<!DOCTYPE html> <!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]--> <!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]--> <head> <meta charset="utf-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Tune GPU Performance — PaddlePaddle documentation</title> <link rel="stylesheet" href="../../_static/css/theme.css" type="text/css" /> <link rel="index" title="Index" href="../../genindex.html"/> <link rel="search" title="Search" href="../../search.html"/> <link rel="top" title="PaddlePaddle documentation" href="../../index.html"/> <link rel="up" title="HOW TO" href="../index_en.html"/> <link rel="next" title="API" href="../../api/index_en.html"/> <link rel="prev" title="RNN Models" href="../deep_model/rnn/index_en.html"/> <link rel="stylesheet" href="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/css/perfect-scrollbar.min.css" type="text/css" /> <link rel="stylesheet" href="../../_static/css/override.css" type="text/css" /> <script> var _hmt = _hmt || []; (function() { var hm = document.createElement("script"); hm.src = "//hm.baidu.com/hm.js?b9a314ab40d04d805655aab1deee08ba"; var s = document.getElementsByTagName("script")[0]; s.parentNode.insertBefore(hm, s); })(); </script> <script src="../../_static/js/modernizr.min.js"></script> </head> <body class="wy-body-for-nav" role="document"> <header class="site-header"> <div class="site-logo"> <a href="/"><img src="../../_static/images/PP_w.png"></a> </div> <div class="site-nav-links"> <div class="site-menu"> <a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Folk me on Github</a> <div class="language-switcher dropdown"> <a type="button" data-toggle="dropdown"> <span>English</span> <i class="fa fa-angle-up"></i> <i class="fa fa-angle-down"></i> </a> <ul class="dropdown-menu"> <li><a href="/doc_cn">中文</a></li> <li><a href="/doc">English</a></li> </ul> </div> <ul class="site-page-links"> <li><a href="/">Home</a></li> </ul> </div> <div class="doc-module"> <ul class="current"> <li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a></li> <li class="toctree-l1 current"><a class="reference internal" href="../index_en.html">HOW TO</a></li> <li class="toctree-l1"><a class="reference internal" href="../../api/index_en.html">API</a></li> <li class="toctree-l1"><a class="reference internal" href="../../about/index_en.html">ABOUT</a></li> </ul> <div role="search"> <form id="rtd-search-form" class="wy-form" action="../../search.html" method="get"> <input type="text" name="q" placeholder="Search docs" /> <input type="hidden" name="check_keywords" value="yes" /> <input type="hidden" name="area" value="default" /> </form> </div> </div> </div> </header> <div class="main-content-wrap"> <nav class="doc-menu-vertical" role="navigation"> <ul class="current"> <li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a><ul> <li class="toctree-l2"><a class="reference internal" href="../../getstarted/build_and_install/index_en.html">Install and Build</a><ul> <li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/docker_install_en.html">PaddlePaddle in Docker Containers</a></li> <li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/ubuntu_install_en.html">Debian Package installation guide</a></li> <li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/build_from_source_en.html">Installing from Sources</a></li> </ul> </li> </ul> </li> <li class="toctree-l1 current"><a class="reference internal" href="../index_en.html">HOW TO</a><ul class="current"> <li class="toctree-l2"><a class="reference internal" href="../usage/cmd_parameter/index_en.html">Set Command-line Parameters</a><ul> <li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/use_case_en.html">Use Case</a></li> <li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/arguments_en.html">Argument Outline</a></li> <li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/detail_introduction_en.html">Detail Description</a></li> </ul> </li> <li class="toctree-l2"><a class="reference internal" href="../usage/cluster/cluster_train_en.html">Run Distributed Training</a></li> <li class="toctree-l2"><a class="reference internal" href="../usage/k8s/k8s_en.html">Paddle On Kubernetes</a></li> <li class="toctree-l2"><a class="reference internal" href="../usage/k8s/k8s_aws_en.html">Distributed PaddlePaddle Training on AWS with Kubernetes</a></li> <li class="toctree-l2"><a class="reference internal" href="../dev/new_layer_en.html">Write New