Works, but with memory leaks

.vscode/launch.json

@@ -12,7 +12,7 @@
             "args": [],
             "stopAtEntry": false,
             "cwd": "${workspaceRoot}/bin/Debug",
-            "console": "internalConsole",
+            "console": "integratedTerminal",
             "environment": [],
             "externalConsole": false,
             "setupCommands": [
@@ -31,7 +31,7 @@
             "args": [],
             "stopAtEntry": false,
             "cwd": "${workspaceRoot}/bin/Debug",
-            "console": "internalConsole",
+            "console": "integratedTerminal",
             "environment": [],
             "symbolSearchPath": "C:\\Program Files (x86)\\Microsoft Visual Studio\\2019\\Community\\VC\\Tools\\MSVC\\14.25.28610\\lib\\x64",
             "requireExactSource": false,
@@ -51,7 +51,7 @@
             "args": [],
             "stopAtEntry": false,
             "cwd": "${workspaceRoot}/bin/Release",
-            "console": "internalConsole",
+            "console": "integratedTerminal",
             "environment": [],
             "symbolSearchPath": "C:\\Program Files (x86)\\Microsoft Visual Studio\\2019\\Community\\VC\\Tools\\MSVC\\14.25.28610\\lib\\x64",
             "requireExactSource": false,
@@ -70,7 +70,7 @@
             "program": "${workspaceRoot}/bin/Debug/Seele_unit_tests",
             "args": [],
             "stopAtEntry": false,
-            "console": "internalConsole",
+            "console": "integratedTerminal",
             "cwd": "${workspaceRoot}/bin/Debug",
             "environment": [],
         },
@@ -81,7 +81,7 @@
             "program": "${workspaceRoot}/bin/Debug/Seele_unit_tests.exe",
             "args": ["--detect_memory_leaks=15533"],
             "stopAtEntry": false,
-            "console": "internalConsole",
+            "console": "integratedTerminal",
             "cwd": "${workspaceRoot}/bin/Debug",
             "visualizerFile": "${workspaceRoot}/Seele.natvis",
             "environment": [],

CMakeLists.txt

@@ -111,7 +111,7 @@ add_subdirectory(src/)
 
 if(MSVC)
   set(_CRT_SECURE_NO_WARNINGS)
-  target_compile_options(Engine PRIVATE /Zi /MP)
+  target_compile_options(Engine PRIVATE /Zi /MP /DEBUG:FASTLINK)
 else()
   target_compile_options(Engine PRIVATE -g -Wall -Wextra -Wno-delete-incomplete -pedantic -fcoroutines)
 endif()

cmake-variants.json

@@ -33,13 +33,6 @@
                 "settings": {
                     "CMAKE_PLATFORM": "x64"
                 }
-            },
-            "x86": {
-                "short": "x86",
-                "long": "x86 platform",
-                "settings": {
-                    "CMAKE_PLATFORM": "x86"
-                }
             }
         }
     },

