feat: RenderGraph GPU timestamp profiling

- RG::SetQueryPool 可选注入；Execute 期每 pass 前后各写一枚
  writeTimestamp2(BOTTOM_OF_PIPE)，索引按帧槽位分区 + 执行序排布
- 每帧录制开头 reset 本槽位查询区间（上一轮结果已在 fence 等待后读取）
- Renderer 在 WaitFence 后读 64bit 结果换算毫秒（timestampPeriod 缓存），
  经 GetPassNames/GetPassGpuMs 透出给 ImGui 面板显示
- 每槽位首帧跳过读取（查询尚未 primed，避免 VUID-09401）

App/main.cpp

@@ -15,6 +15,7 @@
 #include <Engine/ECS/EntityHandle.h>
 #include <Engine/ECS/Components.h>
 #include <Engine/Assets/AssetManager.h>
+#include <Engine/Renderer/Renderer.h>
 
 #include <imgui.h>
 
@@ -23,6 +24,8 @@
 
 #include <GLFW/glfw3.h>
 
+#include "Engine/ImGui/ImGuiLayer.h"
+
 using namespace Engine;
 using namespace Engine::ECS;
 
@@ -114,6 +117,8 @@ protected:
     {
         if (!m_OverlayVisible) return;
 
+        ImGui::ShowDemoWindow();
+
         ImGui::Begin("Scene Debug");
 
         // 阴影光源方向：拖动滑条实时看影子变化
@@ -126,6 +131,38 @@ protected:
             }
         }
 
+        // Tonemap + Bloom 参数（HDR 后处理输出端）
+        auto& renderer = GetRenderer();
+        ImGui::Separator();
+        ImGui::Text("Tonemap");
+        f32 exposure = renderer.GetExposure();
+        if (ImGui::SliderFloat("exposure", &exposure, 0.1f, 4.0f)) {
+            renderer.SetExposure(exposure);
+        }
+        bool aces = renderer.GetUseACES();
+        if (ImGui::Checkbox("ACES", &aces)) {
+            renderer.SetUseACES(aces);
+        }
+
+        ImGui::Text("Bloom");
+        f32 threshold = renderer.GetBloomThreshold();
+        if (ImGui::SliderFloat("threshold", &threshold, 0.0f, 2.0f)) {
+            renderer.SetBloomThreshold(threshold);
+        }
+        f32 strength = renderer.GetBloomStrength();
+        if (ImGui::SliderFloat("strength", &strength, 0.0f, 2.0f)) {
+            renderer.SetBloomStrength(strength);
+        }
+
+        // Pass 级 GPU 耗时（RG timestamp）
+        ImGui::Separator();
+        ImGui::Text("GPU Timings");
+        const auto& names = renderer.GetPassNames();
+        const auto& times = renderer.GetPassGpuMs();
+        for (size_t i = 0; i < names.size() && i < times.size(); ++i) {
+            ImGui::Text("  %-12s %.3f ms", names[i].c_str(), times[i]);
+        }
+
         ImGui::Separator();
         ImGui::Text("fps: %.1f  frame: %llu",
                     1.0f / GetDeltaTime(),

Engine/CMakeLists.txt

@@ -9,6 +9,7 @@ set(IMGUI_SOURCES
         ${IMGUI_DIR}/imgui_draw.cpp
         ${IMGUI_DIR}/imgui_tables.cpp
         ${IMGUI_DIR}/imgui_widgets.cpp
+        ${IMGUI_DIR}/imgui_demo.cpp      # ShowDemoWindow 的实现（调试期保留）
         ${IMGUI_DIR}/backends/imgui_impl_glfw.cpp
         ${IMGUI_DIR}/backends/imgui_impl_vulkan.cpp
 )

Engine/include/Engine/Core/Window.h

@@ -33,6 +33,17 @@ namespace Engine {
         using EventCallbackFn = std::function<void(const Event&)>;
         void SetEventCallback(EventCallbackFn callback) { m_EventCallback = std::move(callback); }
 
+        // 输入捕获：返回 true = 该类输入被 UI 占用（如 ImGui 窗口交互）。
+        // 生效范围：类型化事件回调（Key/Mouse*）+ 轮询接口
+        //（IsKeyPressed/GetMouseDelta）。窗口类事件（Resize/Close/Focus）不受影响。
+        // 未设置时全部放行（行为与无 UI 时一致）。
+        using CaptureFn = std::function<bool()>;
+        void SetInputCapture(CaptureFn keyboard, CaptureFn mouse)
+        {
+            m_KeyboardCapture = std::move(keyboard);
+            m_MouseCapture = std::move(mouse);
+        }
+
         vk::SurfaceKHR CreateSurface(vk::Instance instance);
         void DestroySurface(vk::Instance instance, vk::SurfaceKHR surface);
 
@@ -65,6 +76,10 @@ namespace Engine {
         bool m_FramebufferResized = false;
         EventCallbackFn m_EventCallback;
 
+        // UI 输入捕获（ImGui 接管时置位）
+        CaptureFn m_KeyboardCapture;
+        CaptureFn m_MouseCapture;
+
         // GLFW 静态回调（通过 window user pointer 找回 this）
         static void FramebufferResizeCallback(GLFWwindow* window, int width, int height);
         static void KeyCallback(GLFWwindow* window, int key, int scancode, int action, int mods);

Engine/include/Engine/RenderGraph/RenderGraph.h

@@ -70,8 +70,9 @@ namespace Engine::RG {
 
     struct TextureDesc {
         vk::Format format = vk::Format::eUndefined;
-        u32 width = 0;   // 0 = 跟随目标（swapchain）尺寸
+        u32 width = 0;   // 0 = 跟随目标（swapchain）尺寸 × sizeScale
         u32 height = 0;
+        f32 sizeScale = 1.0f;   // 相对尺寸（后处理降采样用，如 bloom 0.5）
     };
 
     // ---------- 物理镜像 ----------
@@ -154,6 +155,7 @@ namespace Engine::RG {
         ExecuteFn exec;
         bool sideEffect = false;              // 无读写但必须执行（compute 写 buffer 等）
         bool culled = false;
+        vk::Extent2D extent{};                // 本 pass 的渲染尺寸（写入附件的最大分辨率，Compile 期解析）
 
         // Compile 期预算好的 barrier：每个资源两个变体
         //   steady = 该槽位已被使用过（从 resting 状态出发）
@@ -225,6 +227,13 @@ namespace Engine::RG {
         // 每帧执行：按编译好的顺序录命令。要求 cmd 处于 recording 状态。
         void Execute(const FrameContext& frame);
 
+        // GPU timestamp（可选）：注入 pool 后，每个 pass 前后各写一枚
+        // 时间戳，索引 = frameSlot * queriesPerFrame + 执行序位置 * 2。
+        // pool 由调用方创建（容量 ≥ framesInFlight × queriesPerFrame），
+        // 读取由调用方在本槽位 fence 等待之后进行。
+        void SetQueryPool(vk::QueryPool pool, u32 queriesPerFrame);
+        std::vector<std::string> GetPassNames() const;
+
         // ---- 统计 ----
 
         u32 GetPassCount() const;
@@ -274,6 +283,10 @@ namespace Engine::RG {
         // 拓扑排序后的执行序（存 pass 下标）
         std::vector<u32> m_ExecutionOrder;
 
