Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
切换分支/标签
查找文件 普通视图历史永久链接Permalink
gpu_profiling_en.html 39.1 KB
Newer Older
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
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 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 


<!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="API" href="../../api/index_en.html"/>
        <link rel="prev" title="RNN Configuration" href="../deep_model/rnn/rnn_config_en.html"/> 

  <link rel="stylesheet" href="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/css/perfect-scrollbar.min.css" type="text/css" />
  <link rel="stylesheet" href="../../_static/css/override.css" type="text/css" />
  <script>
  var _hmt = _hmt || [];
  (function() {
    var hm = document.createElement("script");
    hm.src = "//hm.baidu.com/hm.js?b9a314ab40d04d805655aab1deee08ba";
    var s = document.getElementsByTagName("script")[0]; 
    s.parentNode.insertBefore(hm, s);
  })();
  </script>

  

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

</head>

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

  
  <header class="site-header">
    <div class="site-logo">
      <a href="/"><img src="../../_static/images/PP_w.png"></a>
    </div>
    <div class="site-nav-links">
      <div class="site-menu">
        <a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Folk me on Github</a>
        <div class="language-switcher dropdown">
          <a type="button" data-toggle="dropdown">
            <span>English</span>
            <i class="fa fa-angle-up"></i>
            <i class="fa fa-angle-down"></i>
          </a>
          <ul class="dropdown-menu">
            <li><a href="/doc_cn">中文</a></li>
            <li><a href="/doc">English</a></li>
          </ul>
        </div>
        <ul class="site-page-links">
          <li><a>Home</a></li>
          <li><a>Get Started</a></li>
          <li class="active"><a>Documentation</a></li>
          <li><a>About Us</a></li>
        </ul>
      </div>
      <div class="doc-module">
        
        <ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../tutorials/index_en.html">TUTORIALS</a></li>
<li class="toctree-l1 current"><a class="reference internal" href="../index_en.html">HOW TO</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../api/index_en.html">API</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../about/index_en.html">ABOUT</a></li>
</ul>

        
<div role="search">
  <form id="rtd-search-form" class="wy-form" action="../../search.html" method="get">
    <input type="text" name="q" placeholder="Search docs" />
    <input type="hidden" name="check_keywords" value="yes" />
    <input type="hidden" name="area" value="default" />
  </form>
</div>        
      </div>
    </div>
  </header>
  
  <div class="main-content-wrap">

    
    <nav class="doc-menu-vertical" role="navigation">
        
          
          <ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../getstarted/build_and_install/index_en.html">Install and Build</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/docker_install_en.html">PaddlePaddle in Docker Containers</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/ubuntu_install_en.html">Debian Package installation guide</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../getstarted/build_and_install/build_from_source_en.html">Installing from Sources</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../../getstarted/basic_usage/index_en.html">Simple Linear Regression</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../tutorials/index_en.html">TUTORIALS</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/quick_start/index_en.html">Quick Start</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/rec/ml_regression_en.html">MovieLens Regression</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/image_classification/index_en.html">Image Classification</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/sentiment_analysis/index_en.html">Sentiment Analysis</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/semantic_role_labeling/index_en.html">Semantic Role Labeling</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/text_generation/index_en.html">Text Generation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/gan/index_en.html">Image Auto-Generation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/imagenet_model/resnet_model_en.html">ImageNet: ResNet</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../tutorials/embedding_model/index_en.html">Embedding: Chinese Word</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="../usage/cmd_parameter/index_en.html">Set Command-line Parameters</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/use_case_en.html">Use Case</a></li>
<li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/arguments_en.html">Argument Outline</a></li>
<li class="toctree-l3"><a class="reference internal" href="../usage/cmd_parameter/detail_introduction_en.html">Detail Description</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../usage/cluster/cluster_train_en.html">Run Distributed Training</a></li>
<li class="toctree-l2"><a class="reference internal" href="../usage/k8s/k8s_en.html">Paddle On Kubernetes</a></li>
<li class="toctree-l2"><a class="reference internal" href="../usage/k8s/k8s_aws_en.html">Distributed PaddlePaddle Training on AWS with Kubernetes</a></li>
<li class="toctree-l2"><a class="reference internal" href="../dev/new_layer_en.html">Write New Layers</a></li>
<li class="toctree-l2"><a class="reference internal" href="../dev/contribute_to_paddle_en.html">Contribute Code</a></li>
<li class="toctree-l2"><a class="reference internal" href="../deep_model/rnn/index_en.html">RNN Models</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../deep_model/rnn/rnn_config_en.html">RNN Configuration</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="../../api/index_en.html">API</a><ul>
T
Deploy to GitHub Pages: 5f2cbce4190fb83883fed45668b5e4d87a577589  
Travis CI 已提交
158 159 160 
<li class="toctree-l2"><a class="reference internal" href="../../api/v2/model_configs.html">Configuration Related API</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../api/v2/data.html">Data Related API</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../api/v2/run_logic.html">Trainer API</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../about/index_en.html">ABOUT</a></li>
</ul>

