https://mr0ptimist.github.io/tags/nvidia/Recent content in NVIDIA on MrOptimistHugozh-cnMon, 20 Apr 2026 10:03:00 +0800https://mr0ptimist.github.io/posts/nvperf_counters_reference/Mon, 20 Apr 2026 10:03:00 +0800https://mr0ptimist.github.io/posts/nvperf_counters_reference/<h1 id="nvperf-gpu-性能计数器参考手册">NvPerf GPU 性能计数器参考手册</h1> <p>本文档基于 NVIDIA 官方文档，对 NvProfAnalyzer 中使用的所有 GPU 性能计数器进行中文解释。</p> <h2 id="参考文档来源">参考文档来源</h2> <ul> <li><a href="https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html">Nsight Compute Profiling Guide</a> — 计数器命名规则、硬件单元、管线定义</li> <li><a href="https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html">Nsight Graphics Advanced Learning</a> — 图形管线各单元的功能说明</li> <li><a href="https://docs.nvidia.com/nsight-graphics/UserGuide/gpu-trace-system-architecture.html">Nsight Graphics System Architecture</a> — GPU 系统架构图解</li> <li><a href="https://developer.nvidia.com/blog/the-peak-performance-analysis-method-for-optimizing-any-gpu-workload/">NVIDIA Peak Performance Analysis Blog</a> — 性能分析方法论</li> <li><a href="https://docs.nvidia.com/nsight-compute/NsightComputeCli/index.html">Nsight Compute CLI</a> — CLI 工具与指标映射表</li> </ul> <hr> <h2 id="一计数器命名规则">一、计数器命名规则</h2> <p>NVIDIA 性能计数器遵循统一的命名格式：</p> <pre tabindex="0"><code>单元__(子单元?)_(管线阶段?)_度量_(限定符?) </code></pre><p>示例解读：</p> <ul> <li> <p><strong><code>sm__inst_executed</code></strong></p> <ul> <li>单元: SM | 度量: inst_executed (指令执行) | 限定符: 无</li> <li>→ SM 执行的 warp 指令总数</li> </ul> </li> <li> <p><strong><code>sm__inst_executed_pipe_fma</code></strong></p> <ul> <li>单元: SM | 度量: inst_executed | 限定符: pipe_fma</li> <li>→ FMA 管线执行的 warp 指令数</li> </ul> </li> <li> <p><strong><code>smsp__thread_inst_executed_pipe_tex_pred_on</code></strong></p>https://mr0ptimist.github.io/posts/nvidia_gpu_performance_counters_complete_zh/Mon, 20 Apr 2026 10:02:00 +0800https://mr0ptimist.github.io/posts/nvidia_gpu_performance_counters_complete_zh/<h1 id="nvidia-gpu性能计数器完整参考手册-nvperfnsight系列">NVIDIA GPU性能计数器完整参考手册 (NvPerf/Nsight系列)</h1> <h2 id="文件信息">文件信息</h2> <ul> <li><strong>CSV文件示例</strong>: <code>Unity_2026.04.02_10.06_frame628066.pagecache.nvperf.csv</code></li> <li><strong>参数总数</strong>: 2958个性能计数器</li> <li><strong>工具演进</strong>: nvperf → Nsight系列工具（推荐）</li> </ul> <hr> <h2 id="一性能计数器命名规则详解">一、性能计数器命名规则详解</h2> <h3 id="11-nsight-compute命名规范">1.1 Nsight Compute命名规范</h3> <p>根据<a href="https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html">Nsight Compute Profiling Guide</a>：</p> <p><strong>基本格式</strong>: <code>unit__(subunit?)_(pipestage?)_quantity_(qualifiers?)</code></p> <p><strong>接口计数器</strong>: <code>unit__(subunit?)_(pipestage?)_(interface)_quantity_(qualifiers?)</code></p> <p><strong>组成部分</strong>:</p> <ul> <li><strong>unit</strong>: GPU逻辑或物理单元（如sm、dram、lts）</li> <li><strong>subunit</strong>: 单元内的子单元（可选）</li> <li><strong>pipestage</strong>: 管线阶段（可选）</li> <li><strong>quantity</strong>: 测量的内容（字节、计数、比率等）</li> <li><strong>qualifiers</strong>: 附加谓词（操作类型、访问模式等）</li> </ul> <h3 id="12-后缀含义">1.2 后缀含义</h3> <ul> <li><strong><code>.avg</code></strong>: 平均值</li> <li><strong><code>.max</code></strong>: 最大值</li> <li><strong><code>.min</code></strong>: 最小值</li> <li><strong><code>.sum</code></strong>: 总和</li> <li><strong><code>(bytes)</code></strong>: 单位标识（字节）</li> <li><strong><code>_op_read</code></strong>: 读取操作</li> <li><strong><code>_op_write</code></strong>: 写入操作</li> <li><strong><code>_lookup_hit</code></strong>: 查找命中</li> <li><strong><code>_lookup_miss</code></strong>: 查找未命中</li> </ul> <hr> <h2 id="二gpu硬件架构单元详解">二、GPU硬件架构单元详解</h2> <h3 id="21-计算核心单元">2.1 计算核心单元</h3> <table> <thead> <tr> <th>单元前缀</th> <th>中文名称</th> <th>功能描述</th> <th>对应文档</th> </tr> </thead> <tbody> <tr> <td><strong><code>sm__</code></strong></td> <td>流多处理器</td> <td>GPU的主要计算单元，包含多个CUDA核心，执行着色器指令</td> <td><a href="https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html">Nsight Compute Profiling