compositor: perform conversions in parallel

Improves throughput of the total convert and blend process and allows
for higher performance across slightly more threads.

Also make use of video aggregator's task pool for blending as well in
order to reduce the number of threads.

Part-of: <https://gitlab.freedesktop.org/gstreamer/gst-plugins-base/-/merge_requests/1129>

docs/plugins/gst_plugins_cache.json

@@ -1926,6 +1926,7 @@
             "GstCompositorPad": {
                 "hierarchy": [
                     "GstCompositorPad",
+                    "GstVideoAggregatorParallelConvertPad",
                     "GstVideoAggregatorConvertPad",
                     "GstVideoAggregatorPad",
                     "GstAggregatorPad",

gst/compositor/compositor.c

@@ -183,7 +183,7 @@ enum
 };
 
 G_DEFINE_TYPE (GstCompositorPad, gst_compositor_pad,
-    GST_TYPE_VIDEO_AGGREGATOR_CONVERT_PAD);
+    GST_TYPE_VIDEO_AGGREGATOR_PARALLEL_CONVERT_PAD);
 
 static void
 gst_compositor_pad_get_property (GObject * object, guint prop_id,
@@ -386,8 +386,8 @@ _pad_obscures_rectangle (GstVideoAggregator * vagg, GstVideoAggregatorPad * pad,
   return TRUE;
 }
 
-static gboolean
-gst_compositor_pad_prepare_frame (GstVideoAggregatorPad * pad,
+static void
+gst_compositor_pad_prepare_frame_start (GstVideoAggregatorPad * pad,
     GstVideoAggregator * vagg, GstBuffer * buffer,
     GstVideoFrame * prepared_frame)
 {
@@ -417,7 +417,7 @@ gst_compositor_pad_prepare_frame (GstVideoAggregatorPad * pad,
 
   if (cpad->alpha == 0.0) {
     GST_DEBUG_OBJECT (pad, "Pad has alpha 0.0, not converting frame");
-    goto done;
+    return;
   }
 
   frame_rect = clamp_rectangle (cpad->xpos, cpad->ypos, width, height,
@@ -426,7 +426,7 @@ gst_compositor_pad_prepare_frame (GstVideoAggregatorPad * pad,
   if (frame_rect.w == 0 || frame_rect.h == 0) {
     GST_DEBUG_OBJECT (pad, "Resulting frame is zero-width or zero-height "
         "(w: %i, h: %i), skipping", frame_rect.w, frame_rect.h);
-    goto done;
+    return;
   }
 
   GST_OBJECT_LOCK (vagg);
@@ -446,16 +446,11 @@ gst_compositor_pad_prepare_frame (GstVideoAggregatorPad * pad,
   GST_OBJECT_UNLOCK (vagg);
 
   if (frame_obscured)
-    goto done;
+    return;
 
-  return
-      GST_VIDEO_AGGREGATOR_PAD_CLASS
-      (gst_compositor_pad_parent_class)->prepare_frame (pad, vagg, buffer,
+  GST_VIDEO_AGGREGATOR_PAD_CLASS
+      (gst_compositor_pad_parent_class)->prepare_frame_start (pad, vagg, buffer,
       prepared_frame);
-
-done:
-
-  return TRUE;
 }
 
 static void
@@ -539,8 +534,8 @@ gst_compositor_pad_class_init (GstCompositorPadClass * klass)
           GST_TYPE_COMPOSITOR_OPERATOR, DEFAULT_PAD_OPERATOR,
           G_PARAM_READWRITE | GST_PARAM_CONTROLLABLE | G_PARAM_STATIC_STRINGS));
 
-  vaggpadclass->prepare_frame =
-      GST_DEBUG_FUNCPTR (gst_compositor_pad_prepare_frame);
+  vaggpadclass->prepare_frame_start =
+      GST_DEBUG_FUNCPTR (gst_compositor_pad_prepare_frame_start);
 
   vaggcpadclass->create_conversion_info =
       GST_DEBUG_FUNCPTR (gst_compositor_pad_create_conversion_info);
@@ -861,123 +856,105 @@ _fixate_caps (GstAggregator * agg, GstCaps * caps)
   return ret;
 }
 
-static gpointer
+static void
 gst_parallelized_task_thread_func (gpointer data)
 {
-  GstParallelizedTaskThread *self = data;
+  GstParallelizedTaskRunner *runner = data;
+  gint idx;
 
-  g_mutex_lock (&self->runner->lock);
-  self->runner->n_done++;
-  if (self->runner->n_done == self->runner->n_threads - 1)
-    g_cond_signal (&self->runner->cond_done);
+  g_mutex_lock (&runner->lock);
+  idx = runner->n_todo--;
+  g_assert (runner->n_todo >= -1);
+  g_mutex_unlock (&runner->lock);
 
-  do {
-    gint idx;
+  g_assert (runner->func != NULL);
 
-    while (self->runner->n_todo == -1 && !self->runner->quit)
-      g_cond_wait (&self->runner->cond_todo, &self->runner->lock);
+  runner->func (runner->task_data[idx]);
+}
 
