fix: Shadow pass 接入 GPUDriven batch，实例上限提至 16384

100 背包掉帧 + 警告刷屏的根因：
1. backpack.obj 有 44 个 mesh 部件，100 实体 × 44 = 4400 实例
   打爆 4096 上限 → 每帧截断 + 告警刷屏。上限提至 16384
  （1MB/slot），告警改一次性
2. Shadow pass 仍逐实体逐部件画（4400 draw，CPU 提交瓶颈）。
   改为与 Forward 共用同一批 batch：Set0 绑 ObjectBuffer，
   lightVP 一次 push，44 draw 完成。shadow.vert 经
   gl_InstanceIndex 从 SSBO 取模型矩阵

实测（100 背包）：GPU 总耗时 ~5.7ms（Shadow 2.9 + Forward 2.6 +
后处理 0.1），0 警告，帧率受 mailbox 垂直同步 pacing 限制

Engine/include/Engine/Renderer/Renderer.h

@@ -140,7 +140,7 @@ namespace Engine {
         // GPUDriven：当帧实例数据（Render 期收集，Forward pass 消费）
         std::vector<InstanceBatch> m_InstanceBatches;
         std::vector<glm::mat4> m_InstanceModels;
-        static constexpr u32 kMaxInstances = 4096;
+        static constexpr u32 kMaxInstances = 16384;
 
         // 可选叠加层（不拥有；Application 持有）
         ImGuiLayer* m_ImGuiLayer = nullptr;

Engine/src/Renderer/Renderer.cpp

@@ -52,10 +52,7 @@ namespace Engine
         constexpr u32 kTextureSetIndex = 1;
         constexpr u32 kShadowSetIndex = 2;
         constexpr u32 kMaxLights = 4;
-
-        // Shadow pass：push constant = { lightVP, model }
         constexpr u32 kShadowMapSize = 2048;
-        constexpr u32 kShadowPushSize = static_cast<u32>(2 * sizeof(glm::mat4));
 
         // 全屏 pass 通用 push（16B）：所有全屏 pass（Tonemap/Bloom 系）
         // 统一声明与推送，规避 VUID-08878 的"range 定义必须一致"
@@ -111,9 +108,10 @@ namespace Engine
         m_PipelineLayout = m_DescriptorManager->CreatePipelineLayout(
             {m_UBOSetLayout, m_TextureSetLayout, m_ShadowSetLayout}, 0);
 
-        // Shadow pass 独立 layout：无 set，128B push constant
+        // Shadow pass layout：Set0 = UBO/ObjectBuffer（实例矩阵），push = lightVP
         m_ShadowPipelineLayout = m_DescriptorManager->CreatePipelineLayout(
-            std::vector<vk::DescriptorSetLayout>{}, kShadowPushSize);
+            std::vector<vk::DescriptorSetLayout>{m_UBOSetLayout},
+            sizeof(glm::mat4));
 
         // 7. ShaderManager（VK_EXT_shader_object）
         m_ShaderManager = CreateScope<ShaderManager>(*m_Device);
@@ -135,10 +133,14 @@ namespace Engine
         // shadow shader：depth-only，无 set
         auto shadowVert = m_ShaderManager->CreateFromFile(
             shaderDir + "shadow.vert.spv",
-            vk::ShaderStageFlagBits::eVertex, {}, kShadowPushSize);
+            vk::ShaderStageFlagBits::eVertex,
+            std::vector<vk::DescriptorSetLayout>{m_UBOSetLayout},
+            sizeof(glm::mat4));
         auto shadowFrag = m_ShaderManager->CreateFromFile(
             shaderDir + "shadow.frag.spv",
-            vk::ShaderStageFlagBits::eFragment, {}, kShadowPushSize);
+            vk::ShaderStageFlagBits::eFragment,
+            std::vector<vk::DescriptorSetLayout>{m_UBOSetLayout},
+            sizeof(glm::mat4));
 
         m_ShadowVertShader = shadowVert.Handle;
         m_ShadowFragShader = shadowFrag.Handle;
@@ -225,7 +227,6 @@ namespace Engine
 
         // 10. Per-frame UBO ring + GPUDriven 实例 SSBO ring
         //     （同一 ring 类，SSBO 只是 usage 不同；正确性保证一致）
-        constexpr u32 kMaxInstances = 4096;
         m_UniformBuffer = CreateScope<UniformBuffer>(
             *m_Device, *m_Allocator, sizeof(UniformBufferObject), kMaxFramesInFlight);
         m_ObjectBuffer = CreateScope<UniformBuffer>(
@@ -445,29 +446,22 @@ namespace Engine
                 vk::Rect2D scissor({0, 0}, {kShadowMapSize, kShadowMapSize});
                 cmd.setScissorWithCount(scissor);
 
