diff options
| author | David Monahan <david.monahan@arm.com> | 2023-03-22 16:48:58 +0000 |
|---|---|---|
| committer | David Monahan <david.monahan@arm.com> | 2023-03-22 16:48:58 +0000 |
| commit | ae050524109f1ce827962665436ef7430f2ac479 (patch) | |
| tree | a087fe0c77570971dd7979f2757426c24e91afc7 /23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml | |
| parent | 8d2ca734165a068478df7cffa46185680b05cd20 (diff) | |
| download | armnn-ae050524109f1ce827962665436ef7430f2ac479.tar.gz | |
IVGCVSW-7255 Update Doxygen Documentation and publish on GitHub.
* Updating Doxygen documentation for 23.02 release.
Signed-off-by: David Monahan <david.monahan@arm.com>
Change-Id: I545574ff7664b4595d2fe6a91a3c35d2ad55df82
Diffstat (limited to '23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml')
| -rw-r--r-- | 23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml | 235 |
1 files changed, 210 insertions, 25 deletions
diff --git a/23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml b/23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml index 9351471ba3..926b084af1 100644 --- a/23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml +++ b/23.02/classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml @@ -8,7 +8,7 @@ <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 1.8.13"/> +<meta name="generator" content="Doxygen 1.8.17"/> <meta name="robots" content="NOINDEX, NOFOLLOW" /> <meta name="viewport" content="width=device-width, initial-scale=1"/> <title>ArmNN: FuseBatchNorm< ConvLayer, ArmnnType, T > Class Template Reference</title> @@ -19,9 +19,6 @@ <script type="text/javascript" src="resize.js"></script> <script type="text/javascript" src="navtreedata.js"></script> <script type="text/javascript" src="navtree.js"></script> -<script type="text/javascript"> - $(document).ready(initResizable); -</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> @@ -30,7 +27,8 @@ extensions: ["tex2jax.js"], jax: ["input/TeX","output/HTML-CSS"], }); -</script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script> +</script> +<script type="text/javascript" async="async" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script> <link href="doxygen.css" rel="stylesheet" type="text/css" /> <link href="stylesheet.css" rel="stylesheet" type="text/css"/> </head> @@ -51,18 +49,21 @@ </table> </div> <!-- end header part --> -<!-- Generated by Doxygen 1.8.13 --> +<!-- Generated by Doxygen 1.8.17 --> <script type="text/javascript"> +/* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ var searchBox = new SearchBox("searchBox", "search",false,'Search'); +/* @license-end */ </script> <script type="text/javascript" src="menudata.js"></script> <script type="text/javascript" src="menu.js"></script> <script type="text/javascript"> +/* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ $(function() { initMenu('',true,false,'search.php','Search'); $(document).ready(function() { init_search(); }); }); -</script> +/* @license-end */</script> <div id="main-nav"></div> </div><!-- top --> <div id="side-nav" class="ui-resizable side-nav-resizable"> @@ -76,7 +77,9 @@ $(function() { </div> </div> <script type="text/javascript"> -$(document).ready(function(){initNavTree('classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml','');}); +/* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ +$(document).ready(function(){initNavTree('classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml',''); initResizable(); }); +/* @license-end */ </script> <div id="doc-content"> <!-- window showing the filter options --> @@ -108,7 +111,7 @@ $(document).ready(function(){initNavTree('classarmnn_1_1optimizations_1_1_fuse_b <tr class="heading"><td colspan="2"><h2 class="groupheader"><a name="pub-methods"></a> Public Member Functions</h2></td></tr> <tr class="memitem:a5a8476ffc04ce7460bb09ad50d1d23de"><td class="memItemLeft" align="right" valign="top">void </td><td class="memItemRight" valign="bottom"><a class="el" href="classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml#a5a8476ffc04ce7460bb09ad50d1d23de">Run</a> (<a class="el" href="classarmnn_1_1_graph.xhtml">Graph</a> &graph, <a class="el" href="classarmnn_1_1_input_slot.xhtml">InputSlot</a> &connection) const</td></tr> -<tr class="memdesc:a5a8476ffc04ce7460bb09ad50d1d23de"><td class="mdescLeft"> </td><td class="mdescRight">Run for every exclusive connection between any base Convolution layer and a child BatchNorm layer for not quantized layers. <a href="#a5a8476ffc04ce7460bb09ad50d1d23de">More...</a><br /></td></tr> +<tr class="memdesc:a5a8476ffc04ce7460bb09ad50d1d23de"><td class="mdescLeft"> </td><td class="mdescRight">Run for every exclusive connection between any base Convolution layer and a child BatchNorm layer for not quantized layers. <a href="classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml#a5a8476ffc04ce7460bb09ad50d1d23de">More...</a><br /></td></tr> <tr class="separator:a5a8476ffc04ce7460bb09ad50d1d23de"><td class="memSeparator" colspan="2"> </td></tr> </table><table class="memberdecls"> <tr class="heading"><td colspan="2"><h2 class="groupheader"><a name="pro-methods"></a> @@ -148,8 +151,6 @@ class armnn::optimizations::FuseBatchNorm< ConvLayer, ArmnnType, T ></h3> </table> </div><div class="memdoc"> -<p class="reference">Referenced by <a class="el" href="_fuse_batch_norm_8hpp_source.xhtml#l00027">FuseBatchNorm< ConvLayer, ArmnnType, T >::Run()</a>.</p> - </div> </div> <a id="a0ff9a790927b898d90261a8ea0e479e6"></a> @@ -175,8 +176,6 @@ class armnn::optimizations::FuseBatchNorm< ConvLayer, ArmnnType, T ></h3> </table> </div><div class="memdoc"> -<p class="reference">Referenced by <a class="el" href="_fuse_batch_norm_8hpp_source.xhtml#l00027">FuseBatchNorm< ConvLayer, ArmnnType, T >::Run()</a>.</p> - </div> </div> <h2 class="groupheader">Member Function Documentation</h2> @@ -218,17 +217,195 @@ class armnn::optimizations::FuseBatchNorm< ConvLayer, ArmnnType, T ></h3> <p>The child will be removed, the base will be removed if it's left unconnected. A new Convolution layer will be added, its weights and bias will be calculated using the weights and bias of the base Convolution layer combined with the parameters of the child BatchNorm layer. </p> <p class="definition">Definition at line <a class="el" href="_fuse_batch_norm_8hpp_source.xhtml#l00027">27</a> of file <a class="el" href="_fuse_batch_norm_8hpp_source.xhtml">FuseBatchNorm.hpp</a>.