ is
- non-NULL), H may point to a stack location. It is up to the underlying
- plugin to ensure that this data is read immediately, rather than at some
- later point when the stack frame will likely have been destroyed. */
+ non-NULL), when the source data is stack or may otherwise be deallocated
+ before the asynchronous copy takes place, EPHEMERAL must be passed as
+ TRUE. The CBUF isn't used for non-ephemeral asynchronous copies, because
+ the host data might not be computed yet (by an earlier asynchronous compute
+ region). */
attribute_hidden void
gomp_copy_host2dev (struct gomp_device_descr *devicep,
struct goacc_asyncqueue *aq,
void *d, const void *h, size_t sz,
- struct gomp_coalesce_buf *cbuf)
+ bool ephemeral, struct gomp_coalesce_buf *cbuf)
{
if (cbuf)
{
}
}
if (__builtin_expect (aq != NULL, 0))
- goacc_device_copy_async (devicep, devicep->openacc.async.host2dev_func,
- "dev", d, "host", h, sz, aq);
+ {
+ if (!devicep->openacc.async.host2dev_func (devicep->target_id, d, h, sz,
+ ephemeral, aq))
+ {
+ gomp_mutex_unlock (&devicep->lock);
+ gomp_fatal ("Copying of host object [%p..%p) to dev object [%p..%p) "
+ "failed", h, h + sz, d, d + sz);
+ }
+ }
else
gomp_device_copy (devicep, devicep->host2dev_func, "dev", d, "host", h, sz);
}
void *h, const void *d, size_t sz)
{
if (__builtin_expect (aq != NULL, 0))
- goacc_device_copy_async (devicep, devicep->openacc.async.dev2host_func,
- "host", h, "dev", d, sz, aq);
+ {
+ if (!devicep->openacc.async.dev2host_func (devicep->target_id, h, d, sz,
+ aq))
+ {
+ gomp_mutex_unlock (&devicep->lock);
+ gomp_fatal ("Copying of dev object [%p..%p) to host object [%p..%p) "
+ "failed", d, d + sz, h, h + sz);
+ }
+ }
else
gomp_device_copy (devicep, devicep->dev2host_func, "host", h, "dev", d, sz);
}
(void *) (oldn->tgt->tgt_start + oldn->tgt_offset
+ newn->host_start - oldn->host_start),
(void *) newn->host_start,
- newn->host_end - newn->host_start, cbuf);
+ newn->host_end - newn->host_start, false, cbuf);
if (oldn->refcount != REFCOUNT_INFINITY)
oldn->refcount++;
cur_node.tgt_offset = (uintptr_t) NULL;
gomp_copy_host2dev (devicep, aq,
(void *) (tgt->tgt_start + target_offset),
- (void *) &cur_node.tgt_offset, sizeof (void *),
- cbuf);
+ (void *) &cur_node.tgt_offset,
+ sizeof (void *), true, cbuf);
return;
}
/* Add bias to the pointer value. */
to initialize the pointer with. */
cur_node.tgt_offset -= bias;
gomp_copy_host2dev (devicep, aq, (void *) (tgt->tgt_start + target_offset),
- (void *) &cur_node.tgt_offset, sizeof (void *), cbuf);
+ (void *) &cur_node.tgt_offset, sizeof (void *), true,
+ cbuf);
}
static void
(void *) (n->tgt->tgt_start + n->tgt_offset), (void *) data);
gomp_copy_host2dev (devicep, aq, (void *) devptr, (void *) &data,
- sizeof (void *), cbufp);
+ sizeof (void *), true, cbufp);
}
else
gomp_debug (1, "%s: attach count for %p -> %u\n", __FUNCTION__,
(void *) target);
gomp_copy_host2dev (devicep, aq, (void *) devptr, (void *) &target,
- sizeof (void *), cbufp);
+ sizeof (void *), true, cbufp);
}
else
gomp_debug (1, "%s: attach count for %p -> %u\n", __FUNCTION__,
for (i = first; i <= last; i++)
{
tgt->list[i].key = NULL;
- if (gomp_to_device_kind_p (get_kind (short_mapkind, kinds, i)
- & typemask))
+ if (!aq
+ && gomp_to_device_kind_p (get_kind (short_mapkind, kinds,
+ i) & typemask))
gomp_coalesce_buf_add (&cbuf,
tgt_size - cur_node.host_end
+ (uintptr_t) hostaddrs[i],
