memory_optimization.html 22.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30


<!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>Memory Optimization &mdash; PaddlePaddle  文档</title>
  

  
  

  

  
  
    

  

  
  
    <link rel="stylesheet" href="../_static/css/theme.css" type="text/css" />
  

  
31

32 33 34 35 36 37 38 39 40 41 42 43 44
  
        <link rel="index" title="索引"
              href="../genindex.html"/>
        <link rel="search" title="搜索" href="../search.html"/>
    <link rel="top" title="PaddlePaddle  文档" href="../index.html"/> 

  
  <script src="../_static/js/modernizr.min.js"></script>

</head>

<body class="wy-body-for-nav" role="document">

45 46 47 48 49 50 51 52 53 54 55 56 57
  <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_cn.html" class="icon icon-home"> PaddlePaddle
          

          
58 59
          </a>

60 61 62 63 64 65
          
            
            
          

          
66 67 68 69 70 71
<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>
72
</div>
73 74

          
75 76 77 78 79 80 81 82 83 84 85 86
        </div>

        <div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
          
            
            
                <ul>
<li class="toctree-l1"><a class="reference internal" href="../getstarted/index_cn.html">新手入门</a></li>
<li class="toctree-l1"><a class="reference internal" href="../build_and_install/index_cn.html">安装与编译</a></li>
<li class="toctree-l1"><a class="reference internal" href="../howto/index_cn.html">进阶使用</a></li>
<li class="toctree-l1"><a class="reference internal" href="../dev/index_cn.html">开发标准</a></li>
<li class="toctree-l1"><a class="reference internal" href="../faq/index_cn.html">FAQ</a></li>
87 88
</ul>

89 90 91 92
            
          
        </div>
      </div>
93 94
    </nav>

95
    <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">
96

97 98 99 100 101
      
      <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_cn.html">PaddlePaddle</a>
      </nav>
102 103


104 105 106 107
      
      <div class="wy-nav-content">
        <div class="rst-content">
          
108

109
 
110 111 112 113 114



<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
115
    <li><a href="../index_cn.html">Docs</a> &raquo;</li>
116 117
      
    <li>Memory Optimization</li>
118 119 120 121 122 123 124
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../_sources/design/memory_optimization.md.txt" rel="nofollow"> View page source</a>
          
        
      </li>
125
  </ul>
126
  <hr/>
127 128 129 130 131 132 133 134 135 136
</div>
          <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
           <div itemprop="articleBody">
            