</p> - -<p class="reference">References <a class="el" href="_graph_8hpp_source.xhtml#l00456">Graph::AddLayer()</a>, <a class="el" href="_assert_8hpp_source.xhtml#l00014">ARMNN_ASSERT</a>, <a class="el" href="_assert_8hpp_source.xhtml#l00015">ARMNN_ASSERT_MSG</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">armnn::BatchNormalization</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">armnn::Convolution2d</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">armnn::DepthwiseConvolution2d</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00120">OutputSlot::Disconnect()</a>, <a class="el" href="classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml#abe49327783cb8bdc12c085c987db14db">FuseBatchNorm< ConvLayer, ArmnnType, T >::FuseBatchNorm()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00056">InputSlot::GetConnectedOutputSlot()</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00313">Layer::GetDataType()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00295">BaseTensor< MemoryType >::GetInfo()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00324">Layer::GetInputSlot()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00319">Layer::GetName()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00303">BaseTensor< MemoryType >::GetNumElements()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00326">Layer::GetOutputSlot()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00053">InputSlot::GetOwningLayer()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00119">OutputSlot::GetOwningLayer()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00191">TensorInfo::GetShape()</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00092">OutputSlot::GetTensorInfo()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00273">Layer::GetType()</a>, <a class="el" href="_ignore_unused_8hpp_source.xhtml#l00014">armnn::IgnoreUnused()</a>, <a class="el" href="_graph_8hpp_source.xhtml#l00471">Graph::InsertNewLayer()</a>, <a class="el" href="_descriptors_8hpp_source.xhtml#l00809">BatchNormalizationDescriptor::m_Eps</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00145">OutputSlot::MoveAllConnections()</a>, <a class="el" href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">armnn::NHWC</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00087">OutputSlot::SetTensorInfo()</a>, and <a class="el" href="classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml#a0ff9a790927b898d90261a8ea0e479e6">FuseBatchNorm< ConvLayer, ArmnnType, T >::~FuseBatchNorm()</a>.</p> -<div class="fragment"><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>  Layer& base = connection.GetConnectedOutputSlot()->GetOwningLayer();</div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span>  Layer& child = connection.GetOwningLayer();</div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> </div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span>  <span class="keywordtype">bool</span> depthwise = (base.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">LayerType::DepthwiseConvolution2d</a>);</div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> </div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span>  <a class="code" href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a>(base.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">LayerType::Convolution2d</a> || depthwise);</div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span>  <a class="code" href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a>(child.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">LayerType::BatchNormalization</a>);</div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> </div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  <span class="keywordflow">if</span> (base.GetDataType() == ArmnnType && child.GetDataType() == ArmnnType)</div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  {</div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  OutputSlot* parentOut = base.GetInputSlot(0).GetConnectedOutputSlot();</div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  <span class="keyword">auto</span> convLayer = PolymorphicDowncast<ConvLayer*>(&base);</div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  <span class="keyword">auto</span> batchNormLayer = PolymorphicDowncast<BatchNormalizationLayer*>(&child);</div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> </div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  <span class="comment">// Read convolution and batch norm parameters</span></div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  BatchNormalizationDescriptor batchNormDescriptor = batchNormLayer->GetParameters();</div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  <span class="keyword">auto</span> epsilon = batchNormDescriptor.m_Eps;</div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  <a class="code" href="namespacearmnn.xhtml#a44affeeb090c3c6a3062830562672e84">IgnoreUnused</a>(epsilon);</div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span> </div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  ConstTensor betaTensor(batchNormLayer->m_Beta->GetTensorInfo(), batchNormLayer->m_Beta->Map(<span class="keyword">true</span>));</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span>  ConstTensor gammaTensor(batchNormLayer->m_Gamma->GetTensorInfo(), batchNormLayer->m_Gamma->Map(<span class="keyword">true</span>));</div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  ConstTensor meanTensor(batchNormLayer->m_Mean->GetTensorInfo(), batchNormLayer->m_Mean->Map(<span class="keyword">true</span>));</div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  ConstTensor varTensor(batchNormLayer->m_Variance->GetTensorInfo(), batchNormLayer->m_Variance->Map(<span class="keyword">true</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> convDescriptor = convLayer->GetParameters();</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  ConstTensor weightsTensor;</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <a class="code" href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a>(convLayer->GetInputSlots()[1].GetConnection() != <span class="keyword">nullptr</span>,</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="stringliteral">"FuseBatchNorm: Weight data should not be null."</span>);</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> </div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  ConstantLayer* weightLayer = PolymorphicDowncast<ConstantLayer*>(</div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  &base.GetInputSlot(1).GetConnectedOutputSlot()->GetOwningLayer());</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> </div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  weightsTensor = ConstTensor(weightLayer->m_LayerOutput->GetTensorInfo(),</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  weightLayer->m_LayerOutput->Map(<span class="keyword">true</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>  <a class="code" href="classarmnn_utils_1_1_data_layout_indexed.xhtml">armnnUtils::DataLayoutIndexed</a> dataLayout(convDescriptor.m_DataLayout);</div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keyword">auto</span> weightsShape = weightsTensor.GetInfo().GetShape();</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> inputChannels = parentOut->GetTensorInfo().GetShape()[dataLayout.GetChannelsIndex()];</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> depthMultiplier = depthwise ? weightsShape[3] / inputChannels : 1;</div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> outputChannels = depthwise ? weightsShape[3] : weightsShape[0];</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsHeight = depthwise ? weightsShape[1] :</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  