Nanite on MrOptimisthttps://mr0ptimist.github.io/tags/nanite/Recent content in Nanite on MrOptimistHugozh-cnTue, 21 Apr 2026 00:00:00 +0000UE5 Nanite 与传统渲染管线的深度源码对比https://mr0ptimist.github.io/posts/ue5-nanite-%E4%B8%8E%E4%BC%A0%E7%BB%9F%E6%B8%B2%E6%9F%93%E7%AE%A1%E7%BA%BF%E7%9A%84%E6%B7%B1%E5%BA%A6%E6%BA%90%E7%A0%81%E5%AF%B9%E6%AF%94/Tue, 21 Apr 2026 00:00:00 +0000https://mr0ptimist.github.io/posts/ue5-nanite-%E4%B8%8E%E4%BC%A0%E7%BB%9F%E6%B8%B2%E6%9F%93%E7%AE%A1%E7%BA%BF%E7%9A%84%E6%B7%B1%E5%BA%A6%E6%BA%90%E7%A0%81%E5%AF%B9%E6%AF%94/<blockquote> <p>本文基于 UE 5.4/5.5 引擎源码，从源码层面深入对比 Nanite 渲染管线与传统网格渲染管线的差异。所有代码引用均标注了引擎内的原始路径。</p> </blockquote> <hr> <h2 id="1-宏观架构概览">1. 宏观架构概览</h2> <p>传统渲染管线和 Nanite 管线的最根本区别在于：<strong>几何处理的主导权从 CPU 转移到了 GPU</strong>，并且<strong>材质着色从 Pixel Shader 迁移到了 Compute Shader</strong>。</p> <table> <thead> <tr> <th>维度</th> <th>传统渲染 (Traditional)</th> <th>Nanite</th> </tr> </thead> <tbody> <tr> <td><strong>几何裁剪</strong></td> <td>CPU-driven Frustum/Occlusion Culling</td> <td>GPU-driven Cluster Culling + Two-Pass Occlusion</td> </tr> <tr> <td><strong>LOD</strong></td> <td>离散 LOD (StaticMesh LOD0~N)</td> <td>连续 LOD (Cluster Hierarchy, Runtime Streaming)</td> </tr> <tr> <td><strong>光栅化</strong></td> <td>硬件光栅化 (Fixed Function RS)</td> <td>软件光栅化 (Compute) + 硬件光栅化 (Mesh/Prim Shader)</td> </tr> <tr> <td><strong>中间表示</strong></td> <td>无 (直接写 GBuffer/FrameBuffer)</td> <td>Visibility Buffer (VisBuffer64)</td> </tr> <tr> <td><strong>材质着色</strong></td> <td>Pixel Shader (<code>BasePassPixelShader.usf</code>)</td> <td>Compute Shader (<code>ComputeShaderOutputCommon.ush</code>)</td> </tr> <tr> <td><strong>GBuffer 输出</strong></td> <td><code>SV_Target</code> MRT</td> <td>UAV (<code>ComputeShadingOutputs.OutTargetN</code>)</td> </tr> <tr> <td><strong>DrawCall</strong></td> <td><code>FMeshDrawCommand</code> (CPU 组装)</td> <td>Indirect Dispatch (GPU 驱动)</td> </tr> </tbody> </table> <hr> <h2 id="2-渲染入口与调度">2. 渲染入口与调度</h2> <h3 id="21-传统渲染的入口">2.1 传统渲染的入口</h3> <p>传统渲染的顶层调度在 <code>FDeferredShadingSceneRenderer::Render()</code> 中，通过 <code>RenderBasePass()</code> 等函数发起。每个 <code>FPrimitiveSceneProxy</code> 会在 <code>FMeshPassProcessor</code> 中被转换为 <code>FMeshDrawCommand</code>，最终由 <code>FParallelMeshDrawCommandPass</code> 提交到 RHI。