https://mr0ptimist.github.io/tags/tbdr/Recent content in TBDR on MrOptimistHugozh-cnMon, 20 Apr 2026 10:04:00 +0800https://mr0ptimist.github.io/posts/%E7%A7%BB%E5%8A%A8%E7%AB%AFgpu%E5%8F%AF%E8%A7%81%E6%80%A7%E5%89%94%E9%99%A4%E6%9C%BA%E5%88%B6/Mon, 20 Apr 2026 10:04:00 +0800https://mr0ptimist.github.io/posts/%E7%A7%BB%E5%8A%A8%E7%AB%AFgpu%E5%8F%AF%E8%A7%81%E6%80%A7%E5%89%94%E9%99%A4%E6%9C%BA%E5%88%B6/<h1 id="1-移动端-gpu-可见性剔除机制对比">1. 移动端 GPU 可见性剔除机制对比</h1> <h2 id="11-概览">1.1 概览</h2> <table> <thead> <tr> <th></th> <th>PowerVR HSR</th> <th>Apple HSR</th> <th>Mali FPK</th> <th>Adreno LRZ</th> </tr> </thead> <tbody> <tr> <td>全称</td> <td>Hidden Surface Removal</td> <td>Hidden Surface Removal</td> <td>Forward Pixel Kill</td> <td>Low Resolution Z</td> </tr> <tr> <td>架构</td> <td>TBDR</td> <td>TBDR</td> <td>TBDR</td> <td>TBDR</td> </tr> <tr> <td>粒度</td> <td>逐像素</td> <td>逐像素</td> <td>逐像素（尽力而为）</td> <td>逐块（8x8 像素）</td> </tr> <tr> <td>保证级</td> <td>不透明物体保证零过度绘制</td> <td>不透明物体保证零过度绘制</td> <td>非保证，尽力剔除</td> <td>非保证，块级粗剔除</td> </tr> <tr> <td>绘制顺序依赖</td> <td>不透明物体顺序无关</td> <td>不透明物体顺序无关</td> <td>正面到背面更优</td> <td>Binning pass 构建后顺序无关，但正面到背面可提升 Early-Z 效率</td> </tr> <tr> <td>AlphaTest</td> <td>失效</td> <td>失效</td> <td>失效</td> <td>失效</td> </tr> <tr> <td>Alpha Blend</td> <td>失效</td> <td>失效</td> <td>失效</td> <td>失效</td> </tr> <tr> <td>gl_FragDepth 写入</td> <td>失效</td> <td>失效</td> <td>失效</td> <td>失效</td> </tr> </tbody> </table> <hr> <h2 id="12-powervr-hsr-imagination">1.2 PowerVR HSR (Imagination)</h2> <h3 id="121-原理">1.2.1 原理</h3> <p>TBDR 架构中，所有几何体先提交到 Tile，HSR 在 PS 执行前对整个 Tile 做可见性解析，只对最终可见像素跑 PS。</p>