feat: 模型异步加载——assimp 解析移后台线程（SunTemple 3s 卡顿消除）

Model 拆两段式：ParseFromDisk（static 纯 CPU，assimp 解析 → 几何 +
材质引用，worker 可调）/ FromParsed（GPU 构建：纹理上传 / mesh 缓冲 /
材质描述符，主线程），同步路径等价两段串行。第二个异步任务类型落地，
线程收割骨架抽成 Core/AsyncTask<T>（thread + atomic + Take，覆盖式
Run；pending 接力语义留调用方）。

实测（SunTemple 1056 mesh）：解析 2.5s 全程在 worker，主线程持续
渲染；GPU 构建 0.56s。assets 层经 IAssetUploader/AssetManager 重载
传预解析产物（路径缓存去重，路径已缓存则产物丢弃）。

已知边界：纹理解码（backpack 4k PNG ~1s）仍在 GPU 段主线程——
Texture 需支持预解码像素注入，后续演进。

App/CMakeLists.txt

@@ -33,7 +33,8 @@ file(GLOB_RECURSE SHADER_SOURCES
         CONFIGURE_DEPENDS                          # 新增 shader 时自动重配
         RELATIVE ${SHADER_SRC_DIR}
         ${SHADER_SRC_DIR}/*.vert
-        ${SHADER_SRC_DIR}/*.frag)
+        ${SHADER_SRC_DIR}/*.frag
+        ${SHADER_SRC_DIR}/*.comp)
 
 foreach (f ${SHADER_SOURCES})
     auto_compile_shader(${f})

App/main.cpp

@@ -11,25 +11,45 @@
 #include <Engine/Core/Application.h>
 #include <Engine/Core/Log.h>
 #include <Engine/Core/Stats.h>
+#include <Engine/Core/AsyncTask.h>
 #include <Engine/ECS/Scene.h>
 #include <Engine/ECS/EntityHandle.h>
 #include <Engine/ECS/Components.h>
 #include <Engine/Assets/AssetManager.h>
 #include <Engine/Renderer/Renderer.h>
+#include <Engine/Renderer/IblGenerator.h>
 
 #include <imgui.h>
 
 #include <glm/glm.hpp>
 #include <glm/gtc/matrix_transform.hpp>
+#include <glm/gtc/type_ptr.hpp>
 
 #include <GLFW/glfw3.h>
 
+#include <atomic>
+#include <thread>
 
 using namespace Engine;
 using namespace Engine::ECS;
 
 namespace
 {
+    // 可切换环境（IBL 动态切换演示）：路径 → AssetManager 缓存去重，
+    // 重复选择同一环境零成本；首选路径触发一次 compute 预积分
+    constexpr const char* kEnvPaths[] = {
+        "assets/textures/venice_dawn_1_4k.hdr",
+        "assets/textures/rooitou_park_4k.hdr",
+        "assets/textures/pink_sunrise_4k.hdr",
+        "assets/textures/birchwood_4k.hdr",
+    };
+    constexpr const char* kEnvNames[] = {
+        "Venice Dawn", "Rooitou Park", "Pink Sunrise", "Birchwood",
+    };
+    constexpr int kEnvCount = 4;
+    constexpr int kDefaultEnv = 0;
+    constexpr const char* kTemplePath = "assets/models/SunTemple/SunTemple.fbx";
+
     class SandboxApp final : public Application
     {
     public:
@@ -38,6 +58,15 @@ namespace
         {
         }
 
+        ~SandboxApp() override
+        {
+            // worker 只做 CPU 解码（不碰 Vulkan），在任何 GPU 资源
+            // 析构之前 join 即安全；派生析构先于基类成员析构执行
+            if (m_EnvWorker.joinable()) {
+                m_EnvWorker.join();
+            }
+        }
+
     protected:
         // ---- 生命周期 ----
 
@@ -48,14 +77,36 @@ namespace
             // ImGui 叠加层（框架只做管线桥接，UI 内容全部在这里定义）
             EnableImGui();
 
+            // 默认环境（IBL 四件套）：经资产层加载（compute 预积分，毫秒级），
+            // 其余环境在 ImGui 里按需 lazy 加载后切换
+            const auto env = GetAssets().LoadEnvironment(
+                kEnvPaths[kDefaultEnv]);
+            if (env.Valid()) {
+                GetRenderer().SetEnvironment(GetAssets().GetEnvironment(env));
+            }
+
             // 模型资产（PBR 验证场景：backpack 为旧式贴图资产，
             // 退化为电介质因子路径——albedo+normal，metallic=0/roughness=0.5）
             const auto model = GetAssets().LoadModel("assets/models/backpack/backpack.obj");
             // const auto model = GetAssets().LoadModel("assets/models/pocket_watch_4k.gltf/pocket_watch_4k.gltf");
 
-            m_Spinners.push_back(scene.CreateEntity("backpack"));
-            m_Spinners[0].AddComponent<MeshRendererComponent>(MeshRendererComponent{model});
-            m_Spinners[0].GetComponent<TransformComponent>().SetScale(glm::vec3{1.0});
+            // 大模型异步加载：assimp 解析在 worker 线程（SunTemple 约 3s），
+            // 完成后 OnUpdate 收割 → GPU 构建 → 建实体（cm 尺度 → 0.001）
+            m_TempleTask.Run([] { return Model::ParseFromDisk(kTemplePath); });
+            m_TempleScale = 0.001f;
+
+
+            int countSize = 0;
+            for (int i = 0; i < countSize; i ++)
+            {
+                for (int j = 0; j < countSize; j ++)
+                {
+                    auto m = scene.CreateEntity();
+                    m.AddComponent<MeshRendererComponent>(MeshRendererComponent{model});
+                    m.GetComponent<TransformComponent>().SetPosition(glm::vec3{(i - countSize * 0.5) * 2, 0, (j - countSize * 0.5) * 2});
+                    m_Spinners.push_back(m);
+                }
+            }
 
             // 地面：程序化纯色平面（半边长 15，中心 y=-2）
             const auto ground = GetAssets().CreatePlane(15.0f, 140, 140, 145);
@@ -112,6 +163,24 @@ namespace
 
         void OnUpdate(f32 deltaTime) override
         {
+            // 异步模型收割：解析完成 → GPU 构建 → 建实体
+            if (m_TempleTask.Done()) {
+                ParsedModel parsed = m_TempleTask.Take();
+                const auto ref = GetAssets().LoadModel(kTemplePath,
+                                                       std::move(parsed));
+                if (ref.Valid()) {
+                    auto temple = GetScene().CreateEntity("SunTemple");
+                    temple.AddComponent<MeshRendererComponent>(
+                        MeshRendererComponent{ref});
+                    temple.GetComponent<TransformComponent>()
+                           .SetScale(glm::vec3{m_TempleScale});
+                    ENGINE_INFO("[ASYNC] SunTemple ready, entity created");
+                }
+            }
+
+            // 异步环境加载：每帧收割完成的解码结果
+            PollEnvironment();
+
             // 暂停时冻结动画时钟（帧历史验证：暂停后 Temporal Diff
             // 视图应在 2 帧内整屏变黑；相机移动同样会被 diff 捕捉到）
             if (!m_PauseAnim)
@@ -135,13 +204,72 @@ namespace
 
         // ---- 事件 ----
 
+        // ---- 异步环境加载（decode 后台线程 / GPU 生成主线程）----
+        // decode 占切换耗时大头（4k ≈ 130ms）；worker 只产出纯 CPU
+        // 像素数据，Vulkan 生成回主线程（设备非线程安全）。
+        // 同一时刻最多一个 in-flight，忙碌期的新请求记为 pending 接力。
+        void RequestEnvironment(int idx)
+        {
+            if (m_EnvLoadBusy)
+            {
+                m_EnvPending = idx;
+                return;
+            }
+            StartEnvDecode(idx);
+        }
+
+        void StartEnvDecode(int idx)
+        {
+            m_EnvLoadBusy = true;
+            m_EnvLoadDone.store(false);
+            m_EnvLoadingPath = kEnvPaths[idx];
+            m_EnvWorker = std::thread([this, idx]() {
+                m_EnvDecoded = IblGenerator::DecodeHdr(kEnvPaths[idx]);
+                m_EnvLoadDone.store(true);   // release：解码结果对主线程可见
+            });
+        }
+
+        void PollEnvironment()
+        {
+            if (!m_EnvLoadBusy || !m_EnvLoadDone.load())
+            {
+                return;
+            }
+
+            m_EnvWorker.join();   // join = 同步点，解码结果此后归主线程
+            m_EnvLoadBusy = false;
+
+            if (m_EnvDecoded.Valid())
+            {
+                const auto ref = GetAssets().LoadEnvironment(
+                    m_EnvLoadingPath, std::move(m_EnvDecoded));
+                if (ref.Valid())
+                {
+                    GetRenderer().SetEnvironment(GetAssets().GetEnvironment(ref));
+                }
+            }
+            m_EnvDecoded = {};
+
+            // 忙碌期间的最新请求接力（rapid switch 只保留最终意图）
+            if (m_EnvPending >= 0)
+            {
+                const int next = m_EnvPending;
+                m_EnvPending = -1;
+                StartEnvDecode(next);
+            }
+        }
+
         // ---- 叠加层 UI（唯一允许出现 ImGui 控件的地方）----
 
         void OnRenderOverlay() override
         {
             ImGui::Begin("Scene Debug");
+            auto& renderer = GetRenderer();
 
-            // 阴影光源方向：拖动滑条实时看影子变化
+            // ---- Scene：场景内容（灯/动画） ----
+            if (ImGui::CollapsingHeader("Scene", ImGuiTreeNodeFlags_DefaultOpen))
+            {
+                // 阴影光源（无开关，控件平铺聚拢）
                 if (auto* light = m_KeyLight.TryGetComponent<LightComponent>())
                 {
                     ImGui::Text("Key Light (shadow caster)");
@@ -154,11 +282,13 @@ namespace
                         light->color.a = 2.0f;
                     }
                 }
+                ImGui::Checkbox("Pause Animation", &m_PauseAnim);
+            }
 
-            // Tonemap + Bloom 参数（HDR 后处理输出端）
-            auto& renderer = GetRenderer();
-            ImGui::Separator();
-            ImGui::Text("Tonemap");
+            // ---- Rendering：渲染管线参数与开关 ----
+            if (ImGui::CollapsingHeader("Rendering", ImGuiTreeNodeFlags_DefaultOpen))
+            {
+                // Tonemap（无开关，平铺）
                 f32 exposure = renderer.GetExposure();
                 if (ImGui::SliderFloat("exposure", &exposure, 0.1f, 4.0f))
                 {
@@ -170,7 +300,14 @@ namespace
                     renderer.SetUseACES(aces);
                 }
 
-            ImGui::Text("Bloom");
+                // Bloom：开关即显隐——勾上才显示参数
+                bool bloom = renderer.GetBloomEnabled();
+                if (ImGui::Checkbox("Bloom", &bloom))
+                {
+                    renderer.SetBloomEnabled(bloom);
+                }
+                if (bloom)
+                {
                     f32 threshold = renderer.GetBloomThreshold();
                     if (ImGui::SliderFloat("threshold", &threshold, 0.0f, 2.0f))
                     {
@@ -181,32 +318,78 @@ namespace
                     {
                         renderer.SetBloomStrength(strength);
                     }
-
-            // 调试视图：帧历史差异（静止=黑，运动边缘发光=链路正确）
-            bool temporalDiff = renderer.GetShowTemporalDiff();
-            if (ImGui::Checkbox("Temporal Diff View", &temporalDiff))
-            {
-                renderer.SetShowTemporalDiff(temporalDiff);
                 }
-            ImGui::Checkbox("Pause Animation", &m_PauseAnim);
 
+                // 无参数的纯开关：横排平铺
+                bool shadows = renderer.GetShadowsEnabled();
+                if (ImGui::Checkbox("Shadows", &shadows))
+                {
+                    renderer.SetShadowsEnabled(shadows);
+                }
+                ImGui::SameLine();
+                bool taa = renderer.GetTAAEnabled();
+                if (ImGui::Checkbox("TAA", &taa))
+                {
+                    renderer.SetTAAEnabled(taa);
+                }
+            }
+
+            // ---- Environment：IBL 环境 + 调试视图（决定"看什么"） ----
+            if (ImGui::CollapsingHeader("Environment", ImGuiTreeNodeFlags_DefaultOpen))
+            {
+                // 动态环境切换（Assets 层加载 → Renderer 换绑；缓存去重）
+                ImGui::Text("Environment (IBL)");
+                bool& DrawSkybox = GetScene().GetSettings().DrawSkybox;
+                ImGui::Checkbox("DrawSkybox", &DrawSkybox);
+                if (DrawSkybox)
+                {
+                    int envIdx = m_EnvIndex;
+                    if (ImGui::Combo("env", &envIdx, kEnvNames, kEnvCount))
+                    {
+                        // 异步：解码后台线程，完成后 OnUpdate 里换绑
+                        m_EnvIndex = envIdx;
+                        RequestEnvironment(envIdx);
+                    }
+                    ImGui::Text(m_EnvLoadBusy ? "(decoding...)" :
+                                "(switch = compute pre-integrate, ms-level)");
+
+                    // 调试视图：单看某个 pass 的中间输出（类 UE pass debug）
+                    const auto& viewNames = renderer.GetDebugViewNames();
+                    int view = renderer.GetDebugView();
+                    if (ImGui::Combo("Debug View", &view, viewNames.data(),
+                                     static_cast<int>(viewNames.size())))
+                    {
+                        renderer.SetDebugView(view);
+                    }
+                }else
+                {
+                    ImGui::ColorEdit4("ClearColor", glm::value_ptr(GetScene().GetSettings().ClearColor));
+                }
+            }
+
+            // ---- GPU Timings：只读遥测（默认折叠，不占面板高度） ----
+            if (ImGui::CollapsingHeader("GPU Timings"))
+            {
                 // Pass 级 GPU 耗时（RG timestamp）
-            ImGui::Separator();
-            ImGui::Text("GPU Timings");
                 ImGui::Text("instances: %u main / %u shadow",
                             renderer.GetMainInstanceCount(),
                             renderer.GetShadowInstanceCount());
+                if (renderer.GetInstancesTruncated()) {
+                    ImGui::TextColored(ImVec4(1.0f, 0.3f, 0.3f, 1.0f),
+                        "TRUNCATED: instance buffer full!");
+                }
                 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]);
                 }
+            }
 
-            // 帧率控制
+            // ---- Frame：帧率控制与显示 ----
+            if (ImGui::CollapsingHeader("Frame", ImGuiTreeNodeFlags_DefaultOpen))
+            {
                 auto& app = *this;
-            ImGui::Separator();
-            ImGui::Text("Frame");
                 bool vsync = app.GetVSync();
                 if (ImGui::Checkbox("VSync", &vsync))
                 {
@@ -225,11 +408,10 @@ namespace
                         app.SetTargetFps(target);
                     }
                 }
-
-            ImGui::Separator();
                 ImGui::Text("fps: %.1f  frame: %llu",
                             1.0f / GetDeltaTime(),
                             static_cast<unsigned long long>(GetFrameCount()));
+            }
 
             ImGui::End();
 
@@ -305,6 +487,19 @@ namespace
         std::vector<EntityHandle> m_Spinners;
         f32 m_Time = 0.0f;
         bool m_PauseAnim = true;   // 冻结动画时钟（帧历史 diff 验证用）
+        int m_EnvIndex = kDefaultEnv;
+
+        // ---- 异步环境加载状态（单 in-flight + pending 接力）----
+        std::thread m_EnvWorker;
+        std::string m_EnvLoadingPath;
+        DecodedEquirect m_EnvDecoded;          // worker 写 → join 后主线程读
+        std::atomic<bool> m_EnvLoadDone{false};
+        bool m_EnvLoadBusy = false;
+        int m_EnvPending = -1;
+
+        // ---- 异步模型加载（SunTemple）----
+        AsyncTask<ParsedModel> m_TempleTask;
+        f32 m_TempleScale = 1.0f;   // cm 尺度资产 → 米（解析后应用）
     };
 }
 

Engine/include/Engine/Assets/AssetManager.h

@@ -27,13 +27,15 @@
 #include <unordered_map>
 #include <vector>
 
-namespace Engine {
-    namespace Assets {
-
+namespace Engine
+{
+    namespace Assets
+    {
         // ---------- 通用资产缓存（不感知 Vulkan / 磁盘细节由 loader 决定） ----------
 
         template <typename T>
-        class AssetCache {
+        class AssetCache
+        {
         public:
             // loader：path → 资产实例（失败返回 nullptr）
             using LoaderFn = std::function<Scope<T>(const std::string&)>;
@@ -41,23 +43,28 @@ namespace Engine {
             // 加载（带路径去重）。失败返回无效句柄。
             AssetRef<T> Load(const std::string& path, const LoaderFn& loader)
             {
-                if (auto it = m_ByPath.find(path); it != m_ByPath.end()) {
+                if (auto it = m_ByPath.find(path); it != m_ByPath.end())
+                {
                     return it->second;
                 }
 
                 Scope<T> asset = loader(path);
-                if (!asset) {
+                if (!asset)
+                {
                     ENGINE_ERROR("AssetCache: failed to load '{}'", path);
                     return {};
                 }
 
                 AssetRef<T> h;
-                if (!m_Free.empty()) {
+                if (!m_Free.empty())
+                {
                     h.index = m_Free.back();
                     m_Free.pop_back();
                     h.generation = m_Entries[h.index].generation;
                     m_Entries[h.index] = Entry{path, h.generation, std::move(asset), true};
-                } else {
+                }
+                else
+                {
                     h.index = static_cast<u32>(m_Entries.size());
                     h.generation = 1;
                     m_Entries.push_back(Entry{path, h.generation, std::move(asset), true});
@@ -70,12 +77,14 @@ namespace Engine {
             // 句柄 → 资产指针。悬空句柄 FATAL（bug 信号）；已卸载返回 nullptr。
             T* Get(AssetRef<T> h) const
             {
-                if (!h.Valid() || h.index >= m_Entries.size()) {
+                if (!h.Valid() || h.index >= m_Entries.size())
+                {
                     ENGINE_FATAL("AssetCache::Get: invalid handle (index={}, gen={})",
                                  h.index, h.generation);
                 }
                 const Entry& e = m_Entries[h.index];
-                if (e.generation != h.generation) {
+                if (e.generation != h.generation)
+                {
                     ENGINE_FATAL("AssetCache::Get: stale handle for '{}' (gen {} != {})",
                                  e.path, h.generation, e.generation);
                 }
@@ -114,7 +123,8 @@ namespace Engine {
             }
 
         private:
-            struct Entry {
+            struct Entry
+            {
                 std::string path;
                 u32 generation = 1;
                 Scope<T> asset;
@@ -128,7 +138,8 @@ namespace Engine {
 
         // ---------- AssetManager：全引擎资产入口 ----------
 
-        class AssetManager {
+        class AssetManager
+        {
         public:
             // 上传服务由组合根（Application）注入：渲染层实现接口，
             // 资产层不感知 Vulkan 具体类
@@ -139,12 +150,23 @@ namespace Engine {
             AssetManager& operator=(const AssetManager&) = delete;
 
             ModelRef LoadModel(const std::string& path);
+            // 异步路径：解析已在后台线程完成（路径已缓存则产物丢弃去重）
+            ModelRef LoadModel(const std::string& path, ParsedModel&& parsed);
             Model* GetModel(ModelRef h) const { return m_Models.Get(h); }
             void UnloadModel(ModelRef h) { m_Models.Unload(h); }
 
             // 程序化平面（纯色贴图）：size 为半边长。带缓存，同参数共享一份。
             ModelRef CreatePlane(f32 halfSize, u8 r, u8 g, u8 b);
 
+            // 环境资产（HDR → IBL 预积分四件套）。路径缓存去重，
+            // 重复 Load 同一路径零成本（动态切换场景直接反复调用）
+            EnvRef LoadEnvironment(const std::string& path);
+            // 异步路径：解码已在后台线程完成（若路径已缓存，解码产物被
+            // 丢弃去重——上层可先查缓存再起 worker，此处不做强约束）
+            EnvRef LoadEnvironment(const std::string& path,
+                                   DecodedEquirect&& decoded);
+            Cubemap* GetEnvironment(EnvRef h) const { return m_Environments.Get(h); }
+
             AssetCache<Model>& Models() { return m_Models; }
 
             // 后续扩展点：TextureHandle / MaterialHandle / ShaderHandle
@@ -154,8 +176,8 @@ namespace Engine {
             IAssetUploader& m_Uploader;
 
             AssetCache<Model> m_Models;
+            AssetCache<Cubemap> m_Environments;
         };
-
     } // namespace Assets
 } // namespace Engine
 

Engine/include/Engine/Assets/IAssetUploader.h

@@ -19,6 +19,7 @@
 
 #include <Engine/Core/Types.h>
 #include <Engine/Renderer/Model.h>
+#include <Engine/Renderer/Cubemap.h>
 
 namespace Engine::Assets {
 
@@ -29,8 +30,19 @@ namespace Engine::Assets {
         // 磁盘模型 → 解析 + GPU 上传（失败返回 nullptr）
         virtual Scope<Model> LoadModel(const std::string& path) = 0;
 
+        // 异步路径：解析已在后台线程完成，此调用只做 GPU 构建（主线程）
+        virtual Scope<Model> LoadModel(const std::string& path,
+                                       ParsedModel&& parsed) = 0;
+
         // 程序化纯色平面（size 为半边长）
         virtual Scope<Model> CreatePlane(f32 halfSize, u8 r, u8 g, u8 b) = 0;
+
+        // 环境 HDR → IBL 预积分四件套（compute 生成，毫秒级；失败返回 nullptr）
+        virtual Scope<Cubemap> LoadEnvironment(const std::string& path) = 0;
+
+        // 异步加载路径：解码已在后台线程完成，此调用只做 GPU 生成（主线程）
+        virtual Scope<Cubemap> LoadEnvironment(const std::string& path,
+                                               DecodedEquirect&& decoded) = 0;
     };
 
 } // namespace Engine::Assets

Engine/include/Engine/Core/AssetRef.h

@@ -19,6 +19,7 @@
 // Engine::Assets::Engine::Model 这类错误嵌套命名空间）
 namespace Engine {
     class Model;
+    class Cubemap;
     class Texture;
     class Material;
 }
@@ -39,6 +40,7 @@ namespace Engine {
 
     // 常用别名（新增资产类型时在这里加一行）
     using ModelRef = AssetRef<Model>;
+    using EnvRef = AssetRef<Cubemap>;   // 环境（IBL 四件套）
 
 } // namespace Engine
 

Engine/include/Engine/Core/AsyncTask.h

@@ -0,0 +1,82 @@
+//
+// Created by atdunbg on 2026/9/24.
+//
+// AsyncTask：单 in-flight 异步任务的通用收割工具（header-only）。
+//
+// 异步加载的公共骨架（已验证两次：环境 DecodeHdr / 模型 ParseFromDisk）：
+//   worker 线程跑纯 CPU 任务 → 原子标志通知 → 主线程每帧轮询收割。
+//   Vulkan 提交/描述符等 GPU 工作全部留在主线程（设备非线程安全）。
+//
+//   AsyncTask<ParsedModel> task;
+//   task.Run([] { return Model::ParseFromDisk(path); });  // 发起
+//   if (task.Done()) { auto p = task.Take(); ... }        // 每帧收割
+//   // 析构自动 join（worker 只做纯 CPU 工作，join 恒安全）
+//
+// 语义：Run 在忙碌时先 join 并丢弃旧结果（覆盖式）；"快速连切只保留
+// 最终意图"的 pending 接力由调用方持有（参照 main.cpp 环境加载）。
+//
+
+#ifndef HELLOVULKAN_ASYNCTASK_H
+#define HELLOVULKAN_ASYNCTASK_H
+
+#include <Engine/Core/Types.h>
+#include <atomic>
+#include <functional>
+#include <thread>
+#include <utility>
+
+namespace Engine {
+
+    template <typename T>
+    class AsyncTask {
+    public:
+        AsyncTask() = default;
+        ~AsyncTask() { Join(); }
+
+        AsyncTask(const AsyncTask&) = delete;
+        AsyncTask& operator=(const AsyncTask&) = delete;
+
+        // 发起任务。忙碌时先等待旧任务完成并放弃其结果。
+        void Run(std::function<T()> job)
+        {
+            Join();
+            m_Busy = true;
+            m_Done.store(false);
+            m_Worker = std::thread([this, j = std::move(job)]() {
+                m_Result = j();
+                m_Done.store(true);   // release：结果对主线程 acquire 可见
+            });
+        }
+
+        bool Busy() const { return m_Busy; }
+        bool Done() const { return m_Busy && m_Done.load(); }
+
+        // 收割：join worker 并移交结果。只在 Done() 为 true 时调用。
+        T Take()
+        {
+            m_Worker.join();   // join = 同步点，此后结果归主线程
+            m_Busy = false;
+            T out = std::move(m_Result);
+            m_Result = T{};
+            return out;
+        }
+
+        // 关闭前兜底等待（析构自动调用）
+        void Join()
+        {
+            if (m_Worker.joinable()) {
+                m_Worker.join();
+            }
+            m_Busy = false;
+        }
+
+    private:
+        std::thread m_Worker;
+        std::atomic<bool> m_Done{false};
+        T m_Result{};
+        bool m_Busy = false;
+    };
+
+} // namespace Engine
+
+#endif //HELLOVULKAN_ASYNCTASK_H

Engine/include/Engine/Core/Log.h

@@ -11,6 +11,12 @@
 
 namespace Engine {
 
+    namespace Logger
+    {
+        void Initialize();
+        void Shutdown();
+    }
+
 #define ENGINE_TRACE(...)  ::spdlog::trace(__VA_ARGS__)
 #define ENGINE_INFO(...)   ::spdlog::info(__VA_ARGS__)
 #define ENGINE_DEBUG(...)   ::spdlog::debug(__VA_ARGS__)