Layers</a></li> <li class="toctree-l2"><a class="reference internal" href="../dev/contribute_to_paddle_en.html">Contribute Code</a></li> <li class="toctree-l2"><a class="reference internal" href="../deep_model/rnn/index_en.html">RNN Models</a></li> <li class="toctree-l2 current"><a class="current reference internal" href="#">Tune GPU Performance</a></li> </ul> </li> <li class="toctree-l1"><a class="reference internal" href="../../api/index_en.html">API</a><ul> <li class="toctree-l2"><a class="reference internal" href="../../api/v2/model_configs.html">Model Configuration</a><ul> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/activation.html">Activation</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/layer.html">Layers</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/evaluators.html">Evaluators</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/optimizer.html">Optimizer</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/pooling.html">Pooling</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/networks.html">Networks</a></li> <li class="toctree-l3"><a class="reference internal" href="../../api/v2/config/attr.html">Parameter Attribute</a></li> </ul> </li> <li class="toctree-l2"><a class="reference internal" href="../../api/v2/data.html">Data Reader Interface and DataSets</a></li> <li class="toctree-l2"><a class="reference internal" href="../../api/v2/run_logic.html">Training and Inference</a></li> </ul> </li> <li class="toctree-l1"><a class="reference internal" href="../../about/index_en.html">ABOUT</a></li> </ul> </nav> <section class="doc-content-wrap"> <div role="navigation" aria-label="breadcrumbs navigation"> <ul class="wy-breadcrumbs"> <li><a href="../index_en.html">HOW TO</a> > </li> <li>Tune GPU Performance</li> </ul> </div> <div class="wy-nav-content" id="doc-content"> <div class="rst-content"> <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article"> <div itemprop="articleBody"> <div class="section" id="tune-gpu-performance"> <h1><a class="toc-backref" href="#id1">Tune GPU Performance</a><a class="headerlink" href="#tune-gpu-performance" title="Permalink to this headline">¶</a></h1> <div class="contents topic" id="contents"> <p class="topic-title first">Contents</p> <ul class="simple"> <li><a class="reference internal" href="#tune-gpu-performance" id="id1">Tune GPU Performance</a><ul> <li><a class="reference internal" href="#what-s-profiling" id="id2">What’s profiling?</a></li> <li><a class="reference internal" href="#why-we-need-profiling" id="id3">Why we need profiling?</a></li> <li><a class="reference internal" href="#how-to-do-profiling" id="id4">How to do profiling?</a></li> <li><a class="reference internal" href="#profiler-tools" id="id5">Profiler Tools</a></li> <li><a class="reference internal" href="#hands-on-approach" id="id6">Hands-on Approach</a><ul> <li><a class="reference internal" href="#built-in-timer" id="id7">Built-in Timer</a></li> <li><a class="reference internal" href="#nvprof-profiler" id="id8">nvprof profiler</a></li> <li><a class="reference internal" href="#nvvp-profiler" id="id9">nvvp profiler</a></li> </ul> </li> <li><a class="reference internal" href="#profiling-tips" id="id10">Profiling tips</a></li> <li><a class="reference internal" href="#reference" id="id11">Reference</a></li> </ul> </li> </ul> </div> <p>This tutorial will guide you step-by-step through how to conduct profiling and performance tuning using built-in timer, <strong>nvprof</strong> and <strong>nvvp</strong>.</p> <ul class="simple"> <li>What is profiling?</li> <li>Why we need profiling?</li> <li>How to do profiling?</li> <li>Profile tools</li> <li>Hands-on Tutorial</li> <li>Profiling tips</li> </ul> <div class="section" id="what-s-profiling"> <h2><a class="toc-backref" href="#id2">What’s profiling?</a><a class="headerlink" href="#what-s-profiling" title="Permalink to this headline">¶</a></h2> <p>In software engineering, profiling is a form of dynamic program analysis that measures the space (memory) or time complexity of a program, the usage of particular instructions, or the frequency and duration of function calls. Most commonly, profiling information serves to aid program optimization.</p> <p>Briefly, profiler is used to measure application performance. Program analysis tools are extremely important for understanding program behavior. Simple profiling can tell you that how long does an operation take? For advanced profiling, it can interpret why does an operation take a long time?