if (tgt_align < align)
tgt_align = align;
tgt_size = (tgt_size + align - 1) & ~(align - 1);
- gomp_coalesce_buf_add (&cbuf, tgt_size,
- cur_node.host_end - cur_node.host_start);
+ if (!aq)
+ gomp_coalesce_buf_add (&cbuf, tgt_size,
+ cur_node.host_end - cur_node.host_start);
tgt_size += cur_node.host_end - cur_node.host_start;
has_firstprivate = true;
continue;
if (tgt_align < align)
tgt_align = align;
tgt_size = (tgt_size + align - 1) & ~(align - 1);
- if (gomp_to_device_kind_p (kind & typemask))
+ if (!aq && gomp_to_device_kind_p (kind & typemask))
gomp_coalesce_buf_add (&cbuf, tgt_size,
cur_node.host_end - cur_node.host_start);
tgt_size += cur_node.host_end - cur_node.host_start;
len = sizes[i];
gomp_copy_host2dev (devicep, aq,
(void *) (tgt->tgt_start + tgt_size),
- (void *) hostaddrs[i], len, cbufp);
+ (void *) hostaddrs[i], len, false, cbufp);
tgt_size += len;
continue;
case GOMP_MAP_FIRSTPRIVATE_INT:
if (cur_node.tgt_offset)
cur_node.tgt_offset -= sizes[i];
gomp_copy_host2dev (devicep, aq,
- (void *) (n->tgt->tgt_start
- + n->tgt_offset
+ (void *) (n->tgt->tgt_start + n->tgt_offset
+ cur_node.host_start
- n->host_start),
(void *) &cur_node.tgt_offset,
- sizeof (void *), cbufp);
+ sizeof (void *), true, cbufp);
cur_node.tgt_offset = n->tgt->tgt_start + n->tgt_offset
+ cur_node.host_start - n->host_start;
continue;
(void *) (tgt->tgt_start
+ k->tgt_offset),
(void *) k->host_start,
- k->host_end - k->host_start, cbufp);
+ k->host_end - k->host_start, false,
+ cbufp);
break;
case GOMP_MAP_POINTER:
gomp_map_pointer (tgt, aq,
(void *) (tgt->tgt_start
+ k->tgt_offset),
(void *) k->host_start,
- k->host_end - k->host_start, cbufp);
+ k->host_end - k->host_start, false,
+ cbufp);
for (j = i + 1; j < mapnum; j++)
if (!GOMP_MAP_POINTER_P (get_kind (short_mapkind, kinds,
(void *) (tgt->tgt_start
+ k->tgt_offset),
(void *) k->host_start,
- sizeof (void *), cbufp);
+ sizeof (void *), false, cbufp);
break;
default:
gomp_mutex_unlock (&devicep->lock);
/* We intentionally do not use coalescing here, as it's not
data allocated by the current call to this function. */
gomp_copy_host2dev (devicep, aq, (void *) n->tgt_offset,
- &tgt_addr, sizeof (void *), NULL);
+ &tgt_addr, sizeof (void *), true, NULL);
}
array++;
}
(void *) tgt->tgt_start
+ k->tgt_offset,
(void *) k->host_start,
- da->data_row_size, cbufp);
+ da->data_row_size, false, cbufp);
array++;
}
target_data_rows[row_start + j] = (void *) target_row_addr;
void *ptrblock = gomp_dynamic_array_create_ptrblock
(da, target_ptrblock, target_data_rows + row_start);
gomp_copy_host2dev (devicep, aq, target_ptrblock, ptrblock,
- da->ptrblock_size, cbufp);
+ da->ptrblock_size, true, cbufp);
free (ptrblock);
}
gomp_copy_host2dev (devicep, aq,
(void *) (tgt->tgt_start + i * sizeof (void *)),
(void *) &cur_node.tgt_offset, sizeof (void *),
- cbufp);
+ true, cbufp);
}
}
(void *) (tgt->tgt_start + cbuf.chunks[c].start),
(char *) cbuf.buf + (cbuf.chunks[c].start
- cbuf.chunks[0].start),
- cbuf.chunks[c].end - cbuf.chunks[c].start, NULL);
+ cbuf.chunks[c].end - cbuf.chunks[c].start, true,
+ NULL);
free (cbuf.buf);
cbuf.buf = NULL;
cbufp = NULL;
if (GOMP_MAP_COPY_TO_P (kind & typemask))
gomp_copy_host2dev (devicep, NULL, devaddr, hostaddr, size,
- NULL);
+ false, NULL);
if (GOMP_MAP_COPY_FROM_P (kind & typemask))
gomp_copy_dev2host (devicep, NULL, hostaddr, devaddr, size);
}
projects / thirdparty / gcc.git / commitdiff