weightsShape[dataLayout.GetHeightIndex()];</div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsWidth = depthwise ? weightsShape[2] :</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  weightsShape[dataLayout.GetWidthIndex()];</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* weightsBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(weightsTensor.GetMemoryArea());</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* betaBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(betaTensor.GetMemoryArea());</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* gammaBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(gammaTensor.GetMemoryArea());</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* meanBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(meanTensor.GetMemoryArea());</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* varBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(varTensor.GetMemoryArea());</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>  std::vector<T> weightsVector (weightsBuffer, weightsBuffer + weightsTensor.GetNumElements());</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  std::vector<T> betaVector (betaBuffer, betaBuffer + betaTensor.GetNumElements());</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  std::vector<T> gammaVector (gammaBuffer, gammaBuffer + gammaTensor.GetNumElements());</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  std::vector<T> meanVector (meanBuffer, meanBuffer + meanTensor.GetNumElements());</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  std::vector<T> varianceVector(varBuffer, varBuffer + varTensor.GetNumElements());</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span> </div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="comment">// fusedWeights = ( gamma * weights ) / ( std - epsilon);</span></div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  std::vector<T> fusedWeightsVector(weightsVector.size());</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cInput = 0; cInput < inputChannels; ++cInput)</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  {</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  {</div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  T mult = gammaVector[cOut] / <span class="keyword">static_cast<</span>T<span class="keyword">></span>(sqrtf(varianceVector[cOut] + epsilon));</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="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> h = 0; h < weightsHeight; ++h)</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  {</div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> w = 0; w < weightsWidth; ++w)</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>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsIdx = 0;</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="keywordflow">if</span> (depthwise)</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  {</div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  cInput = cOut / depthMultiplier;</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  weightsIdx = w * outputChannels + cOut +</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  h * weightsWidth * outputChannels;</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  }</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span> (convDescriptor.m_DataLayout == <a class="code" href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">DataLayout::NHWC</a>)</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>  weightsIdx = cOut * weightsHeight * weightsWidth * inputChannels +</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  h * weightsWidth * inputChannels +</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  w * inputChannels +</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  cInput;</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  }</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  <span class="keywordflow">else</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>  weightsIdx = cOut * weightsWidth * weightsHeight * inputChannels +</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  cInput * weightsWidth * weightsHeight +</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  h * weightsWidth +</div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  w;</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  }</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  fusedWeightsVector[weightsIdx] = mult * weightsVector[weightsIdx];</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  }</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  }</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  }</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  }</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  ConstTensor fusedWeightsTensor(weightsTensor.GetInfo(), fusedWeightsVector);</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span> </div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="comment">// fusedBias = (gamma * (bias - mean)) / (variance - epsilon) + beta;</span></div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  std::vector<T> fusedBiasVector(outputChannels);</div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordtype">bool</span> biasWasEnabledBeforeOpt = convDescriptor.m_BiasEnabled;</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  <span class="keywordflow">if</span> (biasWasEnabledBeforeOpt)</div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  {</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  ConstTensor biasTensor;</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  <a class="code" href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a>(convLayer->GetInputSlots()[2].GetConnection() != <span class="keyword">nullptr</span>,</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  <span class="stringliteral">"FuseBatchNorm: Bias data should not be null if bias is enabled."</span>);</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span> </div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  ConstantLayer* biasLayer = PolymorphicDowncast<ConstantLayer*>(</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  &base.GetInputSlot(2).GetConnectedOutputSlot()->GetOwningLayer());</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span> </div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  biasTensor = ConstTensor(biasLayer->m_LayerOutput->GetTensorInfo(),</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  biasLayer->m_LayerOutput->Map(<span class="keyword">true</span>));</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span> </div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* biasBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(biasTensor.GetMemoryArea());</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  std::vector<T> biasVector(biasBuffer, biasBuffer + biasTensor.GetNumElements());</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  {</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  fusedBiasVector[cOut] = ((gammaVector[cOut] * (biasVector[cOut] - meanVector[cOut])) /</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  