+        // GPU timestamp（可选注入）
+        vk::QueryPool m_QueryPool{};
+        u32 m_QueriesPerFrame = 0;
+
         // 物理资源池：key = (format, extent, usage)，value = 空闲镜像列表
         struct PoolKey {
             u32 format;

Engine/include/Engine/Renderer/Renderer.h

@@ -64,6 +64,8 @@ namespace Engine {
         void BuildGraph();          // 声明 pass 结构 + 首次 Compile
         void HandleResize();
         void UpdateShadowDescriptors();
+        void UpdateFullscreenDescriptors();
+        void ReadGpuTimes(u32 frameSlot);
 
         Window& m_Window;
 
@@ -96,10 +98,30 @@ namespace Engine {
         vk::ShaderEXT m_ShadowVertShader{};
         vk::ShaderEXT m_ShadowFragShader{};
 
+        // Tonemap pass 装配：Set0 = {HDR 颜色图, Bloom 图}，push = 16B
+        //（全屏 pass 通用：exposure/aces/bloomStrength/pad）
+        vk::PipelineLayout m_TonemapPipelineLayout;
+        vk::DescriptorSetLayout m_TonemapSetLayout;
+        std::vector<vk::DescriptorSet> m_TonemapDescriptorSets; // 每帧槽位一个
+        vk::Sampler m_TonemapSampler;
+        vk::ShaderEXT m_TonemapVertShader{};
+        vk::ShaderEXT m_TonemapFragShader{};
+
+        // Bloom：Extract(HDR→A) → BlurH(A→B) → BlurV(B→A)，半分辨率
+        // 与 Tonemap 共用 set layout/pipeline layout/vertex shader
+        std::vector<vk::DescriptorSet> m_BloomExtractSets;  // 每帧槽位
+        std::vector<vk::DescriptorSet> m_BloomBlurHSets;
+        std::vector<vk::DescriptorSet> m_BloomBlurVSets;
+        vk::ShaderEXT m_BloomExtractFrag{};
+        vk::ShaderEXT m_BloomBlurFrag{};
+
         // ---- RenderGraph 资源句柄 ----
         RG::TextureHandle m_Backbuffer{};
         RG::TextureHandle m_Depth{};
         RG::TextureHandle m_ShadowMap{};
+        RG::TextureHandle m_HDRColor{};   // Forward 输出（线性 HDR），Tonemap 归一
+        RG::TextureHandle m_BloomA{};     // 半分辨率：提取 + 垂直模糊结果
+        RG::TextureHandle m_BloomB{};     // 半分辨率：水平模糊中间结果
 
         // ---- 当帧绑定（Render() 期间有效，供 pass 回调读取）----
         ECS::Scene* m_BoundScene = nullptr;
@@ -109,6 +131,36 @@ namespace Engine {
         // 可选叠加层（不拥有；Application 持有）
         ImGuiLayer* m_ImGuiLayer = nullptr;
 
+        // ---- Tonemap/Bloom 参数（ImGui 可调）----
+        f32 m_Exposure = 1.0f;
+        bool m_UseACES = true;
+        f32 m_BloomThreshold = 1.0f;   // HDR 域亮度阈值
+        f32 m_BloomStrength = 0.8f;    // 叠加强度
+
+        // GPU timestamp（pass 级耗时，ImGui 可读）
+        vk::QueryPool m_QueryPool;
+        u32 m_QueriesPerFrame = 0;              // 每帧槽位的查询数 = pass 数 × 2
+        f32 m_TimestampPeriod = 1.0f;           // ns/tick（设备属性缓存）
+        std::vector<std::string> m_PassNames;   // 执行序 pass 名
+        std::vector<f32> m_PassGpuMs;           // 最近一帧各 pass 耗时
+        bool m_TimesPrimed[2] = {false, false}; // 每槽位首帧跳过读取
+
+    public:
+        f32 GetExposure() const { return m_Exposure; }
+        void SetExposure(f32 e) { m_Exposure = e; }
+        bool GetUseACES() const { return m_UseACES; }
+        void SetUseACES(bool b) { m_UseACES = b; }
+        f32 GetBloomThreshold() const { return m_BloomThreshold; }
+        void SetBloomThreshold(f32 v) { m_BloomThreshold = v; }
+        f32 GetBloomStrength() const { return m_BloomStrength; }
+        void SetBloomStrength(f32 v) { m_BloomStrength = v; }
+
+        // pass 级 GPU 耗时（最近完成帧，毫秒；与 GetPassGpuMs 下标对应）
+        const std::vector<std::string>& GetPassNames() const { return m_PassNames; }
+        const std::vector<f32>& GetPassGpuMs() const { return m_PassGpuMs; }
+
+    private:
+
         vk::Extent2D m_Extent{};
         u32 m_CurrentFrame = 0;
     };

Engine/include/Engine/Vulkan/DescriptorManager.h

@@ -42,7 +42,10 @@ namespace Engine {
         vk::PipelineLayout CreatePipelineLayout(vk::DescriptorSetLayout layout,
                                                  u32 pushConstantSize);
 
-        // 多 set layout 的新接口
+        // 多 set layout 的新接口。
+        // push constant range 按程序级 vertex|fragment 声明（与
+        // ShaderManager 的 shader EXT 声明一致，VUID-08878）；
+        // push 调用点必须带同样的双 stage（VUID-01796）
         vk::PipelineLayout CreatePipelineLayout(
             const std::vector<vk::DescriptorSetLayout>& layouts,
             u32 pushConstantSize);

Engine/src/Core/Application.cpp

@@ -11,6 +11,8 @@
 #include <Engine/Vulkan/Device.h>
 #include <Engine/ImGui/ImGuiLayer.h>
 
+#include <imgui.h>
+
 #include <algorithm>
 
 namespace Engine {
@@ -57,6 +59,12 @@ namespace Engine {
 
         m_ImGuiLayer = CreateScope<ImGuiLayer>(GetWindow(), *m_Renderer);
         m_Renderer->SetImGuiLayer(m_ImGuiLayer.get());
+
+        // 输入隔离：ImGui 想要的输入不再透给游戏侧
+        //（事件回调与轮询接口在 Window 层统一门控）
+        GetWindow().SetInputCapture(
+            [] { return ImGui::GetIO().WantCaptureKeyboard; },
+            [] { return ImGui::GetIO().WantCaptureMouse; });
     }
 
     void Application::Run()

Engine/src/Core/Window.cpp

@@ -29,6 +29,9 @@ namespace Engine
         auto* self = static_cast<Window*>(glfwGetWindowUserPointer(window));
         if (!self->m_EventCallback || key == GLFW_KEY_UNKNOWN) return;
 
