<!-- HTML header for doxygen 1.8.14-->
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml">
<head>
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=9"/>
<meta name="generator" content="Doxygen 1.8.11"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>Caffe2 - C++ API: caffe2/mobile/contrib/nnapi/nnapi_benchmark.cc Source File</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
<link rel="icon" href="/static/favicon.png" type="image/x-icon">
<script type="text/javascript" src="jquery.js"></script>
<script type="text/javascript" src="dynsections.js"></script>
<link href="search/search.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="search/searchdata.js"></script>
<script type="text/javascript" src="search/search.js"></script>
<script type="text/javascript">
$(document).ready(function() { init_search(); });
</script>
<link href="stylesheet.css" rel="stylesheet" type="text/css" />
<link href="main.css" rel="stylesheet" type="text/css"/>
</head>
<body>
<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
<div id="titlearea">
<table cellspacing="0" cellpadding="0">
<tbody>
<tr style="height: 56px;">
<td id="projectlogo" width="56"><a href="/"><img alt="Logo" src="Caffe2-with-name-55-tall.png"/></a></td>
<td id="projectalign" style="padding-left: 0.5em;">
<div id="projectname">Caffe2 - C++ API
</div>
<div id="projectbrief">A deep learning, cross platform ML framework</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.8.11 -->
<script type="text/javascript">
var searchBox = new SearchBox("searchBox", "search",false,'Search');
</script>
<div id="navrow1" class="tabs">
<ul class="tablist">
<li><a href="pages.html"><span>Related Pages</span></a></li>
<li><a href="modules.html"><span>Modules</span></a></li>
<li><a href="annotated.html"><span>Data Structures</span></a></li>
<li class="current"><a href="files.html"><span>Files</span></a></li>
<li><a href="/doxygen-c/html/classes.html"><span>C++ API</span></a></li>
<li><a href="/doxygen-python/html/annotated.html"><span>Python API</span></a></li>
<li><a href="https://github.com/caffe2/caffe2"><span>GitHub</span></a></li>
<li>
<div id="MSearchBox" class="MSearchBoxInactive">
<span class="left">
<img id="MSearchSelect" src="search/mag_sel.png"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
alt=""/>
<input type="text" id="MSearchField" value="Search" accesskey="S"
onfocus="searchBox.OnSearchFieldFocus(true)"
onblur="searchBox.OnSearchFieldFocus(false)"
onkeyup="searchBox.OnSearchFieldChange(event)"/>
</span><span class="right">
<a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.png" alt=""/></a>
</span>
</div>
</li>
</ul>
</div>
<div id="navrow2" class="tabs2">
<ul class="tablist">
<li><a href="files.html"><span>File List</span></a></li>
<li><a href="globals.html"><span>Globals</span></a></li>
</ul>
</div>
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
onkeydown="return searchBox.OnSearchSelectKey(event)">
</div>
<!-- iframe showing the search results (closed by default) -->
<div id="MSearchResultsWindow">
<iframe src="javascript:void(0)" frameborder="0"
name="MSearchResults" id="MSearchResults">
</iframe>
</div>
<div id="nav-path" class="navpath">
<ul>
<li class="navelem"><a class="el" href="dir_20697b8f204bdfcab31e6b1a416f3ab8.html">caffe2</a></li><li class="navelem"><a class="el" href="dir_ea54d3f22d239c9b391b50ac4ed72021.html">mobile</a></li><li class="navelem"><a class="el" href="dir_79d23c47dbbd20a7aad1bf3dd1c5b17e.html">contrib</a></li><li class="navelem"><a class="el" href="dir_9180558cfb4a7acc352ba8a14826ebc2.html">nnapi</a></li> </ul>
</div>
</div><!-- top -->
<div class="header">
<div class="headertitle">
<div class="title">nnapi_benchmark.cc</div> </div>
</div><!--header-->
<div class="contents">
<div class="fragment"><div class="line"><a name="l00001"></a><span class="lineno"> 1</span> </div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="preprocessor">#include "caffe2/core/init.h"</span></div><div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="preprocessor">#include "caffe2/core/operator.h"</span></div><div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="preprocessor">#include "caffe2/core/tensor.h"</span></div><div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="preprocessor">#include "caffe2/core/timer.h"</span></div><div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="preprocessor">#include "caffe2/utils/math.h"</span></div><div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="preprocessor">#include "caffe2/utils/proto_utils.h"</span></div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="preprocessor">#include "nnapi.h"</span></div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span> </div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="keyword">namespace </span><a class="code" href="namespacecaffe2.html">caffe2</a> {</div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> </div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="keyword">namespace </span>{</div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> </div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="keyword">static</span> <span class="keywordtype">double</span> benchmark_conv_caffe2(</div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span>  Workspace* ws,</div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span>  <span class="keywordtype">int</span> N,</div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span>  <span class="keywordtype">int</span> C,</div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span>  <span class="keywordtype">int</span> H,</div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span>  <span class="keywordtype">int</span> W,</div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span>  <span class="keywordtype">int</span> K,</div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span>  <span class="keywordtype">int</span> kernel,</div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  <span class="keywordtype">int</span> group,</div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  <span class="keywordtype">int</span> warmup = 5,</div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  <span class="keywordtype">int</span> run = 10,</div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  std::string engine = <span class="stringliteral">"NNPACK"</span>) {</div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  <a class="code" href="classcaffe2_1_1_workspace.html">caffe2::Workspace</a> localWs;</div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  <span class="keywordflow">if</span> (!ws) {</div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  ws = &localWs;</div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  }</div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  {</div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  <span class="keyword">auto</span>* t = ws-><a class="code" href="classcaffe2_1_1_workspace.html#a224e2d844e235c2db1804b7f45cd6822">CreateBlob</a>(<span class="stringliteral">"X_cpu"</span>)-><a class="code" href="classcaffe2_1_1_blob.html#a355cff5bfcdfce83ac53ce2a2eef9ee4">GetMutable</a><TensorCPU>();</div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(N, C, H, W);</div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  CPUContext ctx;</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  }</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  {</div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"W"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  <span class="keywordflow">if</span> (group == 1) {</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(K, C, kernel, kernel);</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  t->Resize(K, 1, kernel, kernel);</div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  }</div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  CPUContext ctx;</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  }</div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  {</div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"B"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(K);</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  CPUContext ctx;</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  }</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> </div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  OperatorDef op;</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  {</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  op.set_type(<span class="stringliteral">"Conv"</span>);</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  op.add_input(<span class="stringliteral">"X_cpu"</span>);</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  op.add_input(<span class="stringliteral">"W"</span>);</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  op.add_input(<span class="stringliteral">"B"</span>);</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  op.add_output(<span class="stringliteral">"Y_cpu"</span>);</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  op.set_engine(engine);</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  {</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  arg.set_name(<span class="stringliteral">"order"</span>);</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  arg.set_s(<span class="stringliteral">"NCHW"</span>);</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  }</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  {</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  arg.set_name(<span class="stringliteral">"convolution_transform_strategy"</span>);</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  arg.set_s(<span class="stringliteral">"PRECOMPUTE"</span>);</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  }</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  {</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  arg.set_name(<span class="stringliteral">"kernel"</span>);</div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  arg.set_i(kernel);</div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  }</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  {</div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  arg.set_name(<span class="stringliteral">"group"</span>);</div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  arg.set_i(group);</div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  }</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  }</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span> </div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  <span class="comment">// NNPack</span></div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  std::unique_ptr<caffe2::OperatorBase> op1(CreateOperator(op, ws));</div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span> </div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  Timer timer;</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  CAFFE_ENFORCE(op1->Run());</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < warmup; i++) {</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  op1->Run();</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  }</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span>  timer.Start();</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < run; i++) {</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  op1->Run();</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  }</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  <span class="keywordflow">return</span> double(timer.MilliSeconds()) / run;</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span> }</div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span> </div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> <span