feat: IBL 预积分 GPU compute 化 + 环境资产入 Assets 层（动态切换）

CPU 逐像素卷积（venice 4k 阻塞 7.36s）→ 4 个 compute shader 一次性
提交：equirect→cube / irradiance / prefilter×5mip / BRDF LUT。
Cubemap 瘦身为纯资源持有类，生成职责移交 IblGenerator；
环境经 AssetManager 缓存（EnvRef），Renderer::SetEnvironment
动态换绑（RenderPass::OnEnvironmentChanged 广播）。
调试视图 +4：EnvCube / Irradiance / Prefilter / BRDF LUT。

性能前后对比（RTX 5060 Laptop，venice_dawn_1_4k.hdr）：
- 改造前 CPU：7.36 s（主线程阻塞）
- 改造后 GPU：HDR decode 127-134ms + 资源 22-35ms + compute 17-19ms
  ≈ 165-190ms（约 40×）；Neutral 兜底 ~6ms
- 运行中动态切换环境同量级（ms 级 GPU，decode 占大头）

冒烟：两构建目录 0 警告；运行期 0 validation（含两次环境切换 +
调试视图 9-12 轮播自动化验证）。

Engine/include/Engine/Assets/AssetManager.h

@@ -145,6 +145,11 @@ namespace Engine {
             // 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);
+            Cubemap* GetEnvironment(EnvRef h) const { return m_Environments.Get(h); }
+
             AssetCache<Model>& Models() { return m_Models; }
 
             // 后续扩展点：TextureHandle / MaterialHandle / ShaderHandle
@@ -154,6 +159,7 @@ namespace Engine {
             IAssetUploader& m_Uploader;
 
             AssetCache<Model> m_Models;
+            AssetCache<Cubemap> m_Environments;
         };
 
     } // namespace Assets

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 {
 
@@ -31,6 +32,9 @@ namespace Engine::Assets {
 
         // 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;
     };
 
 } // 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/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
@@ -19,7 +19,6 @@
 
 #include <Engine/Core/Types.h>
 #include <Engine/Vulkan/MemoryAllocator.h>
-#include <glm/glm.hpp>
 #include <vulkan/vulkan.hpp>
 
 namespace Engine {
@@ -37,18 +36,6 @@ 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 +54,8 @@ 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 +1,101 @@
+//
+// 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/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;
+
+        // 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);
+
+        // 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/Passes/FullscreenGpu.h

@@ -52,6 +52,10 @@ namespace Engine {
             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
         };
 

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/TonemapPass.h

@@ -15,39 +15,50 @@
 #include <vulkan/vulkan.hpp>
 #include <vector>
 
-namespace Engine::Passes {
+namespace Engine {
 
-    class TonemapPass final : public RenderPass {
-    public:
-        TonemapPass(FullscreenGpu& gpu, PostProcessParams& params,
-                    FrameSceneData& frame,
-                    RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
-                    RG::TextureHandle temporalDiff, RG::TextureHandle backbuffer,
-                    RG::TextureHandle rawHdr, RG::TextureHandle velocity,
-                    RG::TextureHandle shadowMap0, RG::TextureHandle shadowMap1,
-                    RG::TextureHandle shadowMap2);
-        ~TonemapPass() override = default;
+    class Cubemap;
+    class DescriptorManager;
 
-        void Register(RG::RenderGraph& rg) override;
-        void OnTargetsRebuilt(RG::RenderGraph& rg) override;
+    namespace Passes {
 
-    private:
-        FullscreenGpu& m_Gpu;
-        PostProcessParams& m_Params;
-        FrameSceneData& m_Frame;   // debugView / bloomEnabled（帧契约）
-        RG::TextureHandle m_HDR;
-        RG::TextureHandle m_A;
-        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];
+        class TonemapPass final : public RenderPass {
+        public:
+            TonemapPass(FullscreenGpu& gpu, DescriptorManager& descriptors,
+                        PostProcessParams& params,
+                        FrameSceneData& frame,
+                        RG::TextureHandle hdrColor, RG::TextureHandle bloomA,
+                        RG::TextureHandle temporalDiff, RG::TextureHandle backbuffer,
+                        RG::TextureHandle rawHdr, RG::TextureHandle velocity,
+                        RG::TextureHandle shadowMap0, RG::TextureHandle shadowMap1,
+                        RG::TextureHandle shadowMap2);
+            ~TonemapPass() override = default;
 
