static void *worker_thread_entry(void *param)
{
work_thread_info *thread = (work_thread_info *)param;
osd_work_queue *queue = thread->queue;
// loop until we exit
for ( ;; )
{
// block waiting for work or exit
// bail on exit, and only wait if there are no pending items in queue
if (!queue->exiting && queue->list == NULL)
{
begin_timing(thread->waittime);
osd_event_wait(thread->wakeevent, INFINITE);
end_timing(thread->waittime);
}
if (queue->exiting)
break;
// indicate that we are live
atomic_exchange32(&thread->active, TRUE);
atomic_increment32(&queue->livethreads);
// process work items
for ( ;; )
{
osd_ticks_t stopspin;
// process as much as we can
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin for a while before giving up
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ && queue->list == NULL)
{
// spin for a while looking for more work
begin_timing(thread->spintime);
stopspin = osd_ticks() + SPIN_LOOP_TIME;
do {
int spin = 10000;
while (--spin && queue->list == NULL)
osd_yield_processor();
} while (queue->list == NULL && osd_ticks() < stopspin);
end_timing(thread->spintime);
}
// if nothing more, release the processor
if (queue->list == NULL)
break;
add_to_stat(&queue->spinloops, 1);
}
// decrement the live thread count
atomic_exchange32(&thread->active, FALSE);
atomic_decrement32(&queue->livethreads);
}
return NULL;
}
static unsigned __stdcall worker_thread_entry(void *param)
{
work_thread_info *thread = param;
osd_work_queue *queue = thread->queue;
// loop until we exit
for ( ;; )
{
// block waiting for work or exit
DWORD result = WAIT_OBJECT_0;
// bail on exit, and only wait if there are no pending items in queue
if (!queue->exiting && queue->list == NULL)
{
begin_timing(thread->waittime);
result = WaitForSingleObject(thread->wakeevent, INFINITE);
end_timing(thread->waittime);
}
if (queue->exiting)
break;
// indicate that we are live
interlocked_exchange32(&thread->active, TRUE);
interlocked_increment(&queue->livethreads);
// process work items
for ( ;; )
{
osd_ticks_t stopspin;
// process as much as we can
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin for a while before giving up
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ)
{
// spin for a while looking for more work
begin_timing(thread->spintime);
stopspin = osd_ticks() + SPIN_LOOP_TIME;
while (queue->list == NULL && osd_ticks() < stopspin)
YieldProcessor();
end_timing(thread->spintime);
}
// if nothing more, release the processor
if (queue->list == NULL)
break;
add_to_stat(&queue->spinloops, 1);
}
// decrement the live thread count
interlocked_exchange32(&thread->active, FALSE);
interlocked_decrement(&queue->livethreads);
}
return 0;
}
static void *worker_thread_entry(void *param)
{
work_thread_info *thread = (work_thread_info *)param;
osd_work_queue &queue = thread->queue;
// loop until we exit
for ( ;; )
{
// block waiting for work or exit
// bail on exit, and only wait if there are no pending items in queue
if (queue.exiting)
break;
if (!queue_has_list_items(&queue))
{
begin_timing(thread->waittime);
thread->wakeevent.wait( OSD_EVENT_WAIT_INFINITE);
end_timing(thread->waittime);
}
if (queue.exiting)
break;
// indicate that we are live
thread->active = TRUE;
++queue.livethreads;
// process work items
for ( ;; )
{
// process as much as we can
worker_thread_process(&queue, thread);
// if we're a high frequency queue, spin for a while before giving up
if (queue.flags & WORK_QUEUE_FLAG_HIGH_FREQ && queue.list.load() == nullptr)
{
// spin for a while looking for more work
begin_timing(thread->spintime);
spin_while<std::atomic<osd_work_item *>, osd_work_item *>(&queue.list, (osd_work_item *)nullptr, SPIN_LOOP_TIME);
end_timing(thread->spintime);
}
// if nothing more, release the processor
if (!queue_has_list_items(&queue))
break;
add_to_stat(queue.spinloops, 1);
}
// decrement the live thread count
thread->active = FALSE;
--queue.livethreads;
}
return nullptr;
}
int osd_work_queue_wait(osd_work_queue *queue, osd_ticks_t timeout)
{
// if no threads, no waiting
if (queue->threads == 0)
return TRUE;
// if no items, we're done
if (queue->items == 0)
return TRUE;
// if this is a multi queue, help out rather than doing nothing
if (queue->flags & WORK_QUEUE_FLAG_MULTI)
{
work_thread_info *thread = &queue->thread[queue->threads];
end_timing(thread->waittime);
// process what we can as a worker thread
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin until done
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ && queue->items != 0)
{
osd_ticks_t stopspin = osd_ticks() + timeout;
// spin until we're done