class="keyword">static</span> <span class="keywordtype">double</span> benchmark_conv_nnapi(</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  Workspace* ws,</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  <span class="keywordtype">int</span> N,</div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  <span class="keywordtype">int</span> C,</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  <span class="keywordtype">int</span> H,</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  <span class="keywordtype">int</span> W,</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  <span class="keywordtype">int</span> K,</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  <span class="keywordtype">int</span> kernel,</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  <span class="keywordtype">int</span> group,</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="keywordtype">int</span> warmup = 5,</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  <span class="keywordtype">int</span> run = 10) {</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  <a class="code" href="classcaffe2_1_1_workspace.html">caffe2::Workspace</a> localWs;</div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="keywordflow">if</span> (!ws) {</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  ws = &localWs;</div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  }</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  {</div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  <span class="keyword">auto</span>* t = ws-><a class="code" href="classcaffe2_1_1_workspace.html#a224e2d844e235c2db1804b7f45cd6822">CreateBlob</a>(<span class="stringliteral">"X_cpu"</span>)-><a class="code" href="classcaffe2_1_1_blob.html#a355cff5bfcdfce83ac53ce2a2eef9ee4">GetMutable</a><TensorCPU>();</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(N, H, W, C);</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  CPUContext ctx;</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  }</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  {</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"W"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="keywordflow">if</span> (group > 1) {</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  CAFFE_ENFORCE_EQ(C, group);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  t->Resize(1, kernel, kernel, C);</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  t->Resize(K, kernel, kernel, C);</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  }</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  CPUContext ctx;</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  }</div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  {</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"B"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(K);</div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  CPUContext ctx;</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  math::RandGaussian<float, CPUContext>(</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  t->size(), 0, 30, t->mutable_data<<span class="keywordtype">float</span>>(), &ctx);</div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  }</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span> </div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  NetDef netdef;</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  {</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  {</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="keyword">auto</span>& op = *(netdef.add_op());</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  op.set_type(<span class="stringliteral">"Conv"</span>);</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  op.add_input(<span class="stringliteral">"X_cpu"</span>);</div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  op.add_input(<span class="stringliteral">"W"</span>);</div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  op.add_input(<span class="stringliteral">"B"</span>);</div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  op.add_output(<span class="stringliteral">"Y_cpu"</span>);</div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  {</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  arg.set_name(<span class="stringliteral">"order"</span>);</div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  arg.set_s(<span class="stringliteral">"NHWC"</span>);</div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  }</div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  {</div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  arg.set_name(<span class="stringliteral">"kernel"</span>);</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  arg.set_i(kernel);</div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  }</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  {</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  arg.set_name(<span class="stringliteral">"group"</span>);</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  arg.set_i(group);</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  }</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  }</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  netdef.add_external_input(<span class="stringliteral">"X_cpu"</span>);</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  netdef.add_external_input(<span class="stringliteral">"W"</span>);</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  netdef.add_external_input(<span class="stringliteral">"B"</span>);</div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  netdef.add_external_output(<span class="stringliteral">"Y_cpu"</span>);</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  }</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span> </div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  <span class="comment">// NN API</span></div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  NetDef initNet;</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  NNApi model(initNet, netdef, ws);</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  std::vector<TensorCPU*> inputs, outputs;</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  inputs.push_back(ws->GetBlob(<span class="stringliteral">"X_cpu"</span>)->GetMutable<TensorCPU>());</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  CAFFE_ENFORCE(model.run(inputs, &outputs));</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span> </div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < warmup; i++) {</div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  model.run(inputs, &outputs);</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  }</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  Timer timer;</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  timer.Start();</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < run; i++) {</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  model.run(inputs, &outputs);</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  }</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  <span class="keywordflow">return</span> double(timer.MilliSeconds()) / run;</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span> }</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span> </div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span> <span class="keyword">static</span> <span class="keywordtype">double</span> benchmark_conv_nnapi_int8(</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  Workspace* ws,</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="keywordtype">int</span> N,</div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="keywordtype">int</span> C,</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  <span class="keywordtype">int</span> H,</div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="keywordtype">int</span> W,</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  <span class="keywordtype">int</span> K,</div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="keywordtype">int</span> kernel,</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  <span class="keywordtype">int</span> group,</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span>  <span class="keywordtype">int</span> warmup = 5,</div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  <span class="keywordtype">int</span> run = 10) {</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  <a class="code" href="classcaffe2_1_1_workspace.html">caffe2::Workspace</a> localWs;</div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  <span class="keywordflow">if</span> (!ws) {</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  ws = &localWs;</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  }</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  {</div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <span class="keyword">auto</span>* t = ws-><a class="code" href="classcaffe2_1_1_workspace.html#a224e2d844e235c2db1804b7f45cd6822">CreateBlob</a>(<span class="stringliteral">"X_cpu"</span>)-><a class="code" href="classcaffe2_1_1_blob.html#a355cff5bfcdfce83ac53ce2a2eef9ee4">GetMutable</a><TensorCPU>();</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(N, H, W, C);</div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < t->size(); i++) {</div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  t->mutable_data<uint8_t>()[i] = rand() % 10;</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  }</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  }</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  {</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"W"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  <span class="keywordflow">if</span> (group > 1) {</div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  CAFFE_ENFORCE_EQ(C, group);</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  t->Resize(1, kernel, kernel, C);</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  t->Resize(K, kernel, kernel, C);</div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  }</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < t->size(); i++) {</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  t->mutable_data<uint8_t>()[i] = rand() % 10;</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  }</div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  }</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> </div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  <span class="comment">// For input tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM type, the bias</span></div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  <span class="comment">// should be of ANEURALNETWORKS_TENSOR_INT32, with zeroPoint of 0 and</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  <span class="comment">// bias_scale == input_scale * filter_scale.