+        // UI 捕获键盘时，游戏侧不收键盘事件（ImGui 后端链式回调自行处理）
+        if (self->m_KeyboardCapture && self->m_KeyboardCapture()) return;
+
         Event e;
         e.key = key;
         e.type = (action == GLFW_PRESS)   ? EventType::KeyPress
@@ -41,6 +44,7 @@ namespace Engine
     {
         auto* self = static_cast<Window*>(glfwGetWindowUserPointer(window));
         if (!self->m_EventCallback) return;
+        if (self->m_MouseCapture && self->m_MouseCapture()) return;
 
         Event e;
         e.button = button;
@@ -53,6 +57,7 @@ namespace Engine
     {
         auto* self = static_cast<Window*>(glfwGetWindowUserPointer(window));
         if (!self->m_EventCallback) return;
+        if (self->m_MouseCapture && self->m_MouseCapture()) return;
 
         Event e;
         e.type = EventType::MouseMove;
@@ -65,6 +70,7 @@ namespace Engine
     {
         auto* self = static_cast<Window*>(glfwGetWindowUserPointer(window));
         if (!self->m_EventCallback) return;
+        if (self->m_MouseCapture && self->m_MouseCapture()) return;
 
         Event e;
         e.type = EventType::MouseScroll;
@@ -165,6 +171,8 @@ namespace Engine
 
     bool Window::IsKeyPressed(int key) const
     {
+        // UI 捕获键盘时，轮询接口同样静默（相机等轮询型消费方）
+        if (m_KeyboardCapture && m_KeyboardCapture()) return false;
         return glfwGetKey(m_Window, key) == GLFW_PRESS;
     }
 
@@ -191,5 +199,11 @@ namespace Engine
 
         lastX = mx;
         lastY = my;
+
+        // UI 捕获鼠标时 delta 归零；lastX/Y 照常更新，避免释放后跳变
+        if (m_MouseCapture && m_MouseCapture())
+        {
+            dx = dy = 0.0f;
+        }
     }
 } // namespace Engine

Engine/src/RenderGraph/RenderGraph.cpp

@@ -8,6 +8,7 @@
 
 #include <algorithm>
 #include <set>
+#include <stdexcept>
 
 namespace Engine::RG {
 
@@ -164,6 +165,22 @@ namespace Engine::RG {
 
         for (auto& entry : m_Textures) ResolveUsage(entry);
 
+        // 每 pass 渲染尺寸 = 其写入附件的最大分辨率（半分辨率后处理的
+        // viewport/渲染区域由此决定）；无写入（如 Present）回退目标尺寸
+        for (auto& pass : m_Passes) {
+            u32 w = 0, h = 0;
+            for (const auto& use : pass.uses) {
+                if (!IsWriteUse(use.use)) continue;
+                const auto& d = m_Textures[use.resIndex].desc;
+                w = std::max(w, d.width ? d.width
+                                        : std::max(1u, u32(m_TargetExtent.width * d.sizeScale)));
+                h = std::max(h, d.height ? d.height
+                                         : std::max(1u, u32(m_TargetExtent.height * d.sizeScale)));
+            }
+            pass.extent = vk::Extent2D(w ? w : m_TargetExtent.width,
+                                       h ? h : m_TargetExtent.height);
+        }
+
         CullPasses();
         Toposort();
         ComputeLifetimes();
@@ -272,34 +289,43 @@ namespace Engine::RG {
 
     void RenderGraph::Toposort()
     {
-        // writer → reader 建边（去重），Kahn 拓扑排序。
-        // 声明顺序通常已是拓扑序，但引擎级实现必须显式校验，
-        // 避免上层重构 pass 顺序时引入环。
+        // 按资源在声明序上的使用序列建边，Kahn 拓扑排序。
+        //   RAW：reader → 依赖声明序上它的上一个 writer
+        //   WAR：writer → 依赖上一次 write 之后出现过的所有 reader
+        //   WAW：writer → 上一个 writer
+        // 多写者（ping-pong 资源）合法：若对 reader 连接"所有"
+        // writer（含声明序在其后的），BlurH 会同时收到 Extract→BlurH 与
+        // BlurV→BlurH，而 BlurH→BlurV 又存在，直接成环。
+        // 声明顺序通常已是拓扑序，但引擎级实现必须显式校验。
 
-        std::vector<std::vector<u32>> writers(m_Textures.size());
-        for (u32 i = 0; i < m_Passes.size(); ++i) {
+        const u32 N = static_cast<u32>(m_Passes.size());
+        std::vector<std::set<u32>> adj(N);
+        std::vector<u32> indegree(N, 0);
+
+        auto addEdge = [&](u32 from, u32 to) {
+            if (from == to) return;
+            if (adj[from].insert(to).second) ++indegree[to];
+        };
+
+        constexpr u32 kNoWriter = UINT32_MAX;
+        std::vector<u32> lastWriter(m_Textures.size(), kNoWriter);
+        std::vector<std::set<u32>> pendingReaders(m_Textures.size());
+
+        for (u32 i = 0; i < N; ++i) {
             if (m_Passes[i].culled) continue;
+
             for (const auto& use : m_Passes[i].uses) {
+                auto& lw = lastWriter[use.resIndex];
+                auto& readers = pendingReaders[use.resIndex];
+
                 if (IsWriteUse(use.use)) {
-                    writers[use.resIndex].push_back(i);
-                }
-            }
-        }
-
-        std::vector<std::set<u32>> adj(m_Passes.size());
-        std::vector<u32> indegree(m_Passes.size(), 0);
-
-        for (u32 q = 0; q < m_Passes.size(); ++q) {
-            if (m_Passes[q].culled) continue;
-            std::set<u32> seenWriters;
-            for (const auto& use : m_Passes[q].uses) {
-                if (IsWriteUse(use.use)) continue;
-                for (u32 p : writers[use.resIndex]) {
-                    if (p == q || m_Passes[p].culled) continue;
-                    if (seenWriters.insert(p).second) {
-                        adj[p].insert(q);
-                        ++indegree[q];
-                    }
+                    for (u32 r : readers) addEdge(r, i);  // WAR
+                    readers.clear();
+                    if (lw != kNoWriter) addEdge(lw, i);  // WAW
+                    lw = i;
+                } else {
+                    if (lw != kNoWriter) addEdge(lw, i);  // RAW
+                    readers.insert(i);
                 }
             }
         }
@@ -309,7 +335,7 @@ namespace Engine::RG {
 
         // 同层按声明序出队，保证确定性
         std::set<u32> ready;
-        for (u32 i = 0; i < m_Passes.size(); ++i) {
+        for (u32 i = 0; i < N; ++i) {
             if (!m_Passes[i].culled && indegree[i] == 0) ready.insert(i);
         }
 
@@ -458,9 +484,14 @@ namespace Engine::RG {
         for (auto& entry : m_Textures) {
             if (entry.imported || entry.firstPass < 0) continue;
 
