ContextPool support for unordered contexts.

aster/CMakeLists.txt

@@ -10,6 +10,7 @@ find_package(fmt CONFIG REQUIRED)
 find_package(VulkanMemoryAllocator CONFIG REQUIRED)
 find_package(EASTL CONFIG REQUIRED)
 find_library(slang NAMES "slang" CONFIG REQUIRED)
+find_package(foonathan_memory CONFIG REQUIRED)
 
 add_library(aster_core STATIC)
 
@@ -28,4 +29,5 @@ target_link_libraries(aster_core PRIVATE fmt::fmt)
 target_link_libraries(aster_core PRIVATE EASTL)
 target_link_libraries(aster_core PUBLIC Vulkan::Headers GPUOpen::VulkanMemoryAllocator)
 target_link_libraries(aster_core PUBLIC ${slang})
+target_link_libraries(aster_core PRIVATE foonathan_memory)
 

aster/include/aster/core/global.h

@@ -13,6 +13,7 @@
 #include <glm/glm.hpp>
 
 #include <fmt/format.h>
+#include <EASTL/shared_ptr.h>
 
 // Macros that can collide with functions.
 #if defined(max)
@@ -252,7 +253,7 @@ struct fmt::formatter<eastl::fixed_string<TType, TCount, TOverflow>> : nested_fo
 };
 
 template <typename T>
-using Ref = std::shared_ptr<T>;
+using Ref = eastl::shared_ptr<T>;
 
 template <typename T>
-using WeakRef = std::weak_ptr<T>;
+using WeakRef = eastl::weak_ptr<T>;

aster/include/aster/systems/CMakeLists.txt

@@ -6,4 +6,5 @@ target_sources(aster_core
 INTERFACE
 	"device.h"
 	"resource.h"
+	"context.h"
 	"commit_manager.h")

aster/include/aster/systems/context.h

@@ -0,0 +1,356 @@
+// =============================================
+//  Aster: context_pool.h
+//  Copyright (c) 2020-2025 Anish Bhobe
+// =============================================
+
+#pragma once
+
+#include "context.h"
+
+#include <aster/aster.h>
+
+#include <aster/core/buffer.h>
+#include <aster/core/image.h>
+#include <aster/core/image_view.h>
+#include <aster/core/physical_device.h>
+#include <aster/core/pipeline.h>
+
+#include <EASTL/intrusive_list.h>
+#include <EASTL/optional.h>
+#include <EASTL/vector.h>
+
+#include <foonathan/memory/memory_pool.hpp>
+#include <foonathan/memory/namespace_alias.hpp>
+
+namespace systems
+{
+
+class Device;
+struct Frame;
+
+namespace _internal
+{
+class GraphicsContextPool;
+class TransferContextPool;
+class ContextPool;
+class OrderlessTransferContextPool;
+} // namespace _internal
+
+#define DEPRECATE_RAW_CALLS
+
+class Context
+{
+  protected:
+    _internal::ContextPool *m_Pool;
+    vk::CommandBuffer m_Cmd;
+
+    friend Device;
+    friend _internal::ContextPool;
+    friend _internal::OrderlessTransferContextPool;
+
+    explicit Context(_internal::ContextPool &pool, const vk::CommandBuffer cmd)
+        : m_Pool{&pool}
+        , m_Cmd{cmd}
+    {
+    }
+
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<Buffer> &buffer);
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<Image> &image);
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<ImageView> &view);
+
+  public:
+    DEPRECATE_RAW_CALLS void Dependency(const vk::DependencyInfo &dependencyInfo);
+
+    void Begin();
+    void End();
+};
+
+class TransferContext : public Context
+{
+  protected:
+    friend Device;
+    friend _internal::TransferContextPool;
+
+    explicit TransferContext(_internal::ContextPool &pool, const vk::CommandBuffer cmd)
+        : Context{pool, cmd}
+    {
+    }
+
+  public:
+    struct ImageData
+    {
+        void *m_Data;
+        usize m_NumBytes;
+    };
+
+    void UploadTexture(const Ref<Image> &image, const ImageData &data);
+
+    TransferContext(TransferContext &&other) noexcept;
+    TransferContext &operator=(TransferContext &&other) noexcept;
+
+    ~TransferContext() = default;
+
+    DISALLOW_COPY_AND_ASSIGN(TransferContext);
+};
+
+class GraphicsContext : public Context
+{
+  protected:
+    friend Device;
+    friend _internal::GraphicsContextPool;
+
+    const Pipeline *m_PipelineInUse;
+
+    explicit GraphicsContext(_internal::ContextPool &pool, const vk::CommandBuffer cmd)
+        : Context{pool, cmd}
+        , m_PipelineInUse{nullptr}
+    {
+    }
+
+  public:
+    DEPRECATE_RAW_CALLS void SetViewport(const vk::Viewport &viewport);
+    void BindVertexBuffer(const Ref<VertexBuffer> &vertexBuffer);
+    void BindPipeline(const Pipeline &pipeline);
+    void
+    PushConstantBlock(auto &block)
+    {
+        if constexpr (sizeof(block) > 128)
+        {
+            WARN("Vulkan only guarantees 128 bytes of Push Constants. Size of PCB is {}", sizeof block);
+        }
+        m_Cmd.pushConstants(m_PipelineInUse->m_Layout, vk::ShaderStageFlagBits::eAll, 0, sizeof block, &block);
+    }
+    void Draw(usize vertexCount);
+    void DrawIndexed(usize indexCount);
+
+    DEPRECATE_RAW_CALLS void BeginRendering(const vk::RenderingInfo &renderingInfo);
+    void EndRendering();
+};
+
+namespace _internal
+{
+
+class ContextPool
+{
+  protected:
+    Device *m_Device;
+    vk::CommandPool m_Pool;
+    eastl::vector<vk::CommandBuffer> m_CommandBuffers;
+    u32 m_BuffersAllocated;
+
+  public:
+    u16 m_ExtraData;
+
+    enum class ManagedBy : u8
+    {
+        eFrame,
+        eDevice,
+    } m_ManagedBy;
+
+  protected:
+    eastl::vector<Ref<Buffer>> m_OwnedBuffers;
+    eastl::vector<Ref<Image>> m_OwnedImages;
+    eastl::vector<Ref<ImageView>> m_OwnedImageViews;
+
+    vk::CommandBuffer AllocateCommandBuffer();
+
+  public:
+    [[nodiscard]] Device &
+    GetDevice() const
+    {
+        assert(m_Device);
+        return *m_Device;
+    }
+
+    eastl::function<void(ContextPool &)> m_ResetCallback;
+
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<Buffer> &buffer);
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<Image> &image);
+    /// Keep the resource alive while the command buffers are acting.
+    void KeepAlive(const Ref<ImageView> &view);
+
+    Context CreateContext();
+
+    void Reset();
+
+    ContextPool() = default;
+    ContextPool(Device &device, u32 queueFamilyIndex, ManagedBy managedBy);
+
+    ContextPool(ContextPool &&other) noexcept;
+    ContextPool &operator=(ContextPool &&other) noexcept;
+
+    bool
+    operator==(const ContextPool &other) const
+    {
+        return m_Pool == other.m_Pool;
+    }
+
+    ~ContextPool();
+
+    DISALLOW_COPY_AND_ASSIGN(ContextPool);
+};
+
+class TransferContextPool : public ContextPool
+{
+  public:
+    TransferContext CreateTransferContext();
+
+    TransferContextPool() = default;
+
+    TransferContextPool(Device &device, const u32 queueFamilyIndex, const ManagedBy managedBy)
+        : ContextPool{device, queueFamilyIndex, managedBy}
+    {
+    }
+
+    TransferContextPool(TransferContextPool &&other) noexcept = default;
+    TransferContextPool &operator=(TransferContextPool &&other) noexcept = default;
+
+    ~TransferContextPool() = default;
+
+    DISALLOW_COPY_AND_ASSIGN(TransferContextPool);
+};
+
+class GraphicsContextPool : public TransferContextPool
+{
+  public:
+    GraphicsContext CreateGraphicsContext();
+
+    GraphicsContextPool() = default;
+    GraphicsContextPool(Device &device, const u32 queueFamilyIndex, const ManagedBy managedBy)
+        : TransferContextPool{device, queueFamilyIndex, managedBy}
+    {
+    }
+
+    GraphicsContextPool(GraphicsContextPool &&other) noexcept = default;
+    GraphicsContextPool &operator=(GraphicsContextPool &&other) noexcept = default;
+
+    ~GraphicsContextPool() = default;
+
+    DISALLOW_COPY_AND_ASSIGN(GraphicsContextPool);
+};
+
+class OrderlessTransferContextPool
+{
+    struct TransferContextEntry : eastl::intrusive_list_node
+    {
+        TransferContextPool m_Pool;
+
+        bool
+        Contains(const _internal::ContextPool &other) const
+        {
+            return m_Pool == other;
+        }
+    };
+
+    using TransferContextList = eastl::intrusive_list<TransferContextEntry>;
+
+    Device *m_Device;
+    memory::memory_pool<> m_TransferContextMemoryPool;
+    TransferContextList m_FreeTransferContexts;
+    TransferContextList m_TransferContexts;
+    u32 m_QueueFamilyIndex;
+
+    constexpr static usize ENTRY_SIZE = sizeof(TransferContextEntry);
+    constexpr static usize ENTRIES_PER_BLOCK = 5;
+    constexpr static usize BLOCK_SIZE = ENTRIES_PER_BLOCK * ENTRY_SIZE;
+
+  public:
+    OrderlessTransferContextPool()
+        : m_Device{nullptr}
+        , m_TransferContextMemoryPool{ENTRY_SIZE, BLOCK_SIZE}
+        , m_QueueFamilyIndex{0}
+    {
+    }
+
+    void
+    Init(Device &device, const u32 queueFamilyIndex)
+    {
+        m_Device = &device;
+        m_QueueFamilyIndex = queueFamilyIndex;
+    }
+
+    TransferContext
+    CreateTransferContext()
+    {
+        if (!m_FreeTransferContexts.empty())
+        {
+            TransferContextEntry &entry = m_FreeTransferContexts.back();
+            m_FreeTransferContexts.pop_back();
+            m_TransferContexts.push_back(entry);
+
+            return entry.m_Pool.CreateTransferContext();
+        }
+
+        TransferContextEntry &entry = *static_cast<TransferContextEntry *>(m_TransferContextMemoryPool.allocate_node());
+        auto pool = TransferContextPool{*m_Device, m_QueueFamilyIndex, ContextPool::ManagedBy::eDevice};
+        pool.m_ResetCallback = [this](ContextPool &resetPool) { this->ReleasePool(resetPool); };
+        new (&entry) TransferContextEntry{
+            .m_Pool = eastl::move(pool),
+        };
+        m_TransferContexts.push_back(entry);
+
+        return entry.m_Pool.CreateTransferContext();
+    }
+
+    void
+    ReleasePool(ContextPool &pool)
+    {
+        auto const found = eastl::find_if(m_TransferContexts.begin(), m_TransferContexts.end(),
+                                          [&pool](const TransferContextEntry &v) { return v.Contains(pool); });
+        auto &v = *found;
+        TransferContextList::remove(v);
+
+        pool.Reset();
+
+        m_FreeTransferContexts.push_back(v);
+    }
+
+    OrderlessTransferContextPool(OrderlessTransferContextPool &&other) noexcept
+        : m_Device{other.m_Device}
+        , m_TransferContextMemoryPool{std::move(other.m_TransferContextMemoryPool)}
+        , m_FreeTransferContexts{other.m_FreeTransferContexts}
+        , m_TransferContexts{other.m_TransferContexts}
+        , m_QueueFamilyIndex{other.m_QueueFamilyIndex}
+    {
+        other.m_FreeTransferContexts.clear();
+        other.m_TransferContexts.clear();
+    }
+
+    OrderlessTransferContextPool &
+    operator=(OrderlessTransferContextPool &&other) noexcept
+    {
+        if (this == &other)
+            return *this;
+        m_Device = other.m_Device;
+        m_TransferContextMemoryPool = std::move(other.m_TransferContextMemoryPool);
+        m_FreeTransferContexts = other.m_FreeTransferContexts;
+        other.m_FreeTransferContexts.clear();
+        m_TransferContexts = other.m_TransferContexts;
+        other.m_TransferContexts.clear();
+        m_QueueFamilyIndex = other.m_QueueFamilyIndex;
+        return *this;
+    }
+
+    ~OrderlessTransferContextPool()
+    {
+        for (auto &entry : m_FreeTransferContexts)
+        {
+            entry.m_Pool.~TransferContextPool();
+        }
+        for (auto &entry : m_TransferContexts)
+        {
+            entry.m_Pool.~TransferContextPool();
+        }
+        // The allocations will 'wink' away.
+    }
+
+    DISALLOW_COPY_AND_ASSIGN(OrderlessTransferContextPool);
+};
+
+} // namespace _internal
+} // namespace systems