src/Engine/Containers/RingBuffer.h

@@ -0,0 +1,414 @@
+#pragma once
+#include "MinimalEngine.h"
+#include "Array.h"
+
+namespace Seele
+{
+template<typename T, Allocator = std::pmr::polymorphic_allocator<T>>
+class RingBuffer
+{
+public:
+    template <typename X>
+    class IteratorBase
+    {
+    public:
+        using iterator_category = std::random_access_iterator_tag;
+        using value_type = X;
+        using difference_type = std::ptrdiff_t;
+        using reference = X&;
+        using pointer = X*;
+
+        IteratorBase(Array<T>& arr, size_t index)
+            : arr(arr)
+            , index(index)
+        {
+        }
+        reference operator*() const
+        {
+            return arr[index];
+        }
+        pointer operator->() const
+        {
+            return &arr[index];
+        }
+        inline bool operator!=(const IteratorBase &other) const
+        {
+            return arr != other.arr && index != other.index;
+        }
+        inline bool operator==(const IteratorBase &other) const
+        {
+            return arr == other.arr && index == other.index;
+        }
+        inline int operator-(const IteratorBase &other) const
+        {
+            return (int)(index - other.index);
+        }
+        IteratorBase &operator++()
+        {
+            index = (index + 1) % arr.size();
+            return *this;
+        }
+        IteratorBase &operator--()
+        {
+            index = (index + 1) % arr.size();
+            return *this;
+        }
+        IteratorBase operator++(int)
+        {
+            IteratorBase tmp(*this);
+            ++*this;
+            return tmp;
+        }
+        IteratorBase operator--(int)
+        {
+            IteratorBase tmp(*this);
+            --*this;
+            return tmp;
+        }
+
+    private:
+        Array<T, Allocator>& arr;
+        size_t index;
+    };
+    
+    using value_type = T;
+    using allocator_type = Allocator;
+    using size_type = std::size_t;
+    using difference_type = std::ptrdiff_t;
+    using pointer = T*;
+    using const_pointer = const T*;
+    using reference = value_type&;
+    using const_reference = const value_type&;
+
+    using Iterator = IteratorBase<T>;
+    using ConstIterator = IteratorBase<const T>;
+    using iterator = Iterator;
+    using const_iterator = ConstIterator;
+    using reverse_iterator = std::reverse_iterator<iterator>;
+    using const_reverse_iterator = std::reverse_iterator<const_iterator>;
+
+    constexpr RingBuffer() noexcept(noexcept(Allocator()))
+    {
+        refreshIterators();
+    }
+
+    constexpr explicit RingBuffer(const allocator_type& alloc)
+        : data(alloc)
+    {
+        refreshIterators();
+    }
+    constexpr RingBuffer(size_type size, const value_type& value, const allocator_type& alloc = allocator_type())
+        : data(size, value, alloc)
+        , end(size)
+    {
+        refreshIterators();
+    }
+    constexpr explicit RingBuffer(size_type size, const allocator_type& alloc = allocator_type())
+        : arraySize(size)
+        , allocated(size)
+        , allocator(alloc)
+    {
+        _data = allocateRingBuffer(size);
+        assert(_data != nullptr);
+        for (size_type i = 0; i < size; ++i)
+        {
+            std::allocator_traits<allocator_type>::construct(allocator, &_data[i]);
+        }
+        markIteratorDirty();
+    }
+    constexpr RingBuffer(std::initializer_list<T> init, const allocator_type& alloc = allocator_type())
+        : arraySize(init.size())
+        , allocated(init.size())
+        , allocator(alloc)
+    {
+        _data = allocateRingBuffer(init.size());
+        assert(_data != nullptr);
+        markIteratorDirty();
+        std::uninitialized_copy(init.begin(), init.end(), begin());
+    }
+    RingBuffer(const RingBuffer &other)
+        : arraySize(other.arraySize)
+        , allocated(other.allocated)
+        , allocator(std::allocator_traits<allocator_type>::select_on_container_copy_construction(other.allocator))
+    {
+        _data = allocateRingBuffer(other.allocated);
+        assert(_data != nullptr);
+        markIteratorDirty();
+        std::uninitialized_copy(other.begin(), other.end(), begin());
+    }
+    RingBuffer(RingBuffer &&other) noexcept
+        : arraySize(std::move(other.arraySize))
+        , allocated(std::move(other.allocated))
+        , allocator(std::move(other.allocator))
+    {
+        _data = other._data;
+        other._data = nullptr;
+        other.allocated = 0;
+        other.arraySize = 0;
+        markIteratorDirty();
+    }
+    RingBuffer &operator=(const RingBuffer &other) noexcept
+    {
+        if (this != &other)
+        {
+            if constexpr (std::allocator_traits<allocator_type>::propagate_on_container_copy_assignment::value )
+            {
+                if (!std::allocator_traits<allocator_type>::is_always_equal::value
+                    && allocator != other.allocator)
+                {
+                    clear();
+                }
+                allocator = other.allocator;
+            }
+            if (other.arraySize > allocated)
+            {
+                clear();
+            }
+            if(_data == nullptr)
+            {
+                _data = allocateRingBuffer(other.allocated);
+                allocated = other.allocated;
+            }
+            arraySize = other.arraySize;
+            markIteratorDirty();
+            std::uninitialized_copy(other.begin(), other.end(), begin());
+        }
+        return *this;
+    }
+    RingBuffer &operator=(RingBuffer &&other) noexcept
+    {
+        if (this != &other)
+        {
+            if constexpr (std::allocator_traits<allocator_type>::propagate_on_container_move_assignment::value)
+            {
+                allocator = std::move(other.allocator);
+            }
+            if (_data != nullptr)
+            {
+                clear();
+            }
+            allocated = std::move(other.allocated);
+            arraySize = std::move(other.arraySize);
+            _data = other._data;
+            other._data = nullptr;
+            markIteratorDirty();
+        }
+        return *this;
+    }
+    ~RingBuffer()
+    {
+        clear();
+    }
+
+    constexpr bool operator==(const RingBuffer &other)
+    {
+        return _data == other._data;
+    }
+
+    constexpr bool operator!=(const RingBuffer &other)
+    {
+        return !(*this == other);
+    }
+
+    constexpr iterator find(const value_type &item)
+    {
+        for (uint32 i = 0; i < arraySize; ++i)
+        {
+            if (_data[i] == item)
+            {
+                return iterator(&_data[i]);
+            }
+        }
+        return endIt;
+    }
+    constexpr iterator find(value_type&& item)
+    {
+        for (uint32 i = 0; i < arraySize; ++i)
+        {
+            if (_data[i] == item)
+            {
+                return iterator(&_data[i]);
+            }
+        }
+        return endIt;	
+    }
+    constexpr allocator_type get_allocator() const
+    {
+        return allocator;
+    }
+    constexpr iterator begin() const
+    {
+        return beginIt;
+    }
+    constexpr iterator end() const
+    {
+        return endIt;
+    }
+    constexpr const_iterator cbegin() const
+    {
+        return beginIt;
+    }
+    constexpr const_iterator cend() const
+    {
+        return endIt;
+    }
+    constexpr reference add(const value_type &item = value_type())
+    {
+        return addInternal(item);
+    }
+    constexpr reference add(value_type&& item)
+    {
+        return addInternal(std::forward<T>(item));
+    }
+    constexpr void addAll(RingBuffer other)
+    {
+        for(auto value : other)
+        {
+            addInternal(value);
+        }
+    }
+    constexpr reference addUnique(const value_type &item = value_type())
+    {
+        iterator it;
+        if((it = std::move(find(item))) != endIt)
+        {
+            return *it;
+        }
+        return addInternal(item);
+    }
+    template<typename... args>
+    constexpr reference emplace(args... arguments)
+    {
+        if (arraySize == allocated)
+        {
+            size_type newSize = arraySize + 1;
+            allocated = calculateGrowth(newSize);
+            T *tempRingBuffer = allocateRingBuffer(allocated);
+            assert(tempRingBuffer != nullptr);
+            
+            std::uninitialized_move(begin(), end(), Iterator(tempRingBuffer));
+            deallocateRingBuffer(_data, arraySize);
+            _data = tempRingBuffer;
+        }
+        std::allocator_traits<allocator_type>::construct(allocator, &_data[arraySize++], arguments...);
+        markIteratorDirty();
+        return _data[arraySize - 1];
+    }
+    constexpr void remove(iterator it, bool keepOrder = true)
+    {
+        remove(it - beginIt, keepOrder);
+    }
+    constexpr void remove(size_type index, bool keepOrder = true)
+    {
+        if (keepOrder)
+        {
+            for(uint32 i = index; i < arraySize-1; ++i)
+            {
+                _data[i] = std::move(_data[i+1]);
+            }
+        }
+        else
+        {
+            _data[index] = std::move(_data[arraySize - 1]);
+        }
+        std::allocator_traits<allocator_type>::destroy(allocator, &_data[--arraySize]);
+        markIteratorDirty();
+    }
+    constexpr void resize(size_type newSize)
+    {
+        resizeInternal(newSize, std::move(T()));
+    }
+    constexpr void resize(size_type newSize, const value_type& value)
+    {
+        resizeInternal(newSize, value);
+    }
+    constexpr void clear()
+    {
+        if(_data == nullptr)
+        {
+            return;
+        }
+        for(size_type i = 0; i < arraySize; ++i)
+        {
+            std::allocator_traits<allocator_type>::destroy(allocator, &_data[i]);
+        }
+        deallocateRingBuffer(_data, allocated);
+        _data = nullptr;
+        arraySize = 0;
+        allocated = 0;
+        markIteratorDirty();
+    }
+    inline size_type indexOf(iterator iterator)
+    {
+        return iterator - beginIt;
+    }
+    inline size_type indexOf(const_iterator iterator) const
+    {
+        return iterator.p - beginIt.p;
+    }
+    inline size_type indexOf(T& t)
+    {
+        return indexOf(find(t));
+    }
+    inline size_type indexOf(const T& t) const
+    {
+        return indexOf(find(t));
+    }
+    inline size_type size() const
+    {
+        return arraySize;
+    }
+    inline size_type empty() const
+    {
+        return arraySize == 0;
+    }
+    inline size_type capacity() const
+    {
+        return allocated;
+    }
+    inline pointer data() const
+    {
+        return _data;
+    }
+    inline reference front() const
+    {
+        assert(arraySize > 0);
+        return _data[0];
+    }
+    inline reference back() const
+    {
+        assert(arraySize > 0);
+        return _data[arraySize - 1];
+    }
+    void pop()
+    {
+        std::allocator_traits<allocator_type>::destroy(allocator, &_data[--arraySize]);
+        markIteratorDirty();
+    }
+    constexpr inline reference operator[](size_type index)
+    {
+        assert(index < arraySize);
+        return _data[index];
+    }
+    constexpr inline const reference operator[](size_type index) const
+    {
+        assert(index < arraySize);
+        return _data[index];
+    }
+private:
+    void refreshIterators()
+    {
+        beginIt = Iterator(data, begin);
+        endIt = Iterator(data, end);
+    }
+    size_type begin = 0;
+    size_type end = 0;
+    iterator beginIt;
+    iterator endIt;
+    Array<T, Allocator> data;
+};
+class StaticRingBuffer
+{
+
+};
+} // namespace Seele