</span></div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  {</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="keyword">auto</span>* t = ws->CreateBlob(<span class="stringliteral">"B"</span>)->GetMutable<TensorCPU>();</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  t-><a class="code" href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">Resize</a>(K);</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < t->size(); i++) {</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  t->mutable_data<int32_t>()[i] = rand() % 10;</div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  }</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  }</div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span> </div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  NetDef netdef;</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  {</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  {</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  <span class="keyword">auto</span>& op = *(netdef.add_op());</div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  op.set_type(<span class="stringliteral">"Conv"</span>);</div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  op.add_input(<span class="stringliteral">"X_cpu"</span>);</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  op.add_input(<span class="stringliteral">"W"</span>);</div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  op.add_input(<span class="stringliteral">"B"</span>);</div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  op.add_output(<span class="stringliteral">"Y_cpu"</span>);</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  {</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  arg.set_name(<span class="stringliteral">"order"</span>);</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  arg.set_s(<span class="stringliteral">"NHWC"</span>);</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  }</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  {</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  arg.set_name(<span class="stringliteral">"kernel"</span>);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  arg.set_i(kernel);</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  }</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  {</div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  arg.set_name(<span class="stringliteral">"group"</span>);</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  arg.set_i(group);</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  }</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span>  <span class="comment">// Hack</span></div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  <span class="comment">// for weight tensor</span></div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  {</div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  arg.set_name(<span class="stringliteral">"weight_scale"</span>);</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  arg.set_f(1.0);</div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  }</div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  {</div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  arg.set_name(<span class="stringliteral">"weight_zero_point"</span>);</div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  arg.set_i(0);</div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  }</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  <span class="comment">// for output tensor</span></div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  <span class="comment">// For output tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM type, the</span></div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <span class="comment">// following condition must be satisfied: output_scale > input_scale *</span></div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <span class="comment">// filter_scale</span></div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  {</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  arg.set_name(<span class="stringliteral">"output_scale"</span>);</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  arg.set_f(2.0);</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  }</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  {</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  <span class="keyword">auto</span>& arg = *(op.add_arg());</div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  arg.set_name(<span class="stringliteral">"output_zero_point"</span>);</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  arg.set_i(0);</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  }</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  }</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span>  netdef.add_external_input(<span class="stringliteral">"X_cpu"</span>);</div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span>  netdef.add_external_input(<span class="stringliteral">"W"</span>);</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  netdef.add_external_input(<span class="stringliteral">"B"</span>);</div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  netdef.add_external_output(<span class="stringliteral">"Y_cpu"</span>);</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  <span class="comment">// scale and zero_point for the input tensor</span></div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  {</div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  <span class="keyword">auto</span>& arg = *(netdef.add_arg());</div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  arg.set_name(<span class="stringliteral">"scale"</span>);</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  arg.set_f(1.0);</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  }</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  {</div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  <span class="keyword">auto</span>& arg = *(netdef.add_arg());</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  arg.set_name(<span class="stringliteral">"zero_point"</span>);</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  arg.set_i(0);</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span>  }</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  }</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span> </div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  <span class="comment">// NN API</span></div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  NetDef initNet;</div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  