  <div class="section" id="memory-optimization">
<span id="memory-optimization"></span><h1>Memory Optimization<a class="headerlink" href="#memory-optimization" title="永久链接至标题"></a></h1>
<div class="section" id="problem">
<span id="problem"></span><h2>Problem<a class="headerlink" href="#problem" title="永久链接至标题"></a></h2>
<p>In a lecture from Andrew Ng, he attributes the recent sucess of AI due to a combination of these:</p>
<ul class="simple">
137 138 139
<li>Availability of Big Data</li>
<li>Supercomputing power to process this Big Data over very large neural networks</li>
<li>Modern algorithms</li>
140 141 142
</ul>
<p>Following graph shows the details:</p>
<p><img alt="" src="../_images/deep_learning.png" /></p>
143
<p>Larger model usually bring better performance. However, GPU memory is limited. For example, the memory size of a GTX TITAN X is only 12GB. To train complex and large models, we have to take care of memory usage. Besides, memory optimization is also necessary in both online/mobile inference.</p>
144 145 146 147 148
</div>
<div class="section" id="solution">
<span id="solution"></span><h2>Solution<a class="headerlink" href="#solution" title="永久链接至标题"></a></h2>
<div class="section" id="basic-strategy">
<span id="basic-strategy"></span><h3>Basic Strategy<a class="headerlink" href="#basic-strategy" title="永久链接至标题"></a></h3>
149
<p>There are some basic strategies to improve memory usage, including in-place operations and memory sharing.</p>
150 151 152 153
<div class="section" id="in-place-operation">
<span id="in-place-operation"></span><h4>In-place Operation<a class="headerlink" href="#in-place-operation" title="永久链接至标题"></a></h4>
<p>In a relu activation operator:</p>
<p>$y = \max(x, 0)$</p>
154
<p>If the variable x is not used in any other operator, we can make an in-place operation. In other words, the memory block of variable y and variable x will be the same. In-place operations will save 50% memory occupancy immediately.</p>
155 156 157 158 159 160 161 162 163 164
</div>
<div class="section" id="memory-sharing">
<span id="memory-sharing"></span><h4>Memory Sharing<a class="headerlink" href="#memory-sharing" title="永久链接至标题"></a></h4>
<p>Not all operators support in-place operations. Memory sharing is a more general strategy.</p>
<p>Following is an example:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">a</span> <span class="o">=</span> <span class="n">op1</span><span class="p">(</span><span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">);</span>
<span class="n">d</span> <span class="o">=</span> <span class="n">op2</span><span class="p">(</span><span class="n">a</span><span class="p">)</span>
<span class="n">e</span> <span class="o">=</span> <span class="n">op3</span><span class="p">(</span><span class="n">d</span><span class="p">,</span> <span class="n">f</span><span class="p">)</span>
</pre></div>
</div>
165
<p>In this case, variable a is no longer used, and op2 does not support in-place operation. After op2 finishes, we can put the memory of variable a to a memory pool. Then, variable e can share the memory of variable a from the pool.</p>
166 167 168 169
</div>
</div>
<div class="section" id="live-variable-analysis">
<span id="live-variable-analysis"></span><h3>Live Variable Analysis<a class="headerlink" href="#live-variable-analysis" title="永久链接至标题"></a></h3>
170
<p>It&#8217;s not enough to only have some basic strategies. The pre-requisite of memory optimization is to know if a variable is still &#8220;live&#8221; after an operation.</p>
171
<p>In our design, the neural network topology is defined as a program. Luckily, <a class="reference external" href="https://en.wikipedia.org/wiki/Live_variable_analysis">live variable analysis</a> is a classic problem in compilers which can be used in many stages, such as register allocation.</p>
172 173
<p>In compilers, the front end of the compiler translates programs into an intermediate language with an unbounded number of temporary variables. This program must run on a machine with a bounded number of registers. Two temporary variables a and b can fit into the same register, if a and b are never &#8220;in use&#8221; at the same time. Thus, many temporary variables can fit in few registers; if they don&#8217;t all fit, the excess tempory variables can be kept in memory.</p>
<p>Therefore, the compiler needs to analyze the intermediate-representation program to determine which temporary variables are in use at the same time. We say a variable is &#8220;live&#8221; if it holds a value that may be needed in the future, so this analysis is called liveness analysis.</p>
174 175 176 177 178 179 180
<p>We can leran these techniques from compilers. There are mainly two stages to make live variable analysis:</p>
<ul class="simple">
<li>construct a control flow graph</li>
<li>solve the dataflow equations</li>
</ul>
<div class="section" id="control-flow-graph">
<span id="control-flow-graph"></span><h4>Control Flow Graph<a class="headerlink" href="#control-flow-graph" title="永久链接至标题"></a></h4>
181
<p>To perform analysis on a program, it is often useful to make a control flow graph. A <a class="reference external" href="https://en.wikipedia.org/wiki/Control_flow_graph">control flow graph</a> (CFG) in computer science is a representation, using graph notation, of all paths that might be traversed through a program during its execution. Each statement in the program is a node in the flow graph; if statemment x can be followed by statement y, there is an egde from x to y.</p>
182 183 184 185 186
<p>Following is the flow graph for a simple loop.</p>
<p><img alt="" src="../_images/control_flow_graph.png" /></p>
</div>
<div class="section" id="dataflow-analysis">
<span id="dataflow-analysis"></span><h4>Dataflow Analysis<a class="headerlink" href="#dataflow-analysis" title="永久链接至标题"></a></h4>
187