src/Engine/Graphics/Vulkan/VulkanCommandBuffer.cpp

@@ -171,12 +171,6 @@ void CmdBuffer::waitForCommand(uint32 timeout)
     refreshFence();
 }
 
-Event CmdBuffer::asyncWait() const
-{
-    return fence->asyncWait();
-}
-
-
 PFence CmdBuffer::getFence()
 {
     return fence;

src/Engine/Graphics/Vulkan/VulkanCommandBuffer.h

@@ -32,7 +32,10 @@ public:
 	void addWaitSemaphore(VkPipelineStageFlags stages, PSemaphore waitSemaphore);
 	void refreshFence();
 	void waitForCommand(uint32 timeToWait = 1000000u);
-	Event asyncWait() const;
+	Fence* operator co_await()
+	{
+		return fence.getHandle();
+	}
 	PFence getFence();
 	PCommandBufferManager getManager();
 	enum State

src/Engine/Graphics/Vulkan/VulkanGraphicsResources.cpp

@@ -80,11 +80,6 @@ void Fence::wait(uint32 timeout)
         break;
     }
 }
-Event Fence::asyncWait() const
-{
-    return signaled;
-}
-
 
 VertexDeclaration::VertexDeclaration(const Array<Gfx::VertexElement>& elementList) 
     : elementList(elementList)

src/Engine/Graphics/Vulkan/VulkanGraphicsResources.h

@@ -43,7 +43,10 @@ public:
         return fence;
     }
     void wait(uint32 timeout);
-    Event asyncWait() const;
+    Event& operator co_await()
+    {
+        return signaled;
+    }
     bool operator<(const Fence &other) const
     {
         return fence < other.fence;

src/Engine/Math/Transform.cpp

@@ -146,15 +146,15 @@ Vector Transform::getScale() const
 
 Vector Transform::getForward() const
 {
-    return Vector(0, 0, 1) * rotation;
+    return glm::normalize(Vector(0, 0, 1) * rotation);
 }
 Vector Transform::getRight() const
 {
-    return Vector(1, 0, 0) * rotation;
+    return glm::normalize(Vector(1, 0, 0) * rotation);
 }
 Vector Transform::getUp() const
 {
-    return Vector(0, 1, 0) * rotation;
+    return glm::normalize(Vector(0, 1, 0) * rotation);
 }
 
 bool Transform::equals(const Transform &other, float tolerance)

src/Engine/Scene/Actor/CameraActor.cpp

@@ -14,6 +14,7 @@ CameraActor::CameraActor()
     cameraComponent->aspectRatio = 1.777778f;
     cameraComponent->setParent(sceneComponent);
     cameraComponent->setOwner(this);
+    sceneComponent->addChildComponent(cameraComponent);
 }
 
 CameraActor::~CameraActor()

src/Engine/Scene/Components/CameraComponent.cpp

@@ -13,8 +13,8 @@ CameraComponent::CameraComponent()
     , projectionMatrix(Matrix4())
     , viewMatrix(Matrix4())
 {
-    rotationX = 0;
-    rotationY = 0;
+    yaw = 0;
+    pitch = 0;
     setRelativeLocation(Vector(0, 10, -50));
 }
 
@@ -22,12 +22,16 @@ CameraComponent::~CameraComponent()
 {
 }
 
-void CameraComponent::mouseMove(float deltaX, float deltaY) 
+void CameraComponent::mouseMove(float deltaYaw, float deltaPitch) 
 {
-    rotationX -= deltaX / 1000.f;
-    rotationY += deltaY / 1000.f;
-    //std::cout << "X:" << rotationX << " Y: " << rotationY << std::endl;
-    setRelativeRotation(Vector(rotationY, rotationX, 0));
+    yaw -= deltaYaw / 500.f;
+    pitch += deltaPitch / 500.f;
+    //std::cout << "Yaw: " << yaw << " Pitch: " << pitch << std::endl;
+    Vector cameraDirection = glm::normalize(Vector(cos(yaw) * cos(pitch), sin(pitch), sin(yaw) * cos(pitch)));
+    Vector xyz = glm::cross(cameraDirection, Vector(0, 0, 1));
+    Quaternion result = Quaternion(glm::dot(cameraDirection, Vector(0, 0, 1)) + 1, xyz.x, xyz.y, xyz.z);
+    //std::cout << "Result " << Vector(0, 0, 1) * glm::normalize(result) << " cameraDirection: " << cameraDirection << std::endl;
+    setRelativeRotation(result);
     bNeedsViewBuild = true;
 }
 
