Optimize scenes loading with Job System

Source/Engine/Core/Cache.cpp

@@ -8,6 +8,7 @@ CollectionPoolCache<ISerializeModifier, Cache::ISerializeModifierClearCallback>
 void Cache::ISerializeModifierClearCallback(::ISerializeModifier* obj)
 {
     obj->EngineBuild = FLAXENGINE_VERSION_BUILD;
+    obj->CurrentInstance = -1;
     obj->IdsMapping.Clear();
 }
 

Source/Engine/Core/Collections/CollectionPoolCache.h

@@ -49,7 +49,7 @@ public:
     private:
         CollectionPoolCache* _pool;
 
-        ScopeCache(CollectionPoolCache* pool, T* value)
+        FORCE_INLINE ScopeCache(CollectionPoolCache* pool, T* value)
         {
             _pool = pool;
             Value = value;
@@ -71,7 +71,7 @@ public:
 
         ~ScopeCache()
         {
-            _pool->Release(Value);
+            _pool->Put(Value);
         }
 
         T* operator->()
@@ -99,13 +99,6 @@ private:
     CriticalSection _locker;
     Array<T*, InlinedAllocation<64>> _pool;
 
-    void Release(T* value)
-    {
-        _locker.Lock();
-        _pool.Add(value);
-        _locker.Unlock();
-    }
-
 public:
     /// <summary>
     /// Finalizes an instance of the <see cref="CollectionPoolCache"/> class.
@@ -120,7 +113,16 @@ public:
     /// Gets the collection instance from the pool. Can reuse the object from the pool or create a new one. Returns collection is always cleared and ready to use.
     /// </summary>
     /// <returns>The collection (cleared).</returns>
-    ScopeCache Get()
+    FORCE_INLINE ScopeCache Get()
+    {
+        return ScopeCache(this, GetUnscoped());
+    }
+
+    /// <summary>
+    /// Gets the collection instance from the pool. Can reuse the object from the pool or create a new one. Returns collection is always cleared and ready to use.
+    /// </summary>
+    /// <returns>The collection (cleared).</returns>
+    T* GetUnscoped()
     {
         T* result;
         _locker.Lock();
@@ -129,10 +131,18 @@ public:
         else
             result = CreateCallback();
         _locker.Unlock();
-
         ClearCallback(result);
+        return result;
+    }
 
-        return ScopeCache(this, result);
+    /// <summary>
+    /// Puts the collection value back to the pool.
+    /// </summary>
+    void Put(T* value)
+    {
+        _locker.Lock();
+        _pool.Add(value);
+        _locker.Unlock();
     }
 
     /// <summary>

Source/Engine/Core/Collections/Dictionary.h

@@ -785,12 +785,11 @@ public:
     /// <param name="other">The other collection to clone.</param>
     void Clone(const Dictionary& other)
     {
+        // TODO: if both key and value are POD types then use raw memory copy for buckets
         Clear();
-        SetCapacity(other.Capacity(), false);
+        EnsureCapacity(other.Capacity(), false);
         for (Iterator i = other.Begin(); i != other.End(); ++i)
             Add(i);
-        ASSERT(Count() == other.Count());
-        ASSERT(Capacity() == other.Capacity());
     }
 
     /// <summary>

Source/Engine/Level/Actor.cpp

@@ -1013,6 +1013,7 @@ void Actor::Deserialize(DeserializeStream& stream, ISerializeModifier* modifier)
             const auto parent = Scripting::FindObject<Actor>(parentId);
             if (_parent != parent)
             {
+                ScopeLock lock(Level::ScenesLock);
                 if (IsDuringPlay())
                 {
                     SetParent(parent, false, false);

Source/Engine/Level/Level.cpp

@@ -18,6 +18,7 @@
 #include "Engine/Debug/Exceptions/JsonParseException.h"
 #include "Engine/Engine/EngineService.h"
 #include "Engine/Threading/Threading.h"
+#include "Engine/Threading/JobSystem.h"
 #include "Engine/Platform/File.h"
 #include "Engine/Platform/FileSystem.h"
 #include "Engine/Profiler/ProfilerCPU.h"
@@ -932,11 +933,9 @@ bool Level::loadScene(rapidjson_flax::Value& data, int32 engineBuild, Scene** ou
         LOG(Error, "Invalid Data member.");
         return true;
     }
-    int32 objectsCount = data.Size();
 
