Differential D5371 Diff 16686 source/blender/blenlib/intern/task.c

Changeset View

Standalone View

source/blender/blenlib/intern/task.c

Show First 20 Lines • Show All 1,048 Lines • ▼ Show 20 Lines	typedef struct ParallelRangeState {
void *userdata;		void *userdata;

TaskParallelRangeFunc func;		TaskParallelRangeFunc func;

int iter;		int iter;
int chunk_size;		int chunk_size;
} ParallelRangeState;		} ParallelRangeState;

BLI_INLINE void task_parallel_range_calc_chunk_size(const TaskParallelSettings *settings,		BLI_INLINE void task_parallel_calc_chunk_size(const TaskParallelSettings *settings,
const int num_tasks,		const int tot_items,
ParallelRangeState *state)		int num_tasks,
		int *r_chunk_size)
{		{
const int tot_items = state->stop - state->start;
int chunk_size = 0;		int chunk_size = 0;

if (settings->min_iter_per_thread > 0) {		if (!settings->use_threading) {
		/* Some users of this helper will still need a valid chunk size in case processing is not
		* threaded... We can use a bigger one than in default threaded case then. */
		sergeyUnsubmitted Done Inline Actions Ellipsis -> fullstop, sergey: Ellipsis -> fullstop,
		chunk_size = 1024;
		num_tasks = 1;
		}
		else if (settings->min_iter_per_thread > 0) {
/* Already set by user, no need to do anything here. */		/* Already set by user, no need to do anything here. */
chunk_size = settings->min_iter_per_thread;		chunk_size = settings->min_iter_per_thread;
}		}
else {		else {
/* Basic heuristic to avoid threading on low amount of items. We could make that limit		/* Basic heuristic to avoid threading on low amount of items. We could make that limit
* configurable in settings too... */		* configurable in settings too... */
if (tot_items > 0 && tot_items < 256) {		if (tot_items > 0 && tot_items < 256) {
chunk_size = tot_items;		chunk_size = tot_items;
Show All 9 Lines	else {
}		}
else {		else {
chunk_size = 32;		chunk_size = 32;
}		}
}		}

BLI_assert(chunk_size > 0);		BLI_assert(chunk_size > 0);

		if (tot_items > 0) {
switch (settings->scheduling_mode) {		switch (settings->scheduling_mode) {
case TASK_SCHEDULING_STATIC:		case TASK_SCHEDULING_STATIC:
state->chunk_size = max_ii(chunk_size, tot_items / (num_tasks));		*r_chunk_size = max_ii(chunk_size, tot_items / num_tasks);
break;		break;
case TASK_SCHEDULING_DYNAMIC:		case TASK_SCHEDULING_DYNAMIC:
state->chunk_size = chunk_size;		*r_chunk_size = chunk_size;
break;		break;
}		}
}		}
		else {
		/* If total amount of items is unknown, we can only use dynamic scheduling. */
		*r_chunk_size = chunk_size;
		}
		}

		BLI_INLINE void task_parallel_range_calc_chunk_size(const TaskParallelSettings *settings,
		const int num_tasks,
		ParallelRangeState *state)
		{
		task_parallel_calc_chunk_size(
		settings, state->stop - state->start, num_tasks, &state->chunk_size);
		}

BLI_INLINE bool parallel_range_next_iter_get(ParallelRangeState *__restrict state,		BLI_INLINE bool parallel_range_next_iter_get(ParallelRangeState *__restrict state,
int *__restrict iter,		int *__restrict iter,
int *__restrict count)		int *__restrict count)
{		{
int previter = atomic_fetch_and_add_int32(&state->iter, state->chunk_size);		int previter = atomic_fetch_and_add_int32(&state->iter, state->chunk_size);

*iter = previter;		*iter = previter;
▲ Show 20 Lines • Show All 140 Lines • ▼ Show 20 Lines	if (settings->func_finalize != NULL) {
userdata_chunk_local = (char )userdata_chunk_array + (userdata_chunk_size i);		userdata_chunk_local = (char )userdata_chunk_array + (userdata_chunk_size i);
settings->func_finalize(userdata, userdata_chunk_local);		settings->func_finalize(userdata, userdata_chunk_local);
}		}
}		}
MALLOCA_FREE(userdata_chunk_array, userdata_chunk_size * num_tasks);		MALLOCA_FREE(userdata_chunk_array, userdata_chunk_size * num_tasks);
}		}
}		}