Engine/include/Engine/ECS/Scene.h

@@ -41,7 +41,9 @@ namespace Engine::ECS {
 
         // ---- 场景级设置 ----
         struct Settings {
-            glm::vec4 ambientColor{0.07f, 0.07f, 0.09f, 0.0f};
+            glm::vec4 ClearColor = {0.1, 0.1, 0.1, 1.0};
+            glm::vec4 AmbientColor{0.07f, 0.07f, 0.09f, 0.0f};
+            bool DrawSkybox = true;
         };
         Settings& GetSettings() { return m_Settings; }
 

Engine/include/Engine/RenderGraph/RenderGraph.h

@@ -212,13 +212,6 @@ namespace Engine::RG {
         // 注意：物理镜像按槽位各一份，槽位变化/重编译后必须重新获取。
         vk::ImageView GetPhysicalView(TextureHandle h, u32 frameSlot) const;
 
-        // 调试/截图：取物理 image + 静息布局（帧尾 barrier 恢复用）。
-        // imported 资源返回 false（无物理实例）
-        bool GetCaptureTarget(TextureHandle h, u32 frameSlot,
-                              vk::Image& outImage, vk::Extent2D& outExtent,
-                              vk::Format& outFormat,
-                              vk::ImageLayout& outResting) const;
-
         // 帧历史槽位 = 上一帧使用的槽位：(slot - 1 + F) mod F
         u32 HistorySlot(u32 frameSlot) const {
             return (frameSlot + m_FramesInFlight - 1) % m_FramesInFlight;

Engine/include/Engine/Renderer/Cubemap.h

@@ -1,17 +1,17 @@
 //
 // Created by atdunbg on 2026/9/23.
 //
-// Cubemap：环境立方体贴图 + IBL 预积分资源（阶段②③，split-sum）。
+// Cubemap：环境立方体贴图 + IBL 预积分资源（split-sum 四件套）。
+//
+// 纯 GPU 资源持有类（对齐 Model 的定位）：只管 image/view/采样器的
+// 同生共死，不做任何生成/上传。预积分由 IblGenerator（compute 链）
+// 产出后注入——环境切换的入口在 Assets 层（AssetManager::LoadEnvironment）。
 //
-// 从等距柱状 HDR（.hdr，polyhaven 等）生成，CPU 一次性预积分：
 //   env cube        —— 重采样 6 面（Skybox 背景）
 //   irradiance cube —— 半球 Lambert 积分（PBR 漫反射 ambient）
 //   prefilter cube  —— GGX 重要性采样 × mip 分级 roughness（镜面 IBL）
 //   BRDF LUT        —— 2D 积分表（纯数学，与环境内容无关）
-// 全部 fp16 上传，运行时采样零积分开销。
-//
-// 后续演进：CPU 预积分 → compute pass（动态环境零卡顿）；本类的
-// 方向表/上传路径届时整体迁移，接口不变。
+// 全部 fp16，运行时采样零积分开销。
 //
 
 #ifndef HELLOVULKAN_CUBEMAP_H
@@ -19,13 +19,24 @@
 
 #include <Engine/Core/Types.h>
 #include <Engine/Vulkan/MemoryAllocator.h>
-#include <glm/glm.hpp>
 #include <vulkan/vulkan.hpp>
+#include <vector>
 
 namespace Engine {
 
     class Device;
 
+    // 环境的 CPU 源数据（等距柱状 HDR 解码产物，RGBA float 行优先）。
+    // 纯数据无线程状态：异步加载时由后台线程产出（DecodeHdr），
+    // 主线程消费（FromDecoded → GPU 生成）。decode 是切换耗时大头，
+    // 搬出主线程后运行时切换只剩 GPU 段（~40ms）。
+    struct DecodedEquirect {
+        std::vector<f32> pixels;
+        u32 w = 0;
+        u32 h = 0;
+        bool Valid() const { return !pixels.empty() && w > 0 && h > 0; }
+    };
+
     // GPU 纹理资源三元组（RHI 惯例）：image/view/显存分配同生共死，
     // 析构统一释放——杜绝"建了 image 丢了 view"类半套泄漏。
     // 上传 helper 直接以此为单位产出，消费端只拿引用计数语义的整单元
@@ -37,18 +48,6 @@ namespace Engine {
 
     class Cubemap {
     public:
-        // 从等距柱状 HDR 生成（CPU 重采样 + irradiance/prefilter/BRDF 预积分）
-        static Scope<Cubemap> CreateFromHdrEquirect(Device& device,
-                                                    MemoryAllocator& allocator,
-                                                    const std::string& path,
-                                                    u32 faceSize);
-
-        // 中性兜底（无 HDR 资产时）：cube 均为常数色；BRDF LUT 仍为
-        // 真实积分（纯数学）。保证引用侧描述符恒有效（validation）
-        static Scope<Cubemap> CreateNeutral(Device& device,
-                                            MemoryAllocator& allocator,
-                                            const glm::vec3& rgb);
-
         ~Cubemap();
 
         Cubemap(const Cubemap&) = delete;
@@ -67,6 +66,8 @@ namespace Engine {
         vk::Sampler GetBrdfLutSampler() const { return m_LutSampler; }
 
     private:
+        friend class IblGenerator;   // 生成端唯一构造入口
+
         Cubemap(Device& device, MemoryAllocator& allocator,
                 EnvTexture env, EnvTexture irr, EnvTexture pref,
                 EnvTexture lut, vk::Sampler envSampler, vk::Sampler lutSampler)

Engine/include/Engine/Renderer/IblGenerator.h

@@ -0,0 +1,111 @@
+//
+// Created by atdunbg on 2026/9/24.
+//
+// IblGenerator：IBL 预积分的 compute 生成端（原 Cubemap CPU 积分路径的
+// GPU 化，"动态环境切换零卡顿"第一步）。
+//
+// 管线：等距柱状 HDR → 上传 equirect 中间纹理（免 CPU 逐像素转换）
+//       → 一次性提交四段 dispatch：
+//         ① equirect_to_cube（硬件过滤重采样 6 面）
+//         ② irradiance（半球 Lambert 卷积，采样 env cube）
+//         ③ prefilter × mip（GGX 重要性采样，push roughness）
+//         ④ BRDF LUT（纯数学）
+//       → fence 等待 → 产出 Cubemap（四件套 fp16 资源）。
+//
+// 为什么不进 RenderGraph：RG 管理的是帧内 2D 瞬时资源（无 cube/数组层/
+// storage 写/条件执行），而预积分是"资产加载时跑一次"的持久资源——
+// 独立一次性提交与原上传路径同形态，侵入最小。残留卡顿 = fence 等待
+// （毫秒级），异步 compute queue / IO 线程解码是后续演进点。
+//
+// 产出与消费端契约（SkyboxPass / SceneGpu set2 binding3-5）不变：
+// view 形态与原 CPU 路径完全一致，绑定代码零改动。
+//
+
+#ifndef HELLOVULKAN_IBLGENERATOR_H
+#define HELLOVULKAN_IBLGENERATOR_H
+
+#include <Engine/Core/Types.h>
+#include <Engine/Renderer/Cubemap.h>   // DecodedEquirect（解码产物数据契约）
+#include <Engine/Vulkan/MemoryAllocator.h>
+#include <glm/glm.hpp>
+#include <vulkan/vulkan.hpp>
+#include <string>
+#include <vector>
+
+namespace Engine {
+
+    class Device;
+    class DescriptorManager;
+    class ShaderManager;
+    class Cubemap;
+
+    class IblGenerator {
+    public:
+        IblGenerator(Device& device, MemoryAllocator& allocator,
+                     DescriptorManager& descriptors, ShaderManager& shaders);
+        ~IblGenerator();
+
+        IblGenerator(const IblGenerator&) = delete;
+        IblGenerator& operator=(const IblGenerator&) = delete;
+
+        // ---- 同步便捷路径（等价 DecodeHdr + FromDecoded）----
+        // 等距柱状 HDR → 环境四件套（失败返回 nullptr，上层走 Neutral 兜底）
+        Scope<Cubemap> FromHdrEquirect(const std::string& path, u32 faceSize);
+
+        // ---- 异步加载两段式（decode 是切换耗时大头，可搬后台线程）----
+        // 纯 CPU：HDR 文件 → 像素数据。无线程状态，worker 线程可安全调用。
+        static DecodedEquirect DecodeHdr(const std::string& path);
+
+        // GPU 生成（必须主线程）：预解码数据 → 四件套。debugName 仅用于日志。
+        Scope<Cubemap> FromDecoded(const DecodedEquirect& decoded, u32 faceSize,
+                                   const std::string& debugName);
+
+        // 中性兜底：常数色环境 + 真实 BRDF LUT（描述符恒有效）
+        Scope<Cubemap> Neutral(const glm::vec3& rgb);
+
+    private:
+        // cube 兼容 image（2D array 写 + cube 采样双 view）
+        struct CubeTarget {
+            vk::Image image{};
+            std::vector<vk::ImageView> storageViews;   // 每 mip 一个（storage 写必须单 mip）
+            vk::ImageView cubeView{};                  // eCube 全 mip（采样端）
+            MemoryAllocator::Allocation allocation;
+        };
+
+        // 从 HDR 像素数据生成（Neutral 走同一链，仅 equirect 为 1×1 常数）
+        Scope<Cubemap> Generate(const f32* pixels, u32 w, u32 h,
+                                u32 faceSize, const std::string& debugName);
+
+        CubeTarget CreateCubeTarget(u32 size, u32 mips) const;
+        void ReleaseStorageViews(const CubeTarget& t) const;
+
+        vk::Sampler CreateEnvSampler(u32 maxLod) const;
+        vk::Sampler CreateLutSampler() const;
+
+        Device& m_Device;
+        MemoryAllocator& m_Allocator;
+        DescriptorManager& m_Descriptors;
+        ShaderManager& m_Shaders;
+
+        // set 布局：A = {0 采样输入, 1 storage 输出}，B = {0 storage}
+        vk::DescriptorSetLayout m_IblSetLayout;
+        vk::DescriptorSetLayout m_LutSetLayout;
+        // 管线布局：无 push（equirect/irradiance/LUT）与 4B push（prefilter）
+        vk::PipelineLayout m_IblLayout;
+        vk::PipelineLayout m_IblLayoutPush;
+        vk::PipelineLayout m_LutLayout;
+
+        vk::ShaderEXT m_ToCubeShader{};
+        vk::ShaderEXT m_IrrShader{};
+        vk::ShaderEXT m_PrefShader{};
+        vk::ShaderEXT m_LutShader{};
+
+        vk::Sampler m_EquirectSampler;   // S=repeat（经度环绕）T=clamp
+        vk::CommandPool m_Pool;          // 持久复用，每次生成 reset
+        vk::CommandBuffer m_Cmd;
+        vk::Fence m_Fence;
+    };
+
+} // namespace Engine
+
+#endif //HELLOVULKAN_IBLGENERATOR_H

Engine/include/Engine/Renderer/Model.h

@@ -12,6 +12,7 @@
 #include <Engine/Renderer/Mesh.h>
 #include <vulkan/vulkan.hpp>
 #include <array>
+#include <filesystem>
 #include <string>
 #include <unordered_map>
 #include <vector>
@@ -28,14 +29,42 @@ namespace Engine
     class Texture;
     class Material;
 
+    // ---- CPU 解析产物（异步加载的跨线程交棒数据）----
+    // 纯数据：几何 + 材质槽位引用（路径 + 因子），无任何 GPU 资源。
+    // worker 线程产出（ParseFromDisk），主线程消费（FromParsed）。
+    struct ParsedMesh {
+        std::vector<Vertex> vertices;
+        std::vector<u32> indices;
+        std::array<std::string, Material::kSlotCount> texPaths;
+        MaterialParams params;
+    };
+    struct ParsedModel {
+        std::string path;
+        bool isGltf = false;
+        std::vector<ParsedMesh> meshes;
+        bool Valid() const { return !meshes.empty(); }
+    };
+
     class Model
     {
     public:
+        // 同步便捷路径（等价 ParseFromDisk + FromParsed，启动期小模型可用）
         Model(Device& device, MemoryAllocator& allocator,
               DescriptorManager& descriptorManager,
               vk::DescriptorSetLayout textureLayout,
               const std::string& path);
 
+        // ---- 异步加载两段式（assimp 解析占大模型耗时大头，可搬后台线程）----
+        // 纯 CPU：读盘 + assimp 解析 → 几何 + 材质引用。无线程状态。
+        static ParsedModel ParseFromDisk(const std::string& path);
+
+        // GPU 构建（必须主线程）：纹理上传 / mesh 缓冲 / 材质描述符。
+        // 解析失败返回 nullptr。
+        static Scope<Model> FromParsed(Device& device, MemoryAllocator& allocator,
+                                       DescriptorManager& descriptorManager,
+                                       vk::DescriptorSetLayout textureLayout,
+                                       ParsedModel&& parsed);
+
         // 程序化模型：单位法线朝上的 XZ 平面（纯色贴图），用于地面等。
         // size 为半边长（平面总边长 = 2*size）。
         struct ProceduralTag {};
@@ -57,9 +86,19 @@ namespace Engine
               vk::DescriptorSetLayout textureLayout,
               ProceduralTag, f32 halfSize, const std::array<u8, 3>& color);
 
+        // 空模型骨架（FromParsed 用）：只建默认纹理，mesh 由调用方填
+        Model(Device& device, MemoryAllocator& allocator,
+              DescriptorManager& descriptorManager,
+              vk::DescriptorSetLayout textureLayout);
+
         void LoadModel(const std::string& path);
-        void ProcessNode(aiNode* node, const aiScene* scene);
-        Scope<Mesh> ProcessMesh(aiMesh* mesh, const aiScene* scene);
+        static void ProcessNode(const aiNode* node, const aiScene* scene,
+                                std::vector<ParsedMesh>& out,
+                                const std::filesystem::path& modelDir,
+                                bool isGltf);
+        static ParsedMesh ProcessMesh(const aiMesh* mesh, const aiScene* scene,
+                                      const std::filesystem::path& modelDir,
+                                      bool isGltf);
         void InitDefaultTextures();
 
         Texture* GetOrLoadTexture(const std::string& path, bool srgb);

Engine/include/Engine/Renderer/Passes/BloomPass.h

@@ -11,6 +11,7 @@
 
 #include <Engine/Renderer/Passes/RenderPass.h>
 #include <Engine/Renderer/Passes/FullscreenGpu.h>
+#include <Engine/Renderer/SceneTypes.h>
 #include <vulkan/vulkan.hpp>
 #include <vector>
 
@@ -19,6 +20,7 @@ namespace Engine::Passes {
     class BloomPass final : public RenderPass {
     public:
         BloomPass(FullscreenGpu& gpu, PostProcessParams& params,
+                  FrameSceneData& frame,
                   RG::TextureHandle hdrColor,
                   RG::TextureHandle bloomA, RG::TextureHandle bloomB);
         ~BloomPass() override = default;
@@ -29,6 +31,7 @@ namespace Engine::Passes {
     private:
         FullscreenGpu& m_Gpu;
         PostProcessParams& m_Params;
+        FrameSceneData& m_Frame;   // bloomEnabled 开关（帧契约）
         RG::TextureHandle m_HDR;
         RG::TextureHandle m_A;   // 提取 + 垂直模糊结果
         RG::TextureHandle m_B;   // 水平模糊中间结果

Engine/include/Engine/Renderer/Passes/FullscreenGpu.h

@@ -31,13 +31,32 @@ namespace Engine {
             f32 param2;
         };
 
-        // 后处理链参数（ImGui 可调；Renderer 持有，pass 按引用读取）
+        // 后处理数值参数（ImGui 可调；Renderer 持有，pass 按引用读取）。
+        // 运行时开关/调试视图不在这里——统一进 FrameSceneData（帧契约）
         struct PostProcessParams {
             f32 exposure = 0.6f;
             bool useACES = true;
             f32 bloomThreshold = 1.0f;   // HDR 域亮度阈值
             f32 bloomStrength = 0.8f;    // 叠加强度
-            bool showTemporalDiff = false;   // 调试视图：显示帧历史差异
+        };
+
+        // 调试视图枚举：TonemapPass 按值把对应中间目标路由上屏。
+        // 与 Renderer 的 DebugViewNames 表下标一一对应
+        enum DebugView : int {
+            kDebugFinal = 0,      // 正常画面（tonemap 输出）
+            kDebugTemporalDiff,   // 帧历史差异（TAA 链路验证）
+            kDebugHDR,            // Forward 原始 HDR（无 bloom/TAA）
+            kDebugTAAColor,       // TAA resolve 后
+            kDebugBloom,          // bloom 模糊结果
+            kDebugVelocity,       // 运动向量（|uv| 增益显示）
+            kDebugShadowMap0,     // CSM 级联 0 深度（近=白）
+            kDebugShadowMap1,     // 级联 1
+            kDebugShadowMap2,     // 级联 2
+            kDebugEnvCube,        // IBL 环境 cube（经纬展开）
+            kDebugIrradiance,     // IBL 漫反射辐照度（32² 低频）
+            kDebugPrefilter,      // IBL pre-filter（纵向 = mip/roughness 分级）
+            kDebugBrdfLut,        // BRDF 积分 LUT（RG 双通道）
+            kDebugViewCount
         };
 
         class FullscreenGpu {

Engine/include/Engine/Renderer/Passes/RenderPass.h

@@ -19,6 +19,10 @@
 
 #include <Engine/RenderGraph/RenderGraph.h>
 
+namespace Engine {
+    class Cubemap;
+}
+
 namespace Engine::Passes {
 
     class RenderPass {
@@ -27,6 +31,10 @@ namespace Engine::Passes {
 
         virtual void Register(RG::RenderGraph& rg) = 0;
         virtual void OnTargetsRebuilt(RG::RenderGraph& rg) { (void)rg; }
+
+        // 环境资产换绑（动态切换）：引用环境描述符的 pass 重绑自己的 set。
+        // 默认无环境影响；env 恒非空（Renderer 有中性兜底）
+        virtual void OnEnvironmentChanged(const Cubemap* /*env*/) {}
     };
 
 } // namespace Engine::Passes

Engine/include/Engine/Renderer/Passes/SkyboxPass.h

@@ -41,6 +41,7 @@ namespace Engine {
             SkyboxPass& operator=(const SkyboxPass&) = delete;
 
             void Register(RG::RenderGraph& rg) override;
+            void OnEnvironmentChanged(const Cubemap* env) override;
 
         private:
             Device& m_Device;

Engine/include/Engine/Renderer/Passes/TAAPass.h

@@ -12,6 +12,7 @@
 
 #include <Engine/Renderer/Passes/RenderPass.h>
 #include <Engine/Renderer/Passes/FullscreenGpu.h>
+#include <Engine/Renderer/SceneTypes.h>
 #include <vulkan/vulkan.hpp>
 #include <vector>
 
@@ -19,7 +20,7 @@ namespace Engine::Passes {
 
     class TAAPass final : public RenderPass {
     public:
-        TAAPass(FullscreenGpu& gpu, f32 historyAlpha,
+        TAAPass(FullscreenGpu& gpu, FrameSceneData& frame, f32 historyAlpha,
                 RG::TextureHandle hdrColor, RG::TextureHandle velocity,
                 RG::TextureHandle taaColor);
         ~TAAPass() override = default;
@@ -29,7 +30,8 @@ namespace Engine::Passes {
 
     private:
         FullscreenGpu& m_Gpu;
-        f32 m_HistoryAlpha;
+        FrameSceneData& m_Frame;   // taaEnabled 开关（帧契约）
+        f32 m_HistoryAlpha;        // 开启时的历史权重
         RG::TextureHandle m_HDR;
         RG::TextureHandle m_Velocity;
         RG::TextureHandle m_TAAColor;   // 输出 = 下一帧的历史

Engine/include/Engine/Renderer/Passes/TonemapPass.h

@@ -11,32 +11,54 @@
 
 #include <Engine/Renderer/Passes/RenderPass.h>
 #include <Engine/Renderer/Passes/FullscreenGpu.h>
+#include <Engine/Renderer/SceneTypes.h>
 #include <vulkan/vulkan.hpp>
 #include <vector>
 
-namespace Engine::Passes {
+namespace Engine {
+
+    class Cubemap;
+    class DescriptorManager;
+
+    namespace Passes {
 
         class TonemapPass final : public RenderPass {
         public:
-        TonemapPass(FullscreenGpu& gpu, PostProcessParams& params,
+            TonemapPass(FullscreenGpu& gpu, DescriptorManager& descriptors,
+                        PostProcessParams& params,
+                        FrameSceneData& frame,
                         RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
-                    RG::TextureHandle temporalDiff, RG::TextureHandle backbuffer);
+                        RG::TextureHandle temporalDiff, RG::TextureHandle backbuffer,
+                        RG::TextureHandle rawHdr, RG::TextureHandle velocity,
+                        RG::TextureHandle shadowMap0, RG::TextureHandle shadowMap1,
+                        RG::TextureHandle shadowMap2);
             ~TonemapPass() override = default;
 
             void Register(RG::RenderGraph& rg) override;
             void OnTargetsRebuilt(RG::RenderGraph& rg) override;
+            void OnEnvironmentChanged(const Cubemap* env) override;
 
         private:
             FullscreenGpu& m_Gpu;
+            DescriptorManager& m_Descriptors;
             PostProcessParams& m_Params;
+            FrameSceneData& m_Frame;   // debugView / bloomEnabled（帧契约）
             RG::TextureHandle m_HDR;
             RG::TextureHandle m_A;
-        RG::TextureHandle m_TemporalDiff;   // 调试视图输入（当前帧槽位）
+            RG::TextureHandle m_TemporalDiff;
             RG::TextureHandle m_Backbuffer;
+            // ---- 调试视图源（binding 3-7，Register 时声明读防止剔除）----
+            // 注：TAAColor 视图直接用 binding0（主输入即 TAA 输出），无需重复绑
+            RG::TextureHandle m_RawHDR;   // Forward 直出（TAA 前）
+            RG::TextureHandle m_Velocity;
+            RG::TextureHandle m_ShadowMaps[3];
+            // ---- 环境调试视图源（binding 8-11；随环境切换重绑）----
+            const Cubemap* m_Env = nullptr;
 
             std::vector<vk::DescriptorSet> m_Sets;   // per-slot
         };
 
-} // namespace Engine::Passes
+    } // namespace Passes
+} // namespace Engine
 
 #endif //HELLOVULKAN_TONEMAPPASS_H

Engine/include/Engine/Renderer/Renderer.h

@@ -42,6 +42,7 @@ namespace Engine {
     class SceneRenderer;
     class VulkanAssetUploader;
     class Cubemap;
+    class IblGenerator;
     class ImGuiLayer;
 
     namespace ECS { class Scene; }
@@ -81,9 +82,25 @@ namespace Engine {
         void SetBloomThreshold(f32 v);
         f32 GetBloomStrength() const;
         void SetBloomStrength(f32 v);
-        // 调试视图：显示帧历史差异（TemporalDiff pass 产出）
-        bool GetShowTemporalDiff() const;
-        void SetShowTemporalDiff(bool b);
+
+        // ---- 特性开关 / 调试视图（转发到 SceneRenderer，每帧落
+        //      FrameSceneData 契约，pass 只读；ImGui 直接对接）----
+        bool GetShadowsEnabled() const;
+        void SetShadowsEnabled(bool b);
+        bool GetTAAEnabled() const;
+        void SetTAAEnabled(bool b);
+        bool GetBloomEnabled() const;
+        void SetBloomEnabled(bool b);
+        int GetDebugView() const;
+        void SetDebugView(int v);   // Passes::DebugView 枚举值
+        // 调试视图名表（下标 = DebugView 枚举值，ImGui combo 直接枚举）
+        const std::vector<const char*>& GetDebugViewNames() const;
+
+        // ---- 动态环境切换（资产归 Assets 层，Renderer 只消费）----
+        // env 传 AssetManager::GetEnvironment(ref) 的指针；nullptr 回中性兜底。
+        // 内部 waitIdle 后统一换绑（引用环境描述符的 pass 全部同步）；
+        // compute 预积分已把生成成本压到毫秒级，waitIdle 尾巴 ≈ 单帧
+        void SetEnvironment(const Cubemap* env);
 
         // 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; }
@@ -92,6 +109,7 @@ namespace Engine {
         // 双视图剔除后的实例数（遥测）
         u32 GetMainInstanceCount() const;
         u32 GetShadowInstanceCount() const;
+        bool GetInstancesTruncated() const;   // 实例预算溢出（遥测红字）
 
         // 垂直同步开关（present mode 切换会重建 swapchain），
         // 只应在帧 UI 构建段调用（Render 之前）
@@ -101,7 +119,6 @@ namespace Engine {
         void BuildGraph();          // 声明 RG 资源 + 注册 pass 对象
         void HandleResize();
         void ReadGpuTimes(u32 frameSlot);
-        void CaptureFrame();        // 调试截图（一次性，写 capture.bmp）
 
         Window& m_Window;
 
@@ -115,7 +132,9 @@ namespace Engine {
         Scope<UniformBuffer> m_PrevObjectBuffer; // 上一帧实例矩阵 ring（TAA velocity）
         Scope<Passes::SceneGpu> m_SceneGpu;     // 场景类 pass 共享装配
         Scope<Passes::FullscreenGpu> m_FullscreenGpu;  // 全屏 pass 共享装配
-        // 环境贴图（IBL 阶段①）：HDR → cube，被 SkyboxPass 引用；
+        // 环境预积分生成端（compute 链），先于 m_Environment/m_Passes 构造
+        Scope<IblGenerator> m_Ibl;
+        // 中性兜底环境（IBL 四件套）：Application 未加载/加载失败时保底；
         // 声明在 m_Passes 之前 → 析构晚于 pass（引用安全）
         Scope<Cubemap> m_Environment;
         Scope<Swapchain> m_Swapchain;
@@ -155,11 +174,6 @@ namespace Engine {
 
         vk::Extent2D m_Extent{};
         u32 m_CurrentFrame = 0;
-
-        // 调试截图（一次性）：第 60 帧把 TAAColor（HDR）落成 capture.bmp
-        u32 m_FrameCounter = 0;
-        bool m_Captured = false;
-        MemoryAllocator::Allocation m_CaptureAlloc{};
     };
 