<p>Liveness of variable &#8220;flows&#8221; around the edges of the control flow graph; determining the live range of each variable is an example of a dataflow problem. <a class="reference external" href="https://en.wikipedia.org/wiki/Data-flow_analysis">Dataflow analysis</a> is a technique for gathering information about the possible set of values calculated at various points in a computer program.</p>
188 189 190 191
<p>A simple way to perform data-flow analysis of programs is to set up dataflow equations for each node of the control flow graph and solve them by repeatedly calculating the output from the input locally at each node until the whole system stabilizes.</p>
<ul class="simple">
<li>Flow Graph Terminology</li>
</ul>
192
<p>A flow graph node has out-edges that lead to sucessor nodes, and in-edges that come from predecessor nodes. The set <em>pred[n]</em> is all the predecessors of node n, and <em>succ[n]</em> is the set of sucessors.
193 194 195 196
In former control flow graph, the out-edges of node 5 are 5 &#8211;&gt; 6 and 5 &#8211;&gt; 2, and <em>succ[5]</em> = {2, 6}. The in-edges of 2 are 5 &#8211;&gt; 2 and 1 &#8211;&gt; 2, and <em>pred[2]</em> = {1, 5}.</p>
<ul class="simple">
<li>Uses and Defs</li>
</ul>
197
<p>An assignmemt to a variable or temporary defines that variable. An occurence of a variable on the right-hand side of an assginment(or in other expressions) uses the variable. We can define the <em>def</em> of a variable as the set of graph nodes that define it; or the <em>def</em> of a graph node as the set of variables that it defines; and the similarly for the <em>use</em> of a variable or graph node. In former control flow graph, <em>def(3)</em> = {c}, <em>use(3)</em> = {b, c}.</p>
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
<ul class="simple">
<li>Liveness</li>
</ul>
<p>A variable is <em>live</em> on an edge if there is a directed path from that edge to a <em>use</em> of the variable that does not go through any <em>def</em>. A variable is <em>live-in</em> at a node if it is live on any of the in-edges of that node; it is <em>live-out</em> at a node if it is live on any of the out-edges of the node.</p>
<p>The calcution of liveness can be solved by iteration until a fixed pointer is reached. Following is the recursive formula:</p>
<p><img alt="" src="../_images/dataflow_equations.png" /></p>
</div>
</div>
<div class="section" id="memory-optimization-transpiler">
<span id="memory-optimization-transpiler"></span><h3>Memory optimization transpiler<a class="headerlink" href="#memory-optimization-transpiler" title="永久链接至标题"></a></h3>
<p>At last, we take basic strategy and liveness analysis techniques learning from compilers to implement our memory optimization transpiler.</p>
<div class="section" id="add-in-place-attribute">
<span id="add-in-place-attribute"></span><h4>add in-place attribute<a class="headerlink" href="#add-in-place-attribute" title="永久链接至标题"></a></h4>
<p>In-place is a built-in attribute of an operator. Since we treat in-place and other operators differently, we have to add an in-place attribute for every operator.</p>
</div>
<div class="section" id="contruct-control-flow-graph">
<span id="contruct-control-flow-graph"></span><h4>contruct control flow graph<a class="headerlink" href="#contruct-control-flow-graph" title="永久链接至标题"></a></h4>
215
<p>Following is the ProgramDesc protobuf of <a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book/test_machine_translation.py">machine translation</a> example.</p>
216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
<ul class="simple">
<li>Block0:</li>
</ul>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">lookup_table</span>
<span class="n">mul</span>
<span class="o">...</span>
<span class="k">while</span><span class="p">(</span><span class="n">sub</span><span class="o">-</span><span class="n">block</span> <span class="n">idx</span> <span class="mi">1</span><span class="p">)</span>
<span class="o">...</span>
<span class="n">array_to_lod_tensor</span>
<span class="n">cross_entropy</span>
<span class="o">...</span>
<span class="n">while_grad</span><span class="p">(</span><span class="n">sub</span><span class="o">-</span><span class="n">block</span> <span class="n">idx</span> <span class="mi">2</span><span class="p">)</span>
<span class="n">read_from_array</span>
<span class="n">array_to_lod_tensor</span>
<span class="o">...</span>
</pre></div>
</div>
<ul class="simple">
<li>Block1</li>
</ul>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">read_from_array</span>
<span class="n">read_from_array</span>
<span class="o">...</span>
<span class="n">write_to_array</span>
<span class="n">increment</span>
<span class="n">write_to_array</span>
<span class="n">less_than</span>
</pre></div>
</div>
<ul class="simple">
<li>Block2</li>
</ul>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">read_from_array</span>
<span class="n">increment</span>
<span class="o">...</span>
<span class="n">write_to_array</span>
<span class="n">write_to_array</span>
</pre></div>
</div>
<p>We can transfer all the operators and variables in ProgramDesc to build a control flow graph.</p>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">ControlFlowGraph</span><span class="p">(</span><span class="nb">object</span><span class="p">):</span>
    <span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">Program</span><span class="p">):</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_sucessors</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_presucessors</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_uses</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_defs</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_live_in</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_live_out</span> <span class="o">=</span> <span class="n">defaultdict</span><span class="p">(</span><span class="nb">set</span><span class="p">)</span>
        <span class="bp">self</span><span class="o">.</span><span class="n">_program</span> <span class="o">=</span> <span class="n">Program</span>
    
