/**
* This is the thread's main entrypoint.
* It's a simple loop:
* 1. wait for work
* 2. do work
* 3. signal that we're done
*/
static PIPE_THREAD_ROUTINE( thread_function, init_data )
{
struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
struct lp_rasterizer *rast = task->rast;
boolean debug = false;
while (1) {
/* wait for work */
if (debug)
debug_printf("thread %d waiting for work\n", task->thread_index);
pipe_semaphore_wait(&task->work_ready);
if (rast->exit_flag)
break;
if (task->thread_index == 0) {
/* thread[0]:
* - get next scene to rasterize
* - map the framebuffer surfaces
*/
lp_rast_begin( rast,
lp_scene_dequeue( rast->full_scenes, TRUE ) );
}
/* Wait for all threads to get here so that threads[1+] don't
* get a null rast->curr_scene pointer.
*/
pipe_barrier_wait( &rast->barrier );
/* do work */
if (debug)
debug_printf("thread %d doing work\n", task->thread_index);
rasterize_scene(task,
rast->curr_scene);
/* wait for all threads to finish with this scene */
pipe_barrier_wait( &rast->barrier );
/* XXX: shouldn't be necessary:
*/
if (task->thread_index == 0) {
lp_rast_end( rast );
}
/* signal done with work */
if (debug)
debug_printf("thread %d done working\n", task->thread_index);
pipe_semaphore_signal(&task->work_done);
}
return NULL;
}
void radeon_drm_ws_queue_cs(struct radeon_drm_winsys *ws, struct radeon_drm_cs *cs)
{
retry:
pipe_mutex_lock(ws->cs_stack_lock);
if (ws->ncs >= RING_LAST) {
/* no room left for a flush */
pipe_mutex_unlock(ws->cs_stack_lock);
goto retry;
}
ws->cs_stack[ws->ncs++] = cs;
pipe_mutex_unlock(ws->cs_stack_lock);
pipe_semaphore_signal(&ws->cs_queued);
}
static PIPE_THREAD_ROUTINE(util_queue_thread_func, param)
{
struct util_queue *queue = (struct util_queue*)param;
unsigned i;
while (1) {
struct util_queue_job job;
pipe_semaphore_wait(&queue->queued);
if (queue->kill_thread)
break;
pipe_mutex_lock(queue->lock);
job = queue->jobs[0];
for (i = 1; i < queue->num_jobs; i++)
queue->jobs[i - 1] = queue->jobs[i];
queue->jobs[--queue->num_jobs].job = NULL;
pipe_mutex_unlock(queue->lock);
pipe_semaphore_signal(&queue->has_space);
if (job.job) {
queue->execute_job(job.job);
pipe_semaphore_signal(&job.fence->done);
}
}
/* signal remaining jobs before terminating */
pipe_mutex_lock(queue->lock);
for (i = 0; i < queue->num_jobs; i++) {
pipe_semaphore_signal(&queue->jobs[i].fence->done);
queue->jobs[i].job = NULL;
}
queue->num_jobs = 0;
pipe_mutex_unlock(queue->lock);
return 0;
}
static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
struct radeon_cs_context *tmp;
radeon_drm_cs_sync_flush(cs);
/* Flip command streams. */
tmp = cs->csc;
cs->csc = cs->cst;
cs->cst = tmp;
/* If the CS is not empty, emit it in a separate thread. */
if (cs->base.cdw) {
unsigned i, crelocs = cs->cst->crelocs;
cs->cst->chunks[0].length_dw = cs->base.cdw;
for (i = 0; i < crelocs; i++) {
/* Update the number of active asynchronous CS ioctls for the buffer. */
p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
}
if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
cs->cst->cs.num_chunks = 3;
cs->cst->flags = RADEON_CS_KEEP_TILING_FLAGS;
} else {
cs->cst->cs.num_chunks = 2;
}
if (cs->thread &&
(flags & RADEON_FLUSH_ASYNC)) {
cs->flush_started = 1;
pipe_semaphore_signal(&cs->flush_queued);
} else {
radeon_drm_cs_emit_ioctl_oneshot(cs->cst);
}
} else {
radeon_cs_context_cleanup(cs->cst);
}
/* Prepare a new CS. */
cs->base.buf = cs->csc->buf;
cs->base.cdw = 0;
}
void
util_queue_add_job(struct util_queue *queue,
void *job,
struct util_queue_fence *fence)
{
/* Set the semaphore to "busy". */
pipe_semaphore_wait(&fence->done);
/* if the queue is full, wait until there is space */
pipe_semaphore_wait(&queue->has_space);
pipe_mutex_lock(queue->lock);
assert(queue->num_jobs < ARRAY_SIZE(queue->jobs));
queue->jobs[queue->num_jobs].job = job;
queue->jobs[queue->num_jobs].fence = fence;
queue->num_jobs++;
pipe_mutex_unlock(queue->lock);
pipe_semaphore_signal(&queue->queued);
}
static void radeon_drm_cs_destroy(struct radeon_winsys_cs *rcs)
{
struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
radeon_drm_cs_sync_flush(cs);
if (cs->thread) {
cs->kill_thread = 1;
pipe_semaphore_signal(&cs->flush_queued);
pipe_semaphore_wait(&cs->flush_completed);
pipe_thread_wait(cs->thread);
}
pipe_semaphore_destroy(&cs->flush_queued);
pipe_semaphore_destroy(&cs->flush_completed);
radeon_cs_context_cleanup(&cs->csc1);
radeon_cs_context_cleanup(&cs->csc2);
p_atomic_dec(&cs->ws->num_cs);
radeon_destroy_cs_context(&cs->csc1);
radeon_destroy_cs_context(&cs->csc2);
FREE(cs);
}
/* Shutdown:
*/
void lp_rast_destroy( struct lp_rasterizer *rast )
{
unsigned i;
/* Set exit_flag and signal each thread's work_ready semaphore.