#undef MALLOCA		typedef struct TaskParallelIteratorState {
#undef MALLOCA_FREE

typedef struct ParallelListbaseState {
void *userdata;		void *userdata;
TaskParallelListbaseFunc func;		TaskParallelIteratorIterFunc iter_func;
		TaskParallelIteratorFunc func;

int chunk_size;		/* * Data used to 'acquire' chunks of items from the iterator. * */
int index;		/* Common data also passed to the generator callback. */
Link *link;		TaskParallelIteratorStateShared iter_shared;
SpinLock lock;		/* Total number of items. If unknown, set it to a negative number. */
} ParallelListState;		int tot_items;
		} TaskParallelIteratorState;

		/* Per-thread specific data passed to the callback. */
		typedef struct TaskParallelIteratorTLS {
		/* 'Public' data passed on to the callback. */
		TaskParallelTLS tls;
		sergeyUnsubmitted Done Inline Actions Maybe `user_tls` ? sergey: Maybe `user_tls` ?

		/* Private data to each thread, used to store chunks of items from the generator callback. */
		TaskParallelIteratorStateTLS iter_tls;
		} TaskParallelIteratorTLS;

BLI_INLINE Link parallel_listbase_next_iter_get(ParallelListState __restrict state,		BLI_INLINE void task_parallel_iterator_calc_chunk_size(const TaskParallelSettings *settings,
int *__restrict index,		const int num_tasks,
int *__restrict count)		TaskParallelIteratorState *state)
{		{
int task_count = 0;		task_parallel_calc_chunk_size(
BLI_spin_lock(&state->lock);		settings, state->tot_items, num_tasks, &state->iter_shared.chunk_size);
Link *result = state->link;		}
if (LIKELY(result != NULL)) {
*index = state->index;
while (state->link != NULL && task_count < state->chunk_size) {
++task_count;
state->link = state->link->next;
}
state->index += task_count;
}
BLI_spin_unlock(&state->lock);
*count = task_count;
return result;
}

static void parallel_listbase_func(TaskPool *__restrict pool,
void *UNUSED(taskdata),
int UNUSED(threadid))
{
ParallelListState *__restrict state = BLI_task_pool_userdata(pool);
Link *link;
int index, count;

while ((link = parallel_listbase_next_iter_get(state, &index, &count)) != NULL) {		static void parallel_iterator_func_do(TaskParallelIteratorState *__restrict state,
for (int i = 0; i < count; ++i) {		void *userdata_chunk,
state->func(state->userdata, link, index + i);		int threadid)
link = link->next;		{
		TaskParallelIteratorTLS tls = {
		.tls.thread_id = threadid,
		.tls.userdata_chunk = userdata_chunk,
		.iter_tls.current_chunk_size = 0,
		};

		const size_t items_size = sizeof((tls.iter_tls.current_chunk_items))
		(size_t)state->iter_shared.chunk_size;
		const size_t indices_size = sizeof((tls.iter_tls.current_chunk_indices))
		(size_t)state->iter_shared.chunk_size;
		tls.iter_tls.current_chunk_items = MALLOCA(items_size);
		tls.iter_tls.current_chunk_indices = MALLOCA(indices_size);

		do {
		state->iter_func(state->userdata, &tls.tls, &state->iter_shared, &tls.iter_tls);
		for (int i = 0; i < tls.iter_tls.current_chunk_size; ++i) {
		state->func(state->userdata,
		tls.iter_tls.current_chunk_items[i],
		tls.iter_tls.current_chunk_indices[i],
		&tls.tls);
}		}
		} while (tls.iter_tls.current_chunk_size != 0);

		MALLOCA_FREE(tls.iter_tls.current_chunk_items, items_size);
		MALLOCA_FREE(tls.iter_tls.current_chunk_indices, indices_size);
}		}

		static void parallel_iterator_func(TaskPool __restrict pool, void userdata_chunk, int threadid)
		{
		TaskParallelIteratorState *__restrict state = BLI_task_pool_userdata(pool);

		parallel_iterator_func_do(state, userdata_chunk, threadid);
}		}

static void task_parallel_listbase_no_threads(struct ListBase *listbase,		static void task_parallel_iterator_no_threads(const TaskParallelSettings *settings,
void *userdata,		TaskParallelIteratorState *state)
TaskParallelListbaseFunc func)
{		{
int i = 0;		/* Prepare user's TLS data. */
for (Link *link = listbase->first; link != NULL; link = link->next, ++i) {		void *userdata_chunk = settings->userdata_chunk;
func(userdata, link, i);		const size_t userdata_chunk_size = settings->userdata_chunk_size;
		void *userdata_chunk_local = NULL;
		const bool use_userdata_chunk = (userdata_chunk_size != 0) && (userdata_chunk != NULL);
		if (use_userdata_chunk) {
		userdata_chunk_local = MALLOCA(userdata_chunk_size);
		memcpy(userdata_chunk_local, userdata_chunk, userdata_chunk_size);
		}