aster/include/aster/systems/device.h

@@ -5,6 +5,7 @@
 
 #pragma once
 
+#include "context.h"
 #include "pipeline_helpers.h"
 #include "resource.h"
 
@@ -20,15 +21,13 @@
 #include "aster/core/size.h"
 #include "aster/core/swapchain.h"
 
-#include "EASTL/deque.h"
 #include <EASTL/hash_map.h>
 #include <EASTL/optional.h>
+#include <EASTL/variant.h>
 
 #include <slang-com-ptr.h>
 #include <slang.h>
 
-#include <variant>
-
 constexpr static u32 MAX_FRAMES_IN_FLIGHT = 3;
 
 struct Window;
@@ -343,127 +342,34 @@ class Receipt
     friend _internal::SyncServer;
 };
 
-class Device;
-struct Frame;
-
-#define DEPRECATE_RAW_CALLS
-
-class Context
-{
-  protected:
-    vk::CommandBuffer m_Cmd;
-
-    friend Device;
-    friend Frame;
-
-    explicit Context(const vk::CommandBuffer cmd)
-        : m_Cmd{cmd}
-    {
-    }
-
-  public:
-    DEPRECATE_RAW_CALLS void Dependency(const vk::DependencyInfo &dependencyInfo);
-
-    void Begin();
-    void End();
-};
-
-class GraphicsContext : public Context
-{
-  protected:
-    friend Device;
-    friend Frame;
-
-    const Pipeline *m_PipelineInUse;
-
-    explicit GraphicsContext(const vk::CommandBuffer cmd)
-        : Context{cmd}
-        , m_PipelineInUse{nullptr}
-    {
-    }
-
-  public:
-    DEPRECATE_RAW_CALLS void SetViewport(const vk::Viewport &viewport);
-    void BindVertexBuffer(const Ref<VertexBuffer> &vertexBuffer);
-    void BindPipeline(const Pipeline &pipeline);
-    void
-    PushConstantBlock(auto &block)
-    {
-        if constexpr (sizeof(block) > 128)
-        {
-            WARN("Vulkan only guarantees 128 bytes of Push Constants. Size of PCB is {}", sizeof block);
-        }
-        m_Cmd.pushConstants(m_PipelineInUse->m_Layout, vk::ShaderStageFlagBits::eAll, 0, sizeof block, &block);
-    }
-    void Draw(usize vertexCount);
-    void DrawIndexed(usize indexCount);
-
-    DEPRECATE_RAW_CALLS void BeginRendering(const vk::RenderingInfo &renderingInfo);
-    void EndRendering();
-};
-
-class TransferContext : public Context
-{
-  protected:
-    friend Device;
-    friend Frame;
-
-    Device *m_Device;
-
-    explicit TransferContext(Device &device, const vk::CommandBuffer cmd)
-        : Context{cmd}
-        , m_Device{&device}
-    {
-    }
-
-    eastl::vector<Ref<Buffer>> m_OwnedBuffers;
-    eastl::vector<Ref<Image>> m_OwnedImages;
-    void Reset();
-
-  public:
-    struct ImageData
-    {
-        void *m_Data;
-        usize m_NumBytes;
-    };
-
-    void UploadTexture(const Ref<Image> &image, const ImageData &data);
-
-    TransferContext(TransferContext &&other) noexcept;
-    TransferContext &operator=(TransferContext &&other) noexcept;
-
-    DISALLOW_COPY_AND_ASSIGN(TransferContext);
-
-    ~TransferContext() = default;
-};
-
 struct Frame
 {
     // Persistent
     Device *m_Device;
+
     // TODO: ThreadSafe
-    vk::CommandPool m_Pool;
+    _internal::GraphicsContextPool m_PrimaryPool;
+    _internal::TransferContextPool m_AsyncTransferPool;
+    _internal::ContextPool m_AsyncComputePool;
+
     vk::Fence m_FrameAvailableFence;
     vk::Semaphore m_ImageAcquireSem;
     vk::Semaphore m_RenderFinishSem;
     u32 m_FrameIdx;
 
-    eastl::vector<vk::CommandBuffer> m_CommandBuffers;
-
     // Transient
     vk::Image m_SwapchainImage;
     vk::ImageView m_SwapchainImageView;
     Size2D m_SwapchainSize;
     u32 m_ImageIdx;
-    u32 m_CommandBuffersAllocated;
 
     void Reset(u32 imageIdx, vk::Image swapchainImage, vk::ImageView swapchainImageView, Size2D swapchainSize);
     GraphicsContext CreateGraphicsContext();
-    TransferContext CreateTransferContext();
+    TransferContext CreateAsyncTransferContext();
     void WaitUntilReady();
 
