Gitiles
Code Review Sign In
gerrit-public.fairphone.software / platform / external / ComputeLibrary / 3d2d44ef55ab6b08afda8be48301ce3c55c7bc67 / . / documentation / depthwise__convolution_8cl_source.xhtml
blob: 0244ecaf3d41cab53332af0c00a2c088d75b3a82 [file] [log] [blame]
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]1<!-- HTML header for doxygen 1.8.9.1-->
2<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
3<html xmlns="http://www.w3.org/1999/xhtml">
4<head>
5<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
6<meta http-equiv="X-UA-Compatible" content="IE=9"/>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]7<meta name="generator" content="Doxygen 1.8.13"/>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]8<meta name="robots" content="NOINDEX, NOFOLLOW" /> <!-- Prevent indexing by search engines -->
9<title>Compute Library: src/core/CL/cl_kernels/depthwise_convolution.cl Source File</title>
10<link href="tabs.css" rel="stylesheet" type="text/css"/>
11<script type="text/javascript" src="jquery.js"></script>
12<script type="text/javascript" src="dynsections.js"></script>
13<link href="navtree.css" rel="stylesheet" type="text/css"/>
14<script type="text/javascript" src="resize.js"></script>
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000[diff] [blame]15<script type="text/javascript" src="navtreedata.js"></script>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]16<script type="text/javascript" src="navtree.js"></script>
17<script type="text/javascript">
18 $(document).ready(initResizable);
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]19</script>
20<link href="search/search.css" rel="stylesheet" type="text/css"/>
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000[diff] [blame]21<script type="text/javascript" src="search/searchdata.js"></script>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]22<script type="text/javascript" src="search/search.js"></script>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]23<script type="text/x-mathjax-config">
24 MathJax.Hub.Config({
25 extensions: ["tex2jax.js"],
26 jax: ["input/TeX","output/HTML-CSS"],
27});
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000[diff] [blame]28</script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]29<link href="doxygen.css" rel="stylesheet" type="text/css" />
30</head>
31<body>
32<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
33<div id="titlearea">
34<table cellspacing="0" cellpadding="0">
35 <tbody>
36 <tr style="height: 56px;">
37 <td style="padding-left: 0.5em;">
38 <div id="projectname">Compute Library
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]39 &#160;<span id="projectnumber">18.11</span>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]40 </div>
41 </td>
42 </tr>
43 </tbody>
44</table>
45</div>
46<!-- end header part -->
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]47<!-- Generated by Doxygen 1.8.13 -->
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]48<script type="text/javascript">
49var searchBox = new SearchBox("searchBox", "search",false,'Search');
50</script>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]51<script type="text/javascript" src="menudata.js"></script>
52<script type="text/javascript" src="menu.js"></script>
53<script type="text/javascript">
54$(function() {
55 initMenu('',true,false,'search.php','Search');
56 $(document).ready(function() { init_search(); });
57});
58</script>
59<div id="main-nav"></div>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]60</div><!-- top -->
61<div id="side-nav" class="ui-resizable side-nav-resizable">
62 <div id="nav-tree">
63 <div id="nav-tree-contents">
64 <div id="nav-sync" class="sync"></div>
65 </div>
66 </div>
67 <div id="splitbar" style="-moz-user-select:none;"
68 class="ui-resizable-handle">
69 </div>
70</div>
71<script type="text/javascript">
72$(document).ready(function(){initNavTree('depthwise__convolution_8cl_source.xhtml','');});
73</script>
74<div id="doc-content">
75<!-- window showing the filter options -->
76<div id="MSearchSelectWindow"
77 onmouseover="return searchBox.OnSearchSelectShow()"
78 onmouseout="return searchBox.OnSearchSelectHide()"
79 onkeydown="return searchBox.OnSearchSelectKey(event)">
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000[diff] [blame]80</div>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]81
82<!-- iframe showing the search results (closed by default) -->
83<div id="MSearchResultsWindow">
84<iframe src="javascript:void(0)" frameborder="0"
85 name="MSearchResults" id="MSearchResults">
86</iframe>
87</div>
88
89<div class="header">
90 <div class="headertitle">
91<div class="title">depthwise_convolution.cl</div> </div>
92</div><!--header-->
93<div class="contents">
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]94<a href="depthwise__convolution_8cl.xhtml">Go to the documentation of this file.</a><div class="fragment"><div class="line"><a name="l00001"></a><span class="lineno"> 1</span>&#160;<span class="comment">/*</span></div><div class="line"><a name="l00002"></a><span class="lineno"> 2</span>&#160;<span class="comment"> * Copyright (c) 2017-2018 ARM Limited.</span></div><div class="line"><a name="l00003"></a><span class="lineno"> 3</span>&#160;<span class="comment"> *</span></div><div class="line"><a name="l00004"></a><span class="lineno"> 4</span>&#160;<span class="comment"> * SPDX-License-Identifier: MIT</span></div><div class="line"><a name="l00005"></a><span class="lineno"> 5</span>&#160;<span class="comment"> *</span></div><div class="line"><a name="l00006"></a><span class="lineno"> 6</span>&#160;<span class="comment"> * Permission is hereby granted, free of charge, to any person obtaining a copy</span></div><div class="line"><a name="l00007"></a><span class="lineno"> 7</span>&#160;<span class="comment"> * of this software and associated documentation files (the &quot;Software&quot;), to</span></div><div class="line"><a name="l00008"></a><span class="lineno"> 8</span>&#160;<span class="comment"> * deal in the Software without restriction, including without limitation the</span></div><div class="line"><a name="l00009"></a><span class="lineno"> 9</span>&#160;<span class="comment"> * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or</span></div><div class="line"><a name="l00010"></a><span class="lineno"> 10</span>&#160;<span class="comment"> * sell copies of the Software, and to permit persons to whom the Software is</span></div><div class="line"><a name="l00011"></a><span class="lineno"> 11</span>&#160;<span class="comment"> * furnished to do so, subject to the following conditions:</span></div><div class="line"><a name="l00012"></a><span class="lineno"> 12</span>&#160;<span class="comment"> *</span></div><div class="line"><a name="l00013"></a><span class="lineno"> 13</span>&#160;<span class="comment"> * The above copyright notice and this permission notice shall be included in all</span></div><div class="line"><a name="l00014"></a><span class="lineno"> 14</span>&#160;<span class="comment"> * copies or substantial portions of the Software.</span></div><div class="line"><a name="l00015"></a><span class="lineno"> 15</span>&#160;<span class="comment"> *</span></div><div class="line"><a name="l00016"></a><span class="lineno"> 16</span>&#160;<span class="comment"> * THE SOFTWARE IS PROVIDED &quot;AS IS&quot;, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR</span></div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span>&#160;<span class="comment"> * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,</span></div><div class="line"><a name="l00018"></a><span class="lineno"> 18</span>&#160;<span class="comment"> * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE</span></div><div class="line"><a name="l00019"></a><span class="lineno"> 19</span>&#160;<span class="comment"> * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER</span></div><div class="line"><a name="l00020"></a><span class="lineno"> 20</span>&#160;<span class="comment"> * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,</span></div><div class="line"><a name="l00021"></a><span class="lineno"> 21</span>&#160;<span class="comment"> * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE</span></div><div class="line"><a name="l00022"></a><span class="lineno"> 22</span>&#160;<span class="comment"> * SOFTWARE.</span></div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span>&#160;<span class="comment"> */</span></div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span>&#160;</div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span>&#160;<span class="preprocessor">#include &quot;<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml">helpers.h</a>&quot;</span></div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span>&#160;</div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span>&#160;<span class="preprocessor">#if defined(DEPTH_MULTIPLIER) &amp;&amp; defined(DST_CHANNELS)</span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span>&#160;<span class="preprocessor">#if defined(CONV_STRIDE_X)</span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span>&#160;</div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span>&#160;<span class="preprocessor">#if CONV_STRIDE_X == 1</span></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span>&#160;<span class="preprocessor">#define convolution1x3 convolution1x3_stride_1</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span>&#160;<span class="preprocessor">#elif CONV_STRIDE_X == 2</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span>&#160;<span class="preprocessor">#define convolution1x3 convolution1x3_stride_2</span></div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span>&#160;<span class="preprocessor">#elif CONV_STRIDE_X == 3</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span>&#160;<span class="preprocessor">#define convolution1x3 convolution1x3_stride_3</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* CONV_STRIDE_X */</span><span class="preprocessor"></span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span>&#160;<span class="preprocessor">#error &quot;Stride not supported&quot;</span></div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* CONV_STRIDE_X */</span><span class="preprocessor"></span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span>&#160;</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span>&#160;<span class="keyword">inline</span> float2 convolution1x3_stride_1(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> left_coeff,</div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> middle_coeff,</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> right_coeff)</div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>&#160;{</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span>&#160; float4 temp = vload4(0, (__global <span class="keywordtype">float</span> *)left_pixel);</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>&#160;</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>&#160; float2 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s01, float2);</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span>&#160; float2 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s12, float2);</div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>&#160; float2 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s23, float2);</div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>&#160;</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span>&#160; <span class="keywordflow">return</span> left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;</div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>&#160;}</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span>&#160;</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>&#160;<span class="keyword">inline</span> float2 convolution1x3_stride_2(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> left_coeff,</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> middle_coeff,</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> right_coeff)</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>&#160;{</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>&#160; float4 temp0 = vload4(0, (__global <span class="keywordtype">float</span> *)left_pixel);</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>&#160; <span class="keywordtype">float</span> temp1 = *((__global <span class="keywordtype">float</span> *)(left_pixel + 4 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>)));</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>&#160;</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>&#160; float2 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s02, float2);</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>&#160; float2 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s13, float2);</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>&#160; float2 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>((float2)(temp0.s2, temp1), float2);</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>&#160;</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>&#160; <span class="keywordflow">return</span> left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>&#160;}</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>&#160;</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>&#160;<span class="keyword">inline</span> float2 convolution1x3_stride_3(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> left_coeff,</div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> middle_coeff,</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> right_coeff)</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>&#160;{</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>&#160; float4 temp0 = vload4(0, (__global <span class="keywordtype">float</span> *)left_pixel);</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span>&#160; float2 temp1 = vload2(0, (__global <span class="keywordtype">float</span> *)(left_pixel + 4 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>)));</div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span>&#160;</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>&#160; float2 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s03, float2);</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>&#160; float2 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>((float2)(temp0.s1, temp1.s0), float2);</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>&#160; float2 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>((float2)(temp0.s2, temp1.s1), float2);</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>&#160;</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>&#160; <span class="keywordflow">return</span> left * (float2)left_coeff + middle * (float2)middle_coeff + right * (float2)right_coeff;</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span>&#160;}</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>&#160;</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>&#160;<span class="keyword">inline</span> float2 <a class="code" href="convolution3x3_8cl.xhtml#afc5fefe72e66f0ae5191fd5b708fade9">convolution3x3</a>(</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>,</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> mat0, <span class="keyword">const</span> <span class="keywordtype">float</span> mat1, <span class="keyword">const</span> <span class="keywordtype">float</span> mat2,</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> mat3, <span class="keyword">const</span> <span class="keywordtype">float</span> mat4, <span class="keyword">const</span> <span class="keywordtype">float</span> mat5,</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>&#160; <span class="keyword">const</span> <span class="keywordtype">float</span> mat6, <span class="keyword">const</span> <span class="keywordtype">float</span> mat7, <span class="keyword">const</span> <span class="keywordtype">float</span> mat8)</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>&#160;{</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>&#160; float2 