</p> </div> <div class="section" id="why-we-need-profiling"> <h2><a class="toc-backref" href="#id3">Why we need profiling?</a><a class="headerlink" href="#why-we-need-profiling" title="Permalink to this headline">¶</a></h2> <p>Since training deep neural network typically take a very long time to get over, performance is gradually becoming the most important thing in deep learning field. The first step to improve performance is to understand what parts are slow. There is no point in improving performance of a region which doesn’t take much time!</p> </div> <div class="section" id="how-to-do-profiling"> <h2><a class="toc-backref" href="#id4">How to do profiling?</a><a class="headerlink" href="#how-to-do-profiling" title="Permalink to this headline">¶</a></h2> <p>To achieve maximum performance, there are five steps you can take to reach your goals.</p> <ul class="simple"> <li>Profile the code</li> <li>Find the slow parts</li> <li>Work out why they’re slow</li> <li>Make them fast</li> <li>Profile the code again</li> </ul> <p>Usually, processor has two key performance limits include float point throughput and memory throughput. For GPU, it also need more parallelism to fulfill its potential. This is why they can be so fast.</p> </div> <div class="section" id="profiler-tools"> <h2><a class="toc-backref" href="#id5">Profiler Tools</a><a class="headerlink" href="#profiler-tools" title="Permalink to this headline">¶</a></h2> <p>For general GPU profiling, a bunch of tools are provided from both NVIDIA and third party.</p> <p><strong>nvprof</strong> is Nvidia profiler and <strong>nvvp</strong> is (GUI based) Nvidia visual profiler. In this tutorial, we will focus on nvprof and nvvp.</p> <p><code class="code docutils literal"><span class="pre">test_GpuProfiler</span></code> from <code class="code docutils literal"><span class="pre">paddle/math/tests</span></code> directory will be used to evaluate above profilers.</p> <div class="highlight-c++"><table class="highlighttable"><tr><td class="linenos"><div class="linenodiv"><pre> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15</pre></div></td><td class="code"><div class="highlight"><pre><span></span><span class="n">TEST</span><span class="p">(</span><span class="n">Profiler</span><span class="p">,</span> <span class="n">testBilinearFwdBwd</span><span class="p">)</span> <span class="p">{</span> <span class="k">auto</span> <span class="n">numSamples</span> <span class="o">=</span> <span class="mi">10</span><span class="p">;</span> <span class="k">auto</span> <span class="n">channels</span> <span class="o">=</span> <span class="mi">16</span><span class="p">;</span> <span class="k">auto</span> <span class="n">imgSize</span> <span class="o">=</span> <span class="mi">64</span><span class="p">;</span> <span class="p">{</span> <span class="c1">// nvprof: GPU Proflier</span> <span class="n">REGISTER_GPU_PROFILER</span><span class="p">(</span><span class="s">"testBilinearFwdBwd"</span><span class="p">);</span> <span class="c1">// Paddle built-in timer</span> <span class="n">REGISTER_TIMER_INFO</span><span class="p">(</span> <span class="s">"testBilinearFwdBwd"</span><span class="p">,</span> <span class="s">"numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"</span><span class="p">);</span> <span class="n">testBilinearFwdBwd</span><span class="p">(</span><span class="n">numSamples</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">channels</span><span class="p">);</span> <span class="p">}</span> <span class="n">globalStat</span><span class="p">.</span><span class="n">printAllStatus</span><span class="p">();</span> <span class="p">}</span> </pre></div> </td></tr></table></div> <p>The above code snippet includes two methods, you can use any of them to profile the regions of interest.</p> <ol class="arabic simple"> <li><code class="code docutils literal"><span class="pre">REGISTER_TIMER_INFO</span></code> is a built-in timer wrapper which can calculate the time overhead of both cpu functions and cuda kernels.</li> </ol> <p>2. <code class="code docutils literal"><span class="pre">REGISTER_GPU_PROFILER</span></code> is a general purpose wrapper object of <code class="code docutils literal"><span class="pre">cudaProfilerStart</span></code> and <code class="code docutils literal"><span class="pre">cudaProfilerStop</span></code> to avoid program crashes when CPU version of PaddlePaddle invokes them.