     Frame() = default;
-    Frame(Device &device, u32 primaryQueueFamily, u32 frameIndex);
+    Frame(Device &device, u32 frameIndex, u32 primaryQueueFamily, u32 asyncTransferQueue, u32 asyncComputeQueue);
     Frame(Frame &&other) noexcept;
 
     Frame &operator=(Frame &&other) noexcept;
@@ -486,12 +392,17 @@ class Device final
     std::unique_ptr<CommitManager> m_CommitManager;
 
     // TODO: This is single-threaded.
-    vk::Queue m_GraphicsQueue;
+    vk::Queue m_PrimaryQueue;
     u32 m_PrimaryQueueFamily;
 
     vk::Queue m_TransferQueue;
     u32 m_TransferQueueFamily;
 
+    vk::Queue m_ComputeQueue;
+    u32 m_ComputeQueueFamily;
+
+    _internal::OrderlessTransferContextPool m_TransferContextPool;
+
     std::array<Frame, MAX_FRAMES_IN_FLIGHT> m_Frames;
     u32 m_CurrentFrameIdx = 0;
 
@@ -603,9 +514,6 @@ class Device final
     // Frames
     // ----------------------------------------------------------------------------------------------------
 
-  private:
-    Frame CreateFrame(u32 frameIndex);
-
   public:
     Frame &GetNextFrame();
     Size2D
@@ -623,11 +531,7 @@ class Device final
     friend GraphicsContext;
     friend TransferContext;
 
-    vk::CommandPool m_TransferPool;
+    TransferContext CreateTransferContext();
-    eastl::deque<TransferContext> m_TransferContexts;
-    eastl::vector<u32> m_TransferContextFreeList;
-
-    TransferContext &CreateTransferContext();
     Receipt Submit(Context &context);
 
     //

aster/include/aster/systems/resource.h

@@ -41,7 +41,7 @@ CastImage(const Ref<TFrom> &from)
 {
     if constexpr (not concepts::ImageInto<TFrom, TTo>)
         assert(TTo::FLAGS & from->m_Flags_);
-    return std::reinterpret_pointer_cast<TTo>(from);
+    return eastl::reinterpret_pointer_cast<TTo>(from);
 }
 #pragma endregion
 
@@ -54,7 +54,7 @@ CastView(const Ref<View<TFrom>> &from)
 {
     if constexpr (not concepts::ImageInto<TFrom, typename TTo::ImageType>)
         assert(TTo::ImageType::FLAGS & from->m_Image->m_Flags_);
-    return std::reinterpret_pointer_cast<TTo>(from);
+    return eastl::reinterpret_pointer_cast<TTo>(from);
 }
 #pragma endregion
 

aster/include/aster/systems/sync_server.h

@@ -4,6 +4,7 @@
 // =============================================
 
 #include "aster/aster.h"
+#include "context.h"
 
 #include <EASTL/deque.h>
 #include <EASTL/intrusive_list.h>
@@ -28,11 +29,20 @@ class SyncServer
     {
         vk::Semaphore m_Semaphore;
         TimelinePoint m_CurrentPoint;
+        ContextPool *m_AttachedPool;
 
-        static Entry Create(Device &device);
+        explicit Entry(Device &device);
         void Destroy(Device &device);
         void Wait(Device &device);
         void Next();
+        void AttachPool(ContextPool *pool);
+
+        Entry(Entry &&) = default;
+        Entry &operator=(Entry &&) = default;
+
+        ~Entry() = default;
+
+        DISALLOW_COPY_AND_ASSIGN(Entry);
     };
 
     Device *m_Device;

aster/src/aster/systems/CMakeLists.txt

@@ -7,4 +7,5 @@ PRIVATE
 "device.cpp"
 "commit_manager.cpp"
 "pipeline_helpers.cpp"
+"context.cpp"
 "sync_server.cpp")

