Refactor ThreadLocal when running on hardware with more cores than PLATFORM_THREADS_LIMIT

Source/Engine/Animations/Graph/AnimGraph.Custom.cpp

@@ -89,7 +89,7 @@ void AnimGraphExecutor::initRuntime()
 void AnimGraphExecutor::ProcessGroupCustom(Box* boxBase, Node* nodeBase, Value& value)
 {
 #if USE_CSHARP
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     if (context.ValueCache.TryGet(boxBase, value))
         return;
     auto box = (AnimGraphBox*)boxBase;

Source/Engine/Animations/Graph/AnimGraph.cpp

@@ -9,7 +9,7 @@
 
 extern void RetargetSkeletonNode(const SkeletonData& sourceSkeleton, const SkeletonData& targetSkeleton, const SkinnedModel::SkeletonMapping& sourceMapping, Transform& node, int32 i);
 
-ThreadLocal<AnimGraphContext> AnimGraphExecutor::Context;
+ThreadLocal<AnimGraphContext*> AnimGraphExecutor::Context;
 
 Transform AnimGraphImpulse::GetNodeModelTransformation(SkeletonData& skeleton, int32 nodeIndex) const
 {
@@ -104,7 +104,7 @@ AnimGraphInstanceData::OutgoingEvent AnimGraphInstanceData::ActiveEvent::End(Ani
 
 AnimGraphImpulse* AnimGraphNode::GetNodes(AnimGraphExecutor* executor)
 {
-    auto& context = AnimGraphExecutor::Context.Get();
+    auto& context = *AnimGraphExecutor::Context.Get();
     const int32 count = executor->_skeletonNodesCount;
     if (context.PoseCacheSize == context.PoseCache.Count())
         context.PoseCache.AddOne();
@@ -204,7 +204,10 @@ void AnimGraphExecutor::Update(AnimGraphInstanceData& data, float dt)
 
     // Initialize
     auto& skeleton = _graph.BaseModel->Skeleton;
-    auto& context = Context.Get();
+    auto& contextPtr = Context.Get();
+    if (!contextPtr)
+        contextPtr = New<AnimGraphContext>();
+    auto& context = *contextPtr;
     {
         ANIM_GRAPH_PROFILE_EVENT("Init");
 
@@ -378,12 +381,12 @@ void AnimGraphExecutor::GetInputValue(Box* box, Value& result)
 
 AnimGraphImpulse* AnimGraphExecutor::GetEmptyNodes()
 {
-    return &Context.Get().EmptyNodes;
+    return &Context.Get()->EmptyNodes;
 }
 
 void AnimGraphExecutor::InitNodes(AnimGraphImpulse* nodes) const
 {
-    const auto& emptyNodes = Context.Get().EmptyNodes;
+    const auto& emptyNodes = Context.Get()->EmptyNodes;
     Platform::MemoryCopy(nodes->Nodes.Get(), emptyNodes.Nodes.Get(), sizeof(Transform) * _skeletonNodesCount);
     nodes->RootMotion = emptyNodes.RootMotion;
     nodes->Position = emptyNodes.Position;
@@ -405,7 +408,7 @@ void AnimGraphExecutor::ResetBuckets(AnimGraphContext& context, AnimGraphBase* g
 
 VisjectExecutor::Value AnimGraphExecutor::eatBox(Node* caller, Box* box)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
 
     // Check if graph is looped or is too deep
     if (context.CallStack.Count() >= ANIM_GRAPH_MAX_CALL_STACK)
@@ -450,6 +453,6 @@ VisjectExecutor::Value AnimGraphExecutor::eatBox(Node* caller, Box* box)
 
 VisjectExecutor::Graph* AnimGraphExecutor::GetCurrentGraph() const
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     return context.GraphStack.Peek();
 }

Source/Engine/Animations/Graph/AnimGraph.h

@@ -819,7 +819,7 @@ private:
     int32 _skeletonNodesCount = 0;
 