</p> <p>You can find more details about how to use both of them in the next session.</p> </div> <div class="section" id="hands-on-approach"> <h2><a class="toc-backref" href="#id6">Hands-on Approach</a><a class="headerlink" href="#hands-on-approach" title="Permalink to this headline">¶</a></h2> <div class="section" id="built-in-timer"> <h3><a class="toc-backref" href="#id7">Built-in Timer</a><a class="headerlink" href="#built-in-timer" title="Permalink to this headline">¶</a></h3> <p>To enable built-in timer in PaddlePaddle, first you have to add <code class="code docutils literal"><span class="pre">REGISTER_TIMER_INFO</span></code> into the regions of you interest. Then, all information could be stamped in the console via <code class="code docutils literal"><span class="pre">printStatus</span></code> or <code class="code docutils literal"><span class="pre">printAllStatus</span></code> function. As a simple example, consider the following:</p> <ol class="arabic"> <li><p class="first">Add <code class="code docutils literal"><span class="pre">REGISTER_TIMER_INFO</span></code> and <code class="code docutils literal"><span class="pre">printAllStatus</span></code> functions (see the emphasize-lines).</p> <blockquote> <div><div class="highlight-c++"><table class="highlighttable"><tr><td class="linenos"><div class="linenodiv"><pre> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15</pre></div></td><td class="code"><div class="highlight"><pre><span></span><span class="n">TEST</span><span class="p">(</span><span class="n">Profiler</span><span class="p">,</span> <span class="n">testBilinearFwdBwd</span><span class="p">)</span> <span class="p">{</span> <span class="k">auto</span> <span class="n">numSamples</span> <span class="o">=</span> <span class="mi">10</span><span class="p">;</span> <span class="k">auto</span> <span class="n">channels</span> <span class="o">=</span> <span class="mi">16</span><span class="p">;</span> <span class="k">auto</span> <span class="n">imgSize</span> <span class="o">=</span> <span class="mi">64</span><span class="p">;</span> <span class="p">{</span> <span class="c1">// nvprof: GPU Proflier</span> <span class="n">REGISTER_GPU_PROFILER</span><span class="p">(</span><span class="s">"testBilinearFwdBwd"</span><span class="p">);</span> <span class="hll"> <span class="c1">// Paddle built-in timer</span> </span><span class="hll"> <span class="n">REGISTER_TIMER_INFO</span><span class="p">(</span> </span><span class="hll"> <span class="s">"testBilinearFwdBwd"</span><span class="p">,</span> </span><span class="hll"> <span class="s">"numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"</span><span class="p">);</span> </span><span class="hll"> <span class="n">testBilinearFwdBwd</span><span class="p">(</span><span class="n">numSamples</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">channels</span><span class="p">);</span> </span> <span class="p">}</span> <span class="hll"> <span class="n">globalStat</span><span class="p">.</span><span class="n">printAllStatus</span><span class="p">();</span> </span><span class="p">}</span> </pre></div> </td></tr></table></div> </div></blockquote> </li> <li><p class="first">Configure cmake with <strong>WITH_TIMER</strong> and recompile PaddlePaddle.</p> <blockquote> <div><div class="highlight-bash"><div class="highlight"><pre><span></span>cmake .. -DWITH_TIMER<span class="o">=</span>ON make </pre></div> </div> </div></blockquote> </li> <li><p class="first">Execute your code and observe the results (see the emphasize-lines).</p> <blockquote> <div><div class="highlight-bash"><div class="highlight"><pre><span></span><span class="hll">> ./paddle/math/tests/test_GpuProfiler </span>I1117 <span class="m">11</span>:13:42.313065 <span class="m">2522362816</span> Util.cpp:155<span class="o">]</span> commandline: ./paddle/math/tests/test_GpuProfiler I1117 <span class="m">11</span>:13:42.845065 <span class="m">2522362816</span> Util.cpp:130<span class="o">]</span> Calling runInitFunctions I1117 <span class="m">11</span>:13:42.845208 <span class="m">2522362816</span> Util.cpp:143<span class="o">]</span> Call runInitFunctions <span class="k">done</span>. <span class="o">[==========]</span> Running <span class="m">1</span> <span class="nb">test</span> from <span class="m">1</span> <span class="nb">test</span> <span