@@ -51,9 +55,9 @@ void CameraComponent::moveY(float amount)
 
 void CameraComponent::buildViewMatrix() 
 {
-    Vector eyePos = getTransform().getPosition();
-    Vector lookAt = eyePos + getTransform().getForward();
-    std::cout << "Eye: " << eyePos << " lookAt: " << lookAt << std::endl;
+    Vector eyePos = getAbsoluteTransform().getPosition();
+    Vector lookAt = eyePos + glm::normalize(Vector(cos(yaw) * cos(pitch), sin(pitch), sin(yaw) * cos(pitch)));
+    //std::cout << "Eye: " << eyePos << " lookAt: " << lookAt << std::endl;
     viewMatrix = glm::lookAt(eyePos, lookAt, Vector(0, 1, 0));
 
     bNeedsViewBuild = false;

src/Engine/Scene/Components/CameraComponent.h

@@ -45,8 +45,8 @@ private:
     //Transforms relative to actor
     Matrix4 viewMatrix;
     Matrix4 projectionMatrix;
-    float rotationX;
-    float rotationY;
+    float yaw;
+    float pitch;
 };
 DEFINE_REF(CameraComponent)
 } // namespace Seele

src/Engine/Scene/Components/Component.cpp

@@ -128,18 +128,22 @@ void Component::notifySceneAttach(PScene scene)
 void Component::setRelativeLocation(Vector location)
 {
     transform = Transform(location, transform.getRotation(), transform.getScale());
+    propagateTransformUpdate();
 }
 void Component::setRelativeRotation(Vector rotation)
 {
     transform = Transform(transform.getPosition(), Quaternion(rotation), transform.getScale());
+    propagateTransformUpdate();
 }
 void Component::setRelativeRotation(Quaternion rotation)
 {
     transform = Transform(transform.getPosition(), rotation, transform.getScale());
+    propagateTransformUpdate();
 }
 void Component::setRelativeScale(Vector scale)
 {
     transform = Transform(transform.getPosition(), transform.getRotation(), scale);
+    propagateTransformUpdate();
 }
 
 //void Component::addAbsoluteTranslation(Vector translation)
@@ -161,14 +165,17 @@ void Component::setRelativeScale(Vector scale)
 void Component::addRelativeLocation(Vector translation)
 {
     transform = Transform(transform.getPosition() + translation, transform.getRotation(), transform.getScale());
+    propagateTransformUpdate();
 }
 void Component::addRelativeRotation(Vector rotation)
 {
-    transform = Transform(transform.getPosition(), transform.getRotation() + Quaternion(rotation), transform.getScale());
+    transform = Transform(transform.getPosition(), transform.getRotation() * Quaternion(rotation), transform.getScale());
+    propagateTransformUpdate();
 }
 void Component::addRelativeRotation(Quaternion rotation)
 {
-    transform = Transform(transform.getPosition(), transform.getRotation(), transform.getScale());
+    transform = Transform(transform.getPosition(), transform.getRotation() * rotation, transform.getScale());
+    propagateTransformUpdate();
 }
 
 Transform Component::getTransform() const
@@ -177,10 +184,22 @@ Transform Component::getTransform() const
 }
 
 Transform Component::getAbsoluteTransform() const
+{
+    return absoluteTransform;
+}
+
+void Component::propagateTransformUpdate()
 {
     if(parent != nullptr)
     {
-        return transform + parent->getAbsoluteTransform();
+        absoluteTransform = transform + parent->getAbsoluteTransform();
+    }
+    else
+    {
+        absoluteTransform = transform;
+    }
+    for(auto child : children)
+    {
+        child->propagateTransformUpdate();
     }
-    return transform;
 }

src/Engine/Scene/Components/Component.h

@@ -74,7 +74,9 @@ private:
         }
         return nullptr;
     }
+    void propagateTransformUpdate();
     Transform transform;
+    Transform absoluteTransform;
     PScene owningScene;
     PActor owner;
     PComponent parent;

src/Engine/ThreadPool.cpp

@@ -3,55 +3,69 @@
 
 using namespace Seele;
 
-
-Event::Event()
-    : flag(std::make_shared<StateStore>())
-{
-}
+std::mutex Seele::promisesLock;
+List<Promise*> Seele::promises;
 
 Event::Event(nullptr_t)
-    : flag(nullptr)
 {
 }
 
-Event::Event(const std::string &name)
-    : flag(std::make_shared<StateStore>())
+Event::Event(const std::string &name, const std::source_location &location)
+    : name(name)
+    , location(location)
 {
-    flag->name = name;
 }
 
-
 Event::Event(const std::source_location &location)
-    : flag(std::make_shared<StateStore>())
+    : name(location.function_name())
+    , location(location)
 {
-    flag->name = location.function_name();
-    flag->location = location;
 }
 
-
 void Event::raise()
 {
-    std::scoped_lock lock(flag->lock);
-    flag->data = 1;
-    getGlobalThreadPool().notify(*this);
+    std::scoped_lock lock(eventLock);
+    data = true;
+    if(waitingJobs.size() > 0)
+    {
+        getGlobalThreadPool().scheduleBatch(waitingJobs);
+    }
+    if(waitingMainJobs.size() > 0)
+    {
+        getGlobalThreadPool().scheduleBatch(waitingMainJobs);
+    }
 }
 void Event::reset()
 {
-    std::scoped_lock lock(flag->lock);
-    flag->data = 0;
+    std::scoped_lock lock(eventLock);
+    data = false;
 }
 
 bool Event::await_ready()
 {
-    flag->lock.lock();
-    bool result = flag->data;
+    eventLock.lock();
+    bool result = data;
     if(result)
     {
-        flag->lock.unlock();
+        eventLock.unlock();
     }
     return result;
 }
 
+void Event::await_suspend(std::coroutine_handle<JobPromiseBase<false>> h)
+{
+    //h.promise().enqueue(this);
+    waitingJobs.add(JobBase<false>(&h.promise()));
+    eventLock.unlock();
+}
+
+void Event::await_suspend(std::coroutine_handle<JobPromiseBase<true>> h)
+{
+    //h.promise().enqueue(this);
+    waitingMainJobs.add(JobBase<true>(&h.promise()));
+    eventLock.unlock();
+}
+
 ThreadPool::ThreadPool(uint32 threadCount)
     : workers(threadCount)
 {
@@ -59,13 +73,24 @@ ThreadPool::ThreadPool(uint32 threadCount)
     for (uint32 i = 0; i < threadCount; ++i)
     {
         workers[i] = std::thread(&ThreadPool::threadLoop, this);
-        workers[i].detach();
     }
 }
 