</p>Nanite VisBuffer 核心概念速查https://mr0ptimist.github.io/posts/nanite_visbuffer%E6%A0%B8%E5%BF%83%E6%A6%82%E5%BF%B5%E9%80%9F%E6%9F%A5/Mon, 20 Apr 2026 14:00:00 +0800https://mr0ptimist.github.io/posts/nanite_visbuffer%E6%A0%B8%E5%BF%83%E6%A6%82%E5%BF%B5%E9%80%9F%E6%9F%A5/<h1 id="nanite-visbuffer-核心概念速查">Nanite VisBuffer 核心概念速查</h1> <p>Nanite 是 UE5 的虚拟化几何系统，其核心创新是用 <strong>Visibility Buffer（可见性缓冲）</strong> 替代传统 G-Buffer，将几何处理与材质着色完全解耦。本文梳理 VisBuffer 管线中涉及的关键概念。</p> <h2 id="visbuffer-整体管线">VisBuffer 整体管线</h2> <pre tabindex="0"><code>Mesh → Instance → Cluster Group → Cluster → Triangle → VisBuffer → Material Pass </code></pre><ol> <li><strong>Culling Pass（计算着色器）</strong>：GPU 端逐 Cluster 做视锥/遮挡/屏幕尺寸剔除</li> <li><strong>Rasterization Pass</strong>：将可见像素写入 VisBuffer（仅存 ID，不做材质计算）</li> <li><strong>Material Pass</strong>：全屏 Pass 读取 VisBuffer，解码 ID，仅对可见像素着色一次</li> </ol> <blockquote> <p>来源：Brian Karis, &ldquo;A Deep Dive into Nanite Virtualized Geometry&rdquo;, SIGGRAPH 2021 (Advances in Real-Time Rendering)</p> </blockquote> <hr> <h2 id="三角形triangle">三角形（Triangle）</h2> <p>Cluster 内的基本渲染单元。每个 Cluster 包含最多 <strong>128 个三角形</strong>。在 VisBuffer 中，三角形 ID 占约 7 bit（2^7 = 128），用于在 Material Pass 中定位该三角形的三个顶点并做重心坐标插值。</p>Nanite: A Deep Divehttps://mr0ptimist.github.io/posts/karis_nanite_siggraph_advances_2021/Mon, 20 Apr 2026 10:00:00 +0800https://mr0ptimist.github.io/posts/karis_nanite_siggraph_advances_2021/<h1 id="nanite-a-deep-dive">Nanite: A Deep Dive</h1> <blockquote> <p><strong>来源</strong>: <a href="https://advances.realtimerendering.com/s2021/Karis_Nanite_SIGGRAPH_Advances_2021_final.pdf">Karis_Nanite_SIGGRAPH_Advances_2021_final</a> — Brian Karis, Rune Stubbe, Graham Wihlidal <strong>会议</strong>: SIGGRAPH 2021 <em>Advances in Real-Time Rendering in Games</em> course <strong>作者主讲</strong>: Brian Karis (Engineering Fellow, Epic Games) <strong>主题</strong>: UE5 全新虚拟几何系统 Nanite 的深度技术解析</p> </blockquote> <hr> <h2 id="目录">目录</h2> <ol> <li><a href="#1-%E6%84%BF%E6%99%AF%E4%B8%8E%E7%8E%B0%E5%AE%9E">愿景与现实</a></li> <li><a href="#2-%E5%8F%AF%E9%80%89%E6%96%B9%E6%A1%88%E7%9A%84%E6%8E%A2%E7%B4%A2">可选方案的探索</a></li> <li><a href="#3-gpu-driven-pipeline">GPU Driven Pipeline</a></li> <li><a href="#4-%E4%B8%89%E8%A7%92%E5%BD%A2-cluster-culling-%E4%B8%8E-occlusion-culling">三角形 Cluster Culling 与 Occlusion Culling</a></li> <li><a