aster/src/aster/systems/context.cpp

@@ -0,0 +1,402 @@
+// =============================================
+//  Aster: context.cpp
+//  Copyright (c) 2020-2025 Anish Bhobe
+// =============================================
+
+#include "aster/systems/context.h"
+
+#include "aster/systems/commit_manager.h"
+#include "systems/device.h"
+
+constexpr static u32
+GetFormatSize(const vk::Format format)
+{
+    switch (format)
+    {
+    case vk::Format::eUndefined:
+        return 0;
+    case vk::Format::eR8Unorm:
+    case vk::Format::eR8Snorm:
+    case vk::Format::eR8Uscaled:
+    case vk::Format::eR8Sscaled:
+    case vk::Format::eR8Uint:
+    case vk::Format::eR8Sint:
+    case vk::Format::eR8Srgb:
+        return 1;
+    case vk::Format::eR8G8Unorm:
+    case vk::Format::eR8G8Snorm:
+    case vk::Format::eR8G8Uscaled:
+    case vk::Format::eR8G8Sscaled:
+    case vk::Format::eR8G8Uint:
+    case vk::Format::eR8G8Sint:
+    case vk::Format::eR8G8Srgb:
+        return 2;
+    case vk::Format::eR8G8B8Unorm:
+    case vk::Format::eR8G8B8Snorm:
+    case vk::Format::eR8G8B8Uscaled:
+    case vk::Format::eR8G8B8Sscaled:
+    case vk::Format::eR8G8B8Uint:
+    case vk::Format::eR8G8B8Sint:
+    case vk::Format::eR8G8B8Srgb:
+    case vk::Format::eB8G8R8Unorm:
+    case vk::Format::eB8G8R8Snorm:
+    case vk::Format::eB8G8R8Uscaled:
+    case vk::Format::eB8G8R8Sscaled:
+    case vk::Format::eB8G8R8Uint:
+    case vk::Format::eB8G8R8Sint:
+    case vk::Format::eB8G8R8Srgb:
+        return 3;
+    case vk::Format::eR8G8B8A8Unorm:
+    case vk::Format::eR8G8B8A8Snorm:
+    case vk::Format::eR8G8B8A8Uscaled:
+    case vk::Format::eR8G8B8A8Sscaled:
+    case vk::Format::eR8G8B8A8Uint:
+    case vk::Format::eR8G8B8A8Sint:
+    case vk::Format::eR8G8B8A8Srgb:
+    case vk::Format::eB8G8R8A8Unorm:
+    case vk::Format::eB8G8R8A8Snorm:
+    case vk::Format::eB8G8R8A8Uscaled:
+    case vk::Format::eB8G8R8A8Sscaled:
+    case vk::Format::eB8G8R8A8Uint:
+    case vk::Format::eB8G8R8A8Sint:
+    case vk::Format::eB8G8R8A8Srgb:
+        return 4;
+    case vk::Format::eR16Unorm:
+    case vk::Format::eR16Snorm:
+    case vk::Format::eR16Uscaled:
+    case vk::Format::eR16Sscaled:
+    case vk::Format::eR16Uint:
+    case vk::Format::eR16Sint:
+    case vk::Format::eR16Sfloat:
+        return 2;
+    case vk::Format::eR16G16Unorm:
+    case vk::Format::eR16G16Snorm:
+    case vk::Format::eR16G16Uscaled:
+    case vk::Format::eR16G16Sscaled:
+    case vk::Format::eR16G16Uint:
+    case vk::Format::eR16G16Sint:
+    case vk::Format::eR16G16Sfloat:
+        return 4;
+    case vk::Format::eR16G16B16Unorm:
+    case vk::Format::eR16G16B16Snorm:
+    case vk::Format::eR16G16B16Uscaled:
+    case vk::Format::eR16G16B16Sscaled:
+    case vk::Format::eR16G16B16Uint:
+    case vk::Format::eR16G16B16Sint:
+    case vk::Format::eR16G16B16Sfloat:
+        return 6;
+    case vk::Format::eR16G16B16A16Unorm:
+    case vk::Format::eR16G16B16A16Snorm:
+    case vk::Format::eR16G16B16A16Uscaled:
+    case vk::Format::eR16G16B16A16Sscaled:
+    case vk::Format::eR16G16B16A16Uint:
+    case vk::Format::eR16G16B16A16Sint:
+    case vk::Format::eR16G16B16A16Sfloat:
+        return 8;
+    case vk::Format::eR32Uint:
+    case vk::Format::eR32Sint:
+    case vk::Format::eR32Sfloat:
+        return 4;
+    case vk::Format::eR32G32Uint:
+    case vk::Format::eR32G32Sint:
+    case vk::Format::eR32G32Sfloat:
+        return 8;
+    case vk::Format::eR32G32B32Uint:
+    case vk::Format::eR32G32B32Sint:
+    case vk::Format::eR32G32B32Sfloat:
+        return 12;
+    case vk::Format::eR32G32B32A32Uint:
+    case vk::Format::eR32G32B32A32Sint:
+    case vk::Format::eR32G32B32A32Sfloat:
+        return 16;
+    case vk::Format::eD16Unorm:
+        return 2;
+    case vk::Format::eD32Sfloat:
+        return 4;
+    case vk::Format::eS8Uint:
+        return 1;
+    case vk::Format::eD16UnormS8Uint:
+        return 6;
+    case vk::Format::eD24UnormS8Uint:
+        return 4;
+    case vk::Format::eD32SfloatS8Uint:
+        return 5;
+    default:
+        TODO("Esoteric Formats");
+    }
+
+    return 0;
+}
+
+void
+systems::Context::KeepAlive(const Ref<Buffer> &buffer)
+{
+    assert(m_Pool);
+    m_Pool->KeepAlive(buffer);
+}
+
+void
+systems::Context::KeepAlive(const Ref<Image> &image)
+{
+    assert(m_Pool);
+    m_Pool->KeepAlive(image);
+}
+
+void
+systems::Context::KeepAlive(const Ref<ImageView> &view)
+{
+    assert(m_Pool);
+    m_Pool->KeepAlive(view);
+}
+
+void
+systems::Context::Dependency(const vk::DependencyInfo &dependencyInfo)
+{
+    m_Cmd.pipelineBarrier2(&dependencyInfo);
+}
+
+void
+systems::Context::Begin()
+{
+    vk::CommandBufferBeginInfo commandBufferBeginInfo = {
+        .flags = vk::CommandBufferUsageFlagBits::eOneTimeSubmit,
+    };
+
+    auto result = m_Cmd.begin(&commandBufferBeginInfo);
+    ERROR_IF(Failed(result), "Could not begin context") THEN_ABORT(result);
+}
+
+void
+systems::Context::End()
+{
+    auto result = m_Cmd.end();
+    ERROR_IF(Failed(result), "Could not end context") THEN_ABORT(result);
+}
+
+void
+systems::GraphicsContext::SetViewport(const vk::Viewport &viewport)
+{
+    m_Cmd.setViewport(0, 1, &viewport);
+}
+
+void
+systems::GraphicsContext::BindVertexBuffer(const Ref<VertexBuffer> &vertexBuffer)
+{
+    constexpr vk::DeviceSize offset = 0;
+    m_Cmd.bindVertexBuffers(0, 1, &vertexBuffer->m_Buffer, &offset);
+}
+
+void
+systems::GraphicsContext::BindPipeline(const Pipeline &pipeline)
+{
+    m_Cmd.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline.m_Pipeline);
+
+    // TODO: Maybe find a smarter place to host this.
+    if (CommitManager::IsInit())
+    {
+        m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, pipeline.m_Layout, 0, 1,
+                                 &CommitManager::Instance().GetDescriptorSet(), 0, nullptr);
+    }
+
+    m_PipelineInUse = &pipeline;
+}
+
+void
+systems::GraphicsContext::Draw(const usize vertexCount)
+{
+    m_Cmd.draw(static_cast<u32>(vertexCount), 1, 0, 0);
+}
+
+void
+systems::GraphicsContext::DrawIndexed(usize indexCount)
+{
+    m_Cmd.drawIndexed(static_cast<u32>(indexCount), 1, 0, 0, 0);
+}
+
+void
+systems::GraphicsContext::BeginRendering(const vk::RenderingInfo &renderingInfo)
+{
+    m_Cmd.beginRendering(&renderingInfo);
+    m_Cmd.setScissor(0, 1, &renderingInfo.renderArea);
+}
+
+void
+systems::GraphicsContext::EndRendering()
+{
+    m_Cmd.endRendering();
+}
+
+void
+systems::TransferContext::UploadTexture(const Ref<Image> &image, const ImageData &data)
+{
+    ERROR_IF(not(image and image->IsValid()), "Invalid image");
+
+    auto [w, h, d] = image->m_Extent;
+    auto formatSize = GetFormatSize(image->m_Format);
+    auto expectedByteSize = static_cast<u64>(w) * static_cast<u64>(h) * static_cast<u64>(d) * formatSize;
+    ERROR_IF(expectedByteSize != data.m_NumBytes, "Mismatch in data size {} vs image size {} ({}x{}x{}x{})",
+             data.m_NumBytes, expectedByteSize, w, h, d, formatSize);
+
+    const Ref<StagingBuffer> stagingBuffer = m_Pool->GetDevice().CreateStagingBuffer(data.m_NumBytes);
+    stagingBuffer->Write(0, data.m_NumBytes, data.m_Data);
+
+    const vk::BufferImageCopy bufferImageCopy = {
+        .bufferOffset = 0,
+        .bufferRowLength = w,
+        .bufferImageHeight = h,
+        .imageSubresource =
+            {
+                .aspectMask = vk::ImageAspectFlagBits::eColor,
+                .mipLevel = 0,
+                .baseArrayLayer = 0,
+                .layerCount = 1,
+            },
+        .imageOffset = {},
+        .imageExtent = image->m_Extent,
+    };
+    m_Cmd.copyBufferToImage(stagingBuffer->m_Buffer, image->m_Image, vk::ImageLayout::eTransferDstOptimal, 1,
+                            &bufferImageCopy);
+
+    KeepAlive(stagingBuffer);
+    KeepAlive(image);
+}
+
+systems::TransferContext::TransferContext(TransferContext &&other) noexcept
+    : Context{std::move(other)}
+{
+}
+
+systems::TransferContext &
+systems::TransferContext::operator=(TransferContext &&other) noexcept
+{
+    if (this == &other)
+        return *this;
+    Context::operator=(std::move(other));
+    return *this;
+}
+
+using namespace systems::_internal;
+
+ContextPool::ContextPool(Device &device, const u32 queueFamilyIndex, const ManagedBy managedBy)
+    : m_Device{&device}
+    , m_BuffersAllocated{0}
+    , m_ExtraData{0}
+    , m_ManagedBy{managedBy}
+    , m_ResetCallback{}
+{
+    const vk::CommandPoolCreateInfo commandPoolCreateInfo = {
+        .flags = vk::CommandPoolCreateFlagBits::eTransient,
+        .queueFamilyIndex = queueFamilyIndex,
+    };
+    AbortIfFailed(device.m_Device->createCommandPool(&commandPoolCreateInfo, nullptr, &m_Pool));
+}
+
+ContextPool::ContextPool(ContextPool &&other) noexcept
+    : m_Device{other.m_Device}
+    , m_Pool{Take(other.m_Pool)}
+    , m_CommandBuffers{std::move(other.m_CommandBuffers)}
+    , m_BuffersAllocated{other.m_BuffersAllocated}
+    , m_ExtraData{other.m_ExtraData}
+    , m_ManagedBy{other.m_ManagedBy}
+    , m_OwnedBuffers{std::move(other.m_OwnedBuffers)}
+    , m_OwnedImages{std::move(other.m_OwnedImages)}
+    , m_OwnedImageViews{std::move(other.m_OwnedImageViews)}
+    , m_ResetCallback{std::move(other.m_ResetCallback)}
+{
+}
+
+ContextPool &
+ContextPool::operator=(ContextPool &&other) noexcept
+{
+    if (this == &other)
+        return *this;
+    using eastl::swap;
+    swap(m_Device, other.m_Device);
+    swap(m_Pool, other.m_Pool);
+    swap(m_CommandBuffers, other.m_CommandBuffers);
+    swap(m_ExtraData, other.m_ExtraData);
+    swap(m_ManagedBy, other.m_ManagedBy);
+    swap(m_BuffersAllocated, other.m_BuffersAllocated);
+    swap(m_OwnedBuffers, other.m_OwnedBuffers);
+    swap(m_OwnedImages, other.m_OwnedImages);
+    swap(m_OwnedImageViews, other.m_OwnedImageViews);
+    swap(m_ResetCallback, other.m_ResetCallback);
+    return *this;
+}
+
+ContextPool::~ContextPool()
+{
+    if (!m_Pool)
+        return;
+
+    m_Device->m_Device->destroy(Take(m_Pool), nullptr);
+}
+
+void
+ContextPool::KeepAlive(const Ref<Buffer> &buffer)
+{
+    m_OwnedBuffers.push_back(buffer);
+}
+
+void
+ContextPool::KeepAlive(const Ref<Image> &image)
+{
+    m_OwnedImages.push_back(image);
+}
+
+void
+ContextPool::KeepAlive(const Ref<ImageView> &view)
+{
+    m_OwnedImageViews.push_back(view);
+}
+
+vk::CommandBuffer
+ContextPool::AllocateCommandBuffer()
+{
+    // Buffers are available.
+    if (m_BuffersAllocated < m_CommandBuffers.size())
+    {
+        return m_CommandBuffers[m_BuffersAllocated++];
+    }
+
+    // Allocate New Buffer.
+    const vk::CommandBufferAllocateInfo allocateInfo = {
+        .commandPool = m_Pool,
+        .level = vk::CommandBufferLevel::ePrimary,
+        .commandBufferCount = 1,
+    };
+    vk::CommandBuffer &cmd = m_CommandBuffers.emplace_back();
+    AbortIfFailed(m_Device->m_Device->allocateCommandBuffers(&allocateInfo, &cmd));
+    return cmd;
+}
+
+systems::Context
+ContextPool::CreateContext()
+{
+    return Context{*this, AllocateCommandBuffer()};
+}
+
+void
+ContextPool::Reset()
+{
+    assert(m_Pool);
+
+    AbortIfFailed(m_Device->m_Device->resetCommandPool(m_Pool, {}));
+    m_BuffersAllocated = 0;
+
+    m_OwnedBuffers.clear();
+    m_OwnedImages.clear();
+    m_OwnedImageViews.clear();
+}
+
+systems::TransferContext
+TransferContextPool::CreateTransferContext()
+{
+    return TransferContext{*this, AllocateCommandBuffer()};
+}
+
+systems::GraphicsContext
+GraphicsContextPool::CreateGraphicsContext()
+{
+    return GraphicsContext{*this, AllocateCommandBuffer()};
+}