		/* Also marking it as non-threaded for the iterator callback. */
		state->iter_shared.spin_lock = NULL;

		parallel_iterator_func_do(state, userdata_chunk, 0);

		if (use_userdata_chunk) {
		if (settings->func_finalize != NULL) {
		settings->func_finalize(state->userdata, userdata_chunk_local);
		}
		MALLOCA_FREE(userdata_chunk_local, userdata_chunk_size);
}		}
}		}

/* NOTE: The idea here is to compensate for rather measurable threading		static void task_parallel_iterator_do(const TaskParallelSettings *settings,
* overhead caused by fetching tasks. With too many CPU threads we are starting		TaskParallelIteratorState *state)
* to spend too much time in those overheads. */
BLI_INLINE int task_parallel_listbasecalc_chunk_size(const int num_threads)
{		{
if (num_threads > 32) {		TaskScheduler *task_scheduler = BLI_task_scheduler_get();
return 128;		const int num_threads = BLI_task_scheduler_num_threads(task_scheduler);

		task_parallel_iterator_calc_chunk_size(settings, num_threads, state);

		if (!settings->use_threading) {
		task_parallel_iterator_no_threads(settings, state);
		return;
}		}
else if (num_threads > 16) {
return 64;		const int chunk_size = state->iter_shared.chunk_size;
		const int tot_items = state->tot_items;
		const size_t num_tasks = tot_items >= 0 ?
		(size_t)min_ii(num_threads, state->tot_items / chunk_size) :
		(size_t)num_threads;

		BLI_assert(num_tasks > 0);
		if (num_tasks == 1) {
		task_parallel_iterator_no_threads(settings, state);
		return;
}		}
return 32;
		SpinLock spin_lock;
		BLI_spin_init(&spin_lock);
		state->iter_shared.spin_lock = &spin_lock;

		void *userdata_chunk = settings->userdata_chunk;
		const size_t userdata_chunk_size = settings->userdata_chunk_size;
		void *userdata_chunk_local = NULL;
		void *userdata_chunk_array = NULL;
		const bool use_userdata_chunk = (userdata_chunk_size != 0) && (userdata_chunk != NULL);

		TaskPool *task_pool = BLI_task_pool_create_suspended(task_scheduler, state);

		if (use_userdata_chunk) {
		userdata_chunk_array = MALLOCA(userdata_chunk_size * num_tasks);
		}

		for (size_t i = 0; i < num_tasks; i++) {
		if (use_userdata_chunk) {
		userdata_chunk_local = (char )userdata_chunk_array + (userdata_chunk_size i);
		memcpy(userdata_chunk_local, userdata_chunk, userdata_chunk_size);
		}
		/* Use this pool's pre-allocated tasks. */
		BLI_task_pool_push_from_thread(task_pool,
		parallel_iterator_func,
		userdata_chunk_local,
		false,
		TASK_PRIORITY_HIGH,
		task_pool->thread_id);
		}

		BLI_task_pool_work_and_wait(task_pool);
		BLI_task_pool_free(task_pool);

		if (use_userdata_chunk) {
		if (settings->func_finalize != NULL) {
		for (size_t i = 0; i < num_tasks; i++) {
		userdata_chunk_local = (char )userdata_chunk_array + (userdata_chunk_size i);
		settings->func_finalize(state->userdata, userdata_chunk_local);
		}
		}
		MALLOCA_FREE(userdata_chunk_array, userdata_chunk_size * num_tasks);
		}

		BLI_spin_end(&spin_lock);
		state->iter_shared.spin_lock = NULL;
		}

		/**
		* This function allows to parallelize for loops using a generic iterator.
		*
		* \param userdata: Common userdata passed to all instances of \a func.
		* \param iter_func: Callback function used to generate chunks of items.
		* \param init_item: The initial item, if necessary (may be NULL if unused).
		* \param init_index: The initial index.
		* \param tot_items: The total amount of items to iterate over
		* (if unkown, set it to a negative number).
		* \param func: Callback function.
		* \param settings: See public API doc of TaskParallelSettings for description of all settings.
		*
		* \note Static scheduling is only available when \a tot_items is >= 0.
		*/

		void BLI_task_parallel_iterator(void *userdata,
		TaskParallelIteratorIterFunc iter_func,
		void *init_item,
		const int init_index,
		const int tot_items,
		TaskParallelIteratorFunc func,
		const TaskParallelSettings *settings)
		{
		TaskParallelIteratorState state = {0};

		state.tot_items = tot_items;
		state.iter_shared.next_index = init_index;
		state.iter_shared.next_item = init_item;
		state.userdata = userdata;
		state.iter_func = iter_func;
		state.func = func;

		task_parallel_iterator_do(settings, &state);
		}

		static void task_parallel_listbase_get(void *__restrict UNUSED(userdata),
		const TaskParallelTLS *__restrict UNUSED(tls),
		TaskParallelIteratorStateShared *__restrict state_shared,
		TaskParallelIteratorStateTLS *__restrict state_tls)
		{
		int num_items = 0;

		if (state_shared->spin_lock != NULL) {
		BLI_spin_lock(state_shared->spin_lock);
		}