-    if (self->runner->quit)
-      break;
+static void
+gst_parallelized_task_runner_join (GstParallelizedTaskRunner * self)
+{
+  gboolean joined = FALSE;
 
-    idx = self->runner->n_todo--;
-    g_assert (self->runner->n_todo >= -1);
-    g_mutex_unlock (&self->runner->lock);
-
-    g_assert (self->runner->func != NULL);
-
-    self->runner->func (self->runner->task_data[idx]);
-
-    g_mutex_lock (&self->runner->lock);
-    self->runner->n_done++;
-    if (self->runner->n_done == self->runner->n_threads - 1)
-      g_cond_signal (&self->runner->cond_done);
-  } while (TRUE);
-
-  g_mutex_unlock (&self->runner->lock);
-
-  return NULL;
+  while (!joined) {
+    g_mutex_lock (&self->lock);
+    if (!(joined = gst_queue_array_is_empty (self->tasks))) {
+      gpointer task = gst_queue_array_pop_head (self->tasks);
+      g_mutex_unlock (&self->lock);
+      gst_task_pool_join (self->pool, task);
+    } else {
+      g_mutex_unlock (&self->lock);
+    }
+  }
 }
 
 static void
 gst_parallelized_task_runner_free (GstParallelizedTaskRunner * self)
 {
-  guint i;
-
-  g_mutex_lock (&self->lock);
-  self->quit = TRUE;
-  g_cond_broadcast (&self->cond_todo);
-  g_mutex_unlock (&self->lock);
-
-  for (i = 1; i < self->n_threads; i++) {
-    if (!self->threads[i].thread)
-      continue;
-
-    g_thread_join (self->threads[i].thread);
-  }
+  gst_parallelized_task_runner_join (self);
 
+  gst_queue_array_free (self->tasks);
+  if (self->own_pool)
+    gst_task_pool_cleanup (self->pool);
+  gst_object_unref (self->pool);
   g_mutex_clear (&self->lock);
-  g_cond_clear (&self->cond_todo);
-  g_cond_clear (&self->cond_done);
-  g_free (self->threads);
   g_free (self);
 }
 
 static GstParallelizedTaskRunner *
-gst_parallelized_task_runner_new (guint n_threads)
+gst_parallelized_task_runner_new (guint n_threads, GstTaskPool * pool,
+    gboolean async_tasks)
 {
   GstParallelizedTaskRunner *self;
-  guint i;
-  GError *err = NULL;
 
   if (n_threads == 0)
     n_threads = g_get_num_processors ();
 
   self = g_new0 (GstParallelizedTaskRunner, 1);
-  self->n_threads = n_threads;
-  self->threads = g_new0 (GstParallelizedTaskThread, n_threads);
 
-  self->quit = FALSE;
+  if (pool) {
+    self->pool = g_object_ref (pool);
+    self->own_pool = FALSE;
+
+    /* No reason to split up the work between more threads than the
+     * pool can spawn */
+    if (GST_IS_SHARED_TASK_POOL (pool))
+      n_threads =
+          MIN (n_threads,
+          gst_shared_task_pool_get_max_threads (GST_SHARED_TASK_POOL (pool)));
+  } else {
+    self->pool = gst_shared_task_pool_new ();
+    self->own_pool = TRUE;
+    gst_shared_task_pool_set_max_threads (GST_SHARED_TASK_POOL (self->pool),
+        n_threads);
+    gst_task_pool_prepare (self->pool, NULL);
+  }
+
+  self->tasks = gst_queue_array_new (n_threads);
+
+  self->n_threads = n_threads;
+
   self->n_todo = -1;
-  self->n_done = 0;
   g_mutex_init (&self->lock);
-  g_cond_init (&self->cond_todo);
-  g_cond_init (&self->cond_done);
 
   /* Set when scheduling a job */
   self->func = NULL;
   self->task_data = NULL;
-
-  for (i = 0; i < n_threads; i++) {
-    self->threads[i].runner = self;
-    self->threads[i].idx = i;
-
-    /* First thread is the one calling run() */
-    if (i > 0) {
-      self->threads[i].thread =
-          g_thread_try_new ("compositor-blend",
-          gst_parallelized_task_thread_func, &self->threads[i], &err);
-      if (!self->threads[i].thread)
-        goto error;
-    }
-  }
-
-  g_mutex_lock (&self->lock);
-  while (self->n_done < self->n_threads - 1)
-    g_cond_wait (&self->cond_done, &self->lock);
-  self->n_done = 0;
-  g_mutex_unlock (&self->lock);
+  self->async_tasks = async_tasks;
 
   return self;
+}
 
-error:
-  {
-    GST_ERROR ("Failed to start thread %u: %s", i, err->message);
-    g_clear_error (&err);
+static void
+gst_parallelized_task_runner_finish (GstParallelizedTaskRunner * self)
+{
+  g_return_if_fail (self->func != NULL);
 