pixels;</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>&#160;</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>&#160; pixels = <a class="code" href="convolution3x3_8cl.xhtml#a92702074338198e81a46c3e309d9b04f">convolution1x3</a>(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 0), mat0, mat1, mat2);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>&#160; pixels += <a class="code" href="convolution3x3_8cl.xhtml#a92702074338198e81a46c3e309d9b04f">convolution1x3</a>(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 1), mat3, mat4, mat5);</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>&#160; pixels += <a class="code" href="convolution3x3_8cl.xhtml#a92702074338198e81a46c3e309d9b04f">convolution1x3</a>(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 2), mat6, mat7, mat8);</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>&#160;</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>&#160; <span class="keywordflow">return</span> pixels;</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>&#160;}</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>&#160;</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3(</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>&#160; ,</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>&#160;)</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>&#160;{</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>&#160;<span class="preprocessor">#endif //defined(HAS_BIAS)</span></div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>&#160;</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>&#160; <span class="comment">// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>&#160; src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>&#160; __global uchar *weights_addr = weights.ptr + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>&#160;</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>&#160; uchar3 <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = (uchar3)(0, 1, 2) * (uchar3)weights_stride_y;</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>&#160; float3 weights_values0 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + offset.s0));</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>&#160; float3 weights_values1 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + offset.s1));</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>&#160; float3 weights_values2 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + offset.s2));</div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>&#160;</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>&#160; float2 pixels = <a class="code" href="convolution3x3_8cl.xhtml#afc5fefe72e66f0ae5191fd5b708fade9">convolution3x3</a>(&amp;src, weights_values0.s0, weights_values0.s1, weights_values0.s2,</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>&#160; weights_values1.s0, weights_values1.s1, weights_values1.s2,</div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>&#160; weights_values2.s0, weights_values2.s1, weights_values2.s2);</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>&#160; pixels += (float2)(*((__global <span class="keywordtype">float</span> *)(biases.ptr + channel * biases_stride_x)));</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>&#160;<span class="preprocessor">#endif //defined(HAS_BIAS)</span></div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>&#160;</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>&#160; vstore2(pixels, 0, (__global <span class="keywordtype">float</span> *)<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr);</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span>&#160;}</div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>&#160;<span class="preprocessor">#endif //defined(CONV_STRIDE_X)</span></div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span>&#160;</div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>&#160;<span class="preprocessor">#define CONVOLUTION1x3_BIFROST2X1_STRIDE1(acc, src0, weights_row0) \</span></div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>&#160;<span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \</span></div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \</span></div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \</span></div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \</span></div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \</span></div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \</span></div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>&#160;<span class="preprocessor"> })</span></div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>&#160;</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>&#160;<span class="preprocessor">#define CONVOLUTION1x3_BIFROST4X1_STRIDE1(acc, src0, weights_row0) \</span></div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span>&#160;<span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \</span></div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \</span></div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \</span></div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s1, weights_row0.s0, acc.s1); \</span></div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s2, weights_row0.s1, acc.s1); \</span></div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s3, weights_row0.s2, acc.s1); \</span></div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s2, weights_row0.s0, acc.s2); \</span></div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s3, weights_row0.s1, acc.s2); \</span></div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s4, weights_row0.s2, acc.s2); \</span></div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>&#160;<span class="preprocessor"> acc.s3 = fma(src0.s3, weights_row0.s0, acc.s3); \</span></div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span>&#160;<span class="preprocessor"> acc.s3 = fma(src0.s4, weights_row0.s1, acc.s3); \</span></div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>&#160;<span class="preprocessor"> acc.s3 = fma(src0.s5, weights_row0.s2, acc.s3); \</span></div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span>&#160;<span class="preprocessor"> })</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>&#160;</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>&#160;<span class="preprocessor">#define CONVOLUTION1x3_BIFROST2X1_STRIDE2(acc, src0, src1, weights_row0) \</span></div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>&#160;<span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \</span></div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \</span></div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \</span></div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \</span></div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \</span></div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span>&#160;<span class="preprocessor"> acc.s1 = fma(src1.s0, weights_row0.s2, acc.s1); \</span></div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>&#160;<span class="preprocessor"> })</span></div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>&#160;</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>&#160;<span class="preprocessor">#define CONVOLUTION1x3_BIFROST4X1_STRIDE2(acc, src0, src1, weights_row0) \</span></div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>&#160;<span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s0, weights_row0.s0, acc.s0); \</span></div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s1, weights_row0.s1, acc.s0); \</span></div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span>&#160;<span class="preprocessor"> acc.s0 = fma(src0.s2, weights_row0.s2, acc.s0); \</span></div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s2, weights_row0.s0, acc.s1); \</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s3, weights_row0.s1, acc.s1); \</span></div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span>&#160;<span class="preprocessor"> acc.s1 = fma(src0.s4, weights_row0.s2, acc.s1); \</span></div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s4, weights_row0.s0, acc.s2); \</span></div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s5, weights_row0.s1, acc.s2); \</span></div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>&#160;<span class="preprocessor"> acc.s2 = fma(src0.s6, weights_row0.s2, acc.s2); \</span></div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>&#160;<span class="preprocessor"> acc.s3 = fma(src0.s6, weights_row0.s0, acc.s3); \</span></div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>&#160;<span class="preprocessor"> acc.s3 = fma(src0.s7, weights_row0.s1, acc.s3); \</span></div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>&#160;<span class="preprocessor"> acc.s3 = fma(src1.s0, weights_row0.s2, acc.s3); \</span></div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>&#160;<span class="preprocessor"> })</span></div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>&#160;</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32(</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>&#160; ,</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>&#160;)</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>&#160;{</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>&#160;</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span>&#160; float2 pixels0 = 0.0f;</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span>&#160; float2 pixels1 = 0.0f;</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>&#160; float2 pixels2 = 0.0f;</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>&#160; float2 pixels3 = 0.0f;</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>&#160;</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>&#160; <span class="comment">// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>&#160; __global uchar *weights_addr = weights.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span>&#160; __global uchar *src_addr = src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>&#160;</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span>&#160; <span class="comment">// Load the weights</span></div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>&#160; float3 weights_row0 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 0 * weights_stride_y));</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>&#160; float3 weights_row1 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>&#160; float3 weights_row2 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>&#160;</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>&#160; <span class="comment">// Note: Since each work-item computes 4x2 elements, we need to load 6 rows from the input tensor</span></div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>&#160; float4 src00 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>&#160; float4 src10 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>&#160; float4 src20 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>&#160; float4 src30 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>&#160; float4 src40 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>&#160; float4 src50 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 5 * src_stride_y)); <span class="comment">// Row5</span></div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>&#160;</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src00, weights_row0);</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src10, weights_row1);</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels0, src20, weights_row2);</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src10, weights_row0);</div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src20, weights_row1);</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels1, src30, weights_row2);</div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src20, weights_row0);</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src30, weights_row1);</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels2, src40, weights_row2);</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src30, weights_row0);</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src40, weights_row1);</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE1(pixels3, src50, weights_row2);</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>&#160;</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>&#160;</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>&#160; <span class="keywordtype">float</span> bias = *((__global <span class="keywordtype">float</span> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&amp;biases, channel)));</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>&#160;</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>&#160; pixels0 += (float2)bias;</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>&#160; pixels1 += (float2)bias;</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>&#160; pixels2 += (float2)bias;</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>&#160; pixels3 += (float2)bias;</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>&#160;</div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>&#160; vstore2(pixels0, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 0 * dst_stride_y));</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>&#160; vstore2(pixels1, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 1 * dst_stride_y));</div><div class="line"><a name="l00368"></a><span class="lineno"> 368</span>&#160; vstore2(pixels2, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 2 * dst_stride_y));</div><div class="line"><a name="l00369"></a><span class="lineno"> 369</span>&#160; vstore2(pixels3, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 3 * dst_stride_y));</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>&#160;}</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>&#160;</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32(</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>&#160; ,</div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>&#160;)</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>&#160;{</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>&#160;</div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span>&#160; float2 pixels0 = 0.0f;</div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>&#160; float2 pixels1 = 0.0f;</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span>&#160;</div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00422"></a><span class="lineno"> 422</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>&#160; <span class="comment">// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span>&#160; __global uchar *weights_addr = weights.