mkl_packed.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>Intel® MKL Packed on PaddlePaddle: Design Doc &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>Intel® MKL Packed on PaddlePaddle: Design Doc</li>
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../../_sources/design/mkl/mkl_packed.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="intel-mkl-packed-on-paddlepaddle-design-doc">
<span id="intel-mkl-packed-on-paddlepaddle-design-doc"></span><h1>Intel® MKL Packed on PaddlePaddle: Design Doc<a class="headerlink" href="#intel-mkl-packed-on-paddlepaddle-design-doc" title="Permalink to this headline">¶</a></h1>
<div class="section" id="contents">
<span id="contents"></span><h2>Contents<a class="headerlink" href="#contents" title="Permalink to this headline">¶</a></h2>
<ul class="simple">
<li><a class="reference external" href="#overview">Overview</a></li>
<li><a class="reference external" href="#key-points">Key Points</a><ul>
<li><a class="reference external" href="#background">Background</a></li>
<li><a class="reference external" href="#solution">Solution</a></li>
</ul>
</li>
<li><a class="reference external" href="#actions">Actions</a><ul>
<li><a class="reference external" href="#cmake">CMake</a></li>
<li><a class="reference external" href="#layers">Layers</a></li>
<li><a class="reference external" href="#unit-tests">Unit Tests</a></li>
<li><a class="reference external" href="#python-api">Python API</a></li>
<li><a class="reference external" href="#benchmarking">Benchmarking</a></li>
</ul>
</li>
</ul>
</div>
<div class="section" id="overview">
<span id="overview"></span><h2>Overview<a class="headerlink" href="#overview" title="Permalink to this headline">¶</a></h2>
<p>我们计划将 Intel® MKL 中引入的 GEMM Packed APIs[<a class="reference external" href="#references">1</a>] 集成到 PaddlePaddle 中，充分发挥英特尔平台的优势，有效提升PaddlePaddle在英特尔架构上的性能。
现阶段的优化主要针对 Recurrent Neural Network（以下简称RNN）相关层（包括<code class="docutils literal"><span class="pre">RecurrentLayer</span></code>, <code class="docutils literal"><span class="pre">GatedRecurrentLayer</span></code>和<code class="docutils literal"><span class="pre">LstmLayer</span></code>）， 以及 PaddlePaddle V1 API。</p>
</div>
<div class="section" id="key-points">
<span id="key-points"></span><h2>Key Points<a class="headerlink" href="#key-points" title="Permalink to this headline">¶</a></h2>
<div class="section" id="background">
<span id="background"></span><h3>Background<a class="headerlink" href="#background" title="Permalink to this headline">¶</a></h3>
<p>目前PaddlePaddle采用了 Intel® MKL库的<a class="reference external" href="https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm">cblas_?gemm</a>函数，这个函数本身会在计算前将原数据转换为更适合英特尔平台的内部格式。</p>
<ol class="simple">
<li>转换耗时 这一数据格式的转换操作（Packing），在问题本身的计算量比较小的时候，显得相对来说较为耗时。例如在DeepSpeech2 [<a class="reference external" href="#references">2</a>] 的Vanilla RNN部分中，矩阵大小是<code class="docutils literal"><span class="pre">batch_size</span> <span class="pre">*</span> <span class="pre">2048</span></code>。</li>