begin_timing(thread->spintime);
do {
int spin = 10000;
while (--spin && queue->items != 0)
osd_yield_processor();
} while (queue->items != 0 && osd_ticks() < stopspin);
end_timing(thread->spintime);
begin_timing(thread->waittime);
return (queue->items == 0);
}
begin_timing(thread->waittime);
}
// reset our done event and double-check the items before waiting
osd_event_reset(queue->doneevent);
atomic_exchange32(&queue->waiting, TRUE);
if (queue->items != 0)
osd_event_wait(queue->doneevent, timeout);
atomic_exchange32(&queue->waiting, FALSE);
// return TRUE if we actually hit 0
return (queue->items == 0);
}
int osd_work_queue_wait(osd_work_queue *queue, osd_ticks_t timeout)
{
// if no threads, no waiting
if (queue->threads == 0)
return TRUE;
// if no items, we're done
if (queue->items == 0)
return TRUE;
// if this is a multi queue, help out rather than doing nothing
if (queue->flags & WORK_QUEUE_FLAG_MULTI)
{
work_thread_info *thread = &queue->thread[queue->threads];
osd_ticks_t stopspin = osd_ticks() + timeout;
end_timing(thread->waittime);
// process what we can as a worker thread
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin until done
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ)
{
// spin until we're done
begin_timing(thread->spintime);
while (queue->items != 0 && osd_ticks() < stopspin)
YieldProcessor();
end_timing(thread->spintime);
begin_timing(thread->waittime);
return (queue->items == 0);
}
begin_timing(thread->waittime);
}
// reset our done event and double-check the items before waiting
ResetEvent(queue->doneevent);
interlocked_exchange32(&queue->waiting, TRUE);
if (queue->items != 0)
WaitForSingleObject(queue->doneevent, timeout * 1000 / osd_ticks_per_second());
interlocked_exchange32(&queue->waiting, FALSE);
// return TRUE if we actually hit 0
return (queue->items == 0);
}
int osd_work_queue_wait(osd_work_queue *queue, osd_ticks_t timeout)
{
// if no threads, no waiting
if (queue->threads == 0)
return TRUE;
// if no items, we're done
if (queue->items == 0)
return TRUE;
// if this is a multi queue, help out rather than doing nothing
if (queue->flags & WORK_QUEUE_FLAG_MULTI)
{
work_thread_info *thread = queue->thread[queue->threads];
end_timing(thread->waittime);
// process what we can as a worker thread
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin until done
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ && queue->items != 0)
{
// spin until we're done
begin_timing(thread->spintime);
spin_while_not<std::atomic<int>,int>(&queue->items, 0, timeout);
end_timing(thread->spintime);
begin_timing(thread->waittime);
return (queue->items == 0);
}
begin_timing(thread->waittime);
}
// reset our done event and double-check the items before waiting
queue->doneevent.reset();
queue->waiting = TRUE;
if (queue->items != 0)
queue->doneevent.wait(timeout);
queue->waiting = FALSE;
// return TRUE if we actually hit 0
return (queue->items == 0);
}
static void *worker_thread_entry(void *param)
{
work_thread_info *thread = (work_thread_info *)param;
osd_work_queue *queue = thread->queue;
#if defined(SDLMAME_MACOSX)
void *arp = NewAutoreleasePool();
#endif
// loop until we exit
for ( ;; )
{
// block waiting for work or exit
// bail on exit, and only wait if there are no pending items in queue
if (queue->exiting)
break;
if (!queue_has_list_items(queue))
{
begin_timing(thread->waittime);
osd_event_wait(thread->wakeevent, OSD_EVENT_WAIT_INFINITE);
end_timing(thread->waittime);
}
if (queue->exiting)
break;
// indicate that we are live
atomic_exchange32(&thread->active, TRUE);
atomic_increment32(&queue->livethreads);
// process work items
for ( ;; )
{
// process as much as we can
worker_thread_process(queue, thread);
// if we're a high frequency queue, spin for a while before giving up
if (queue->flags & WORK_QUEUE_FLAG_HIGH_FREQ && queue->list == NULL)
{
// spin for a while looking for more work
begin_timing(thread->spintime);
spin_while(&queue->list, (osd_work_item *)NULL, SPIN_LOOP_TIME);
end_timing(thread->spintime);
}
// if nothing more, release the processor
if (!queue_has_list_items(queue))
break;
add_to_stat(&queue->spinloops, 1);
}
// decrement the live thread count
atomic_exchange32(&thread->active, FALSE);
atomic_decrement32(&queue->livethreads);
}
#if defined(SDLMAME_MACOSX)
ReleaseAutoreleasePool(arp);
#endif
return NULL;
}
osd_work_item *osd_work_item_queue_multiple(osd_work_queue *queue, osd_work_callback callback, INT32 numitems, void *parambase, INT32 paramstep, UINT32 flags)
{
osd_work_item *itemlist = NULL, *lastitem = NULL;
osd_work_item **item_tailptr = &itemlist;