 } // namespace Engine

Engine/include/Engine/Renderer/SceneRenderer.h

@@ -49,6 +49,14 @@ namespace Engine {
         }
         f32 GetCollectMs() const { return m_Frame.collectMs; }
 
+        // ---- 运行时开关/调试视图（ImGui 经 Renderer 门面调用）。
+        //      状态唯一持有者在此，BuildFrame 每帧落入 FrameSceneData；
+        //      TAA 开关同时控制投影抖动，同一持有者天然同步 ----
+        void SetShadowsEnabled(bool b) { m_ShadowsEnabled = b; }
+        void SetTAAEnabled(bool b) { m_TAAEnabled = b; }
+        void SetBloomEnabled(bool b) { m_BloomEnabled = b; }
+        void SetDebugView(int v) { m_DebugView = v; }
+
     private:
         // CSM：把相机视锥按实用分割切 N 段，每段独立拟合光源正交投影
         //（含纹素对齐防阴影边缘闪烁）。写入（Shadow pass push）与采样
@@ -60,16 +68,19 @@ namespace Engine {
         // ---- 剔除中间量（每帧 clear 复用，拼接进 m_Frame 后保留容量）----
         struct ViewScratch {
             CullResult result;
-            std::vector<glm::mat4> worlds;
-            std::vector<glm::mat4> prevWorlds;
+            // 每 batch 自己的实例列表（按 mesh 分组）。GPU 端 batch 语义 =
+            // firstInstance + 连续区间，实体序收集不满足（多实体共享 mesh
+            // 时实例分散各处），扁平化成连续数组推迟到 appendView
+            std::vector<std::vector<glm::mat4>> batchWorlds;
+            std::vector<std::vector<glm::mat4>> batchPrevWorlds;
             std::unordered_map<Mesh*, u32> lookup;
 
             void Clear()
             {
                 result.batches.clear();
                 result.instanceCount = 0;
-                worlds.clear();
-                prevWorlds.clear();
+                batchWorlds.clear();
+                batchPrevWorlds.clear();
                 lookup.clear();
             }
         };
@@ -86,6 +97,12 @@ namespace Engine {
         std::unordered_map<u32, glm::mat4> m_CurWorlds;
 
         bool m_TruncationWarned = false;   // 实例缓冲打满只告警一次
+
+        // ---- 运行时开关/调试视图状态（本类是唯一写入者）----
+        bool m_ShadowsEnabled = true;
+        bool m_TAAEnabled = true;
+        bool m_BloomEnabled = true;
+        int m_DebugView = 0;
     };
 
 } // namespace Engine

Engine/include/Engine/Renderer/SceneTypes.h

@@ -33,7 +33,11 @@ namespace Engine {
     };
 
     constexpr u32 kMaxLights = 4;
-    constexpr u32 kMaxInstances = 16384;          // 实例 SSBO 容量（矩阵数）
+    // 实例 SSBO 容量（矩阵数）。四视图（main + 3 级联）共享同一预算：
+    // 需求 = Σ各视图剔除后实例数，大场景斜光时阴影实例可膨胀到主视图
+    // 的 2-3 倍。65536 = 4MB/槽位，覆盖 ~800 实体 × 79 mesh 全级联；
+    // 再往上就该上 LOD / GPU culling，而不是继续加预算
+    constexpr u32 kMaxInstances = 65536;
     constexpr u32 kMaxCascades = 3;               // CSM 级联数
     constexpr u32 kCascadeRes = 1024;             // 每级联阴影图边长
 
@@ -52,7 +56,8 @@ namespace Engine {
         glm::mat4 view;
         glm::mat4 proj;         // 抖动后的投影（光栅化用）
         glm::mat4 lightSpace[kMaxCascades];  // CSM 各级联光空间 VP（[0,1] 重映射）
-        glm::vec4 cascadeSplits;  // xyz = 级联 0..2 远边界（视图深度，正值），w 未用
+        glm::vec4 cascadeSplits;  // xyz = 级联 0..2 远边界（视图深度，正值）
+                                  // w = shadowsEnabled（1.0/0.0，frag 阴影短路用）
         glm::vec4 ambientColor;
         LightData lights[kMaxLights];
 
@@ -68,6 +73,9 @@ namespace Engine {
     struct FrameSceneData {
         UniformBufferObject ubo{};
 
+        glm::vec4 ClearColor{0.1, 0.1, 0.1, 1.0};
+        bool DrawSkybox = true;
+
         // 天空盒反投影矩阵：inverse(projNoJitter × mat4(mat3(view)))
         //（视图去平移只留旋转 → 反投影结果即世界方向；Skybox pass push）
         glm::mat4 skyInvVP{1.0f};
@@ -82,6 +90,17 @@ namespace Engine {
         // 级联盒沿光方向加了 caster pad，影子不会被剔除裁掉）
         CullResult shadowCascades[kMaxCascades];
 
+        // ---- 运行时开关/调试视图（ImGui 经 Renderer 门面写入，
+        //      SceneRenderer 唯一落帧，各 pass 只读；每帧确定性）----
+        bool shadowsEnabled = true;   // ShadowPass 空转 + frag ShadowFactor 短路
+        bool taaEnabled = true;       // TAAPass 直通(alpha=0) + 停投影抖动
+        bool bloomEnabled = true;     // BloomPass 子 pass 空转 + tonemap 强度归零
+        int  debugView = 0;           // Passes::DebugView 枚举（tonemap 路由上屏）
+
+        // 实例预算溢出（kMaxInstances）：本轮收集有实例被静默丢弃——
+        // 必须遥测可见，"只告警一次"的大场景截断等于没有告警
+        bool instancesTruncated = false;
+
         f32 collectMs = 0.0f;   // 收集+剔除 CPU 耗时（遥测）
     };
 

Engine/include/Engine/Renderer/VulkanAssetUploader.h

@@ -19,23 +19,32 @@ namespace Engine {
     class Device;
     class MemoryAllocator;
     class DescriptorManager;
+    class IblGenerator;
 
     class VulkanAssetUploader final : public Assets::IAssetUploader {
     public:
         VulkanAssetUploader(Device& device, MemoryAllocator& allocator,
                             DescriptorManager& descriptors,
-                            vk::DescriptorSetLayout textureLayout)
+                            vk::DescriptorSetLayout textureLayout,
+                            IblGenerator& ibl)
             : m_Device(device), m_Allocator(allocator),
-              m_Descriptors(descriptors), m_TextureLayout(textureLayout) {}
+              m_Descriptors(descriptors), m_TextureLayout(textureLayout),
+              m_Ibl(ibl) {}
 
         Scope<Model> LoadModel(const std::string& path) override;
+        Scope<Model> LoadModel(const std::string& path,
+                               ParsedModel&& parsed) override;
         Scope<Model> CreatePlane(f32 halfSize, u8 r, u8 g, u8 b) override;
+        Scope<Cubemap> LoadEnvironment(const std::string& path) override;
+        Scope<Cubemap> LoadEnvironment(const std::string& path,
+                                       DecodedEquirect&& decoded) override;
 
     private:
         Device& m_Device;
         MemoryAllocator& m_Allocator;
         DescriptorManager& m_Descriptors;
         vk::DescriptorSetLayout m_TextureLayout;
+        IblGenerator& m_Ibl;   // 环境 IBL 预积分的 compute 生成端
     };
 
 } // namespace Engine

Engine/include/Engine/Vulkan/DescriptorManager.h

@@ -40,17 +40,25 @@ namespace Engine {
         void UpdateTexture(vk::DescriptorSet set, u32 binding,
                            vk::Sampler sampler, vk::ImageView view) const;
 
+        // storage image 写入（compute 输出；layout = eGeneral）
+        void UpdateStorageImage(vk::DescriptorSet set, u32 binding,
+                                vk::ImageView view) const;
+
         // 单个 set layout 的旧接口（保留）
         vk::PipelineLayout CreatePipelineLayout(vk::DescriptorSetLayout layout,
                                                  u32 pushConstantSize) const;
 
         // 多 set layout 的新接口。
-        // push constant range 按程序级 vertex|fragment 声明（与
+        // push constant range 按程序级声明（默认 vertex|fragment，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）
+        // push 调用点必须带同样的 stage（VUID-01796）。
+        // compute 管线传 vk::ShaderStageFlagBits::eCompute
         vk::PipelineLayout CreatePipelineLayout(
             const std::vector<vk::DescriptorSetLayout>& layouts,
-            u32 pushConstantSize) const;
+            u32 pushConstantSize,
+            vk::ShaderStageFlags pushStages =
+                vk::ShaderStageFlagBits::eVertex |
+                vk::ShaderStageFlagBits::eFragment) const;
 
     private:
         Device& m_Device;

Engine/src/Assets/AssetManager.cpp

@@ -20,6 +20,17 @@ namespace Engine::Assets {
             [this](const std::string& p) { return m_Uploader.LoadModel(p); });
     }
 
+    ModelRef AssetManager::LoadModel(const std::string& path,
+                                     ParsedModel&& parsed)
+    {
+        // 预解析数据移交 loader（Load 同步调用 lambda，引用捕获安全）；
+        // 路径已缓存则 loader 不执行，parsed 自然丢弃（去重）
+        return m_Models.Load(path,
+            [this, &parsed](const std::string& p) {
+                return m_Uploader.LoadModel(p, std::move(parsed));
+            });
+    }
+
     ModelRef AssetManager::CreatePlane(f32 halfSize, u8 r, u8 g, u8 b)
     {
         // 程序化资产复用同一缓存：键 = 伪路径 + 参数
@@ -30,4 +41,22 @@ namespace Engine::Assets {
             [&](const std::string&) { return m_Uploader.CreatePlane(halfSize, r, g, b); });
     }
 
+    EnvRef AssetManager::LoadEnvironment(const std::string& path)
+    {
+        // GPU 生成（compute 预积分）经接口发生在渲染层实现里
+        return m_Environments.Load(path,
+            [this](const std::string& p) { return m_Uploader.LoadEnvironment(p); });
+    }
+
+    EnvRef AssetManager::LoadEnvironment(const std::string& path,
+                                         DecodedEquirect&& decoded)
+    {
+        // 预解码数据移交 loader（Load 同步调用 lambda，引用捕获安全）；
+        // 路径已缓存则 loader 不执行，decoded 自然丢弃（去重）
+        return m_Environments.Load(path,
+            [this, &decoded](const std::string& p) {
+                return m_Uploader.LoadEnvironment(p, std::move(decoded));
+            });
+    }
+
 } // namespace Engine::Assets

Engine/src/Core/Application.cpp

@@ -22,6 +22,8 @@ namespace Engine {
 
     Application::Application(const std::string& title, u32 width, u32 height)
     {
+        Logger::Initialize();
+
         m_Window = CreateScope<Window>(title, width, height);
 
         // 事件通道：GLFW 回调 → 本类 DispatchEvent → 派生类钩子
@@ -51,6 +53,10 @@ namespace Engine {
         if (m_Renderer) {
             m_Renderer->GetDevice().GetDevice().waitIdle();
         }
+
+        if (m_ImGuiLayer) m_ImGuiLayer.reset();
+
+        Logger::Shutdown();
     }
 
     void Application::EnableImGui()
@@ -102,16 +108,16 @@ namespace Engine {
             m_Scene->FlushDestroyQueue();
             m_Scene->UpdateTransforms();   // UE transform-update 阶段：渲染前保证 world 新鲜
 
-            const auto uiStart = std::chrono::steady_clock::now();
 
             // ---- UI 构建段：OnRenderOverlay 只"声明"控件，不直接碰 GPU ----
             if (m_ImGuiLayer && m_ImGuiLayer->IsEnabled())
             {
+                const auto uiStart = std::chrono::steady_clock::now();
                 m_ImGuiLayer->NewFrame();
                 OnRenderOverlay();
                 m_ImGuiLayer->BuildDrawData();
-            }
                 const auto uiEnd = std::chrono::steady_clock::now();
+            }
 
             // ---- 主渲染（框架行为）：RG 顺序 Forward → ImGui(loadOp=Load) →
             //      Present，draw data 在 ImGui pass 里才真正提交 GPU ----

Engine/src/Core/Log.cpp

@@ -1,3 +1,51 @@
 //
 // Created by atdunbg on 2026/9/20.
 //
+#include <Engine/Core/Log.h>
+
+#include "spdlog/sinks/basic_file_sink.h"
+#include "spdlog/sinks/stdout_color_sinks.h"
+
+namespace Engine::Logger
+{
+    static bool IsInited = false;
+
+    void Initialize()
+    {
+        if (IsInited)
+        {
+            throw std::runtime_error("spdlog already inited");
+        }
+        IsInited = true;
+
+        // 控制台彩色 sink (多线程安全)
+        auto console = std::make_shared<spdlog::sinks::stderr_color_sink_mt>();
+        console->set_pattern("%^[%T.%e] [%l] [%n] %v%$");
+
+        // 文件
+        auto file = std::make_shared<spdlog::sinks::basic_file_sink_mt>("HelloVulkan.log", true);
+        file->set_pattern("[%Y-%m-%d %T.%e] [%l] [%t] %v");
+
+        // 多sink 组合默认 logger
+
+        // auto logger = std::make_shared<spdlog::logger>("hello", spdlog::sinks_init_list{console, file});
+        auto logger = std::make_shared<spdlog::logger>("APP", spdlog::sinks_init_list{console, file});
+        logger->set_level(spdlog::level::trace);
+        logger->flush_on(spdlog::level::warn); // warn 及以上立即 flush,避免崩溃丢日志
+
+        spdlog::set_default_logger(logger);
+    }
+
+
+    void Shutdown()
+    {
+        if (!IsInited)
+        {
+            throw std::runtime_error("spdlog already shutdown or not init");
+        }
+        IsInited = false;
+
+        spdlog::shutdown();
+    }
+}
+

Engine/src/RenderGraph/RenderGraph.cpp

@@ -694,24 +694,6 @@ namespace Engine::RG {
         return entry.physical[frameSlot]->view;
     }
 
-    bool RenderGraph::GetCaptureTarget(TextureHandle h, u32 frameSlot,
-                                       vk::Image& outImage, vk::Extent2D& outExtent,
-                                       vk::Format& outFormat,
-                                       vk::ImageLayout& outResting) const
-    {
-        const auto& entry = GetEntry(h);
-        if (entry.imported || frameSlot >= entry.physical.size() ||
-            !entry.physical[frameSlot]) {
-            return false;
-        }
-        const auto& phys = *entry.physical[frameSlot];
-        outImage = phys.image;
-        outExtent = phys.extent;
-        outFormat = phys.format;
-        outResting = entry.restingLayout;
-        return true;
-    }
-
     void RenderGraph::SetQueryPool(vk::QueryPool pool, u32 queriesPerFrame)
     {
         m_QueryPool = pool;

Engine/src/Renderer/Cubemap.cpp

@@ -1,753 +1,17 @@
 //
 // Created by atdunbg on 2026/9/23.
 //
-// Cubemap 实现：HDR 等距柱状 → CPU 预积分四件套 → fp16 上传。
-//
-// 一次性加载成本（秒级 CPU），换取运行时零积分开销；硬件 Cube 采样
-//（filtering/边缘接缝由 driver 处理）。split-sum 预积分（LearnOpenGL/
-// UE 拆分方案的离线端）：
-//   irradiance  —— 半球 Lambert 积分，漫反射 ambient
-//   prefilter   —— GGX 重要性采样（Hammersley 序列），mip 分级 roughness
-//   BRDF LUT    —— (NdotV, roughness) → (F 缩放, F 偏置)，纯数学表
+// Cubemap 实现：纯资源持有（构造注入 + 成对析构）。
+// 预积分生成见 IblGenerator（compute 链）；CPU 积分路径已整体移除，
+// 环境切换从秒级阻塞降到毫秒级 GPU 提交。
 //
 
 #include <Engine/Renderer/Cubemap.h>
 #include <Engine/Vulkan/Device.h>
 #include <Engine/Vulkan/MemoryAllocator.h>
-#include <Engine/Core/Log.h>
-
-#include <stb/stb_image.h>
-
-#include <glm/glm.hpp>
-#include <glm/gtc/constants.hpp>
-
-#include <algorithm>
-#include <bit>
-#include <cmath>
-#include <cstring>
-#include <vector>
 