class="k">case</span>. <span class="o">[</span>----------<span class="o">]</span> Global <span class="nb">test</span> environment set-up. <span class="o">[</span>----------<span class="o">]</span> <span class="m">1</span> <span class="nb">test</span> from Profiler <span class="o">[</span> RUN <span class="o">]</span> Profiler.BilinearFwdBwd I1117 <span class="m">11</span>:13:42.845310 <span class="m">2522362816</span> test_GpuProfiler.cpp:114<span class="o">]</span> Enable GPU Profiler Stat: <span class="o">[</span>testBilinearFwdBwd<span class="o">]</span> <span class="s2">"numSamples = 10, channels = 16, im</span> <span class="s2">gSizeX = 64, imgSizeY = 64"</span> I1117 <span class="m">11</span>:13:42.850154 <span class="m">2522362816</span> ThreadLocal.cpp:37<span class="o">]</span> thread use undeterministic rand seed:20659751 <span class="hll">I1117 <span class="m">11</span>:13:42.981501 <span class="m">2522362816</span> Stat.cpp:130<span class="o">]</span> <span class="o">=======</span> StatSet: <span class="o">[</span>GlobalStatInfo<span class="o">]</span> <span class="nv">status</span> <span class="o">======</span> </span><span class="hll">I1117 <span class="m">11</span>:13:42.981539 <span class="m">2522362816</span> Stat.cpp:133<span class="o">]</span> <span class="nv">Stat</span><span class="o">=</span>testBilinearFwdBwd <span class="nv">total</span><span class="o">=</span><span class="m">136</span>.141 <span class="nv">avg</span><span class="o">=</span><span class="m">136</span>.141 <span class="nv">max</span><span class="o">=</span><span class="m">136</span>.141 <span class="nv">min</span><span class="o">=</span><span class="m">136</span>.141 <span class="nv">count</span><span class="o">=</span><span class="m">1</span> </span><span class="hll">I1117 <span class="m">11</span>:13:42.981572 <span class="m">2522362816</span> Stat.cpp:141<span class="o">]</span> <span class="o">=======</span> BarrierStatSet <span class="nv">status</span> <span class="o">======</span> </span><span class="hll">I1117 <span class="m">11</span>:13:42.981575 <span class="m">2522362816</span> Stat.cpp:154<span class="o">]</span> -------------------------------------------------- </span><span class="o">[</span> OK <span class="o">]</span> Profiler.BilinearFwdBwd <span class="o">(</span><span class="m">136</span> ms<span class="o">)</span> <span class="o">[</span>----------<span class="o">]</span> <span class="m">1</span> <span class="nb">test</span> from Profiler <span class="o">(</span><span class="m">136</span> ms total<span class="o">)</span> <span class="o">[</span>----------<span class="o">]</span> Global <span class="nb">test</span> environment tear-down <span class="o">[==========]</span> <span class="m">1</span> <span class="nb">test</span> from <span class="m">1</span> <span class="nb">test</span> <span class="k">case</span> ran. <span class="o">(</span><span class="m">136</span> ms total<span class="o">)</span> <span class="o">[</span> PASSED <span class="o">]</span> <span class="m">1</span> test. </pre></div> </div> </div></blockquote> </li> </ol> </div> <div class="section" id="nvprof-profiler"> <h3><a class="toc-backref" href="#id8">nvprof profiler</a><a class="headerlink" href="#nvprof-profiler" title="Permalink to this headline">¶</a></h3> <p>To use this command line profiler <strong>nvprof</strong>, you can simply issue the following command:</p> <ol class="arabic"> <li><p class="first">Add <code class="code docutils literal"><span class="pre">REGISTER_GPU_PROFILER</span></code> function (see the emphasize-lines).</p> <blockquote> <div><div class="highlight-c++"><table class="highlighttable"><tr><td class="linenos"><div class="linenodiv"><pre> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15</pre></div></td><td class="code"><div class="highlight"><pre><span></span><span class="n">TEST</span><span class="p">(</span><span class="n">Profiler</span><span class="p">,</span> <span class="n">testBilinearFwdBwd</span><span class="p">)</span> <span class="p">{</span> <span class="k">auto</span> <span class="n">numSamples</span> <span class="o">=</span> <span class="mi">10</span><span class="p">;</span> <span class="k">auto</span> <span class="n">channels</span> <span class="o">=</span> <span class="mi">16</span><span class="p">;</span> <span class="k">auto</span> <span class="n">imgSize</span> <span class="o">=</span> <span class="mi">64</span><span