documentation/gemm_8cl_source.xhtml

   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"/>
   7 <meta name="generator" content="Doxygen 1.8.11"/>
   8 <meta name="robots" content="NOINDEX, NOFOLLOW" /> <!-- Prevent indexing by search engines -->
   9 <title>Compute Library: src/core/CL/cl_kernels/gemm.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>
  15 <script type="text/javascript" src="navtreedata.js"></script>
  16 <script type="text/javascript" src="navtree.js"></script>
  17 <script type="text/javascript">
  18   $(document).ready(initResizable);
  19   $(window).load(resizeHeight);
  20 </script>
  21 <link href="search/search.css" rel="stylesheet" type="text/css"/>
  22 <script type="text/javascript" src="search/searchdata.js"></script>
  23 <script type="text/javascript" src="search/search.js"></script>
  24 <script type="text/javascript">
  25   $(document).ready(function() { init_search(); });
  26 </script>
  27 <script type="text/x-mathjax-config">
  28   MathJax.Hub.Config({
  29     extensions: ["tex2jax.js"],
  30     jax: ["input/TeX","output/HTML-CSS"],
  31 });
  32 </script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script>
  33 <link href="doxygen.css" rel="stylesheet" type="text/css" />
  34 </head>
  35 <body>
  36 <div id="top"><!-- do not remove this div, it is closed by doxygen! -->
  37 <div id="titlearea">
  38 <table cellspacing="0" cellpadding="0">
  39  <tbody>
  40  <tr style="height: 56px;">
  41   <td style="padding-left: 0.5em;">
  42    <div id="projectname">Compute Library
  43    &#160;<span id="projectnumber">18.05</span>
  44    </div>
  45   </td>
  46  </tr>
  47  </tbody>
  48 </table>
  49 </div>
  50 <!-- end header part -->
  51 <!-- Generated by Doxygen 1.8.11 -->
  52 <script type="text/javascript">
  53 var searchBox = new SearchBox("searchBox", "search",false,'Search');
  54 </script>
  55   <div id="navrow1" class="tabs">
  56     <ul class="tablist">
  57       <li><a href="index.xhtml"><span>Main&#160;Page</span></a></li>
  58       <li><a href="pages.xhtml"><span>Related&#160;Pages</span></a></li>
  59       <li><a href="namespaces.xhtml"><span>Namespaces</span></a></li>
  60       <li><a href="annotated.xhtml"><span>Data&#160;Structures</span></a></li>
  61       <li class="current"><a href="files.xhtml"><span>Files</span></a></li>
  62       <li>
  63         <div id="MSearchBox" class="MSearchBoxInactive">
  64         <span class="left">
  65           <img id="MSearchSelect" src="search/mag_sel.png"
  66                onmouseover="return searchBox.OnSearchSelectShow()"
  67                onmouseout="return searchBox.OnSearchSelectHide()"
  68                alt=""/>
  69           <input type="text" id="MSearchField" value="Search" accesskey="S"
  70                onfocus="searchBox.OnSearchFieldFocus(true)"
  71                onblur="searchBox.OnSearchFieldFocus(false)"
  72                onkeyup="searchBox.OnSearchFieldChange(event)"/>
  73           </span><span class="right">
  74             <a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.png" alt=""/></a>
  75           </span>
  76         </div>
  77       </li>
  78     </ul>
  79   </div>
  80   <div id="navrow2" class="tabs2">
  81     <ul class="tablist">
  82       <li><a href="files.xhtml"><span>File&#160;List</span></a></li>
  83       <li><a href="globals.xhtml"><span>Globals</span></a></li>
  84     </ul>
  85   </div>
  86 </div><!-- top -->
  87 <div id="side-nav" class="ui-resizable side-nav-resizable">
  88   <div id="nav-tree">
  89     <div id="nav-tree-contents">
  90       <div id="nav-sync" class="sync"></div>
  91     </div>
  92   </div>
  93   <div id="splitbar" style="-moz-user-select:none;"
  94        class="ui-resizable-handle">
  95   </div>
  96 </div>
  97 <script type="text/javascript">
  98 $(document).ready(function(){initNavTree('gemm_8cl_source.xhtml','');});
  99 </script>
 100 <div id="doc-content">
 101 <!-- window showing the filter options -->
 102 <div id="MSearchSelectWindow"
 103      onmouseover="return searchBox.OnSearchSelectShow()"
 104      onmouseout="return searchBox.OnSearchSelectHide()"
 105      onkeydown="return searchBox.OnSearchSelectKey(event)">
 106 </div>
 107
 108 <!-- iframe showing the search results (closed by default) -->
 109 <div id="MSearchResultsWindow">
 110 <iframe src="javascript:void(0)" frameborder="0"
 111         name="MSearchResults" id="MSearchResults">
 112 </iframe>
 113 </div>
 114
 115 <div class="header">
 116   <div class="headertitle">
 117 <div class="title">gemm.cl</div>  </div>
 118 </div><!--header-->
 119 <div class="contents">