-    gst_parallelized_task_runner_free (self);
-    return NULL;
-  }
+  gst_parallelized_task_runner_join (self);
+
+  self->func = NULL;
+  self->task_data = NULL;
 }
 
 static void
@@ -989,32 +966,43 @@ gst_parallelized_task_runner_run (GstParallelizedTaskRunner * self,
   self->func = func;
   self->task_data = task_data;
 
-  if (n_threads > 1) {
+  if (n_threads > 1 || self->async_tasks) {
+    guint i = 0;
     g_mutex_lock (&self->lock);
-    self->n_todo = self->n_threads - 2;
-    self->n_done = 0;
-    g_cond_broadcast (&self->cond_todo);
+    self->n_todo = self->n_threads - 1;
+    if (!self->async_tasks) {
+      /* if not async, perform one of the functions in the current thread */
+      self->n_todo--;
+      i = 1;
+    }
+    for (; i < n_threads; i++) {
+      gpointer task =
+          gst_task_pool_push (self->pool, gst_parallelized_task_thread_func,
+          self, NULL);
+
+      /* The return value of push() is nullable but NULL is only returned
+       * with the shared task pool when gst_task_pool_prepare() has not been
+       * called and would thus be a programming error that we should hard-fail
+       * on.
+       */
+      g_assert (task != NULL);
+      gst_queue_array_push_tail (self->tasks, task);
+    }
     g_mutex_unlock (&self->lock);
   }
 
-  self->func (self->task_data[self->n_threads - 1]);
+  if (!self->async_tasks) {
+    self->func (self->task_data[self->n_threads - 1]);
 
-  if (n_threads > 1) {
-    g_mutex_lock (&self->lock);
-    while (self->n_done < self->n_threads - 1)
-      g_cond_wait (&self->cond_done, &self->lock);
-    self->n_done = 0;
-    g_mutex_unlock (&self->lock);
+    gst_parallelized_task_runner_finish (self);
   }
-
-  self->func = NULL;
-  self->task_data = NULL;
 }
 
 static gboolean
 _negotiated_caps (GstAggregator * agg, GstCaps * caps)
 {
   GstCompositor *compositor = GST_COMPOSITOR (agg);
+  GstVideoAggregator *vagg = GST_VIDEO_AGGREGATOR (agg);
   GstVideoInfo v_info;
   guint n_threads;
 
@@ -1041,8 +1029,12 @@ _negotiated_caps (GstAggregator * agg, GstCaps * caps)
     gst_parallelized_task_runner_free (compositor->blend_runner);
     compositor->blend_runner = NULL;
   }
-  if (!compositor->blend_runner)
-    compositor->blend_runner = gst_parallelized_task_runner_new (n_threads);
+  if (!compositor->blend_runner) {
+    GstTaskPool *pool = gst_video_aggregator_get_execution_task_pool (vagg);
+    compositor->blend_runner =
+        gst_parallelized_task_runner_new (n_threads, pool, FALSE);
+    gst_clear_object (&pool);
+  }
 
   return GST_AGGREGATOR_CLASS (parent_class)->negotiated_src_caps (agg, caps);
 }

gst/compositor/compositor.h

@@ -24,6 +24,7 @@
 #include <gst/gst.h>
 #include <gst/video/video.h>
 #include <gst/video/gstvideoaggregator.h>
+#include <gst/base/base.h>
 
 #include "blend.h"
 
@@ -35,7 +36,7 @@ G_DECLARE_FINAL_TYPE (GstCompositor, gst_compositor, GST, COMPOSITOR,
 
 #define GST_TYPE_COMPOSITOR_PAD (gst_compositor_pad_get_type())
 G_DECLARE_FINAL_TYPE (GstCompositorPad, gst_compositor_pad, GST, COMPOSITOR_PAD,
-    GstVideoAggregatorConvertPad)
+    GstVideoAggregatorParallelConvertPad)
 
 /**
  * GstCompositorBackground:
@@ -80,28 +81,22 @@ typedef enum
 typedef void (*GstParallelizedTaskFunc) (gpointer user_data);
 
 typedef struct _GstParallelizedTaskRunner GstParallelizedTaskRunner;
-typedef struct _GstParallelizedTaskThread GstParallelizedTaskThread;
-
-struct _GstParallelizedTaskThread
-{
-  GstParallelizedTaskRunner *runner;
-  guint idx;
-  GThread *thread;
-};
 
 struct _GstParallelizedTaskRunner
 {
+  GstTaskPool *pool;
+  gboolean own_pool;
   guint n_threads;
 
-  GstParallelizedTaskThread *threads;
+  GstQueueArray *tasks;
 
   GstParallelizedTaskFunc func;
   gpointer *task_data;
 
   GMutex lock;
-  GCond cond_todo, cond_done;
-  gint n_todo, n_done;
-  gboolean quit;
+  gint n_todo;
+
+  gboolean async_tasks;
 };
 
 /**
@@ -139,7 +134,7 @@ struct _GstCompositor
  */
 struct _GstCompositorPad
 {
-  GstVideoAggregatorConvertPad parent;
+  GstVideoAggregatorParallelConvertPad parent;
 
   /* properties */
   gint xpos, ypos;