NNApi model(initNet, netdef, ws);</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  std::vector<TensorCPU*> inputs, outputs;</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  inputs.push_back(ws->GetBlob(<span class="stringliteral">"X_cpu"</span>)->GetMutable<TensorCPU>());</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  CAFFE_ENFORCE(model.run(inputs, &outputs));</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span> </div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < warmup; i++) {</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span>  model.run(inputs, &outputs);</div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>  }</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  Timer timer;</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  timer.Start();</div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < run; i++) {</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  model.run(inputs, &outputs);</div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  }</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  <span class="keywordflow">return</span> double(timer.MilliSeconds()) / run;</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span> }</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span> </div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span> } <span class="comment">// namespace</span></div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span> </div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span> } <span class="comment">// namespace caffe2</span></div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span> </div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span> <span class="keywordtype">int</span> main(<span class="keywordtype">int</span> argc, <span class="keywordtype">char</span>** argv) {</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  <a class="code" href="classcaffe2_1_1_workspace.html">caffe2::Workspace</a> ws;</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  ws.GetThreadPool()->setMinWorkSize(0);</div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span> </div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  <span class="keywordtype">int</span> warmup = 2, mainrun = 10;</div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="comment">// float32</span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> space : {14, 26, 52, 104}) {</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> input_channel : {64, 128, 256, 512}) {</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> kernel : {1, 3}) {</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  <span class="keywordtype">int</span> output_channel = input_channel;</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> cpu_time = caffe2::benchmark_conv_caffe2(</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  &ws,</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  1,</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  input_channel,</div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  space,</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  space,</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  output_channel,</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  kernel,</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  1,</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  warmup,</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  mainrun,</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  <span class="stringliteral">"NNPACK"</span>);</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> nn_time_fp32 = caffe2::benchmark_conv_nnapi(</div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  &ws,</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  1,</div><div class="line"><a name="l00368"></a><span class="lineno"> 368</span>  input_channel,</div><div class="line"><a name="l00369"></a><span class="lineno"> 369</span>  space,</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  space,</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  output_channel,</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  kernel,</div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  1,</div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  warmup,</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  mainrun);</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> nn_time_int8 = caffe2::benchmark_conv_nnapi_int8(</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  &ws,</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  1,</div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  input_channel,</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  space,</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  space,</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  output_channel,</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  kernel,</div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  1,</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  warmup,</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  mainrun);</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> flops = double(input_channel) * output_channel * kernel *</div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  kernel * (kernel == 1 ? space : space - 2) *</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  (kernel == 1 ? space : space - 2) * 2;</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  printf(</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  <span class="stringliteral">"Conv: X: %ix%i \tC: %i -> %i\tK: %ix%i\t32b"</span></div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  <span class="stringliteral">"NNPACK GFLOPS: %.2f\t32b"</span></div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  <span class="stringliteral">"NN-API GFLOPS: %.2f\t8b"</span></div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  <span class="stringliteral">"NN-API GOPS: %.2f\n"</span>,</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  space,</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  space,</div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  input_channel,</div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  output_channel,</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  kernel,</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  kernel,</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  flops / cpu_time / 1E6,</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  flops / nn_time_fp32 / 1E6,</div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  flops / nn_time_int8 / 1E6);</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  }</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  }</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  }</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span>  fflush(stdout);</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span> </div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <span class="comment">// depthwise</span></div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> space : {14, 26, 52, 104}) {</div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> channel : {64, 128, 256, 512}) {</div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> kernel : {3}) {</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> cpu_time = caffe2::benchmark_conv_caffe2(</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  &ws,</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span>  1,</div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span>  channel,</div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  space,</div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span>  space,</div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  channel,</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span>  kernel,</div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span>  channel,</div><div class="line"><a name="l00422"></a><span class="lineno"> 422</span>  warmup,</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  mainrun,</div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  <span class="stringliteral">"DEPTHWISE_3x3"</span>);</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> nn_time_fp32_dwise = caffe2::benchmark_conv_nnapi(</div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  &ws,</div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  1,</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span>  channel,</div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  space,</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  space,</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  channel,</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span>  kernel,</div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span>  channel,</div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span>  warmup,</div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  mainrun);</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> nn_time_int8_dwise = caffe2::benchmark_conv_nnapi_int8(</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  &ws,</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  1,</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span>  channel,</div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  space,</div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  space,</div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  channel,</div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  kernel,</div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  channel,</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  warmup,</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  mainrun);</div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <span class="keyword">const</span> <span class="keywordtype">double</span> dwise_bandwidth = <span class="keyword">sizeof</span>(float) * <span class="keywordtype">double</span>(channel) *</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  (2 * (space - 2) * (space - 2) + kernel * kernel);</div><div class="line"><a name="l00449"></a><span class="lineno"> 449</span>  printf(</div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  <span class="stringliteral">"Conv: X: %ix%i \tC: %i -> %i\tK: %ix%i\t32b"</span></div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="stringliteral">"Caffe2 Dwise GB/s: %.2f\t32b"</span></div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  <span class="stringliteral">"NN-API Dwise GB/s: %.2f\t8b"</span></div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="stringliteral">"NN-API Dwise GB/s: %.2f\n"</span>,</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  space,</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  space,</div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>  channel,</div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  channel,</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  kernel,</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  kernel,</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  dwise_bandwidth / cpu_time / 1E6,</div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span>  dwise_bandwidth / nn_time_fp32_dwise / 1E6,</div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  dwise_bandwidth / <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>) / nn_time_int8_dwise / 1E6);</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  }</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  }</div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  }</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span> }</div><div class="ttc" id="classcaffe2_1_1_workspace_html_a224e2d844e235c2db1804b7f45cd6822"><div class="ttname"><a href="classcaffe2_1_1_workspace.html#a224e2d844e235c2db1804b7f45cd6822">caffe2::Workspace::CreateBlob</a></div><div class="ttdeci">Blob * CreateBlob(const string &name)</div><div class="ttdoc">Creates a blob of the given name. </div><div class="ttdef"><b>Definition:</b> <a href="workspace_8cc_source.html#l00104">workspace.cc:104</a></div></div>
