gpu_profiling_en.html



<!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 &mdash; 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="Development" href="../../dev/index_en.html"/>
        <link rel="prev" title="API comparision between RNN and hierarchical RNN" href="../rnn/hrnn_rnn_api_compare_en.html"/> 
<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">

  <div class="wy-grid-for-nav">

    
    <nav data-toggle="wy-nav-shift" class="wy-nav-side">
      <div class="wy-side-scroll">
        <div class="wy-side-nav-search">
          

            <a href="../../index_en.html" class="icon icon-home"> PaddlePaddle
          

          </a>

          
<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 class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
          
<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/quickstart_en.html">Quick Start</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../getstarted/concepts/use_concepts_en.html">Basic Concept</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../build_and_install/index_en.html">Install and Build</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/pip_install_en.html">Install using pip</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/docker_install_en.html">Run in Docker Containers</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/build_from_source_en.html">Build from Sources</a></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="../cmd_parameter/index_en.html">Set Command-line Parameters</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/use_case_en.html">Use Case</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/arguments_en.html">Argument Outline</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/detail_introduction_en.html">Detail Description</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../cluster/index_en.html">Distributed Training</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../cluster/preparations_en.html">Preparations</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cluster/cmd_argument_en.html">Command-line arguments</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cluster/multi_cluster/index_en.html">Use different clusters</a><ul>
<li class="toctree-l4"><a class="reference internal" href="../cluster/multi_cluster/fabric_en.html">Fabric</a></li>
<li class="toctree-l4"><a class="reference internal" href="../cluster/multi_cluster/openmpi_en.html">OpenMPI</a></li>
<li class="toctree-l4"><a class="reference internal" href="../cluster/multi_cluster/k8s_en.html">Kubernetes</a></li>
<li class="toctree-l4"><a class="reference internal" href="../cluster/multi_cluster/k8s_aws_en.html">Kubernetes on AWS</a></li>
</ul>
</li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../capi/index_en.html">C-API Prediction Library</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../capi/compile_paddle_lib_en.html">Install and Build</a></li>
<li class="toctree-l3"><a class="reference internal" href="../capi/organization_of_the_inputs_en.html">Input/Output Data Organization</a></li>
<li class="toctree-l3"><a class="reference internal" href="../capi/workflow_of_capi_en.html">C-API Workflow</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../rnn/index_en.html">RNN Models</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../rnn/rnn_config_en.html">RNN Configuration</a></li>
<li class="toctree-l3"><a class="reference internal" href="../rnn/recurrent_group_en.html">Recurrent Group Tutorial</a></li>
<li class="toctree-l3"><a class="reference internal" href="../rnn/hierarchical_layer_en.html">Layers supporting hierarchical sequence as input</a></li>
<li class="toctree-l3"><a class="reference internal" href="../rnn/hrnn_rnn_api_compare_en.html">API comparision between RNN and hierarchical RNN</a></li>
</ul>
</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="../../dev/index_en.html">Development</a><ul>
<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="../../dev/write_docs_en.html">Contribute Documentation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../dev/new_layer_en.html">Write New Layers</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../faq/index_en.html">FAQ</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../faq/build_and_install/index_en.html">Install, Build and Unit test</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/model/index_en.html">Model Configuration</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/parameter/index_en.html">Parameter Setting</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/local/index_en.html">Local Training and Prediction</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/cluster/index_en.html">Cluster Training and Prediction</a></li>
</ul>
</li>
</ul>

</nav>

        </div>
      </div>
    </nav>

    <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">

      
      <nav class="wy-nav-top" role="navigation" aria-label="top navigation">
        <i data-toggle="wy-nav-top" class="fa fa-bars"></i>
        <a href="../../index_en.html">PaddlePaddle</a>
      </nav>


      <div class="wy-nav-content">
        <div class="rst-content">
          

<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
    <li><a href="../../index_en.html">Docs</a> &raquo;</li>
      
          <li><a href="../index_en.html">HOW TO</a> &raquo;</li>
      
    <li>Tune GPU Performance</li>
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../../_sources/howto/optimization/gpu_profiling_en.rst.txt" rel="nofollow"> View page source</a>
          
        
      </li>
  </ul>
  <hr/>
</div>
          <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&#8217;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&#8217;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">&quot;testBilinearFwdBwd&quot;</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">&quot;testBilinearFwdBwd&quot;</span><span class="p">,</span>
        <span class="s">&quot;numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64&quot;</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">&quot;testBilinearFwdBwd&quot;</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">&quot;testBilinearFwdBwd&quot;</span><span class="p">,</span>
</span><span class="hll">        <span class="s">&quot;numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64&quot;</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">&gt; ./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">&quot;numSamples = 10, channels = 16, im</span>
<span class="s2">gSizeX = 64, imgSizeY = 64&quot;</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">&quot;testBilinearFwdBwd&quot;</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">&quot;testBilinearFwdBwd&quot;</span><span class="p">,</span>
        <span class="s">&quot;numSamples = 10, channels = 16, imgSizeX = 64, imgSizeY = 64&quot;</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&#8217;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="../../dev/index_en.html" class="btn btn-neutral float-right" title="Development" accesskey="n">Next <span class="fa fa-arrow-circle-right"></span></a>
      
      
        <a href="../rnn/hrnn_rnn_api_compare_en.html" class="btn btn-neutral" title="API comparision between RNN and hierarchical RNN" accesskey="p"><span class="fa fa-arrow-circle-left"></span> Previous</a>
      
    </div>
  

  <hr/>

  <div role="contentinfo">
    <p>
        &copy; 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
        };
    </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 type="text/javascript">
      jQuery(function () {
          SphinxRtdTheme.StickyNav.enable();
      });
  </script>
   

</body>
</html>