 namespace Engine
 {
-    namespace
-    {
-        constexpr u32 kFaceCount = 6;
-        constexpr u32 kIrradianceFaceSize = 32;   // 漫反射辐照度面边长（低频）
-        constexpr u32 kPhiSteps = 32;             // 方位角采样数（512 采样/像素）
-        constexpr u32 kThetaSteps = 16;           // 极角采样数
-        constexpr u32 kPrefilterFaceSize = 128;   // pre-filter 面边长（mip0）
-        constexpr u32 kPrefilterMips = 5;         // mip 分级 roughness 数
-        constexpr u32 kBrdfLutSize = 128;         // BRDF LUT 边长
-        constexpr u32 kBrdfLutSamples = 128;      // LUT 每像素采样数
-
-        // ---- GL/Vulkan cube 面方向表（与 GL spec 立方体选择表互逆）----
-        // (u,v)∈[0,1]（u 向右 v 向下）→ s,t∈[-1,1] → 单位方向
-        glm::vec3 FaceDir(u32 face, f32 u, f32 v)
-        {
-            const f32 s = u * 2.0f - 1.0f;
-            const f32 t = v * 2.0f - 1.0f;
-            glm::vec3 dir;
-            switch (face) {
-                case 0:  dir = { 1.0f, -t, -s }; break;   // +X（面片顶行 = +Y）
-                case 1:  dir = {-1.0f, -t,  s }; break;   // -X
-                case 2:  dir = { s, 1.0f,  t };  break;   // +Y
-                case 3:  dir = { s,-1.0f, -t };  break;   // -Y
-                case 4:  dir = { s, -t, 1.0f };  break;   // +Z
-                default: dir = {-s, -t,-1.0f };  break;   // -Z
-            }
-            return glm::normalize(dir);   // equirect 查表要求单位方向
-        }
-
-        // 等距柱状双线性采样（x 环绕跨经度接缝，y 钳制到极点）。
-        // 图像行 0 = 天顶（+Y）：v = acos(y)/π，与 stbi 行序一致，无需翻转
-        glm::vec3 SampleEquirect(const f32* img, u32 w, u32 h, const glm::vec3& dir)
-        {
-            const f32 u = std::atan2(dir.z, dir.x) * (0.5f / glm::pi<f32>()) + 0.5f;
-            const f32 v = std::acos(glm::clamp(dir.y, -1.0f, 1.0f)) / glm::pi<f32>();
-
-            const f32 fx = u * static_cast<f32>(w) - 0.5f;
-            const f32 fy = v * static_cast<f32>(h) - 0.5f;
-            const u32 x0 = static_cast<u32>(std::floor(fx));
-            const u32 y0 = static_cast<u32>(std::floor(fy));
-            const f32 tx = fx - std::floor(fx);
-            const f32 ty = fy - std::floor(fy);
-
-            auto fetch = [&](u32 x, u32 y) {
-                const u32 xc = ((x % w) + w) % w;           // 经度环绕
-                const u32 yc = glm::clamp(y, 0u, h - 1u);   // 纬度钳制
-                const size_t off = (static_cast<size_t>(yc) * w + xc) * 4;
-                return glm::vec3(img[off + 0], img[off + 1], img[off + 2]);
-            };
-
-            const glm::vec3 c00 = fetch(x0,     y0);
-            const glm::vec3 c10 = fetch(x0 + 1, y0);
-            const glm::vec3 c01 = fetch(x0,     y0 + 1);
-            const glm::vec3 c11 = fetch(x0 + 1, y0 + 1);
-
-            return glm::mix(glm::mix(c00, c10, tx), glm::mix(c01, c11, tx), ty);
-        }
-
-        // 漫反射辐照度半球积分（Lambert）：E = ∫ L·cosθ·sinθ dθ dφ。
-        // 切线基 up×N 构造；均匀网格 ≈2048 采样/像素，对 32² 低频图足够
-        glm::vec3 ConvolveIrradiance(const f32* img, u32 w, u32 h,
-                                     const glm::vec3& N)
-        {
-            const glm::vec3 up = glm::abs(N.y) > 0.999f
-                ? glm::vec3(0.0f, 0.0f, 1.0f) : glm::vec3(0.0f, 1.0f, 0.0f);
-            const glm::vec3 T = glm::normalize(glm::cross(up, N));
-            const glm::vec3 B = glm::cross(N, T);
-
-            const f32 dPhi = 2.0f * glm::pi<f32>() / static_cast<f32>(kPhiSteps);
-            const f32 dTheta = 0.5f * glm::pi<f32>() / static_cast<f32>(kThetaSteps);
-
-            glm::vec3 irradiance(0.0f);
-            for (u32 p = 0; p < kPhiSteps; ++p) {
-                const f32 phi = (static_cast<f32>(p) + 0.5f) * dPhi;
-                for (u32 t = 0; t < kThetaSteps; ++t) {
-                    const f32 theta = (static_cast<f32>(t) + 0.5f) * dTheta;
-                    const f32 sinT = std::sin(theta);
-                    const f32 cosT = std::cos(theta);
-                    const glm::vec3 sampleDir =
-                        T * (sinT * std::cos(phi)) +
-                        B * (sinT * std::sin(phi)) +
-                        N * cosT;
-                    irradiance += SampleEquirect(img, w, h, sampleDir) *
-                                  (cosT * sinT);
-                }
-            }
-            return irradiance * (dPhi * dTheta);
-        }
-
-        // ---- GGX 重要性采样（split-sum 镜面端）----
-
-        // van der Corput 基底反转（位反转技巧）。
-        // 注意是整数值转浮点（static_cast），不是位型重解读——
-        // bit_cast 会把反转后的整数当 IEEE754 解读，序列完全错乱
-        f32 RadicalInverseVdC(u32 bits)
-        {
-            bits = (bits << 16u) | (bits >> 16u);
-            bits = ((bits & 0x55555555u) << 1u) | ((bits & 0xAAAAAAAAu) >> 1u);
-            bits = ((bits & 0x33333333u) << 2u) | ((bits & 0xCCCCCCCCu) >> 2u);
-            bits = ((bits & 0x0F0F0F0Fu) << 4u) | ((bits & 0xF0F0F0F0u) >> 4u);
-            bits = ((bits & 0x00FF00FFu) << 8u) | ((bits & 0xFF00FF00u) >> 8u);
-            return static_cast<f32>(bits) * 2.3283064365386963e-10f;
-        }
-
-        // n = 当次积分总采样数（xi.x = i/n 必须覆盖 [0,1] 全程）
-        glm::vec2 Hammersley(u32 i, u32 n)
-        {
-            return { static_cast<f32>(i) / static_cast<f32>(n),
-                     RadicalInverseVdC(i) };
-        }
-
-        // GGX NDF 按概率密度取半程向量 H（切线空间 → 世界）
-        glm::vec3 ImportanceSampleGGX(const glm::vec2& xi, const glm::vec3& N,
-                                      f32 roughness)
-        {
-            const f32 a = roughness * roughness;
-            const f32 phi = 2.0f * glm::pi<f32>() * xi.x;
-            const f32 cosTheta =
-                std::sqrt((1.0f - xi.y) / (1.0f + (a * a - 1.0f) * xi.y));
-            const f32 sinTheta = std::sqrt(1.0f - cosTheta * cosTheta);
-
-            const glm::vec3 H(sinTheta * std::cos(phi),
-                              sinTheta * std::sin(phi), cosTheta);
-
-            const glm::vec3 up = glm::abs(N.z) > 0.999f
-                ? glm::vec3(1.0f, 0.0f, 0.0f) : glm::vec3(0.0f, 1.0f, 0.0f);
-            const glm::vec3 T = glm::normalize(glm::cross(up, N));
-            const glm::vec3 B = glm::cross(N, T);
-            return glm::normalize(T * H.x + B * H.y + N * H.z);
-        }
-
-        // pre-filter 卷积：R(N) ≈ Σ L·NdotL / Σ NdotL（GGX 加权镜面均值）
-        glm::vec3 Prefilter(const f32* img, u32 w, u32 h,
-                            const glm::vec3& N, f32 roughness, u32 samples)
-        {
-            const glm::vec3 V = N;   // split-sum 约定：V 沿 N
-            glm::vec3 acc(0.0f);
-            f32 weight = 0.0f;
-            for (u32 i = 0; i < samples; ++i) {
-                const glm::vec3 H = ImportanceSampleGGX(
-                    Hammersley(i, samples), N, roughness);
-                const glm::vec3 L = glm::normalize(2.0f * glm::dot(V, H) * H - V);
-                const f32 nDotL = glm::max(glm::dot(N, L), 0.0f);
-                if (nDotL > 0.0f) {
-                    acc += SampleEquirect(img, w, h, L) * nDotL;
-                    weight += nDotL;
-                }
-            }
-            return weight > 0.0f ? acc / weight : glm::vec3(0.0f);
-        }
-
-        // BRDF LUT 积分：(NdotV, roughness) → (A, B)。
-        // A/B 是 Fresnel 的缩放/偏置：specular = prefiltered·(F·A + B)
-        glm::vec2 IntegrateBrdf(f32 nDotV, f32 roughness)
-        {
-            // V 在以 N 为 +Z 的切线空间里（积分只依赖这两角度）
-            const glm::vec3 V(std::sqrt(1.0f - nDotV * nDotV), 0.0f, nDotV);
-
-            f32 a = 0.0f, b = 0.0f;
-            const f32 k = roughness * roughness * 0.5f;   // IBL 版 Smith k
-
-            for (u32 i = 0; i < kBrdfLutSamples; ++i) {
-                const glm::vec2 xi = Hammersley(i, kBrdfLutSamples);
-                const glm::vec3 H = ImportanceSampleGGX(xi, glm::vec3(0, 0, 1),
-                                                        roughness);
-                const glm::vec3 L = glm::normalize(
-                    2.0f * glm::dot(V, H) * H - V);
-
-                const f32 nDotL = glm::max(L.z, 0.0f);
-                const f32 nDotH = glm::max(H.z, 0.0f);
-                const f32 vDotH = glm::max(glm::dot(V, H), 0.0f);
-                if (nDotL <= 0.0f) continue;
-
-                const f32 g = (nDotV / (nDotV * (1.0f - k) + k)) *
-                              (nDotL / (nDotL * (1.0f - k) + k));
-                // 高 roughness 时部分样本 H.z 恰为 0（g 也为 0），
-                // 0/0 = NaN 会累进 LUT——分母钳 epsilon
-                const f32 gVis = g * vDotH /
-                                 glm::max(nDotH * nDotV, 1e-6f);
-                const f32 fc = 1.0f - vDotH;
-                const f32 f = fc * fc * fc * fc * fc;   // (1-vDotH)⁵，乘法链
-
-                a += (1.0f - f) * gVis;
-                b += f * gVis;
-            }
-            return { a / static_cast<f32>(kBrdfLutSamples),
-                     b / static_cast<f32>(kBrdfLutSamples) };
-        }
-
-        // IEEE754 f32 → fp16（round-to-nearest-even；溢出饱和为 Inf）
-        u16 FloatToHalf(f32 value)
-        {
-            const u32 x = std::bit_cast<u32>(value);
-            const u16 sign = static_cast<u16>((x >> 16) & 0x8000u);
-            const u32 absVal = x & 0x7fffffffu;
-
-            if (absVal > 0x7f800000u) {
-                return static_cast<u16>(sign | 0x7e00u);            // NaN
-            }
-            if (absVal >= 0x47800000u) {
-                return static_cast<u16>(sign | 0x7c00u);            // Inf / 溢出饱和
-            }
-            if (absVal >= 0x38800000u) {
-                const u32 bits = absVal - 0x38000000u;
-                const u32 round = 0x0fffu + ((bits >> 13) & 1u);
-                return static_cast<u16>(sign | ((bits + round) >> 13));
-            }
-            if (absVal < 0x33000000u) {
-                return sign;   // < 2^-25：舍入为 0
-            }
-            const u32 mantissa = (absVal & 0x007fffffu) | 0x00800000u;
-            const u32 shift = 126u - (absVal >> 23u);
-            const u32 m = (mantissa + (1u << (shift - 1))) >> shift;
-            return static_cast<u16>(sign | std::min(m, 0x400u));
-        }
-
-        // ---- 上传 ----
-
-        vk::ImageView CreateCubeView(Device& device, vk::Image image, u32 mips)
-        {
-            return device.GetDevice().createImageView(
-                vk::ImageViewCreateInfo{}
-                    .setImage(image)
-                    .setViewType(vk::ImageViewType::eCube)
-                    .setFormat(vk::Format::eR16G16B16A16Sfloat)
-                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                          0, mips, 0, kFaceCount}));
-        }
-
-        vk::ImageView CreateLutView(Device& device, vk::Image image)
-        {
-            return device.GetDevice().createImageView(
-                vk::ImageViewCreateInfo{}
-                    .setImage(image)
-                    .setViewType(vk::ImageViewType::e2D)
-                    .setFormat(vk::Format::eR16G16Sfloat)
-                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                          0, 1, 0, 1}));
-        }
-
-        // cube mip 链数据：[mip0: 6 面][mip1: 6 面]...，每面 faceSize²×8B
-        struct CubeMipData {
-            const u16* data;   // 6 面连续
-            u32 faceSize;
-        };
-
-        // 上传 cube（含 mip 链）为 eCubeCompatible image（fp16），
-        // 返回完整资源单元（image + view + 显存分配）
-        EnvTexture UploadCube(Device& device, MemoryAllocator& allocator,
-                              const std::vector<CubeMipData>& mips,
-                              u32 viewMipCount)
-        {
-            const u32 mipCount = static_cast<u32>(mips.size());
-
-            // staging：各 mip 紧凑排布，记偏移
-            vk::DeviceSize total = 0;
-            for (const auto& mip : mips) {
-                total += static_cast<vk::DeviceSize>(mip.faceSize) *
-                         mip.faceSize * 8 * kFaceCount;
-            }
-
-            auto stagingAlloc = allocator.CreateBuffer(
-                total,
-                vk::BufferUsageFlagBits::eTransferSrc,
-                vk::MemoryPropertyFlagBits::eHostVisible |
-                vk::MemoryPropertyFlagBits::eHostCoherent);
-            {
-                auto* dst = static_cast<u8*>(stagingAlloc.MappedPtr);
-                vk::DeviceSize off = 0;
-                for (const auto& mip : mips) {
-                    const vk::DeviceSize bytes =
-                        static_cast<vk::DeviceSize>(mip.faceSize) *
-                        mip.faceSize * 8 * kFaceCount;
-                    std::memcpy(dst + off, mip.data,
-                                static_cast<size_t>(bytes));
-                    off += bytes;
-                }
-            }
-
-            const u32 baseSize = mips[0].faceSize;
-            vk::ImageCreateInfo imageInfo;
-            imageInfo.setImageType(vk::ImageType::e2D)
-                     .setFormat(vk::Format::eR16G16B16A16Sfloat)
-                     .setExtent({baseSize, baseSize, 1})
-                     .setMipLevels(mipCount)
-                     .setArrayLayers(kFaceCount)
-                     .setSamples(vk::SampleCountFlagBits::e1)
-                     .setTiling(vk::ImageTiling::eOptimal)
-                     .setFlags(vk::ImageCreateFlagBits::eCubeCompatible)
-                     .setUsage(vk::ImageUsageFlagBits::eTransferDst |
-                               vk::ImageUsageFlagBits::eSampled)
-                     .setSharingMode(vk::SharingMode::eExclusive)
-                     .setInitialLayout(vk::ImageLayout::eUndefined);
-
-            auto imageAlloc = allocator.CreateImage(imageInfo,
-                vk::MemoryPropertyFlagBits::eDeviceLocal);
-            const vk::Image image = imageAlloc.Image;
-
-            vk::CommandPoolCreateInfo poolInfo;
-            poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
-                    .setFlags(vk::CommandPoolCreateFlagBits::eTransient);
-            vk::CommandPool pool = device.GetDevice().createCommandPool(poolInfo);
-
-            vk::CommandBuffer cmd = device.GetDevice().allocateCommandBuffers(
-                vk::CommandBufferAllocateInfo{}
-                    .setCommandPool(pool)
-                    .setLevel(vk::CommandBufferLevel::ePrimary)
-                    .setCommandBufferCount(1))[0];
-
-            cmd.begin(vk::CommandBufferBeginInfo{
-                vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
-
-            // 全 mip × 全层一起转 layout
-            vk::ImageMemoryBarrier2 toTransfer;
-            toTransfer.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
-                      .setSrcAccessMask(vk::AccessFlagBits2::eNone)
-                      .setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                      .setDstAccessMask(vk::AccessFlagBits2::eTransferWrite)
-                      .setOldLayout(vk::ImageLayout::eUndefined)
-                      .setNewLayout(vk::ImageLayout::eTransferDstOptimal)
-                      .setImage(image)
-                      .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                            0, mipCount, 0, kFaceCount});
-            cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
-
-            // 逐 mip 逐面拷贝；buffer 偏移必须按 mips[0..mip) 字节数累积
-            //（mip 面尺寸递减，用当前 dim 乘 mip 序号会错位 → 读到
-            // staging 未初始化数据 → fp16 NaN 污染整屏）
-            vk::DeviceSize offset = 0;
-            for (u32 mip = 0; mip < mipCount; ++mip) {
-                const u32 dim = mips[mip].faceSize;
-                for (u32 f = 0; f < kFaceCount; ++f) {
-                    vk::BufferImageCopy region;
-                    region.setBufferOffset(
-                               offset + static_cast<vk::DeviceSize>(dim) *
-                                        dim * 8 * f)
-                          .setBufferRowLength(0)
-                          .setBufferImageHeight(0)
-                          .setImageSubresource({vk::ImageAspectFlagBits::eColor,
-                                                mip, f, 1})
-                          .setImageOffset({0, 0, 0})
-                          .setImageExtent({dim, dim, 1});
-                    cmd.copyBufferToImage(stagingAlloc.Buffer, image,
-                                          vk::ImageLayout::eTransferDstOptimal,
-                                          region);
-                }
-                offset += static_cast<vk::DeviceSize>(dim) * dim * 8 *
-                          kFaceCount;
-            }
-
-            vk::ImageMemoryBarrier2 toShader;
-            toShader.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                    .setSrcAccessMask(vk::AccessFlagBits2::eTransferWrite)
-                    .setDstStageMask(vk::PipelineStageFlagBits2::eFragmentShader)
-                    .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
-                    .setOldLayout(vk::ImageLayout::eTransferDstOptimal)
-                    .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
-                    .setImage(image)
-                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                          0, mipCount, 0, kFaceCount});
-            cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
-
-            cmd.end();
-
-            vk::Fence fence = device.GetDevice().createFence({});
-            device.GetGraphicsQueue().submit(
-                vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
-            if (device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
-                vk::Result::eSuccess) {
-                ENGINE_ERROR("Cubemap upload: waitForFences failed");
-            }
-
-            device.GetDevice().destroyFence(fence);
-            device.GetDevice().destroyCommandPool(pool);
-            allocator.Destroy(stagingAlloc);
-
-            return {imageAlloc.Image,
-                    CreateCubeView(device, imageAlloc.Image, viewMipCount),
-                    imageAlloc};
-        }
-
-        // 上传 BRDF LUT（2D，R16G16Sfloat，8B/px），返回完整资源单元
-        EnvTexture UploadTexture2D(Device& device,
-                                   MemoryAllocator& allocator,
-                                   const std::vector<u16>& texels,
-                                   u32 size, vk::Format format)
-        {
-            const vk::DeviceSize total =
-                static_cast<vk::DeviceSize>(size) * size * 8;
-
-            auto stagingAlloc = allocator.CreateBuffer(
-                total,
-                vk::BufferUsageFlagBits::eTransferSrc,
-                vk::MemoryPropertyFlagBits::eHostVisible |
-                vk::MemoryPropertyFlagBits::eHostCoherent);
-            std::memcpy(stagingAlloc.MappedPtr, texels.data(),
-                        static_cast<size_t>(total));
-
-            vk::ImageCreateInfo imageInfo;
-            imageInfo.setImageType(vk::ImageType::e2D)
-                     .setFormat(format)
-                     .setExtent({size, size, 1})
-                     .setMipLevels(1)
-                     .setArrayLayers(1)
-                     .setSamples(vk::SampleCountFlagBits::e1)
-                     .setTiling(vk::ImageTiling::eOptimal)
-                     .setUsage(vk::ImageUsageFlagBits::eTransferDst |
-                               vk::ImageUsageFlagBits::eSampled)
-                     .setSharingMode(vk::SharingMode::eExclusive)
-                     .setInitialLayout(vk::ImageLayout::eUndefined);
-
-            auto imageAlloc = allocator.CreateImage(imageInfo,
-                vk::MemoryPropertyFlagBits::eDeviceLocal);
-            const vk::Image image = imageAlloc.Image;
-
-            vk::CommandPoolCreateInfo poolInfo;
-            poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
-                    .setFlags(vk::CommandPoolCreateFlagBits::eTransient);
-            vk::CommandPool pool = device.GetDevice().createCommandPool(poolInfo);
-
-            vk::CommandBuffer cmd = device.GetDevice().allocateCommandBuffers(
-                vk::CommandBufferAllocateInfo{}
-                    .setCommandPool(pool)
-                    .setLevel(vk::CommandBufferLevel::ePrimary)
-                    .setCommandBufferCount(1))[0];
-
-            cmd.begin(vk::CommandBufferBeginInfo{
-                vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
-
-            vk::ImageMemoryBarrier2 toTransfer;
-            toTransfer.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
-                      .setSrcAccessMask(vk::AccessFlagBits2::eNone)
-                      .setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                      .setDstAccessMask(vk::AccessFlagBits2::eTransferWrite)
-                      .setOldLayout(vk::ImageLayout::eUndefined)
-                      .setNewLayout(vk::ImageLayout::eTransferDstOptimal)
-                      .setImage(image)
-                      .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                            0, 1, 0, 1});
-            cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(toTransfer));
-
-            vk::BufferImageCopy region;
-            region.setBufferOffset(0)
-                  .setBufferRowLength(0)
-                  .setBufferImageHeight(0)
-                  .setImageSubresource({vk::ImageAspectFlagBits::eColor,
-                                        0, 0, 1})
-                  .setImageOffset({0, 0, 0})
-                  .setImageExtent({size, size, 1});
-            cmd.copyBufferToImage(stagingAlloc.Buffer, image,
-                                  vk::ImageLayout::eTransferDstOptimal, region);
-
-            vk::ImageMemoryBarrier2 toShader;
-            toShader.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                    .setSrcAccessMask(vk::AccessFlagBits2::eTransferWrite)
-                    .setDstStageMask(vk::PipelineStageFlagBits2::eFragmentShader)
-                    .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
-                    .setOldLayout(vk::ImageLayout::eTransferDstOptimal)
-                    .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
-                    .setImage(image)
-                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                          0, 1, 0, 1});
-            cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(toShader));
-
-            cmd.end();
-
-            vk::Fence fence = device.GetDevice().createFence({});
-            device.GetGraphicsQueue().submit(
-                vk::SubmitInfo{}.setCommandBuffers(cmd), fence);
-            if (device.GetDevice().waitForFences(fence, VK_TRUE, UINT64_MAX) !=
-                vk::Result::eSuccess) {
-                ENGINE_ERROR("BRDF LUT upload: waitForFences failed");
-            }
-
-            device.GetDevice().destroyFence(fence);
-            device.GetDevice().destroyCommandPool(pool);
-            allocator.Destroy(stagingAlloc);
-
-            return {imageAlloc.Image,
-                    CreateLutView(device, imageAlloc.Image), imageAlloc};
-        }
-
-        // fp16 NaN 自检（指数全 1 且尾数非 0）：预积分数据上传前的
-        // 最后一道防线——NaN 进 GPU 会污染整屏输出
-        bool ContainsNaNHalf(const u16* data, size_t count)
-        {
-            for (size_t i = 0; i < count; ++i) {
-                const u32 abs = data[i] & 0x7FFFu;
-                if (abs > 0x7C00u) return true;
-            }
-            return false;
-        }
-
-        vk::Sampler CreateEnvSampler(Device& device, u32 maxLod)
-        {
-            vk::SamplerCreateInfo samplerInfo;
-            samplerInfo.setMagFilter(vk::Filter::eLinear)
-                       .setMinFilter(vk::Filter::eLinear)
-                       .setMipmapMode(vk::SamplerMipmapMode::eNearest)
-                       .setAddressModeU(vk::SamplerAddressMode::eClampToEdge)
-                       .setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
-                       .setAddressModeW(vk::SamplerAddressMode::eClampToEdge)
-                       .setAnisotropyEnable(VK_FALSE)
-                       .setBorderColor(vk::BorderColor::eIntOpaqueBlack)
-                       .setUnnormalizedCoordinates(VK_FALSE)
-                       .setCompareEnable(VK_FALSE)
-                       .setMinLod(0.0f)
-                       .setMaxLod(static_cast<f32>(maxLod));
-            return device.GetDevice().createSampler(samplerInfo);
-        }
-
-        // BRDF LUT 专用：clamp-to-edge（共享 border 白采样器会在 uv=1.0
-        // 处混入白色 border，高 roughness 镜面偏大）
-        vk::Sampler CreateLutSampler(Device& device)
-        {
-            return CreateEnvSampler(device, 0);
-        }
-    }
-
-    Scope<Cubemap> Cubemap::CreateFromHdrEquirect(Device& device,
-                                                  MemoryAllocator& allocator,
-                                                  const std::string& path,
-                                                  u32 faceSize)
-    {
-        // ---- 1. CPU 加载 HDR（float RGBA）----
-        int w = 0, h = 0, channels = 0;
-        f32* pixels = stbi_loadf(path.c_str(), &w, &h, &channels, STBI_rgb_alpha);
-        if (!pixels) {
-            ENGINE_ERROR("Cubemap: failed to load HDR: {}", path);
-            return nullptr;
-        }
-        const u32 srcW = static_cast<u32>(w);
-        const u32 srcH = static_cast<u32>(h);
-
-        // ---- 2a. 环境 cube：重采样 6 面 ----
-        const u32 facePixels = faceSize * faceSize;
-        std::vector<u16> envFaces(kFaceCount * facePixels * 4);
-        for (u32 f = 0; f < kFaceCount; ++f) {
-            u16* dst = envFaces.data() + f * facePixels * 4;
-            for (u32 y = 0; y < faceSize; ++y) {
-                for (u32 x = 0; x < faceSize; ++x) {
-                    const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(faceSize);
-                    const f32 v = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(faceSize);
-                    const glm::vec3 rgb =
-                        SampleEquirect(pixels, srcW, srcH, FaceDir(f, u, v));
-                    u16* px = dst + (static_cast<size_t>(y) * faceSize + x) * 4;
-                    px[0] = FloatToHalf(rgb.r);
-                    px[1] = FloatToHalf(rgb.g);
-                    px[2] = FloatToHalf(rgb.b);
-                    px[3] = 0x3C00;   // 1.0
-                }
-            }
-        }
-
-        // ---- 2b. irradiance cube：半球 Lambert 积分 ----
-        const u32 irrSize = kIrradianceFaceSize;
-        const u32 irrPixels = irrSize * irrSize;
-        std::vector<u16> irrFaces(kFaceCount * irrPixels * 4);
-        for (u32 f = 0; f < kFaceCount; ++f) {
-            u16* dst = irrFaces.data() + f * irrPixels * 4;
-            for (u32 y = 0; y < irrSize; ++y) {
-                for (u32 x = 0; x < irrSize; ++x) {
-                    const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(irrSize);
-                    const f32 v = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(irrSize);
-                    const glm::vec3 rgb = ConvolveIrradiance(
-                        pixels, srcW, srcH, FaceDir(f, u, v));
-                    u16* px = dst + (static_cast<size_t>(y) * irrSize + x) * 4;
-                    px[0] = FloatToHalf(rgb.r);
-                    px[1] = FloatToHalf(rgb.g);
-                    px[2] = FloatToHalf(rgb.b);
-                    px[3] = 0x3C00;
-                }
-            }
-        }
-
-        // ---- 2c. pre-filter cube：GGX 重要性采样 × mip 分级 roughness ----
-        // 采样数随 mip 递增（高 roughness 分布更宽需要更多样本）。
-        // 注意：必须一次性分配总容量——逐 mip resize 会 realloc，使
-        // prefMips 里已记录的 data 指针悬空（上传读到垃圾 = NaN 全黑）
-        u32 prefTotal = 0;
-        for (u32 mip = 0; mip < kPrefilterMips; ++mip) {
-            const u32 dim = kPrefilterFaceSize >> mip;
-            prefTotal += kFaceCount * dim * dim * 4;
-        }
-        std::vector<u16> prefFaces(prefTotal);
-        std::vector<CubeMipData> prefMips;
-        prefMips.reserve(kPrefilterMips);
-        size_t prefOffset = 0;
-        for (u32 mip = 0; mip < kPrefilterMips; ++mip) {
-            const u32 dim = kPrefilterFaceSize >> mip;
-            const f32 roughness =
-                static_cast<f32>(mip) / static_cast<f32>(kPrefilterMips - 1);
-            const u32 samples = std::min(32u << mip, 256u);
-
-            u16* base = prefFaces.data() + prefOffset;
-            for (u32 f = 0; f < kFaceCount; ++f) {
-                u16* dst = base + f * dim * dim * 4;
-                for (u32 y = 0; y < dim; ++y) {
-                    for (u32 x = 0; x < dim; ++x) {
-                        const f32 u = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(dim);
-                        const f32 v = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(dim);
-                        const glm::vec3 rgb = Prefilter(
-                            pixels, srcW, srcH, FaceDir(f, u, v),
-                            roughness, samples);
-                        u16* px = dst + (static_cast<size_t>(y) * dim + x) * 4;
-                        px[0] = FloatToHalf(rgb.r);
-                        px[1] = FloatToHalf(rgb.g);
-                        px[2] = FloatToHalf(rgb.b);
-                        px[3] = 0x3C00;
-                    }
-                }
-            }
-            prefMips.push_back({base, dim});
-            prefOffset += kFaceCount * dim * dim * 4;
-        }
-
-        stbi_image_free(pixels);   // 源数据消费完毕
-
-        // ---- 2d. BRDF LUT：纯数学积分表（与环境内容无关）----
-        std::vector<u16> lutTexels(kBrdfLutSize * kBrdfLutSize * 2);
-        for (u32 y = 0; y < kBrdfLutSize; ++y) {
-            const f32 roughness = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(kBrdfLutSize);
-            for (u32 x = 0; x < kBrdfLutSize; ++x) {
-                const f32 nDotV = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(kBrdfLutSize);
-                const glm::vec2 ab = IntegrateBrdf(nDotV, roughness);
-                u16* px = lutTexels.data() + (static_cast<size_t>(y) * kBrdfLutSize + x) * 2;
-                px[0] = FloatToHalf(ab.x);
-                px[1] = FloatToHalf(ab.y);
-            }
-        }
-
-        // ---- 3. 上传：env / irradiance / prefilter / LUT（NaN 自检）----
-        if (ContainsNaNHalf(envFaces.data(), envFaces.size()) ||
-            ContainsNaNHalf(irrFaces.data(), irrFaces.size()) ||
-            ContainsNaNHalf(prefFaces.data(), prefFaces.size())) {
-            ENGINE_ERROR("Cubemap: NaN in pre-integrated data, "
-                         "skipping IBL upload");
-            return nullptr;
-        }
-        if (ContainsNaNHalf(lutTexels.data(), lutTexels.size())) {
-            ENGINE_ERROR("Cubemap: NaN in BRDF LUT");
-            return nullptr;
-        }
-
-        EnvTexture env = UploadCube(device, allocator,
-            {{envFaces.data(), faceSize}}, 1);
-        EnvTexture irr = UploadCube(device, allocator,
-            {{irrFaces.data(), irrSize}}, 1);
-        EnvTexture pref = UploadCube(device, allocator, prefMips,
-                                     kPrefilterMips);
-        EnvTexture lut = UploadTexture2D(device, allocator, lutTexels,
-                                         kBrdfLutSize,
-                                         vk::Format::eR16G16Sfloat);
-
-        envFaces.clear(); envFaces.shrink_to_fit();
-        irrFaces.clear(); irrFaces.shrink_to_fit();
-        prefFaces.clear(); prefFaces.shrink_to_fit();
-
-        // cube sampler 共享：maxLod 覆盖 pre-filter mip 链（textureLod 显式
-        // 采样会被 clamp 到 maxLod，必须 ≥ mipCount-1）
-        vk::Sampler sampler = CreateEnvSampler(device, kPrefilterMips - 1);
-        vk::Sampler lutSampler = CreateLutSampler(device);
-
-        ENGINE_INFO("Environment IBL: {} ({}x6 env, {}x6 irradiance, "
-                    "{}x6 prefilter {} mips, {}² BRDF LUT)",
-                    path, faceSize, irrSize, kPrefilterFaceSize,
-                    kPrefilterMips, kBrdfLutSize);
-
-        return Scope<Cubemap>(new Cubemap(device, allocator,
-            env, irr, pref, lut, sampler, lutSampler));
-    }
-
-    Scope<Cubemap> Cubemap::CreateNeutral(Device& device,
-                                          MemoryAllocator& allocator,
-                                          const glm::vec3& rgb)
-    {
-        // 1×1×6 常数 cube（prefilter 仅 1 mip——1px 图不能建 mip 链，
-        // 采样越界 lod 会被 view range clamp 到 mip0，效果等同）；
-        // BRDF LUT 仍为真实积分（纯数学，无环境依赖）——描述符恒有效
-        const u16 px[4] = { FloatToHalf(rgb.r), FloatToHalf(rgb.g),
-                            FloatToHalf(rgb.b), 0x3C00 };
-        std::vector<u16> faces(kFaceCount * 4, 0);
-        for (u32 f = 0; f < kFaceCount; ++f) {
-            std::memcpy(faces.data() + f * 4, px, sizeof(px));
-        }
-
-        std::vector<u16> prefFaces(kFaceCount * 4, 0);
-        std::vector<CubeMipData> prefMips = {{prefFaces.data(), 1}};
-
-        std::vector<u16> lutTexels(kBrdfLutSize * kBrdfLutSize * 2);
-        for (u32 y = 0; y < kBrdfLutSize; ++y) {
-            const f32 roughness = (static_cast<f32>(y) + 0.5f) / static_cast<f32>(kBrdfLutSize);
-            for (u32 x = 0; x < kBrdfLutSize; ++x) {
-                const f32 nDotV = (static_cast<f32>(x) + 0.5f) / static_cast<f32>(kBrdfLutSize);
-                const glm::vec2 ab = IntegrateBrdf(nDotV, roughness);
-                u16* t = lutTexels.data() + (static_cast<size_t>(y) * kBrdfLutSize + x) * 2;
-                t[0] = FloatToHalf(ab.x);
-                t[1] = FloatToHalf(ab.y);
-            }
-        }
-
-        EnvTexture env = UploadCube(device, allocator, {{faces.data(), 1}}, 1);
-        EnvTexture irr = UploadCube(device, allocator, {{faces.data(), 1}}, 1);
-        EnvTexture pref = UploadCube(device, allocator, prefMips, 1);
-        EnvTexture lut = UploadTexture2D(device, allocator, lutTexels,
-                                         kBrdfLutSize, vk::Format::eR16G16Sfloat);
-
-        vk::Sampler sampler = CreateEnvSampler(device, kPrefilterMips - 1);
-        vk::Sampler lutSampler = CreateLutSampler(device);
-
-        ENGINE_INFO("Neutral IBL environment (constant fallback, real BRDF LUT)");
-        return Scope<Cubemap>(new Cubemap(device, allocator,
-            env, irr, pref, lut, sampler, lutSampler));
-    }
-
     Cubemap::~Cubemap()
     {
         const auto& vkDevice = m_Device.GetDevice();