class="p">;</span> <span class="p">{</span> <span class="hll"> <span class="c1">// nvprof: GPU Proflier</span> </span><span class="hll"> <span class="n">REGISTER_GPU_PROFILER</span><span class="p">(</span><span class="s">"testBilinearFwdBwd"</span><span class="p">);</span> </span> <span class="c1">// Paddle built-in timer</span> <span class="n">REGISTER_TIMER_INFO</span><span class="p">(</span> <span class="s">"testBilinearFwdBwd"</span><span class="p">,</span> <span class="s">"numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64"</span><span class="p">);</span> <span class="n">testBilinearFwdBwd</span><span class="p">(</span><span class="n">numSamples</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">imgSize</span><span class="p">,</span> <span class="n">channels</span><span class="p">);</span> <span class="p">}</span> <span class="n">globalStat</span><span class="p">.</span><span class="n">printAllStatus</span><span class="p">();</span> <span class="p">}</span> </pre></div> </td></tr></table></div> </div></blockquote> </li> <li><p class="first">Configure cmake with <strong>WITH_PROFILER</strong> and recompile PaddlePaddle.</p> <blockquote> <div><div class="highlight-bash"><div class="highlight"><pre><span></span>cmake .. -DWITH_PROFILER<span class="o">=</span>ON make </pre></div> </div> </div></blockquote> </li> <li><p class="first">Use Nvidia profiler <strong>nvprof</strong> to profile the binary.</p> <blockquote> <div><div class="highlight-bash"><div class="highlight"><pre><span></span>nvprof ./paddle/math/tests/test_GpuProfiler </pre></div> </div> </div></blockquote> </li> </ol> <p>Then, you can get the following profiling result:</p> <div class="highlight-bash"><div class="highlight"><pre><span></span><span class="o">==</span><span class="nv">78544</span><span class="o">==</span> Profiling application: ./paddle/math/tests/test_GpuProfiler <span class="o">==</span><span class="nv">78544</span><span class="o">==</span> Profiling result: Time<span class="o">(</span>%<span class="o">)</span> Time Calls Avg Min Max Name <span class="m">27</span>.60% <span class="m">9</span>.6305ms <span class="m">5</span> <span class="m">1</span>.9261ms <span class="m">3</span>.4560us <span class="m">6</span>.4035ms <span class="o">[</span>CUDA memcpy HtoD<span class="o">]</span> <span class="m">26</span>.07% <span class="m">9</span>.0957ms <span class="m">1</span> <span class="m">9</span>.0957ms <span class="m">9</span>.0957ms <span class="m">9</span>.0957ms KeBilinearInterpBw <span class="m">23</span>.78% <span class="m">8</span>.2977ms <span class="m">1</span> <span class="m">8</span>.2977ms <span class="m">8</span>.2977ms <span class="m">8</span>.2977ms KeBilinearInterpFw <span class="m">22</span>.55% <span class="m">7</span>.8661ms <span class="m">2</span> <span class="m">3</span>.9330ms <span class="m">1</span>.5798ms <span class="m">6</span>.2863ms <span class="o">[</span>CUDA memcpy DtoH<span class="o">]</span> <span class="o">==</span><span class="nv">78544</span><span class="o">==</span> API calls: Time<span class="o">(</span>%<span class="o">)</span> Time Calls Avg Min Max Name <span class="m">46</span>.85% <span class="m">682</span>.28ms <span class="m">8</span> <span class="m">85</span>.285ms <span class="m">12</span>.639us <span class="m">682</span>.03ms cudaStreamCreateWithFlags <span class="m">39</span>.83% <span class="m">580</span>.00ms <span class="m">4</span> <span class="m">145</span>.00ms 302ns <span class="m">550</span>.27ms cudaFree <span class="m">9</span>.82% <span class="m">143</span>.03ms <span class="m">9</span> <span class="m">15</span>.892ms <span class="m">8</span>.7090us <span class="m">142</span>.78ms cudaStreamCreate <span class="m">1</span>.23% <span class="m">17</span>.983ms <span class="m">7</span> <span class="m">2</span>.5690ms <span class="m">23</span>.210us <span class="m">6</span>.4563ms cudaMemcpy <span class="m">1</span>.23% <span class="m">17</span>.849ms <span class="m">2</span> <span class="m">8</span>.9247ms <span class="m">8</span>.4726ms <span class="m">9</span>.3768ms cudaStreamSynchronize <span class="m">0</span>.66% <span class="m">9</span>.5969ms <span class="m">7</span> <span class="m">1</span>.3710ms <span class="m">288</span>.43us <span class="m">2</span>.4279ms cudaHostAlloc <span class="m">0</span>.13% <span class="m">1</span>.9530ms <span class="m">11</span> <span class="m">177</span>.54us <span class="m">7</span>.6810us <span class="m">591</span>.06us cudaMalloc <span class="m">0</span>.07% <span class="m">1</span>.0424ms <span class="m">8</span> <span class="m">130</span>.30us <span class="m">1</span>.6970us <span class="m">453</span>.72us cudaGetDevice <span class="m">0</span>.04% <span class="m">527</span>.90us <span class="m">40</span> <span class="m">13</span>.197us 525ns <span class="m">253</span>.99us cudaEventCreateWithFlags <span class="m">0</span>.03% <span class="m">435</span>.73us <span class="m">348</span> <span class="m">1</span>.2520us 124ns <span class="m">42</span>.704us cuDeviceGetAttribute <span class="m">0</span>.03% <span class="m">419</span>.36us <span class="m">1</span> <span class="m">419</span>.36us <span class="m">419</span>.36us <span class="m">419</span>.36us cudaGetDeviceCount <span class="m">0</span>.02% <span class="m">260</span>.75us <span class="m">2</span> <span class="m">130</span>.38us <span class="m">129</span>.32us <span class="m">131</span>.43us cudaGetDeviceProperties <span class="m">0</span>.02% <span class="m">222</span>.32us <span class="m">2</span> <span class="m">111</span>.16us <span class="m">106</span>.94us <span class="m">115</span>.39us cudaLaunch <span class="m">0</span>.01% <span class="m">214</span>.06us <span class="m">4</span> <span class="m">53</span>.514us <span class="m">28</span>.586us <span class="m">77</span>.655us cuDeviceGetName <span class="m">0</span>.01% <span class="m">115</span>.45us <span class="m">4</span> <span class="m">28</span>.861us <span class="m">9</span>.8250us <span class="m">44</span>.526us cuDeviceTotalMem <span class="m">0</span>.01% <span class="m">83</span>.988us <span class="m">4</span> <span class="m">20</span>.997us 578ns <span class="m">77</span>.760us cudaSetDevice <span class="m">0</span>.00% <span class="m">38</span>.918us <span class="m">1</span> <span class="m">38</span>.918us <span class="m">38</span>.918us <span class="m">38</span>.918us cudaEventCreate <span class="m">0</span>.00% <span class="m">34</span>.573us <span class="m">31</span> <span class="m">1</span>.1150us 279ns <span class="m">12</span>.784us cudaDeviceGetAttribute <span class="m">0</span>.00% <span class="m">17</span>.767us <span class="m">1</span> <span class="m">17</span>.767us <span class="m">17</span>.767us <span class="m">17</span>.767us cudaProfilerStart <span class="m">0</span>.00% <span class="m">15</span>.228us <span class="m">2</span> <span class="m">7</span>.6140us <span class="m">3</span>.5460us <span class="m">11</span>.682us cudaConfigureCall <span class="m">0</span>.00% <span class="m">14</span>.536us <span class="m">2</span> <span class="m">7</span>.2680us <span class="m">1</span>.1490us <span class="m">13</span>.387us cudaGetLastError <span class="m">0</span>.00% <span class="m">8</span>.6080us <span class="m">26</span> 331ns 173ns 783ns cudaSetupArgument <span class="m">0</span>.00% <span class="m">5</span>.5470us <span class="m">6</span> 924ns 215ns <span class="m">2</span>.6780us cuDeviceGet <span class="m">0</span>.00% <span class="m">5</span>.4090us <span class="m">6</span> 901ns 328ns <span class="m">3</span>.3320us cuDeviceGetCount <span class="m">0</span>.00% <span class="m">4</span>.1770us <span class="m">3</span> <span class="m">1</span>.3920us <span class="m">1</span>.0630us <span class="m">1</span>.8300us cuDriverGetVersion <span class="m">0</span>.00% <span class="m">3</span>.4650us <span class="m">3</span> <span class="m">1</span>.1550us <span class="m">1</span>.0810us <span class="m">1</span>.2680us cuInit <span class="m">0</span>.00% 830ns <span class="m">1</span> 830ns 830ns 830ns cudaRuntimeGetVersion </pre></div> </div> </div> <div class="section" id="nvvp-profiler"> <h3><a class="toc-backref" href="#id9">nvvp profiler</a><a class="headerlink" href="#nvvp-profiler" title="Permalink to this headline">¶</a></h3> <p>For visual profiler <strong>nvvp</strong>, you can either import the output of <code class="code docutils literal"><span class="pre">nvprof</span> <span class="pre">–o</span> <span class="pre">...</span></code> or run application through GUI.</p> <p><strong>Note: nvvp also support CPU profiling</strong> (Click the box in nvvp to enable profile execution on CPU).