+            // 显式尺寸优先；否则 = 目标尺寸 × sizeScale（后处理降采样）
             vk::Extent2D extent{
-                entry.desc.width ? entry.desc.width : m_TargetExtent.width,
-                entry.desc.height ? entry.desc.height : m_TargetExtent.height
+                entry.desc.width
+                    ? entry.desc.width
+                    : std::max(1u, u32(m_TargetExtent.width * entry.desc.sizeScale)),
+                entry.desc.height
+                    ? entry.desc.height
+                    : std::max(1u, u32(m_TargetExtent.height * entry.desc.sizeScale))
             };
 
             for (u32 slot = 0; slot < m_FramesInFlight; ++slot) {
@@ -567,20 +598,55 @@ namespace Engine::RG {
         const auto& entry = GetEntry(h);
         if (entry.imported) return entry.importedView;
         if (frameSlot >= entry.physical.size() || !entry.physical[frameSlot]) {
-            ENGINE_FATAL("RenderGraph: no physical image for '{}' slot {} "
-                         "(compile first?)", entry.name, frameSlot);
+            // ENGINE_FATAL 只记日志不中止，这里必须 throw 才能拦住空指针
+            //（Application::Run 的 catch 会接住并干净退出）
+            throw std::runtime_error(
+                "RenderGraph: no physical image for '" + entry.name +
+                "' slot " + std::to_string(frameSlot));
         }
         return entry.physical[frameSlot]->view;
     }
 
+    void RenderGraph::SetQueryPool(vk::QueryPool pool, u32 queriesPerFrame)
+    {
+        m_QueryPool = pool;
+        m_QueriesPerFrame = queriesPerFrame;
+    }
+
+    std::vector<std::string> RenderGraph::GetPassNames() const
+    {
+        std::vector<std::string> names;
+        names.reserve(m_ExecutionOrder.size());
+        for (const u32 passIdx : m_ExecutionOrder) {
+            names.push_back(m_Passes[passIdx].name);
+        }
+        return names;
+    }
+
     // ---------- 执行 ----------
 
     void RenderGraph::Execute(const FrameContext& frame)
     {
+        // GPU timestamp：可选。池由 Renderer 创建注入；每帧开头重置本槽位
+        // 区间（上一轮结果已在 fence 等待后被读取），每个 pass 前后各一枚
+        // 时间戳（BOTTOM_OF_PIPE），索引按执行序排布
+        u32 tsPos = 0;
+        if (m_QueryPool) {
+            frame.cmd.resetQueryPool(m_QueryPool,
+                                     frame.frameSlot * m_QueriesPerFrame,
+                                     m_QueriesPerFrame);
+        }
+
         for (const u32 passIdx : m_ExecutionOrder) {
             auto& pass = m_Passes[passIdx];
             if (pass.culled) continue;
 
+            if (m_QueryPool) {
+                frame.cmd.writeTimestamp2(
+                    vk::PipelineStageFlagBits2::eBottomOfPipe,
+                    m_QueryPool, frame.frameSlot * m_QueriesPerFrame + tsPos * 2);
+            }
+
             // barrier：按槽位状态挑选变体，批量提交
             m_ScratchBarriers.clear();
             for (const auto& bp : pass.barriers) {
@@ -614,10 +680,17 @@ namespace Engine::RG {
                 RenderGraphContext ctx;
                 ctx.cmd = frame.cmd;
                 ctx.frame = frame;
-                ctx.extent = m_TargetExtent;
+                ctx.extent = pass.extent;
                 ctx.graph = this;
                 pass.exec(ctx);
             }
+
+            if (m_QueryPool) {
+                frame.cmd.writeTimestamp2(
+                    vk::PipelineStageFlagBits2::eBottomOfPipe,
+                    m_QueryPool, frame.frameSlot * m_QueriesPerFrame + tsPos * 2 + 1);
+            }
+            ++tsPos;
         }
     }
 

Engine/src/Renderer/Renderer.cpp

@@ -56,6 +56,16 @@ namespace Engine
         constexpr u32 kShadowMapSize = 2048;
         constexpr u32 kShadowPushSize = static_cast<u32>(2 * sizeof(glm::mat4));
 
+        // 全屏 pass 通用 push（16B）：所有全屏 pass（Tonemap/Bloom 系）
+        // 统一声明与推送，规避 VUID-08878 的"range 定义必须一致"
+        struct FullscreenPush {
+            f32 exposure;
+            f32 param0;   // 各 pass 自定语义（tonemap: aces 开关）
+            f32 param1;
+            f32 param2;
+        };
+        constexpr u32 kFullscreenPushSize = static_cast<u32>(sizeof(FullscreenPush));
+
         // 光源正交包围盒（覆盖当前场景；后续按场景 AABB 自适应）
         constexpr f32 kLightOrthoExtent = 15.0f;
         constexpr f32 kLightDistance = 30.0f;
@@ -147,6 +157,60 @@ namespace Engine
                          .setCompareEnable(VK_FALSE);
         m_ShadowSampler = m_Device->GetDevice().createSampler(shadowSamplerInfo);
 
+        // 7.5 全屏后处理装配（Tonemap + Bloom 共用一套 layout/sampler）：
+        // Set0 = {HDR 颜色图, Bloom 图}，push = 16B 通用结构
+        std::vector<DescriptorBinding> fullscreenBindings = {
+            {0, vk::DescriptorType::eCombinedImageSampler,
+             vk::ShaderStageFlagBits::eFragment},
+            {1, vk::DescriptorType::eCombinedImageSampler,
+             vk::ShaderStageFlagBits::eFragment},
+        };
+        m_TonemapSetLayout = m_DescriptorManager->CreateLayout(fullscreenBindings);
+
+        m_TonemapPipelineLayout = m_DescriptorManager->CreatePipelineLayout(
+            std::vector<vk::DescriptorSetLayout>{m_TonemapSetLayout},
+            kFullscreenPushSize);
+
+        m_TonemapSampler = m_Device->GetDevice().createSampler(
+            vk::SamplerCreateInfo{}
+                .setMagFilter(vk::Filter::eLinear)
+                .setMinFilter(vk::Filter::eLinear)
+                .setMipmapMode(vk::SamplerMipmapMode::eNearest)
+                .setAddressModeU(vk::SamplerAddressMode::eClampToEdge)
+                .setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
+                .setAddressModeW(vk::SamplerAddressMode::eClampToEdge));
+
+        auto tonemapVert = m_ShaderManager->CreateFromFile(
+            shaderDir + "tonemap.vert.spv",
+            vk::ShaderStageFlagBits::eVertex,
+            std::vector<vk::DescriptorSetLayout>{m_TonemapSetLayout},
+            kFullscreenPushSize);
+        auto tonemapFrag = m_ShaderManager->CreateFromFile(
+            shaderDir + "tonemap.frag.spv",
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{m_TonemapSetLayout},
+            kFullscreenPushSize);
+        auto bloomExtractFrag = m_ShaderManager->CreateFromFile(
+            shaderDir + "bloom_extract.frag.spv",
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{m_TonemapSetLayout},
+            kFullscreenPushSize);
+        auto bloomBlurFrag = m_ShaderManager->CreateFromFile(
+            shaderDir + "bloom_blur.frag.spv",
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{m_TonemapSetLayout},
+            kFullscreenPushSize);
+
+        m_TonemapVertShader = tonemapVert.Handle;
+        m_TonemapFragShader = tonemapFrag.Handle;
+        m_BloomExtractFrag = bloomExtractFrag.Handle;
+        m_BloomBlurFrag = bloomBlurFrag.Handle;
+
+        if (!m_TonemapVertShader || !m_TonemapFragShader ||
+            !m_BloomExtractFrag || !m_BloomBlurFrag) {
+            ENGINE_FATAL("Failed to create fullscreen shader objects");
+        }
+
         // 8. Swapchain（只管呈现；深度图由 RenderGraph 创建）
         m_Swapchain = CreateScope<Swapchain>(*m_Device, *m_Allocator, m_Window);
         m_Extent = m_Swapchain->GetExtent();
@@ -162,6 +226,10 @@ namespace Engine
 