aster/src/aster/systems/device.cpp

@@ -12,6 +12,7 @@
 #include "aster/systems/sync_server.h"
 #include "aster/util/files.h"
 #include "systems/commit_manager.h"
+#include "systems/context.h"
 
 #include <EASTL/vector_map.h>
 #include <fmt/ranges.h>
@@ -20,126 +21,6 @@ static constexpr QueueSupportFlags REQUIRED_QUEUE_SUPPORT =
     QueueSupportFlags{} | QueueSupportFlagBits::eGraphics | QueueSupportFlagBits::eCompute |
     QueueSupportFlagBits::ePresent | QueueSupportFlagBits::eTransfer;
 
-constexpr static u32
-GetFormatSize(const vk::Format format)
-{
-    switch (format)
-    {
-    case vk::Format::eUndefined:
-        return 0;
-    case vk::Format::eR8Unorm:
-    case vk::Format::eR8Snorm:
-    case vk::Format::eR8Uscaled:
-    case vk::Format::eR8Sscaled:
-    case vk::Format::eR8Uint:
-    case vk::Format::eR8Sint:
-    case vk::Format::eR8Srgb:
-        return 1;
-    case vk::Format::eR8G8Unorm:
-    case vk::Format::eR8G8Snorm:
-    case vk::Format::eR8G8Uscaled:
-    case vk::Format::eR8G8Sscaled:
-    case vk::Format::eR8G8Uint:
-    case vk::Format::eR8G8Sint:
-    case vk::Format::eR8G8Srgb:
-        return 2;
-    case vk::Format::eR8G8B8Unorm:
-    case vk::Format::eR8G8B8Snorm:
-    case vk::Format::eR8G8B8Uscaled:
-    case vk::Format::eR8G8B8Sscaled:
-    case vk::Format::eR8G8B8Uint:
-    case vk::Format::eR8G8B8Sint:
-    case vk::Format::eR8G8B8Srgb:
-    case vk::Format::eB8G8R8Unorm:
-    case vk::Format::eB8G8R8Snorm:
-    case vk::Format::eB8G8R8Uscaled:
-    case vk::Format::eB8G8R8Sscaled:
-    case vk::Format::eB8G8R8Uint:
-    case vk::Format::eB8G8R8Sint:
-    case vk::Format::eB8G8R8Srgb:
-        return 3;
-    case vk::Format::eR8G8B8A8Unorm:
-    case vk::Format::eR8G8B8A8Snorm:
-    case vk::Format::eR8G8B8A8Uscaled:
-    case vk::Format::eR8G8B8A8Sscaled:
-    case vk::Format::eR8G8B8A8Uint:
-    case vk::Format::eR8G8B8A8Sint:
-    case vk::Format::eR8G8B8A8Srgb:
-    case vk::Format::eB8G8R8A8Unorm:
-    case vk::Format::eB8G8R8A8Snorm:
-    case vk::Format::eB8G8R8A8Uscaled:
-    case vk::Format::eB8G8R8A8Sscaled:
-    case vk::Format::eB8G8R8A8Uint:
-    case vk::Format::eB8G8R8A8Sint:
-    case vk::Format::eB8G8R8A8Srgb:
-        return 4;
-    case vk::Format::eR16Unorm:
-    case vk::Format::eR16Snorm:
-    case vk::Format::eR16Uscaled:
-    case vk::Format::eR16Sscaled:
-    case vk::Format::eR16Uint:
-    case vk::Format::eR16Sint:
-    case vk::Format::eR16Sfloat:
-        return 2;
-    case vk::Format::eR16G16Unorm:
-    case vk::Format::eR16G16Snorm:
-    case vk::Format::eR16G16Uscaled:
-    case vk::Format::eR16G16Sscaled:
-    case vk::Format::eR16G16Uint:
-    case vk::Format::eR16G16Sint:
-    case vk::Format::eR16G16Sfloat:
-        return 4;
-    case vk::Format::eR16G16B16Unorm:
-    case vk::Format::eR16G16B16Snorm:
-    case vk::Format::eR16G16B16Uscaled:
-    case vk::Format::eR16G16B16Sscaled:
-    case vk::Format::eR16G16B16Uint:
-    case vk::Format::eR16G16B16Sint:
-    case vk::Format::eR16G16B16Sfloat:
-        return 6;
-    case vk::Format::eR16G16B16A16Unorm:
-    case vk::Format::eR16G16B16A16Snorm:
-    case vk::Format::eR16G16B16A16Uscaled:
-    case vk::Format::eR16G16B16A16Sscaled:
-    case vk::Format::eR16G16B16A16Uint:
-    case vk::Format::eR16G16B16A16Sint:
-    case vk::Format::eR16G16B16A16Sfloat:
-        return 8;
-    case vk::Format::eR32Uint:
-    case vk::Format::eR32Sint:
-    case vk::Format::eR32Sfloat:
-        return 4;
-    case vk::Format::eR32G32Uint:
-    case vk::Format::eR32G32Sint:
-    case vk::Format::eR32G32Sfloat:
-        return 8;
-    case vk::Format::eR32G32B32Uint:
-    case vk::Format::eR32G32B32Sint:
-    case vk::Format::eR32G32B32Sfloat:
-        return 12;
-    case vk::Format::eR32G32B32A32Uint:
-    case vk::Format::eR32G32B32A32Sint:
-    case vk::Format::eR32G32B32A32Sfloat:
-        return 16;
-    case vk::Format::eD16Unorm:
-        return 2;
-    case vk::Format::eD32Sfloat:
-        return 4;
-    case vk::Format::eS8Uint:
-        return 1;
-    case vk::Format::eD16UnormS8Uint:
-        return 6;
-    case vk::Format::eD24UnormS8Uint:
-        return 4;
-    case vk::Format::eD32SfloatS8Uint:
-        return 5;
-    default:
-        TODO("Esoteric Formats");
-    }
-
-    return 0;
-}
-
 PhysicalDevice
 systems::DefaultPhysicalDeviceSelector(const PhysicalDevices &physicalDevices)
 {
@@ -182,7 +63,7 @@ systems::Device::CreateStorageBuffer(const usize size, const cstr name)
                                                      VMA_ALLOCATION_CREATE_HOST_ACCESS_ALLOW_TRANSFER_INSTEAD_BIT |
                                                      VMA_ALLOCATION_CREATE_MAPPED_BIT;
     constexpr VmaMemoryUsage memoryUsage = VMA_MEMORY_USAGE_AUTO;
-    return std::make_shared<StorageBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
+    return eastl::make_shared<StorageBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
 }
 
 Ref<UniformBuffer>
@@ -193,7 +74,7 @@ systems::Device::CreateUniformBuffer(const usize size, const cstr name)
                                                      VMA_ALLOCATION_CREATE_HOST_ACCESS_ALLOW_TRANSFER_INSTEAD_BIT |
                                                      VMA_ALLOCATION_CREATE_MAPPED_BIT;
     constexpr VmaMemoryUsage memoryUsage = VMA_MEMORY_USAGE_AUTO;
-    return std::make_shared<UniformBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
+    return eastl::make_shared<UniformBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
 }
 
 Ref<StagingBuffer>
@@ -203,7 +84,7 @@ systems::Device::CreateStagingBuffer(const usize size, const cstr name)
     constexpr VmaAllocationCreateFlags createFlags =
         VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT | VMA_ALLOCATION_CREATE_MAPPED_BIT;
     constexpr VmaMemoryUsage memoryUsage = VMA_MEMORY_USAGE_AUTO;
-    return std::make_shared<StagingBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
+    return eastl::make_shared<StagingBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
 }
 