Engine/src/Renderer/IblGenerator.cpp

@@ -0,0 +1,549 @@
+//
+// Created by atdunbg on 2026/9/24.
+//
+// IblGenerator 实现：一次性 compute 提交完成全部预积分。
+//
+// 同步与生命周期模型：
+//   - 提交用独立 command buffer + fence，与帧管线（RG 提交）互不干扰；
+//   - fence 等待后所有产物 layout 恒为 eShaderReadOnlyOptimal，后续
+//     帧提交经队列序自动可见（无需额外 barrier）；
+//   - 中间资源（equirect 纹理 / staging / storage view）用完即毁；
+//   - 描述符集不回收（growable pool 吸收；环境加载是低频事件）。
+//
+// 计时打点输出各段耗时，作为 CPU→GPU 迁移的前后对比依据。
+//
+
+#include <Engine/Renderer/IblGenerator.h>
+#include <Engine/Renderer/Cubemap.h>
+#include <Engine/Vulkan/Device.h>
+#include <Engine/Vulkan/DescriptorManager.h>
+#include <Engine/Vulkan/ShaderManager.h>
+#include <Engine/Core/Log.h>
+
+#include <stb/stb_image.h>
+
+#include <chrono>
+#include <cstring>
+#include <filesystem>
+
+namespace Engine
+{
+    namespace
+    {
+        constexpr u32 kFaceCount = 6;
+        constexpr u32 kIrradianceFaceSize = 32;   // 漫反射辐照度面边长（低频）
+        constexpr u32 kPrefilterFaceSize = 128;   // pre-filter 面边长（mip0）
+        constexpr u32 kPrefilterMips = 5;         // mip 分级 roughness 数
+        constexpr u32 kBrdfLutSize = 128;         // BRDF LUT 边长
+        constexpr u32 kGroupSize = 8;             // compute local_size（shader 同步）
+
+        f32 MsSince(const std::chrono::steady_clock::time_point& t0)
+        {
+            return std::chrono::duration<f32, std::milli>(
+                std::chrono::steady_clock::now() - t0).count();
+        }
+    }
+
+    IblGenerator::IblGenerator(Device& device, MemoryAllocator& allocator,
+                               DescriptorManager& descriptors,
+                               ShaderManager& shaders)
+        : m_Device(device), m_Allocator(allocator),
+          m_Descriptors(descriptors), m_Shaders(shaders)
+    {
+        // Set A：{0 采样输入（cube/2D 共用 combined 槽），1 storage 输出}
+        const std::vector<DescriptorBinding> iblBindings = {
+            {0, vk::DescriptorType::eCombinedImageSampler,
+             vk::ShaderStageFlagBits::eCompute},
+            {1, vk::DescriptorType::eStorageImage,
+             vk::ShaderStageFlagBits::eCompute},
+        };
+        m_IblSetLayout = descriptors.CreateLayout(iblBindings);
+
+        // Set B：LUT 只有 storage 输出
+        const std::vector<DescriptorBinding> lutBindings = {
+            {0, vk::DescriptorType::eStorageImage,
+             vk::ShaderStageFlagBits::eCompute},
+        };
+        m_LutSetLayout = descriptors.CreateLayout(lutBindings);
+
+        // 管线布局：push 仅 prefilter 用（compute 单独按 eCompute 声明）
+        m_IblLayout = descriptors.CreatePipelineLayout(
+            {m_IblSetLayout}, 0, vk::ShaderStageFlagBits::eCompute);
+        m_IblLayoutPush = descriptors.CreatePipelineLayout(
+            {m_IblSetLayout}, sizeof(f32), vk::ShaderStageFlagBits::eCompute);
+        m_LutLayout = descriptors.CreatePipelineLayout(
+            {m_LutSetLayout}, 0, vk::ShaderStageFlagBits::eCompute);
+
+        const std::string dir = "shaders/";
+        m_ToCubeShader = shaders.CreateFromFile(
+            dir + "ibl_equirect_to_cube.comp.spv",
+            vk::ShaderStageFlagBits::eCompute,
+            {m_IblSetLayout}).Handle;
+        m_IrrShader = shaders.CreateFromFile(
+            dir + "ibl_irradiance.comp.spv",
+            vk::ShaderStageFlagBits::eCompute,
+            {m_IblSetLayout}).Handle;
+        m_PrefShader = shaders.CreateFromFile(
+            dir + "ibl_prefilter.comp.spv",
+            vk::ShaderStageFlagBits::eCompute,
+            {m_IblSetLayout}, sizeof(f32)).Handle;
+        m_LutShader = shaders.CreateFromFile(
+            dir + "ibl_brdf_lut.comp.spv",
+            vk::ShaderStageFlagBits::eCompute,
+            {m_LutSetLayout}).Handle;
+
+        if (!m_ToCubeShader || !m_IrrShader || !m_PrefShader || !m_LutShader) {
+            ENGINE_FATAL("IblGenerator: failed to create compute shaders");
+        }
+
+        // equirect 中间纹理采样器：硬件双线性，S=repeat 吃掉经度接缝
+        vk::SamplerCreateInfo sInfo;
+        sInfo.setMagFilter(vk::Filter::eLinear)
+             .setMinFilter(vk::Filter::eLinear)
+             .setMipmapMode(vk::SamplerMipmapMode::eNearest)
+             .setAddressModeU(vk::SamplerAddressMode::eRepeat)
+             .setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
+             .setAddressModeW(vk::SamplerAddressMode::eClampToEdge);
+        m_EquirectSampler = device.GetDevice().createSampler(sInfo);
+
+        // 持久命令设施：每次生成 reset 复用（避免反复建池）
+        vk::CommandPoolCreateInfo poolInfo;
+        poolInfo.setQueueFamilyIndex(device.GetGraphicsFamily())
+                .setFlags(vk::CommandPoolCreateFlagBits::eTransient |
+                          vk::CommandPoolCreateFlagBits::eResetCommandBuffer);
+        m_Pool = device.GetDevice().createCommandPool(poolInfo);
+        m_Cmd = device.GetDevice().allocateCommandBuffers(
+            vk::CommandBufferAllocateInfo{}
+                .setCommandPool(m_Pool)
+                .setLevel(vk::CommandBufferLevel::ePrimary)
+                .setCommandBufferCount(1))[0];
+        m_Fence = device.GetDevice().createFence(
+            vk::FenceCreateInfo{vk::FenceCreateFlagBits::eSignaled});
+    }
+
+    IblGenerator::~IblGenerator()
+    {
+        const auto& vkDevice = m_Device.GetDevice();
+
+        for (const vk::ShaderEXT s : {m_ToCubeShader, m_IrrShader,
+                                      m_PrefShader, m_LutShader}) {
+            if (s) vkDevice.destroyShaderEXT(s);
+        }
+        vkDevice.destroySampler(m_EquirectSampler);
+        vkDevice.destroyPipelineLayout(m_IblLayout);
+        vkDevice.destroyPipelineLayout(m_IblLayoutPush);
+        vkDevice.destroyPipelineLayout(m_LutLayout);
+        vkDevice.destroyDescriptorSetLayout(m_IblSetLayout);
+        vkDevice.destroyDescriptorSetLayout(m_LutSetLayout);
+        vkDevice.destroyFence(m_Fence);
+        vkDevice.freeCommandBuffers(m_Pool, m_Cmd);
+        vkDevice.destroyCommandPool(m_Pool);
+    }
+
+    // ---- cube 兼容目标：storage（2DArray per-mip view）+ 采样（eCube view）----
+
+    IblGenerator::CubeTarget IblGenerator::CreateCubeTarget(u32 size, u32 mips) const
+    {
+        // fp16 storage 支持（桌面 GPU 普遍具备，显式校验防退化平台）
+        const vk::Format format = m_Device.FindSupportedFormat(
+            {vk::Format::eR16G16B16A16Sfloat},
+            vk::ImageTiling::eOptimal,
+            vk::FormatFeatureFlagBits::eStorageImage);
+
+        vk::ImageCreateInfo info;
+        info.setImageType(vk::ImageType::e2D)
+             .setFormat(format)
+             .setExtent({size, size, 1})
+             .setMipLevels(mips)
+             .setArrayLayers(kFaceCount)
+             .setSamples(vk::SampleCountFlagBits::e1)
+             .setTiling(vk::ImageTiling::eOptimal)
+             .setFlags(vk::ImageCreateFlagBits::eCubeCompatible)
+             .setUsage(vk::ImageUsageFlagBits::eStorage |
+                       vk::ImageUsageFlagBits::eSampled)
+             .setInitialLayout(vk::ImageLayout::eUndefined);
+
+        CubeTarget t;
+        t.allocation = m_Allocator.CreateImage(info,
+            vk::MemoryPropertyFlagBits::eDeviceLocal);
+        t.image = t.allocation.Image;
+
+        // storage 写：2DArray view，每 mip 一个（storage 必须单 mip 层）
+        for (u32 mip = 0; mip < mips; ++mip) {
+            t.storageViews.push_back(m_Device.GetDevice().createImageView(
+                vk::ImageViewCreateInfo{}
+                    .setImage(t.image)
+                    .setViewType(vk::ImageViewType::e2DArray)
+                    .setFormat(format)
+                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                          mip, 1, 0, kFaceCount})));
+        }
+        // 采样读：eCube view 全 mip 链（消费端契约与原 CPU 路径一致）
+        t.cubeView = m_Device.GetDevice().createImageView(
+            vk::ImageViewCreateInfo{}
+                .setImage(t.image)
+                .setViewType(vk::ImageViewType::eCube)
+                .setFormat(format)
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, mips, 0, kFaceCount}));
+        return t;
+    }
+
+    // 只释放 storage 中转 view（image/cubeView/显存交给 Cubemap 接管）
+    void IblGenerator::ReleaseStorageViews(const CubeTarget& t) const
+    {
+        for (const vk::ImageView v : t.storageViews) {
+            m_Device.GetDevice().destroyImageView(v);
+        }
+    }
+
+    DecodedEquirect IblGenerator::DecodeHdr(const std::string& path)
+    {
+        // stbi_loadf 无共享可变状态（失败原因每次调用独立），可并发调用
+        int w = 0, h = 0, channels = 0;
+        f32* pixels = stbi_loadf(path.c_str(), &w, &h, &channels, STBI_rgb_alpha);
+        if (!pixels) {
+            ENGINE_ERROR("IblGenerator: failed to load HDR: {}", path);
+            return {};
+        }
+
+        DecodedEquirect decoded;
+        decoded.w = static_cast<u32>(w);
+        decoded.h = static_cast<u32>(h);
+        decoded.pixels.assign(pixels,
+                              pixels + static_cast<size_t>(w) * h * 4);
+        stbi_image_free(pixels);
+        return decoded;
+    }
+
+    Scope<Cubemap> IblGenerator::FromDecoded(const DecodedEquirect& decoded,
+                                             u32 faceSize,
+                                             const std::string& debugName)
+    {
+        if (!decoded.Valid()) {
+            ENGINE_ERROR("IblGenerator: invalid decoded equirect ({})",
+                         debugName);
+            return nullptr;
+        }
+        return Generate(decoded.pixels.data(), decoded.w, decoded.h,
+                        faceSize, debugName);
+    }
+
+    Scope<Cubemap> IblGenerator::FromHdrEquirect(const std::string& path,
+                                                 u32 faceSize)
+    {
+        const auto t0 = std::chrono::steady_clock::now();
+        DecodedEquirect decoded = DecodeHdr(path);
+        if (!decoded.Valid()) {
+            return nullptr;
+        }
+        ENGINE_INFO("IblGenerator: HDR decode {:.2f} ms", MsSince(t0));
+        return FromDecoded(decoded, faceSize, path);
+    }
+
+    Scope<Cubemap> IblGenerator::Neutral(const glm::vec3& rgb)
+    {
+        // 1×1 常数 equirect：全链复用（LUT 仍是真实积分，纯数学）
+        const f32 pixel[4] = {rgb.r, rgb.g, rgb.b, 1.0f};
+        return Generate(pixel, 1, 1, 1, "neutral");
+    }
+
+    Scope<Cubemap> IblGenerator::Generate(const f32* pixels, u32 w, u32 h,
+                                          u32 faceSize,
+                                          const std::string& debugName)
+    {
+        const auto t0 = std::chrono::steady_clock::now();
+        const auto& vkDevice = m_Device.GetDevice();
+
+        // ---- 1. equirect 中间纹理（R32G32B32A32，免 CPU fp16 转换）----
+        const vk::DeviceSize eqBytes =
+            static_cast<vk::DeviceSize>(w) * h * 16;
+        auto staging = m_Allocator.CreateBuffer(
+            eqBytes,
+            vk::BufferUsageFlagBits::eTransferSrc,
+            vk::MemoryPropertyFlagBits::eHostVisible |
+            vk::MemoryPropertyFlagBits::eHostCoherent);
+        std::memcpy(staging.MappedPtr, pixels, static_cast<size_t>(eqBytes));
+
+        vk::ImageCreateInfo eqInfo;
+        eqInfo.setImageType(vk::ImageType::e2D)
+              .setFormat(vk::Format::eR32G32B32A32Sfloat)
+              .setExtent({w, h, 1})
+              .setMipLevels(1)
+              .setArrayLayers(1)
+              .setSamples(vk::SampleCountFlagBits::e1)
+              .setTiling(vk::ImageTiling::eOptimal)
+              .setUsage(vk::ImageUsageFlagBits::eTransferDst |
+                        vk::ImageUsageFlagBits::eSampled)
+              .setInitialLayout(vk::ImageLayout::eUndefined);
+        auto eqAlloc = m_Allocator.CreateImage(eqInfo,
+            vk::MemoryPropertyFlagBits::eDeviceLocal);
+        vk::ImageView eqView = vkDevice.createImageView(
+            vk::ImageViewCreateInfo{}
+                .setImage(eqAlloc.Image)
+                .setViewType(vk::ImageViewType::e2D)
+                .setFormat(vk::Format::eR32G32B32A32Sfloat)
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, 1, 0, 1}));
+
+        // ---- 2. 输出资源 + 消费端采样器 ----
+        CubeTarget env = CreateCubeTarget(faceSize, 1);
+        CubeTarget irr = CreateCubeTarget(kIrradianceFaceSize, 1);
+        CubeTarget pref = CreateCubeTarget(kPrefilterFaceSize, kPrefilterMips);
+
+        const vk::Sampler envSampler = CreateEnvSampler(kPrefilterMips - 1);
+        const vk::Sampler lutSampler = CreateLutSampler();
+
+        const vk::Format lutFormat = m_Device.FindSupportedFormat(
+            {vk::Format::eR16G16Sfloat},
+            vk::ImageTiling::eOptimal,
+            vk::FormatFeatureFlagBits::eStorageImage);
+        vk::ImageCreateInfo lutInfo;
+        lutInfo.setImageType(vk::ImageType::e2D)
+               .setFormat(lutFormat)
+               .setExtent({kBrdfLutSize, kBrdfLutSize, 1})
+               .setMipLevels(1)
+               .setArrayLayers(1)
+               .setSamples(vk::SampleCountFlagBits::e1)
+               .setTiling(vk::ImageTiling::eOptimal)
+               .setUsage(vk::ImageUsageFlagBits::eStorage |
+                         vk::ImageUsageFlagBits::eSampled)
+               .setInitialLayout(vk::ImageLayout::eUndefined);
+        auto lutAlloc = m_Allocator.CreateImage(lutInfo,
+            vk::MemoryPropertyFlagBits::eDeviceLocal);
+        vk::ImageView lutView = vkDevice.createImageView(
+            vk::ImageViewCreateInfo{}
+                .setImage(lutAlloc.Image)
+                .setViewType(vk::ImageViewType::e2D)
+                .setFormat(lutFormat)
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, 1, 0, 1}));
+
+        const auto resourceMs = MsSince(t0);
+
+        // ---- 3. 描述符（低频事件，growable pool 吸收，不回收）----
+        vk::DescriptorSet setToCube =
+            m_Descriptors.AllocateDescriptorSet(m_IblSetLayout);
+        vk::DescriptorSet setIrr =
+            m_Descriptors.AllocateDescriptorSet(m_IblSetLayout);
+        vk::DescriptorSet setLut =
+            m_Descriptors.AllocateDescriptorSet(m_LutSetLayout);
+        std::array<vk::DescriptorSet, kPrefilterMips> setPref{};
+        for (u32 m = 0; m < kPrefilterMips; ++m) {
+            setPref[m] = m_Descriptors.AllocateDescriptorSet(m_IblSetLayout);
+        }
+
+        m_Descriptors.UpdateTexture(setToCube, 0, m_EquirectSampler, eqView);
+        m_Descriptors.UpdateStorageImage(setToCube, 1, env.storageViews[0]);
+        // cube 采样必须用带 mip 的 env 采样器（textureLod 被 clamp 到 maxLod）
+        m_Descriptors.UpdateTexture(setIrr, 0, envSampler, env.cubeView);
+        m_Descriptors.UpdateStorageImage(setIrr, 1, irr.storageViews[0]);
+        for (u32 m = 0; m < kPrefilterMips; ++m) {
+            m_Descriptors.UpdateTexture(setPref[m], 0, envSampler, env.cubeView);
+            m_Descriptors.UpdateStorageImage(setPref[m], 1, pref.storageViews[m]);
+        }
+        m_Descriptors.UpdateStorageImage(setLut, 0, lutView);
+
+        // ---- 4. 录制 + 一次性提交 ----
+        vkDevice.resetFences(m_Fence);
+        vkDevice.resetCommandPool(m_Pool);
+        m_Cmd.begin(vk::CommandBufferBeginInfo{
+            vk::CommandBufferUsageFlagBits::eOneTimeSubmit});
+
+        // 4a. equirect staging → image → shader 读 layout
+        vk::ImageMemoryBarrier2 eqToTransfer;
+        eqToTransfer.setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
+                    .setSrcAccessMask(vk::AccessFlagBits2::eNone)
+                    .setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
+                    .setDstAccessMask(vk::AccessFlagBits2::eTransferWrite)
+                    .setOldLayout(vk::ImageLayout::eUndefined)
+                    .setNewLayout(vk::ImageLayout::eTransferDstOptimal)
+                    .setImage(eqAlloc.Image)
+                    .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                          0, 1, 0, 1});
+        m_Cmd.pipelineBarrier2(
+            vk::DependencyInfo{}.setImageMemoryBarriers(eqToTransfer));
+
+        vk::BufferImageCopy eqCopy;
+        eqCopy.setBufferOffset(0)
+              .setBufferRowLength(0)
+              .setBufferImageHeight(0)
+              .setImageSubresource({vk::ImageAspectFlagBits::eColor, 0, 0, 1})
+              .setImageOffset({0, 0, 0})
+              .setImageExtent({w, h, 1});
+        m_Cmd.copyBufferToImage(staging.Buffer, eqAlloc.Image,
+                                vk::ImageLayout::eTransferDstOptimal, eqCopy);
+
+        vk::ImageMemoryBarrier2 eqToShader;
+        eqToShader.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
+                  .setSrcAccessMask(vk::AccessFlagBits2::eTransferWrite)
+                  .setDstStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                  .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
+                  .setOldLayout(vk::ImageLayout::eTransferDstOptimal)
+                  .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
+                  .setImage(eqAlloc.Image)
+                  .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                        0, 1, 0, 1});
+        m_Cmd.pipelineBarrier2(
+            vk::DependencyInfo{}.setImageMemoryBarriers(eqToShader));
+
+        // 4b. 输出资源 Undefined → General（compute storage 写的要求）
+        std::array<vk::ImageMemoryBarrier2, 4> toGeneral{};
+        const vk::Image images[4] = {env.image, irr.image, pref.image,
+                                     lutAlloc.Image};
+        const u32 mipCounts[4] = {1, 1, kPrefilterMips, 1};
+        for (u32 i = 0; i < 4; ++i) {
+            toGeneral[i]
+                .setSrcStageMask(vk::PipelineStageFlagBits2::eNone)
+                .setSrcAccessMask(vk::AccessFlagBits2::eNone)
+                .setDstStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                .setDstAccessMask(vk::AccessFlagBits2::eShaderWrite)
+                .setOldLayout(vk::ImageLayout::eUndefined)
+                .setNewLayout(vk::ImageLayout::eGeneral)
+                .setImage(images[i])
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, mipCounts[i], 0, kFaceCount});
+        }
+        toGeneral[3].setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                          0, 1, 0, 1});   // LUT 是 2D
+        m_Cmd.pipelineBarrier2(
+            vk::DependencyInfo{}.setImageMemoryBarriers(toGeneral));
+
+        // 4c. ① equirect → env cube
+        std::array<vk::ShaderStageFlagBits, 1> computeStage = {
+            vk::ShaderStageFlagBits::eCompute};
+        std::array<vk::ShaderEXT, 1> toCubeShaders = {m_ToCubeShader};
+        m_Cmd.bindShadersEXT(computeStage, toCubeShaders);
+        m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                 m_IblLayout, 0, setToCube, {});
+        m_Cmd.dispatch((faceSize + kGroupSize - 1) / kGroupSize,
+                       (faceSize + kGroupSize - 1) / kGroupSize, kFaceCount);
+
+        // env 转采样读（②③的输入）
+        vk::ImageMemoryBarrier2 envToRead;
+        envToRead.setSrcStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                 .setSrcAccessMask(vk::AccessFlagBits2::eShaderWrite)
+                 .setDstStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                 .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
+                 .setOldLayout(vk::ImageLayout::eGeneral)
+                 .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
+                 .setImage(env.image)
+                 .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                       0, 1, 0, kFaceCount});
+        m_Cmd.pipelineBarrier2(
+            vk::DependencyInfo{}.setImageMemoryBarriers(envToRead));
+
+        // 4d. ② irradiance（env → irr）
+        std::array<vk::ShaderEXT, 1> irrShaders = {m_IrrShader};
+        m_Cmd.bindShadersEXT(computeStage, irrShaders);
+        m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                 m_IblLayout, 0, setIrr, {});
+        m_Cmd.dispatch((kIrradianceFaceSize + kGroupSize - 1) / kGroupSize,
+                       (kIrradianceFaceSize + kGroupSize - 1) / kGroupSize,
+                       kFaceCount);
+
+        // 4e. ③ prefilter（env → pref，每 mip 一次 dispatch + push roughness）
+        std::array<vk::ShaderEXT, 1> prefShaders = {m_PrefShader};
+        m_Cmd.bindShadersEXT(computeStage, prefShaders);
+        for (u32 m = 0; m < kPrefilterMips; ++m) {
+            const u32 dim = kPrefilterFaceSize >> m;
+            const f32 roughness =
+                static_cast<f32>(m) / static_cast<f32>(kPrefilterMips - 1);
+            m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                     m_IblLayoutPush, 0, setPref[m], {});
+            m_Cmd.pushConstants<f32>(m_IblLayoutPush,
+                                     vk::ShaderStageFlagBits::eCompute,
+                                     0, roughness);
+            m_Cmd.dispatch((dim + kGroupSize - 1) / kGroupSize,
+                           (dim + kGroupSize - 1) / kGroupSize, kFaceCount);
+        }
+
+        // 4f. ④ BRDF LUT（独立纯数学）
+        std::array<vk::ShaderEXT, 1> lutShaders = {m_LutShader};
+        m_Cmd.bindShadersEXT(computeStage, lutShaders);
+        m_Cmd.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                 m_LutLayout, 0, setLut, {});
+        m_Cmd.dispatch((kBrdfLutSize + kGroupSize - 1) / kGroupSize,
+                       (kBrdfLutSize + kGroupSize - 1) / kGroupSize, 1);
+
+        // 4g. irr/pref/lut 收尾：General → ShaderReadOnly（消费端采样 layout；
+        //     fence 等待后对后续队列提交自动可见）
+        std::array<vk::ImageMemoryBarrier2, 3> toRead{};
+        const vk::Image readImages[3] = {irr.image, pref.image,
+                                         lutAlloc.Image};
+        const u32 readMips[3] = {1, kPrefilterMips, 1};
+        const u32 readLayers[3] = {kFaceCount, kFaceCount, 1};   // LUT 是 2D
+        for (u32 i = 0; i < 3; ++i) {
+            toRead[i]
+                .setSrcStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                .setSrcAccessMask(vk::AccessFlagBits2::eShaderWrite)
+                .setDstStageMask(vk::PipelineStageFlagBits2::eComputeShader)
+                .setDstAccessMask(vk::AccessFlagBits2::eShaderRead)
+                .setOldLayout(vk::ImageLayout::eGeneral)
+                .setNewLayout(vk::ImageLayout::eShaderReadOnlyOptimal)
+                .setImage(readImages[i])
+                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
+                                      0, readMips[i], 0, readLayers[i]});
+        }
+        m_Cmd.pipelineBarrier2(
+            vk::DependencyInfo{}.setImageMemoryBarriers(toRead));
+
+        m_Cmd.end();
+
+        m_Device.GetGraphicsQueue().submit(
+            vk::SubmitInfo{}.setCommandBuffers(m_Cmd), m_Fence);
+        if (vkDevice.waitForFences(m_Fence, VK_TRUE, UINT64_MAX) !=
+            vk::Result::eSuccess) {
+            ENGINE_ERROR("IblGenerator: waitForFences failed ({})", debugName);
+        }
+
+        // ---- 5. 中间资源即毁；image/cubeView/显存交棒 Cubemap ----
+        vkDevice.destroyImageView(eqView);
+        m_Allocator.Destroy(eqAlloc);
+        m_Allocator.Destroy(staging);
+        ReleaseStorageViews(env);
+        ReleaseStorageViews(irr);
+        ReleaseStorageViews(pref);
+
+        const f32 computeMs = MsSince(t0) - resourceMs;
+        ENGINE_INFO("IblGenerator[{}]: {}x6 env, {}x6 irradiance, "
+                    "{}x6 prefilter {} mips, {}² BRDF LUT — "
+                    "resources {:.2f} ms, compute+wait {:.2f} ms",
+                    debugName, faceSize, kIrradianceFaceSize,
+                    kPrefilterFaceSize, kPrefilterMips, kBrdfLutSize,
+                    resourceMs, computeMs);
+
+        return Scope<Cubemap>(new Cubemap(
+            m_Device, m_Allocator,
+            EnvTexture{env.image, env.cubeView, env.allocation},
+            EnvTexture{irr.image, irr.cubeView, irr.allocation},
+            EnvTexture{pref.image, pref.cubeView, pref.allocation},
+            EnvTexture{lutAlloc.Image, lutView, lutAlloc},
+            envSampler, lutSampler));
+    }
+
+    vk::Sampler IblGenerator::CreateEnvSampler(u32 maxLod) const
+    {
+        vk::SamplerCreateInfo sInfo;
+        sInfo.setMagFilter(vk::Filter::eLinear)
+             .setMinFilter(vk::Filter::eLinear)
+             .setMipmapMode(vk::SamplerMipmapMode::eNearest)
+             .setAddressModeU(vk::SamplerAddressMode::eClampToEdge)
+             .setAddressModeV(vk::SamplerAddressMode::eClampToEdge)
+             .setAddressModeW(vk::SamplerAddressMode::eClampToEdge)
+             .setAnisotropyEnable(VK_FALSE)
+             .setBorderColor(vk::BorderColor::eIntOpaqueBlack)
+             .setUnnormalizedCoordinates(VK_FALSE)
+             .setCompareEnable(VK_FALSE)
+             .setMinLod(0.0f)
+             .setMaxLod(static_cast<f32>(maxLod));
+        return m_Device.GetDevice().createSampler(sInfo);
+    }
+
+    vk::Sampler IblGenerator::CreateLutSampler() const
+    {
+        // clamp-to-edge：共享 border 白采样器会在 uv=1.0 处混入白色 border
+        return CreateEnvSampler(0);
+    }
+
+} // namespace Engine