 ThreadPool::~ThreadPool()
 {
-    workers.clear();
+    running.store(false);
+    {
+        std::unique_lock lock(mainJobLock);
+        mainJobCV.notify_all();
+    }
+    {
+        std::unique_lock lock(jobQueueLock);
+        jobQueueCV.notify_all();
+    }
+    for(auto& worker : workers)
+    {
+        worker.join();
+    }
 }
 
 void ThreadPool::waitIdle()
@@ -73,108 +98,63 @@ void ThreadPool::waitIdle()
     while(true)
     {
         std::unique_lock lock(numIdlingLock);
-        numIdlingIncr.wait(lock);
         if(numIdling == workers.size())
         {
             return;
         }
+        numIdlingIncr.wait(lock);
     }
 }
-void ThreadPool::enqueueWaiting(Event &event, Promise* job)
+void ThreadPool::scheduleJob(Job job)
 {
-    assert(!job->done());
-    std::scoped_lock lock(waitingLock);
-    //std::cout << "Job " << job->finishedEvent.name << " waiting on event " << event.name << std::endl;
-    waitingJobs[event].add(job);
-    job->addRef();
-}
-void ThreadPool::enqueueWaiting(Event &event, MainPromise* job)
-{
-    assert(!job->done());
-    std::scoped_lock lock(waitingMainLock);
-    //std::cout << job->finishedEvent.name << " waiting on event " << event.name << std::endl;
-    waitingMainJobs[event].add(job);
-    job->addRef();
-}
-void ThreadPool::scheduleJob(Promise* job)
-{
-    assert(!job->done());
+    assert(!job.done());
     std::scoped_lock lock(jobQueueLock);
-    //std::cout << "Queueing job " << job->finishedEvent << std::endl;
-    jobQueue.add(job);
+    jobQueue.add(std::move(job));
     jobQueueCV.notify_one();
-    job->addRef();
 }
-void ThreadPool::scheduleJob(MainPromise* job)
+void ThreadPool::scheduleJob(MainJob job)
 {
-    assert(!job->done());
+    assert(!job.done());
     std::scoped_lock lock(mainJobLock);
-    //std::cout << "Queueing job " << job->finishedEvent << std::endl;
-    mainJobs.add(job);
+    mainJobs.add(std::move(job));
     mainJobCV.notify_one();
-    job->addRef();
 }
-void ThreadPool::notify(Event &event)
-{
-    //std::cout << "Event " << event.name << " raised" << std::endl;
-    {
-        std::scoped_lock lock(jobQueueLock, waitingLock);
-        List<Promise*> jobs = std::move(waitingJobs[event]);
-        waitingJobs.erase(event);
-        for (auto &job : jobs)
-        {
-            //assert(job.id != -1ull);
-            //std::cout << "Waking up " << job->finishedEvent.name << std::endl;
-            job->state = Promise::State::SCHEDULED;
-            jobQueue.add(job);
-            jobQueueCV.notify_one();
-        }
-    }
-    {
-        std::scoped_lock lock(mainJobLock, waitingMainLock);
-        List<MainPromise*> jobs = std::move(waitingMainJobs[event]);
-        waitingMainJobs.erase(event);
-        for (auto &job : jobs)
-        {
-            //assert(job.id != -1ull);
-            //std::cout << "Waking up main " << job->finishedEvent.name << std::endl;
-            job->state = MainPromise::State::SCHEDULED;
-            mainJobs.add(job);
-            mainJobCV.notify_one();
-        }
-    }
-}
-
 void ThreadPool::mainLoop()
 {
-    while(true)
+    while(running.load())
     {
-        MainPromise* job;
+        MainJob job;
         {
             std::unique_lock lock(mainJobLock);
             if(mainJobs.empty())
             {
                 mainJobCV.wait(lock);
             }
-            job = mainJobs.front();
-            mainJobs.popFront();
+            [[likely]]
+            if(!mainJobs.empty())
+            {
+                job = mainJobs.front();
+                mainJobs.popFront();
+            }
+            else
+            {
+                continue;
+            }
         }
-        job->resume();
-        job->removeRef();
+        job.resume();
     }
 }
 
 void ThreadPool::threadLoop()
 {
-    List<Promise*> localQueue;
-    while (true)
+    List<Job> localQueue;
+    while (running.load())
     {
         [[likely]]
         if(!localQueue.empty())
         {
-            Promise* job = localQueue.retrieve();
-            job->resume();
-            job->removeRef();
+            Job job = localQueue.retrieve();
+            job.resume();
         }
         else
         {
@@ -194,7 +174,8 @@ void ThreadPool::threadLoop()
             }
             // take 1/numThreads jobs, maybe make this a parameter that
             // adjusts based on past workload
-            uint32 numTaken = std::max(jobQueue.size() / workers.size(), 1ull);
+            uint32 partitionedWorkload = (uint32)(jobQueue.size() / workers.size());
+            uint32 numTaken = std::clamp(partitionedWorkload, 1u, localQueueSize);
             while (!jobQueue.empty() && localQueue.size() < numTaken)
             {
                 localQueue.add(jobQueue.retrieve());

src/Engine/ThreadPool.h

@@ -10,66 +10,65 @@ namespace Seele
 {
 extern class ThreadPool& getGlobalThreadPool();
 template<bool MainJob>      
+struct JobBase;
+template<bool MainJob>
 struct JobPromiseBase;
 struct Event
 {
 public:
-    Event();
     Event(nullptr_t);
-    Event(const std::string& name);
-    Event(const std::source_location& location);
+    Event(const std::string& name, const std::source_location& location = std::source_location::current());
+    Event(const std::source_location& location = std::source_location::current());
+    Event(const Event& other) = delete;
+    Event(Event&& other) = default;
     ~Event() = default;
+    Event& operator=(const Event& other) = delete;
+    Event& operator=(Event&& other) = default;
     auto operator<=>(const Event& other) const
     {
-        return flag <=> other.flag;
+        return name <=> other.name;
     }
     bool operator==(const Event& other) const
     {
-        return flag == other.flag;
-    }
-    Event operator co_await()
-    {
-        return *this;
+        return name == other.name;
     }
     operator bool()
     {
-        std::scoped_lock lock(flag->lock);
-        return flag->data;
+        std::scoped_lock lock(eventLock);
+        return data;
     }
     
