kernel_selection.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>Background &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"/> 
<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>
<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"><a class="reference internal" href="../howto/index_en.html">HOW TO</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../howto/cmd_parameter/index_en.html">Set Command-line Parameters</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/cmd_parameter/use_case_en.html">Use Case</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/cmd_parameter/arguments_en.html">Argument Outline</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/cmd_parameter/detail_introduction_en.html">Detail Description</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../howto/cluster/index_en.html">Distributed Training</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/cluster/preparations_en.html">Preparations</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/cluster/cmd_argument_en.html">Command-line arguments</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/cluster/multi_cluster/index_en.html">Use different clusters</a><ul>
<li class="toctree-l4"><a class="reference internal" href="../howto/cluster/multi_cluster/fabric_en.html">Fabric</a></li>
<li class="toctree-l4"><a class="reference internal" href="../howto/cluster/multi_cluster/openmpi_en.html">OpenMPI</a></li>
<li class="toctree-l4"><a class="reference internal" href="../howto/cluster/multi_cluster/k8s_en.html">Kubernetes</a></li>
<li class="toctree-l4"><a class="reference internal" href="../howto/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="../howto/rnn/index_en.html">RNN Models</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/rnn/rnn_config_en.html">RNN Configuration</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/rnn/recurrent_group_en.html">Recurrent Group Tutorial</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/rnn/hierarchical_layer_en.html">Layers supporting hierarchical sequence as input</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/rnn/hrnn_rnn_api_compare_en.html">API comparision between RNN and hierarchical RNN</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../howto/optimization/gpu_profiling_en.html">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>
</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>Background</li>
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../_sources/design/kernel_selection.md.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="background">
<span id="background"></span><h1>Background<a class="headerlink" href="#background" title="Permalink to this headline">¶</a></h1>
<p>Every operator has many kernels because there are multiple data types, places, data layout, library type that Fluid supports. We use the <code class="docutils literal"><span class="pre">OpKernelType</span></code> to describe kernel types that operators can hold.</p>
<p>The <code class="docutils literal"><span class="pre">OpKernelType</span></code> is as follows:</p>
<div class="highlight-cpp"><div class="highlight"><pre><span></span><span class="k">struct</span> <span class="n">OpKernelType</span> <span class="p">{</span>
  <span class="n">Place</span> <span class="n">place_</span><span class="p">;</span>
  <span class="n">DataType</span> <span class="n">data_type_</span><span class="p">;</span>
  <span class="n">DataLayout</span> <span class="n">data_layout_</span><span class="p">;</span>
  <span class="n">LibraryType</span> <span class="n">library_type_</span><span class="p">;</span>
<span class="p">};</span>
</pre></div>
</div>
<ul class="simple">
<li>The <code class="docutils literal"><span class="pre">place_</span></code> is a descriptor of the device, e.g., CPUPlace, CUDAPlace.</li>
<li>The <code class="docutils literal"><span class="pre">data_type_</span></code> is the data type that this kernel performs on, e.g., <code class="docutils literal"><span class="pre">FP32</span></code>, <code class="docutils literal"><span class="pre">INT64</span></code>. Note that one kernel may have inputs with different data types. However, it will be a major <code class="docutils literal"><span class="pre">data_type</span></code>. For example, the <code class="docutils literal"><span class="pre">cross_entropy</span></code> takes <code class="docutils literal"><span class="pre">int64</span></code> as it label, and <code class="docutils literal"><span class="pre">double</span></code>/<code class="docutils literal"><span class="pre">float</span></code> as its input logit and output cost. The major <code class="docutils literal"><span class="pre">data_type</span></code> of <code class="docutils literal"><span class="pre">cross_entropy</span></code> is <code class="docutils literal"><span class="pre">float</span></code> or <code class="docutils literal"><span class="pre">double</span></code>.</li>
<li>The <code class="docutils literal"><span class="pre">data_layout_</span></code> is useful for some computational library. One example is that MKLDNN uses many kinds of layout, such as <code class="docutils literal"><span class="pre">nChw8c</span></code>. Each kind of layout will invoke the different kernel.</li>