Engine/src/Renderer/Model.cpp

@@ -23,14 +23,22 @@ Model::Model(Device& device, MemoryAllocator& allocator,
              DescriptorManager& descriptorManager,
              vk::DescriptorSetLayout textureLayout,
              const std::string& path)
+    : Model(device, allocator, descriptorManager, textureLayout)
+{
+    // 同步便捷路径 = 两段式串行（等价异步收割后的主线程消费）
+    LoadModel(path);
+}
+
+// 空模型骨架：只建默认纹理；mesh/材质由 FromParsed 填充
+Model::Model(Device& device, MemoryAllocator& allocator,
+             DescriptorManager& descriptorManager,
+             vk::DescriptorSetLayout textureLayout)
     : m_Device(device)
     , m_Allocator(allocator)
     , m_DescriptorManager(descriptorManager)
     , m_TextureLayout(textureLayout)
 {
     InitDefaultTextures();
-
-    LoadModel(path);
 }
 
 Model::~Model() = default;
@@ -92,9 +100,10 @@ Scope<Model> Model::CreatePlane(Device& device, MemoryAllocator& allocator,
                                   halfSize, std::array<u8, 3>{r, g, b}));
 }
 
-void Model::LoadModel(const std::string& path)
+ParsedModel Model::ParseFromDisk(const std::string& path)
 {
-    m_ModelPath = path;
+    ParsedModel out;
+    out.path = path;
 
     // UV 翻转按格式区分：
     //   OBJ/FBX   —— v 原点在底部，需翻到左上惯例（配合"v=0=图像顶"采样）
@@ -106,36 +115,84 @@ void Model::LoadModel(const std::string& path)
                        [](unsigned char c) { return std::tolower(c); });
         return s;
     }();
-    m_IsGltf = lower.size() >= 5 &&
+    out.isGltf = lower.size() >= 5 &&
         (lower.rfind(".gltf") == lower.size() - 5 ||
          lower.rfind(".glb") == lower.size() - 4);
 
     Assimp::Importer importer;
     const aiScene* scene = importer.ReadFile(path,
         aiProcess_Triangulate |
-        (m_IsGltf ? 0 : aiProcess_FlipUVs) |
+        (out.isGltf ? 0 : aiProcess_FlipUVs) |
         aiProcess_GenSmoothNormals |
         aiProcess_CalcTangentSpace |   // 法线贴图 TBN
         aiProcess_JoinIdenticalVertices);
 
     if (!scene || (scene->mFlags & AI_SCENE_FLAGS_INCOMPLETE) || !scene->mRootNode) {
         ENGINE_ERROR("Assimp error: {}", importer.GetErrorString());
-        return;
+        return out;   // meshes 空 = 无效
     }
 
-    ProcessNode(scene->mRootNode, scene);
+    const std::filesystem::path modelDir =
+        std::filesystem::path(path).parent_path();
+    ProcessNode(scene->mRootNode, scene, out.meshes, modelDir, out.isGltf);
+    ENGINE_INFO("Model parsed: {} ({} meshes)", path, out.meshes.size());
+    return out;
+}
+
+Scope<Model> Model::FromParsed(Device& device, MemoryAllocator& allocator,
+                               DescriptorManager& descriptorManager,
+                               vk::DescriptorSetLayout textureLayout,
+                               ParsedModel&& parsed)
+{
+    if (!parsed.Valid()) {
+        ENGINE_ERROR("Model::FromParsed: invalid parsed model ({})", parsed.path);
+        return nullptr;
+    }
+
+    auto model = Scope<Model>(new Model(device, allocator,
+                                        descriptorManager, textureLayout));
+    model->m_ModelPath = parsed.path;
+    model->m_IsGltf = parsed.isGltf;
+
+    // 材质缓存去重 + 纹理上传 + mesh 缓冲（全部 GPU，主线程）
+    for (auto& pm : parsed.meshes) {
+        Material* mat = model->GetOrCreateMaterial(pm.texPaths.data(), pm.params);
+        model->m_Meshes.push_back(CreateScope<Mesh>(
+            device, allocator, std::move(pm.vertices), std::move(pm.indices), mat));
+    }
+    ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
+                parsed.path, model->m_Meshes.size(), model->m_TextureCache.size());
+    return model;
+}
+
+void Model::LoadModel(const std::string& path)
+{
+    // 同步路径：解析 + GPU 构建串行（与 FromParsed 相同的消费逻辑）
+    ParsedModel parsed = ParseFromDisk(path);
+    if (!parsed.Valid()) {
+        return;
+    }
+    m_ModelPath = parsed.path;
+    m_IsGltf = parsed.isGltf;
+    for (auto& pm : parsed.meshes) {
+        Material* mat = GetOrCreateMaterial(pm.texPaths.data(), pm.params);
+        m_Meshes.push_back(CreateScope<Mesh>(
+            m_Device, m_Allocator, std::move(pm.vertices), std::move(pm.indices), mat));
+    }
     ENGINE_INFO("Model loaded: {} ({} meshes, {} textures)",
                 path, m_Meshes.size(), m_TextureCache.size());
 }
 
-void Model::ProcessNode(aiNode* node, const aiScene* scene)
+void Model::ProcessNode(const aiNode* node, const aiScene* scene,
+                        std::vector<ParsedMesh>& out,
+                        const std::filesystem::path& modelDir, bool isGltf)
 {
     for (u32 i = 0; i < node->mNumMeshes; ++i) {
         aiMesh* mesh = scene->mMeshes[node->mMeshes[i]];
-        m_Meshes.push_back(ProcessMesh(mesh, scene));
+        out.push_back(ProcessMesh(mesh, scene, modelDir, isGltf));
     }
     for (u32 i = 0; i < node->mNumChildren; ++i) {
-        ProcessNode(node->mChildren[i], scene);
+        ProcessNode(node->mChildren[i], scene, out, modelDir, isGltf);
     }
 }
 
@@ -193,10 +250,13 @@ Material* Model::GetOrCreateMaterial(const std::string paths[Material::kSlotCoun
     return ptr;
 }
 
-Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
+ParsedMesh Model::ProcessMesh(const aiMesh* mesh, const aiScene* scene,
+                              const std::filesystem::path& modelDir,
+                              bool isGltf)
 {
-    std::vector<Vertex> vertices;
-    std::vector<u32> indices;
+    ParsedMesh out;
+    std::vector<Vertex>& vertices = out.vertices;
+    std::vector<u32>& indices = out.indices;
 
     vertices.reserve(mesh->mNumVertices);
 
@@ -248,9 +308,7 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
     MaterialParams params;
 
     if (mesh->mMaterialIndex >= 0) {
-        aiMaterial* material = scene->mMaterials[mesh->mMaterialIndex];
-        const std::filesystem::path modelDir =
-            std::filesystem::path(m_ModelPath).parent_path();
+        const aiMaterial* material = scene->mMaterials[mesh->mMaterialIndex];
 
         // glTF2 贴图在 assimp 5.x 的落点：MR 合图 → METALNESS(15)/
         // DIFFUSE_ROUGHNESS(16)（同图双槽）；baseColor → BASE_COLOR(12)。
@@ -304,7 +362,7 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
         //     roughness 有独立贴图 → 因子 1（G 通道直达）；
         //     无贴图 → Blinn-Phong Ns 换算 R=sqrt(2/(Ns+2))，
         //     补偿默认 MR 贴图 G=0.5 → 因子 = 2R
-        if (m_IsGltf) {
+        if (isGltf) {
             ai_real metallic = 1.0f, roughness = 1.0f;
             material->Get(AI_MATKEY_METALLIC_FACTOR, metallic);
             material->Get(AI_MATKEY_ROUGHNESS_FACTOR, roughness);
@@ -323,11 +381,12 @@ Scope<Mesh> Model::ProcessMesh(aiMesh* mesh, const aiScene* scene)
         }
     }
 
-    // 材质缓存：同贴图组合+因子 → 同一 Material（一个描述符集）
-    Material* mat = GetOrCreateMaterial(texPaths, params);
-
-    return CreateScope<Mesh>(
-        m_Device, m_Allocator, vertices, indices, mat);
+    // 材质创建延迟到 GPU 构建段（FromParsed）：解析期只留引用
+    for (u32 slot = 0; slot < Material::kSlotCount; ++slot) {
+        out.texPaths[slot] = std::move(texPaths[slot]);
+    }
+    out.params = params;
+    return out;
 }
 
 } // namespace Engine

Engine/src/Renderer/Passes/BloomPass.cpp

@@ -8,9 +8,10 @@
 namespace Engine::Passes
 {
     BloomPass::BloomPass(FullscreenGpu& gpu, PostProcessParams& params,
+                         FrameSceneData& frame,
                          RG::TextureHandle hdrColor,
                          RG::TextureHandle bloomA, RG::TextureHandle bloomB)
-        : m_Gpu(gpu), m_Params(params),
+        : m_Gpu(gpu), m_Params(params), m_Frame(frame),
           m_HDR(hdrColor), m_A(bloomA), m_B(bloomB)
     {
         m_ExtractSets.resize(kMaxFramesInFlight);
@@ -25,9 +26,13 @@ namespace Engine::Passes
 
     void BloomPass::Register(RG::RenderGraph& rg)
     {
-        // Extract：HDR → 半分辨率 A（软阈值由 frag 的 param0 传入）
+        // bloom 关闭：三个子 pass 全部空转（省两次模糊 + 一次提取），
+        // tonemap 侧强度同步归零不采样叠加。
+        // 注意开关判断必须在录制 lambda 内（每帧评估），Register 只在
+        // RG 编译时跑一次
         rg.AddPass("BloomExtract",
             [this](RG::RenderGraphContext& ctx) {
+                if (!m_Frame.bloomEnabled) return;
                 m_Gpu.RecordFullscreen(ctx, ctx.View(m_A), m_Gpu.extractFrag,
                                        m_ExtractSets[ctx.frame.frameSlot],
                                        FullscreenPush{0.0f, m_Params.bloomThreshold, 0.0f, 0.0f});
@@ -39,6 +44,7 @@ namespace Engine::Passes
         // BlurH：A → B（水平步长经 push 传入）
         rg.AddPass("BloomBlurH",
             [this](RG::RenderGraphContext& ctx) {
+                if (!m_Frame.bloomEnabled) return;
                 m_Gpu.RecordFullscreen(ctx, ctx.View(m_B), m_Gpu.blurFrag,
                                        m_BlurHSets[ctx.frame.frameSlot],
                                        FullscreenPush{0.0f,
@@ -52,6 +58,7 @@ namespace Engine::Passes
         // BlurV：B → A（垂直步长）
         rg.AddPass("BloomBlurV",
             [this](RG::RenderGraphContext& ctx) {
+                if (!m_Frame.bloomEnabled) return;
                 m_Gpu.RecordFullscreen(ctx, ctx.View(m_A), m_Gpu.blurFrag,
                                        m_BlurVSets[ctx.frame.frameSlot],
                                        FullscreenPush{0.0f, 0.0f,

Engine/src/Renderer/Passes/FullscreenGpu.cpp

@@ -16,17 +16,17 @@ namespace Engine::Passes
                                  ShaderManager& shaders)
         : m_Device(device), m_Descriptors(descriptors)
     {
-        // Set0 = {HDR 颜色图, Bloom 图, 调试输入}，push = 16B 通用结构
-        //（binding2 只被 tonemap 的调试视图使用；bloom 系 shader 静态
-        //  使用的 binding 是布局的子集，合法）
-        std::vector<DescriptorBinding> bindings = {
-            {0, vk::DescriptorType::eCombinedImageSampler,
-             vk::ShaderStageFlagBits::eFragment},
-            {1, vk::DescriptorType::eCombinedImageSampler,
-             vk::ShaderStageFlagBits::eFragment},
-            {2, vk::DescriptorType::eCombinedImageSampler,
-             vk::ShaderStageFlagBits::eFragment},
-        };
+        // Set0 = {HDR 颜色图, Bloom 图, 调试输入, TAA, Velocity,
+        //         ShadowMap0/1/2}，push = 16B 通用结构（VUID-08878）。
+        // 后 6 个 binding 供 Tonemap 调试视图使用；其余全屏 shader
+        // 静态使用的 binding 是布局的子集，合法
+        std::vector<DescriptorBinding> bindings;
+        // 0-7 帧 RG 中间目标 + 8-11 环境调试视图（cube 三件套 + LUT）
+        for (u32 b = 0; b < 12; ++b) {
+            bindings.push_back(
+                {b, vk::DescriptorType::eCombinedImageSampler,
+                 vk::ShaderStageFlagBits::eFragment});
+        }
         setLayout = descriptors.CreateLayout(bindings);
 
         layout = descriptors.CreatePipelineLayout(

Engine/src/Renderer/Passes/SceneGpu.cpp

@@ -116,7 +116,8 @@ namespace Engine::Passes
                          .setBorderColor(vk::BorderColor::eFloatOpaqueWhite)
                          .setUnnormalizedCoordinates(VK_FALSE)
                          .setCompareEnable(VK_TRUE)
-                         .setCompareOp(vk::CompareOp::eGreater);   // 【诊断】方向对照
+                         // ref < stored → lit=1：片元比遮挡物近即被照亮
+                         .setCompareOp(vk::CompareOp::eLess);
         shadowSampler = device.GetDevice().createSampler(shadowSamplerInfo);
 
         // per-slot 描述符分配 + Set0 绑定（buffer 句柄全程稳定，不用重绑）

Engine/src/Renderer/Passes/ShadowPass.cpp

@@ -16,6 +16,10 @@ namespace Engine::Passes
         // 材质/实体数据驱动剔除，而非 pass 逻辑特判。
         // （每级联独立剔除是后续优化点：shadowView 细分为 per-cascade 列表）
         rg.AddPass("Shadow", [this](RG::RenderGraphContext& ctx) {
+            // 阴影关闭：整 pass 空转（省一整趟场景绘制）。图仍是
+            // Clear=1.0，frag 侧 cascadeSplits.w 短路根本不采样
+            if (!m_Frame.shadowsEnabled) return;
+
             const vk::CommandBuffer cmd = ctx.cmd;
 
             for (u32 c = 0; c < kMaxCascades; ++c) {

Engine/src/Renderer/Passes/SkyboxPass.cpp

@@ -74,15 +74,30 @@ namespace Engine::Passes
         vkDevice.destroyDescriptorSetLayout(m_SetLayout);
     }
 
+    void SkyboxPass::OnEnvironmentChanged(const Cubemap* env)
+    {
+        m_Env = env;
+        if (!m_Env) return;
+
+        // cube view 全程稳定语义是"单个环境内"；切换环境 = 新 view，
+        // per-slot set 全部重绑
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Descriptors.UpdateTexture(m_Sets[slot], 0,
+                                        m_Env->GetSampler(),
+                                        m_Env->GetView());
+        }
+    }
+
     void SkyboxPass::Register(RG::RenderGraph& rg)
     {
         rg.AddPass("Skybox", [this](RG::RenderGraphContext& ctx) {
             ZoneScopedN("Skybox");
             const vk::CommandBuffer cmd = ctx.cmd;
 
+            const glm::vec4& cc = m_Frame.ClearColor;
             vk::ClearValue clear;
             clear.color = vk::ClearColorValue(
-                std::array<float, 4>{0.1f, 0.1f, 0.1f, 1.0f});
+                std::array<float, 4>{cc.r, cc.g, cc.b, cc.a});
 
             vk::RenderingAttachmentInfo color;
             color.setImageView(ctx.View(m_HDRColor))
@@ -98,13 +113,13 @@ namespace Engine::Passes
 
             cmd.beginRendering(renderingInfo);
 
-            if (m_Env) {
+            if (m_Env && m_Frame.DrawSkybox) {
                 // 无环境贴图：只清屏（Forward loadOp=Load 依赖这里先写）
-                std::array<vk::ShaderStageFlagBits, 2> stages = {
+                const std::array<vk::ShaderStageFlagBits, 2> stages = {
                     vk::ShaderStageFlagBits::eVertex,
                     vk::ShaderStageFlagBits::eFragment
                 };
-                std::array<vk::ShaderEXT, 2> shaders = {m_Vert, m_Frag};
+                const std::array<vk::ShaderEXT, 2> shaders = {m_Vert, m_Frag};
                 cmd.bindShadersEXT(stages, shaders);
 
                 cmd.bindDescriptorSets(vk::PipelineBindPoint::eGraphics,
@@ -135,12 +150,12 @@ namespace Engine::Passes
                 cmd.setVertexInputEXT({}, {});
 
                 cmd.setRasterizationSamplesEXT(vk::SampleCountFlagBits::e1);
-                vk::SampleMask sampleMask = 0xFFFFFFFF;
+                constexpr 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,
+                const vk::Viewport viewport(0.0f, 0.0f,
                                       static_cast<float>(ctx.extent.width),
                                       static_cast<float>(ctx.extent.height),
                                       0.0f, 1.0f);
@@ -156,8 +171,8 @@ namespace Engine::Passes
                              .setDstAlphaBlendFactor(vk::BlendFactor::eZero)
                              .setAlphaBlendOp(vk::BlendOp::eAdd);
                 // ArrayProxyNoTemporaries 拒绝临时量：具名局部变量传入
-                vk::Bool32 blendEnable = VK_FALSE;
-                vk::ColorComponentFlags colorWriteMask =
+                constexpr vk::Bool32 blendEnable = VK_FALSE;
+                constexpr vk::ColorComponentFlags colorWriteMask =
                     vk::ColorComponentFlagBits::eR |
                     vk::ColorComponentFlagBits::eG |
                     vk::ColorComponentFlagBits::eB |

Engine/src/Renderer/Passes/TAAPass.cpp

@@ -7,10 +7,10 @@
 
 namespace Engine::Passes
 {
-    TAAPass::TAAPass(FullscreenGpu& gpu, f32 historyAlpha,
+    TAAPass::TAAPass(FullscreenGpu& gpu, FrameSceneData& frame, f32 historyAlpha,
                      RG::TextureHandle hdrColor, RG::TextureHandle velocity,
                      RG::TextureHandle taaColor)
-        : m_Gpu(gpu), m_HistoryAlpha(historyAlpha),
+        : m_Gpu(gpu), m_Frame(frame), m_HistoryAlpha(historyAlpha),
           m_HDR(hdrColor), m_Velocity(velocity), m_TAAColor(taaColor)
     {
         m_Sets.resize(kMaxFramesInFlight);
@@ -23,9 +23,12 @@ namespace Engine::Passes
     {
         rg.AddPass("TAA",
             [this](RG::RenderGraphContext& ctx) {
+                // TAA 关闭：alpha=0 → mix(cur, hist, 0) 直通当前帧
+                //（抖动已由 SceneRenderer 同步停掉，画面静止无锯齿晃动）
+                const f32 alpha = m_Frame.taaEnabled ? m_HistoryAlpha : 0.0f;
                 m_Gpu.RecordFullscreen(ctx, ctx.View(m_TAAColor), m_Gpu.taaFrag,
                                        m_Sets[ctx.frame.frameSlot],
-                                       FullscreenPush{0.0f, m_HistoryAlpha, 0.0f, 0.0f});
+                                       FullscreenPush{0.0f, alpha, 0.0f, 0.0f});
             })
             .Read(m_HDR, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment)

Engine/src/Renderer/Passes/TonemapPass.cpp

@@ -3,22 +3,35 @@
 //
 
 #include <Engine/Renderer/Passes/TonemapPass.h>
+#include <Engine/Renderer/Cubemap.h>
+#include <Engine/Vulkan/DescriptorManager.h>
 #include <Engine/Core/FrameContext.h>
 
 namespace Engine::Passes
 {
-    TonemapPass::TonemapPass(FullscreenGpu& gpu, PostProcessParams& params,
+    TonemapPass::TonemapPass(FullscreenGpu& gpu, DescriptorManager& descriptors,
+                             PostProcessParams& params,
+                             FrameSceneData& frame,
                              RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
                              RG::TextureHandle temporalDiff,
-                             RG::TextureHandle backbuffer)
-        : m_Gpu(gpu), m_Params(params),
+                             RG::TextureHandle backbuffer,
+                             RG::TextureHandle rawHdr,
+                             RG::TextureHandle velocity,
+                             RG::TextureHandle shadowMap0,
+                             RG::TextureHandle shadowMap1,
+                             RG::TextureHandle shadowMap2)
+        : m_Gpu(gpu), m_Descriptors(descriptors), m_Params(params), m_Frame(frame),
           m_HDR(hdrColor), m_A(bloomA), m_TemporalDiff(temporalDiff),
-          m_Backbuffer(backbuffer)
+          m_Backbuffer(backbuffer),
+          m_RawHDR(rawHdr), m_Velocity(velocity),
+          m_ShadowMaps{shadowMap0, shadowMap1, shadowMap2}
     {
         m_Sets.resize(kMaxFramesInFlight);
         for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
             m_Sets[slot] = m_Gpu.AllocateSet();
         }
+        // 环境调试视图源（binding 8-11）由 OnEnvironmentChanged 填充
+        //（Renderer 构造尾部统一广播一次）
     }
 
     void TonemapPass::Register(RG::RenderGraph& rg)
@@ -30,17 +43,31 @@ namespace Engine::Passes
                                        m_Sets[ctx.frame.frameSlot],
                                        FullscreenPush{m_Params.exposure,
                                                       m_Params.useACES ? 1.0f : 0.0f,
-                                                      m_Params.bloomStrength,
-                                                      m_Params.showTemporalDiff ? 1.0f : 0.0f});
+                                                      m_Frame.bloomEnabled
+                                                          ? m_Params.bloomStrength
+                                                          : 0.0f,
+                                                      static_cast<f32>(
+                                                          m_Frame.debugView)});
             })
             .Write(m_Backbuffer, RG::ResourceUse::ColorAttachment)
             .Read(m_HDR, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment)
             .Read(m_A, RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment)
-            // shader 静态采样 binding2（调试视图开关在 push constant 里），
-            // 声明读否则本 pass 的产出无人消费会被剔除、镜像也不会分配
+            // 调试视图源全部声明读：路由开关在 push constant（每帧变），
+            // 绑定集与 barrier 必须静态成立——这是"声明超集、运行时路由"
+            // 模式，避免切换视图时重绑描述符
             .Read(m_TemporalDiff, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_RawHDR, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_Velocity, RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_ShadowMaps[0], RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_ShadowMaps[1], RG::ResourceUse::ShaderRead,
+                  vk::ShaderStageFlagBits::eFragment)
+            .Read(m_ShadowMaps[2], RG::ResourceUse::ShaderRead,
                   vk::ShaderStageFlagBits::eFragment);
     }
 