     friend std::ostream& operator<<(std::ostream& stream, const Event& event)
     {
         stream 
-            << event.flag->location.file_name() 
+            << event.location.file_name() 
             << "(" 
-            << event.flag->location.line() 
+            << event.location.line() 
             << ":" 
-            << event.flag->location.column() 
+            << event.location.column() 
             << "): " 
-            << event.flag->location.function_name();
+            << event.location.function_name();
         return stream;
     }
 
     void raise();
     void reset();
     bool await_ready();
-    template<bool MainJob>
-    constexpr void await_suspend(std::coroutine_handle<JobPromiseBase<MainJob>> h);
+    void await_suspend(std::coroutine_handle<JobPromiseBase<false>> h);
+    void await_suspend(std::coroutine_handle<JobPromiseBase<true>> h);
     constexpr void await_resume() {}
 private:
-    struct StateStore
-    {
-        std::mutex lock;
-        std::string name;
-        std::source_location location;
-        bool data;
-    };
-    std::shared_ptr<StateStore> flag;
+    std::mutex eventLock;
+    std::string name;
+    std::source_location location;
+    bool data = false;
+    Array<JobBase<false>> waitingJobs;
+    Array<JobBase<true>> waitingMainJobs;
     friend class ThreadPool;
 };
 
-template<bool MainJob>      
-struct JobBase;
+extern std::mutex promisesLock;
+extern List<JobPromiseBase<false>*> promises;
 template<bool MainJob>
 struct JobPromiseBase
 {
@@ -82,12 +81,24 @@ struct JobPromiseBase
         DONE
     };
     JobPromiseBase(const std::source_location& location = std::source_location::current())
+        : handle(std::coroutine_handle<JobPromiseBase<MainJob>>::from_promise(*this))
+        , finishedEvent(Event(location))
     {
-        handle = std::coroutine_handle<JobPromiseBase<MainJob>>::from_promise(*this);
-        finishedEvent = Event(location);
+        if constexpr(!MainJob)
+        {
+            std::scoped_lock lock(promisesLock);
+            promises.add(this);
+        }
     }
     ~JobPromiseBase()
-    {}
+    {
+        if constexpr (!MainJob)
+        {
+            std::scoped_lock lock(promisesLock);
+            promises.remove(promises.find(this));
+        }
+    }
+
     JobBase<MainJob> get_return_object() noexcept;
 
     inline auto initial_suspend() noexcept;
@@ -102,7 +113,6 @@ struct JobPromiseBase
 
     void resume()
     {
-        std::scoped_lock lock(promiseLock);
         if(!handle || handle.done() || executing())
         {
             return;
@@ -116,17 +126,16 @@ struct JobPromiseBase
         finishedEvent.raise();
         if(continuation)
         {
-            std::scoped_lock lock(continuation->promiseLock);
-            getGlobalThreadPool().scheduleJob(continuation);
+            getGlobalThreadPool().scheduleJob(JobBase<MainJob>(continuation));
             continuation->removeRef();
         }
     }
     void setContinuation(JobPromiseBase* cont)
     {
-        std::scoped_lock lock(promiseLock, cont->promiseLock);
         assert(cont->ready());
         continuation = cont;
         cont->state = State::SCHEDULED;
+        cont->waitingFor = &finishedEvent;
         cont->addRef();
     }
     bool done()
@@ -149,29 +158,25 @@ struct JobPromiseBase
     {
         return state == State::READY;
     }
-    void reset()
-    {
-        std::scoped_lock lock(promiseLock);
-        finishedEvent.reset();
-    }
-    void enqueue(Event& event)
+    void enqueue(Event* event)
     {
         if(!handle || handle.done() || waiting() || scheduled())
         {
             return;
         }
         state = State::WAITING;
-        getGlobalThreadPool().enqueueWaiting(event, this);
+        waitingFor = event;
+        getGlobalThreadPool().enqueueWaiting(event, std::move(JobBase<MainJob>(this)));
     }
-    void schedule()
+    bool schedule()
     {
-        std::scoped_lock lock(promiseLock);
         if(!handle || done() || !ready())
         {
-            return;
+            return false;
         }
         state = State::SCHEDULED;
-        getGlobalThreadPool().scheduleJob(this);
+        getGlobalThreadPool().scheduleJob(std::move(JobBase<MainJob>(this)));
+        return true;
     }
     void addRef()
     {
@@ -179,25 +184,23 @@ struct JobPromiseBase
     }
     void removeRef()
     {
-        if(--numRefs < 1)
+        numRefs--;
+        if(numRefs == 0)
         {
-            if(done())
+            if(!schedule())
             {
                 handle.destroy();
             }
-            else
-            {
-                schedule();
-            }
         } 
     }
-
-    std::mutex promiseLock;
+    uint64 pad0 = 0x7472617453;
     std::coroutine_handle<JobPromiseBase> handle;
+    Event* waitingFor = nullptr;
     JobPromiseBase* continuation = nullptr;
-    std::atomic_uint64_t numRefs = 0;
+    uint64 numRefs = 0;
     Event finishedEvent;
     State state = State::READY;
+    uint64 pad1 = 0x646E45;
 };
 