     // Peek scene node value (it's the first actor serialized)
-    auto& sceneValue = data[0];
+    auto sceneId = JsonTools::GetGuid(data[0], "ID");
-    auto sceneId = JsonTools::GetGuid(sceneValue, "ID");
     if (!sceneId.IsValid())
     {
         LOG(Error, "Invalid scene id.");
@@ -957,59 +956,103 @@ bool Level::loadScene(rapidjson_flax::Value& data, int32 engineBuild, Scene** ou
     auto scene = New<Scene>(ScriptingObjectSpawnParams(sceneId, Scene::TypeInitializer));
     scene->LoadTime = startTime;
     scene->RegisterObject();
-    scene->Deserialize(sceneValue, modifier.Value);
+    scene->Deserialize(data[0], modifier.Value);
 
     // Fire event
     CallSceneEvent(SceneEventType::OnSceneLoading, scene, sceneId);
 
     // Loaded scene objects list
     CollectionPoolCache<ActorsCache::SceneObjectsListType>::ScopeCache sceneObjects = ActorsCache::SceneObjectsListCache.Get();
+    const int32 objectsCount = (int32)data.Size();
     sceneObjects->Resize(objectsCount);
     sceneObjects->At(0) = scene;
 
+    // Spawn all scene objects
     SceneObjectsFactory::Context context(modifier.Value);
+    context.Async = JobSystem::GetThreadsCount() > 1 && objectsCount > 10;
     {
         PROFILE_CPU_NAMED("Spawn");
-
+        SceneObject** objects = sceneObjects->Get();
-        // Spawn all scene objects
+        if (context.Async)
-        for (int32 i = 1; i < objectsCount; i++) // start from 1. at index [0] was scene
         {
-            auto& stream = data[i];
+            JobSystem::Execute([&](int32 i)
-            auto obj = SceneObjectsFactory::Spawn(context, stream);
+            {
-            sceneObjects->At(i) = obj;
+                i++; // Start from 1. at index [0] was scene
-            if (obj)
+                auto& stream = data[i];
-                obj->RegisterObject();
+                auto obj = SceneObjectsFactory::Spawn(context, stream);
-            else
+                objects[i] = obj;
-                SceneObjectsFactory::HandleObjectDeserializationError(stream);
+                if (obj)
+                {
+                    obj->RegisterObject();
+#if USE_EDITOR
+                    // Auto-create C# objects for all actors in Editor during scene load when running in async (so main thread already has all of them)
+                    obj->CreateManaged();
+#endif
+                }
+                else
+                    SceneObjectsFactory::HandleObjectDeserializationError(stream);
+            }, objectsCount - 1);
+        }
+        else
+        {
+            for (int32 i = 1; i < objectsCount; i++) // start from 1. at index [0] was scene
+            {
+                auto& stream = data[i];
+                auto obj = SceneObjectsFactory::Spawn(context, stream);
+                sceneObjects->At(i) = obj;
+                if (obj)
+                    obj->RegisterObject();
+                else
+                    SceneObjectsFactory::HandleObjectDeserializationError(stream);
+            }
         }
     }
 
+    // Capture prefab instances in a scene to restore any missing objects (eg. newly added objects to prefab that are missing in scene file)
     SceneObjectsFactory::PrefabSyncData prefabSyncData(*sceneObjects.Value, data, modifier.Value);
-
     SceneObjectsFactory::SetupPrefabInstances(context, prefabSyncData);
-
     // TODO: resave and force sync scenes during game cooking so this step could be skipped in game
     SceneObjectsFactory::SynchronizeNewPrefabInstances(context, prefabSyncData);
 
     // /\ all above this has to be done on an any thread
     // \/ all below this has to be done on multiple threads at once
 