INT32 lockslot;
int itemnum;
// loop over items, building up a local list of work
for (itemnum = 0; itemnum < numitems; itemnum++)
{
osd_work_item *item;
// first allocate a new work item; try the free list first
INT32 lockslot = osd_scalable_lock_acquire(queue->lock);
do
{
item = (osd_work_item *)queue->free;
} while (item != NULL && compare_exchange_ptr((PVOID volatile *)&queue->free, item, item->next) != item);
osd_scalable_lock_release(queue->lock, lockslot);
// if nothing, allocate something new
if (item == NULL)
{
// allocate the item
item = (osd_work_item *)osd_malloc(sizeof(*item));
if (item == NULL)
return NULL;
item->event = NULL;
item->queue = queue;
item->done = FALSE;
}
else
{
atomic_exchange32(&item->done, FALSE); // needs to be set this way to prevent data race/usage of uninitialized memory on Linux
}
// fill in the basics
item->next = NULL;
item->callback = callback;
item->param = parambase;
item->result = NULL;
item->flags = flags;
// advance to the next
lastitem = item;
*item_tailptr = item;
item_tailptr = &item->next;
parambase = (UINT8 *)parambase + paramstep;
}
// enqueue the whole thing within the critical section
lockslot = osd_scalable_lock_acquire(queue->lock);
*queue->tailptr = itemlist;
queue->tailptr = item_tailptr;
osd_scalable_lock_release(queue->lock, lockslot);
// increment the number of items in the queue
atomic_add32(&queue->items, numitems);
add_to_stat(&queue->itemsqueued, numitems);
// look for free threads to do the work
if (queue->livethreads < queue->threads)
{
int threadnum;
// iterate over all the threads
for (threadnum = 0; threadnum < queue->threads; threadnum++)
{
work_thread_info *thread = &queue->thread[threadnum];
// if this thread is not active, wake him up
if (!thread->active)
{
osd_event_set(thread->wakeevent);
add_to_stat(&queue->setevents, 1);
// for non-shared, the first one we find is good enough
if (--numitems == 0)
break;
}
}
}
// if no threads, run the queue now on this thread
if (queue->threads == 0)
{
end_timing(queue->thread[0].waittime);
worker_thread_process(queue, &queue->thread[0]);
begin_timing(queue->thread[0].waittime);
}
// only return the item if it won't get released automatically
return (flags & WORK_ITEM_FLAG_AUTO_RELEASE) ? NULL : lastitem;
}
osd_work_item *osd_work_item_queue_multiple(osd_work_queue *queue, osd_work_callback callback, INT32 numitems, void *parambase, INT32 paramstep, UINT32 flags)
{
osd_work_item *itemlist = nullptr, *lastitem = nullptr;
osd_work_item **item_tailptr = &itemlist;
int itemnum;
// loop over items, building up a local list of work
for (itemnum = 0; itemnum < numitems; itemnum++)
{
osd_work_item *item;
// first allocate a new work item; try the free list first
{
std::lock_guard<std::mutex> lock(queue->lock);
do
{
item = (osd_work_item *)queue->free;
} while (item != nullptr && !queue->free.compare_exchange_weak(item, item->next, std::memory_order_release, std::memory_order_relaxed));
}
// if nothing, allocate something new
if (item == nullptr)
{
// allocate the item
item = new osd_work_item(*queue);
if (item == nullptr)
return nullptr;
}
else
{
item->done = FALSE; // needs to be set this way to prevent data race/usage of uninitialized memory on Linux
}
// fill in the basics
item->next = nullptr;
item->callback = callback;
item->param = parambase;
item->result = nullptr;
item->flags = flags;
// advance to the next
lastitem = item;
*item_tailptr = item;
item_tailptr = &item->next;
parambase = (UINT8 *)parambase + paramstep;
}
// enqueue the whole thing within the critical section
{
std::lock_guard<std::mutex> lock(queue->lock);
*queue->tailptr = itemlist;
queue->tailptr = item_tailptr;
}
// increment the number of items in the queue
queue->items += numitems;
add_to_stat(queue->itemsqueued, numitems);
// look for free threads to do the work
if (queue->livethreads < queue->threads)
{
int threadnum;
// iterate over all the threads
for (threadnum = 0; threadnum < queue->threads; threadnum++)
{
work_thread_info *thread = queue->thread[threadnum];
// if this thread is not active, wake him up
if (!thread->active)
{
thread->wakeevent.set();
add_to_stat(queue->setevents, 1);
// for non-shared, the first one we find is good enough
if (--numitems == 0)
break;
}
}
}
// if no threads, run the queue now on this thread
if (queue->threads == 0)
{
end_timing(queue->thread[0]->waittime);
worker_thread_process(queue, queue->thread[0]);
begin_timing(queue->thread[0]->waittime);
}
// only return the item if it won't get released automatically
return (flags & WORK_ITEM_FLAG_AUTO_RELEASE) ? nullptr : lastitem;
}