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span>&#160; __global uchar *src_addr = src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>&#160;</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span>&#160; <span class="comment">// Load the weights</span></div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>&#160; float3 weights_row0 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 0 * weights_stride_y));</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>&#160; float3 weights_row1 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>&#160; float3 weights_row2 = vload3(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span>&#160;</div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span>&#160; <span class="comment">// Note: Since each work-item computes 4x2 elements, we need to load 5 rows from the input tensor</span></div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span>&#160; float4 src00 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>&#160; float2 src01 = vload2(2, (__global <span class="keywordtype">float</span> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>&#160; float4 src10 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>&#160; float2 src11 = vload2(2, (__global <span class="keywordtype">float</span> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>&#160; float4 src20 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span>&#160; float2 src21 = vload2(2, (__global <span class="keywordtype">float</span> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>&#160; float4 src30 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>&#160; float2 src31 = vload2(2, (__global <span class="keywordtype">float</span> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>&#160; float4 src40 = vload4(0, (__global <span class="keywordtype">float</span> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>&#160; float2 src41 = vload2(2, (__global <span class="keywordtype">float</span> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>&#160;</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src00, src01, weights_row0);</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src10, src11, weights_row1);</div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels0, src20, src21, weights_row2);</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src20, src21, weights_row0);</div><div class="line"><a name="l00449"></a><span class="lineno"> 449</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src30, src31, weights_row1);</div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>&#160; CONVOLUTION1x3_BIFROST2X1_STRIDE2(pixels1, src40, src41, weights_row2);</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>&#160;</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>&#160;</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>&#160; <span class="keywordtype">float</span> bias = *((__global <span class="keywordtype">float</span> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&amp;biases, channel)));</div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>&#160;</div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>&#160; pixels0 += (float2)bias;</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>&#160; pixels1 += (float2)bias;</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>&#160;</div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span>&#160; vstore2(pixels0, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 0 * dst_stride_y));</div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>&#160; vstore2(pixels1, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>.ptr + 1 * dst_stride_y));</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>&#160;}</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>&#160;</div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>&#160;<span class="preprocessor">#endif // defined(DEPTH_MULTIPLIER) &amp;&amp; defined(DST_CHANNELS)</span></div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>&#160;</div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span>&#160;<span class="preprocessor">#if defined(NCHW)</span></div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>&#160;<span class="preprocessor">#define in_stride_x src_stride_x</span></div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>&#160;<span class="preprocessor">#define in_stride_y src_stride_y</span></div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>&#160;<span class="preprocessor">#define in_stride_z src_stride_z</span></div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>&#160;<span class="preprocessor">#define out_stride_x dst_stride_x</span></div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>&#160;<span class="preprocessor">#define out_stride_y dst_stride_y</span></div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>&#160;<span class="preprocessor">#define out_stride_z dst_stride_z</span></div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>&#160;<span class="preprocessor">#else //defined(NCHW)</span></div><div class="line"><a name="l00475"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#a5447e2c4e1aae8fff9290b775fe813ff"> 475</a></span>&#160;<span class="preprocessor">#define in_stride_x src_stride_y</span></div><div class="line"><a name="l00476"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#ac353c0687abd8c71a4fa096d322b39ec"> 476</a></span>&#160;<span class="preprocessor">#define in_stride_y src_stride_z</span></div><div class="line"><a name="l00477"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#a32137dd220c1110408a7b498108de2c5"> 477</a></span>&#160;<span class="preprocessor">#define in_stride_z src_stride_x</span></div><div class="line"><a name="l00478"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#a68fce1f0ee4080ed89c9427496d625f2"> 478</a></span>&#160;<span class="preprocessor">#define out_stride_x dst_stride_y</span></div><div class="line"><a name="l00479"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#ac8b6b732372937bba276e6391944d1e1"> 479</a></span>&#160;<span class="preprocessor">#define out_stride_y dst_stride_z</span></div><div class="line"><a name="l00480"></a><span class="lineno"><a class="line" href="depthwise__convolution_8cl.xhtml#a128f47cb6aacde29e07fde2c4b9f5dd2"> 480</a></span>&#160;<span class="preprocessor">#define out_stride_z dst_stride_x</span></div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span>&#160;<span class="preprocessor">#endif //defined(NCHW)</span></div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span>&#160;</div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>&#160;<span class="preprocessor">#if defined(SRC_WIDTH) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>&#160;</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_weights_reshape(</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>)</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>&#160;#ifdef HAS_BIAS</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>&#160; ,</div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>&#160;#endif <span class="comment">/* HAS_BIAS */</span></div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>&#160;)</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>&#160;{</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* HAS_BIAS */</span><span class="preprocessor"></span></div><div class="line"><a name="l00519"></a><span class="lineno"> 519</span>&#160;</div><div class="line"><a name="l00520"></a><span class="lineno"> 520</span>&#160; __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * <a class="code" href="depthwise__convolution_8cl.xhtml#ac353c0687abd8c71a4fa096d322b39ec">in_stride_y</a> + get_global_id(2) * <a class="code" href="depthwise__convolution_8cl.xhtml#a32137dd220c1110408a7b498108de2c5">in_stride_z</a>;</div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span>&#160; __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(1) * SRC_WIDTH * dst_stride_x + get_global_id(2) * dst_stride_y;</div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>&#160;</div><div class="line"><a name="l00523"></a><span class="lineno"> 523</span>&#160; <span class="keywordflow">for</span>(<span class="keywordtype">int</span> i = 0; i &lt; SRC_WIDTH; ++i, input_ptr += <a class="code" href="depthwise__convolution_8cl.xhtml#a5447e2c4e1aae8fff9290b775fe813ff">in_stride_x</a>)</div><div class="line"><a name="l00524"></a><span class="lineno"> 524</span>&#160; {</div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span>&#160; *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(output_ptr + i * dst_stride_x)) = *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)input_ptr);</div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span>&#160; }</div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span>&#160;</div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span>&#160; <span class="keywordflow">if</span>(get_global_id(1) == 0)</div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>&#160; {</div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span>&#160; *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(output_ptr + SRC_WIDTH * get_global_size(1) * dst_stride_x)) = *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(2) * biases_stride_x));</div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span>&#160; }</div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>&#160;<span class="preprocessor">#endif // defined(HAS_BIAS)</span></div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span>&#160;}</div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span>&#160;<span class="preprocessor">#endif //defined(SRC_WIDTH) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>&#160;</div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>&#160;<span class="preprocessor">#if defined(STRIDE_X) &amp;&amp; defined(STRIDE_Y) &amp;&amp; defined(PAD_LEFT) &amp;&amp; defined(PAD_TOP) &amp;&amp; defined(PAD_RIGHT) &amp;&amp; defined(PAD_BOTTOM) &amp;&amp; defined(KERNEL_WIDTH) &amp;&amp; defined(KERNEL_HEIGHT) &amp;&amp; defined(SRC_WIDTH) &amp;&amp; defined(SRC_HEIGHT) &amp;&amp; defined(DATA_TYPE) &amp;&amp; defined(PAD_VALUE) &amp;&amp; defined(DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>&#160;</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_im2col(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src), <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>))</div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span>&#160;{</div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>&#160;</div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> src_pixel_linear = get_global_id(1) * STRIDE_X;</div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> full_length = SRC_WIDTH + PAD_LEFT + PAD_RIGHT;</div><div class="line"><a name="l00566"></a><span class="lineno"> 566</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> max_initial_x = STRIDE_X * (((full_length - KERNEL_WIDTH) / STRIDE_X) + 1);</div><div class="line"><a name="l00567"></a><span class="lineno"> 567</span>&#160;</div><div class="line"><a name="l00568"></a><span class="lineno"> 568</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> src_x = -PAD_LEFT + src_pixel_linear % max_initial_x;</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> src_y = -PAD_TOP + src_pixel_linear / max_initial_x * STRIDE_Y;</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> src_z = get_global_id(2) / DEPTH_MULTIPLIER;</div><div class="line"><a name="l00571"></a><span class="lineno"> 571</span>&#160;</div><div class="line"><a name="l00572"></a><span class="lineno"> 572</span>&#160; __global uchar *input_ptr = src_ptr + src_offset_first_element_in_bytes + src_z * <a class="code" href="depthwise__convolution_8cl.xhtml#a32137dd220c1110408a7b498108de2c5">in_stride_z</a>;</div><div class="line"><a name="l00573"></a><span class="lineno"> 573</span>&#160; __global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *output_ptr = ((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>));</div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span>&#160;</div><div class="line"><a name="l00575"></a><span class="lineno"> 575</span>&#160; <span class="keywordflow">for</span>(<span class="keywordtype">int</span> y = src_y; y &lt; src_y + KERNEL_HEIGHT; ++y)</div><div class="line"><a name="l00576"></a><span class="lineno"> 576</span>&#160; {</div><div class="line"><a name="l00577"></a><span class="lineno"> 577</span>&#160; <span class="keywordflow">for</span>(<span class="keywordtype">int</span> x = src_x; x &lt; src_x + KERNEL_WIDTH; ++x, ++output_ptr)</div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span>&#160; {</div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>&#160; <span class="keywordflow">if</span>(x &lt; 0 || x &gt;= SRC_WIDTH || y &lt; 0 || y &gt;= SRC_HEIGHT)</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span>&#160; {</div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span>&#160; *output_ptr = PAD_VALUE;</div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span>&#160; }</div><div class="line"><a name="l00583"></a><span class="lineno"> 583</span>&#160; <span class="keywordflow">else</span></div><div class="line"><a name="l00584"></a><span class="lineno"> 584</span>&#160; {</div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span>&#160; *output_ptr = *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(input_ptr + x * <a class="code" href="depthwise__convolution_8cl.xhtml#a5447e2c4e1aae8fff9290b775fe813ff">in_stride_x</a> + y * <a class="code" href="depthwise__convolution_8cl.xhtml#ac353c0687abd8c71a4fa096d322b39ec">in_stride_y</a>));</div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>&#160; }</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span>&#160; }</div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span>&#160; }</div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span>&#160; *output_ptr = (<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>)(1);</div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>&#160;<span class="preprocessor">#endif // defined(HAS_BIAS)</span></div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>&#160;}</div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span>&#160;</div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span>&#160;<span class="preprocessor">#endif //defined(STRIDE_X) &amp;&amp; defined(STRIDE_Y) &amp;&amp; defined(PAD_LEFT) &amp;&amp; defined(PAD_TOP) &amp;&amp; defined(PAD_RIGHT) &amp;&amp; defined(PAD_BOTTOM) &amp;&amp; defined(KERNEL_WIDTH) &amp;&amp; defined(KERNEL_HEIGHT) &amp;&amp; defined(SRC_WIDTH) &amp;&amp; defined(DATA_TYPE) &amp;&amp; defined(PAD_VALUE) &amp;&amp; defined(DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>&#160;</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span>&#160;<span class="preprocessor">#if defined(CONV_WIDTH) &amp;&amp; defined(CONV_HEIGHT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>&#160;</div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_vector_to_tensor(</div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(src),</div><div class="line"><a name="l00618"></a><span class="lineno"> 618</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>))</div><div class="line"><a name="l00619"></a><span class="lineno"> 619</span>&#160;{</div><div class="line"><a name="l00620"></a><span class="lineno"> 620</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(src);</div><div class="line"><a name="l00621"></a><span class="lineno"> 621</span>&#160;</div><div class="line"><a name="l00622"></a><span class="lineno"> 622</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> patch_size = CONV_WIDTH * CONV_HEIGHT;</div><div class="line"><a name="l00623"></a><span class="lineno"> 623</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> id0 = get_global_id(0);</div><div class="line"><a name="l00624"></a><span class="lineno"> 624</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> z = id0 / patch_size;</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> index2D = id0 - z * patch_size;</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span>&#160;</div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span>&#160; __global uchar *out_ptr = dst_ptr + dst_offset_first_element_in_bytes + index2D % CONV_WIDTH * <a class="code" href="depthwise__convolution_8cl.xhtml#a68fce1f0ee4080ed89c9427496d625f2">out_stride_x</a> + index2D / CONV_WIDTH * <a class="code" href="depthwise__convolution_8cl.xhtml#ac8b6b732372937bba276e6391944d1e1">out_stride_y</a> + z * <a class="code" href="depthwise__convolution_8cl.xhtml#a128f47cb6aacde29e07fde2c4b9f5dd2">out_stride_z</a>;</div><div class="line"><a name="l00628"></a><span class="lineno"> 628</span>&#160; *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)out_ptr) = *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)src.