sqrtf(varianceVector[cOut] + epsilon)) + betaVector[cOut];</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>  }</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  {</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  convDescriptor.m_BiasEnabled = <span class="keyword">true</span>;</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  std::vector<T> biasVector(outputChannels, T(0));</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  {</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  fusedBiasVector[cOut] = ((gammaVector[cOut] * (biasVector[cOut] - meanVector[cOut])) /</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  sqrtf(varianceVector[cOut] + epsilon)) + betaVector[cOut];</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  }</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  }</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  ConstTensor fusedBiasTensor(TensorInfo({outputChannels}, ArmnnType, 0.0f, 0, <span class="keyword">true</span>), fusedBiasVector);</div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span> </div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// Insert the new convolution layer that has batch norm parameters fused into</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="keyword">const</span> std::string name = std::string(<span class="stringliteral">"fused-"</span>) + child.GetName() + std::string(<span class="stringliteral">"-into-"</span>) + base.GetName();</div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="keyword">auto</span>& newConv2dLayer = *graph.InsertNewLayer<ConvLayer>(base.GetInputSlot(0),</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  convDescriptor,</div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  name.c_str());</div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span> </div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  <span class="comment">// Connect weights and bias from old to new Conv2d layer</span></div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  <span class="comment">// This optimization will always have 3 input slots on the Conv2d base layer</span></div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <span class="keywordflow">if</span> (newConv2dLayer.GetNumInputSlots() > 1)</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  {</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <span class="comment">// Remove old connection and connect to new layer2d</span></div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  weightLayer->GetOutputSlot(0).Disconnect(base.GetInputSlot(1));</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  weightLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(1));</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  weightLayer->m_LayerOutput = std::make_unique<ScopedTensorHandle>(fusedWeightsTensor);</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span> </div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="comment">// Move bias const layers as normal if it was enabled before the optimisation</span></div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  ConstantLayer* biasLayer;</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="keywordflow">if</span> (biasWasEnabledBeforeOpt)</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  {</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  biasLayer = PolymorphicDowncast<ConstantLayer*>(</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  &base.GetInputSlot(2).GetConnectedOutputSlot()->GetOwningLayer());</div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="comment">// Remove old connection and connect to new layer2d</span></div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  biasLayer->GetOutputSlot(0).Disconnect(base.GetInputSlot(2));</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  biasLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(2));</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span> </div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  }</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  <span class="comment">// Otherwise create a new bias layer and add to the new convolution2d</span></div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  {</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  <span class="comment">// Add in bias constant layer</span></div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  biasLayer = graph.AddLayer<ConstantLayer>(<span class="stringliteral">"Bias"</span>);</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  biasLayer->GetOutputSlot(0).SetTensorInfo(fusedBiasTensor.GetInfo());</div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  biasLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(2));</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>  biasLayer->m_LayerOutput = std::make_unique<ScopedTensorHandle>(ConstTensor(fusedBiasTensor));</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  }</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span> </div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span> </div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  <span class="comment">// Reconnects with original parent.</span></div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  newConv2dLayer.GetOutputSlot().MoveAllConnections(*parentOut);</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  <span class="comment">// Parent is now the new convolution2d layer.</span></div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  parentOut = &newConv2dLayer.GetOutputSlot();</div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span> </div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <span class="comment">// Moves connections in child output to parent layer.</span></div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="comment">// Child layer will be removed as it's left unconnected.</span></div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="comment">// Base layer will be removed if left unconnected.</span></div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  child.GetOutputSlot().MoveAllConnections(*parentOut);</div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  }</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  }</div><div class="ttc" id="namespacearmnn_xhtml_a44affeeb090c3c6a3062830562672e84"><div class="ttname"><a href="namespacearmnn.xhtml#a44affeeb090c3c6a3062830562672e84">armnn::IgnoreUnused</a></div><div class="ttdeci">void IgnoreUnused(Ts &&...)</div><div class="ttdef"><b>Definition:</b> <a href="_ignore_unused_8hpp_source.xhtml#l00014">IgnoreUnused.hpp:14</a></div></div> -<div class="ttc" id="_assert_8hpp_xhtml_a91c4dfde57907d7698c7531785690a7f"><div class="ttname"><a href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a></div><div class="ttdeci">#define ARMNN_ASSERT_MSG(COND, MSG)</div><div class="ttdef"><b>Definition:</b> <a href="_assert_8hpp_source.xhtml#l00015">Assert.hpp:15</a></div></div> -<div class="ttc" id="classarmnn_utils_1_1_data_layout_indexed_xhtml"><div class="ttname"><a href="classarmnn_utils_1_1_data_layout_indexed.xhtml">armnnUtils::DataLayoutIndexed</a></div><div class="ttdoc">Provides access to the appropriate indexes for Channels, Height and Width based on DataLayout...