 template<bool MainJob = false>
@@ -213,16 +216,15 @@ public:
     explicit JobBase(JobPromiseBase<MainJob>* promise)
         : promise(promise)
     {
+        promise->addRef();
     }
     JobBase(const JobBase& other)
     {
-        std::scoped_lock lock(other.promise->promiseLock);
         promise = other.promise;
         promise->addRef();
     }
     JobBase(JobBase&& other)
     {
-        std::scoped_lock lock(other.promise->promiseLock);
         promise = other.promise;
         other.promise = nullptr;
     }
@@ -238,7 +240,10 @@ public:
     {
         if(this != &other)
         {
-            std::scoped_lock lock(other.promise->promiseLock);
+            if(promise != nullptr)
+            {
+                promise->removeRef();
+            }
             promise = other.promise;
             promise->addRef();
         }
@@ -248,7 +253,10 @@ public:
     {
         if(this != &other)
         {
-            std::scoped_lock lock(other.promise->promiseLock);
+            if(promise != nullptr)
+            {
+                promise->removeRef();
+            }
             promise = other.promise;
             other.promise = nullptr;
         }
@@ -273,7 +281,7 @@ public:
     {
         return promise->done();
     }
-    Event operator co_await() const
+    Event& operator co_await()
     {
         // the co_await operator keeps a reference to this, it won't 
         // be scheduled from the destructor
@@ -311,8 +319,7 @@ public:
         Array<JobBase> jobs;
         for(auto&& param : params)
         {
-            JobBase base = func(param);
-            jobs.add(base);
+            jobs.add(func(param));
         }
         getGlobalThreadPool().scheduleBatch(jobs);
         for(auto job : jobs)
@@ -333,15 +340,11 @@ using Promise = JobPromiseBase<false>;
 class ThreadPool
 {
 public:
-    ThreadPool(uint32 threadCount = std::thread::hardware_concurrency());
+    ThreadPool(uint32 threadCount = 1);//std::thread::hardware_concurrency());
     virtual ~ThreadPool();
     void waitIdle();
-    // Adds a job to the waiting queue for event
-    void enqueueWaiting(Event& event, Promise* job);
-    // Adds a job to the waiting queue for event
-    void enqueueWaiting(Event& event, MainPromise* job);
-    void scheduleJob(Promise* job);
-    void scheduleJob(MainPromise* job);
+    void scheduleJob(Job job);
+    void scheduleJob(MainJob job);
     template<std::ranges::range Iterable>
     requires std::same_as<std::ranges::range_value_t<Iterable>, MainJob>
     void scheduleBatch(Iterable jobs)
@@ -349,9 +352,9 @@ public:
         std::scoped_lock lock(mainJobLock);
         for(auto job : jobs)
         {
-            job.promise->addRef();
+            //job.promise->addRef();
             job.promise->state = JobPromiseBase<true>::State::SCHEDULED;
-            mainJobs.add(job.promise);
+            mainJobs.add(job);
         }
         mainJobCV.notify_one();
     }
@@ -362,13 +365,13 @@ public:
         std::scoped_lock lock(jobQueueLock);
         for(auto job : jobs)
         {
-            job.promise->addRef();
+            //job.promise->addRef();
             job.promise->state = JobPromiseBase<false>::State::SCHEDULED;
-            jobQueue.add(job.promise);
+            jobQueue.add(job);
         }
         jobQueueCV.notify_all();
     }
-    void notify(Event& event);
+    void notify(Event* event);
     void mainLoop();
     void threadLoop();
 private:
@@ -378,27 +381,20 @@ private:
     uint32 numIdling;
     Array<std::thread> workers;
 
-    List<MainPromise*> mainJobs;
+    List<MainJob> mainJobs;
     std::mutex mainJobLock;
     std::condition_variable mainJobCV;
     
-    List<Promise*> jobQueue;
+    List<Job> jobQueue;
     std::mutex jobQueueLock;
     std::condition_variable jobQueueCV;
 
-    Map<Event, List<MainPromise*>> waitingMainJobs;
-    std::mutex waitingMainLock;
-
-    Map<Event, List<Promise*>> waitingJobs;
-    std::mutex waitingLock;
-
-    uint32 maxLocalQueueSize = 50;
+    uint32 localQueueSize = 50;
 };
 
 template<bool MainJob>
 inline JobBase<MainJob> JobPromiseBase<MainJob>::get_return_object() noexcept 
 {
-    numRefs++;
     return JobBase<MainJob>(this);
 }
 
@@ -415,11 +411,4 @@ inline auto JobPromiseBase<MainJob>::final_suspend() noexcept
     return std::suspend_always{};
 }
 
-template<bool MainJob>
-inline constexpr void Event::await_suspend(std::coroutine_handle<JobPromiseBase<MainJob>> h)
-{
-    h.promise().enqueue(*this);
-    flag->lock.unlock();
-}
-
 } // namespace Seele

src/Engine/Window/View.cpp

@@ -25,3 +25,8 @@ void View::setFocused()
 {
 	owner->setFocused(this);	
 }
+
+const std::string& View::getName()
+{
+	return name;
+}

src/Engine/Window/View.h

@@ -25,6 +25,8 @@ public:
 	void applyArea(URect area);
 	void setFocused();
 
+	const std::string& getName();
+
 protected:
 	Gfx::PGraphics graphics;
 	Gfx::PViewport viewport;

src/Engine/Window/Window.cpp

@@ -16,9 +16,7 @@ Window::~Window()
 
 void Window::addView(PView view)
 {
-    WindowView* windowView = new WindowView();
-    windowView->view = view;
-    windowView->updateFinished = Event(view->name);
+    WindowView* windowView = new WindowView(view);
     //windowView->worker = std::thread(&Window::viewWorker, this, windowView);
     views.add(windowView);
     viewWorker(views.size() - 1);

src/Engine/Window/Window.h

@@ -10,6 +10,10 @@ struct WindowView
     PView view;
     Event updateFinished;
     std::mutex workerMutex;
+    WindowView(PView view)
+        : view(view)
+        , updateFinished(view->getName())
+    {}
 };
 DEFINE_REF(WindowView)
 DECLARE_REF(WindowManager)

test/Engine/EngineTest.cpp

@@ -1,112 +1,120 @@
 #include "EngineTest.h"
 #include "MinimalEngine.h"
+#include "ThreadPool.h"
 #define BOOST_TEST_MODULE SeeleEngine
 #include <boost/test/unit_test.hpp>
+//#include <vld.h>
 
 using namespace Seele;
+
+Seele::GlobalFixture::~GlobalFixture()
+{
+    getGlobalThreadPool().waitIdle();
+}
+
 BOOST_TEST_GLOBAL_FIXTURE(GlobalFixture);
 BOOST_AUTO_TEST_SUITE(RefPtr)
 
 struct TestStruct
 {
-	TestStruct()
-		: data(10)
-	{
-	}
-	virtual ~TestStruct()
-	{
-		data = 15;
-	}
-	uint32 data;
+    TestStruct()
+        : data(10)
+    {
+    }
+    virtual ~TestStruct()
+    {
+        data = 15;
+    }
+    uint32 data;
 };
 