<li>The <code class="docutils literal"><span class="pre">library_type_</span></code> describes the computational library, e.g., <code class="docutils literal"><span class="pre">MKLDNN</span></code>, <code class="docutils literal"><span class="pre">CUDNN</span></code>.</li>
</ul>
</div>
<div class="section" id="problem">
<span id="problem"></span><h1>Problem<a class="headerlink" href="#problem" title="Permalink to this headline">¶</a></h1>
<p>We register a kernel for every operator and every kernel type ideally. However, it is impracticable for the following situations.</p>
<ol class="simple">
<li>Some operators, like CRF, are complicated and inefficient to be implemented on GPU. The CRF operator will only have a CPU kernel.</li>
<li>Some operators will take too many memory. It is better to force them into CPU. However, the rest of operators in this neural network will be performed on GPU, i.e., model parallel problem.</li>
<li>Some layout and place are particular. One example is that MKLDNN uses <code class="docutils literal"><span class="pre">nChw8</span></code> and there is no other library uses <code class="docutils literal"><span class="pre">nChw8c</span></code>.</li>
</ol>
<p>Take one situation to give a detailed explanation, if we have two Operators: OP1 and OP2, OP1 has one output <code class="docutils literal"><span class="pre">op1_to_op2</span></code>, and <code class="docutils literal"><span class="pre">op1_to_op2</span></code> is the input of OP2.</p>
<p>If OP1 and OP2 run on the same place(for example CPUPlace), then <code class="docutils literal"><span class="pre">op1_2_op2</span></code> can be used directly by OP2.</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">OP1</span><span class="p">(</span><span class="n">CPUPlace</span><span class="p">)</span>
     <span class="o">|</span>
 <span class="n">op1_2_op2</span>
     <span class="o">|</span>
<span class="n">OP2</span><span class="p">(</span><span class="n">CPUPlace</span><span class="p">)</span>
</pre></div>
</div>
<p>If OP1 and OP2 run one different place, then OP2 cannot <code class="docutils literal"><span class="pre">use</span> <span class="pre">op1_2_op2</span></code> directly.</p>
<p>Problems under these situations are similar. We can formalize this problem as follow.</p>
<p>We register kernels with types $KT = {kt_1, kt_2, kt_3, ...}$ for one operator. The inputs of this operator should be run on kernel type $kt_{?}$, which the $kt_{?} \notin KT$. How to cast the input of this operator from $kt_{?}$ to any of kernel type in $KT$.</p>
</div>
<div class="section" id="solution-data-transform">
<span id="solution-data-transform"></span><h1>Solution: data transform<a class="headerlink" href="#solution-data-transform" title="Permalink to this headline">¶</a></h1>
<p>It is clear that transforming inputs of an operator to adapt another kernel type is not related to the particular operator. So we should register these transformation methods as global methods.</p>
<p>We can infer kernel type for each input of an operator. We let this kernel type as <code class="docutils literal"><span class="pre">actual</span> <span class="pre">kernel</span> <span class="pre">type</span> <span class="pre">for</span> <span class="pre">var</span></code>, which means this kernel type is the kernel type that can process this input variable.</p>
<p>We can get a kernel type by 1) The configuration of operator description. (Users may want to force use <code class="docutils literal"><span class="pre">MKL</span></code> for <code class="docutils literal"><span class="pre">conv</span></code> operator). 2) The place of the current executor. (Executor is running on GPU). This kernel type is what we expect the operator will be performed on. We let this kernel type as <code class="docutils literal"><span class="pre">expect</span> <span class="pre">kernel</span> <span class="pre">type</span></code>.</p>
<p>We transform the input data from <code class="docutils literal"><span class="pre">actual</span></code> to <code class="docutils literal"><span class="pre">expect</span></code> if the actual kernel type is not as same as expect kernel type.</p>
<p>The algorithm is described as following</p>
<div class="highlight-cpp"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="n">OperatorWithKernel</span><span class="o">::</span><span class="n">Run</span><span class="p">(</span>
        <span class="k">const</span> <span class="n">Scope</span><span class="o">&amp;</span> <span class="n">scope</span><span class="p">,</span>
        <span class="k">const</span> <span class="n">platform</span><span class="o">::</span><span class="n">Place</span><span class="o">&amp;</span> <span class="n">place</span><span class="p">)</span> <span class="k">const</span> <span class="p">{</span>
  <span class="n">ExecutionContext</span> <span class="n">ctx</span><span class="p">(...);</span>
  <span class="k">auto</span> <span class="n">expected_kernel_key</span> <span class="o">=</span> <span class="k">this</span><span class="o">-&gt;</span><span class="n">GetExpectedKernelType</span><span class="p">(</span><span class="n">ctx</span><span class="p">);</span>