</div><div class="ttdef"><b>Definition:</b> <a href="_data_layout_indexed_8hpp_source.xhtml#l00017">DataLayoutIndexed.hpp:17</a></div></div> -<div class="ttc" id="namespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">armnn::LayerType::Convolution2d</a></div></div> -<div class="ttc" id="_assert_8hpp_xhtml_a5698be69cbd5dfe6c28fcd9867e8cbed"><div class="ttname"><a href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a></div><div class="ttdeci">#define ARMNN_ASSERT(COND)</div><div class="ttdef"><b>Definition:</b> <a href="_assert_8hpp_source.xhtml#l00014">Assert.hpp:14</a></div></div> -<div class="ttc" id="namespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">armnn::LayerType::DepthwiseConvolution2d</a></div></div> -<div class="ttc" id="namespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">armnn::LayerType::BatchNormalization</a></div></div> -<div class="ttc" id="namespacearmnn_xhtml_ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51"><div class="ttname"><a href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">armnn::DataLayout::NHWC</a></div></div> +<div class="fragment"><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>  Layer& base = connection.GetConnectedOutputSlot()->GetOwningLayer();</div> +<div class="line"><a name="l00030"></a><span class="lineno"> 30</span>  Layer& child = connection.GetOwningLayer();</div> +<div class="line"><a name="l00031"></a><span class="lineno"> 31</span>  </div> +<div class="line"><a name="l00032"></a><span class="lineno"> 32</span>  <span class="keywordtype">bool</span> depthwise = (base.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">LayerType::DepthwiseConvolution2d</a>);</div> +<div class="line"><a name="l00033"></a><span class="lineno"> 33</span>  </div> +<div class="line"><a name="l00034"></a><span class="lineno"> 34</span>  <a class="code" href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a>(base.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">LayerType::Convolution2d</a> || depthwise);</div> +<div class="line"><a name="l00035"></a><span class="lineno"> 35</span>  <a class="code" href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a>(child.GetType() == <a class="code" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">LayerType::BatchNormalization</a>);</div> +<div class="line"><a name="l00036"></a><span class="lineno"> 36</span>  </div> +<div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  <span class="keywordflow">if</span> (base.GetDataType() == ArmnnType && child.GetDataType() == ArmnnType)</div> +<div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  {</div> +<div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  OutputSlot* parentOut = base.GetInputSlot(0).GetConnectedOutputSlot();</div> +<div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  <span class="keyword">auto</span> convLayer = PolymorphicDowncast<ConvLayer*>(&base);</div> +<div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  <span class="keyword">auto</span> batchNormLayer = PolymorphicDowncast<BatchNormalizationLayer*>(&child);</div> +<div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  </div> +<div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  <span class="comment">// Read convolution and batch norm parameters</span></div> +<div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  BatchNormalizationDescriptor batchNormDescriptor = batchNormLayer->GetParameters();</div> +<div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  <span class="keyword">auto</span> epsilon = batchNormDescriptor.m_Eps;</div> +<div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  <a class="code" href="namespacearmnn.xhtml#a44affeeb090c3c6a3062830562672e84">IgnoreUnused</a>(epsilon);</div> +<div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  </div> +<div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  ConstTensor betaTensor(batchNormLayer->m_Beta->GetTensorInfo(), batchNormLayer->m_Beta->Map(<span class="keyword">true</span>));</div> +<div class="line"><a name="l00049"></a><span class="lineno"> 49</span>  ConstTensor gammaTensor(batchNormLayer->m_Gamma->GetTensorInfo(), batchNormLayer->m_Gamma->Map(<span class="keyword">true</span>));</div> +<div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  ConstTensor meanTensor(batchNormLayer->m_Mean->GetTensorInfo(), batchNormLayer->m_Mean->Map(<span class="keyword">true</span>));</div> +<div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  ConstTensor varTensor(batchNormLayer->m_Variance->GetTensorInfo(), batchNormLayer->m_Variance->Map(<span class="keyword">true</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> convDescriptor = convLayer->GetParameters();</div> +<div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  ConstTensor weightsTensor;</div> +<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <a class="code" href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a>(convLayer->GetInputSlots()[1].GetConnection() != <span class="keyword">nullptr</span>,</div> +<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="stringliteral">"FuseBatchNorm: Weight data should not be null."</span>);</div> +<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  </div> +<div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  ConstantLayer* weightLayer = PolymorphicDowncast<ConstantLayer*>(</div> +<div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  &base.GetInputSlot(1).GetConnectedOutputSlot()->GetOwningLayer());</div> +<div class="line"><a name="l00060"></a><span class="lineno"> 60</span>  </div> +<div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  weightsTensor = ConstTensor(weightLayer->m_LayerOutput->GetTensorInfo(),</div> +<div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  weightLayer->m_LayerOutput->Map(<span class="keyword">true</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>  <a class="code" href="classarmnn_utils_1_1_data_layout_indexed.xhtml">armnnUtils::DataLayoutIndexed</a> dataLayout(convDescriptor.m_DataLayout);</div> +<div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keyword">auto</span> weightsShape = weightsTensor.GetInfo().GetShape();</div> +<div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> inputChannels = parentOut->GetTensorInfo().GetShape()[dataLayout.GetChannelsIndex()];</div> +<div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> depthMultiplier = depthwise ? weightsShape[3] / inputChannels : 1;</div> +<div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> outputChannels = depthwise ? weightsShape[3] : weightsShape[0];</div> +<div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsHeight = depthwise ? weightsShape[1] :</div> +<div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  weightsShape[dataLayout.GetHeightIndex()];</div> +<div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsWidth = depthwise ? weightsShape[2] :</div> +<div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  weightsShape[dataLayout.GetWidthIndex()];</div> +<div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  </div> +<div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* weightsBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(weightsTensor.GetMemoryArea());</div> +<div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* betaBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(betaTensor.GetMemoryArea());</div> +<div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* gammaBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(gammaTensor.GetMemoryArea());</div> +<div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* meanBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(meanTensor.GetMemoryArea());</div> +<div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* varBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(varTensor.GetMemoryArea());</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>  std::vector<T> weightsVector (weightsBuffer, weightsBuffer + weightsTensor.GetNumElements());</div> +<div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  std::vector<T> betaVector (betaBuffer, betaBuffer + betaTensor.GetNumElements());</div> +<div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  std::vector<T> gammaVector (gammaBuffer, gammaBuffer + gammaTensor.GetNumElements());</div> +<div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  std::vector<T> meanVector (meanBuffer, meanBuffer + meanTensor.GetNumElements());</div> +<div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  std::vector<T> varianceVector(varBuffer, varBuffer + varTensor.GetNumElements());</div> +<div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  </div> +<div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="comment">// fusedWeights = ( gamma * weights ) / ( std - epsilon);</span></div> +<div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  std::vector<T> fusedWeightsVector(weightsVector.size());</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cInput = 0; cInput < inputChannels; ++cInput)</div> +<div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  {</div> +<div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div> +<div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  {</div> +<div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  T mult = gammaVector[cOut] / <span class="keyword">static_cast<</span>T<span class="keyword">></span>(sqrtf(varianceVector[cOut] + epsilon));</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="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> h = 0; h < weightsHeight; ++h)</div> +<div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  {</div> +<div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> w = 0; w < weightsWidth; ++w)</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>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weightsIdx = 0;</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="keywordflow">if</span> (depthwise)</div> +<div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  {</div> +<div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  cInput = cOut / depthMultiplier;</div> +<div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  weightsIdx = w * outputChannels + cOut +</div> +<div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  h * weightsWidth * outputChannels;</div> +<div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  }</div> +<div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span> (convDescriptor.m_DataLayout == <a class="code" href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">DataLayout::NHWC</a>)</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>  weightsIdx = cOut * weightsHeight * weightsWidth * inputChannels +</div> +<div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  h * weightsWidth * inputChannels +</div> +<div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  w * inputChannels +</div> +<div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  cInput;</div> +<div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  }</div> +<div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  <span class="keywordflow">else</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>  weightsIdx = cOut * weightsWidth * weightsHeight * inputChannels +</div> +<div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  cInput * weightsWidth * weightsHeight +</div> +<div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  h * weightsWidth +</div> +<div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  w;</div> +<div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  }</div> +<div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  fusedWeightsVector[weightsIdx] = mult * weightsVector[weightsIdx];</div> +<div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  }</div> +<div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  }</div> +<div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  }</div> +<div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  }</div> +<div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  ConstTensor fusedWeightsTensor(weightsTensor.GetInfo(), fusedWeightsVector);</div> +<div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  </div> +<div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="comment">// fusedBias = (gamma * (bias - mean)) / (variance - epsilon) + beta;</span></div> +<div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  std::vector<T> fusedBiasVector(outputChannels);</div> +<div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordtype">bool</span> biasWasEnabledBeforeOpt = convDescriptor.m_BiasEnabled;</div> +<div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  <span class="keywordflow">if</span> (biasWasEnabledBeforeOpt)</div> +<div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  {</div> +<div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  ConstTensor biasTensor;</div> +<div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  <a class="code" href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a>(convLayer->GetInputSlots()[2].GetConnection() != <span class="keyword">nullptr</span>,</div> +<div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  <span class="stringliteral">"FuseBatchNorm: Bias data should not be null if bias is enabled."