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00629"></a><span class="lineno"> 629</span>&#160;}</div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span>&#160;</div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>&#160;<span class="preprocessor">#endif //defined(CONV_WIDTH) &amp;&amp; defined(CONV_HEIGHT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span>&#160;</div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span>&#160;<span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) &amp;&amp; defined(DEPTH_MULTIPLIER) &amp;&amp; defined(DST_CHANNELS)</span></div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span>&#160;<span class="preprocessor">#if defined(CONV_STRIDE_X)</span></div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>&#160;<span class="preprocessor">#if CONV_STRIDE_X == 1</span></div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span>&#160;<span class="preprocessor">#define convolution1x3_f16 convolution1x3_stride_1_f16</span></div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>&#160;<span class="preprocessor">#elif CONV_STRIDE_X == 2</span></div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>&#160;<span class="preprocessor">#define convolution1x3_f16 convolution1x3_stride_2_f16</span></div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>&#160;<span class="preprocessor">#elif CONV_STRIDE_X == 3</span></div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>&#160;<span class="preprocessor">#define convolution1x3_f16 convolution1x3_stride_3_f16</span></div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* CONV_STRIDE_X */</span><span class="preprocessor"></span></div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span>&#160;<span class="preprocessor">#error &quot;Stride not supported&quot;</span></div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* CONV_STRIDE_X */</span><span class="preprocessor"></span></div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span>&#160;</div><div class="line"><a name="l00654"></a><span class="lineno"> 654</span>&#160;<span class="keyword">inline</span> half4 convolution1x3_stride_1_f16(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> left_coeff,</div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> middle_coeff,</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> right_coeff)</div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>&#160;{</div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span>&#160; half8 temp = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)left_pixel);</div><div class="line"><a name="l00660"></a><span class="lineno"> 660</span>&#160;</div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span>&#160; half4 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s0123, half4);</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>&#160; half4 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s1234, half4);</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span>&#160; half4 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp.s2345, half4);</div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span>&#160;</div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span>&#160; <span class="keywordflow">return</span> left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;</div><div class="line"><a name="l00666"></a><span class="lineno"> 666</span>&#160;}</div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span>&#160;</div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>&#160;<span class="keyword">inline</span> half4 convolution1x3_stride_2_f16(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> left_coeff,</div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> middle_coeff,</div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> right_coeff)</div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>&#160;{</div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span>&#160; half8 temp0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)left_pixel);</div><div class="line"><a name="l00683"></a><span class="lineno"> 683</span>&#160; <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> temp1 = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(left_pixel + 8 * <span class="keyword">sizeof</span>(<a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a>)));</div><div class="line"><a name="l00684"></a><span class="lineno"> 684</span>&#160;</div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span>&#160; half4 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s0246, half4);</div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span>&#160; half4 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s1357, half4);</div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>&#160; half4 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>((half4)(temp0.s246, temp1), half4);</div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span>&#160;</div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span>&#160; <span class="keywordflow">return</span> left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;</div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span>&#160;}</div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>&#160;</div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span>&#160;<span class="keyword">inline</span> half4 convolution1x3_stride_3_f16(__global <span class="keyword">const</span> uchar *left_pixel,</div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> left_coeff,</div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> middle_coeff,</div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> right_coeff)</div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span>&#160;{</div><div class="line"><a name="l00706"></a><span class="lineno"> 706</span>&#160; half16 temp0 = vload16(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)left_pixel);</div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span>&#160;</div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>&#160; half4 left = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s0369, half4);</div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span>&#160; half4 middle = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s147A, half4);</div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>&#160; half4 right = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(temp0.s258B, half4);</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span>&#160;</div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>&#160; <span class="keywordflow">return</span> left * (half4)left_coeff + middle * (half4)middle_coeff + right * (half4)right_coeff;</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span>&#160;}</div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span>&#160;</div><div class="line"><a name="l00737"></a><span class="lineno"> 737</span>&#160;<span class="keyword">inline</span> half4 convolution3x3_f16(</div><div class="line"><a name="l00738"></a><span class="lineno"> 738</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> *src,</div><div class="line"><a name="l00739"></a><span class="lineno"> 739</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat0, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat1, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat2,</div><div class="line"><a name="l00740"></a><span class="lineno"> 740</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat3, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat4, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat5,</div><div class="line"><a name="l00741"></a><span class="lineno"> 741</span>&#160; <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat6, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat7, <span class="keyword">const</span> <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> mat8)</div><div class="line"><a name="l00742"></a><span class="lineno"> 742</span>&#160;{</div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span>&#160; half4 pixels;</div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span>&#160;</div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span>&#160; pixels = convolution1x3_f16(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 0), mat0, mat1, mat2);</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span>&#160; pixels += convolution1x3_f16(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 1), mat3, mat4, mat5);</div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>&#160; pixels += convolution1x3_f16(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(src, 0, 2), mat6, mat7, mat8);</div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span>&#160;</div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span>&#160; <span class="keywordflow">return</span> pixels;</div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>&#160;}</div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span>&#160;</div><div class="line"><a name="l00752"></a><span class="lineno"> 752</span>&#160;<span class="preprocessor">#if defined(DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00753"></a><span class="lineno"> 753</span>&#160;</div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_f16(</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00787"></a><span class="lineno"> 787</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00788"></a><span class="lineno"> 788</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00789"></a><span class="lineno"> 789</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00790"></a><span class="lineno"> 790</span>&#160; ,</div><div class="line"><a name="l00791"></a><span class="lineno"> 791</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00792"></a><span class="lineno"> 792</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00793"></a><span class="lineno"> 793</span>&#160;)</div><div class="line"><a name="l00794"></a><span class="lineno"> 794</span>&#160;{</div><div class="line"><a name="l00795"></a><span class="lineno"> 795</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00796"></a><span class="lineno"> 796</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00797"></a><span class="lineno"> 797</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00798"></a><span class="lineno"> 798</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00799"></a><span class="lineno"> 799</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00800"></a><span class="lineno"> 800</span>&#160;<span class="preprocessor">#endif //defined(HAS_BIAS)</span></div><div class="line"><a name="l00801"></a><span class="lineno"> 801</span>&#160;</div><div class="line"><a name="l00802"></a><span class="lineno"> 802</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00803"></a><span class="lineno"> 803</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00804"></a><span class="lineno"> 804</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00805"></a><span class="lineno"> 805</span>&#160; <span class="comment">// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00806"></a><span class="lineno"> 806</span>&#160; src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> -= batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z + (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00807"></a><span class="lineno"> 807</span>&#160; __global uchar *weights_addr = weights.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00808"></a><span class="lineno"> 808</span>&#160;</div><div class="line"><a name="l00809"></a><span class="lineno"> 809</span>&#160; uchar3 offset = (uchar3)(0, 1, 2) * (uchar3)weights_stride_y;</div><div class="line"><a name="l00810"></a><span class="lineno"> 810</span>&#160; half3 weights_values0 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + offset.s0));</div><div class="line"><a name="l00811"></a><span class="lineno"> 811</span>&#160; half3 weights_values1 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + offset.s1));</div><div class="line"><a name="l00812"></a><span class="lineno"> 812</span>&#160; half3 weights_values2 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + offset.s2));</div><div class="line"><a name="l00813"></a><span class="lineno"> 813</span>&#160;</div><div class="line"><a name="l00814"></a><span class="lineno"> 814</span>&#160; half4 pixels = convolution3x3_f16(&amp;src, weights_values0.s0, weights_values0.s1, weights_values0.s2,</div><div class="line"><a name="l00815"></a><span class="lineno"> 815</span>&#160; weights_values1.s0, weights_values1.s1, weights_values1.s2,</div><div class="line"><a name="l00816"></a><span class="lineno"> 816</span>&#160; weights_values2.s0, weights_values2.s1, weights_values2.s2);</div><div class="line"><a name="l00817"></a><span class="lineno"> 817</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l00818"></a><span class="lineno"> 818</span>&#160; pixels += (half4)(*((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + channel * biases_stride_x)));</div><div class="line"><a name="l00819"></a><span class="lineno"> 819</span>&#160;<span class="preprocessor">#endif //defined(HAS_BIAS)</span></div><div class="line"><a name="l00820"></a><span class="lineno"> 820</span>&#160;</div><div class="line"><a name="l00821"></a><span class="lineno"> 821</span>&#160; vstore4(pixels, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00822"></a><span class="lineno"> 822</span>&#160;}</div><div class="line"><a name="l00823"></a><span class="lineno"> 823</span>&#160;<span class="preprocessor">#endif // defined(DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00824"></a><span class="lineno"> 824</span>&#160;<span class="preprocessor">#endif // defined(CONV_STRIDE_X)</span></div><div class="line"><a name="l00825"></a><span class="lineno"> 825</span>&#160;</div><div class="line"><a name="l00858"></a><span class="lineno"> 858</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16(</div><div class="line"><a name="l00859"></a><span class="lineno"> 859</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00860"></a><span class="lineno"> 860</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00861"></a><span class="lineno"> 861</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00862"></a><span class="lineno"> 862</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00863"></a><span class="lineno"> 863</span>&#160; ,</div><div class="line"><a name="l00864"></a><span class="lineno"> 864</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00865"></a><span class="lineno"> 865</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00866"></a><span class="lineno"> 866</span>&#160;)</div><div class="line"><a name="l00867"></a><span class="lineno"> 867</span>&#160;{</div><div class="line"><a name="l00868"></a><span class="lineno"> 868</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00869"></a><span class="lineno"> 869</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00870"></a><span class="lineno"> 870</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00871"></a><span class="lineno"> 871</span>&#160;</div><div class="line"><a name="l00872"></a><span class="lineno"> 872</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00873"></a><span class="lineno"> 873</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00874"></a><span class="lineno"> 874</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00875"></a><span class="lineno"> 875</span>&#160;</div><div class="line"><a name="l00876"></a><span class="lineno"> 876</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00877"></a><span class="lineno"> 877</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00878"></a><span class="lineno"> 878</span>&#160;</div><div class="line"><a name="l00879"></a><span class="lineno"> 879</span>&#160; <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> bias = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&amp;biases, channel)));</div><div class="line"><a name="l00880"></a><span class="lineno"> 880</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00881"></a><span class="lineno"> 881</span>&#160;</div><div class="line"><a name="l00882"></a><span class="lineno"> 882</span>&#160; half4 pixels0 = 0.0f;</div><div class="line"><a name="l00883"></a><span class="lineno"> 883</span>&#160; half4 pixels1 = 0.0f;</div><div class="line"><a name="l00884"></a><span class="lineno"> 884</span>&#160; half4 pixels2 = 0.0f;</div><div class="line"><a name="l00885"></a><span class="lineno"> 885</span>&#160; half4 pixels3 = 0.0f;</div><div class="line"><a name="l00886"></a><span class="lineno"> 886</span>&#160;</div><div class="line"><a name="l00887"></a><span class="lineno"> 887</span>&#160; <span class="comment">// Load relevant input and weights data (Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00888"></a><span class="lineno"> 888</span>&#160; __global uchar *weights_addr = weights.