         m_UBODescriptorSets.resize(kMaxFramesInFlight);
         m_ShadowDescriptorSets.resize(kMaxFramesInFlight);
+        m_TonemapDescriptorSets.resize(kMaxFramesInFlight);
+        m_BloomExtractSets.resize(kMaxFramesInFlight);
+        m_BloomBlurHSets.resize(kMaxFramesInFlight);
+        m_BloomBlurVSets.resize(kMaxFramesInFlight);
         for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
             m_UBODescriptorSets[slot] =
                 m_DescriptorManager->AllocateDescriptorSet(m_UBOSetLayout);
@@ -171,13 +239,37 @@ namespace Engine
 
             m_ShadowDescriptorSets[slot] =
                 m_DescriptorManager->AllocateDescriptorSet(m_ShadowSetLayout);
+
+            m_TonemapDescriptorSets[slot] =
+                m_DescriptorManager->AllocateDescriptorSet(m_TonemapSetLayout);
+
+            m_BloomExtractSets[slot] =
+                m_DescriptorManager->AllocateDescriptorSet(m_TonemapSetLayout);
+            m_BloomBlurHSets[slot] =
+                m_DescriptorManager->AllocateDescriptorSet(m_TonemapSetLayout);
+            m_BloomBlurVSets[slot] =
+                m_DescriptorManager->AllocateDescriptorSet(m_TonemapSetLayout);
         }
 
         // 11. RenderGraph：声明 + 编译（场景在 Render(Scene&) 时按帧绑定）
         BuildGraph();
 
-        // Compile 已分配物理镜像：把各槽位的阴影图 view 绑进描述符
+        // Compile 已分配物理镜像：把各槽位的阴影图/HDR view 绑进描述符
         UpdateShadowDescriptors();
+        UpdateFullscreenDescriptors();
+
+        // 12. GPU timestamp：每 pass 两枚（前后），按帧槽位分区
+        m_QueriesPerFrame = static_cast<u32>(m_RenderGraph->GetPassCount()) * 2;
+        m_PassNames = m_RenderGraph->GetPassNames();
+        m_PassGpuMs.assign(m_RenderGraph->GetPassCount(), 0.0f);
+
+        vk::QueryPoolCreateInfo qpInfo;
+        qpInfo.setQueryType(vk::QueryType::eTimestamp)
+              .setQueryCount(m_QueriesPerFrame * kMaxFramesInFlight);
+        m_QueryPool = m_Device->GetDevice().createQueryPool(qpInfo);
+        m_RenderGraph->SetQueryPool(m_QueryPool, m_QueriesPerFrame);
+        m_TimestampPeriod = static_cast<f32>(
+            m_Device->GetPhysicalDevice().getProperties().limits.timestampPeriod);
 
         ENGINE_INFO("Renderer initialized");
     }
@@ -200,13 +292,31 @@ namespace Engine
         if (m_ShadowFragShader) {
             vkDevice.destroyShaderEXT(m_ShadowFragShader);
         }
+        if (m_TonemapVertShader) {
+            vkDevice.destroyShaderEXT(m_TonemapVertShader);
+        }
+        if (m_TonemapFragShader) {
+            vkDevice.destroyShaderEXT(m_TonemapFragShader);
+        }
+        if (m_BloomExtractFrag) {
+            vkDevice.destroyShaderEXT(m_BloomExtractFrag);
+        }
+        if (m_BloomBlurFrag) {
+            vkDevice.destroyShaderEXT(m_BloomBlurFrag);
+        }
 
         vkDevice.destroySampler(m_ShadowSampler);
+        vkDevice.destroySampler(m_TonemapSampler);
+        if (m_QueryPool) {
+            vkDevice.destroyQueryPool(m_QueryPool);
+        }
         vkDevice.destroyPipelineLayout(m_PipelineLayout);
         vkDevice.destroyPipelineLayout(m_ShadowPipelineLayout);
+        vkDevice.destroyPipelineLayout(m_TonemapPipelineLayout);
         vkDevice.destroyDescriptorSetLayout(m_UBOSetLayout);
         vkDevice.destroyDescriptorSetLayout(m_TextureSetLayout);
         vkDevice.destroyDescriptorSetLayout(m_ShadowSetLayout);
+        vkDevice.destroyDescriptorSetLayout(m_TonemapSetLayout);
     }
 
     vk::Format Renderer::GetSwapchainFormat() const
@@ -243,6 +353,21 @@ namespace Engine
             "ShadowMap",
             RG::TextureDesc{vk::Format::eD32Sfloat, kShadowMapSize, kShadowMapSize});
 
+        // HDR 颜色图：Forward 的输出（线性光照值），Tonemap 显式完成
+        // 显示变换后写入 Backbuffer。对齐 UE/Unity 管线：
+        // SceneColor(HDR) → tonemapper → UI（UI 在 tonemap 之后叠加）
+        m_HDRColor = m_RenderGraph->CreateTexture(
+            "HDRColor",
+            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat});
+
+        // Bloom ping-pong：半分辨率（sizeScale 相对目标尺寸，resize 自动跟随）
+        m_BloomA = m_RenderGraph->CreateTexture(
+            "BloomA",
+            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat, 0, 0, 0.5f});
+        m_BloomB = m_RenderGraph->CreateTexture(
+            "BloomB",
+            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat, 0, 0, 0.5f});
+
         // ---- Shadow pass：光源视角写深度（所有渲染体都写入，不做实体级开关——
         //      castShadow 应由材质/实体数据驱动剔除，而非 pass 逻辑特判）----
         m_RenderGraph->AddPass("Shadow",
@@ -321,8 +446,10 @@ namespace Engine
                         if (!model) return;
 
                         const ShadowPush push{m_LightSpace, transform.world};
+                        // stage 必须覆盖 layout range 声明的全部 stage（VUID-01796）
                         cmd.pushConstants(m_ShadowPipelineLayout,
-                                          vk::ShaderStageFlagBits::eVertex,
+                                          vk::ShaderStageFlagBits::eVertex |
+                                          vk::ShaderStageFlagBits::eFragment,
                                           0, kShadowPushSize, &push);
 