     // Per-thread context to allow async execution
-    static ThreadLocal<AnimGraphContext> Context;
+    static ThreadLocal<AnimGraphContext*> Context;
 
 public:
     /// <summary>

Source/Engine/Animations/Graph/AnimGroup.Animation.cpp

@@ -87,7 +87,7 @@ void AnimGraphExecutor::ProcessAnimEvents(AnimGraphNode* node, bool loop, float
     if (anim->Events.Count() == 0)
         return;
     ANIM_GRAPH_PROFILE_EVENT("Events");
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     float eventTimeMin = animPrevPos;
     float eventTimeMax = animPos;
     if (loop && context.DeltaTime * speed < 0)
@@ -231,7 +231,7 @@ void AnimGraphExecutor::ProcessAnimation(AnimGraphImpulse* nodes, AnimGraphNode*
     const float animPrevPos = GetAnimSamplePos(length, anim, prevPos, speed);
 
     // Add to trace
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     if (context.Data->EnableTracing)
     {
         auto& trace = context.AddTraceEvent(node);
@@ -655,7 +655,7 @@ void ComputeMultiBlendLength(float& length, AnimGraphNode* node)
 
 void AnimGraphExecutor::ProcessGroupParameters(Box* box, Node* node, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     switch (node->TypeID)
     {
     // Get
@@ -766,7 +766,7 @@ void AnimGraphExecutor::ProcessGroupParameters(Box* box, Node* node, Value& valu
 
 void AnimGraphExecutor::ProcessGroupTools(Box* box, Node* nodeBase, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     auto node = (AnimGraphNode*)nodeBase;
     switch (node->TypeID)
     {
@@ -790,7 +790,7 @@ void AnimGraphExecutor::ProcessGroupTools(Box* box, Node* nodeBase, Value& value
 
 void AnimGraphExecutor::ProcessGroupAnimation(Box* boxBase, Node* nodeBase, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     if (context.ValueCache.TryGet(boxBase, value))
         return;
     auto box = (AnimGraphBox*)boxBase;
@@ -2272,7 +2272,7 @@ void AnimGraphExecutor::ProcessGroupAnimation(Box* boxBase, Node* nodeBase, Valu
 
 void AnimGraphExecutor::ProcessGroupFunction(Box* boxBase, Node* node, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     if (context.ValueCache.TryGet(boxBase, value))
         return;
     switch (node->TypeID)

Source/Engine/Content/Storage/FlaxStorage.cpp

@@ -211,7 +211,10 @@ FlaxStorage::~FlaxStorage()
 
 #if USE_EDITOR
     // Ensure to close any outstanding file handles to prevent file locking in case it failed to load
-    _file.DeleteAll();
+    Array<FileReadStream*> streams;
+    _file.GetValues(streams);
+    for (FileReadStream* stream : streams)
+        Delete(stream);
 #endif
 }
 
@@ -1264,7 +1267,6 @@ bool FlaxStorage::LoadAssetHeader(const Entry& e, AssetInitData& data)
     }
 
 #if ASSETS_LOADING_EXTRA_VERIFICATION
-
     // Validate loaded header (asset ID and type ID must be the same)
     if (e.ID != data.Header.ID)
     {
@@ -1274,7 +1276,6 @@ bool FlaxStorage::LoadAssetHeader(const Entry& e, AssetInitData& data)
     {
         LOG(Error, "Loading asset header data mismatch! Expected Type Name: {0}, loaded header: {1}.\nSource: {2}", e.TypeName, data.Header.ToString(), ToString());
     }
-
 #endif
 
     return false;
@@ -1337,7 +1338,11 @@ bool FlaxStorage::CloseFileHandles()
         return true; // Failed, someone is still accessing the file
 
     // Close file handles (from all threads)
-    _file.DeleteAll();
+    Array<FileReadStream*> streams;
+    _file.GetValues(streams);
+    for (FileReadStream* stream : streams)
+        Delete(stream);
+    _file.Clear();
     return false;
 }
 

Source/Engine/Content/Storage/FlaxStorage.h

@@ -94,7 +94,7 @@ protected:
     CriticalSection _loadLocker;
 
     // Storage
-    ThreadLocalObject<FileReadStream> _file;
+    ThreadLocal<FileReadStream*> _file;
     Array<FlaxChunk*> _chunks;
 
     // Metadata

Source/Engine/Core/Collections/Sorting.cpp

@@ -5,11 +5,14 @@
 #include "Engine/Threading/ThreadLocal.h"
 