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00889"></a><span class="lineno"> 889</span>&#160; __global uchar *src_addr = src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00890"></a><span class="lineno"> 890</span>&#160;</div><div class="line"><a name="l00891"></a><span class="lineno"> 891</span>&#160; <span class="comment">// Load the weights</span></div><div class="line"><a name="l00892"></a><span class="lineno"> 892</span>&#160; half3 weights_row0 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 0 * weights_stride_y));</div><div class="line"><a name="l00893"></a><span class="lineno"> 893</span>&#160; half3 weights_row1 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00894"></a><span class="lineno"> 894</span>&#160; half3 weights_row2 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00895"></a><span class="lineno"> 895</span>&#160;</div><div class="line"><a name="l00896"></a><span class="lineno"> 896</span>&#160; <span class="comment">// Note: Since each work-item computes 4x4 elements, we need to load 6 rows from the input tensor</span></div><div class="line"><a name="l00897"></a><span class="lineno"> 897</span>&#160; half8 src00 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l00898"></a><span class="lineno"> 898</span>&#160; half8 src10 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l00899"></a><span class="lineno"> 899</span>&#160; half8 src20 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l00900"></a><span class="lineno"> 900</span>&#160; half8 src30 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l00901"></a><span class="lineno"> 901</span>&#160; half8 src40 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l00902"></a><span class="lineno"> 902</span>&#160; half8 src50 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 5 * src_stride_y)); <span class="comment">// Row5</span></div><div class="line"><a name="l00903"></a><span class="lineno"> 903</span>&#160;</div><div class="line"><a name="l00904"></a><span class="lineno"> 904</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src00, weights_row0);</div><div class="line"><a name="l00905"></a><span class="lineno"> 905</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src10, weights_row1);</div><div class="line"><a name="l00906"></a><span class="lineno"> 906</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels0, src20, weights_row2);</div><div class="line"><a name="l00907"></a><span class="lineno"> 907</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src10, weights_row0);</div><div class="line"><a name="l00908"></a><span class="lineno"> 908</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src20, weights_row1);</div><div class="line"><a name="l00909"></a><span class="lineno"> 909</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels1, src30, weights_row2);</div><div class="line"><a name="l00910"></a><span class="lineno"> 910</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src20, weights_row0);</div><div class="line"><a name="l00911"></a><span class="lineno"> 911</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src30, weights_row1);</div><div class="line"><a name="l00912"></a><span class="lineno"> 912</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels2, src40, weights_row2);</div><div class="line"><a name="l00913"></a><span class="lineno"> 913</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src30, weights_row0);</div><div class="line"><a name="l00914"></a><span class="lineno"> 914</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src40, weights_row1);</div><div class="line"><a name="l00915"></a><span class="lineno"> 915</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE1(pixels3, src50, weights_row2);</div><div class="line"><a name="l00916"></a><span class="lineno"> 916</span>&#160;</div><div class="line"><a name="l00917"></a><span class="lineno"> 917</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00918"></a><span class="lineno"> 918</span>&#160; pixels0 += (half4)bias;</div><div class="line"><a name="l00919"></a><span class="lineno"> 919</span>&#160; pixels1 += (half4)bias;</div><div class="line"><a name="l00920"></a><span class="lineno"> 920</span>&#160; pixels2 += (half4)bias;</div><div class="line"><a name="l00921"></a><span class="lineno"> 921</span>&#160; pixels3 += (half4)bias;</div><div class="line"><a name="l00922"></a><span class="lineno"> 922</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00923"></a><span class="lineno"> 923</span>&#160;</div><div class="line"><a name="l00924"></a><span class="lineno"> 924</span>&#160; vstore4(pixels0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l00925"></a><span class="lineno"> 925</span>&#160; vstore4(pixels1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l00926"></a><span class="lineno"> 926</span>&#160; vstore4(pixels2, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l00927"></a><span class="lineno"> 927</span>&#160; vstore4(pixels3, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l00928"></a><span class="lineno"> 928</span>&#160;}</div><div class="line"><a name="l00929"></a><span class="lineno"> 929</span>&#160;</div><div class="line"><a name="l00962"></a><span class="lineno"> 962</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16(</div><div class="line"><a name="l00963"></a><span class="lineno"> 963</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00964"></a><span class="lineno"> 964</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00965"></a><span class="lineno"> 965</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights)</div><div class="line"><a name="l00966"></a><span class="lineno"> 966</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l00967"></a><span class="lineno"> 967</span>&#160; ,</div><div class="line"><a name="l00968"></a><span class="lineno"> 968</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l00969"></a><span class="lineno"> 969</span>&#160;#endif <span class="comment">//defined(HAS_BIAS)</span></div><div class="line"><a name="l00970"></a><span class="lineno"> 970</span>&#160;)</div><div class="line"><a name="l00971"></a><span class="lineno"> 971</span>&#160;{</div><div class="line"><a name="l00972"></a><span class="lineno"> 972</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00973"></a><span class="lineno"> 973</span>&#160; <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00974"></a><span class="lineno"> 974</span>&#160; <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00975"></a><span class="lineno"> 975</span>&#160;</div><div class="line"><a name="l00976"></a><span class="lineno"> 976</span>&#160; <span class="comment">// Extract channel and linearized batch indices</span></div><div class="line"><a name="l00977"></a><span class="lineno"> 977</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> channel = get_global_id(2) % DST_CHANNELS;</div><div class="line"><a name="l00978"></a><span class="lineno"> 978</span>&#160; <span class="keyword">const</span> <span class="keywordtype">int</span> batch = get_global_id(2) / DST_CHANNELS;</div><div class="line"><a name="l00979"></a><span class="lineno"> 979</span>&#160;</div><div class="line"><a name="l00980"></a><span class="lineno"> 980</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00981"></a><span class="lineno"> 981</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00982"></a><span class="lineno"> 982</span>&#160;</div><div class="line"><a name="l00983"></a><span class="lineno"> 983</span>&#160; <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> bias = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&amp;biases, channel)));</div><div class="line"><a name="l00984"></a><span class="lineno"> 984</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00985"></a><span class="lineno"> 985</span>&#160;</div><div class="line"><a name="l00986"></a><span class="lineno"> 986</span>&#160; half4 pixels0 = 0.0f;</div><div class="line"><a name="l00987"></a><span class="lineno"> 987</span>&#160; half4 pixels1 = 0.0f;</div><div class="line"><a name="l00988"></a><span class="lineno"> 988</span>&#160;</div><div class="line"><a name="l00989"></a><span class="lineno"> 989</span>&#160; <span class="comment">// Load relevant input and weights data ( Accounts depth multiplier when indexing input, OFM = IFM * DEPTH_MULTIPLIER)</span></div><div class="line"><a name="l00990"></a><span class="lineno"> 990</span>&#160; __global uchar *weights_addr = weights.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * weights_step_x + get_global_id(1) * weights_step_y + channel * weights_step_z;</div><div class="line"><a name="l00991"></a><span class="lineno"> 991</span>&#160; __global uchar *src_addr = src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> - batch * (DST_CHANNELS / DEPTH_MULTIPLIER) * (DEPTH_MULTIPLIER - 1) * src_step_z - (channel - (channel / DEPTH_MULTIPLIER)) * src_step_z;</div><div class="line"><a name="l00992"></a><span class="lineno"> 992</span>&#160;</div><div class="line"><a name="l00993"></a><span class="lineno"> 993</span>&#160; <span class="comment">// Load the weights</span></div><div class="line"><a name="l00994"></a><span class="lineno"> 994</span>&#160; half3 weights_row0 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 0 * weights_stride_y));</div><div class="line"><a name="l00995"></a><span class="lineno"> 995</span>&#160; half3 weights_row1 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00996"></a><span class="lineno"> 996</span>&#160; half3 weights_row2 = vload3(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00997"></a><span class="lineno"> 997</span>&#160;</div><div class="line"><a name="l00998"></a><span class="lineno"> 998</span>&#160; <span class="comment">// Note: Since each work-item computes 2x4 elements, we need to load 5 rows from the input tensor</span></div><div class="line"><a name="l00999"></a><span class="lineno"> 999</span>&#160; half8 src00 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l01000"></a><span class="lineno"> 1000</span>&#160; half2 src01 = vload2(4, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 0 * src_stride_y)); <span class="comment">// Row0</span></div><div class="line"><a name="l01001"></a><span class="lineno"> 1001</span>&#160; half8 src10 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l01002"></a><span class="lineno"> 1002</span>&#160; half2 src11 = vload2(4, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 1 * src_stride_y)); <span class="comment">// Row1</span></div><div class="line"><a name="l01003"></a><span class="lineno"> 1003</span>&#160; half8 src20 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l01004"></a><span class="lineno"> 1004</span>&#160; half2 src21 = vload2(4, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 2 * src_stride_y)); <span class="comment">// Row2</span></div><div class="line"><a name="l01005"></a><span class="lineno"> 1005</span>&#160; half8 src30 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l01006"></a><span class="lineno"> 1006</span>&#160; half2 src31 = vload2(4, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 3 * src_stride_y)); <span class="comment">// Row3</span></div><div class="line"><a name="l01007"></a><span class="lineno"> 1007</span>&#160; half8 src40 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l01008"></a><span class="lineno"> 1008</span>&#160; half2 src41 = vload2(4, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src_addr + 4 * src_stride_y)); <span class="comment">// Row4</span></div><div class="line"><a name="l01009"></a><span class="lineno"> 1009</span>&#160;</div><div class="line"><a name="l01010"></a><span class="lineno"> 1010</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src00, src01, weights_row0);</div><div class="line"><a name="l01011"></a><span class="lineno"> 1011</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src10, src11, weights_row1);</div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels0, src20, src21, weights_row2);</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src20, src21, weights_row0);</div><div class="line"><a name="l01014"></a><span class="lineno"> 1014</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src30, src31, weights_row1);</div><div class="line"><a name="l01015"></a><span class="lineno"> 1015</span>&#160; CONVOLUTION1x3_BIFROST4X1_STRIDE2(pixels1, src40, src41, weights_row2);</div><div class="line"><a name="l01016"></a><span class="lineno"> 1016</span>&#160;</div><div class="line"><a name="l01017"></a><span class="lineno"> 1017</span>&#160;<span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l01018"></a><span class="lineno"> 1018</span>&#160; pixels0 += (half4)bias;</div><div class="line"><a name="l01019"></a><span class="lineno"> 1019</span>&#160; pixels1 += (half4)bias;</div><div class="line"><a name="l01020"></a><span class="lineno"> 1020</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01021"></a><span class="lineno"> 1021</span>&#160;</div><div class="line"><a name="l01022"></a><span class="lineno"> 1022</span>&#160; vstore4(pixels0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l01023"></a><span class="lineno"> 1023</span>&#160; vstore4(pixels1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l01024"></a><span class="lineno"> 1024</span>&#160;}</div><div class="line"><a name="l01025"></a><span class="lineno"> 1025</span>&#160;<span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) &amp;&amp; defined(DEPTH_MULTIPLIER) &amp;&amp; defined(DST_CHANNELS)</span></div><div class="line"><a name="l01026"></a><span class="lineno"> 1026</span>&#160;</div><div class="line"><a name="l01027"></a><span class="lineno"> 1027</span>&#160;<span class="preprocessor">#if defined(VEC_SIZE) &amp;&amp; defined(SRC_DIM_2) &amp;&amp; defined(CONV_PAD_TOP) &amp;&amp; defined(CONV_PAD_LEFT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l01028"></a><span class="lineno"> 1028</span>&#160;</div><div class="line"><a name="l01029"></a><span class="lineno"> 1029</span>&#160;<span class="preprocessor">#if DATA_TYPE != float || DATA_TYPE != half</span></div><div class="line"><a name="l01030"></a><span class="lineno"> 1030</span>&#160;<span class="preprocessor">#error &quot;Unsupported data type&quot;</span></div><div class="line"><a name="l01031"></a><span class="lineno"> 1031</span>&#160;<span class="preprocessor">#endif // DATA_TYPE != float || DATA_TYPE != half</span></div><div class="line"><a name="l01032"></a><span class="lineno"> 1032</span>&#160;</div><div class="line"><a name="l01033"></a><span class="lineno"> 1033</span>&#160;<span class="preprocessor">#define VEC_FLOAT VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)</span></div><div class="line"><a name="l01034"></a><span class="lineno"> 1034</span>&#160;</div><div class="line"><a name="l01035"></a><span class="lineno"> 1035</span>&#160;<span class="preprocessor">#if defined(CONV_STRIDE_X) &amp;&amp; defined(CONV_STRIDE_Y)</span></div><div class="line"><a name="l01036"></a><span class="lineno"> 1036</span>&#160;</div><div class="line"><a name="l01080"></a><span class="lineno"> 1080</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_nhwc(</div><div class="line"><a name="l01081"></a><span class="lineno"> 1081</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a>(src),</div><div class="line"><a name="l01082"></a><span class="lineno"> 1082</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l01083"></a><span class="lineno"> 1083</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights),</div><div class="line"><a name="l01084"></a><span class="lineno"> 1084</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l01085"></a><span class="lineno"> 1085</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l01086"></a><span class="lineno"> 