@@ -50,6 +77,34 @@ namespace Engine::Passes
             m_Gpu.UpdateSetInput(m_Sets[slot], 0, rg, m_HDR, slot);
             m_Gpu.UpdateSetInput(m_Sets[slot], 1, rg, m_A, slot);
             m_Gpu.UpdateSetInput(m_Sets[slot], 2, rg, m_TemporalDiff, slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 3, rg, m_RawHDR, slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 4, rg, m_Velocity, slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 5, rg, m_ShadowMaps[0], slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 6, rg, m_ShadowMaps[1], slot);
+            m_Gpu.UpdateSetInput(m_Sets[slot], 7, rg, m_ShadowMaps[2], slot);
+        }
+    }
+
+    void TonemapPass::OnEnvironmentChanged(const Cubemap* env)
+    {
+        m_Env = env;
+        if (!m_Env) return;
+
+        // 环境调试视图源（binding 8-11）：cube 三件套 + LUT。
+        // 静态绑定超集模式——只在切换环境时重绑，每帧零成本
+        for (u32 slot = 0; slot < kMaxFramesInFlight; ++slot) {
+            m_Descriptors.UpdateTexture(m_Sets[slot], 8,
+                                        m_Env->GetSampler(),
+                                        m_Env->GetView());
+            m_Descriptors.UpdateTexture(m_Sets[slot], 9,
+                                        m_Env->GetSampler(),
+                                        m_Env->GetIrradianceView());
+            m_Descriptors.UpdateTexture(m_Sets[slot], 10,
+                                        m_Env->GetSampler(),
+                                        m_Env->GetPrefilteredView());
+            m_Descriptors.UpdateTexture(m_Sets[slot], 11,
+                                        m_Env->GetBrdfLutSampler(),
+                                        m_Env->GetBrdfLutView());
         }
     }
 

Engine/src/Renderer/Renderer.cpp

@@ -16,6 +16,7 @@
 #include <Engine/Renderer/Passes/ShadowPass.h>
 #include <Engine/Renderer/Passes/ForwardPass.h>
 #include <Engine/Renderer/Cubemap.h>
+#include <Engine/Renderer/IblGenerator.h>
 #include <Engine/Renderer/Passes/TAAPass.h>
 #include <Engine/Renderer/Passes/BloomPass.h>
 #include <Engine/Renderer/Passes/TonemapPass.h>
@@ -40,85 +41,10 @@
 #include <glm/glm.hpp>
 
 #include <array>
-#include <filesystem>
 #include <limits>
-#include <bit>
-#include <fstream>
 
 namespace Engine
 {
-    namespace
-    {
-        // fp16 → f32（截图用，规范路径足够）
-        f32 HalfToFloat(u16 h)
-        {
-            const u32 sign = (h & 0x8000u) << 16u;
-            const u32 exp = (h >> 10u) & 0x1Fu;
-            const u32 man = h & 0x3FFu;
-            u32 bits;
-            if (exp == 0) {
-                if (man == 0) return std::bit_cast<f32>(sign);
-                u32 e = 127 - 15 + 1;
-                u32 m = man;
-                while (!(m & 0x400u)) { m <<= 1u; --e; }
-                bits = sign | (e << 23u) | ((m & 0x3FFu) << 13u);
-            } else if (exp == 31) {
-                bits = sign | 0x7F800000u | (man << 13u);
-            } else {
-                bits = sign | ((exp - 15u + 127u) << 23u) | (man << 13u);
-            }
-            return std::bit_cast<f32>(bits);
-        }
-
-        // RGBA16F 线性数据 → 24bpp BMP（行倒序；depthMode 读 f32 深度）
-        void WriteHdrBMP(const std::string& path, const void* data,
-                         u32 w, u32 h, bool depthMode = false)
-        {
-            const auto* src = static_cast<const u16*>(data);
-            const u32 rowBytes = w * 3u;
-            const u32 pad = (4u - rowBytes % 4u) % 4u;
-            const u32 stride = rowBytes + pad;
-            const u32 size = 54u + stride * h;
-
-            std::vector<u8> bmp(size, 0);
-            auto put = [&bmp](u32 off, u32 v, u32 n) {
-                for (u32 i = 0; i < n; ++i)
-                    bmp[off + i] = static_cast<u8>((v >> (8u * i)) & 0xFFu);
-            };
-            bmp[0] = 'B'; bmp[1] = 'M';
-            put(2, size, 4); put(10, 54, 4); put(14, 40, 4);
-            put(18, w, 4); put(22, h, 4); put(26, 1, 2); put(28, 24, 2);
-            put(34, stride * h, 4);
-
-            for (u32 y = 0; y < h; ++y) {
-                const u32 dstRow = h - 1 - y;   // BMP 底行在前
-                for (u32 x = 0; x < w; ++x) {
-                    const u16* px = src + (static_cast<size_t>(y) * w + x) * 4;
-                    const u32 off = 54u + dstRow * stride + x * 3u;
-                    for (u32 c = 0; c < 3; ++c) {
-                        f32 v;
-                        if (depthMode) {
-                            // D32 buffer：f32 深度，反相显示（近=白）
-                            v = 1.0f - reinterpret_cast<const f32*>(
-                                data)[static_cast<size_t>(y) * w + x];
-                        } else {
-                            v = HalfToFloat(px[c]) * 0.6f;   // 匹配默认 exposure
-                            // ACES 近似（与 TonemapPass 同族），还原屏幕观感
-                            v = (v * (2.51f * v + 0.03f)) /
-                                (v * (2.43f * v + 0.59f) + 0.14f);
-                            v = glm::pow(glm::clamp(v, 0.0f, 1.0f),
-                                         1.0f / 2.2f);
-                        }
-                        bmp[off + c] = static_cast<u8>(v * 255.0f);
-                    }
-                }
-            }
-
-            std::ofstream file(path, std::ios::binary);
-            file.write(reinterpret_cast<const char*>(bmp.data()),
-                       static_cast<std::streamsize>(bmp.size()));
-        }
-    }
     Renderer::Renderer(Window& window)
         : m_Window(window)
     {
@@ -148,17 +74,12 @@ namespace Engine
         m_FullscreenGpu = CreateScope<Passes::FullscreenGpu>(
             *m_Device, *m_DescriptorManager, *m_ShaderManager);
 
-        // 3.5 环境贴图（IBL）：HDR 等距柱状 → cube + 预积分。
-        //     加载失败/文件缺失一律中性常数兜底（描述符恒有效）
-        const std::string envPath = "assets/textures/venice_dawn_1_4k.hdr";
-        if (std::filesystem::exists(envPath)) {
-            m_Environment = Cubemap::CreateFromHdrEquirect(
-                *m_Device, *m_Allocator, envPath, 1024);
-        }
-        if (!m_Environment) {
-            m_Environment = Cubemap::CreateNeutral(
-                *m_Device, *m_Allocator, glm::vec3(0.1f));
-        }
+        // 3.5 环境预积分生成端（compute 链）+ 中性兜底四件套。
+        //     真实环境由 Application 经 AssetManager 加载后 SetEnvironment
+        //     （资产路径归资产层，Renderer 不感知具体文件）
+        m_Ibl = CreateScope<IblGenerator>(
+            *m_Device, *m_Allocator, *m_DescriptorManager, *m_ShaderManager);
+        m_Environment = m_Ibl->Neutral(glm::vec3(0.1f));
 
         // 4. Swapchain（只管呈现；渲染目标统一归 RenderGraph）
         m_Swapchain = CreateScope<Swapchain>(*m_Device, *m_Allocator, m_Window);
@@ -175,22 +96,22 @@ namespace Engine
         // 7. 资产上传服务（渲染层对外提供的 IAssetUploader 实现）
         m_AssetUploader = CreateScope<VulkanAssetUploader>(
             *m_Device, *m_Allocator, *m_DescriptorManager,
-            m_SceneGpu->textureSetLayout);
+            m_SceneGpu->textureSetLayout, *m_Ibl);
 
         // 8. RenderGraph：资源声明 + pass 注册 + 首次编译
-        ENGINE_INFO("[diag] BuildGraph begin");
         m_PostParams = CreateScope<Passes::PostProcessParams>();
         BuildGraph();
-        ENGINE_INFO("[diag] BuildGraph done");
 
         // 首次编译后物理镜像就位：各 pass 绑定自己的输入描述符
         for (auto& pass : m_Passes) {
             pass->OnTargetsRebuilt(*m_RenderGraph);
         }
-        ENGINE_INFO("[diag] targets rebuilt");
-        // 环境辐照度绑定（view 全程稳定，绑一次）
+        // 环境辐照度绑定（view 在单个环境内全程稳定，绑一次）
         m_SceneGpu->BindEnvironment(*m_Environment);
-        ENGINE_INFO("[diag] environment bound");
+        // 环境消费端 pass（Skybox 背景 / Tonemap 调试视图）同步初始绑定
+        for (auto& pass : m_Passes) {
+            pass->OnEnvironmentChanged(m_Environment.get());
+        }
 
         // 9. GPU timestamp：每 pass 两枚（前后），按帧槽位分区
         m_QueriesPerFrame = static_cast<u32>(m_RenderGraph->GetPassCount()) * 2;
@@ -246,14 +167,59 @@ namespace Engine
     void Renderer::SetBloomThreshold(f32 v) { m_PostParams->bloomThreshold = v; }
     f32 Renderer::GetBloomStrength() const { return m_PostParams->bloomStrength; }
     void Renderer::SetBloomStrength(f32 v) { m_PostParams->bloomStrength = v; }
-    bool Renderer::GetShowTemporalDiff() const { return m_PostParams->showTemporalDiff; }
-    void Renderer::SetShowTemporalDiff(bool b) { m_PostParams->showTemporalDiff = b; }
+
+    // ---- 特性开关 / 调试视图：转发 SceneRenderer（状态唯一持有者）----
+    bool Renderer::GetShadowsEnabled() const
+    {
+        return m_SceneRenderer->Frame().shadowsEnabled;
+    }
+    void Renderer::SetShadowsEnabled(bool b) { m_SceneRenderer->SetShadowsEnabled(b); }
+    bool Renderer::GetTAAEnabled() const { return m_SceneRenderer->Frame().taaEnabled; }
+    void Renderer::SetTAAEnabled(bool b) { m_SceneRenderer->SetTAAEnabled(b); }
+    bool Renderer::GetBloomEnabled() const
+    {
+        return m_SceneRenderer->Frame().bloomEnabled;
+    }
+    void Renderer::SetBloomEnabled(bool b) { m_SceneRenderer->SetBloomEnabled(b); }
+    int Renderer::GetDebugView() const { return m_SceneRenderer->Frame().debugView; }
+    void Renderer::SetDebugView(int v) { m_SceneRenderer->SetDebugView(v); }
+
+    void Renderer::SetEnvironment(const Cubemap* env)
+    {
+        if (!env) env = m_Environment.get();
+
+        // 描述符 host 更新要求 set 不被 in-flight 帧使用（validation 严格
+        // 校验）：waitIdle 两槽位全停飞后统一换绑。环境切换低频，等待
+        // 成本 ≈ 单帧 GPU 尾巴；生成端（compute 预积分）已在资产加载期完成
+        m_Device->GetDevice().waitIdle();
+        m_SceneGpu->BindEnvironment(*env);
+        for (auto& pass : m_Passes) {
+            pass->OnEnvironmentChanged(env);
+        }
+    }
+
+    const std::vector<const char*>& Renderer::GetDebugViewNames() const
+    {
+        static const std::vector<const char*> kNames = {
+            "Final", "TemporalDiff", "HDR", "TAAColor", "Bloom",
+            "Velocity", "ShadowMap0", "ShadowMap1", "ShadowMap2",
+            "EnvCube", "Irradiance", "Prefilter", "BrdfLut",
+        };
+        static_assert(int(Passes::kDebugViewCount) == 13,
+                      "DebugView 枚举与名表同步");
+        return kNames;
+    }
 
     u32 Renderer::GetMainInstanceCount() const
     {
         return m_SceneRenderer->GetMainInstanceCount();
     }
 
+    bool Renderer::GetInstancesTruncated() const
+    {
+        return m_SceneRenderer->Frame().instancesTruncated;
+    }
+
     u32 Renderer::GetShadowInstanceCount() const
     {
         return m_SceneRenderer->GetShadowInstanceCount();
@@ -290,11 +256,9 @@ namespace Engine
 
         // HDR 颜色图：Forward 的输出（线性光照值 + 抖动），TAA 显式完成
         // 时域累积后写入 TAAColor（对齐 UE/Unity 管线）
-        // extraUsage：调试截图需要 TRANSFER_SRC（第 60 帧 copy 到 buffer）
         m_HDRColor = m_RenderGraph->CreateTexture(
             "HDRColor",
-            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat, 0, 0, 1.0f,
-                            vk::ImageUsageFlagBits::eTransferSrc});
+            RG::TextureDesc{vk::Format::eR16G16B16A16Sfloat});
 
         // velocity：clip 空间运动向量（TAA 反推历史像素位置）
         m_Velocity = m_RenderGraph->CreateTexture(
@@ -332,15 +296,17 @@ namespace Engine
         m_Passes.push_back(CreateScope<Passes::ForwardPass>(
             *m_SceneGpu, frame, m_HDRColor, m_Velocity, m_Depth, m_ShadowMaps));
         m_Passes.push_back(CreateScope<Passes::TAAPass>(
-            *m_FullscreenGpu, /*historyAlpha*/ 0.9f, m_HDRColor, m_Velocity,
-            m_TAAColor));
+            *m_FullscreenGpu, frame, /*historyAlpha*/ 0.9f, m_HDRColor,
+            m_Velocity, m_TAAColor));
         m_Passes.push_back(CreateScope<Passes::TemporalDiffPass>(
             *m_FullscreenGpu, /*gain*/ 8.0f, m_HDRColor, m_TemporalDiff));
         m_Passes.push_back(CreateScope<Passes::BloomPass>(
-            *m_FullscreenGpu, *m_PostParams, m_TAAColor, m_BloomA, m_BloomB));
+            *m_FullscreenGpu, *m_PostParams, frame, m_TAAColor, m_BloomA,
+            m_BloomB));
         m_Passes.push_back(CreateScope<Passes::TonemapPass>(
-            *m_FullscreenGpu, *m_PostParams, m_TAAColor, m_BloomA,
-            m_TemporalDiff, m_Backbuffer));
+            *m_FullscreenGpu, *m_DescriptorManager, *m_PostParams, frame,
+            m_TAAColor, m_BloomA, m_TemporalDiff, m_Backbuffer, m_HDRColor,
+            m_Velocity, m_ShadowMaps[0], m_ShadowMaps[1], m_ShadowMaps[2]));
         m_Passes.push_back(CreateScope<Passes::ImGuiPass>(
             &m_ImGuiLayer, m_Backbuffer));
         m_Passes.push_back(CreateScope<Passes::PresentPass>(m_Backbuffer));
@@ -472,68 +438,6 @@ namespace Engine
 
         m_RenderGraph->Execute(frameCtx);
 
-        // 调试截图：帧尾把 TAAColor 拷进 host buffer（一次）
-        ++m_FrameCounter;
-        const bool doCapture = !m_Captured && m_FrameCounter == 60;
-        vk::Image capImage{};
-        vk::Extent2D capExtent{};
-        vk::Format capFormat{};
-        vk::ImageLayout capResting{};
-        if (doCapture &&
-            m_RenderGraph->GetCaptureTarget(m_HDRColor, frameSlot,
-                capImage, capExtent, capFormat, capResting)) {
-            const vk::DeviceSize capBytes = static_cast<vk::DeviceSize>(
-                capExtent.width) * capExtent.height * 8;   // RGBA16F
-            const bool capDepth = capFormat == vk::Format::eD32Sfloat;
-
-            m_CaptureAlloc = m_Allocator->CreateBuffer(
-                capBytes,
-                vk::BufferUsageFlagBits::eTransferDst,
-                vk::MemoryPropertyFlagBits::eHostVisible |
-                vk::MemoryPropertyFlagBits::eHostCoherent);
-
-            // resting → TRANSFER_SRC → 拷贝 → 恢复 resting
-            const vk::ImageAspectFlagBits capAspect =
-                capDepth ? vk::ImageAspectFlagBits::eDepth
-                         : vk::ImageAspectFlagBits::eColor;
-            vk::ImageMemoryBarrier2 toSrc;
-            toSrc.setSrcStageMask(vk::PipelineStageFlagBits2::eAllCommands)
-                  .setSrcAccessMask(vk::AccessFlagBits2::eMemoryWrite)
-                  .setDstStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                  .setDstAccessMask(vk::AccessFlagBits2::eTransferRead)
-                  .setOldLayout(capResting)
-                  .setNewLayout(vk::ImageLayout::eTransferSrcOptimal)
-                  .setImage(capImage)
-                  .setSubresourceRange({capAspect, 0, 1, 0, 1});
-            frameCtx.cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(toSrc));
-
-            vk::BufferImageCopy capRegion;
-            capRegion.setBufferOffset(0)
-                     .setBufferRowLength(0)
-                     .setBufferImageHeight(0)
-                     .setImageSubresource({capAspect, 0, 0, 1})
-                     .setImageOffset({0, 0, 0})
-                     .setImageExtent({capExtent.width, capExtent.height, 1});
-            frameCtx.cmd.copyImageToBuffer(
-                capImage, vk::ImageLayout::eTransferSrcOptimal,
-                m_CaptureAlloc.Buffer, capRegion);
-
-            vk::ImageMemoryBarrier2 back;
-            back.setSrcStageMask(vk::PipelineStageFlagBits2::eTransfer)
-                .setSrcAccessMask(vk::AccessFlagBits2::eTransferRead)
-                .setDstStageMask(vk::PipelineStageFlagBits2::eAllCommands)
-                .setDstAccessMask(vk::AccessFlagBits2::eMemoryRead |
-                                  vk::AccessFlagBits2::eMemoryWrite)
-                .setOldLayout(vk::ImageLayout::eTransferSrcOptimal)
-                .setNewLayout(capResting)
-                .setImage(capImage)
-                .setSubresourceRange({vk::ImageAspectFlagBits::eColor,
-                                      0, 1, 0, 1});
-            frameCtx.cmd.pipelineBarrier2(
-                vk::DependencyInfo{}.setImageMemoryBarriers(back));
-        }
-
         m_CommandManager->EndFrame(frameSlot);
 
         // 5. 提交
@@ -577,21 +481,6 @@ namespace Engine
             HandleResize();
         }
 
-        // 调试截图落盘：等 fence（copy 已提交完成）→ fp16 → BMP
-        if (doCapture) {
-            // 诊断截图落盘：等 fence（copy 已提交完成）→ BMP（返回值仅超时标志）
-            (void)m_Device->GetDevice().waitForFences(
-                m_SyncManager->GetFrameFence(frameSlot), VK_TRUE,
-                UINT64_MAX);
-            WriteHdrBMP("capture.bmp", m_CaptureAlloc.MappedPtr,
-                        capExtent.width, capExtent.height, false);
-            m_Allocator->Destroy(m_CaptureAlloc);
-            m_Captured = true;
-            ENGINE_INFO("Debug capture written: capture.bmp (HDRColor) "
-                        "shadowInstances={}",
-                        m_SceneRenderer->GetShadowInstanceCount());
-        }
-
         m_CurrentFrame = (m_CurrentFrame + 1) % kMaxFramesInFlight;
     }
 

Engine/src/Renderer/SceneRenderer.cpp

@@ -157,8 +157,8 @@ namespace Engine
         ubo.projNoJitter = ubo.proj;
 
         // 子像素抖动（Halton×8）：光栅化采样位置逐帧偏移，TAA 累积去锯齿。
-        // NDC 偏移 = 像素偏移 × 2 / 尺寸
-        {
+        // NDC 偏移 = 像素偏移 × 2 / 尺寸。TAA 关闭时停抖（直通画面静止）
+        if (m_TAAEnabled) {
             const u32 idx = static_cast<u32>(m_FrameIndex % 8);
             const f32 jx = Halton(idx + 1, 2) - 0.5f;
             const f32 jy = Halton(idx + 1, 3) - 0.5f;
@@ -166,7 +166,16 @@ namespace Engine
             ubo.proj[2][1] += jy * 2.0f / static_cast<f32>(targetExtent.height);
         }
 
-        ubo.ambientColor = scene.GetSettings().ambientColor;
+        ubo.ambientColor = scene.GetSettings().AmbientColor;
+
+        // ---- 运行时开关/调试视图落帧（各 pass 与 frag 只读契约字段）----
+        m_Frame.ClearColor = scene.GetSettings().ClearColor;
+        m_Frame.DrawSkybox = scene.GetSettings().DrawSkybox;
+        m_Frame.shadowsEnabled = m_ShadowsEnabled;
+        m_Frame.taaEnabled = m_TAAEnabled;
+        m_Frame.bloomEnabled = m_BloomEnabled;
+        m_Frame.debugView = m_DebugView;
+        ubo.cascadeSplits.w = m_ShadowsEnabled ? 1.0f : 0.0f;
 
         // 天空盒：视图去平移（只留旋转）× 投影 → 逆矩阵，push 给 Skybox pass；
         // 反投影远平面点 = 相机相对世界方向，frag normalize 后采样 cube
@@ -296,16 +305,24 @@ namespace Engine
                             instanceTruncated = true;
                             return;
                         }
+                        // lookup 值 = batches 索引（此前误存 worlds 索引：
+                        // 单实体时两序号恰好恒等，多实体共享 mesh 后分叉
+                        // → batches[worldsIdx] 越界）。batch.first（GPU
+                        // firstInstance 基址）才是 worlds 索引，语义别混
                         const auto [it, inserted] = s.lookup.emplace(
-                            mesh.get(), static_cast<u32>(s.worlds.size()));
+                            mesh.get(), static_cast<u32>(s.result.batches.size()));
                         if (inserted) {
-                            s.result.batches.push_back({mesh.get(),
-                                static_cast<u32>(s.worlds.size()), 0});
+                            s.result.batches.push_back({mesh.get(), 0, 0});
+                            s.batchWorlds.emplace_back();
+                            s.batchPrevWorlds.emplace_back();
                         }
-                        s.result.batches[it->second].count++;
+                        // 实例归入自己的 batch（mesh 分组），保证同 batch
+                        // 实例在扁平化后连续——firstInstance 语义的前提
+                        const u32 batchIdx = it->second;
+                        s.result.batches[batchIdx].count++;
+                        s.batchWorlds[batchIdx].push_back(t.World());
+                        s.batchPrevWorlds[batchIdx].push_back(prevWorld);
                         s.result.instanceCount++;
-                        s.worlds.push_back(t.World());
-                        s.prevWorlds.push_back(prevWorld);
                         ++totalInstances;
                     };
 
@@ -316,21 +333,25 @@ namespace Engine
                 }
             });
 
-        // ---- 4. 拼接：[主段|级联0|级联1|级联2]，batch.first 加段基址 ----
+        // ---- 4. 拼接：[主段|级联0|级联1|级联2]，batch 内实例连续扁平化 ----
         auto appendView = [&](CullResult& dst, ViewScratch& s) {
-            const u32 base = static_cast<u32>(m_Frame.instanceModels.size());
+            u32 base = static_cast<u32>(m_Frame.instanceModels.size());
             dst.batches.clear();
             dst.batches.reserve(s.result.batches.size());
-            for (auto& b : s.result.batches) {
-                b.first += base;
+            for (size_t i = 0; i < s.result.batches.size(); ++i) {
+                InstanceBatch b = s.result.batches[i];
+                b.first = base;          // 扁平化后的连续区间基址
+                base += b.count;
                 dst.batches.push_back(b);
+                m_Frame.instanceModels.insert(m_Frame.instanceModels.end(),
+                                              s.batchWorlds[i].begin(),
+                                              s.batchWorlds[i].end());
+                m_Frame.prevInstanceModels.insert(
+                    m_Frame.prevInstanceModels.end(),
+                    s.batchPrevWorlds[i].begin(),
+                    s.batchPrevWorlds[i].end());
             }
             dst.instanceCount = s.result.instanceCount;
-            m_Frame.instanceModels.insert(m_Frame.instanceModels.end(),
-                                          s.worlds.begin(), s.worlds.end());
-            m_Frame.prevInstanceModels.insert(m_Frame.prevInstanceModels.end(),
-                                              s.prevWorlds.begin(),
-                                              s.prevWorlds.end());
         };
         appendView(m_Frame.mainView, m_MainScratch);
         for (u32 c = 0; c < kMaxCascades; ++c) {
@@ -344,11 +365,13 @@ namespace Engine
         m_LastProjNoJitter = ubo.projNoJitter;
         ++m_FrameIndex;
 
+        // 截断状态每帧落遥测（ImGui 可见）——"只告警一次"的日志对
+        // 大场景不够：预算是动态穿越的，用户必须实时看到截断发生
+        m_Frame.instancesTruncated = instanceTruncated;
         if (instanceTruncated && !m_TruncationWarned) {
-            // 只告警一次（每帧告警会刷屏）
             m_TruncationWarned = true;
-            ENGINE_WARN("GPUDriven: instance buffer full ({}), entities truncated",
-                        kMaxInstances);
+            ENGINE_WARN("GPUDriven: instance buffer full ({}), entities truncated "
+                        "(见 Scene Debug 面板红字)", kMaxInstances);
         }
 
         m_Frame.collectMs = std::chrono::duration<f32, std::milli>(

Engine/src/Renderer/VulkanAssetUploader.cpp

@@ -4,6 +4,7 @@
 
 #include <Engine/Renderer/VulkanAssetUploader.h>
 #include <Engine/Renderer/Model.h>
+#include <Engine/Renderer/IblGenerator.h>
 
 namespace Engine {
 
@@ -14,10 +15,31 @@ namespace Engine {
             m_Device, m_Allocator, m_Descriptors, m_TextureLayout, path);
     }
 