<li>转换冗余 由于在现有的某些情况下（例如RNN），多次调用 cblas_?gemm 会使用相同的原数据，因此，每次调用时对原数据的重复Packing便成为了冗余。</li>
</ol>
<p>为了最大程度减少多次调用 cblas_?gemm 在Packing上的耗时，Intel® MKL 引入了以下四个API:</p>
<div class="toctree-wrapper compound">
<ul>
<li class="toctree-l1"><a class="reference external" href="https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-alloc">cblas</a></li>
<li class="toctree-l1"><a class="reference external" href="https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-pack">cblas</a></li>
<li class="toctree-l1"><a class="reference external" href="https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-compute">cblas</a></li>
<li class="toctree-l1"><a class="reference external" href="https://software.intel.com/en-us/mkl-developer-reference-c-cblas-gemm-free">cblas</a></li>
</ul>
</div>
<p>通过使用这些API，我们可以先完成对原数据的Packing操作，再把已转换为Packed格式的数据传递给那些复用同一数据的gemm_compute函数，从而避免了Packing冗余。</p>
</div>
<div class="section" id="solution">
<span id="solution"></span><h3>Solution<a class="headerlink" href="#solution" title="Permalink to this headline">¶</a></h3>
<p>在RNN的情况下，同一次前向、后向（forward/backward）过程中所有时间步（time step）共享同一个权重（weight）。当只做推断（inference）时，各次前向之间也都使用了相同的权重，没有必要在每次前向中每个时间步的计算时对权重进行重复的Packing操作。</p>
<p>我们通过使用新引入的GEMM Packed APIs，在层初始化的时候，先完成对权重的Packing操作，然后在前向，后向时复用已经转换过的权重，并在每次权重更新后，对新的权重进行转换用于下次迭代。</p>
<ul class="simple">
<li>优化前，对于序列长度（sequence length）为<code class="docutils literal"><span class="pre">T</span></code>的网络模型（model）, <code class="docutils literal"><span class="pre">N</span></code>次迭代执行的转换次数为：<ul>
<li><code class="docutils literal"><span class="pre">inference</span></code>： <code class="docutils literal"><span class="pre">N</span> <span class="pre">*</span> <span class="pre">T</span></code></li>
<li><code class="docutils literal"><span class="pre">training</span></code>： <code class="docutils literal"><span class="pre">2</span> <span class="pre">*</span> <span class="pre">N</span> <span class="pre">*</span> <span class="pre">T</span></code></li>
</ul>
</li>
<li>优化后，对于同样设置的网络模型，其转换次数减少至：<ul>
<li><code class="docutils literal"><span class="pre">inference</span></code>： <code class="docutils literal"><span class="pre">1</span></code></li>
<li><code class="docutils literal"><span class="pre">training</span></code>： <code class="docutils literal"><span class="pre">2</span> <span class="pre">*</span> <span class="pre">N</span></code></li>
</ul>
</li>
</ul>
</div>
</div>
<div class="section" id="actions">
<span id="actions"></span><h2>Actions<a class="headerlink" href="#actions" title="Permalink to this headline">¶</a></h2>
<p>添加的相关文件和目录结构如下：</p>
<div class="highlight-txt"><div class="highlight"><pre><span></span>PaddlePaddle/Paddle
├── ...
└── paddle/
    ├── ...
    └── gserver/
        ├── ...
        ├── layers/
        │   ├── ...
        │   ├── MKLPackedRecurrentLayer.*
        |   ├── MKLPackedGatedRecurrentLayer.*
        |   ├── MKLPackedLstmLayer.*
        |   └── MKLPackedGemm.h
        └── tests/
            ├── ...