 Ref<VertexBuffer>
@@ -213,7 +94,7 @@ systems::Device::CreateVertexBuffer(const usize size, const cstr name)
     constexpr VmaAllocationCreateFlags createFlags =
         VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT | VMA_ALLOCATION_CREATE_MAPPED_BIT;
     constexpr VmaMemoryUsage memoryUsage = VMA_MEMORY_USAGE_AUTO;
-    return std::make_shared<VertexBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
+    return eastl::make_shared<VertexBuffer>(Buffer{&m_Device, size, usage, createFlags, memoryUsage, name});
 }
 #pragma endregion
 
@@ -265,8 +146,8 @@ systems::Device::CreateTexture2D(const Texture2DCreateInfo &createInfo)
 
     m_Device.SetName(image, createInfo.m_Name);
 
-    return std::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format, flags,
+    return eastl::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format,
-                                   layerCount, mipLevels);
+                                     flags, layerCount, mipLevels);
 }
 
 Ref<ImageCube>
@@ -297,7 +178,7 @@ systems::Device::CreateTextureCube(const TextureCubeCreateInfo &createInfo)
 
     m_Device.SetName(image, createInfo.m_Name);
 
-    return CastImage<ImageCube>(std::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent,
+    return CastImage<ImageCube>(eastl::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent,
                                                           imageCreateInfo.format, flags, layerCount, mipLevels));
 }
 
@@ -324,7 +205,7 @@ systems::Device::CreateAttachment(const AttachmentCreateInfo &createInfo)
 
     m_Device.SetName(image, createInfo.m_Name);
 
-    return std::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format,
+    return eastl::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format,
                                      Image::Flags{}, layerCount, mipLevels);
 }
 
@@ -351,7 +232,7 @@ systems::Device::CreateDepthStencilImage(const DepthStencilImageCreateInfo &crea
 
     m_Device.SetName(image, createInfo.m_Name);
 
-    return std::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format,
+    return eastl::make_shared<Image>(&m_Device, image, allocation, imageCreateInfo.extent, imageCreateInfo.format,
                                      Image::Flags{}, layerCount, mipLevels);
 }
 
@@ -470,7 +351,7 @@ systems::Device::CreateView(const ViewCreateInfo<Image> &createInfo)
 
     m_Device.SetName(view, createInfo.m_Name);
 
-    return std::make_shared<ImageView>(createInfo.m_Image, view, createInfo.m_Image->m_Extent, createInfo.m_BaseLayer,
+    return eastl::make_shared<ImageView>(createInfo.m_Image, view, createInfo.m_Image->m_Extent, createInfo.m_BaseLayer,
                                          layerCount, createInfo.m_BaseMipLevel, mipCount);
 }
 
@@ -541,7 +422,7 @@ systems::Device::CreateSampler(const SamplerCreateInfo &createInfo)
         return iter->second.lock();
     }
 
-    auto object = std::make_shared<Sampler>(&m_Device, vkCreateInfo, createInfo.m_Name ? createInfo.m_Name : nullptr);
+    auto object = eastl::make_shared<Sampler>(&m_Device, vkCreateInfo, createInfo.m_Name ? createInfo.m_Name : nullptr);
     m_HashToSamplerIdx.emplace(vkCreateInfo, object);
 
     return object;
@@ -1029,14 +910,27 @@ systems::Device::Device(const DeviceCreateInfo &createInfo)
         m_TransferQueueFamily = m_PrimaryQueueFamily;
     }
 
-    // TODO: Async Compute
+    u32 computeQueueIndex;
+    if (const auto asyncCompute = FindAsyncComputeQueue(physicalDevice, m_PrimaryQueueFamily))
+    {
+        const QueueAllocation allocation = asyncCompute.value();
+        queueAllocations.push_back(allocation);
+        m_ComputeQueueFamily = allocation.m_Family;
+        computeQueueIndex = 0;
+    }
+    else
+    {
+        computeQueueIndex = primaryQueue.m_Count;
+        ++primaryQueue.m_Count;
+        m_ComputeQueueFamily = m_PrimaryQueueFamily;
+    }
 
     m_Device = ::Device{m_Instance,       physicalDevice, features, queueAllocations, createInfo.m_PipelineCacheData,
                         createInfo.m_Name};
 
-    m_GraphicsQueue = m_Device.GetQueue(m_PrimaryQueueFamily, primaryQueueIndex);
+    m_PrimaryQueue = m_Device.GetQueue(m_PrimaryQueueFamily, primaryQueueIndex);
-
     m_TransferQueue = m_Device.GetQueue(m_TransferQueueFamily, transferQueueIndex);
+    m_ComputeQueue = m_Device.GetQueue(m_ComputeQueueFamily, computeQueueIndex);
 
     m_Swapchain = Swapchain{m_Surface, m_Device, m_Window.get().GetSize()};
 
@@ -1045,11 +939,7 @@ systems::Device::Device(const DeviceCreateInfo &createInfo)
         m_CommitManager = std::make_unique<CommitManager>(this, 1000, 1000, 1000, CreateSampler({}));
     }
 
-    const vk::CommandPoolCreateInfo transferPoolCreateInfo = {
+    m_TransferContextPool.Init(*this, m_TransferQueueFamily);
-        .flags = vk::CommandPoolCreateFlagBits::eTransient | vk::CommandPoolCreateFlagBits::eResetCommandBuffer,
-        .queueFamilyIndex = m_TransferQueueFamily,
-    };
-    AbortIfFailed(m_Device->createCommandPool(&transferPoolCreateInfo, nullptr, &m_TransferPool));
 
     constexpr SlangGlobalSessionDesc globalSessionDesc = {};
     auto result = slang::createGlobalSession(&globalSessionDesc, m_GlobalSlangSession.writeRef());
@@ -1091,7 +981,7 @@ systems::Device::Device(const DeviceCreateInfo &createInfo)
     u32 index = 0;
     for (auto &frame : m_Frames)
     {
-        frame = Frame(*this, m_PrimaryQueueFamily, index++);
+        frame = Frame(*this, index++, m_PrimaryQueueFamily, m_TransferQueueFamily, m_ComputeQueueFamily);
     }
 }
 
@@ -1102,10 +992,7 @@ systems::Device::~Device()
         m_Device->destroy(Take(frame.m_FrameAvailableFence), nullptr);
         m_Device->destroy(Take(frame.m_ImageAcquireSem), nullptr);
         m_Device->destroy(Take(frame.m_RenderFinishSem), nullptr);
-        m_Device->destroy(Take(frame.m_Pool), nullptr);
     }
-    m_Device->destroy(Take(m_TransferPool), nullptr);
-    m_TransferContexts.clear();
     m_HashToSamplerIdx.clear();
 }
 
@@ -1164,7 +1051,7 @@ systems::Device::Present(Frame &frame, GraphicsContext &graphicsContext)
         .signalSemaphoreCount = 1,
         .pSignalSemaphores = &frame.m_RenderFinishSem,
     };
-    vk::Result result = m_GraphicsQueue.submit(1, &submitInfo, frame.m_FrameAvailableFence);
+    vk::Result result = m_PrimaryQueue.submit(1, &submitInfo, frame.m_FrameAvailableFence);
     ERROR_IF(Failed(result), "Command queue submit failed. Cause: {}", result)
     THEN_ABORT(result);
 
@@ -1176,7 +1063,7 @@ systems::Device::Present(Frame &frame, GraphicsContext &graphicsContext)
         .pImageIndices = &frame.m_ImageIdx,
         .pResults = nullptr,
     };
-    switch (result = m_GraphicsQueue.presentKHR(&presentInfo))
+    switch (result = m_PrimaryQueue.presentKHR(&presentInfo))
     {
     case vk::Result::eSuccess:
         break;
@@ -1190,36 +1077,20 @@ systems::Device::Present(Frame &frame, GraphicsContext &graphicsContext)
     }
 }
 