 // Use a cached storage for the sorting (one per thread to reduce locking)
-ThreadLocal<Sorting::SortingStack> SortingStacks;
+ThreadLocal<Sorting::SortingStack*> SortingStacks;
 
 Sorting::SortingStack& Sorting::SortingStack::Get()
 {
-    return SortingStacks.Get();
+    SortingStack*& stack = SortingStacks.Get();
+    if (!stack)
+        stack = New<SortingStack>();
+    return *stack;
 }
 
 Sorting::SortingStack::SortingStack()

Source/Engine/Engine/EngineService.cpp

@@ -72,9 +72,6 @@ void EngineService::OnInit()
 
     // Init services from front to back
     auto& services = GetServices();
-#if TRACY_ENABLE
-    Char nameBuffer[100];
-#endif
     for (int32 i = 0; i < services.Count(); i++)
     {
         const auto service = services[i];
@@ -82,6 +79,7 @@ void EngineService::OnInit()
 #if TRACY_ENABLE
         ZoneScoped;
         int32 nameBufferLength = 0;
+        Char nameBuffer[100];
         for (int32 j = 0; j < name.Length(); j++)
             if (name[j] != ' ')
                 nameBuffer[nameBufferLength++] = name[j];
@@ -114,6 +112,18 @@ void EngineService::OnDispose()
         const auto service = services[i];
         if (service->IsInitialized)
         {
+#if TRACY_ENABLE
+            ZoneScoped;
+            const StringView name(service->Name);
+            int32 nameBufferLength = 0;
+            Char nameBuffer[100];
+            for (int32 j = 0; j < name.Length(); j++)
+                if (name[j] != ' ')
+                    nameBuffer[nameBufferLength++] = name[j];
+            Platform::MemoryCopy(nameBuffer + nameBufferLength, TEXT("::Dispose"), 10 * sizeof(Char));
+            nameBufferLength += 10;
+            ZoneName(nameBuffer, nameBufferLength);
+#endif
             service->IsInitialized = false;
             service->Dispose();
         }

Source/Engine/Level/SceneObjectsFactory.cpp

@@ -63,7 +63,7 @@ SceneObjectsFactory::Context::~Context()
 {
     if (Async)
     {
-        Array<ISerializeModifier*, FixedAllocation<PLATFORM_THREADS_LIMIT>> modifiers;
+        Array<ISerializeModifier*, InlinedAllocation<PLATFORM_THREADS_LIMIT>> modifiers;
         Modifiers.GetValues(modifiers);
         for (ISerializeModifier* e : modifiers)
         {

Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp

@@ -52,7 +52,7 @@ namespace
 int32 ParticleEmitterGraphCPUExecutor::ProcessSpawnModule(int32 index)
 {
     const auto node = _graph.SpawnModules[index];
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     auto& data = context.Data->SpawnModulesData[index];
 
     // Accumulate the previous frame fraction
@@ -120,7 +120,7 @@ int32 ParticleEmitterGraphCPUExecutor::ProcessSpawnModule(int32 index)
 
 void ParticleEmitterGraphCPUExecutor::ProcessModule(ParticleEmitterGraphCPUNode* node, int32 particlesStart, int32 particlesEnd)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     auto stride = context.Data->Buffer->Stride;
     auto start = context.Data->Buffer->GetParticleCPU(particlesStart);
 

Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp

@@ -12,7 +12,7 @@
 
 void ParticleEmitterGraphCPUExecutor::ProcessGroupParameters(Box* box, Node* node, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     switch (node->TypeID)
     {
     // Get
@@ -168,7 +168,7 @@ void ParticleEmitterGraphCPUExecutor::ProcessGroupTextures(Box* box, Node* node,
 
 void ParticleEmitterGraphCPUExecutor::ProcessGroupTools(Box* box, Node* node, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     switch (node->TypeID)
     {
     // Linearize Depth
@@ -202,7 +202,7 @@ void ParticleEmitterGraphCPUExecutor::ProcessGroupTools(Box* box, Node* node, Va
 
 void ParticleEmitterGraphCPUExecutor::ProcessGroupParticles(Box* box, Node* nodeBase, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     auto node = (ParticleEmitterGraphCPUNode*)nodeBase;
     switch (node->TypeID)
     {
@@ -468,7 +468,7 @@ void ParticleEmitterGraphCPUExecutor::ProcessGroupParticles(Box* box, Node* node
 
 void ParticleEmitterGraphCPUExecutor::ProcessGroupFunction(Box* box, Node* node, Value& value)
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     switch (node->TypeID)
     {
     // Function Input

Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp

@@ -8,7 +8,7 @@
 #include "Engine/Engine/Time.h"
 #include "Engine/Profiler/ProfilerCPU.h"
 
-ThreadLocal<ParticleEmitterGraphCPUContext> ParticleEmitterGraphCPUExecutor::Context;
+ThreadLocal<ParticleEmitterGraphCPUContext*> ParticleEmitterGraphCPUExecutor::Context;
 
 namespace
 {
@@ -122,7 +122,10 @@ ParticleEmitterGraphCPUExecutor::ParticleEmitterGraphCPUExecutor(ParticleEmitter
 
 void ParticleEmitterGraphCPUExecutor::Init(ParticleEmitter* emitter, ParticleEffect* effect, ParticleEmitterInstance& data, float dt)
 {
-    auto& context = Context.Get();
+    auto& contextPtr = Context.Get();
+    if (!contextPtr)
+        contextPtr = New<ParticleEmitterGraphCPUContext>();
+    auto& context = *contextPtr;
     context.GraphStack.Clear();
     context.GraphStack.Push(&_graph);
     context.Data = &data;
@@ -252,8 +255,8 @@ bool ParticleEmitterGraphCPUExecutor::ComputeBounds(ParticleEmitter* emitter, Pa
             case 401:
             {
                 // Prepare graph data
-                auto& context = Context.Get();
                 Init(emitter, effect, data);
+                auto& context = *Context.Get();
 
                 // Find the maximum radius of the particle light
                 float maxRadius = 0.0f;
@@ -377,7 +380,7 @@ void ParticleEmitterGraphCPUExecutor::Draw(ParticleEmitter* emitter, ParticleEff
 
     // Prepare graph data
     Init(emitter, effect, data);
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
 
     // Draw lights
     for (int32 moduleIndex = 0; moduleIndex < emitter->Graph.LightModules.Count(); moduleIndex++)
@@ -571,7 +574,6 @@ int32 ParticleEmitterGraphCPUExecutor::UpdateSpawn(ParticleEmitter* emitter, Par
     PROFILE_CPU_NAMED("Spawn");
 
     // Prepare data
-    auto& context = Context.Get();
     Init(emitter, effect, data, dt);
 
     // Spawn particles
@@ -587,7 +589,7 @@ int32 ParticleEmitterGraphCPUExecutor::UpdateSpawn(ParticleEmitter* emitter, Par
 VisjectExecutor::Value ParticleEmitterGraphCPUExecutor::eatBox(Node* caller, Box* box)
 {
     // Check if graph is looped or is too deep
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     if (context.CallStackSize >= PARTICLE_EMITTER_MAX_CALL_STACK)
     {
         OnError(caller, box, TEXT("Graph is looped or too deep!"));
@@ -618,6 +620,6 @@ VisjectExecutor::Value ParticleEmitterGraphCPUExecutor::eatBox(Node* caller, Box
 
 VisjectExecutor::Graph* ParticleEmitterGraphCPUExecutor::GetCurrentGraph() const
 {
-    auto& context = Context.Get();
+    auto& context = *Context.Get();
     return (Graph*)context.GraphStack.Peek();
 }

Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h

@@ -133,7 +133,7 @@ private:
     ParticleEmitterGraphCPU& _graph;
 
     // Per-thread context to allow async execution
-    static ThreadLocal<ParticleEmitterGraphCPUContext> Context;
+    static ThreadLocal<ParticleEmitterGraphCPUContext*> Context;
 
 public:
     /// <summary>

Source/Engine/Platform/Base/PlatformBase.cpp

@@ -41,6 +41,10 @@ static_assert(sizeof(bool) == 1, "Invalid bool type size.");
 static_assert(sizeof(float) == 4, "Invalid float type size.");
 static_assert(sizeof(double) == 8, "Invalid double type size.");
 