  <span class="n">Scope</span><span class="o">&amp;</span> <span class="n">new_scope</span> <span class="o">=</span> <span class="n">scope</span><span class="p">.</span><span class="n">NewScope</span><span class="p">();</span>

  <span class="k">for</span> <span class="p">(</span><span class="k">auto</span><span class="o">&amp;</span> <span class="nl">var_name</span> <span class="p">:</span> <span class="k">this</span><span class="o">-&gt;</span><span class="n">Inputs</span><span class="p">())</span> <span class="p">{</span>
    <span class="k">auto</span><span class="o">*</span> <span class="n">tensor_in</span> <span class="o">=</span> <span class="n">GetTensor</span><span class="p">(</span><span class="n">var_name</span><span class="p">);</span>
    <span class="k">auto</span> <span class="n">kernel_type_for_var</span> <span class="o">=</span> <span class="k">this</span><span class="o">-&gt;</span><span class="n">GetKernelTypeForVar</span><span class="p">(...);</span>
    <span class="k">if</span> <span class="p">(</span><span class="n">kernel_type_for_var</span><span class="p">.</span><span class="n">place_</span> <span class="o">!=</span> <span class="n">expected_kernel_key</span><span class="p">.</span><span class="n">place_</span><span class="p">)</span> <span class="p">{</span>
      <span class="k">auto</span><span class="o">*</span> <span class="n">trans_var</span> <span class="o">=</span> <span class="n">new_scope</span><span class="p">.</span><span class="n">Var</span><span class="p">(</span><span class="n">var_name</span><span class="p">);</span>
      <span class="k">auto</span><span class="o">*</span> <span class="n">out</span> <span class="o">=</span> <span class="n">DataTransform</span><span class="p">(</span><span class="n">expected_kernel_key</span><span class="p">,</span>
                                <span class="n">kernel_type_for_var</span><span class="p">,</span>
                                <span class="o">*</span><span class="n">tensor_in</span><span class="p">);</span>
      <span class="n">CopyVariableWithTensor</span><span class="p">(...);</span>
    <span class="p">}</span>
  <span class="p">}</span>

  <span class="k">auto</span> <span class="n">kernel</span> <span class="o">=</span> <span class="n">kernels</span><span class="p">.</span><span class="n">find</span><span class="p">(</span><span class="n">expected_kernel_key</span><span class="p">);</span>
  <span class="n">kernel</span><span class="o">-&gt;</span><span class="n">Compute</span><span class="p">(</span><span class="n">ExecutionContext</span><span class="p">(...));</span>
<span class="p">}</span>
</pre></div>
</div>
<p>then the actual process for the multi-device above will be:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">OP1</span><span class="p">(</span><span class="n">CPUPlace</span><span class="p">)</span>
     <span class="o">|</span>
<span class="n">op1_2_op2</span><span class="p">(</span><span class="n">on</span> <span class="n">CPU</span><span class="p">)</span>
     <span class="o">|</span>
<span class="p">[</span><span class="n">transform</span><span class="p">](</span><span class="kn">from</span> <span class="nn">CPU</span> <span class="n">to</span> <span class="n">GPU</span><span class="p">)</span>
     <span class="o">|</span>
<span class="n">op1_2_op2</span><span class="p">(</span><span class="n">on</span> <span class="n">GPU</span><span class="p">)</span>
     <span class="o">|</span>
<span class="n">OP2</span><span class="p">(</span><span class="n">CUDAPlace</span><span class="p">)</span>
</pre></div>
</div>
</div>


           </div>
          </div>
          <footer>
  

  <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>