+    // Load all scene objects
     {
         PROFILE_CPU_NAMED("Deserialize");
-
-        // TODO: at this point we would probably spawn a few thread pool tasks which will load deserialize scene object but only if scene is big enough
-
-        // Load all scene objects
-        Scripting::ObjectsLookupIdMapping.Set(&modifier.Value->IdsMapping);
         SceneObject** objects = sceneObjects->Get();
-        for (int32 i = 1; i < objectsCount; i++) // start from 1. at index [0] was scene
+        bool wasAsync = context.Async;
+        context.Async = false; // TODO: fix Actor's Scripts and Children order when loading objects data out of order via async jobs
+        if (context.Async)
         {
-            auto& objData = data[i];
+            ScenesLock.Unlock(); // Unlock scenes from Main Thread so Job Threads can use it to safely setup actors hierarchy (see Actor::Deserialize)
-            auto obj = objects[i];
+            JobSystem::Execute([&](int32 i)
-            if (obj)
+            {
-                SceneObjectsFactory::Deserialize(context, obj, objData);
+                i++; // Start from 1. at index [0] was scene
+                auto obj = objects[i];
+                if (obj)
+                {
+                    auto& idMapping = Scripting::ObjectsLookupIdMapping.Get();
+                    idMapping = &context.GetModifier()->IdsMapping;
+                    SceneObjectsFactory::Deserialize(context, obj, data[i]);
+                    idMapping = nullptr;
+                }
+            }, objectsCount - 1);
+            ScenesLock.Lock();
         }
-        Scripting::ObjectsLookupIdMapping.Set(nullptr);
+        else
+        {
+            Scripting::ObjectsLookupIdMapping.Set(&modifier.Value->IdsMapping);
+            for (int32 i = 1; i < objectsCount; i++) // start from 1. at index [0] was scene
+            {
+                auto& objData = data[i];
+                auto obj = objects[i];
+                if (obj)
+                    SceneObjectsFactory::Deserialize(context, obj, objData);
+            }
+            Scripting::ObjectsLookupIdMapping.Set(nullptr);
+        }
+        context.Async = wasAsync;
     }
 
     // /\ all above this has to be done on multiple threads at once
@@ -1031,7 +1074,7 @@ bool Level::loadScene(rapidjson_flax::Value& data, int32 engineBuild, Scene** ou
         PROFILE_CPU_NAMED("Initialize");
 
         SceneObject** objects = sceneObjects->Get();
-        for (int32 i = 0; i < sceneObjects->Count(); i++)
+        for (int32 i = 0; i < objectsCount; i++)
         {
             SceneObject* obj = objects[i];
             if (obj)

Source/Engine/Level/SceneObjectsFactory.cpp

@@ -4,6 +4,7 @@
 #include "Engine/Level/Actor.h"
 #include "Engine/Level/Prefabs/Prefab.h"
 #include "Engine/Content/Content.h"
+#include "Engine/Core/Cache.h"
 #include "Engine/Core/Log.h"
 #include "Engine/Scripting/Scripting.h"
 #include "Engine/Serialization/JsonTools.h"
@@ -18,16 +19,49 @@ SceneObjectsFactory::Context::Context(ISerializeModifier* modifier)
 {
 }
 