                         for (auto& mesh : model->GetMeshes()) {
@@ -345,7 +472,7 @@ namespace Engine
                     std::array<float, 4>{0.1f, 0.1f, 0.1f, 1.0f});
 
                 vk::RenderingAttachmentInfo colorAttachment;
-                colorAttachment.setImageView(ctx.View(m_Backbuffer))
+                colorAttachment.setImageView(ctx.View(m_HDRColor))
                                .setImageLayout(vk::ImageLayout::eColorAttachmentOptimal)
                                .setLoadOp(vk::AttachmentLoadOp::eClear)
                                .setStoreOp(vk::AttachmentStoreOp::eStore)
@@ -447,8 +574,10 @@ namespace Engine
                         Model* model = m_BoundAssets->GetModel(meshRenderer.model);
                         if (!model) return;
 
+                        // stage 必须覆盖 layout range 声明的全部 stage（VUID-01796）
                         cmd.pushConstants(m_PipelineLayout,
-                                          vk::ShaderStageFlagBits::eVertex,
+                                          vk::ShaderStageFlagBits::eVertex |
+                                          vk::ShaderStageFlagBits::eFragment,
                                           0, sizeof(glm::mat4), &transform.world);
 
                         for (auto& mesh : model->GetMeshes()) {
@@ -461,11 +590,137 @@ namespace Engine
 
                 cmd.endRendering();
             })
-            .Write(m_Backbuffer, RG::ResourceUse::ColorAttachment)
+            .Write(m_HDRColor, RG::ResourceUse::ColorAttachment)
             .Write(m_Depth, RG::ResourceUse::DepthAttachment)
             .Read(m_ShadowMap, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment);
 
+        // ---- 全屏 pass 公共录制：动态状态/draw 全一致，仅附件 view、
+        //      输入 set、push 值不同（Extract/BlurH/BlurV/Tonemap 复用）----
+        auto recordFullscreen =
+            [this](RG::RenderGraphContext& ctx, vk::ImageView view,
+                   vk::ShaderEXT frag, const vk::DescriptorSet& set,
+                   const FullscreenPush& push) {
+                const vk::CommandBuffer cmd = ctx.cmd;
+
+                vk::ClearValue clearValue;
+                clearValue.color = vk::ClearColorValue(
+                    std::array<float, 4>{0.0f, 0.0f, 0.0f, 1.0f});
+
+                vk::RenderingAttachmentInfo colorAttachment;
+                colorAttachment.setImageView(view)
+                               .setImageLayout(vk::ImageLayout::eColorAttachmentOptimal)
+                               .setLoadOp(vk::AttachmentLoadOp::eClear)
+                               .setStoreOp(vk::AttachmentStoreOp::eStore)
+                               .setClearValue(clearValue);
+
+                vk::RenderingInfo renderingInfo;
+                renderingInfo.setRenderArea({{0, 0}, ctx.extent})
+                             .setLayerCount(1)
+                             .setColorAttachments(colorAttachment);
+
+                cmd.beginRendering(renderingInfo);
+
+                std::array<vk::ShaderStageFlagBits, 2> stages = {
+                    vk::ShaderStageFlagBits::eVertex,
+                    vk::ShaderStageFlagBits::eFragment
+                };
+                std::array<vk::ShaderEXT, 2> shaders = {
+                    m_TonemapVertShader, frag};
+                cmd.bindShadersEXT(stages, shaders);
+
+                cmd.bindDescriptorSets(vk::PipelineBindPoint::eGraphics,
+                                       m_TonemapPipelineLayout, 0, set, {});
+
+                // ---- 动态状态（无深度附件、无顶点输入）----
+                cmd.setRasterizerDiscardEnable(VK_FALSE);
+                cmd.setDepthTestEnable(VK_FALSE);
+                cmd.setDepthWriteEnable(VK_FALSE);
+                cmd.setDepthBoundsTestEnable(VK_FALSE);
+                cmd.setStencilTestEnable(VK_FALSE);
+                cmd.setDepthBiasEnable(VK_FALSE);
+                cmd.setPolygonModeEXT(vk::PolygonMode::eFill);
+                cmd.setCullMode(vk::CullModeFlagBits::eNone);
+                cmd.setFrontFace(vk::FrontFace::eCounterClockwise);
+                cmd.setPrimitiveTopology(vk::PrimitiveTopology::eTriangleList);
+                cmd.setPrimitiveRestartEnable(VK_FALSE);
+                cmd.setVertexInputEXT(
+                    std::array<vk::VertexInputBindingDescription2EXT, 0>{},
+                    std::array<vk::VertexInputAttributeDescription2EXT, 0>{});
+                cmd.setRasterizationSamplesEXT(vk::SampleCountFlagBits::e1);
+                vk::SampleMask sampleMask = 0xFFFFFFFF;
+                cmd.setSampleMaskEXT(vk::SampleCountFlagBits::e1, sampleMask);
+                cmd.setAlphaToCoverageEnableEXT(VK_FALSE);
+                cmd.setAlphaToOneEnableEXT(VK_FALSE);
+
+                vk::Viewport viewport(0.0f, 0.0f,
+                                      static_cast<float>(ctx.extent.width),
+                                      static_cast<float>(ctx.extent.height),
+                                      0.0f, 1.0f);
+                cmd.setViewportWithCount(viewport);
+                cmd.setScissorWithCount(vk::Rect2D({0, 0}, ctx.extent));
+
+                cmd.pushConstants(m_TonemapPipelineLayout,
+                                  vk::ShaderStageFlagBits::eVertex |
+                                  vk::ShaderStageFlagBits::eFragment,
+                                  0, kFullscreenPushSize, &push);
+
+                cmd.draw(3, 1, 0, 0);
+                cmd.endRendering();
+            };
+
+        // ---- Bloom：Extract（HDR 软阈值）→ BlurH → BlurV，全部半分辨率 ----
+        m_RenderGraph->AddPass("BloomExtract",
+            [this, recordFullscreen](RG::RenderGraphContext& ctx) {
+                recordFullscreen(ctx, ctx.View(m_BloomA), m_BloomExtractFrag,
+                                 m_BloomExtractSets[ctx.frame.frameSlot],
+                                 FullscreenPush{0.0f, m_BloomThreshold, 0.0f, 0.0f});
+            })
+            .Read(m_HDRColor, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Write(m_BloomA, RG::ResourceUse::ColorAttachment);
+
+        m_RenderGraph->AddPass("BloomBlurH",
+            [this, recordFullscreen](RG::RenderGraphContext& ctx) {
+                recordFullscreen(ctx, ctx.View(m_BloomB), m_BloomBlurFrag,
+                                 m_BloomBlurHSets[ctx.frame.frameSlot],
+                                 FullscreenPush{0.0f,
+                                                1.0f / ctx.extent.width, 0.0f,
+                                                0.0f});
+            })
+            .Read(m_BloomA, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Write(m_BloomB, RG::ResourceUse::ColorAttachment);
+
+        m_RenderGraph->AddPass("BloomBlurV",
+            [this, recordFullscreen](RG::RenderGraphContext& ctx) {
+                recordFullscreen(ctx, ctx.View(m_BloomA), m_BloomBlurFrag,
+                                 m_BloomBlurVSets[ctx.frame.frameSlot],
+                                 FullscreenPush{0.0f, 0.0f,
+                                                1.0f / ctx.extent.height,
+                                                0.0f});
+            })
+            .Read(m_BloomB, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Write(m_BloomA, RG::ResourceUse::ColorAttachment);
+
+        // ---- Tonemap pass：HDR（+bloom）→ 显示空间，全屏三角无顶点缓冲 ----
+        // 显示变换在此显式建模（此前隐式依赖 Srgb 附件，是设计欠账）。
+        // Backbuffer 从此只接收"显示空间"颜色：tonemap 直接写、ImGui 叠加。
+        m_RenderGraph->AddPass("Tonemap",
+            [this, recordFullscreen](RG::RenderGraphContext& ctx) {
+                recordFullscreen(ctx, ctx.View(m_Backbuffer), m_TonemapFragShader,
+                                 m_TonemapDescriptorSets[ctx.frame.frameSlot],
+                                 FullscreenPush{m_Exposure,
+                                                m_UseACES ? 1.0f : 0.0f,
+                                                m_BloomStrength, 0.0f});
+            })
+            .Write(m_Backbuffer, RG::ResourceUse::ColorAttachment)
+            .Read(m_HDRColor, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_BloomA, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment);
+
         // ---- ImGui pass：叠加层（loadOp=Load，不清屏）----
         // 框架只负责把 draw data 录进 cmd；UI 内容全部由派生类
         // 在 OnRenderOverlay() 里构建，框架自身不画任何 ImGui 控件。
@@ -520,8 +775,9 @@ namespace Engine
         // RG 重编译：物理镜像按新尺寸轮换
         m_RenderGraph->SetTargetExtent(m_Extent);
 