</p> <a class="reference internal image-reference" href="../../_images/nvvp1.png"><img alt="../../_images/nvvp1.png" class="align-center" src="../../_images/nvvp1.png" style="width: 932.58px; height: 399.3px;" /></a> <p>From the perspective of kernel functions, <strong>nvvp</strong> can even illustrate why does an operation take a long time? As shown in the following figure, kernel’s block usage, register usage and shared memory usage from <code class="code docutils literal"><span class="pre">nvvp</span></code> allow us to fully utilize all warps on the GPU.</p> <a class="reference internal image-reference" href="../../_images/nvvp2.png"><img alt="../../_images/nvvp2.png" class="align-center" src="../../_images/nvvp2.png" style="width: 727.32px; height: 475.86px;" /></a> <p>From the perspective of application, <strong>nvvp</strong> can give you some suggestions to address performance bottleneck. For instance, some advice in data movement and compute utilization from the below figure can guide you to tune performance.</p> <a class="reference internal image-reference" href="../../_images/nvvp3.png"><img alt="../../_images/nvvp3.png" class="align-center" src="../../_images/nvvp3.png" style="width: 717.42px; height: 213.84px;" /></a> <a class="reference internal image-reference" href="../../_images/nvvp4.png"><img alt="../../_images/nvvp4.png" class="align-center" src="../../_images/nvvp4.png" style="width: 718.08px; height: 197.34px;" /></a> </div> </div> <div class="section" id="profiling-tips"> <h2><a class="toc-backref" href="#id10">Profiling tips</a><a class="headerlink" href="#profiling-tips" title="Permalink to this headline">¶</a></h2> <ul class="simple"> <li>The <strong>nvprof</strong> and <strong>nvvp</strong> output is a very good place to start.</li> <li>The timeline is a good place to go next.</li> <li>Only dig deep into a kernel if it’s taking a significant amount of your time.</li> <li><dl class="first docutils"> <dt>Where possible, try to match profiler output with theory.</dt> <dd><ol class="first last arabic"> <li>For example, if I know I’m moving 1GB, and my kernel takes 10ms, I expect the profiler to report 100GB/s.</li> <li>Discrepancies are likely to mean your application isn’t doing what you thought it was.</li> </ol> </dd> </dl> </li> <li>Know your hardware: If your GPU can do 6 TFLOPs, and you’re already doing 5.5 TFLOPs, you won’t go much faster!</li> </ul> <p>Profiling is a key step in optimization. Sometimes quite simple changes can lead to big improvements in performance. Your mileage may vary!</p> </div> <div class="section" id="reference"> <h2><a class="toc-backref" href="#id11">Reference</a><a class="headerlink" href="#reference" title="Permalink to this headline">¶</a></h2> <p>Jeremy Appleyard, <a class="reference external" href="http://www.robots.ox.ac.uk/~seminars/seminars/Extra/2015_10_08_JeremyAppleyard.pdf">GPU Profiling for Deep Learning</a>, 2015</p> </div> </div> </div> </div> <footer> <div class="rst-footer-buttons" role="navigation" aria-label="footer navigation"> <a href="../../api/index_en.html" class="btn btn-neutral float-right" title="API" accesskey="n">Next <span class="fa fa-arrow-circle-right"></span></a> <a href="../deep_model/rnn/index_en.html" class="btn btn-neutral" title="RNN Models" accesskey="p"><span class="fa fa-arrow-circle-left"></span> Previous</a> </div> <hr/> <div role="contentinfo"> <p> © Copyright 2016, PaddlePaddle developers. </p> </div> Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/snide/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>. </footer> </div> </div> </section> </div> <script type="text/javascript"> var DOCUMENTATION_OPTIONS = { URL_ROOT:'../../', VERSION:'', COLLAPSE_INDEX:false, FILE_SUFFIX:'.html', HAS_SOURCE: true, SOURCELINK_SUFFIX: ".txt", }; </script> <script type="text/javascript" src="../../_static/jquery.js"></script> <script type="text/javascript" src="../../_static/underscore.js"></script> <script type="text/javascript" src="../../_static/doctools.js"></script> <script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script> <script type="text/javascript" src="../../_static/js/theme.js"></script> <script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.7/js/bootstrap.min.js" integrity="sha384-Tc5IQib027qvyjSMfHjOMaLkfuWVxZxUPnCJA7l2mCWNIpG9mGCD8wGNIcPD7Txa" crossorigin="anonymous"></script> <script src="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/js/perfect-scrollbar.jquery.min.js"></script> <script src="../../_static/js/paddle_doc_init.js"></script> </body> </html>