 120 <a href="gemm_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;<span class="preprocessor">#include &quot;<a class="code" href="helpers_8h.xhtml">helpers.h</a>&quot;</span></div><div class="line"><a name="l00025"></a><span class="lineno">   25</span>&#160;</div><div class="line"><a name="l00026"></a><span class="lineno">   26</span>&#160;<span class="preprocessor">#ifdef FIXED_POINT_POSITION</span></div><div class="line"><a name="l00027"></a><span class="lineno">   27</span>&#160;<span class="preprocessor">#include &quot;<a class="code" href="fixed__point_8h.xhtml">fixed_point.h</a>&quot;</span></div><div class="line"><a name="l00028"></a><span class="lineno">   28</span>&#160;<span class="preprocessor">#endif // FIXED_POINT_POSITION</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 defined(TRANSPOSE_W) &amp;&amp; defined(MULT_TRANSPOSE1XW_WIDTH)</span></div><div class="line"><a name="l00031"></a><span class="lineno">   31</span>&#160;</div><div class="line"><a name="l00032"></a><span class="lineno">   32</span>&#160;<span class="preprocessor">#if ELEMENT_SIZE == 1</span></div><div class="line"><a name="l00033"></a><span class="lineno">   33</span>&#160;<span class="preprocessor">#define DATA_TYPE uchar</span></div><div class="line"><a name="l00034"></a><span class="lineno">   34</span>&#160;<span class="preprocessor">#elif ELEMENT_SIZE == 2</span></div><div class="line"><a name="l00035"></a><span class="lineno">   35</span>&#160;<span class="preprocessor">#define DATA_TYPE ushort</span></div><div class="line"><a name="l00036"></a><span class="lineno">   36</span>&#160;<span class="preprocessor">#elif ELEMENT_SIZE == 4</span></div><div class="line"><a name="l00037"></a><span class="lineno">   37</span>&#160;<span class="preprocessor">#define DATA_TYPE uint</span></div><div class="line"><a name="l00038"></a><span class="lineno">   38</span>&#160;<span class="preprocessor">#else // ELEMENT_SIZE == 1</span></div><div class="line"><a name="l00039"></a><span class="lineno">   39</span>&#160;<span class="preprocessor">#error &quot;Element size not supported&quot;</span></div><div class="line"><a name="l00040"></a><span class="lineno">   40</span>&#160;<span class="preprocessor">#endif // ELEMENT_SIZE</span></div><div class="line"><a name="l00041"></a><span class="lineno">   41</span>&#160;</div><div class="line"><a name="l00064"></a><span class="lineno">   64</span>&#160;__kernel <span class="keywordtype">void</span> gemm_transpose1xW(<a class="code" href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>),</div><div class="line"><a name="l00065"></a><span class="lineno">   65</span>&#160;                                <a class="code" href="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="l00066"></a><span class="lineno">   66</span>&#160;{</div><div class="line"><a name="l00067"></a><span class="lineno">   67</span>&#160;    uint x = get_global_id(0);</div><div class="line"><a name="l00068"></a><span class="lineno">   68</span>&#160;    uint y = get_global_id(1);</div><div class="line"><a name="l00069"></a><span class="lineno">   69</span>&#160;    uint z = get_global_id(2);</div><div class="line"><a name="l00070"></a><span class="lineno">   70</span>&#160;</div><div class="line"><a name="l00071"></a><span class="lineno">   71</span>&#160;    <span class="comment">// Compute address for Matrix B - source</span></div><div class="line"><a name="l00072"></a><span class="lineno">   72</span>&#160;    <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a> = <a class="code" href="helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l00073"></a><span class="lineno">   73</span>&#160;</div><div class="line"><a name="l00074"></a><span class="lineno">   74</span>&#160;    <span class="comment">// Compute address for Matrix B transposed - destination. X and Y are swapped</span></div><div class="line"><a name="l00075"></a><span class="lineno">   75</span>&#160;    uint dst_addr_in_bytes = dst_offset_first_element_in_bytes + y * TRANSPOSE_W * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * MULT_TRANSPOSE1XW_WIDTH + (x / MULT_TRANSPOSE1XW_WIDTH) * dst_stride_y +</div><div class="line"><a name="l00076"></a><span class="lineno">   76</span>&#160;                             (x % MULT_TRANSPOSE1XW_WIDTH) * TRANSPOSE_W * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>);</div><div class="line"><a name="l00077"></a><span class="lineno">   77</span>&#160;</div><div class="line"><a name="l00078"></a><span class="lineno">   78</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00079"></a><span class="lineno">   79</span>&#160;    dst_addr_in_bytes += z * dst_stride_z;</div><div class="line"><a name="l00080"></a><span class="lineno">   80</span>&#160;</div><div class="line"><a name="l00081"></a><span class="lineno">   81</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, TRANSPOSE_W)</div><div class="line"><a name="l00082"></a><span class="lineno">   82</span>&#160;    b0 = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(TRANSPOSE_W)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)src.ptr);</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;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(TRANSPOSE_W)</div><div class="line"><a name="l00085"></a><span class="lineno">   85</span>&#160;    (b0, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00086"></a><span class="lineno">   86</span>&#160;}</div><div class="line"><a name="l00087"></a><span class="lineno">   87</span>&#160;<span class="preprocessor">#endif // defined(TRANSPOSE_W) &amp;&amp; defined(MULT_TRANSPOSE1XW_WIDTH)</span></div><div