        
    </nav>
    
    <nav class="local-toc"><ul>
<li><a class="reference internal" href="#">Tune GPU Performance</a><ul>
<li><a class="reference internal" href="#what-s-profiling">What&#8217;s profiling?</a></li>
<li><a class="reference internal" href="#why-we-need-profiling">Why we need profiling?</a></li>
<li><a class="reference internal" href="#how-to-do-profiling">How to do profiling?</a></li>
<li><a class="reference internal" href="#profiler-tools">Profiler Tools</a></li>
<li><a class="reference internal" href="#hands-on-approach">Hands-on Approach</a><ul>
<li><a class="reference internal" href="#built-in-timer">Built-in Timer</a></li>
<li><a class="reference internal" href="#nvprof-profiler">nvprof profiler</a></li>
<li><a class="reference internal" href="#nvvp-profiler">nvvp profiler</a></li>
</ul>
</li>
<li><a class="reference internal" href="#profiling-tips">Profiling tips</a></li>
<li><a class="reference internal" href="#reference">Reference</a></li>
</ul>
</li>
</ul>
</nav>
    
    <section class="doc-content-wrap">

      

 







<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
      
        <li><a href="../index_en.html">HOW TO</a> > </li>
      
    <li>Tune GPU Performance</li>
  </ul>
</div>
      
      <div class="wy-nav-content" id="doc-content">
        <div class="rst-content">
          <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
           <div itemprop="articleBody">
            
  <div class="section" id="tune-gpu-performance">
<h1><a class="toc-backref" href="#id1">Tune GPU Performance</a><a class="headerlink" href="#tune-gpu-performance" title="Permalink to this headline"></a></h1>
<div class="contents topic" id="contents">
<p class="topic-title first">Contents</p>
<ul class="simple">
<li><a class="reference internal" href="#tune-gpu-performance" id="id1">Tune GPU Performance</a><ul>
<li><a class="reference internal" href="#what-s-profiling" id="id2">What&#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="../../api/index_en.html" class="btn btn-neutral float-right" title="API" accesskey="n">Next <span class="fa fa-arrow-circle-right"></span></a>
      
      
        <a href="../deep_model/rnn/rnn_config_en.html" class="btn btn-neutral" title="RNN Configuration" 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://cdn.mathjax.org/mathjax/latest/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>
       
  

  
  
    <script type="text/javascript" src="../../_static/js/theme.js"></script>
  
  
  <script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.7/js/bootstrap.min.js" integrity="sha384-Tc5IQib027qvyjSMfHjOMaLkfuWVxZxUPnCJA7l2mCWNIpG9mGCD8wGNIcPD7Txa" crossorigin="anonymous"></script>
  <script src="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/js/perfect-scrollbar.jquery.min.js"></script>
  <script src="../../_static/js/paddle_doc_init.js"></script> 

</body>
</html>