-void SceneObjectsFactory::Context::SetupIdsMapping(const SceneObject* obj)
+SceneObjectsFactory::Context::~Context()
+{
+    if (Async)
+    {
+        Array<ISerializeModifier*, FixedAllocation<PLATFORM_THREADS_LIMIT>> modifiers;
+        Modifiers.GetValues(modifiers);
+        for (ISerializeModifier* e : modifiers)
+            Cache::ISerializeModifier.Put(e);
+    }
+}
+
+ISerializeModifier* SceneObjectsFactory::Context::GetModifier()
+{
+    ISerializeModifier* modifier = Modifier;
+    if (Async)
+    {
+        // When using context in async then use one ISerializeModifier per-thread
+        ISerializeModifier*& modifierThread = Modifiers.Get();
+        if (!modifierThread)
+        {
+            modifierThread = Cache::ISerializeModifier.GetUnscoped();
+            Modifiers.Set(modifierThread);
+            Locker.Lock();
+            modifierThread->EngineBuild = modifier->EngineBuild;
+            modifierThread->CurrentInstance = modifier->CurrentInstance;
+            modifierThread->IdsMapping = modifier->IdsMapping;
+            Locker.Unlock();
+        }
+        modifier = modifierThread;
+    }
+    return modifier;
+}
+
+void SceneObjectsFactory::Context::SetupIdsMapping(const SceneObject* obj, ISerializeModifier* modifier)
 {
     int32 instanceIndex;
-    if (ObjectToInstance.TryGet(obj->GetID(), instanceIndex) && instanceIndex != CurrentInstance)
+    if (ObjectToInstance.TryGet(obj->GetID(), instanceIndex) && instanceIndex != modifier->CurrentInstance)
     {
         // Apply the current prefab instance objects ids table to resolve references inside a prefab properly
-        CurrentInstance = instanceIndex;
+        modifier->CurrentInstance = instanceIndex;
         auto& instance = Instances[instanceIndex];
         for (auto& e : instance.IdsMapping)
-            Modifier->IdsMapping[e.Key] = e.Value;
+            modifier->IdsMapping[e.Key] = e.Value;
     }
 }
 
@@ -35,13 +69,13 @@ SceneObject* SceneObjectsFactory::Spawn(Context& context, const ISerializable::D
 {
     // Get object id
     Guid id = JsonTools::GetGuid(stream, "ID");
-    context.Modifier->IdsMapping.TryGet(id, id);
+    ISerializeModifier* modifier = context.GetModifier();
+    modifier->IdsMapping.TryGet(id, id);
     if (!id.IsValid())
     {
         LOG(Warning, "Invalid object id.");
         return nullptr;
     }
-
     SceneObject* obj = nullptr;
 
     // Check for prefab instance
@@ -78,7 +112,7 @@ SceneObject* SceneObjectsFactory::Spawn(Context& context, const ISerializable::D
         }
 
         // Map prefab object ID to the deserialized instance ID
-        context.Modifier->IdsMapping[prefabObjectId] = id;
+        modifier->IdsMapping[prefabObjectId] = id;
 
         // Create prefab instance (recursive prefab loading to support nested prefabs)
         obj = Spawn(context, *prefabData);
@@ -169,6 +203,7 @@ void SceneObjectsFactory::Deserialize(Context& context, SceneObject* obj, ISeria
 #if ENABLE_ASSERTION
     CHECK(obj);
 #endif
+    ISerializeModifier* modifier = context.GetModifier();
 
     // Check for prefab instance
     Guid prefabObjectId;
@@ -204,24 +239,16 @@ void SceneObjectsFactory::Deserialize(Context& context, SceneObject* obj, ISeria
         }
 
         // Deserialize prefab data (recursive prefab loading to support nested prefabs)
-        const auto prevVersion = context.Modifier->EngineBuild;
+        const auto prevVersion = modifier->EngineBuild;
-        context.Modifier->EngineBuild = prefab->DataEngineBuild;
+        modifier->EngineBuild = prefab->DataEngineBuild;
         Deserialize(context, obj, *(ISerializable::DeserializeStream*)prefabData);
-        context.Modifier->EngineBuild = prevVersion;
+        modifier->EngineBuild = prevVersion;
     }
 
-    int32 instanceIndex;
+    context.SetupIdsMapping(obj, modifier);
-    if (context.ObjectToInstance.TryGet(obj->GetID(), instanceIndex) && instanceIndex != context.CurrentInstance)
-    {
-        // Apply the current prefab instance objects ids table to resolve references inside a prefab properly
-        context.CurrentInstance = instanceIndex;
-        auto& instance = context.Instances[instanceIndex];
-        for (auto& e : instance.IdsMapping)
-            context.Modifier->IdsMapping[e.Key] = e.Value;
-    }
 