            └── test_MKLPacked.cpp
</pre></div>
</div>
<div class="section" id="cmake">
<span id="cmake"></span><h3>CMake<a class="headerlink" href="#cmake" title="Permalink to this headline">¶</a></h3>
<p>在对应的<code class="docutils literal"><span class="pre">CMakeLists.txt</span></code>中根据<code class="docutils literal"><span class="pre">WITH_MKL</span></code>是否打开，来决定是否开启MKL Packed相关功能。</p>
</div>
<div class="section" id="layers">
<span id="layers"></span><h3>Layers<a class="headerlink" href="#layers" title="Permalink to this headline">¶</a></h3>
<p>所有的<code class="docutils literal"><span class="pre">MKLPacked*Layer</span></code>都继承于PaddlePaddle的基类<code class="docutils literal"><span class="pre">Layer</span></code>, 并添加头文件 <code class="docutils literal"><span class="pre">MKLPackedGemm.h</span></code>，该文件对相关GEMM Packed APIs做了封装。</p>
</div>
<div class="section" id="unit-tests">
<span id="unit-tests"></span><h3>Unit Tests<a class="headerlink" href="#unit-tests" title="Permalink to this headline">¶</a></h3>
<p>我们会添加<code class="docutils literal"><span class="pre">test_MKLPacked.cpp</span></code>用于MKL Packed优化后layer的测试。
对于每一个新加的RNN layer，我们会对比如下2个方面：</p>
<ol class="simple">
<li>对比优化后layer自身，sequence mode（<code class="docutils literal"><span class="pre">rnn_use_batch=false</span></code>）与batch mode(<code class="docutils literal"><span class="pre">rnn_use_batch=true</span></code>)的结果。</li>
<li>对比优化后layer与相对应的PaddlePaddle原有layer, 在batch mode下的结果。</li>
</ol>
</div>
<div class="section" id="python-api">
<span id="python-api"></span><h3>Python API<a class="headerlink" href="#python-api" title="Permalink to this headline">¶</a></h3>
<p>计划在<code class="docutils literal"><span class="pre">paddle/utils.Flags</span></code>中添加<code class="docutils literal"><span class="pre">use_mkl_packed</span></code>的flag，用于选择是否使用相关功能，并且当编译时<code class="docutils literal"><span class="pre">WITH_MKL=ON</span></code>的情况下，默认设置为<code class="docutils literal"><span class="pre">true</span></code>。</p>
<p>同时，在<code class="docutils literal"><span class="pre">python/paddle/trainer/config_parser.py</span></code>中对应的layer处，添加<code class="docutils literal"><span class="pre">use_mkl_packed</span></code>这个选择，方便用户在Python端选择是否启用这个功能。</p>
<p>具体实现方式比如：</p>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="n">use_mkl_packed</span> <span class="o">=</span> <span class="nb">bool</span><span class="p">(</span><span class="nb">int</span><span class="p">(</span><span class="n">g_command_config_args</span><span class="o">.</span><span class="n">get</span><span class="p">(</span><span class="s2">&quot;use_mkl_packed&quot;</span><span class="p">,</span> <span class="mi">0</span><span class="p">)))</span>
<span class="k">if</span> <span class="n">use_mkl_packed</span><span class="p">:</span>
    <span class="bp">self</span><span class="o">.</span><span class="n">layer_type</span> <span class="o">=</span> <span class="n">mkl_packed_</span><span class="o">*</span>
</pre></div>
</div>
<p>所有相关的<code class="docutils literal"><span class="pre">layer_type</span></code>会以*mkl_packed_*开头，这些会在<code class="docutils literal"><span class="pre">MKLPacked*Layer</span></code>注册layer的时候保证，以示区分。</p>
</div>
<div class="section" id="benchmarking">
<span id="benchmarking"></span><h3>Benchmarking<a class="headerlink" href="#benchmarking" title="Permalink to this headline">¶</a></h3>
<p>会添加相应的脚本用于测试和对比在使用MKL Packed recurrent layers 前后的网络性能。</p>
</div>
</div>
<div class="section" id="references">
<span id="references"></span><h2>References<a class="headerlink" href="#references" title="Permalink to this headline">¶</a></h2>
<div class="toctree-wrapper compound">
<ul>
<li class="toctree-l1"><a class="reference external" href="https://software.intel.com/en-us/articles/introducing-the-new-packed-apis-for-gemm">Introducing the new Packed APIs for GEMM</a></li>
<li class="toctree-l1"><a class="reference external" href="https://github.com/PaddlePaddle/DeepSpeech#deepspeech2-on-paddlepaddle">DeepSpeech2 on PaddlePaddle</a></li>
</ul>
</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>