		/* Get current status. */
		int index = state_shared->next_index;
		Link *link = state_shared->next_item;

		/* 'Acquire' a chunk of items from the list. */
		for (int i = 0; link != NULL && num_items < state_shared->chunk_size;
		link = link->next, num_items++, index++, i++) {
		state_tls->current_chunk_items[i] = link;
		state_tls->current_chunk_indices[i] = index;
		}

		/* Update current status. */
		state_shared->next_index += num_items;
		state_shared->next_item = link;

		if (state_shared->spin_lock != NULL) {
		BLI_spin_unlock(state_shared->spin_lock);
		}

		state_tls->current_chunk_size = num_items;
}		}

/**		/**
* This function allows to parallelize for loops over ListBase items.		* This function allows to parallelize for loops over ListBase items.
*		*
* \param listbase: The double linked list to loop over.		* \param listbase: The double linked list to loop over.
* \param userdata: Common userdata passed to all instances of \a func.		* \param userdata: Common userdata passed to all instances of \a func.
* \param func: Callback function.		* \param func: Callback function.
* \param use_threading: If \a true, actually split-execute loop in threads,		* \param settings: See public API doc of ParallelRangeSettings for description of all settings.
* else just do a sequential forloop
* (allows caller to use any kind of test to switch on parallelization or not).
*		*
* \note There is no static scheduling here,		* \note There is no static scheduling here,
* since it would need another full loop over items to count them.		* since it would need another full loop over items to count them.
*/		*/
void BLI_task_parallel_listbase(struct ListBase *listbase,		void BLI_task_parallel_listbase(ListBase *listbase,
void *userdata,		void *userdata,
TaskParallelListbaseFunc func,		TaskParallelIteratorFunc func,
const bool use_threading)		const TaskParallelSettings *settings)
{		{
if (BLI_listbase_is_empty(listbase)) {		if (BLI_listbase_is_empty(listbase)) {
return;		return;
}		}
if (!use_threading) {
task_parallel_listbase_no_threads(listbase, userdata, func);
return;
}
TaskScheduler *task_scheduler = BLI_task_scheduler_get();
const int num_threads = BLI_task_scheduler_num_threads(task_scheduler);
/* TODO(sergey): Consider making chunk size configurable. */
const int chunk_size = task_parallel_listbasecalc_chunk_size(num_threads);
const int num_tasks = min_ii(num_threads, BLI_listbase_count(listbase) / chunk_size);
if (num_tasks <= 1) {
task_parallel_listbase_no_threads(listbase, userdata, func);
return;
}

ParallelListState state;		TaskParallelIteratorState state = {0};
TaskPool *task_pool = BLI_task_pool_create_suspended(task_scheduler, &state);

state.index = 0;		state.tot_items = BLI_listbase_count(listbase);
state.link = listbase->first;		state.iter_shared.next_index = 0;
		state.iter_shared.next_item = listbase->first;
state.userdata = userdata;		state.userdata = userdata;
		state.iter_func = task_parallel_listbase_get;
state.func = func;		state.func = func;
state.chunk_size = chunk_size;
BLI_spin_init(&state.lock);

BLI_assert(num_tasks > 0);		task_parallel_iterator_do(settings, &state);
for (int i = 0; i < num_tasks; i++) {
/* Use this pool's pre-allocated tasks. */
BLI_task_pool_push_from_thread(
task_pool, parallel_listbase_func, NULL, false, TASK_PRIORITY_HIGH, task_pool->thread_id);
}		}

BLI_task_pool_work_and_wait(task_pool);		#undef MALLOCA
BLI_task_pool_free(task_pool);		#undef MALLOCA_FREE

BLI_spin_end(&state.lock);
}

typedef struct ParallelMempoolState {		typedef struct ParallelMempoolState {
void *userdata;		void *userdata;
TaskParallelMempoolFunc func;		TaskParallelMempoolFunc func;
} ParallelMempoolState;		} ParallelMempoolState;

static void parallel_mempool_func(TaskPool __restrict pool, void taskdata, int UNUSED(threadid))		static void parallel_mempool_func(TaskPool __restrict pool, void taskdata, int UNUSED(threadid))
{		{
▲ Show 20 Lines • Show All 77 Lines • Show Last 20 Lines