+// Check configuration
+static_assert((PLATFORM_THREADS_LIMIT & (PLATFORM_THREADS_LIMIT - 1)) == 0, "Threads limit must be power of two.");
+static_assert(PLATFORM_THREADS_LIMIT % 4 == 0, "Threads limit must be multiple of 4.");
+
 float PlatformBase::CustomDpiScale = 1.0f;
 Array<User*, FixedAllocation<8>> PlatformBase::Users;
 Delegate<User*> PlatformBase::UserAdded;

Source/Engine/Scripting/Scripting.cpp

@@ -115,7 +115,7 @@ Action Scripting::ScriptsLoaded;
 Action Scripting::ScriptsUnload;
 Action Scripting::ScriptsReloading;
 Action Scripting::ScriptsReloaded;
-ThreadLocal<Scripting::IdsMappingTable*, PLATFORM_THREADS_LIMIT, true> Scripting::ObjectsLookupIdMapping;
+ThreadLocal<Scripting::IdsMappingTable*, PLATFORM_THREADS_LIMIT> Scripting::ObjectsLookupIdMapping;
 ScriptingService ScriptingServiceInstance;
 
 bool initFlaxEngine();

Source/Engine/Scripting/Scripting.h

@@ -6,7 +6,7 @@
 #include "Engine/Scripting/ScriptingType.h"
 #include "Types.h"
 
-template<typename T, int32 MaxThreads, bool ClearMemory>
+template<typename T, int32 MaxThreads>
 class ThreadLocal;
 
 /// <summary>
@@ -114,7 +114,7 @@ public:
     /// <summary>
     /// The objects lookup identifier mapping used to override the object ids on FindObject call (used by the object references deserialization).
     /// </summary>
-    static ThreadLocal<IdsMappingTable*, PLATFORM_THREADS_LIMIT, true> ObjectsLookupIdMapping;
+    static ThreadLocal<IdsMappingTable*, PLATFORM_THREADS_LIMIT> ObjectsLookupIdMapping;
 
     /// <summary>
     /// Finds the object by the given identifier. Searches registered scene objects and optionally assets. Logs warning if fails.

Source/Engine/Serialization/FileReadStream.h

@@ -12,7 +12,6 @@
 class FLAXENGINE_API FileReadStream : public ReadStream
 {
 private:
-
     File* _file;
     uint32 _virtualPosInBuffer; // Current position in the buffer (index)
     uint32 _bufferSize; // Amount of loaded bytes from the file to the buffer
@@ -33,11 +32,9 @@ public:
     ~FileReadStream();
 
 public:
-
     /// <summary>
     /// Gets the file handle.
     /// </summary>
-    /// <returns>File</returns>
     FORCE_INLINE const File* GetFile() const
     {
         return _file;
@@ -49,7 +46,6 @@ public:
     void Unlink();
 
 public:
-
     /// <summary>
     /// Open file to write data to it
     /// </summary>
@@ -58,7 +54,6 @@ public:
     static FileReadStream* Open(const StringView& path);
 
 public:
-
     // [ReadStream]
     void Flush() final override;
     void Close() final override;

Source/Engine/Threading/JobSystem.cpp

@@ -93,7 +93,7 @@ struct TIsPODType<JobContext>
 namespace
 {
     JobSystemService JobSystemInstance;
-    Thread* Threads[PLATFORM_THREADS_LIMIT] = {};
+    Thread* Threads[PLATFORM_THREADS_LIMIT / 2] = {};
     int32 ThreadsCount = 0;
     bool JobStartingOnDispatch = true;
     volatile int64 ExitFlag = 0;

Source/Engine/Threading/ThreadLocal.h

@@ -6,14 +6,14 @@
 #include "Engine/Platform/Platform.h"
 
 /// <summary>
-/// Per-thread local variable storage.
+/// Per-thread local variable storage for basic types (POD). Implemented using atomic with per-thread storage indexed via thread id hashing. Consider using 'THREADLOCAL' define before the variable instead.
-/// Implemented using atomic with per-thread storage indexed via thread id hashing.
-/// ForConsider using 'THREADLOCAL' define before the variable instead.
 /// </summary>
-template<typename T, int32 MaxThreads = PLATFORM_THREADS_LIMIT, bool ClearMemory = true>
+template<typename T, int32 MaxThreads = PLATFORM_THREADS_LIMIT>
 class ThreadLocal
 {
 protected:
+    constexpr static int32 DynamicMaxThreads = 1024;
+    static_assert(TIsPODType<T>::Value, "Only POD types are supported");
 
     struct Bucket
     {
@@ -21,34 +21,29 @@ protected:
         T Value;
     };
 