-        std::vector<vk::DescriptorSet> m_Sets;   // per-slot
-    };
+            void Register(RG::RenderGraph& rg) override;
+            void OnTargetsRebuilt(RG::RenderGraph& rg) override;
+            void OnEnvironmentChanged(const Cubemap* env) override;
 
-} // namespace Engine::Passes
+        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;
+            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 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; }
@@ -95,6 +96,12 @@ namespace Engine {
         // 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; }
         const std::vector<f32>& GetPassGpuMs() const { return m_PassGpuMs; }
@@ -124,7 +131,9 @@ 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;

Engine/include/Engine/Renderer/VulkanAssetUploader.h

@@ -19,23 +19,28 @@ 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> CreatePlane(f32 halfSize, u8 r, u8 g, u8 b) override;
+        Scope<Cubemap> LoadEnvironment(const std::string& path) 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

@@ -30,4 +30,11 @@ 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); });
+    }
+
 } // namespace Engine::Assets

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 +1,527 @@
+//
+// 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);
+        }
+    }
+
+    Scope<Cubemap> IblGenerator::FromHdrEquirect(const std::string& path,
+                                                 u32 faceSize)
+    {
+        const auto t0 = std::chrono::steady_clock::now();
+
+        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 nullptr;
+        }
+
+        const f32 decodeMs = MsSince(t0);
+        auto cubemap = Generate(pixels, static_cast<u32>(w),
+                                static_cast<u32>(h), faceSize, path);
+        stbi_image_free(pixels);
+
+        ENGINE_INFO("IblGenerator: HDR decode {:.2f} ms",
+                    decodeMs);
+        return cubemap;
+    }
+
+    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/Passes/FullscreenGpu.cpp

@@ -21,7 +21,8 @@ namespace Engine::Passes
         // 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;
-        for (u32 b = 0; b < 8; ++b) {
+        // 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});

Engine/src/Renderer/Passes/SkyboxPass.cpp

@@ -74,6 +74,20 @@ 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) {

Engine/src/Renderer/Passes/TonemapPass.cpp

@@ -3,11 +3,14 @@
 //
 
 #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,
@@ -17,7 +20,7 @@ namespace Engine::Passes
                              RG::TextureHandle shadowMap0,
                              RG::TextureHandle shadowMap1,
                              RG::TextureHandle shadowMap2)
-        : m_Gpu(gpu), m_Params(params), m_Frame(frame),
+        : 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_RawHDR(rawHdr), m_Velocity(velocity),
@@ -27,6 +30,8 @@ namespace Engine::Passes
         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)
@@ -80,4 +85,27 @@ namespace Engine::Passes
         }
     }
 
+    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());
+        }
+    }
+
 } // namespace Engine::Passes

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>
@@ -40,7 +41,6 @@
 #include <glm/glm.hpp>
 
 #include <array>
-#include <filesystem>
 #include <limits>
 
 namespace Engine
@@ -74,17 +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);
@@ -101,7 +96,7 @@ 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
         m_PostParams = CreateScope<Passes::PostProcessParams>();
@@ -111,8 +106,12 @@ namespace Engine
         for (auto& pass : m_Passes) {
             pass->OnTargetsRebuilt(*m_RenderGraph);
         }
-        // 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);
+        // 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;
@@ -185,13 +184,28 @@ namespace Engine
     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) == 9,
+        static_assert(int(Passes::kDebugViewCount) == 13,
                       "DebugView 枚举与名表同步");
         return kNames;
     }
@@ -285,9 +299,9 @@ namespace Engine
             *m_FullscreenGpu, *m_PostParams, frame, m_TAAColor, m_BloomA,
             m_BloomB));
         m_Passes.push_back(CreateScope<Passes::TonemapPass>(
-            *m_FullscreenGpu, *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_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));

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 {
 
@@ -20,4 +21,10 @@ namespace Engine {
                                   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);
+    }
+
 } // 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)
@@ -60,9 +61,12 @@ 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 |
-                                    vk::ShaderStageFlagBits::eFragment)
+        // 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);
         createInfo.setPushConstantRanges(pushRanges);