 struct DeclStruct;
 BOOST_AUTO_TEST_CASE(basic_refcount)
 {
-	{
-		Seele::RefPtr<TestStruct> ptr = new TestStruct();
-		BOOST_REQUIRE_EQUAL(ptr->data, 10);
-		{
-			Seele::RefPtr<TestStruct> secondPtr = ptr;
-			BOOST_REQUIRE_EQUAL(secondPtr->data, 10);
-			BOOST_REQUIRE_EQUAL(ptr->data, 10);
-		}
-		BOOST_REQUIRE_EQUAL(ptr->data, 10);
-	}
+    {
+        Seele::RefPtr<TestStruct> ptr = new TestStruct();
+        BOOST_REQUIRE_EQUAL(ptr->data, 10);
+        {
+            Seele::RefPtr<TestStruct> secondPtr = ptr;
+            BOOST_REQUIRE_EQUAL(secondPtr->data, 10);
+            BOOST_REQUIRE_EQUAL(ptr->data, 10);
+        }
+        BOOST_REQUIRE_EQUAL(ptr->data, 10);
+    }
 }
 struct DeclStruct
 {
-	~DeclStruct()
-	{
-		data = 10;
-	}
-	uint32 data = 20;
+    ~DeclStruct()
+    {
+        data = 10;
+    }
+    uint32 data = 20;
 };
 
 struct DerivedStruct : public TestStruct
 {
-	DerivedStruct()
-		:data2(20)
-	{
+    DerivedStruct()
+        :data2(20)
+    {
 
-	}
-	~DerivedStruct()
-	{
-		data2 = 30;
-	}
-	uint32 data2;
+    }
+    ~DerivedStruct()
+    {
+        data2 = 30;
+    }
+    uint32 data2;
 };
 
 BOOST_AUTO_TEST_CASE(inheritance_cast)
 {
-	Seele::RefPtr<DerivedStruct> backCast;
-	{
-		Seele::RefPtr<DerivedStruct> derived = new DerivedStruct();
-		Seele::RefPtr<TestStruct> base = derived;
-		BOOST_REQUIRE_EQUAL(base->data, 10);
-		backCast = base.cast<DerivedStruct>();
-		BOOST_REQUIRE_EQUAL(backCast->data, 10);
-		BOOST_REQUIRE_EQUAL(backCast->data2, 20);
-	}
-	BOOST_REQUIRE_EQUAL(backCast->data, 10);
-	BOOST_REQUIRE_EQUAL(backCast->data2, 20);
+    Seele::RefPtr<DerivedStruct> backCast;
+    {
+        Seele::RefPtr<DerivedStruct> derived = new DerivedStruct();
+        Seele::RefPtr<TestStruct> base = derived;
+        BOOST_REQUIRE_EQUAL(base->data, 10);
+        backCast = base.cast<DerivedStruct>();
+        BOOST_REQUIRE_EQUAL(backCast->data, 10);
+        BOOST_REQUIRE_EQUAL(backCast->data2, 20);
+    }
+    BOOST_REQUIRE_EQUAL(backCast->data, 10);
+    BOOST_REQUIRE_EQUAL(backCast->data2, 20);
 }
 
 BOOST_AUTO_TEST_CASE(unique_ptr)
 {
-	Seele::UniquePtr<TestStruct> uptr = new TestStruct();
-	Seele::UniquePtr<TestStruct> uptr2 = std::move(uptr);
-	BOOST_REQUIRE_EQUAL(uptr2->data, 10);
-	BOOST_REQUIRE_EQUAL(uptr.isValid(), false);
+    Seele::UniquePtr<TestStruct> uptr = new TestStruct();
+    Seele::UniquePtr<TestStruct> uptr2 = std::move(uptr);
+    BOOST_REQUIRE_EQUAL(uptr2->data, 10);
+    BOOST_REQUIRE_EQUAL(uptr.isValid(), false);
 }
 struct ThisReference
 {
-	ThisReference()
-	{
-	}
-	~ThisReference()
-	{
-		data = 12;
-	}
-	Seele::RefPtr<ThisReference> getRef()
-	{
-		return this;
-	}
-	uint32 data = 1;
+    ThisReference()
+    {
+    }
+    ~ThisReference()
+    {
+        data = 12;
+    }
+    Seele::RefPtr<ThisReference> getRef()
+    {
+        return this;
+    }
+    uint32 data = 1;
 };
 
 BOOST_AUTO_TEST_CASE(this_reference)
 {
-	Seele::RefPtr<ThisReference> ptr2;
-	{
-		Seele::RefPtr<ThisReference> ptr = new ThisReference();
-		BOOST_REQUIRE_EQUAL(ptr->data, 1);
-		ptr2 = ptr->getRef();
-		BOOST_REQUIRE_EQUAL(ptr2->data, 1);
-		ptr2->data = 5;
-		BOOST_REQUIRE_EQUAL(ptr->data, 5);
-	}
-	BOOST_REQUIRE_EQUAL(ptr2->data, 5);
+    Seele::RefPtr<ThisReference> ptr2;
+    {
+        Seele::RefPtr<ThisReference> ptr = new ThisReference();
+        BOOST_REQUIRE_EQUAL(ptr->data, 1);
+        ptr2 = ptr->getRef();
+        BOOST_REQUIRE_EQUAL(ptr2->data, 1);
+        ptr2->data = 5;
+        BOOST_REQUIRE_EQUAL(ptr->data, 5);
+    }
+    BOOST_REQUIRE_EQUAL(ptr2->data, 5);
 }
 
 BOOST_AUTO_TEST_SUITE_END()

test/Engine/EngineTest.h

@@ -1,6 +1,5 @@
 #pragma once
 #include <vld.h>
-#include "ThreadPool.h"
 
 namespace Seele
 {
@@ -10,10 +9,7 @@ namespace Seele
         {
             //_CrtSetDbgFlag(_CRTDBG_ALLOC_MEM_DF | _CRTDBG_LEAK_CHECK_DF);
         }
-        ~GlobalFixture()
-        {
-            getGlobalThreadPool().waitIdle();
-        }
+        ~GlobalFixture();
         void setup()
         {
             //Fibers::JobQueue::initJobQueues();

test/Engine/ThreadPool.cpp

@@ -42,6 +42,7 @@ Job basicAwaitBase()
     basicAwaitState = 25;
     co_await basicAwaitThird();
     BOOST_REQUIRE_EQUAL(basicAwaitState, 30);
+    co_return;
 }
 
 BOOST_AUTO_TEST_CASE(basic_coawait)
@@ -211,7 +212,7 @@ Job launchStressTest()
 
 BOOST_AUTO_TEST_CASE(stress_test)
 {
-    //launchStressTest();
+    launchStressTest();
 }
 
 BOOST_AUTO_TEST_SUITE_END()