[Bf-blender-cvs] [12a20b7] master: Depsgraph: Don't leave active thread if there's only one children node

[Sergey Sharybin]

Commit: 12a20b78d3d9b0f3d5ef9cc6a81ae2619c82db88
Author: Sergey Sharybin
Date:   Tue May 10 10:39:38 2016 +0200
Branches: master
https://developer.blender.org/rB12a20b78d3d9b0f3d5ef9cc6a81ae2619c82db88

Depsgraph: Don't leave active thread if there's only one children node

This reduces stress on the task scheduler and avoids some unwanted overhead
caused by all the threading business in the cases when there's only one
children node. We try to immediately switch to it's evaluation now, keeping
active thread up and running.

This bumps FPS from 58 to 64 on the blenrig test file from jpbouza.

===================================================================

M	source/blender/depsgraph/intern/depsgraph_eval.cc

===================================================================

diff --git a/source/blender/depsgraph/intern/depsgraph_eval.cc b/source/blender/depsgraph/intern/depsgraph_eval.cc
index e96ee61..8031f46 100644
--- a/source/blender/depsgraph/intern/depsgraph_eval.cc
+++ b/source/blender/depsgraph/intern/depsgraph_eval.cc
@@ -140,29 +140,101 @@ static void deg_task_run_func(TaskPool *pool,
 
 	BLI_assert(!node->is_noop() && "NOOP nodes should not actually be scheduled");
 
-	/* Get context. */
-	// TODO: who initialises this? "Init" operations aren't able to initialise it!!!
-	/* TODO(sergey): Wedon't use component contexts at this moment. */
-	/* ComponentDepsNode *comp = node->owner; */
-	BLI_assert(node->owner != NULL);
-	
-	/* Take note of current time. */
-	double start_time = PIL_check_seconds_timer();
-	DepsgraphDebug::task_started(state->graph, node);
-	
 	/* Should only be the case for NOOPs, which never get to this point. */
 	BLI_assert(node->evaluate);
-	
-	/* Perform operation. */
-	node->evaluate(state->eval_ctx);
-	
-	/* Note how long this took. */
-	double end_time = PIL_check_seconds_timer();
-	DepsgraphDebug::task_completed(state->graph,
-	                               node,
-	                               end_time - start_time);
 
-	schedule_children(pool, state->graph, node, state->layers, thread_id);
+	while (true) {
+		/* Get context. */
+		// TODO: who initialises this? "Init" operations aren't able to initialise it!!!
+		/* TODO(sergey): We don't use component contexts at this moment. */
+		/* ComponentDepsNode *comp = node->owner; */
+		BLI_assert(node->owner != NULL);
+
+		/* Since we're not leaving the thread for until the graph branches it is
+		 * possible to have NO-OP on the way. for which evaluate() will be NULL.
+		 * but that's all fine, we'll just scheduler it's children.
+		 */
+		if (node->evaluate) {
+			/* Take note of current time. */
+			double start_time = PIL_check_seconds_timer();
+			DepsgraphDebug::task_started(state->graph, node);
+
+			/* Perform operation. */
+			node->evaluate(state->eval_ctx);
+
+			/* Note how long this took. */
+			double end_time = PIL_check_seconds_timer();
+			DepsgraphDebug::task_completed(state->graph,
+			                               node,
+			                               end_time - start_time);
+		}
+
+		/* If there's only one outgoing link we try to immediately switch to
+		 * that node evaluation, without leaving the thread.
+		 *
+		 * It's only doable if the child don't have extra relations or all they
+		 * are satisfied.
+		 *
+		 * TODO(sergey): Checks here can be de-duplicated with the ones from
+		 * schedule_node(), however, how to do it nicely?
+		 */
+		if (node->outlinks.size() == 1) {
+			DepsRelation *rel = node->outlinks[0];
+			OperationDepsNode *child = (OperationDepsNode *)rel->to;
+			BLI_assert(child->type == DEPSNODE_TYPE_OPERATION);
+			if (!child->scheduled) {
+				int id_layers = child->owner->owner->layers;
+				if (!((child->flag & DEPSOP_FLAG_NEEDS_UPDATE) != 0 &&
+				      (id_layers & state->layers) != 0))
+				{
+					/* Node does not need an update, so can;t continue with the
+					 * chain and need to switch to another one by leaving the
+					 * thread.
+					 */
+					break;
+				}
+				if ((rel->flag & DEPSREL_FLAG_CYCLIC) == 0) {
+					BLI_assert(child->num_links_pending > 0);
+					atomic_sub_uint32(&child->num_links_pending, 1);
+				}
+				if (child->num_links_pending == 0) {
+					bool is_scheduled = atomic_fetch_and_or_uint8((uint8_t *)&child->scheduled, (uint8_t)true);
+					if (!is_scheduled) {
+						/* Node was not scheduled, switch to it! */
+						node = child;
+					}
+					else {
+						/* Someone else scheduled the node, leaving us
+						 * unemployed in this thread, we're leaving.
+						 */
+						break;
+					}
+				}
+				else {
+					/* There are other dependencies on the child, can't do
+					 * anything in the current thread.
+					 */
+					break;
+				}
+			}
+			else {
+				/* Happens when having cyclic dependencies.
+				 *
+				 * Nothing to do here, single child was already scheduled, we
+				 * can leave the thread now.
+				 */
+				break;
+			}
+		}
+		else {
+			/* TODO(sergey): It's possible to use one of the outgoing relations
+			 * as a chain which we'll try to keep alive, but it's a bit more
+			 * involved change.
+			 */
+			schedule_children(pool, state->graph, node, state->layers, thread_id);
+			break;
+		}
+	}
 }
 
 static void calculate_pending_parents(Depsgraph *graph, int layers)