1086</span>&#160;#endif <span class="comment">/* defined(HAS_BIAS) */</span></div><div class="line"><a name="l01087"></a><span class="lineno"> 1087</span>&#160; <span class="keywordtype">int</span> max_offset)</div><div class="line"><a name="l01088"></a><span class="lineno"> 1088</span>&#160;{</div><div class="line"><a name="l01089"></a><span class="lineno"> 1089</span>&#160; <span class="keywordtype">int</span> x = get_global_id(0); <span class="comment">// channels</span></div><div class="line"><a name="l01090"></a><span class="lineno"> 1090</span>&#160; <span class="keywordtype">int</span> y = get_global_id(1); <span class="comment">// spatial coordinate x</span></div><div class="line"><a name="l01091"></a><span class="lineno"> 1091</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01092"></a><span class="lineno"> 1092</span>&#160; <span class="keywordtype">int</span> z = get_global_id(2) % (int)DST_DEPTH; <span class="comment">// spatial coordinate y</span></div><div class="line"><a name="l01093"></a><span class="lineno"> 1093</span>&#160; <span class="keywordtype">int</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> = get_global_id(2) / (int)DST_DEPTH; <span class="comment">// batch</span></div><div class="line"><a name="l01094"></a><span class="lineno"> 1094</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01095"></a><span class="lineno"> 1095</span>&#160; <span class="keywordtype">int</span> z = get_global_id(2); <span class="comment">// spatial coordinate y</span></div><div class="line"><a name="l01096"></a><span class="lineno"> 1096</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01097"></a><span class="lineno"> 1097</span>&#160;</div><div class="line"><a name="l01098"></a><span class="lineno"> 1098</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(weights);</div><div class="line"><a name="l01099"></a><span class="lineno"> 1099</span>&#160;</div><div class="line"><a name="l01100"></a><span class="lineno"> 1100</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01101"></a><span class="lineno"> 1101</span>&#160; __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * VEC_SIZE + b * src_stride_w;</div><div class="line"><a name="l01102"></a><span class="lineno"> 1102</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01103"></a><span class="lineno"> 1103</span>&#160; __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * VEC_SIZE;</div><div class="line"><a name="l01104"></a><span class="lineno"> 1104</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01105"></a><span class="lineno"> 1105</span>&#160;</div><div class="line"><a name="l01106"></a><span class="lineno"> 1106</span>&#160; <span class="keywordtype">int</span> z_coord = 0;</div><div class="line"><a name="l01107"></a><span class="lineno"> 1107</span>&#160; int4 offset = 0;</div><div class="line"><a name="l01108"></a><span class="lineno"> 1108</span>&#160; int4 y_offset = ((int4)(y * CONV_STRIDE_X) + (int4)(0, 1, 2, 3) - CONV_PAD_LEFT) * (int4)src_stride_y;</div><div class="line"><a name="l01109"></a><span class="lineno"> 1109</span>&#160;</div><div class="line"><a name="l01110"></a><span class="lineno"> 1110</span>&#160; <span class="comment">// We compute 2x1x1 [C,W,H] elements</span></div><div class="line"><a name="l01111"></a><span class="lineno"> 1111</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> acc = 0;</div><div class="line"><a name="l01112"></a><span class="lineno"> 1112</span>&#160;</div><div class="line"><a name="l01113"></a><span class="lineno"> 1113</span>&#160; <span class="comment">// Load weights</span></div><div class="line"><a name="l01114"></a><span class="lineno"> 1114</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01115"></a><span class="lineno"> 1115</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01116"></a><span class="lineno"> 1116</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w2 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01117"></a><span class="lineno"> 1117</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w3 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01118"></a><span class="lineno"> 1118</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w4 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01119"></a><span class="lineno"> 1119</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w5 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01120"></a><span class="lineno"> 1120</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w6 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01121"></a><span class="lineno"> 1121</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w7 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01122"></a><span class="lineno"> 1122</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w8 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01123"></a><span class="lineno"> 1123</span>&#160;</div><div class="line"><a name="l01124"></a><span class="lineno"> 1124</span>&#160; <span class="comment">// Load input values</span></div><div class="line"><a name="l01125"></a><span class="lineno"> 1125</span>&#160; <span class="comment">// z == 0</span></div><div class="line"><a name="l01126"></a><span class="lineno"> 1126</span>&#160; <span class="comment">// Clamp z_coord as for z = 0, it can be negative</span></div><div class="line"><a name="l01127"></a><span class="lineno"> 1127</span>&#160; <span class="comment">// z_coord is casted to unsigned int in order to use just a min() operation</span></div><div class="line"><a name="l01128"></a><span class="lineno"> 1128</span>&#160; <span class="comment">// A &quot;-1&quot; 32 bit signed variable converted to unsigned gives 4294967295</span></div><div class="line"><a name="l01129"></a><span class="lineno"> 1129</span>&#160; z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP;</div><div class="line"><a name="l01130"></a><span class="lineno"> 1130</span>&#160; z_coord = min((uint)z_coord, (uint)SRC_DIM_2);</div><div class="line"><a name="l01131"></a><span class="lineno"> 1131</span>&#160; offset = y_offset + (int4)(z_coord * src_stride_z);</div><div class="line"><a name="l01132"></a><span class="lineno"> 1132</span>&#160; offset = min(offset, (int4)max_offset);</div><div class="line"><a name="l01133"></a><span class="lineno"> 1133</span>&#160;</div><div class="line"><a name="l01134"></a><span class="lineno"> 1134</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01135"></a><span class="lineno"> 1135</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01136"></a><span class="lineno"> 1136</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values2 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01137"></a><span class="lineno"> 1137</span>&#160;</div><div class="line"><a name="l01138"></a><span class="lineno"> 1138</span>&#160; <span class="comment">// z == 1</span></div><div class="line"><a name="l01139"></a><span class="lineno"> 1139</span>&#160; <span class="comment">// z_coord can be only negative for z = 0 so we do not need to clamp it</span></div><div class="line"><a name="l01140"></a><span class="lineno"> 1140</span>&#160; <span class="comment">// Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset</span></div><div class="line"><a name="l01141"></a><span class="lineno"> 1141</span>&#160; z_coord = z * CONV_STRIDE_Y - (int)CONV_PAD_TOP + 1;</div><div class="line"><a name="l01142"></a><span class="lineno"> 1142</span>&#160; offset = y_offset + (int4)(z_coord * src_stride_z);</div><div class="line"><a name="l01143"></a><span class="lineno"> 1143</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values3 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01144"></a><span class="lineno"> 1144</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values4 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01145"></a><span class="lineno"> 1145</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values5 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01146"></a><span class="lineno"> 1146</span>&#160;</div><div class="line"><a name="l01147"></a><span class="lineno"> 1147</span>&#160; <span class="comment">// z == 2</span></div><div class="line"><a name="l01148"></a><span class="lineno"> 1148</span>&#160; <span class="comment">// After z = 1 we can simply add src_stride_z to offset without updating z_coord</span></div><div class="line"><a name="l01149"></a><span class="lineno"> 1149</span>&#160; <span class="comment">// However offset can be out-of-bound so we need to check if it is greater than max_offset</span></div><div class="line"><a name="l01150"></a><span class="lineno"> 1150</span>&#160; offset += (int4)src_stride_z;</div><div class="line"><a name="l01151"></a><span class="lineno"> 1151</span>&#160; offset = min(offset, (int4)max_offset);</div><div class="line"><a name="l01152"></a><span class="lineno"> 1152</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values6 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01153"></a><span class="lineno"> 1153</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values7 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01154"></a><span class="lineno"> 1154</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values8 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01155"></a><span class="lineno"> 1155</span>&#160;</div><div class="line"><a name="l01156"></a><span class="lineno"> 1156</span>&#160; acc = fma(values0, w0, acc);</div><div class="line"><a name="l01157"></a><span class="lineno"> 1157</span>&#160; acc = fma(values1, w1, acc);</div><div class="line"><a name="l01158"></a><span class="lineno"> 1158</span>&#160; acc = fma(values2, w2, acc);</div><div class="line"><a name="l01159"></a><span class="lineno"> 1159</span>&#160;</div><div class="line"><a name="l01160"></a><span class="lineno"> 1160</span>&#160; acc = fma(values3, w3, acc);</div><div class="line"><a name="l01161"></a><span class="lineno"> 1161</span>&#160; acc = fma(values4, w4, acc);</div><div class="line"><a name="l01162"></a><span class="lineno"> 1162</span>&#160; acc = fma(values5, w5, acc);</div><div class="line"><a name="l01163"></a><span class="lineno"> 1163</span>&#160;</div><div class="line"><a name="l01164"></a><span class="lineno"> 1164</span>&#160; acc = fma(values6, w6, acc);</div><div class="line"><a name="l01165"></a><span class="lineno"> 1165</span>&#160; acc = fma(values7, w7, acc);</div><div class="line"><a name="l01166"></a><span class="lineno"> 1166</span>&#160; acc = fma(values8, w8, acc);</div><div class="line"><a name="l01167"></a><span class="lineno"> 1167</span>&#160;</div><div class="line"><a name="l01168"></a><span class="lineno"> 1168</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l01169"></a><span class="lineno"> 1169</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l01170"></a><span class="lineno"> 1170</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> bias_values = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l01171"></a><span class="lineno"> 1171</span>&#160; acc += bias_values;</div><div class="line"><a name="l01172"></a><span class="lineno"> 1172</span>&#160;<span class="preprocessor">#endif // defined(HAS_BIAS)</span></div><div class="line"><a name="l01173"></a><span class="lineno"> 1173</span>&#160;</div><div class="line"><a name="l01174"></a><span class="lineno"> 1174</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01175"></a><span class="lineno"> 1175</span>&#160; __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z + b * dst_stride_w;</div><div class="line"><a name="l01176"></a><span class="lineno"> 1176</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01177"></a><span class="lineno"> 1177</span>&#160; __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + z * dst_step_z;</div><div class="line"><a name="l01178"></a><span class="lineno"> 1178</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01179"></a><span class="lineno"> 1179</span>&#160;</div><div class="line"><a name="l01180"></a><span class="lineno"> 1180</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VEC_SIZE)</div><div class="line"><a name="l01181"></a><span class="lineno"> 1181</span>&#160; (acc, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_addr));</div><div class="line"><a name="l01182"></a><span class="lineno"> 1182</span>&#160;}</div><div class="line"><a name="l01183"></a><span class="lineno"> 1183</span>&#160;<span class="preprocessor">#endif // defined(CONV_STRIDE_X) &amp;&amp; defined(CONV_STRIDE_Y)</span></div><div class="line"><a name="l01184"></a><span class="lineno"> 1184</span>&#160;</div><div class="line"><a name="l01185"></a><span class="lineno"> 1185</span>&#160;<span class="preprocessor">#if defined(NUM_ROWS_PROCESSED) &amp;&amp; defined(NUM_PLANES_PROCESSED)</span></div><div class="line"><a name="l01186"></a><span class="lineno"> 1186</span>&#160;</div><div class="line"><a name="l01229"></a><span class="lineno"> 1229</span>&#160;__kernel <span class="keywordtype">void</span> depthwise_convolution_3x3_nhwc_stride1(</div><div class="line"><a name="l01230"></a><span class="lineno"> 1230</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a>(src),</div><div class="line"><a name="l01231"></a><span class="lineno"> 1231</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l01232"></a><span class="lineno"> 1232</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights),</div><div class="line"><a name="l01233"></a><span class="lineno"> 1233</span>&#160;#<span class="keywordflow">if</span> defined(HAS_BIAS)</div><div class="line"><a name="l01234"></a><span class="lineno"> 1234</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l01235"></a><span class="lineno"> 1235</span>&#160;#endif <span class="comment">/* defined(HAS_BIAS) */</span></div><div class="line"><a name="l01236"></a><span class="lineno"> 1236</span>&#160; <span class="keywordtype">int</span> max_offset)</div><div class="line"><a name="l01237"></a><span class="lineno"> 1237</span>&#160;{</div><div class="line"><a name="l01238"></a><span class="lineno"> 1238</span>&#160; <span class="keywordtype">int</span> x = get_global_id(0); <span class="comment">// channels</span></div><div class="line"><a name="l01239"></a><span class="lineno"> 1239</span>&#160; <span class="keywordtype">int</span> y = get_global_id(1); <span class="comment">// spatial coordinate x</span></div><div class="line"><a name="l01240"></a><span class="lineno"> 1240</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01241"></a><span class="lineno"> 1241</span>&#160; <span class="keywordtype">int</span> z = get_global_id(2) % (int)DST_DEPTH; <span class="comment">// spatial coordinate y</span></div><div class="line"><a name="l01242"></a><span class="lineno"> 1242</span>&#160; <span class="keywordtype">int</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">b</a> = get_global_id(2) / (int)DST_DEPTH; <span class="comment">// batch</span></div><div class="line"><a name="l01243"></a><span class="lineno"> 1243</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01244"></a><span class="lineno"> 1244</span>&#160; <span class="keywordtype">int</span> z = get_global_id(2); <span class="comment">// spatial coordinate y</span></div><div class="line"><a name="l01245"></a><span class="lineno"> 1245</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01246"></a><span class="lineno"> 1246</span>&#160;</div><div class="line"><a name="l01247"></a><span class="lineno"> 1247</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(weights);</div><div class="line"><a name="l01248"></a><span class="lineno"> 1248</span>&#160;</div><div class="line"><a name="l01249"></a><span class="lineno"> 1249</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01250"></a><span class="lineno"> 1250</span>&#160; __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * VEC_SIZE + b * src_stride_w;</div><div class="line"><a name="l01251"></a><span class="lineno"> 1251</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01252"></a><span class="lineno"> 1252</span>&#160; __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * VEC_SIZE;</div><div