* Each thread will be woken up, notice that the exit_flag is set and
* break out of its main loop. The thread will then exit.
*/
rast->exit_flag = TRUE;
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_signal(&rast->tasks[i].work_ready);
}
/* Wait for threads to terminate before cleaning up per-thread data.
* We don't actually call pipe_thread_wait to avoid dead lock on Windows
* per https://bugs.freedesktop.org/show_bug.cgi?id=76252 */
for (i = 0; i < rast->num_threads; i++) {
#ifdef _WIN32
pipe_semaphore_wait(&rast->tasks[i].work_done);
#else
thrd_join(rast->threads[i], NULL);
#endif
}
/* Clean up per-thread data */
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_destroy(&rast->tasks[i].work_ready);
pipe_semaphore_destroy(&rast->tasks[i].work_done);
}
for (i = 0; i < MAX2(1, rast->num_threads); i++) {
align_free(rast->tasks[i].thread_data.cache);
}
/* for synchronizing rasterization threads */
if (rast->num_threads > 0) {
pipe_barrier_destroy( &rast->barrier );
}
lp_scene_queue_destroy(rast->full_scenes);
FREE(rast);
}
/**
* Called by setup module when it has something for us to render.
*/
void
lp_rast_queue_scene( struct lp_rasterizer *rast,
struct lp_scene *scene)
{
LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
if (rast->num_threads == 0) {
/* no threading */
unsigned fpstate = util_fpstate_get();
/* Make sure that denorms are treated like zeros. This is
* the behavior required by D3D10. OpenGL doesn't care.
*/
util_fpstate_set_denorms_to_zero(fpstate);
lp_rast_begin( rast, scene );
rasterize_scene( &rast->tasks[0], scene );
lp_rast_end( rast );
util_fpstate_set(fpstate);
rast->curr_scene = NULL;
}
else {
/* threaded rendering! */
unsigned i;
lp_scene_enqueue( rast->full_scenes, scene );
/* signal the threads that there's work to do */
for (i = 0; i < rast->num_threads; i++) {
pipe_semaphore_signal(&rast->tasks[i].work_ready);
}
}
LP_DBG(DEBUG_SETUP, "%s done \n", __FUNCTION__);
}
static void radeon_winsys_destroy(struct radeon_winsys *rws)
{
struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
if (ws->thread) {
ws->kill_thread = 1;
pipe_semaphore_signal(&ws->cs_queued);
pipe_thread_wait(ws->thread);
}
pipe_semaphore_destroy(&ws->cs_queued);
pipe_mutex_destroy(ws->hyperz_owner_mutex);
pipe_mutex_destroy(ws->cmask_owner_mutex);
pipe_mutex_destroy(ws->cs_stack_lock);
ws->cman->destroy(ws->cman);
ws->kman->destroy(ws->kman);
if (ws->gen >= DRV_R600) {
radeon_surface_manager_free(ws->surf_man);
}
FREE(rws);
}
/**
* This is the thread's main entrypoint.
* It's a simple loop:
* 1. wait for work
* 2. do work
* 3. signal that we're done
*/
static int
thread_function(void *init_data)
{
struct lp_rasterizer_task *task = (struct lp_rasterizer_task *) init_data;
struct lp_rasterizer *rast = task->rast;
boolean debug = false;
char thread_name[16];
unsigned fpstate;
util_snprintf(thread_name, sizeof thread_name, "llvmpipe-%u", task->thread_index);
u_thread_setname(thread_name);
/* Make sure that denorms are treated like zeros. This is
* the behavior required by D3D10. OpenGL doesn't care.
*/
fpstate = util_fpstate_get();
util_fpstate_set_denorms_to_zero(fpstate);
while (1) {
/* wait for work */
if (debug)
debug_printf("thread %d waiting for work\n", task->thread_index);
pipe_semaphore_wait(&task->work_ready);
if (rast->exit_flag)
break;
if (task->thread_index == 0) {
/* thread[0]:
* - get next scene to rasterize
* - map the framebuffer surfaces
*/
lp_rast_begin( rast,
lp_scene_dequeue( rast->full_scenes, TRUE ) );
}
/* Wait for all threads to get here so that threads[1+] don't
* get a null rast->curr_scene pointer.
*/
pipe_barrier_wait( &rast->barrier );
/* do work */
if (debug)
debug_printf("thread %d doing work\n", task->thread_index);
rasterize_scene(task,
rast->curr_scene);
/* wait for all threads to finish with this scene */
pipe_barrier_wait( &rast->barrier );
/* XXX: shouldn't be necessary:
*/
if (task->thread_index == 0) {
lp_rast_end( rast );
}
/* signal done with work */
if (debug)
debug_printf("thread %d done working\n", task->thread_index);
pipe_semaphore_signal(&task->work_done);
}
#ifdef _WIN32
pipe_semaphore_signal(&task->work_done);
#endif
return 0;
}