-    Bucket _buckets[MaxThreads];
+    Bucket _staticBuckets[MaxThreads];
+    Bucket* _dynamicBuckets = nullptr;
 
 public:
-
     ThreadLocal()
     {
-        // Clear buckets
+        Platform::MemoryClear(_staticBuckets, sizeof(_staticBuckets));
-        if (ClearMemory)
+    }
-        {
+
-            Platform::MemoryClear(_buckets, sizeof(_buckets));
+    ~ThreadLocal()
-        }
+    {
-        else
+        Platform::Free(_dynamicBuckets);
-        {
-            for (int32 i = 0; i < MaxThreads; i++)
-                _buckets[i].ThreadID = 0;
-        }
     }
 
 public:
-
+    FORCE_INLINE T& Get()
-    T& Get()
     {
-        return _buckets[GetIndex()].Value;
+        return GetBucket().Value;
     }
 
-    void Set(const T& value)
+    FORCE_INLINE void Set(const T& value)
     {
-        _buckets[GetIndex()].Value = value;
+        GetBucket().Value = value;
     }
 
     int32 Count() const
@@ -56,9 +51,17 @@ public:
         int32 result = 0;
         for (int32 i = 0; i < MaxThreads; i++)
         {
-            if (Platform::AtomicRead((int64 volatile*)&_buckets[i].ThreadID) != 0)
+            if (Platform::AtomicRead((int64 volatile*)&_staticBuckets[i].ThreadID) != 0)
                 result++;
         }
+        if (auto dynamicBuckets = (Bucket*)Platform::AtomicRead((intptr volatile*)&_dynamicBuckets))
+        {
+            for (int32 i = 0; i < MaxThreads; i++)
+            {
+                if (Platform::AtomicRead((int64 volatile*)&dynamicBuckets[i].ThreadID) != 0)
+                    result++;
+            }
+        }
         return result;
     }
 
@@ -67,89 +70,72 @@ public:
     {
         for (int32 i = 0; i < MaxThreads; i++)
         {
-            if (Platform::AtomicRead((int64 volatile*)&_buckets[i].ThreadID) != 0)
+            if (Platform::AtomicRead((int64 volatile*)&_staticBuckets[i].ThreadID) != 0)
-                result.Add(_buckets[i].Value);
+                result.Add(_staticBuckets[i].Value);
         }
+        if (auto dynamicBuckets = (Bucket*)Platform::AtomicRead((intptr volatile*)&_dynamicBuckets))
+        {
+            for (int32 i = 0; i < MaxThreads; i++)
+            {
+                if (Platform::AtomicRead((int64 volatile*)&dynamicBuckets[i].ThreadID) != 0)
+                    result.Add(dynamicBuckets[i].Value);
+            }
+        }
+    }
+
+    void Clear()
+    {
+        Platform::MemoryClear(_staticBuckets, sizeof(_staticBuckets));
+        Platform::Free(_dynamicBuckets);
+        _dynamicBuckets = nullptr;
     }
 
 protected:
-
+    Bucket& GetBucket()
-    FORCE_INLINE static int32 Hash(const int64 value)
     {
-        return value & (MaxThreads - 1);
+        const int64 key = (int64)Platform::GetCurrentThreadID();
-    }
 
-    FORCE_INLINE int32 GetIndex()
+        // Search statically allocated buckets
-    {
+        int32 index = (int32)(key & (MaxThreads - 1));
-        int64 key = (int64)Platform::GetCurrentThreadID();
+        int32 spaceLeft = MaxThreads;
-        auto index = Hash(key);
+        while (spaceLeft)
-        while (true)
         {
-            const int64 value = Platform::AtomicRead(&_buckets[index].ThreadID);
+            const int64 value = Platform::AtomicRead(&_staticBuckets[index].ThreadID);
             if (value == key)
-                break;
+                return _staticBuckets[index];
-            if (value == 0 && Platform::InterlockedCompareExchange(&_buckets[index].ThreadID, key, 0) == 0)
+            if (value == 0 && Platform::InterlockedCompareExchange(&_staticBuckets[index].ThreadID, key, 0) == 0)
-                break;
+                return _staticBuckets[index];
-            index = Hash(index + 1);
+            index = (index + 1) & (MaxThreads - 1);
+            spaceLeft--;
         }
-        return index;
-    }
-};
 