Guide</a></td> </tr> <tr> <td><strong><code>smsp__</code></strong></td> <td>SM子分区</td> <td>SM内的四个子分区，各含调度器、寄存器文件和执行单元</td> <td>同上</td> </tr> <tr> <td><strong><code>tpc__</code></strong></td> <td>纹理处理集群</td> <td>包含多个SM和纹理单元的处理集群</td> <td><a href="https://docs.nvidia.com/nsight-graphics/UserGuide/gpu-trace-system-architecture.html">Nsight Graphics System Architecture</a></td> </tr> <tr> <td><strong><code>vpc__</code></strong></td> <td>顶点处理集群</td> <td>处理顶点着色相关任务的集群</td> <td>同上</td> </tr> </tbody> </table> <h3 id="22-图形管线单元">2.2 图形管线单元</h3> <table> <thead> <tr> <th>单元前缀</th> <th>中文名称</th> <th>功能描述</th> <th>对应文档</th> </tr> </thead> <tbody> <tr> <td><strong><code>fe__</code></strong></td> <td>前端单元</td> <td>图形管线的初始阶段，处理命令分发</td> <td><a href="https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html">Nsight Graphics Advanced Learning</a></td> </tr> <tr> <td><strong><code>gr__</code></strong></td> <td>图形渲染单元</td> <td>图形渲染相关操作</td> <td>同上</td> </tr> <tr> <td><strong><code>raster__</code></strong></td> <td>光栅化单元</td> <td>将图元转换为像素片段</td> <td>同上</td> </tr> <tr> <td><strong><code>pes__</code></strong></td> <td>图元引擎状态</td> <td>协调顶点、曲面细分、几何等阶段</td> <td>同上</td> </tr> </tbody> </table> <h3 id="23-内存系统单元">2.3 内存系统单元</h3> <table> <thead> <tr> <th>单元前缀</th> <th>中文名称</th> <th>功能描述</th> <th>对应文档</th> </tr> </thead> <tbody> <tr> <td><strong><code>dram__</code></strong></td> <td>DRAM内存控制器</td> <td>设备主内存（GDDR6/GDDR5X）访问控制器</td> <td><a href="https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html">Nsight Compute Profiling Guide</a></td> </tr> <tr> <td><strong><code>fbpa__</code></strong></td> <td>帧缓冲区分区</td> <td>帧缓冲区内存分区管理</td> <td><a href="https://docs.nvidia.com/nsight-graphics/UserGuide/gpu-trace-system-architecture.html">Nsight Graphics System Architecture</a></td> </tr> <tr> <td><strong><code>lts__</code></strong></td> <td>本地纹理存储</td> <td>纹理数据的本地存储</td> <td>同上</td> </tr> <tr> <td><strong><code>l1tex__</code></strong></td> <td>L1纹理缓存</td> <td>包含L1数据缓存和纹理处理两个并行管线</td> <td>同上</td> </tr> </tbody> </table> <h3 id="24-缓存系统单元">2.4 缓存系统单元</h3> <table> <thead> <tr> <th>单元前缀</th> <th>中文名称</th> <th>功能描述</th> <th>对应文档</th> </tr> </thead> <tbody> <tr> <td><strong><code>gcc__</code></strong></td> <td>图形命令缓存</td> <td>图形命令的缓存系统</td> <td><a href="https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html">Nsight Graphics Advanced Learning</a></td> </tr> <tr> <td><strong><code>l2__</code></strong></td> <td>L2缓存</td> <td>为GPU所有单元提供服务，一致性的中心点</td> <td><a href="https://docs.nvidia.com/nsight-graphics/UserGuide/gpu-trace-system-architecture.html">Nsight Graphics System Architecture</a></td> </tr> <tr> <td><strong><code>syslts__</code></strong></td> <td>系统本地纹理存储</td> <td>系统级的纹理存储管理</td> <td>同上</td> </tr> </tbody> </table> <h3 id="25-其他系统单元">2.5 其他系统单元</h3> <table> <thead> <tr> <th>单元前缀</th> <th>中文名称</th> <th>功能描述</th> <th>对应文档</th> </tr> </thead> <tbody> <tr> <td><strong><code>idc__</code></strong></td> <td>指令分发单元</td> <td>指令分发相关操作</td> <td><a href="https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html">Nsight Compute Profiling Guide</a></td> </tr> <tr> <td><strong><code>pcie__</code></strong></td> <td>PCI Express总线</td> <td>CPU-GPU数据传输总线</td> <td><a href="https://docs.nvidia.com/nsight-graphics/UserGuide/gpu-trace-system-architecture.html">Nsight Graphics System Architecture</a></td> </tr> <tr> <td><strong><code>prop__</code></strong></td> <td>预ROP单元</td> <td>协调深度和颜色像素处理，管理API顺序</td> <td><a href="https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html">Nsight Graphics Advanced Learning</a></td> </tr> <tr> <td><strong><code>rtcore__</code></strong></td> <td>光线追踪核心</td> <td>专用光线追踪处理单元</td> <td>同上</td> </tr> </tbody> </table> <hr> <h2 id="三图形管线处理阶段详解">三、图形管线处理阶段详解</h2> <h3 id="31-前端处理world-pipe">3.1 前端处理（World Pipe）</h3> <p>根据<a href="https://docs.nvidia.com/nsight-graphics/AdvancedLearning/index.html">Nsight Graphics Advanced Learning</a>：</p>