-systems::TransferContext &
+systems::TransferContext
 systems::Device::CreateTransferContext()
 {
-    if (!m_TransferContextFreeList.empty())
+    return m_TransferContextPool.CreateTransferContext();
-    {
-        u32 freeIndex = m_TransferContextFreeList.back();
-        m_TransferContextFreeList.pop_back();
-
-        TransferContext &context = m_TransferContexts[freeIndex];
-
-        return context;
-    }
-
-    vk::CommandBuffer cmd;
-    vk::CommandBufferAllocateInfo allocateInfo = {
-        .commandPool = m_TransferPool,
-        .level = vk::CommandBufferLevel::ePrimary,
-        .commandBufferCount = 1,
-    };
-    AbortIfFailed(m_Device->allocateCommandBuffers(&allocateInfo, &cmd));
-
-    m_TransferContexts.push_back(TransferContext{*this, cmd});
-    return m_TransferContexts.back();
 }
 
 systems::Receipt
 systems::Device::Submit(Context &context)
 {
     const auto rect = m_SyncServer->Allocate();
-    auto entry = m_SyncServer->GetEntry(rect);
+    auto &entry = m_SyncServer->GetEntry(rect);
+
+    if (context.m_Pool->m_ManagedBy == _internal::ContextPool::ManagedBy::eDevice)
+        entry.AttachPool(context.m_Pool);
 
     auto [wait, next] = entry.m_CurrentPoint;
     vk::TimelineSemaphoreSubmitInfo timelineSubmit = {
@@ -1241,7 +1112,7 @@ systems::Device::Submit(Context &context)
         .signalSemaphoreCount = 1,
         .pSignalSemaphores = &entry.m_Semaphore,
     };
-    vk::Result result = m_GraphicsQueue.submit(1, &submitInfo, {});
+    vk::Result result = m_PrimaryQueue.submit(1, &submitInfo, {});
     ERROR_IF(Failed(result), "Command queue submit failed. Cause: {}", result)
     THEN_ABORT(result);
 
@@ -1259,8 +1130,9 @@ systems::Frame::Reset(u32 imageIdx, vk::Image swapchainImage, vk::ImageView swap
 {
     AbortIfFailedMV(m_Device->m_Device->resetFences(1, &m_FrameAvailableFence), "Fence {} reset failed.", m_FrameIdx);
 
-    AbortIfFailedMV(m_Device->m_Device->resetCommandPool(m_Pool, {}), "Command pool {} reset failed.", m_FrameIdx);
+    m_PrimaryPool.Reset();
-    m_CommandBuffersAllocated = 0;
+    m_AsyncTransferPool.Reset();
+    m_AsyncComputePool.Reset();
 
     m_ImageIdx = imageIdx;
     m_SwapchainImage = swapchainImage;
@@ -1271,47 +1143,13 @@ systems::Frame::Reset(u32 imageIdx, vk::Image swapchainImage, vk::ImageView swap
 systems::GraphicsContext
 systems::Frame::CreateGraphicsContext()
 {
-    vk::CommandBuffer cmd;
+    return m_PrimaryPool.CreateGraphicsContext();
-    if (m_CommandBuffers.size() > m_CommandBuffersAllocated)
-    {
-        cmd = m_CommandBuffers[m_CommandBuffersAllocated++];
-    }
-    else
-    {
-        const vk::CommandBufferAllocateInfo allocateInfo{
-            .commandPool = m_Pool,
-            .level = vk::CommandBufferLevel::ePrimary,
-            .commandBufferCount = 1,
-        };
-        AbortIfFailedMV(m_Device->m_Device->allocateCommandBuffers(&allocateInfo, &cmd),
-                        "Command buffer {} alloc failed.", m_FrameIdx);
-        m_CommandBuffers.push_back(cmd);
-    }
-
-    return GraphicsContext{cmd};
 }
 
 systems::TransferContext
-systems::Frame::CreateTransferContext()
+systems::Frame::CreateAsyncTransferContext()
 {
-    vk::CommandBuffer cmd;
+    return m_AsyncTransferPool.CreateTransferContext();
-    if (m_CommandBuffers.size() > m_CommandBuffersAllocated)
-    {
-        cmd = m_CommandBuffers[m_CommandBuffersAllocated++];
-    }
-    else
-    {
-        const vk::CommandBufferAllocateInfo allocateInfo{
-            .commandPool = m_Pool,
-            .level = vk::CommandBufferLevel::ePrimary,
-            .commandBufferCount = 1,
-        };
-        AbortIfFailedMV(m_Device->m_Device->allocateCommandBuffers(&allocateInfo, &cmd),
-                        "Command buffer {} alloc failed.", m_FrameIdx);
-        m_CommandBuffers.push_back(cmd);
-    }
-
-    return TransferContext{*m_Device, cmd};
 }
 
 void
@@ -1321,21 +1159,38 @@ systems::Frame::WaitUntilReady()
                     "Waiting for fence {} failed.", m_FrameIdx);
 }
 
-systems::Frame::Frame(Device &device, u32 primaryQueueFamily, u32 frameIndex)
+systems::Frame &
+systems::Frame::operator=(Frame &&other) noexcept
+{
+    if (this == &other)
+        return *this;
+    m_Device = other.m_Device;
+    m_PrimaryPool = Take(other.m_PrimaryPool);
+    m_AsyncTransferPool = Take(other.m_AsyncTransferPool);
+    m_AsyncComputePool = Take(other.m_AsyncComputePool);
+    m_FrameAvailableFence = Take(other.m_FrameAvailableFence);
+    m_ImageAcquireSem = Take(other.m_ImageAcquireSem);
+    m_RenderFinishSem = Take(other.m_RenderFinishSem);
+    m_FrameIdx = other.m_FrameIdx;
+    m_SwapchainImage = Take(other.m_SwapchainImage);
+    m_SwapchainImageView = Take(other.m_SwapchainImageView);
+    m_SwapchainSize = Take(other.m_SwapchainSize);
+    m_ImageIdx = other.m_ImageIdx;
+    return *this;
+}
+
+systems::Frame::Frame(Device &device, u32 frameIndex, u32 const primaryQueueFamily, u32 const asyncTransferQueue,
+                      u32 const asyncComputeQueue)
     : m_Device{&device}
+    , m_PrimaryPool{device, primaryQueueFamily, _internal::ContextPool::ManagedBy::eFrame}
+    , m_AsyncTransferPool{device, asyncTransferQueue, _internal::ContextPool::ManagedBy::eFrame}
+    , m_AsyncComputePool{device, asyncComputeQueue, _internal::ContextPool::ManagedBy::eFrame}
     , m_FrameIdx{frameIndex}
     , m_ImageIdx{0}
 {
     NameString name = "Frame ";
     name += static_cast<char>(frameIndex + '0');
 
-    const vk::CommandPoolCreateInfo commandPoolCreateInfo = {
-        .flags = vk::CommandPoolCreateFlagBits::eTransient,
-        .queueFamilyIndex = primaryQueueFamily,
-    };
-    AbortIfFailedMV(device.m_Device->createCommandPool(&commandPoolCreateInfo, nullptr, &m_Pool),
-                    "Could not command pool for frame {}", frameIndex);
-
     constexpr vk::FenceCreateInfo fenceCreateInfo = {.flags = vk::FenceCreateFlagBits::eSignaled};
     AbortIfFailedMV(device.m_Device->createFence(&fenceCreateInfo, nullptr, &m_FrameAvailableFence),
                     "Could not create a fence for frame {}", frameIndex);
@@ -1346,7 +1201,6 @@ systems::Frame::Frame(Device &device, u32 primaryQueueFamily, u32 frameIndex)
     AbortIfFailedMV(device.m_Device->createSemaphore(&semaphoreCreateInfo, nullptr, &m_RenderFinishSem),
                     "Could not create RF semaphore for frame {}.", frameIndex);
 
-    m_Device->SetName(m_Pool, name.c_str());
     m_Device->SetName(m_FrameAvailableFence, name.c_str());
     m_Device->SetName(m_ImageAcquireSem, name.c_str());
     m_Device->SetName(m_RenderFinishSem, name.c_str());
@@ -1354,186 +1208,20 @@ systems::Frame::Frame(Device &device, u32 primaryQueueFamily, u32 frameIndex)
     DEBUG("Frame {} created successfully.", frameIndex);
 }
 
-systems::Frame &
-systems::Frame::operator=(Frame &&other) noexcept
-{
-    if (this == &other)
-        return *this;
-    m_Device = other.m_Device;
-    m_Pool = Take(other.m_Pool);
-    m_FrameAvailableFence = Take(other.m_FrameAvailableFence);
-    m_ImageAcquireSem = Take(other.m_ImageAcquireSem);
-    m_RenderFinishSem = Take(other.m_RenderFinishSem);
-    m_FrameIdx = other.m_FrameIdx;
-    m_CommandBuffers = Take(other.m_CommandBuffers);
-    m_SwapchainImage = Take(other.m_SwapchainImage);
-    m_SwapchainImageView = Take(other.m_SwapchainImageView);
-    m_SwapchainSize = Take(other.m_SwapchainSize);
-    m_ImageIdx = other.m_ImageIdx;
-    m_CommandBuffersAllocated = other.m_CommandBuffersAllocated;
-    return *this;
-}
-
 systems::Frame::Frame(Frame &&other) noexcept
     : m_Device{Take(other.m_Device)}
-    , m_Pool{Take(other.m_Pool)}
+    , m_PrimaryPool{std::move(other.m_PrimaryPool)}
+    , m_AsyncTransferPool{std::move(other.m_AsyncTransferPool)}
+    , m_AsyncComputePool{std::move(other.m_AsyncComputePool)}
     , m_FrameAvailableFence{Take(other.m_FrameAvailableFence)}
     , m_ImageAcquireSem{Take(other.m_ImageAcquireSem)}
     , m_RenderFinishSem{Take(other.m_RenderFinishSem)}
     , m_FrameIdx{other.m_FrameIdx}
-    , m_CommandBuffers{Take(other.m_CommandBuffers)}
     , m_SwapchainImage{Take(other.m_SwapchainImage)}
     , m_SwapchainImageView{Take(other.m_SwapchainImageView)}
     , m_SwapchainSize{Take(other.m_SwapchainSize)}
     , m_ImageIdx{other.m_ImageIdx}
-    , m_CommandBuffersAllocated{other.m_CommandBuffersAllocated}
 {
 }
 