-/// <summary>
+        // Allocate dynamic buckets if missing
-/// Per thread local object
+    DYNAMIC:
-/// </summary>
+        auto dynamicBuckets = (Bucket*)Platform::AtomicRead((intptr volatile*)&_dynamicBuckets);
-template<typename T, int32 MaxThreads = PLATFORM_THREADS_LIMIT>
+        if (!dynamicBuckets)
-class ThreadLocalObject : public ThreadLocal<T*, MaxThreads>
-{
-public:
-
-    typedef ThreadLocal<T*, MaxThreads> Base;
-
-public:
-
-    void Delete()
-    {
-        auto value = Base::Get();
-        Base::SetAll(nullptr);
-        ::Delete(value);
-    }
-
-    void DeleteAll()
-    {
-        for (int32 i = 0; i < MaxThreads; i++)
         {
-            auto& bucket = Base::_buckets[i];
+            dynamicBuckets = (Bucket*)Platform::Allocate(DynamicMaxThreads * sizeof(Bucket), 16);
-            if (bucket.Value != nullptr)
+            Platform::MemoryClear(dynamicBuckets, DynamicMaxThreads * sizeof(Bucket));
+            if (Platform::InterlockedCompareExchange((intptr volatile*)&_dynamicBuckets, (intptr)dynamicBuckets, 0) != 0)
             {
-                ::Delete(bucket.Value);
+                Platform::Free(dynamicBuckets);
-                bucket.ThreadID = 0;
+                goto DYNAMIC;
-                bucket.Value = nullptr;
             }
         }
-    }
 
-    template<typename AllocationType = HeapAllocation>
+        // Search dynamically allocated buckets
-    void GetNotNullValues(Array<T*, AllocationType>& result) const
+        index = (int32)(key & (DynamicMaxThreads - 1));
-    {
+        spaceLeft = DynamicMaxThreads;
-        result.EnsureCapacity(MaxThreads);
+        while (spaceLeft)
-        for (int32 i = 0; i < MaxThreads; i++)
         {
-            if (Base::_buckets[i].Value != nullptr)
+            const int64 value = Platform::AtomicRead(&dynamicBuckets[index].ThreadID);
-            {
+            if (value == key)
-                result.Add(Base::_buckets[i].Value);
+                return dynamicBuckets[index];
-            }
+            if (value == 0 && Platform::InterlockedCompareExchange(&dynamicBuckets[index].ThreadID, key, 0) == 0)
+                return dynamicBuckets[index];
+            index = (index + 1) & (DynamicMaxThreads - 1);
+            spaceLeft--;
         }
-    }
+        return *(Bucket*)nullptr;
-
-    int32 CountNotNullValues() const
-    {
-        int32 result = 0;
-        for (int32 i = 0; i < MaxThreads; i++)
-        {
-            if (Base::_buckets[i].Value != nullptr)
-                result++;
-        }
-        return result;
     }
 };

Source/Engine/Threading/ThreadPool.cpp

@@ -58,7 +58,7 @@ ThreadPoolService ThreadPoolServiceInstance;
 bool ThreadPoolService::Init()
 {
     // Spawn threads
-    const int32 numThreads = Math::Clamp<int32>(Platform::GetCPUInfo().ProcessorCoreCount - 1, 2, PLATFORM_THREADS_LIMIT);
+    const int32 numThreads = Math::Clamp<int32>(Platform::GetCPUInfo().ProcessorCoreCount - 1, 2, PLATFORM_THREADS_LIMIT / 2);
     LOG(Info, "Spawning {0} Thread Pool workers", numThreads);
     for (int32 i = ThreadPoolImpl::Threads.Count(); i < numThreads; i++)
     {