-        // 重编译后物理镜像可能变化，阴影描述符重新绑定
+        // 重编译后物理镜像可能变化，描述符重新绑定
         UpdateShadowDescriptors();
+        UpdateFullscreenDescriptors();
     }
 
     void Renderer::UpdateShadowDescriptors()
@@ -535,6 +791,66 @@ namespace Engine
         }
     }
 
+    void Renderer::UpdateFullscreenDescriptors()
+    {
+        // 全屏 pass 的输入绑定（per-slot 物理镜像）：
+        //   Tonemap   = {0: HDRColor, 1: BloomA}
+        //   Extract   = {0: HDRColor}        （binding1 不用，可不写）
+        //   BlurH     = {0: BloomA}
+        //   BlurV     = {0: BloomB}
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            const vk::ImageView hdr =
+                m_RenderGraph->GetPhysicalView(m_HDRColor, slot);
+            const vk::ImageView bloomA =
+                m_RenderGraph->GetPhysicalView(m_BloomA, slot);
+            const vk::ImageView bloomB =
+                m_RenderGraph->GetPhysicalView(m_BloomB, slot);
+
+            m_DescriptorManager->UpdateTexture(
+                m_TonemapDescriptorSets[slot], 0, m_TonemapSampler, hdr);
+            m_DescriptorManager->UpdateTexture(
+                m_TonemapDescriptorSets[slot], 1, m_TonemapSampler, bloomA);
+
+            m_DescriptorManager->UpdateTexture(
+                m_BloomExtractSets[slot], 0, m_TonemapSampler, hdr);
+            m_DescriptorManager->UpdateTexture(
+                m_BloomBlurHSets[slot], 0, m_TonemapSampler, bloomA);
+            m_DescriptorManager->UpdateTexture(
+                m_BloomBlurVSets[slot], 0, m_TonemapSampler, bloomB);
+        }
+    }
+
+    void Renderer::ReadGpuTimes(u32 frameSlot)
+    {
+        if (!m_QueryPool || m_QueriesPerFrame == 0) return;
+
+        // 每槽位首帧：该区间的查询尚未被 reset/写入，跳过一次读取
+        if (!m_TimesPrimed[frameSlot]) {
+            m_TimesPrimed[frameSlot] = true;
+            return;
+        }
+
+        const u32 passCount = m_QueriesPerFrame / 2;
+        std::vector<u64> ticks(m_QueriesPerFrame, 0);
+
+        // fence 已等待本槽位上一轮完成，结果必然可用（阻塞式 + 64bit）
+        const vk::Result r = m_Device->GetDevice().getQueryPoolResults(
+            m_QueryPool,
+            frameSlot * m_QueriesPerFrame, m_QueriesPerFrame,
+            sizeof(u64) * m_QueriesPerFrame,
+            ticks.data(), sizeof(u64),
+            vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait);
+        if (r != vk::Result::eSuccess) return;
+
+        for (u32 i = 0; i < passCount; ++i) {
+            const u64 start = ticks[i * 2];
+            const u64 end = ticks[i * 2 + 1];
+            m_PassGpuMs[i] = (end >= start)
+                ? static_cast<f32>(end - start) * m_TimestampPeriod / 1.0e6f
+                : 0.0f;
+        }
+    }
+
     void Renderer::Render(ECS::Scene& scene, Assets::AssetManager& assets)
     {
         m_BoundScene = &scene;
@@ -611,6 +927,7 @@ namespace Engine
         // 此时本槽位的 fence 尚未 reset——先等它（保证 GPU 已完成本槽位上一轮），
         // 再写 UBO，写-读竞争关闭。
         m_SyncManager->WaitFence(frameSlot);
+        ReadGpuTimes(frameSlot);   // fence 已等：上一轮该槽位的查询已就绪
         m_UniformBuffer->Update(frameSlot, &ubo, sizeof(ubo));
         m_SyncManager->ResetFence(frameSlot);
 

Engine/src/Vulkan/DescriptorManager.cpp

@@ -115,14 +115,19 @@ namespace Engine
         return m_Device.GetDevice().createPipelineLayout(info);
     }
 
-    vk::PipelineLayout DescriptorManager::CreatePipelineLayout(const std::vector<vk::DescriptorSetLayout>& layouts,u32 pushConstantSize)
+    vk::PipelineLayout DescriptorManager::CreatePipelineLayout(const std::vector<vk::DescriptorSetLayout>& layouts,
+                                                               u32 pushConstantSize)
     {
         vk::PipelineLayoutCreateInfo info;
         info.setSetLayouts(layouts);
 
         std::array<vk::PushConstantRange, 1> pushRanges;
         if (pushConstantSize > 0) {
-            pushRanges[0].setStageFlags(vk::ShaderStageFlagBits::eVertex)
+            // 与 ShaderManager 一致：程序级 vertex|fragment 声明。
+            // 注意 vkCmdPushConstants 的 stageFlags 必须覆盖 range 的
+            // 全部 stage（VUID-01796），因此 push 调用点也要带双 stage
+            pushRanges[0].setStageFlags(vk::ShaderStageFlagBits::eVertex |
+                                        vk::ShaderStageFlagBits::eFragment)
                          .setOffset(0)
                          .setSize(pushConstantSize);
             info.setPushConstantRanges(pushRanges);