<div class="ttc" id="classcaffe2_1_1_workspace_html"><div class="ttname"><a href="classcaffe2_1_1_workspace.html">caffe2::Workspace</a></div><div class="ttdoc">Workspace is a class that holds all the related objects created during runtime: (1) all blobs...</div><div class="ttdef"><b>Definition:</b> <a href="workspace_8h_source.html#l00047">workspace.h:47</a></div></div>
<div class="ttc" id="classcaffe2_1_1_tensor_html_a359b5ed5cfd9beaf7f62a5561d939c3b"><div class="ttname"><a href="classcaffe2_1_1_tensor.html#a359b5ed5cfd9beaf7f62a5561d939c3b">caffe2::Tensor::Resize</a></div><div class="ttdeci">void Resize(Ts...dim_source)</div><div class="ttdoc">Resizes a tensor. </div><div class="ttdef"><b>Definition:</b> <a href="tensor_8h_source.html#l00288">tensor.h:288</a></div></div>
<div class="ttc" id="namespacecaffe2_html"><div class="ttname"><a href="namespacecaffe2.html">caffe2</a></div><div class="ttdoc">A global dictionary that holds information about what Caffe2 modules have been loaded in the current ...</div><div class="ttdef"><b>Definition:</b> <a href="convert__encoded__to__raw__leveldb_8cc_source.html#l00047">convert_encoded_to_raw_leveldb.cc:47</a></div></div>
<div class="ttc" id="classcaffe2_1_1_blob_html_a355cff5bfcdfce83ac53ce2a2eef9ee4"><div class="ttname"><a href="classcaffe2_1_1_blob.html#a355cff5bfcdfce83ac53ce2a2eef9ee4">caffe2::Blob::GetMutable</a></div><div class="ttdeci">T * GetMutable(bool *is_new_object=nullptr)</div><div class="ttdoc">Gets a mutable pointer to the stored object. </div><div class="ttdef"><b>Definition:</b> <a href="blob_8h_source.html#l00101">blob.h:101</a></div></div>
</div><!-- fragment --></div><!-- contents -->
<!-- HTML footer for doxygen 1.8.14-->
<!-- start footer part -->
<hr class="footer"/><address class="footer"><small>
Generated on Thu Apr 19 2018 13:03:54 for Caffe2 - C++ API by  <a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/>
</a> 1.8.11
</small></address>
<div class="footerContainer">
<div id="footer_wrap" class="wrapper footerWrapper">
<div class="footerBlocks">
<div id="fb_oss" class="footerSection fbOpenSourceFooter">
<svg class="facebookOSSLogoSvg" viewBox="0 0 1133.9 1133.9" x="0px" y="0px" height=50 width=50>
<g>
<path class="logoRing outerRing" d="M 498.3 3.7 c 153.6 88.9 307.3 177.7 461.1 266.2 c 7.6 4.4 10.3 9.1 10.3 17.8 c -0.3 179.1 -0.2 358.3 0 537.4 c 0 8.1 -2.4 12.8 -9.7 17.1 c -154.5 88.9 -308.8 178.1 -462.9 267.5 c -9 5.2 -15.5 5.3 -24.6 0.1 c -153.9 -89.2 -307.9 -178 -462.1 -266.8 C 3 838.8 0 833.9 0 825.1 c 0.3 -179.1 0.2 -358.3 0 -537.4 c 0 -8.6 2.6 -13.6 10.2 -18 C 164.4 180.9 318.4 92 472.4 3 C 477 -1.5 494.3 -0.7 498.3 3.7 Z M 48.8 555.3 c 0 79.9 0.2 159.9 -0.2 239.8 c -0.1 10 3 15.6 11.7 20.6 c 137.2 78.8 274.2 157.8 411 237.3 c 9.9 5.7 17 5.7 26.8 0.1 c 137.5 -79.8 275.2 -159.2 412.9 -238.5 c 7.4 -4.3 10.5 -8.9 10.5 -17.8 c -0.3 -160.2 -0.3 -320.5 0 -480.7 c 0 -8.8 -2.8 -13.6 -10.3 -18 C 772.1 218 633.1 137.8 494.2 57.4 c -6.5 -3.8 -11.5 -4.5 -18.5 -0.5 C 336.8 137.4 197.9 217.7 58.8 297.7 c -7.7 4.4 -10.2 9.2 -10.2 17.9 C 48.9 395.5 48.8 475.4 48.8 555.3 Z" />
<path class="logoRing middleRing" d="M 184.4 555.9 c 0 -33.3 -1 -66.7 0.3 -100 c 1.9 -48 24.1 -86 64.7 -110.9 c 54.8 -33.6 110.7 -65.5 167 -96.6 c 45.7 -25.2 92.9 -24.7 138.6 1 c 54.4 30.6 108.7 61.5 162.2 93.7 c 44 26.5 67.3 66.8 68 118.4 c 0.9 63.2 0.9 126.5 0 189.7 c -0.7 50.6 -23.4 90.7 -66.6 116.9 c -55 33.4 -110.8 65.4 -167.1 96.5 c -43.4 24 -89 24.2 -132.3 0.5 c -57.5 -31.3 -114.2 -64 -170 -98.3 c -41 -25.1 -62.9 -63.7 -64.5 -112.2 C 183.5 621.9 184.3 588.9 184.4 555.9 Z M 232.9 556.3 c 0 29.5 0.5 59.1 -0.1 88.6 c -0.8 39.2 16.9 67.1 50.2 86.2 c 51.2 29.4 102.2 59.2 153.4 88.4 c 31.4 17.9 63.6 18.3 95 0.6 c 53.7 -30.3 107.1 -61.2 160.3 -92.5 c 29.7 -17.5 45 -44.5 45.3 -78.8 c 0.6 -61.7 0.5 -123.5 0 -185.2 c -0.3 -34.4 -15.3 -61.5 -44.9 -79 C 637.7 352.6 583 320.8 527.9 290 c -27.5 -15.4 -57.2 -16.1 -84.7 -0.7 c -56.9 31.6 -113.4 64 -169.1 97.6 c -26.4 15.9 -40.7 41.3 -41.1 72.9 C 232.6 491.9 232.9 524.1 232.9 556.3 Z" />
<path class="logoRing innerRing" d="M 484.9 424.4 c 69.8 -2.8 133.2 57.8 132.6 132 C 617 630 558.5 688.7 484.9 689.1 c -75.1 0.4 -132.6 -63.6 -132.7 -132.7 C 352.1 485 413.4 421.5 484.9 424.4 Z M 401.3 556.7 c -3.4 37.2 30.5 83.6 83 84.1 c 46.6 0.4 84.8 -37.6 84.9 -84 c 0.1 -46.6 -37.2 -84.4 -84.2 -84.6 C 432.2 472.1 397.9 518.3 401.3 556.7 Z" />
</g>
</svg>
<h2>Facebook Open Source</h2>
</div>
<div class="footerSection">
<a class="footerLink" href="https://code.facebook.com/projects/" target="_blank">Open Source Projects</a>
<a class="footerLink" href="https://github.com/facebook/" target="_blank">GitHub</a>
<a class="footerLink" href="https://twitter.com/fbOpenSource" target="_blank">Twitter</a>
</div>
<div class="footerSection rightAlign">
<a class="footerLink" href="https://github.com/caffe2/caffe2" target="_blank">Contribute to this project on GitHub</a>
</div>
</div>
</div>
</div>
<script type="text/javascript" src="/js/jekyll-link-anchors.js"></script>
<script>
(function(i,s,o,g,r,a,m){i['GoogleAnalyticsObject']=r;i[r]=i[r]||function(){
(i[r].q=i[r].q||[]).push(arguments)},i[r].l=1*new Date();a=s.createElement(o),
m=s.getElementsByTagName(o)[0];a.async=1;a.src=g;m.parentNode.insertBefore(a,m)
})(window,document,'script','//www.google-analytics.com/analytics.js','ga');
ga('create', '{{ site.gacode }}', 'auto');
ga('send', 'pageview');
</script>
</body>
</html>