 #pragma endregion
-
-// ====================================================================================================
-#pragma region Context Impl
-// ====================================================================================================
-
-void
-systems::Context::Dependency(const vk::DependencyInfo &dependencyInfo)
-{
-    m_Cmd.pipelineBarrier2(&dependencyInfo);
-}
-
-void
-systems::Context::Begin()
-{
-    vk::CommandBufferBeginInfo commandBufferBeginInfo = {
-        .flags = vk::CommandBufferUsageFlagBits::eOneTimeSubmit,
-    };
-
-    auto result = m_Cmd.begin(&commandBufferBeginInfo);
-    ERROR_IF(Failed(result), "Could not begin context") THEN_ABORT(result);
-}
-
-void
-systems::Context::End()
-{
-    auto result = m_Cmd.end();
-    ERROR_IF(Failed(result), "Could not end context") THEN_ABORT(result);
-}
-
-void
-systems::GraphicsContext::SetViewport(const vk::Viewport &viewport)
-{
-    m_Cmd.setViewport(0, 1, &viewport);
-}
-
-void
-systems::GraphicsContext::BindVertexBuffer(const Ref<VertexBuffer> &vertexBuffer)
-{
-    constexpr vk::DeviceSize offset = 0;
-    m_Cmd.bindVertexBuffers(0, 1, &vertexBuffer->m_Buffer, &offset);
-}
-
-void
-systems::GraphicsContext::BindPipeline(const Pipeline &pipeline)
-{
-    m_Cmd.bindPipeline(vk::PipelineBindPoint::eGraphics, pipeline.m_Pipeline);
-
-    // TODO: Maybe find a smarter place to host this.
-    if (CommitManager::IsInit())
-    {
-        m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eGraphics, pipeline.m_Layout, 0, 1,
-                                 &CommitManager::Instance().GetDescriptorSet(), 0, nullptr);
-    }
-
-    m_PipelineInUse = &pipeline;
-}
-
-void
-systems::GraphicsContext::Draw(const usize vertexCount)
-{
-    m_Cmd.draw(static_cast<u32>(vertexCount), 1, 0, 0);
-}
-
-void
-systems::GraphicsContext::DrawIndexed(usize indexCount)
-{
-    m_Cmd.drawIndexed(static_cast<u32>(indexCount), 1, 0, 0, 0);
-}
-
-void
-systems::GraphicsContext::BeginRendering(const vk::RenderingInfo &renderingInfo)
-{
-    m_Cmd.beginRendering(&renderingInfo);
-    m_Cmd.setScissor(0, 1, &renderingInfo.renderArea);
-}
-
-void
-systems::GraphicsContext::EndRendering()
-{
-    m_Cmd.endRendering();
-}
-
-void
-systems::TransferContext::Reset()
-{
-    m_OwnedImages.clear();
-    m_OwnedBuffers.clear();
-    AbortIfFailed(m_Cmd.reset({}));
-}
-
-void
-systems::TransferContext::UploadTexture(const Ref<Image> &image, const ImageData &data)
-{
-    ERROR_IF(not(image and image->IsValid()), "Invalid image");
-
-    auto [w, h, d] = image->m_Extent;
-    auto formatSize = GetFormatSize(image->m_Format);
-    auto expectedByteSize = static_cast<u64>(w) * static_cast<u64>(h) * static_cast<u64>(d) * formatSize;
-    ERROR_IF(expectedByteSize != data.m_NumBytes, "Mismatch in data size {} vs image size {} ({}x{}x{}x{})",
-             data.m_NumBytes, expectedByteSize, w, h, d, formatSize);
-
-    const Ref<StagingBuffer> stagingBuffer = m_Device->CreateStagingBuffer(data.m_NumBytes);
-    stagingBuffer->Write(0, data.m_NumBytes, data.m_Data);
-
-    const vk::BufferImageCopy bufferImageCopy = {
-        .bufferOffset = 0,
-        .bufferRowLength = w,
-        .bufferImageHeight = h,
-        .imageSubresource =
-            {
-                .aspectMask = vk::ImageAspectFlagBits::eColor,
-                .mipLevel = 0,
-                .baseArrayLayer = 0,
-                .layerCount = 1,
-            },
-        .imageOffset = {},
-        .imageExtent = image->m_Extent,
-    };
-    m_Cmd.copyBufferToImage(stagingBuffer->m_Buffer, image->m_Image, vk::ImageLayout::eTransferDstOptimal, 1,
-                            &bufferImageCopy);
-
-    m_OwnedBuffers.push_back(stagingBuffer);
-    m_OwnedImages.push_back(image);
-}
-
-systems::TransferContext::TransferContext(TransferContext &&other) noexcept
-    : Context{other.m_Cmd}
-    , m_Device{Take(other.m_Device)}
-    , m_OwnedBuffers{std::move(other.m_OwnedBuffers)}
-    , m_OwnedImages{std::move(other.m_OwnedImages)}
-{
-}
-
-systems::TransferContext &
-systems::TransferContext::operator=(TransferContext &&other) noexcept
-{
-    if (this == &other)
-        return *this;
-    m_Cmd = other.m_Cmd;
-    m_Device = Take(other.m_Device);
-    m_OwnedBuffers = std::move(other.m_OwnedBuffers);
-    m_OwnedImages = std::move(other.m_OwnedImages);
-    return *this;
-}
-
-#pragma endregion

aster/src/aster/systems/sync_server.cpp

@@ -9,8 +9,9 @@
 
 using namespace systems::_internal;
 
-SyncServer::Entry
+SyncServer::Entry::Entry(Device &device)
-SyncServer::Entry::Create(Device &device)
+    : m_CurrentPoint{0, 1}
+    , m_AttachedPool{nullptr}
 {
     constexpr static vk::SemaphoreTypeCreateInfo TYPE_CREATE_INFO = {
         .semaphoreType = vk::SemaphoreType::eTimeline,
@@ -18,13 +19,7 @@ SyncServer::Entry::Create(Device &device)
     };
     constexpr static vk::SemaphoreCreateInfo SEMAPHORE_CREATE_INFO = {.pNext = &TYPE_CREATE_INFO};
 
-    vk::Semaphore semaphore;
+    AbortIfFailed(device.m_Device->createSemaphore(&SEMAPHORE_CREATE_INFO, nullptr, &m_Semaphore));
-    AbortIfFailed(device.m_Device->createSemaphore(&SEMAPHORE_CREATE_INFO, nullptr, &semaphore));
-
-    return {
-        .m_Semaphore = semaphore,
-        .m_CurrentPoint = {.m_WaitValue = 0, .m_NextValue = 1},
-    };
 }
 
 void
@@ -51,6 +46,11 @@ SyncServer::Entry::Wait(Device &device)
     // Thus, this is safe.
     m_CurrentPoint.m_WaitValue = m_CurrentPoint.m_NextValue;
     m_CurrentPoint.m_WaitValue = m_CurrentPoint.m_NextValue + 1;
+    if (m_AttachedPool)
+    {
+        m_AttachedPool->Reset();
+        m_AttachedPool = nullptr;
+    }
 }
 
 void
@@ -60,6 +60,13 @@ SyncServer::Entry::Next()
     ++m_CurrentPoint.m_NextValue;
 }
 
+void
+SyncServer::Entry::AttachPool(ContextPool *pool)
+{
+    assert(!m_AttachedPool);
+    m_AttachedPool = pool;
+}
+
 systems::Receipt
 SyncServer::Allocate()
 {
@@ -91,8 +98,7 @@ SyncServer::AllocateEntry()
         return alloc;
     }
 
-    m_Allocations.push_back(Entry::Create(*m_Device));
+    return m_Allocations.emplace_back(*m_Device);
-    return m_Allocations.back();
 }
 
 void

samples/02_box/box.cpp

@@ -246,7 +246,7 @@ main(int, char **)
             .pImageMemoryBarriers = &imageReadyToRead,
         };
 
-        auto &context = device.CreateTransferContext();
+        auto context = device.CreateTransferContext();
         context.Begin();
 
         context.Dependency(imageReadyToWriteDependency);