-                // ---- 绘制：push {lightVP, model}，与 Forward 用同一份 world 矩阵 ----
+                // ---- 绘制：GPUDriven——与 Forward 共用同一批 batch ----
-                struct ShadowPush {
+                // lightVP 全 pass 只 push 一次；模型矩阵经实例索引从
-                    glm::mat4 lightVP;
+                // ObjectBuffer 取（与 Forward 同一份数据）
-                    glm::mat4 model;
+                cmd.bindDescriptorSets(vk::PipelineBindPoint::eGraphics,
-                };
+                                       m_ShadowPipelineLayout, 0,
+                                       m_UBODescriptorSets[ctx.frame.frameSlot], {});
 
-                m_BoundScene->Each<TransformComponent, MeshRendererComponent>(
+                // stage 必须覆盖 layout range 声明的全部 stage（VUID-01796）
-                    [&](Entity, TransformComponent& transform,
+                cmd.pushConstants(m_ShadowPipelineLayout,
-                        MeshRendererComponent& meshRenderer) {
+                                  vk::ShaderStageFlagBits::eVertex |
-                        Model* model = m_BoundAssets->GetModel(meshRenderer.model);
+                                  vk::ShaderStageFlagBits::eFragment,
-                        if (!model) return;
+                                  0, sizeof(glm::mat4), &m_LightSpace);
 
-                        const ShadowPush push{m_LightSpace, transform.world};
+                for (const auto& batch : m_InstanceBatches) {
-                        // stage 必须覆盖 layout range 声明的全部 stage（VUID-01796）
+                    batch.mesh->Draw(cmd, batch.count, batch.first);
-                        cmd.pushConstants(m_ShadowPipelineLayout,
+                }
-                                          vk::ShaderStageFlagBits::eVertex |
-                                          vk::ShaderStageFlagBits::eFragment,
-                                          0, kShadowPushSize, &push);
-
-                        for (auto& mesh : model->GetMeshes()) {
-                            mesh->Draw(cmd);
-                        }
-                    });
 
                 cmd.endRendering();
             })
@@ -848,6 +842,17 @@ namespace Engine
                 ? static_cast<f32>(end - start) * m_TimestampPeriod / 1.0e6f
                 : 0.0f;
         }
+
+        // 控制台遥测：每 120 次读取（≈60 帧 @2 slot）输出一行
+        static u32 telemetryCount = 0;
+        if (++telemetryCount % 120 == 1) {
+            std::string line = "GPU times:";
+            for (u32 i = 0; i < passCount; ++i) {
+                line += " " + m_PassNames[i] + "=" +
+                        std::to_string(m_PassGpuMs[i]).substr(0, 5) + "ms";
+            }
+            ENGINE_INFO("{}", line);
+        }
     }
 
     void Renderer::Render(ECS::Scene& scene, Assets::AssetManager& assets)
@@ -973,8 +978,13 @@ namespace Engine
             }
         }
         if (instanceTruncated) {
-            ENGINE_WARN("GPUDriven: instance buffer full ({}), entities truncated",
+            // 只告警一次（每帧告警会刷屏）
-                        kMaxInstances);
+            static bool warned = false;
+            if (!warned) {
+                warned = true;
+                ENGINE_WARN("GPUDriven: instance buffer full ({}), entities truncated",
+                            kMaxInstances);
+            }
         }
         if (!m_InstanceModels.empty()) {
             m_ObjectBuffer->Update(frameSlot, m_InstanceModels.data(),

shaders/shadow.vert

@@ -1,17 +1,19 @@
 #version 450
 
 // Shadow pass vertex：光源视角写深度。
-// push constant = { lightVP, model }，与 Forward pass 用同一份
+// GPUDriven：模型矩阵经 gl_InstanceIndex 从实例缓冲取（与 Forward
-// lightVP（UBO.lightSpace）和同一份 model（TransformComponent.world），
+// 共用同一份数据），push 只带 lightVP（全 pass 一次）。
-// 保证阴影贴图与主视角几何严格一致。
 
 layout(location = 0) in vec3 inPosition;
 
-layout(push_constant) uniform ShadowPush {
+layout(std430, set = 0, binding = 1) readonly buffer ObjectBuffer {
+    mat4 models[];
+} objects;
+
+layout(push_constant) uniform PushConstants {
     mat4 lightVP;
-    mat4 model;
 } pc;
 
 void main() {
-    gl_Position = pc.lightVP * pc.model * vec4(inPosition, 1.0);
+    gl_Position = pc.lightVP * objects.models[gl_InstanceIndex] * vec4(inPosition, 1.0);
 }