class="line"><a name="l01253"></a><span class="lineno"> 1253</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01254"></a><span class="lineno"> 1254</span>&#160;</div><div class="line"><a name="l01255"></a><span class="lineno"> 1255</span>&#160; <span class="keywordtype">int</span> z_coord = 0;</div><div class="line"><a name="l01256"></a><span class="lineno"> 1256</span>&#160; int4 offset = 0;</div><div class="line"><a name="l01257"></a><span class="lineno"> 1257</span>&#160; int4 y_offset = ((int4)(y * NUM_ROWS_PROCESSED) + (int4)(0, 1, 2, 3) - (int)CONV_PAD_LEFT) * (int4)src_stride_y;</div><div class="line"><a name="l01258"></a><span class="lineno"> 1258</span>&#160;</div><div class="line"><a name="l01259"></a><span class="lineno"> 1259</span>&#160; <span class="comment">// We compute 2x2x2 [C,W,H] elements</span></div><div class="line"><a name="l01260"></a><span class="lineno"> 1260</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> acc0 = 0;</div><div class="line"><a name="l01261"></a><span class="lineno"> 1261</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> acc1 = 0;</div><div class="line"><a name="l01262"></a><span class="lineno"> 1262</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> acc2 = 0;</div><div class="line"><a name="l01263"></a><span class="lineno"> 1263</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> acc3 = 0;</div><div class="line"><a name="l01264"></a><span class="lineno"> 1264</span>&#160;</div><div class="line"><a name="l01265"></a><span class="lineno"> 1265</span>&#160; <span class="comment">// Load weights</span></div><div class="line"><a name="l01266"></a><span class="lineno"> 1266</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01267"></a><span class="lineno"> 1267</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01268"></a><span class="lineno"> 1268</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w2 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 0 * weights_stride_z));</div><div class="line"><a name="l01269"></a><span class="lineno"> 1269</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w3 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01270"></a><span class="lineno"> 1270</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w4 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01271"></a><span class="lineno"> 1271</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w5 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 1 * weights_stride_z));</div><div class="line"><a name="l01272"></a><span class="lineno"> 1272</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w6 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01273"></a><span class="lineno"> 1273</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w7 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01274"></a><span class="lineno"> 1274</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> w8 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * weights_stride_y + 2 * weights_stride_z));</div><div class="line"><a name="l01275"></a><span class="lineno"> 1275</span>&#160;</div><div class="line"><a name="l01276"></a><span class="lineno"> 1276</span>&#160; <span class="comment">// Load input values</span></div><div class="line"><a name="l01277"></a><span class="lineno"> 1277</span>&#160; <span class="comment">// z == 0</span></div><div class="line"><a name="l01278"></a><span class="lineno"> 1278</span>&#160; <span class="comment">// Clamp z_coord as for z = 0, it can be negative</span></div><div class="line"><a name="l01279"></a><span class="lineno"> 1279</span>&#160; <span class="comment">// z_coord is casted to unsigned int in order to use just a min() operation</span></div><div class="line"><a name="l01280"></a><span class="lineno"> 1280</span>&#160; <span class="comment">// A &quot;-1&quot; 32 bit signed variable converted to unsigned gives 4294967295</span></div><div class="line"><a name="l01281"></a><span class="lineno"> 1281</span>&#160; z_coord = z * (int)NUM_PLANES_PROCESSED - (<span class="keywordtype">int</span>)CONV_PAD_TOP;</div><div class="line"><a name="l01282"></a><span class="lineno"> 1282</span>&#160; z_coord = min((uint)z_coord, (uint)SRC_DIM_2);</div><div class="line"><a name="l01283"></a><span class="lineno"> 1283</span>&#160; offset = y_offset + (int4)(z_coord * src_stride_z);</div><div class="line"><a name="l01284"></a><span class="lineno"> 1284</span>&#160; offset = min(offset, (int4)max_offset);</div><div class="line"><a name="l01285"></a><span class="lineno"> 1285</span>&#160;</div><div class="line"><a name="l01286"></a><span class="lineno"> 1286</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01287"></a><span class="lineno"> 1287</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01288"></a><span class="lineno"> 1288</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values2 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01289"></a><span class="lineno"> 1289</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values3 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s3));</div><div class="line"><a name="l01290"></a><span class="lineno"> 1290</span>&#160;</div><div class="line"><a name="l01291"></a><span class="lineno"> 1291</span>&#160; <span class="comment">// z == 1</span></div><div class="line"><a name="l01292"></a><span class="lineno"> 1292</span>&#160; <span class="comment">// z_coord can be only negative for z = 0 so we do not need to clamp it</span></div><div class="line"><a name="l01293"></a><span class="lineno"> 1293</span>&#160; <span class="comment">// Moreover z_coord cannot be out-of-bound for z = 1 so we do not need to clamp the offset</span></div><div class="line"><a name="l01294"></a><span class="lineno"> 1294</span>&#160; z_coord = z * (int)NUM_PLANES_PROCESSED - (<span class="keywordtype">int</span>)CONV_PAD_TOP + 1;</div><div class="line"><a name="l01295"></a><span class="lineno"> 1295</span>&#160; offset = y_offset + (int4)(z_coord * src_stride_z);</div><div class="line"><a name="l01296"></a><span class="lineno"> 1296</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values4 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01297"></a><span class="lineno"> 1297</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values5 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01298"></a><span class="lineno"> 1298</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values6 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01299"></a><span class="lineno"> 1299</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values7 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s3));</div><div class="line"><a name="l01300"></a><span class="lineno"> 1300</span>&#160;</div><div class="line"><a name="l01301"></a><span class="lineno"> 1301</span>&#160; <span class="comment">// z == 2</span></div><div class="line"><a name="l01302"></a><span class="lineno"> 1302</span>&#160; <span class="comment">// After z = 1 we can simply add src_stride_z to offset without updating z_coord</span></div><div class="line"><a name="l01303"></a><span class="lineno"> 1303</span>&#160; <span class="comment">// However offset can be out-of-bound so we need to check if it is greater than max_offset</span></div><div class="line"><a name="l01304"></a><span class="lineno"> 1304</span>&#160; offset += (int4)src_stride_z;</div><div class="line"><a name="l01305"></a><span class="lineno"> 1305</span>&#160; offset = min(offset, (int4)max_offset);</div><div class="line"><a name="l01306"></a><span class="lineno"> 1306</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values8 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01307"></a><span class="lineno"> 1307</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values9 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01308"></a><span class="lineno"> 1308</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values10 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01309"></a><span class="lineno"> 1309</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values11 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s3));</div><div class="line"><a name="l01310"></a><span class="lineno"> 1310</span>&#160;</div><div class="line"><a name="l01311"></a><span class="lineno"> 1311</span>&#160; <span class="comment">// z == 3</span></div><div class="line"><a name="l01312"></a><span class="lineno"> 1312</span>&#160; <span class="comment">// After z = 1 we can simply add src_stride_z to offset without updating z_coord</span></div><div class="line"><a name="l01313"></a><span class="lineno"> 1313</span>&#160; <span class="comment">// However offset can be out-of-bound so we need to check if it is greater than max_offset</span></div><div class="line"><a name="l01314"></a><span class="lineno"> 1314</span>&#160; offset += (int4)src_stride_z;</div><div class="line"><a name="l01315"></a><span class="lineno"> 1315</span>&#160; offset = min(offset, (int4)max_offset);</div><div class="line"><a name="l01316"></a><span class="lineno"> 1316</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values12 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s0));</div><div class="line"><a name="l01317"></a><span class="lineno"> 1317</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values13 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s1));</div><div class="line"><a name="l01318"></a><span class="lineno"> 1318</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values14 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s2));</div><div class="line"><a name="l01319"></a><span class="lineno"> 1319</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> values15 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + offset.s3));</div><div class="line"><a name="l01320"></a><span class="lineno"> 1320</span>&#160;</div><div class="line"><a name="l01321"></a><span class="lineno"> 1321</span>&#160; acc0 = fma(values0, w0, acc0);</div><div class="line"><a name="l01322"></a><span class="lineno"> 1322</span>&#160; acc0 = fma(values1, w1, acc0);</div><div class="line"><a name="l01323"></a><span class="lineno"> 1323</span>&#160; acc0 = fma(values2, w2, acc0);</div><div class="line"><a name="l01324"></a><span class="lineno"> 1324</span>&#160; acc1 = fma(values1, w0, acc1);</div><div class="line"><a name="l01325"></a><span class="lineno"> 1325</span>&#160; acc1 = fma(values2, w1, acc1);</div><div class="line"><a name="l01326"></a><span class="lineno"> 1326</span>&#160; acc1 = fma(values3, w2, acc1);</div><div class="line"><a name="l01327"></a><span class="lineno"> 1327</span>&#160;</div><div class="line"><a name="l01328"></a><span class="lineno"> 1328</span>&#160; acc0 = fma(values4, w3, acc0);</div><div class="line"><a name="l01329"></a><span class="lineno"> 1329</span>&#160; acc0 = fma(values5, w4, acc0);</div><div class="line"><a name="l01330"></a><span class="lineno"> 1330</span>&#160; acc0 = fma(values6, w5, acc0);</div><div class="line"><a name="l01331"></a><span class="lineno"> 1331</span>&#160; acc1 = fma(values5, w3, acc1);</div><div class="line"><a name="l01332"></a><span class="lineno"> 1332</span>&#160; acc1 = fma(values6, w4, acc1);</div><div class="line"><a name="l01333"></a><span class="lineno"> 1333</span>&#160; acc1 = fma(values7, w5, acc1);</div><div class="line"><a name="l01334"></a><span class="lineno"> 1334</span>&#160;</div><div class="line"><a name="l01335"></a><span class="lineno"> 1335</span>&#160; acc0 = fma(values8, w6, acc0);</div><div class="line"><a name="l01336"></a><span class="lineno"> 1336</span>&#160; acc0 = fma(values9, w7, acc0);</div><div class="line"><a name="l01337"></a><span class="lineno"> 1337</span>&#160; acc0 = fma(values10, w8, acc0);</div><div class="line"><a name="l01338"></a><span class="lineno"> 1338</span>&#160; acc1 = fma(values9, w6, acc1);</div><div class="line"><a name="l01339"></a><span class="lineno"> 1339</span>&#160; acc1 = fma(values10, w7, acc1);</div><div class="line"><a name="l01340"></a><span class="lineno"> 1340</span>&#160; acc1 = fma(values11, w8, acc1);</div><div class="line"><a name="l01341"></a><span class="lineno"> 1341</span>&#160;</div><div class="line"><a name="l01342"></a><span class="lineno"> 1342</span>&#160; acc2 = fma(values4, w0, acc2);</div><div class="line"><a name="l01343"></a><span class="lineno"> 1343</span>&#160; acc2 = fma(values5, w1, acc2);</div><div class="line"><a name="l01344"></a><span class="lineno"> 1344</span>&#160; acc2 = fma(values6, w2, acc2);</div><div class="line"><a name="l01345"></a><span class="lineno"> 1345</span>&#160; acc3 = fma(values5, w0, acc3);</div><div class="line"><a name="l01346"></a><span class="lineno"> 1346</span>&#160; acc3 = fma(values6, w1, acc3);</div><div class="line"><a name="l01347"></a><span class="lineno"> 1347</span>&#160; acc3 = fma(values7, w2, acc3);</div><div class="line"><a name="l01348"></a><span class="lineno"> 1348</span>&#160;</div><div class="line"><a name="l01349"></a><span class="lineno"> 1349</span>&#160; acc2 = fma(values8, w3, acc2);</div><div class="line"><a name="l01350"></a><span class="lineno"> 1350</span>&#160; acc2 = fma(values9, w4, acc2);</div><div class="line"><a name="l01351"></a><span class="lineno"> 1351</span>&#160; acc2 = fma(values10, w5, acc2);</div><div class="line"><a name="l01352"></a><span class="lineno"> 1352</span>&#160; acc3 = fma(values9, w3, acc3);</div><div class="line"><a name="l01353"></a><span class="lineno"> 1353</span>&#160; acc3 = fma(values10, w4, acc3);</div><div class="line"><a name="l01354"></a><span class="lineno"> 1354</span>&#160; acc3 = fma(values11, w5, acc3);</div><div class="line"><a name="l01355"></a><span class="lineno"> 1355</span>&#160;</div><div class="line"><a name="l01356"></a><span class="lineno"> 1356</span>&#160; acc2 = fma(values12, w6, acc2);</div><div class="line"><a name="l01357"></a><span class="lineno"> 1357</span>&#160; acc2 = fma(values13, w7, acc2);</div><div class="line"><a name="l01358"></a><span class="lineno"> 1358</span>&#160; acc2 = fma(values14, w8, acc2);</div><div class="line"><a name="l01359"></a><span class="lineno"> 1359</span>&#160; acc3 = fma(values13, w6, acc3);</div><div class="line"><a name="l01360"></a><span class="lineno"> 1360</span>&#160; acc3 = fma(values14, w7, acc3);</div><div class="line"><a name="l01361"></a><span class="lineno"> 1361</span>&#160; acc3 = fma(values15, w8, acc3);</div><div class="line"><a name="l01362"></a><span class="lineno"> 1362</span>&#160;</div><div class="line"><a name="l01363"></a><span class="lineno"> 1363</span>&#160;<span class="preprocessor">#if defined(HAS_BIAS)</span></div><div class="line"><a name="l01364"></a><span class="lineno"> 1364</span>&#160; <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l01365"></a><span class="lineno"> 1365</span>&#160;</div><div class="line"><a name="l01366"></a><span class="lineno"> 1366</span>&#160; <a class="code" href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a> bias_values = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VEC_SIZE)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l01367"></a><span class="lineno"> 1367</span>&#160;</div><div class="line"><a name="l01368"></a><span class="lineno"> 1368</span>&#160; acc0 += bias_values;</div><div class="line"><a name="l01369"></a><span class="lineno"> 1369</span>&#160; acc1 += bias_values;</div><div class="line"><a name="l01370"></a><span class="lineno"> 1370</span>&#160; acc2 += bias_values;</div><div class="line"><a name="l01371"></a><span class="lineno"> 1371</span>&#160; acc3 += bias_values;</div><div class="line"><a name="l01372"></a><span class="lineno"> 1372</span>&#160;<span class="preprocessor">#endif // defined(HAS_BIAS)</span></div><div class="line"><a name="l01373"></a><span class="lineno"> 1373</span>&#160;</div><div class="line"><a name="l01374"></a><span class="lineno"> 1374</span>&#160;<span class="preprocessor">#if defined(DST_DEPTH)</span></div><div class="line"><a name="l01375"></a><span class="lineno"> 1375</span>&#160; __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z + b * dst_stride_w;</div><div class="line"><a name="l01376"></a><span class="lineno"> 1376</span>&#160;<span class="preprocessor">#else </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01377"></a><span class="lineno"> 1377</span>&#160; __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x * dst_step_x + y * dst_step_y + (z * NUM_PLANES_PROCESSED) * dst_step_z;</div><div class="line"><a name="l01378"></a><span class="lineno"> 1378</span>&#160;<span class="preprocessor">#endif </span><span class="comment">/* defined(DST_DEPTH) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01379"></a><span class="lineno"> 1379</span>&#160;</div><div class="line"><a name="l01380"></a><span class="lineno"> 1380</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VEC_SIZE)</div><div class="line"><a name="l01381"></a><span class="lineno"> 1381</span>&#160; (acc0, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01382"></a><span class="lineno"> 1382</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VEC_SIZE)</div><div class="line"><a name="l01383"></a><span class="lineno"> 1383</span>&#160; (acc1, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01384"></a><span class="lineno"> 1384</span>&#160;</div><div class="line"><a name="l01385"></a><span class="lineno"> 1385</span>&#160;<span class="preprocessor">#if((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)</span></div><div class="line"><a name="l01386"></a><span class="lineno"> 1386</span>&#160; <span class="keywordflow">if</span>((z * NUM_PLANES_PROCESSED + 1) &lt; DST_DIM_2)</div><div class="line"><a name="l01387"></a><span class="lineno"> 1387</span>&#160;#endif <span class="comment">// ((DST_DIM_2 % NUM_PLANES_PROCESSED) != 0)</span></div><div class="line"><a name="l01388"></a><span class="lineno"> 1388</span>&#160; {</div><div class="line"><a name="l01389"></a><span class="lineno"> 1389</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VEC_SIZE)</div><div class="line"><a name="l01390"></a><span class="lineno"> 1390</span>&#160; (acc2, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_addr + 0 * dst_stride_y + 1 * dst_stride_z));</div><div class="line"><a name="l01391"></a><span class="lineno"> 1391</span>&#160; <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VEC_SIZE)</div><div class="line"><a name="l01392"></a><span class="lineno"> 1392</span>&#160; (acc3, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_addr + 1 * dst_stride_y + 1 * dst_stride_z));</div><div class="line"><a name="l01393"></a><span class="lineno"> 1393</span>&#160; }</div><div class="line"><a name="l01394"></a><span class="lineno"> 1394</span>&#160;}</div><div class="line"><a name="l01395"></a><span class="lineno"> 1395</span>&#160;</div><div class="line"><a name="l01396"></a><span class="lineno"> 1396</span>&#160;<span class="preprocessor">#endif // defined(NUM_ROWS_PROCESSED) &amp;&amp; defined(NUM_PLANES_PROCESSED)</span></div><div class="line"><a name="l01397"></a><span class="lineno"> 1397</span>&#160;<span class="preprocessor">#endif // defined(VEC_SIZE) &amp;&amp; defined(SRC_DIM_2) &amp;&amp; defined(CONV_PAD_TOP) &amp;&amp; defined(CONV_PAD_LEFT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="ttc" id="struct_vector_xhtml"><div class="ttname"><a href="struct_vector.xhtml">Vector</a></div><div class="ttdoc">Structure to hold Vector information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00140">helpers.h:140</a></div></div>
95<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a009469e4d9b8fce3b6d5e97d2077827d"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a></div><div class="ttdeci">__global uchar * offset(const Image *img, int x, int y)</div><div class="ttdoc">Get the pointer position of a Image. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00309">helpers.h:309</a></div></div>
96<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_aa8d95ba04fc73845abc6045952cae5be"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a></div><div class="ttdeci">#define CONVERT(x, type)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00060">helpers.h:60</a></div></div>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]97<div class="ttc" id="convolution3x3_8cl_xhtml_afb8c72ce35c4a1f4a2588d6573e54aa1"><div class="ttname"><a href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a></div><div class="ttdeci">#define DATA_TYPE</div><div class="ttdef"><b>Definition:</b> <a href="convolution3x3_8cl_source.xhtml#l00027">convolution3x3.cl:27</a></div></div>
Anthony Barbier06ea0482018-02-22 15:45:35 +0000[diff] [blame]98<div class="ttc" id="namespacearm__compute_xhtml_a73e2825fd61d349c5ca2f5313e3c8ea1"><div class="ttname"><a href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">arm_compute::half</a></div><div class="ttdeci">half_float::half half</div><div class="ttdoc">16-bit floating point type </div><div class="ttdef"><b>Definition:</b> <a href="arm__compute_2core_2_types_8h_source.xhtml#l00044">Types.h:44</a></div></div>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]99<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a22f42fcf2077d951271df83b55c1a71a"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a></div><div class="ttdeci">#define IMAGE_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00074">helpers.h:74</a></div></div>
100<div class="ttc" id="struct_tensor3_d_xhtml"><div class="ttname"><a href="struct_tensor3_d.xhtml">Tensor3D</a></div><div class="ttdoc">Structure to hold 3D tensor information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00157">helpers.h:157</a></div></div>
101<div class="ttc" id="depthwise__convolution_8cl_xhtml_ac353c0687abd8c71a4fa096d322b39ec"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#ac353c0687abd8c71a4fa096d322b39ec">in_stride_y</a></div><div class="ttdeci">#define in_stride_y</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00476">depthwise_convolution.cl:476</a></div></div>
102<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a79e8e562daa6599317d2d1cd86ef1bf2"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a></div><div class="ttdeci">#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00129">helpers.h:129</a></div></div>
103<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a527bfdf5eeb306f1cf01c4a8e29f38e0"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_VECTOR_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00104">helpers.h:104</a></div></div>
104<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a541f8db866a0fa93ee67d58ea31a7d0c"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a541f8db866a0fa93ee67d58ea31a7d0c">CONVERT_TENSOR3D_TO_IMAGE_STRUCT</a></div><div class="ttdeci">#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00122">helpers.h:122</a></div></div>
105<div class="ttc" id="depthwise__convolution_8cl_xhtml_a128f47cb6aacde29e07fde2c4b9f5dd2"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#a128f47cb6aacde29e07fde2c4b9f5dd2">out_stride_z</a></div><div class="ttdeci">#define out_stride_z</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00480">depthwise_convolution.cl:480</a></div></div>
106<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a40a6eb9f2a7712f08d6bb8ff6c9e6ca7"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a></div><div class="ttdeci">#define VECTOR_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00068">helpers.h:68</a></div></div>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]107<div class="ttc" id="convolution3x3_8cl_xhtml_afc5fefe72e66f0ae5191fd5b708fade9"><div class="ttname"><a href="convolution3x3_8cl.xhtml#afc5fefe72e66f0ae5191fd5b708fade9">convolution3x3</a></div><div class="ttdeci">short8 convolution3x3(Image *src, const short mat0, const short mat1, const short mat2, const short mat3, const short mat4, const short mat5, const short mat6, const short mat7, const short mat8, uint scale)</div><div class="ttdoc">Apply a 3x3 convolution matrix to a single channel U8 input image and return the result. </div><div class="ttdef"><b>Definition:</b> <a href="convolution3x3_8cl_source.xhtml#l00082">convolution3x3.cl:82</a></div></div>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]108<div class="ttc" id="struct_vector_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">Vector::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00142">helpers.h:142</a></div></div>
Jenkins52ba29e2018-08-29 15:32:11 +0000[diff] [blame]109<div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_adbf67dcee294e673cf796f1ed8aeb6a4"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">arm_compute::test::validation::dst</a></div><div class="ttdeci">CLTensor dst</div><div class="ttdef"><b>Definition:</b> <a href="_c_l_2_arithmetic_division_8cpp_source.xhtml#l00100">ArithmeticDivision.cpp:100</a></div></div>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]110<div class="ttc" id="struct_image_xhtml"><div class="ttname"><a href="struct_image.xhtml">Image</a></div><div class="ttdoc">Structure to hold Image information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00148">helpers.h:148</a></div></div>
111<div class="ttc" id="depthwise__convolution_8cl_xhtml_a5447e2c4e1aae8fff9290b775fe813ff"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#a5447e2c4e1aae8fff9290b775fe813ff">in_stride_x</a></div><div class="ttdeci">#define in_stride_x</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00475">depthwise_convolution.cl:475</a></div></div>
112<div class="ttc" id="struct_image_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">Image::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00150">helpers.h:150</a></div></div>
113<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a31c8c760f08fb1a331b16b7c204321dc"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_TENSOR3D_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00125">helpers.h:125</a></div></div>
114<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a7e4940407322d6f0ccb8b6b86b856019"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a></div><div class="ttdeci">__global const uchar * vector_offset(const Vector *vec, int x)</div><div class="ttdoc">Get the pointer position of a Vector. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00298">helpers.h:298</a></div></div>
115<div class="ttc" id="depthwise__convolution_8cl_xhtml_a32137dd220c1110408a7b498108de2c5"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#a32137dd220c1110408a7b498108de2c5">in_stride_z</a></div><div class="ttdeci">#define in_stride_z</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00477">depthwise_convolution.cl:477</a></div></div>
116<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml">helpers.h</a></div></div>
117<div class="ttc" id="depthwise__convolution_8cl_xhtml_ac8b6b732372937bba276e6391944d1e1"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#ac8b6b732372937bba276e6391944d1e1">out_stride_y</a></div><div class="ttdeci">#define out_stride_y</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00479">depthwise_convolution.cl:479</a></div></div>
118<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_acb282042d1edeeaa3cc979a206f78b54"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a></div><div class="ttdeci">#define VSTORE(size)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00051">helpers.h:51</a></div></div>
119<div class="ttc" id="depthwise__convolution_8cl_xhtml_a68fce1f0ee4080ed89c9427496d625f2"><div class="ttname"><a href="depthwise__convolution_8cl.xhtml#a68fce1f0ee4080ed89c9427496d625f2">out_stride_x</a></div><div class="ttdeci">#define out_stride_x</div><div class="ttdef"><b>Definition:</b> <a href="depthwise__convolution_8cl_source.xhtml#l00478">depthwise_convolution.cl:478</a></div></div>
120<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a481bdc6d61b3df9dcdbdb244f0f97790"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a></div><div class="ttdeci">#define TENSOR4D_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00092">helpers.h:92</a></div></div>
121<div class="ttc" id="struct_tensor3_d_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">Tensor3D::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00159">helpers.h:159</a></div></div>
122<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a287e2fc366c312b468382c95bb90f91f"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a></div><div class="ttdeci">#define VLOAD(size)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00048">helpers.h:48</a></div></div>
123<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a6b83038822d1ae7ab619b684ed3b7fc0"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a></div><div class="ttdeci">#define TENSOR3D_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00082">helpers.h:82</a></div></div>
124<div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_a7b8004eef325a40dd43eb80755610fff"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#a7b8004eef325a40dd43eb80755610fff">arm_compute::test::validation::b</a></div><div class="ttdeci">CLTensor b</div><div class="ttdef"><b>Definition:</b> <a href="validation_2_c_l_2_g_e_m_m_8cpp_source.xhtml#l00091">GEMM.cpp:91</a></div></div>
125<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a64d779f80eeb923e0ab2313433f7b40b"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a></div><div class="ttdeci">#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00107">helpers.h:107</a></div></div>
Jenkins52ba29e2018-08-29 15:32:11 +0000[diff] [blame]126<div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_a6743f0a130e8311e6f5b1a23df102472"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">arm_compute::test::validation::src</a></div><div class="ttdeci">convolution configure &amp; src</div><div class="ttdef"><b>Definition:</b> <a href="validation_2_c_l_2_convolution_8cpp_source.xhtml#l00147">Convolution.cpp:147</a></div></div>
Jenkins52ba29e2018-08-29 15:32:11 +0000[diff] [blame]127<div class="ttc" id="activation__layer__qa8_8cl_xhtml_ade2e33e6f303ce93468eef7e56d95c0c"><div class="ttname"><a href="activation__layer__qa8_8cl.xhtml#ade2e33e6f303ce93468eef7e56d95c0c">VEC_FLOAT</a></div><div class="ttdeci">#define VEC_FLOAT</div><div class="ttdef"><b>Definition:</b> <a href="activation__layer__qa8_8cl_source.xhtml#l00027">activation_layer_qa8.cl:27</a></div></div>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]128<div class="ttc" id="convolution3x3_8cl_xhtml_a92702074338198e81a46c3e309d9b04f"><div class="ttname"><a href="convolution3x3_8cl.xhtml#a92702074338198e81a46c3e309d9b04f">convolution1x3</a></div><div class="ttdeci">short8 convolution1x3(__global const uchar *left_pixel, const short left_coeff, const short middle_coeff, const short right_coeff)</div><div class="ttdoc">Compute a 1D horizontal convolution of size 3 for 8 bytes assuming the input is made of 1 channel of ...</div><div class="ttdef"><b>Definition:</b> <a href="convolution3x3_8cl_source.xhtml#l00043">convolution3x3.cl:43</a></div></div>
129</div><!-- fragment --></div><!-- contents -->
130</div><!-- doc-content -->
131<!-- start footer part -->
132<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
133 <ul>
Anthony Barbier8140e1e2017-12-14 23:48:46 +0000[diff] [blame]134 <li class="navelem"><a class="el" href="dir_68267d1309a1af8e8297ef4c3efbcdba.xhtml">src</a></li><li class="navelem"><a class="el" href="dir_aebb8dcc11953d78e620bbef0b9e2183.xhtml">core</a></li><li class="navelem"><a class="el" href="dir_8c278f79c760e5c5fbd911f9870614c1.xhtml">CL</a></li><li class="navelem"><a class="el" href="dir_25885286e9dad4fa105b7b25a8031bbf.xhtml">cl_kernels</a></li><li class="navelem"><a class="el" href="depthwise__convolution_8cl.xhtml">depthwise_convolution.cl</a></li>
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]135 <li class="footer">Generated on Thu Nov 22 2018 11:57:38 for Compute Library by
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]136 <a href="http://www.doxygen.org/index.html">
Jenkinsb9abeae2018-11-22 11:58:08 +0000[diff] [blame]137 <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li>
Kaizen8938bd32017-09-28 14:38:23 +0100[diff] [blame]138 </ul>
139</div>
140</body>
141</html>