Engine/src/Vulkan/ShaderManager.cpp

@@ -57,7 +57,12 @@ ShaderObject ShaderManager::CreateFromSPIRV(const std::vector<u32>& spirv,
 
     std::array<vk::PushConstantRange, 1> pushRanges;
     if (pushConstantSize > 0) {
-        pushRanges[0].setStageFlags(vk::ShaderStageFlagBits::eVertex)
+        // push range 是"程序级"声明：shader object 要求同一程序绑定的
+        // 全部 shader（vert+frag）push range 定义完全一致（VUID-08878），
+        // 因此统一按 vertex|fragment 声明；shader 实际读取哪些字节由
+        // GLSL 的 push_constant block 决定，声明范围只是能力上限
+        pushRanges[0].setStageFlags(vk::ShaderStageFlagBits::eVertex |
+                                    vk::ShaderStageFlagBits::eFragment)
                      .setOffset(0)
                      .setSize(pushConstantSize);
         createInfo.setPushConstantRanges(pushRanges);

Engine/src/Vulkan/Swapchain.cpp

@@ -134,9 +134,13 @@ namespace Engine
 
     vk::SurfaceFormatKHR Swapchain::ChooseSurfaceFormat(const std::vector<vk::SurfaceFormatKHR>& formats)
     {
+        // 用 Unorm 而非 Srgb：当前管线（场景光照 + ImGui）都是把
+        // "显示空间"颜色直接写进 framebuffer，若用 Srgb 附件硬件会再
+        // 做一次 linear→sRGB 提亮 → 整体泛白（ImGui 原生蓝 #4296FA
+        // 会被洗成 #8CCBFC）。等上 HDR/tonemap 管线后再切回 Srgb。
         for (const auto& f : formats)
         {
-            if (f.format == vk::Format::eB8G8R8A8Srgb &&
+            if (f.format == vk::Format::eB8G8R8A8Unorm &&
                 f.colorSpace == vk::ColorSpaceKHR::eSrgbNonlinear)
             {
                 return f;

shaders/bloom_blur.frag

@@ -0,0 +1,29 @@
+#version 450
+
+// Bloom 高斯模糊：可分离 9-tap，方向由 push 给出（H/V 各跑一次）。
+// 源图为半分辨率，线性采样 + 权重对应的偏移天然抗锯齿。
+
+layout(location = 0) in vec2 fragUV;
+layout(location = 0) out vec4 outColor;
+
+layout(set = 0, binding = 0) uniform sampler2D srcImage;
+
+layout(push_constant) uniform PushConstants {
+    float exposure;   // 未用（程序级统一布局）
+    float stepX;      // param0: UV 步长 x（水平 pass = 1/width，垂直 = 0）
+    float stepY;      // param1: UV 步长 y
+    float _pad0;
+} pc;
+
+const float weights[5] = float[](0.227027, 0.1945946, 0.1216216, 0.054054, 0.016216);
+
+void main() {
+    vec2 off = vec2(pc.stepX, pc.stepY);
+
+    vec3 c = texture(srcImage, fragUV).rgb * weights[0];
+    for (int i = 1; i < 5; ++i) {
+        c += texture(srcImage, fragUV + off * float(i)).rgb * weights[i];
+        c += texture(srcImage, fragUV - off * float(i)).rgb * weights[i];
+    }
+    outColor = vec4(c, 1.0);
+}

shaders/bloom_extract.frag

@@ -0,0 +1,25 @@
+#version 450
+
+// Bloom 提取：从 HDR 颜色图中滤出超过阈值的部分（软阈值，保留超出比例）。
+// 全屏 pass 通用 push（16B）：param0 = 亮度阈值（HDR 域，1.0 = 仅高光溢出）。
+
+layout(location = 0) in vec2 fragUV;
+layout(location = 0) out vec4 outColor;
+
+layout(set = 0, binding = 0) uniform sampler2D hdrImage;
+
+layout(push_constant) uniform PushConstants {
+    float exposure;   // 未用（程序级统一布局）
+    float threshold;  // param0
+    float _pad0;
+    float _pad1;
+} pc;
+
+void main() {
+    vec3 c = texture(hdrImage, fragUV).rgb;
+    float l = max(c.r, max(c.g, c.b));
+
+    // 软阈值：只保留超出 threshold 的能量比例，避免硬切产生的块状边缘
+    float k = max(l - pc.threshold, 0.0) / max(l, 1e-4);
+    outColor = vec4(c * k, 1.0);
+}

shaders/tonemap.frag

@@ -0,0 +1,38 @@
+#version 450
+
+// Tonemap pass：HDR 线性颜色 → 显示空间。
+// 这是显示变换的显式建模（对齐 UE/Unity：HDR SceneColor → tonemapper → UI）。
+// ACES = Narkowicz 拟合（线性域映射到 [0,1]），可经 push constant 关闭
+// 退回纯 gamma（旧观感）；bloom 等后处理在 HDR 域（Tonemap 之前）插 pass。
+
+layout(location = 0) in vec2 fragUV;
+layout(location = 0) out vec4 outColor;
+
+layout(set = 0, binding = 0) uniform sampler2D hdrImage;
+layout(set = 0, binding = 1) uniform sampler2D bloomImage;
+
+// 全屏 pass 通用 push（16B，程序级统一声明，VUID-08878）
+layout(push_constant) uniform PushConstants {
+    float exposure;      // 曝光系数
+    float aces;          // >0.5 启用 ACES 曲线，否则纯 gamma
+    float bloomStrength; // bloom 叠加强度（0 = 关闭）
+    float _pad0;
+} pc;
+
+vec3 acesFilm(vec3 x) {
+    return clamp((x * (2.51 * x + 0.03)) / (x * (2.43 * x + 0.59) + 0.14), 0.0, 1.0);
+}
+
+void main() {
+    // bloom 在 HDR 域叠加（半分辨率模糊图线性采样上采样）
+    vec3 hdr = texture(hdrImage, fragUV).rgb
+             + texture(bloomImage, fragUV).rgb * pc.bloomStrength;
+
+    hdr *= pc.exposure;
+
+    vec3 mapped = (pc.aces > 0.5)
+        ? acesFilm(hdr)
+        : clamp(hdr, 0.0, 1.0);
+
+    outColor = vec4(pow(mapped, vec3(1.0 / 2.2)), 1.0);
+}

shaders/tonemap.vert

@@ -0,0 +1,23 @@
+#version 450
+
+// 全屏三角形：无顶点缓冲，从 gl_VertexIndex 派生 UV。
+//  0 → (-1,3)? 经典三角覆盖整个屏幕：
+//  idx0:(-1,-1) idx1:(3,-1) idx2:(-1,3)，UV 相应 [0..2]，
+//  光栅化插值后屏幕内 UV 恰好落在 [0,1]。
+layout(location = 0) out vec2 fragUV;
+
+// 全屏 pass 通用 push（16B，与 tonemap.frag 及所有全屏 pass 一致，VUID-08878）
+layout(push_constant) uniform PushConstants {
+    float exposure;
+    float aces;
+    float _pad0;
+    float _pad1;
+} pc;
+
+void main() {
+    vec2 p = vec2(
+        float((gl_VertexIndex << 1) & 2),
+        float(gl_VertexIndex & 2));
+    fragUV = p;
+    gl_Position = vec4(p * 2.0 - 1.0, 0.0, 1.0);
+}