</span>);</div> +<div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  </div> +<div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  ConstantLayer* biasLayer = PolymorphicDowncast<ConstantLayer*>(</div> +<div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  &base.GetInputSlot(2).GetConnectedOutputSlot()->GetOwningLayer());</div> +<div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  </div> +<div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  biasTensor = ConstTensor(biasLayer->m_LayerOutput->GetTensorInfo(),</div> +<div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  biasLayer->m_LayerOutput->Map(<span class="keyword">true</span>));</div> +<div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  </div> +<div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="keyword">const</span> <span class="keyword">auto</span>* biasBuffer = <span class="keyword">static_cast<</span><span class="keyword">const </span>T*<span class="keyword">></span>(biasTensor.GetMemoryArea());</div> +<div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  std::vector<T> biasVector(biasBuffer, biasBuffer + biasTensor.GetNumElements());</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div> +<div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  {</div> +<div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  fusedBiasVector[cOut] = ((gammaVector[cOut] * (biasVector[cOut] - meanVector[cOut])) /</div> +<div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  sqrtf(varianceVector[cOut] + epsilon)) + betaVector[cOut];</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>  }</div> +<div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="keywordflow">else</span></div> +<div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  {</div> +<div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  convDescriptor.m_BiasEnabled = <span class="keyword">true</span>;</div> +<div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  std::vector<T> biasVector(outputChannels, T(0));</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>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cOut = 0; cOut < outputChannels; ++cOut)</div> +<div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  {</div> +<div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  fusedBiasVector[cOut] = ((gammaVector[cOut] * (biasVector[cOut] - meanVector[cOut])) /</div> +<div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  sqrtf(varianceVector[cOut] + epsilon)) + betaVector[cOut];</div> +<div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  }</div> +<div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  }</div> +<div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  ConstTensor fusedBiasTensor(TensorInfo({outputChannels}, ArmnnType, 0.0f, 0, <span class="keyword">true</span>), fusedBiasVector);</div> +<div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  </div> +<div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// Insert the new convolution layer that has batch norm parameters fused into</span></div> +<div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="keyword">const</span> std::string name = std::string(<span class="stringliteral">"fused-"</span>) + child.GetName() + std::string(<span class="stringliteral">"-into-"</span>) + base.GetName();</div> +<div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="keyword">auto</span>& newConv2dLayer = *graph.InsertNewLayer<ConvLayer>(base.GetInputSlot(0),</div> +<div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  convDescriptor,</div> +<div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  name.c_str());</div> +<div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  </div> +<div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  <span class="comment">// Connect weights and bias from old to new Conv2d layer</span></div> +<div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  <span class="comment">// This optimization will always have 3 input slots on the Conv2d base layer</span></div> +<div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <span class="keywordflow">if</span> (newConv2dLayer.GetNumInputSlots() > 1)</div> +<div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  {</div> +<div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <span class="comment">// Remove old connection and connect to new layer2d</span></div> +<div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  weightLayer->GetOutputSlot(0).Disconnect(base.GetInputSlot(1));</div> +<div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  weightLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(1));</div> +<div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  weightLayer->m_LayerOutput = std::make_unique<ScopedTensorHandle>(fusedWeightsTensor);</div> +<div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  </div> +<div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="comment">// Move bias const layers as normal if it was enabled before the optimisation</span></div> +<div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  ConstantLayer* biasLayer;</div> +<div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="keywordflow">if</span> (biasWasEnabledBeforeOpt)</div> +<div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  {</div> +<div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  biasLayer = PolymorphicDowncast<ConstantLayer*>(</div> +<div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  &base.GetInputSlot(2).GetConnectedOutputSlot()->GetOwningLayer());</div> +<div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="comment">// Remove old connection and connect to new layer2d</span></div> +<div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  biasLayer->GetOutputSlot(0).Disconnect(base.GetInputSlot(2));</div> +<div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  biasLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(2));</div> +<div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  </div> +<div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  }</div> +<div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  <span class="comment">// Otherwise create a new bias layer and add to the new convolution2d</span></div> +<div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  <span class="keywordflow">else</span></div> +<div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  {</div> +<div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  <span class="comment">// Add in bias constant layer</span></div> +<div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  biasLayer = graph.AddLayer<ConstantLayer>(<span class="stringliteral">"Bias"</span>);</div> +<div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  biasLayer->GetOutputSlot(0).SetTensorInfo(fusedBiasTensor.GetInfo());</div> +<div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  biasLayer->GetOutputSlot(0).Connect(newConv2dLayer.GetInputSlot(2));</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>  biasLayer->m_LayerOutput = std::make_unique<ScopedTensorHandle>(ConstTensor(fusedBiasTensor));</div> +<div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  }</div> +<div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  </div> +<div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  </div> +<div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  <span class="comment">// Reconnects with original parent.</span></div> +<div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  newConv2dLayer.GetOutputSlot().MoveAllConnections(*parentOut);</div> +<div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  <span class="comment">// Parent is now the new convolution2d layer.