     // Load data
-    obj->Deserialize(stream, context.Modifier);
+    obj->Deserialize(stream, modifier);
 }
 
 void SceneObjectsFactory::HandleObjectDeserializationError(const ISerializable::DeserializeStream& value)
@@ -518,7 +545,7 @@ void SceneObjectsFactory::SynchronizePrefabInstances(Context& context, PrefabSyn
             LOG(Info, "Object {0} has invalid parent object {4} -> {5} (PrefabObjectID: {1}, PrefabID: {2}, Path: {3})", obj->GetSceneObjectId(), prefabObjectId, prefab->GetID(), prefab->GetPath(), parentPrefabObjectId, actualParentPrefabId);
 
             // Map actual prefab object id to the current scene objects collection
-            context.SetupIdsMapping(obj);
+            context.SetupIdsMapping(obj, data.Modifier);
             data.Modifier->IdsMapping.TryGet(actualParentPrefabId, actualParentPrefabId);
 
             // Find parent

Source/Engine/Level/SceneObjectsFactory.h

@@ -4,6 +4,8 @@
 
 #include "SceneObject.h"
 #include "Engine/Core/Collections/Dictionary.h"
+#include "Engine/Platform/CriticalSection.h"
+#include "Engine/Threading/ThreadLocal.h"
 
 /// <summary>
 /// Helper class for scene objects creation and deserialization utilities.
@@ -21,13 +23,17 @@ public:
     struct Context
     {
         ISerializeModifier* Modifier;
-        int32 CurrentInstance = -1;
+        bool Async = false;
         Array<PrefabInstance> Instances;
         Dictionary<Guid, int32> ObjectToInstance;
+        CriticalSection Locker;
+        ThreadLocal<ISerializeModifier*> Modifiers;
 
         Context(ISerializeModifier* modifier);
+        ~Context();
 
-        void SetupIdsMapping(const SceneObject* obj);
+        ISerializeModifier* GetModifier();
+        void SetupIdsMapping(const SceneObject* obj, ISerializeModifier* modifier);
     };
 
     /// <summary>

Source/Engine/Scripting/Script.cpp

@@ -327,6 +327,7 @@ void Script::Deserialize(DeserializeStream& stream, ISerializeModifier* modifier
             const auto parent = Scripting::FindObject<Actor>(parentId);
             if (_parent != parent)
             {
+                ScopeLock lock(Level::ScenesLock);
                 if (IsDuringPlay())
                 {
                     SetParent(parent, false);

Source/Engine/Serialization/ISerializeModifier.h

@@ -18,6 +18,9 @@ public:
     /// </summary>
     uint32 EngineBuild = FLAXENGINE_VERSION_BUILD;
 
+    // Utility for scene deserialization to track currently mapped in Prefab Instance object IDs into IdsMapping.
+    int32 CurrentInstance = -1;
+
     /// <summary>
     /// The object IDs mapping. Key is a serialized object id, value is mapped value to use.
     /// </summary>

Source/Engine/Threading/JobSystem.cpp

@@ -228,6 +228,7 @@ int32 JobSystemThread::Run()
 
 void JobSystem::Execute(const Function<void(int32)>& job, int32 jobCount)
 {
+#if JOB_SYSTEM_ENABLED
     // TODO: disable async if called on job thread? or maybe Wait should handle waiting in job thread to do the processing?
     if (jobCount > 1)
     {
@@ -235,7 +236,8 @@ void JobSystem::Execute(const Function<void(int32)>& job, int32 jobCount)
         const int64 jobWaitHandle = Dispatch(job, jobCount);
         Wait(jobWaitHandle);
     }
-    else if (jobCount > 0)
+    else
+#endif
     {
         // Sync
         for (int32 i = 0; i < jobCount; i++)

Source/Engine/Threading/ThreadLocal.h

@@ -65,10 +65,10 @@ public:
     template<typename AllocationType = HeapAllocation>
     void GetValues(Array<T, AllocationType>& result) const
     {
-        result.EnsureCapacity(MaxThreads);
         for (int32 i = 0; i < MaxThreads; i++)
         {
-            result.Add(_buckets[i].Value);
+            if (Platform::AtomicRead((int64 volatile*)&_buckets[i].ThreadID) != 0)
+                result.Add(_buckets[i].Value);
         }
     }