href="#5-visibility-buffer-%E4%B8%8E%E5%8F%AF%E8%A7%81%E6%80%A7%E6%9D%90%E8%B4%A8%E8%A7%A3%E8%80%A6">Visibility Buffer 与可见性/材质解耦</a></li> <li><a href="#6-%E6%AC%A1%E7%BA%BF%E6%80%A7%E6%89%A9%E5%B1%95%E4%B8%8E-cluster-%E5%B1%82%E6%AC%A1%E7%BB%93%E6%9E%84">次线性扩展与 Cluster 层次结构</a></li> <li><a href="#7-lod-%E8%A3%82%E7%BC%9D%E9%97%AE%E9%A2%98%E4%B8%8E-dag-%E6%9E%84%E5%BB%BA">LOD 裂缝问题与 DAG 构建</a></li> <li><a href="#8-%E6%9E%84%E5%BB%BA%E6%B5%81%E7%A8%8B%E8%AF%A6%E8%A7%A3build-operations">构建流程详解（Build Operations）</a></li> <li><a href="#9-%E7%AE%80%E5%8C%96%E7%AE%97%E6%B3%95%E4%B8%8E%E8%AF%AF%E5%B7%AE%E5%BA%A6%E9%87%8F">简化算法与误差度量</a></li> <li><a href="#10-%E8%BF%90%E8%A1%8C%E6%97%B6%E8%A7%86%E7%9B%B8%E5%85%B3-lod-%E9%80%89%E6%8B%A9">运行时视相关 LOD 选择</a></li> <li><a href="#11-%E5%B9%B6%E8%A1%8C-lod-%E9%80%89%E6%8B%A9%E4%B8%8E%E5%B1%82%E6%AC%A1%E8%A3%81%E5%89%AA">并行 LOD 选择与层次裁剪</a></li> <li><a href="#12-persistent-threads-%E4%B8%8E%E4%B8%A4-pass-occlusion-culling">Persistent Threads 与两 Pass Occlusion Culling</a></li> <li><a href="#13-%E5%85%89%E6%A0%85%E5%8C%96%E8%BD%AF%E4%BB%B6--%E7%A1%AC%E4%BB%B6%E6%B7%B7%E5%90%88">光栅化（软件 + 硬件混合）</a></li> <li><a href="#14-%E5%B0%8F%E4%B8%89%E8%A7%92%E5%BD%A2%E4%B8%8E%E5%BE%AE%E5%A4%9A%E8%BE%B9%E5%BD%A2%E8%BD%AF%E5%85%89%E6%A0%85%E5%99%A8">小三角形与微多边形软光栅器</a></li> <li><a href="#15-%E5%B0%8F%E5%AE%9E%E4%BE%8Btiny-instances%E4%B8%8E-imposter">小实例（Tiny Instances）与 Imposter</a></li> <li><a href="#16-%E5%BB%B6%E8%BF%9F%E6%9D%90%E8%B4%A8%E6%B1%82%E5%80%BCdeferred-material-evaluation">延迟材质求值（Deferred Material Evaluation）</a></li> <li><a href="#17-%E6%B5%81%E6%B0%B4%E7%BA%BF%E6%80%A7%E8%83%BD%E6%95%B0%E6%8D%AE">流水线性能数据</a></li> <li><a href="#18-%E9%98%B4%E5%BD%B1virtual-shadow-maps">阴影：Virtual Shadow Maps</a></li> <li><a href="#19-streaming%E5%87%A0%E4%BD%95%E6%B5%81%E9%80%81">Streaming（几何流送）</a></li> <li><a href="#20-%E5%8E%8B%E7%BC%A9%E5%86%85%E5%AD%98%E8%A1%A8%E7%A4%BA%E4%B8%8E%E7%A3%81%E7%9B%98%E8%A1%A8%E7%A4%BA">压缩：内存表示与磁盘表示</a></li> <li><a href="#21-%E7%BB%93%E6%9E%9C%E4%B8%8E%E6%9C%AA%E6%9D%A5%E5%B7%A5%E4%BD%9C">结果与未来工作</a></li> <li><a href="#22-%E8%87%B4%E8%B0%A2%E4%B8%8E%E5%8F%82%E8%80%83%E6%96%87%E7%8C%AE">致谢与参考文献</a></li> </ol> <hr> <h2 id="1-愿景与现实">1. 愿景与现实</h2> <h3 id="11-the-dream梦想">1.1 The Dream（梦想）</h3> <p>像 Virtual Texturing 那样<strong>虚拟化几何</strong>：</p>