+    Scope<Model> VulkanAssetUploader::LoadModel(const std::string& path,
+                                                ParsedModel&& parsed)
+    {
+        // 异步路径：assimp 解析已在 worker 线程完成，此处只剩 GPU 构建
+        return Model::FromParsed(m_Device, m_Allocator, m_Descriptors,
+                                 m_TextureLayout, std::move(parsed));
+    }
+
     Scope<Model> VulkanAssetUploader::CreatePlane(f32 halfSize, u8 r, u8 g, u8 b)
     {
         return Model::CreatePlane(m_Device, m_Allocator, m_Descriptors,
                                   m_TextureLayout, halfSize, r, g, b);
     }
 
+    Scope<Cubemap> VulkanAssetUploader::LoadEnvironment(const std::string& path)
+    {
+        // compute 预积分（毫秒级）：HDR → cube + irradiance/prefilter/LUT
+        return m_Ibl.FromHdrEquirect(path, 1024);
+    }
+
+    Scope<Cubemap> VulkanAssetUploader::LoadEnvironment(
+        const std::string& path, DecodedEquirect&& decoded)
+    {
+        // 异步路径：decode 已在 worker 线程完成，此处只剩 GPU 段
+        return m_Ibl.FromDecoded(decoded, 1024, path);
+    }
+
 } // namespace Engine

Engine/src/Vulkan/DescriptorManager.cpp

@@ -25,10 +25,11 @@ namespace Engine
 
     vk::DescriptorPool DescriptorManager::CreateChunk() const
     {
-        std::array<vk::DescriptorPoolSize, 3> poolSizes = {
+        std::array<vk::DescriptorPoolSize, 4> poolSizes = {
             vk::DescriptorPoolSize(vk::DescriptorType::eUniformBuffer, kChunkSets),
             vk::DescriptorPoolSize(vk::DescriptorType::eCombinedImageSampler, kChunkSets),
             vk::DescriptorPoolSize(vk::DescriptorType::eStorageBuffer, kChunkSets),
+            vk::DescriptorPoolSize(vk::DescriptorType::eStorageImage, kChunkSets),
         };
 
         vk::DescriptorPoolCreateInfo poolInfo;
@@ -133,6 +134,26 @@ namespace Engine
         m_Device.GetDevice().updateDescriptorSets(write, {});
     }
 
+    void DescriptorManager::UpdateStorageImage(const vk::DescriptorSet set,
+                                               const u32 binding,
+                                               const vk::ImageView view) const
+    {
+        // storage image 写（compute 输出）：layout 必须 eGeneral
+        vk::DescriptorImageInfo imageInfo;
+        imageInfo.setImageView(view)
+                 .setImageLayout(vk::ImageLayout::eGeneral);
+
+        vk::WriteDescriptorSet write;
+        write.setDstSet(set)
+             .setDstBinding(binding)
+             .setDstArrayElement(0)
+             .setDescriptorType(vk::DescriptorType::eStorageImage)
+             .setDescriptorCount(1)
+             .setImageInfo(imageInfo);
+
+        m_Device.GetDevice().updateDescriptorSets(write, {});
+    }
+
     vk::PipelineLayout DescriptorManager::CreatePipelineLayout(vk::DescriptorSetLayout layout,
                                                                const u32 pushConstantSize) const
     {
@@ -154,18 +175,18 @@ namespace Engine
     }
 
     vk::PipelineLayout DescriptorManager::CreatePipelineLayout(const std::vector<vk::DescriptorSetLayout>& layouts,
-                                                               const u32 pushConstantSize) const
+                                                               const u32 pushConstantSize,
+                                                               const vk::ShaderStageFlags pushStages) const
     {
         vk::PipelineLayoutCreateInfo info;
         info.setSetLayouts(layouts);
 
         std::array<vk::PushConstantRange, 1> pushRanges;
         if (pushConstantSize > 0) {
-            // 与 ShaderManager 一致：程序级 vertex|fragment 声明。
+            // 与 ShaderManager 一致：程序级声明（默认 vertex|fragment）。
             // 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)
+            // 全部 stage（VUID-01796），因此 push 调用点也要带同样的 stage
+            pushRanges[0].setStageFlags(pushStages)
                          .setOffset(0)
                          .setSize(pushConstantSize);
             info.setPushConstantRanges(pushRanges);

Engine/src/Vulkan/ShaderManager.cpp

@@ -42,6 +42,7 @@ ShaderObject ShaderManager::CreateFromSPIRV(const std::vector<u32>& spirv,
     if (stage == vk::ShaderStageFlagBits::eVertex) {
         nextStage = vk::ShaderStageFlagBits::eFragment;
     }
+    // compute 单阶段：无 nextStage（VUID-09183 只约束 graphics 链）
 
     vk::ShaderCreateInfoEXT createInfo;
     createInfo.setStage(stage)
@@ -60,8 +61,11 @@ ShaderObject ShaderManager::CreateFromSPIRV(const std::vector<u32>& spirv,
         // 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 |
+        // GLSL 的 push_constant block 决定，声明范围只是能力上限。
+        // compute 无程序链，单独按 eCompute 声明（layout/push 调用点一致）
+        pushRanges[0].setStageFlags(stage == vk::ShaderStageFlagBits::eCompute
+            ? vk::ShaderStageFlagBits::eCompute
+            : vk::ShaderStageFlagBits::eVertex |
               vk::ShaderStageFlagBits::eFragment)
                      .setOffset(0)
                      .setSize(pushConstantSize);

shaders/ibl_brdf_lut.comp

@@ -0,0 +1,54 @@
+#version 450
+
+// IBL ④：BRDF 积分 LUT（2D，纯数学，与环境内容无关）。
+// (NdotV, roughness) → (F 缩放 A, F 偏置 B)：specular = prefiltered·(F·A + B)。
+// 与环境切换无关，但随生成链一起跑（GPU 亚毫秒级），省去跨环境缓存。
+
+#include "ibl_common.glsl"
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(set = 0, binding = 0, rg16f) uniform restrict writeonly image2D outLut;
+
+const uint SAMPLES = 128;
+
+void main()
+{
+    ivec2 gid = ivec2(gl_GlobalInvocationID);
+    if (gid.x >= imageSize(outLut).x || gid.y >= imageSize(outLut).y) return;
+
+    vec2 uv = (vec2(gid) + 0.5) / vec2(imageSize(outLut));
+    float nDotV = uv.x;
+    float roughness = uv.y;
+
+    // V 在以 N 为 +Z 的切线空间里（积分只依赖这两角度）
+    vec3 V = vec3(sqrt(1.0 - nDotV * nDotV), 0.0, nDotV);
+    vec3 Nz = vec3(0.0, 0.0, 1.0);
+
+    float a = 0.0, b = 0.0;
+    float k = roughness * roughness * 0.5;   // IBL 版 Smith k
+
+    for (uint i = 0u; i < SAMPLES; ++i) {
+        vec3 H = iblImportanceSampleGGX(iblHammersley(i, SAMPLES), Nz, roughness);
+        vec3 L = normalize(2.0 * dot(V, H) * H - V);
+
+        float lDotN = L.z;
+        float nDotL = max(lDotN, 0.0);
+        float nDotH = max(H.z, 0.0);
+        float vDotH = max(dot(V, H), 0.0);
+        if (nDotL <= 0.0) continue;
+
+        float g = (nDotV / (nDotV * (1.0 - k) + k)) *
+                  (nDotL / (nDotL * (1.0 - k) + k));
+        // 高 roughness 时部分样本 H.z 恰为 0（g 也为 0），0/0 = NaN
+        // 会累进 LUT——分母钳 epsilon（与 CPU 版一致）
+        float gVis = g * vDotH / max(nDotH * nDotV, 1e-6);
+        float fc = 1.0 - vDotH;
+        float f = fc * fc * fc * fc * fc;   // (1-vDotH)⁵
+
+        a += (1.0 - f) * gVis;
+        b += f * gVis;
+    }
+
+    imageStore(outLut, gid, vec4(a / float(SAMPLES), b / float(SAMPLES), 0.0, 1.0));
+}

shaders/ibl_common.glsl

@@ -0,0 +1,60 @@
+// ibl_common：IBL compute 预积分共享数学（与原 CPU 实现逐式对齐）。
+//
+// 被 4 个 .comp 包含（glslc 相对 include）。改造自 Cubemap.cpp 的
+// CPU 积分路径——公式一致保证视觉无回归，可直接对比验证。
+
+#ifndef IBL_COMMON_GLSL
+#define IBL_COMMON_GLSL
+
+const float IBL_PI = 3.14159265358979;
+
+// ---- GL/Vulkan cube 面方向表（与 GL spec 立方体选择表互逆）----
+// (u,v)∈[0,1]（u 向右 v 向下）→ s,t∈[-1,1] → 单位方向
+vec3 iblFaceDir(uint face, float u, float v)
+{
+    float s = u * 2.0 - 1.0;
+    float t = v * 2.0 - 1.0;
+    vec3 dir;
+    if (face == 0u)      dir = vec3( 1.0, -t, -s);   // +X（面片顶行 = +Y）
+    else if (face == 1u) dir = vec3(-1.0, -t,  s);   // -X
+    else if (face == 2u) dir = vec3( s, 1.0,  t);    // +Y
+    else if (face == 3u) dir = vec3( s,-1.0, -t);    // -Y
+    else if (face == 4u) dir = vec3( s, -t, 1.0);    // +Z
+    else                 dir = vec3(-s, -t,-1.0);    // -Z
+    return normalize(dir);   // equirect 查表要求单位方向
+}
+
+// 单位方向 → 等距柱状 uv。行 0 = 天顶（+Y）：v = acos(y)/π，
+// 与 stbi 行序一致；u 经度环绕交给采样器 S = eRepeat 处理接缝
+vec2 iblDirToEquirectUv(vec3 dir)
+{
+    float u = atan(dir.z, dir.x) / (2.0 * IBL_PI) + 0.5;
+    float v = acos(clamp(dir.y, -1.0, 1.0)) / IBL_PI;
+    return vec2(u, v);
+}
+
+// van der Corput 基底反转。硬件位反转指令替代 CPU 手写交换链；
+// 整数值转浮点（不是位型重解读），序列与 CPU 版一致
+vec2 iblHammersley(uint i, uint n)
+{
+    return vec2(float(i) / float(n),
+                float(bitfieldReverse(i)) * 2.3283064365386963e-10);
+}
+
+// GGX NDF 按概率密度取半程向量 H（切线空间 → 世界）
+vec3 iblImportanceSampleGGX(vec2 xi, vec3 N, float roughness)
+{
+    float a = roughness * roughness;
+    float phi = 2.0 * IBL_PI * xi.x;
+    float cosTheta = sqrt((1.0 - xi.y) / (1.0 + (a * a - 1.0) * xi.y));
+    float sinTheta = sqrt(1.0 - cosTheta * cosTheta);
+
+    vec3 H = vec3(sinTheta * cos(phi), sinTheta * sin(phi), cosTheta);
+
+    vec3 up = abs(N.z) > 0.999 ? vec3(1.0, 0.0, 0.0) : vec3(0.0, 1.0, 0.0);
+    vec3 T = normalize(cross(up, N));
+    vec3 B = cross(N, T);
+    return normalize(T * H.x + B * H.y + N * H.z);
+}
+
+#endif // IBL_COMMON_GLSL

shaders/ibl_equirect_to_cube.comp

@@ -0,0 +1,27 @@
+#version 450
+
+// IBL ①：等距柱状 HDR → 环境 cube（6 面重采样）。
+// 替代原 CPU 逐像素 SampleEquirect 双线性查表——GPU 上直接用硬件
+// 线性过滤采样（经度接缝由 S=eRepeat 处理），观感与 CPU 版一致。
+// 输出为 image2DArray（layer = face），图像本体带 CUBE_COMPATIBLE，
+// 采样端另建 eCube view。
+
+#include "ibl_common.glsl"
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(set = 0, binding = 0) uniform sampler2D equirect;
+layout(set = 0, binding = 1, rgba16f) uniform restrict writeonly image2DArray outCube;
+
+void main()
+{
+    ivec3 gid = ivec3(gl_GlobalInvocationID);
+    ivec2 size = imageSize(outCube).xy;
+    if (gid.x >= size.x || gid.y >= size.y) return;
+
+    vec2 uv = (vec2(gid.xy) + 0.5) / vec2(size);
+    vec3 dir = iblFaceDir(uint(gid.z), uv.x, uv.y);
+    vec3 rgb = texture(equirect, iblDirToEquirectUv(dir)).rgb;
+
+    imageStore(outCube, gid, vec4(rgb, 1.0));
+}

shaders/ibl_irradiance.comp

@@ -0,0 +1,48 @@
+#version 450
+
+// IBL ②：环境 cube → 漫反射辐照度 cube（半球 Lambert 卷积）。
+// E = ∫ L·cosθ·sinθ dθ dφ，切线基 up×N 构造。64×32 = 2048 采样/像素
+// （与 CPU 版同级精度），32² 低频输出上 GPU 耗时可忽略。
+
+#include "ibl_common.glsl"
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(set = 0, binding = 0) uniform samplerCube env;
+layout(set = 0, binding = 1, rgba16f) uniform restrict writeonly image2DArray outCube;
+
+const int PHI_STEPS = 64;
+const int THETA_STEPS = 32;
+
+void main()
+{
+    ivec3 gid = ivec3(gl_GlobalInvocationID);
+    ivec2 size = imageSize(outCube).xy;
+    if (gid.x >= size.x || gid.y >= size.y) return;
+
+    vec2 uv = (vec2(gid.xy) + 0.5) / vec2(size);
+    vec3 N = iblFaceDir(uint(gid.z), uv.x, uv.y);
+
+    vec3 up = abs(N.y) > 0.999 ? vec3(0.0, 0.0, 1.0) : vec3(0.0, 1.0, 0.0);
+    vec3 T = normalize(cross(up, N));
+    vec3 B = cross(N, T);
+
+    float dPhi = 2.0 * IBL_PI / float(PHI_STEPS);
+    float dTheta = 0.5 * IBL_PI / float(THETA_STEPS);
+
+    vec3 irradiance = vec3(0.0);
+    for (int p = 0; p < PHI_STEPS; ++p) {
+        float phi = (float(p) + 0.5) * dPhi;
+        for (int t = 0; t < THETA_STEPS; ++t) {
+            float theta = (float(t) + 0.5) * dTheta;
+            float sinT = sin(theta);
+            float cosT = cos(theta);
+            vec3 sampleDir = T * (sinT * cos(phi)) +
+                             B * (sinT * sin(phi)) +
+                             N * cosT;
+            irradiance += texture(env, sampleDir).rgb * (cosT * sinT);
+        }
+    }
+
+    imageStore(outCube, gid, vec4(irradiance * (dPhi * dTheta), 1.0));
+}

shaders/ibl_prefilter.comp

@@ -0,0 +1,44 @@
+#version 450
+
+// IBL ③：环境 cube → pre-filtered cube（GGX 重要性采样，split-sum 镜面端）。
+// R(N) ≈ Σ L·NdotL / Σ NdotL。每 mip 一次 dispatch（storage view 必须
+// 单 mip），roughness 经 push constant 传入；256 采样统一（高 roughness
+// 分布更宽，取 CPU 版采样数上限保证收敛）。
+
+#include "ibl_common.glsl"
+
+layout(local_size_x = 8, local_size_y = 8) in;
+
+layout(set = 0, binding = 0) uniform samplerCube env;
+layout(set = 0, binding = 1, rgba16f) uniform restrict writeonly image2DArray outCube;
+
+layout(push_constant) uniform Push {
+    float roughness;
+} pc;
+
+const uint SAMPLES = 256;
+
+void main()
+{
+    ivec3 gid = ivec3(gl_GlobalInvocationID);
+    ivec2 size = imageSize(outCube).xy;
+    if (gid.x >= size.x || gid.y >= size.y) return;
+
+    vec2 uv = (vec2(gid.xy) + 0.5) / vec2(size);
+    vec3 N = iblFaceDir(uint(gid.z), uv.x, uv.y);
+    vec3 V = N;   // split-sum 约定：V 沿 N
+
+    vec3 acc = vec3(0.0);
+    float weight = 0.0;
+    for (uint i = 0u; i < SAMPLES; ++i) {
+        vec3 H = iblImportanceSampleGGX(iblHammersley(i, SAMPLES), N, pc.roughness);
+        vec3 L = normalize(2.0 * dot(V, H) * H - V);
+        float nDotL = max(dot(N, L), 0.0);
+        if (nDotL > 0.0) {
+            acc += texture(env, L).rgb * nDotL;
+            weight += nDotL;
+        }
+    }
+
+    imageStore(outCube, gid, vec4(weight > 0.0 ? acc / weight : vec3(0.0), 1.0));
+}

shaders/simple_shader.frag

@@ -57,6 +57,10 @@ layout(set = 2, binding = 5) uniform sampler2D brdfLut;           // BRDF 积分
 // 等效 36 次比较）→ 软阴影边缘；采样器 clamp-to-border 白色：
 // 光视锥外自动 1.0（无阴影），省出界分支
 float ShadowFactor(vec3 N, float viewDepth) {
+    // 阴影总开关（ubo.cascadeSplits.w 由 CPU 每帧写入）：关闭即全亮，
+    // 配合 ShadowPass 空转（深度图无内容也读不到这里）
+    if (ubo.cascadeSplits.w < 0.5) return 1.0;
+
     int cascade = 2;
     for (int i = 0; i < 3; ++i) {
         if (viewDepth < ubo.cascadeSplits[i]) { cascade = i; break; }
@@ -142,7 +146,7 @@ void main() {
     vec3 F0 = mix(vec3(0.04), albedo, metallic);   // 电介质基准反射率 4%
 
     vec3 Lo = vec3(0.0);
-    float shade0 = 1.0f;   // 【诊断】阴影系数可视化用
+    float shade0 = 1.0f;   // 第 0 盏光的阴影因子（环境光同步衰减用）
 
     for (int i = 0; i < MAX_LIGHTS; ++i) {
         // 按类型分支：方向光无衰减；点光方向指向光源 + 平方距离衰减
@@ -213,8 +217,6 @@ void main() {
     vec3 ambient = (diffuse + specular) * ubo.ambientColor.rgb
                    * mix(0.4, 1.0, shade0);   // 阴影区环境同步衰减（简化遮蔽）
 
-    vec3 color = ambient + Lo + emissive;
-
-    outColor = vec4(color, 1.0);
+    outColor = vec4(ambient + Lo + emissive, 1.0);
     outVelocity = fragVelocity;
 }

shaders/tonemap.frag

@@ -1,47 +1,123 @@
 #version 450
 
-// Tonemap pass：HDR 线性颜色 → 显示空间。
+// Tonemap pass：HDR（+bloom）→ 显示空间，兼调试视图路由上屏。
 // 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。
+// 退回纯 gamma（旧观感）。
+//
+// 调试视图：pc.debugView 整型路由到已绑定的中间目标（绑定超集静态成立，
+// 运行时零重绑）。HDR 域资源过同一套 exposure/tonemap 保证观感可比；
+// LDR 类（velocity/深度）直接可视化。
 
 layout(location = 0) in vec2 fragUV;
 layout(location = 0) out vec4 outColor;
 
-layout(set = 0, binding = 0) uniform sampler2D hdrImage;
+layout(set = 0, binding = 0) uniform sampler2D taaImage;         // TAA 后 HDR（主输入）
 layout(set = 0, binding = 1) uniform sampler2D bloomImage;
-layout(set = 0, binding = 2) uniform sampler2D temporalDiffImage;   // 帧历史差异（调试视图）
+layout(set = 0, binding = 2) uniform sampler2D temporalDiffImage;
+layout(set = 0, binding = 3) uniform sampler2D rawHdrImage;      // Forward 直出（TAA 前）
+layout(set = 0, binding = 4) uniform sampler2D velocityImage;
+layout(set = 0, binding = 5) uniform sampler2D shadowMap0Image;  // D32 深度（r 分量）
+layout(set = 0, binding = 6) uniform sampler2D shadowMap1Image;
+layout(set = 0, binding = 7) uniform sampler2D shadowMap2Image;
+// 环境调试视图源（随环境切换重绑，静态绑定超集）
+layout(set = 0, binding = 8) uniform samplerCube envCubeImage;
+layout(set = 0, binding = 9) uniform samplerCube irradianceImage;
+layout(set = 0, binding = 10) uniform samplerCube prefilterImage;
+layout(set = 0, binding = 11) uniform sampler2D brdfLutImage;
+
+// 与 Passes::DebugView 枚举同步（FullscreenGpu.h）
+const int VIEW_FINAL        = 0;
+const int VIEW_TEMPORALDIFF = 1;
+const int VIEW_RAW_HDR      = 2;
+const int VIEW_TAA          = 3;
+const int VIEW_BLOOM        = 4;
+const int VIEW_VELOCITY     = 5;
+const int VIEW_SHADOW0      = 6;
+const int VIEW_SHADOW1      = 7;
+const int VIEW_SHADOW2      = 8;
+const int VIEW_ENV_CUBE     = 9;
+const int VIEW_IRRADIANCE   = 10;
+const int VIEW_PREFILTER    = 11;
+const int VIEW_BRDF_LUT     = 12;
 
 // 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 debugView;     // >0.5 显示帧历史差异视图（TemporalDiff）
+    float debugView;     // 调试视图枚举（整型经 float 传）
 } 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);
 }
 
+vec3 tonemap(vec3 hdr) {
+    hdr *= pc.exposure;
+    vec3 mapped = (pc.aces > 0.5) ? acesFilm(hdr) : clamp(hdr, 0.0, 1.0);
+    return pow(mapped, vec3(1.0 / 2.2));
+}
+
+// 深度可视化：反相（近=白，与调试截图风格一致）
+vec3 depthView(sampler2D depthImage) {
+    float d = texture(depthImage, fragUV).r;
+    return vec3(pow(1.0 - d, 2.2));
+}
+
+// 全屏 uv → 球面方向（cube 调试视图的经纬展开：横向经度、纵向极角）
+vec3 latLongDir(vec2 uv) {
+    float phi = (uv.x * 2.0 - 1.0) * 3.14159265358979;     // -π..π 经度
+    float theta = (1.0 - uv.y) * 3.14159265358979;         // 0..π 极角（顶=天顶）
+    return vec3(sin(theta) * cos(phi), cos(theta), sin(theta) * sin(phi));
+}
+
+// cube 视图统一走 tonemap（HDR 域，观感与场景一致）；
+// prefilter 纵向映射 mip 链（上锐下糙，直观验证 roughness 分级）
+vec3 cubeView(samplerCube cube, float lod) {
+    return tonemap(textureLod(cube, latLongDir(fragUV), lod).rgb);
+}
+
 void main() {
-    // 调试视图：帧历史差异（静止=黑，运动边缘发光）
-    if (pc.debugView > 0.5) {
+    int view = int(pc.debugView + 0.5);
+
+    // ---- 调试视图路由 ----
+    if (view != VIEW_FINAL) {
+        switch (view) {
+        case VIEW_TEMPORALDIFF: {
+            // 帧历史差异：静止=黑，运动边缘发光
             vec3 d = clamp(texture(temporalDiffImage, fragUV).rgb, 0.0, 1.0);
             outColor = vec4(pow(d, vec3(1.0 / 2.2)), 1.0);
             return;
         }
-
-    // bloom 在 HDR 域叠加（半分辨率模糊图线性采样上采样）
-    vec4 hdrTex = texture(hdrImage, fragUV);
-    vec3 hdr = hdrTex.rgb + texture(bloomImage, fragUV).rgb * pc.bloomStrength;
-    float alpha = hdrTex.a;
-
-    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)) * alpha, alpha);
+        case VIEW_RAW_HDR:   outColor = vec4(tonemap(texture(rawHdrImage, fragUV).rgb), 1.0); return;
+        case VIEW_TAA:       outColor = vec4(tonemap(texture(taaImage, fragUV).rgb), 1.0); return;
+        case VIEW_BLOOM:     outColor = vec4(tonemap(texture(bloomImage, fragUV).rgb * 4.0), 1.0); return;
+        case VIEW_VELOCITY: {
+            // |velocity| × 增益：静止=黑，运动越快越亮
+            vec2 vel = texture(velocityImage, fragUV).xy;
+            float mag = clamp(length(vel) * 8.0, 0.0, 1.0);
+            outColor = vec4(vec3(mag), 1.0);
+            return;
+        }
+        case VIEW_SHADOW0:   outColor = vec4(depthView(shadowMap0Image), 1.0); return;
+        case VIEW_SHADOW1:   outColor = vec4(depthView(shadowMap1Image), 1.0); return;
+        case VIEW_SHADOW2:   outColor = vec4(depthView(shadowMap2Image), 1.0); return;
+        case VIEW_ENV_CUBE:    outColor = vec4(cubeView(envCubeImage, 0.0), 1.0); return;
+        case VIEW_IRRADIANCE:  outColor = vec4(cubeView(irradianceImage, 0.0), 1.0); return;
+        case VIEW_PREFILTER:   outColor = vec4(cubeView(prefilterImage, fragUV.y * 4.0), 1.0); return;
+        case VIEW_BRDF_LUT: {
+            // LUT 是 (F 缩放, F 偏置) 双通道，R/G 直读可视化
+            vec2 ab = texture(brdfLutImage, fragUV).rg;
+            outColor = vec4(ab, 0.0, 1.0);
+            return;
+        }
+        }
+    }
+
+    // ---- 正常路径：TAA 后 HDR + bloom 叠加 → tonemap ----
+    vec4 hdrTex = texture(taaImage, fragUV);
+    vec3 hdr = hdrTex.rgb + texture(bloomImage, fragUV).rgb * pc.bloomStrength;
+
+    outColor = vec4(tonemap(hdr) * hdrTex.a, hdrTex.a);
 }