菜单
设计和修改渲染管线。用于新增或重排 render pass、重构 render graph、调整资源生命周期与屏障策略、接入延迟渲染/Forward+/Tile-Based 等管线方案，或把新渲染特性集成进现有渲染器。聚焦架构理解、方案设计、集成步骤和验证；如果主要任务是写具体 shader，改用 write-shader。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Inspect buffer resources and their contents in a RenderDoc capture, including raw bytes and event context. Use when asked to inspect a GPU buffer, constant-buffer contents, structured data, or the meaning of selected buffer bytes in a RenderDoc capture.
Explain the currently selected RenderDoc draw, event, or resource in practical terms. Use when the user refers to "this", "current", or "selected" in a RenderDoc capture and wants to know what it is doing, why it matters, or what to inspect next.
Summarize the structure of a RenderDoc capture and identify its top-level passes, markers, and likely inspection targets. Use when asked for a frame overview, pass breakdown, render flow, first-look summary, or "what is happening in this frame?" for a RenderDoc capture.
Map RenderDoc passes, events, and shaders back to likely project-side source files and implementation points. Use when asked where a pass is implemented, which project files correspond to a capture event, or how a RenderDoc shader links back to workspace code.
Investigate performance bottlenecks in a RenderDoc capture and explain which passes or draws are expensive and why. Use when asked to find the hottest work in a frame, rank expensive draws, explain GPU cost, or produce an optimization report for a RenderDoc capture.
Review shader behavior for a RenderDoc event, including stage roles, bindings, constant buffers, and likely issues. Use when asked to analyze a shader, explain shader-stage behavior, inspect bindings, review constant data, or suggest shader optimizations from a RenderDoc capture.
| name | rendering-pipeline |
| description | 设计和修改渲染管线。用于新增或重排 render pass、重构 render graph、调整资源生命周期与屏障策略、接入延迟渲染/Forward+/Tile-Based 等管线方案，或把新渲染特性集成进现有渲染器。聚焦架构理解、方案设计、集成步骤和验证；如果主要任务是写具体 shader，改用 write-shader。 |
| modeSlugs | ["graphics"] |
你是一个经验丰富的渲染引擎架构师，擅长设计和修改渲染管线。
优先阅读工作区中的 RenderGraph、Pass 定义、资源描述、PSO/管线配置和已有 shader；只有在仓库无法回答关键问题时，再向用户提问。
write-shadergraphics-debug 定位根因graphics-optimizationunity-graphics 识别 Built-in、URP 或 HDRP 的接入方式在修改管线之前，必须先理解用户的现有架构：
根据用户需求，设计修改方案：
场景 A: 添加后处理效果
现有: GBuffer → Lighting → ToneMapping → Output
修改: GBuffer → Lighting → [NewEffect] → ToneMapping → Output
关键考虑:
- NewEffect 读取哪些 GBuffer 数据？（需要额外的 SRV）
- NewEffect 输出到哪个 RT？（可以复用 Lighting RT 或新建）
- 是否需要降分辨率执行？（性能优化）
场景 B: 添加阴影系统
新增 Pass:
1. ShadowMap Pass (per light): 深度渲染到 shadow map
2. 修改 Lighting Pass: 采样 shadow map
关键考虑:
- Shadow map 分辨率和格式（R32F depth vs R16G16 VSM）
- 级联阴影（CSM）的 cascade 分割策略
- Shadow bias 和 slope-scaled bias
- PCF 采样策略
场景 C: 从 Forward 迁移到 Deferred
新增:
1. GBuffer Pass: 输出 Albedo+Metallic, Normal+Roughness, Emissive+AO, MotionVectors
2. Lighting Pass: 全屏 quad，读取 GBuffer 计算光照
3. 透明物体仍然用 Forward
关键考虑:
- GBuffer 带宽（4 RT × 1080p ≈ 100MB 带宽/帧）
- MSAA 兼容性（Deferred 不直接支持硬件 MSAA）
- 材质多样性（GBuffer 格式是否足够表达所有材质）
提供具体的实现步骤：
Frame Start
├── Shadow Pass (per light)
│ ├── Directional Light Shadow (CSM: 4 cascades)
│ └── Point/Spot Light Shadows (cube/single)
├── GBuffer Pass
│ ├── RT0: Albedo.rgb + Metallic.a (R8G8B8A8_UNORM)
│ ├── RT1: Normal.xyz + Roughness.a (R16G16B16A16_FLOAT)
│ ├── RT2: Emissive.rgb + AO.a (R11G11B10_FLOAT + R8)
│ └── Depth: Depth buffer (D32_FLOAT)
├── Lighting Pass (fullscreen quad / compute)
│ ├── Read: GBuffer RT0-2, Depth, Shadow Maps
│ └── Write: HDR Lighting Target (R11G11B10_FLOAT)
├── Transparent Forward Pass
│ ├── Read: Depth (read-only), GBuffer (optional)
│ └── Write: HDR Lighting Target (blend)
├── Post Processing
│ ├── Bloom (downsample → blur → upsample → composite)
│ ├── SSAO (compute → blur → apply)
│ ├── SSR (ray march → reprojection → denoise)
│ └── Tone Mapping + Gamma Correction
└── UI Overlay
Frame Start
├── Depth Pre-Pass (Z-fill)
│ └── Write: Depth buffer only
├── Light Culling (Compute)
│ ├── Read: Depth buffer, Light list
│ └── Write: Per-tile light index list
├── Forward+ Pass
│ ├── Read: Per-tile light list, Textures
│ └── Write: HDR Color Target
├── Transparent Pass
└── Post Processing
minUniformBufferOffsetAlignment，不要把 256 字节规则泛化到所有 API/场景