    <span class="k">def</span> <span class="nf">build</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
        <span class="k">pass</span>
    
    <span class="k">def</span> <span class="nf">dataflow_analysis</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
        <span class="k">pass</span>
        
    <span class="k">def</span> <span class="nf">memory_optimization</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
        <span class="k">pass</span>
        
    <span class="k">def</span> <span class="nf">get_program</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
        <span class="k">return</span> <span class="bp">self</span><span class="o">.</span><span class="n">_program</span>
</pre></div>
</div>
</div>
<div class="section" id="make-dataflow-analysis">
281 282
<span id="make-dataflow-analysis"></span><h4>Make dataflow analysis<a class="headerlink" href="#make-dataflow-analysis" title="永久链接至标题"></a></h4>
<p>We follow the guide from compilers and try to solve the dataflow equation to get liveness of every variable. If the live-in of an operator node is different from the live-out, then we can make memory sharing.</p>
283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360
<p>For example:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">a</span> <span class="o">=</span> <span class="n">op1</span><span class="p">(</span><span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">);</span>
<span class="n">d</span> <span class="o">=</span> <span class="n">op2</span><span class="p">(</span><span class="n">a</span><span class="p">)</span>
<span class="n">e</span> <span class="o">=</span> <span class="n">op3</span><span class="p">(</span><span class="n">d</span><span class="p">,</span> <span class="n">f</span><span class="p">)</span>
</pre></div>
</div>
<p>The dataflow analysis result is:</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">live_in</span><span class="p">(</span><span class="n">op1</span><span class="p">)</span> <span class="o">=</span> <span class="p">{</span><span class="n">b</span><span class="p">,</span> <span class="n">c</span><span class="p">,</span> <span class="n">f</span><span class="p">}</span>
<span class="n">live_out</span><span class="p">(</span><span class="n">op1</span><span class="p">)</span> <span class="o">=</span> <span class="p">{</span><span class="n">a</span><span class="p">,</span> <span class="n">f</span><span class="p">}</span>

<span class="n">live_in</span><span class="p">(</span><span class="n">op2</span><span class="p">)</span> <span class="o">=</span> <span class="p">{</span><span class="n">a</span><span class="p">,</span> <span class="n">f</span><span class="p">}</span>
<span class="n">live_out</span><span class="p">(</span><span class="n">op2</span><span class="p">)</span> <span class="o">=</span> <span class="p">{</span><span class="n">d</span><span class="p">,</span> <span class="n">f</span><span class="p">}</span>

<span class="n">live_in</span><span class="p">(</span><span class="n">op3</span><span class="p">)</span> <span class="o">=</span> <span class="p">{</span><span class="n">d</span><span class="p">,</span> <span class="n">f</span><span class="p">}</span>
<span class="n">live_out</span><span class="p">(</span><span class="n">op3</span><span class="p">)</span> <span class="o">=</span> <span class="p">{}</span>
</pre></div>
</div>
<p>After op1, we can process variable b and variable c; After op2, we can process variable a. After op3, we can process variable d and variable f.</p>
</div>
<div class="section" id="memory-sharing-policy">
<span id="memory-sharing-policy"></span><h4>memory sharing policy<a class="headerlink" href="#memory-sharing-policy" title="永久链接至标题"></a></h4>
<p>A memory pool will be mantained in the stage of memory optimization. Each operator node will be scanned to determine memory optimization is done or not. If an operator satifies the requirement, following policy will be taken to handle input/output variables.</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="k">if</span> <span class="n">op</span><span class="o">.</span><span class="n">support_inplace</span><span class="p">():</span>
    <span class="n">i</span> <span class="o">--&gt;</span> <span class="n">pool</span>
    <span class="n">pool</span> <span class="o">--&gt;</span> <span class="n">o</span>
<span class="k">else</span><span class="p">:</span>
    <span class="n">pool</span> <span class="o">--&gt;</span> <span class="n">o</span>
    <span class="n">i</span> <span class="o">--&gt;</span> <span class="n">pool</span>
</pre></div>
</div>
</div>
</div>
</div>
<div class="section" id="reference">
<span id="reference"></span><h2>Reference<a class="headerlink" href="#reference" title="永久链接至标题"></a></h2>
<ul class="simple">
<li><a class="reference external" href="https://manavsehgal.com/lecture-notes-from-artificial-intelligence-is-the-new-electricity-by-andrew-ng-4712dcbf26e5">Lecture Notes From Artificial Intelligence Is The New Electricity By Andrew Ng</a></li>
<li>Modern compiler implementation in ML, by Andrew W. Appel</li>
<li><a class="reference external" href="https://mxnet.incubator.apache.org/architecture/note_memory.html">Optimizing Memory Consumption in Deep learning</a></li>
</ul>
</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',
361
            HAS_SOURCE:  true
362 363 364 365 366 367 368
        };
    </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="../_static/translations.js"></script>
      <script type="text/javascript" src="https://cdn.bootcss.com/mathjax/2.7.0/MathJax.js"></script>
369

370 371 372 373 374 375
  

  
  
    <script type="text/javascript" src="../_static/js/theme.js"></script>
  
376

377
  
378 379 380 381 382 383 384
  
  <script type="text/javascript">
      jQuery(function () {
          SphinxRtdTheme.StickyNav.enable();
      });
  </script>
   
385 386 387

</body>
</html>