</span></div> +<div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  parentOut = &newConv2dLayer.GetOutputSlot();</div> +<div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  </div> +<div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <span class="comment">// Moves connections in child output to parent layer.</span></div> +<div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="comment">// Child layer will be removed as it's left unconnected.</span></div> +<div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="comment">// Base layer will be removed if left unconnected.</span></div> +<div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  child.GetOutputSlot().MoveAllConnections(*parentOut);</div> +<div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  }</div> +<div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  }</div> </div><!-- fragment --> +<p class="reference">References <a class="el" href="_graph_8hpp_source.xhtml#l00456">Graph::AddLayer()</a>, <a class="el" href="_assert_8hpp_source.xhtml#l00014">ARMNN_ASSERT</a>, <a class="el" href="_assert_8hpp_source.xhtml#l00015">ARMNN_ASSERT_MSG</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">armnn::BatchNormalization</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00112">OutputSlot::Connect()</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">armnn::Convolution2d</a>, <a class="el" href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">armnn::DepthwiseConvolution2d</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00120">OutputSlot::Disconnect()</a>, <a class="el" href="_data_layout_indexed_8hpp_source.xhtml#l00023">DataLayoutIndexed::GetChannelsIndex()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00056">InputSlot::GetConnectedOutputSlot()</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00313">Layer::GetDataType()</a>, <a class="el" href="_data_layout_indexed_8hpp_source.xhtml#l00024">DataLayoutIndexed::GetHeightIndex()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00295">BaseTensor< MemoryType >::GetInfo()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00324">Layer::GetInputSlot()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00305">BaseTensor< MemoryType >::GetMemoryArea()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00319">Layer::GetName()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00303">BaseTensor< MemoryType >::GetNumElements()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00326">Layer::GetOutputSlot()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00053">InputSlot::GetOwningLayer()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00119">OutputSlot::GetOwningLayer()</a>, <a class="el" href="_tensor_8hpp_source.xhtml#l00191">TensorInfo::GetShape()</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00092">OutputSlot::GetTensorInfo()</a>, <a class="el" href="_layer_8hpp_source.xhtml#l00273">Layer::GetType()</a>, <a class="el" href="_data_layout_indexed_8hpp_source.xhtml#l00025">DataLayoutIndexed::GetWidthIndex()</a>, <a class="el" href="_ignore_unused_8hpp_source.xhtml#l00014">armnn::IgnoreUnused()</a>, <a class="el" href="_graph_8hpp_source.xhtml#l00471">Graph::InsertNewLayer()</a>, <a class="el" href="_descriptors_8hpp_source.xhtml#l00809">BatchNormalizationDescriptor::m_Eps</a>, <a class="el" href="_constant_layer_8hpp_source.xhtml#l00044">ConstantLayer::m_LayerOutput</a>, <a class="el" href="_layer_8cpp_source.xhtml#l00145">OutputSlot::MoveAllConnections()</a>, <a class="el" href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">armnn::NHWC</a>, and <a class="el" href="_layer_8cpp_source.xhtml#l00087">OutputSlot::SetTensorInfo()</a>.</p> + </div> </div> <hr/>The documentation for this class was generated from the following file:<ul> @@ -236,13 +413,21 @@ class armnn::optimizations::FuseBatchNorm< ConvLayer, ArmnnType, T ></h3> </ul> </div><!-- contents --> </div><!-- doc-content --> +<div class="ttc" id="anamespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4adb033d2f81b68f9a17e8f62de69fed4a">armnn::LayerType::Convolution2d</a></div><div class="ttdeci">@ Convolution2d</div></div> +<div class="ttc" id="anamespacearmnn_xhtml_a44affeeb090c3c6a3062830562672e84"><div class="ttname"><a href="namespacearmnn.xhtml#a44affeeb090c3c6a3062830562672e84">armnn::IgnoreUnused</a></div><div class="ttdeci">void IgnoreUnused(Ts &&...)</div><div class="ttdef"><b>Definition:</b> <a href="_ignore_unused_8hpp_source.xhtml#l00014">IgnoreUnused.hpp:14</a></div></div> +<div class="ttc" id="anamespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4ae4743c3ec15d1d84169b17264634692e">armnn::LayerType::BatchNormalization</a></div><div class="ttdeci">@ BatchNormalization</div></div> +<div class="ttc" id="anamespacearmnn_xhtml_a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7"><div class="ttname"><a href="namespacearmnn.xhtml#a56943a0946e5f15e5e58054b8e7a04a4af97adbfc88b7012a0243215b1076e7e7">armnn::LayerType::DepthwiseConvolution2d</a></div><div class="ttdeci">@ DepthwiseConvolution2d</div></div> +<div class="ttc" id="anamespacearmnn_xhtml_ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51"><div class="ttname"><a href="namespacearmnn.xhtml#ad1d5cce2d9e9a5d61c243e5c989112e0ad066db54b89b0912e7e7c6da51e2da51">armnn::DataLayout::NHWC</a></div><div class="ttdeci">@ NHWC</div></div> +<div class="ttc" id="a_assert_8hpp_xhtml_a91c4dfde57907d7698c7531785690a7f"><div class="ttname"><a href="_assert_8hpp.xhtml#a91c4dfde57907d7698c7531785690a7f">ARMNN_ASSERT_MSG</a></div><div class="ttdeci">#define ARMNN_ASSERT_MSG(COND, MSG)</div><div class="ttdef"><b>Definition:</b> <a href="_assert_8hpp_source.xhtml#l00015">Assert.hpp:15</a></div></div> +<div class="ttc" id="a_assert_8hpp_xhtml_a5698be69cbd5dfe6c28fcd9867e8cbed"><div class="ttname"><a href="_assert_8hpp.xhtml#a5698be69cbd5dfe6c28fcd9867e8cbed">ARMNN_ASSERT</a></div><div class="ttdeci">#define ARMNN_ASSERT(COND)</div><div class="ttdef"><b>Definition:</b> <a href="_assert_8hpp_source.xhtml#l00014">Assert.hpp:14</a></div></div> +<div class="ttc" id="aclassarmnn_utils_1_1_data_layout_indexed_xhtml"><div class="ttname"><a href="classarmnn_utils_1_1_data_layout_indexed.xhtml">armnnUtils::DataLayoutIndexed</a></div><div class="ttdoc">Provides access to the appropriate indexes for Channels, Height and Width based on DataLayout.</div><div class="ttdef"><b>Definition:</b> <a href="_data_layout_indexed_8hpp_source.xhtml#l00017">DataLayoutIndexed.hpp:17</a></div></div> <!-- start footer part --> <div id="nav-path" class="navpath"><!-- id is needed for treeview function! --> <ul> <li class="navelem"><a class="el" href="namespacearmnn.xhtml">armnn</a></li><li class="navelem"><a class="el" href="namespacearmnn_1_1optimizations.xhtml">optimizations</a></li><li class="navelem"><a class="el" href="classarmnn_1_1optimizations_1_1_fuse_batch_norm.xhtml">FuseBatchNorm</a></li> - <li class="footer">Generated on Fri Feb 24 2023 10:24:34 for ArmNN by + <li class="footer">Generated on Wed Mar 22 2023 15:53:09 for ArmNN by <a href="http://www.doxygen.org/index.html"> - <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li> + <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.17 </li> </ul> </div> </body> |