class="line"><a name="l00088"></a><span class="lineno">   88</span>&#160;</div><div class="line"><a name="l00089"></a><span class="lineno">   89</span>&#160;<span class="preprocessor">#if defined(MULT_INTERLEAVE4X4_HEIGHT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00090"></a><span class="lineno">   90</span>&#160;</div><div class="line"><a name="l00113"></a><span class="lineno">  113</span>&#160;__kernel <span class="keywordtype">void</span> gemm_interleave4x4(<a class="code" href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l00114"></a><span class="lineno">  114</span>&#160;                                 <a class="code" href="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="l00115"></a><span class="lineno">  115</span>&#160;{</div><div class="line"><a name="l00116"></a><span class="lineno">  116</span>&#160;    <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l00117"></a><span class="lineno">  117</span>&#160;    uint x = get_global_id(0);</div><div class="line"><a name="l00118"></a><span class="lineno">  118</span>&#160;    uint y = get_global_id(1);</div><div class="line"><a name="l00119"></a><span class="lineno">  119</span>&#160;    uint z = get_global_id(2);</div><div class="line"><a name="l00120"></a><span class="lineno">  120</span>&#160;</div><div class="line"><a name="l00121"></a><span class="lineno">  121</span>&#160;    <span class="comment">// Compute address for source tensor</span></div><div class="line"><a name="l00122"></a><span class="lineno">  122</span>&#160;    <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> src = <a class="code" href="helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l00123"></a><span class="lineno">  123</span>&#160;</div><div class="line"><a name="l00124"></a><span class="lineno">  124</span>&#160;    <span class="comment">// Compute address for Matrix B transposed - destination. X and Y are swapped</span></div><div class="line"><a name="l00125"></a><span class="lineno">  125</span>&#160;    uint dst_addr_in_bytes = dst_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) * 16 * MULT_INTERLEAVE4X4_HEIGHT + (y / MULT_INTERLEAVE4X4_HEIGHT) * dst_stride_y +</div><div class="line"><a name="l00126"></a><span class="lineno">  126</span>&#160;                             (y % MULT_INTERLEAVE4X4_HEIGHT) * 4 * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>);</div><div class="line"><a name="l00127"></a><span class="lineno">  127</span>&#160;</div><div class="line"><a name="l00128"></a><span class="lineno">  128</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00129"></a><span class="lineno">  129</span>&#160;    dst_addr_in_bytes += z * dst_stride_z;</div><div class="line"><a name="l00130"></a><span class="lineno">  130</span>&#160;</div><div class="line"><a name="l00131"></a><span class="lineno">  131</span>&#160;    __global uchar *input_ptr = src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>;</div><div class="line"><a name="l00132"></a><span class="lineno">  132</span>&#160;</div><div class="line"><a name="l00133"></a><span class="lineno">  133</span>&#160;    <span class="comment">// Load values from Matrix A</span></div><div class="line"><a name="l00134"></a><span class="lineno">  134</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 4)</div><div class="line"><a name="l00135"></a><span class="lineno">  135</span>&#160;    a0 = vload4(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(input_ptr + 0 * src_stride_y));</div><div class="line"><a name="l00136"></a><span class="lineno">  136</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 4)</div><div class="line"><a name="l00137"></a><span class="lineno">  137</span>&#160;    a1 = vload4(0, (__global DATA_TYPE *)(input_ptr + 1 * src_stride_y));</div><div class="line"><a name="l00138"></a><span class="lineno">  138</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00139"></a><span class="lineno">  139</span>&#160;    a2 = vload4(0, (__global DATA_TYPE *)(input_ptr + 2 * src_stride_y));</div><div class="line"><a name="l00140"></a><span class="lineno">  140</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00141"></a><span class="lineno">  141</span>&#160;    a3 = vload4(0, (__global DATA_TYPE *)(input_ptr + 3 * src_stride_y));</div><div class="line"><a name="l00142"></a><span class="lineno">  142</span>&#160;</div><div class="line"><a name="l00143"></a><span class="lineno">  143</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00144"></a><span class="lineno">  144</span>&#160;    val0 = (<a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s0, a1.s0, a2.s0, a3.s0);</div><div class="line"><a name="l00145"></a><span class="lineno">  145</span>&#160;    vstore4(val0, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 0 * MULT_INTERLEAVE4X4_HEIGHT));</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;    val0 = (<a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s1, a1.s1, a2.s1, a3.s1);</div><div class="line"><a name="l00148"></a><span class="lineno">  148</span>&#160;    vstore4(val0, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 4 * MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00149"></a><span class="lineno">  149</span>&#160;</div><div class="line"><a name="l00150"></a><span class="lineno">  150</span>&#160;    val0 = (<a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s2, a1.s2, a2.s2, a3.s2);</div><div class="line"><a name="l00151"></a><span class="lineno">  151</span>&#160;    vstore4(val0, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 8 * MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00152"></a><span class="lineno">  152</span>&#160;</div><div class="line"><a name="l00153"></a><span class="lineno">  153</span>&#160;    val0 = (<a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s3, a1.s3, a2.s3, a3.s3);</div><div class="line"><a name="l00154"></a><span class="lineno">  154</span>&#160;    vstore4(val0, 0, ((__global DATA_TYPE *)(dst_ptr + dst_addr_in_bytes) + 12 * MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00155"></a><span class="lineno">  155</span>&#160;}</div><div class="line"><a name="l00156"></a><span class="lineno">  156</span>&#160;<span class="preprocessor">#endif // defined(MULT_INTERLEAVE4X4_HEIGHT) &amp;&amp; defined(DATA_TYPE)</span></div><div class="line"><a name="l00157"></a><span class="lineno">  157</span>&#160;</div><div class="line"><a name="l00158"></a><span class="lineno">  158</span>&#160;<span class="preprocessor">#if defined(COLS_B) &amp;&amp; defined(MULT_TRANSPOSE1XW_WIDTH) &amp;&amp; defined(MULT_INTERLEAVE4X4_HEIGHT)</span></div><div class="line"><a name="l00159"></a><span class="lineno">  159</span>&#160;</div><div class="line"><a name="l00187"></a><span class="lineno">  187</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00188"></a><span class="lineno">  188</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00189"></a><span class="lineno">  189</span>&#160;                                                 <a class="code" href="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="l00190"></a><span class="lineno">  190</span>&#160;                                                 uint src0_stride_z,</div><div class="line"><a name="l00191"></a><span class="lineno">  191</span>&#160;                                                 uint src1_stride_z,</div><div class="line"><a name="l00192"></a><span class="lineno">  192</span>&#160;                                                 uint dst_stride_z)</div><div class="line"><a name="l00193"></a><span class="lineno">  193</span>&#160;{</div><div class="line"><a name="l00194"></a><span class="lineno">  194</span>&#160;    <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00195"></a><span class="lineno">  195</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00196"></a><span class="lineno">  196</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00197"></a><span class="lineno">  197</span>&#160;</div><div class="line"><a name="l00198"></a><span class="lineno">  198</span>&#160;    <span class="comment">// Offset</span></div><div class="line"><a name="l00199"></a><span class="lineno">  199</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00200"></a><span class="lineno">  200</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 4;</div><div class="line"><a name="l00201"></a><span class="lineno">  201</span>&#160;</div><div class="line"><a name="l00202"></a><span class="lineno">  202</span>&#160;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00203"></a><span class="lineno">  203</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00204"></a><span class="lineno">  204</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l00205"></a><span class="lineno">  205</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l00206"></a><span class="lineno">  206</span>&#160;</div><div class="line"><a name="l00207"></a><span class="lineno">  207</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00208"></a><span class="lineno">  208</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00209"></a><span class="lineno">  209</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00210"></a><span class="lineno">  210</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00211"></a><span class="lineno">  211</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00212"></a><span class="lineno">  212</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00213"></a><span class="lineno">  213</span>&#160;</div><div class="line"><a name="l00214"></a><span class="lineno">  214</span>&#160;    __global <span class="keywordtype">float</span> *src_addr_a = (__global <span class="keywordtype">float</span> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l00215"></a><span class="lineno">  215</span>&#160;    __global <span class="keywordtype">float</span> *src_addr_b = (__global <span class="keywordtype">float</span> *)(src1_ptr + src1_addr_in_bytes);</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="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00218"></a><span class="lineno">  218</span>&#160;    __global <span class="keywordtype">float</span> *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00219"></a><span class="lineno">  219</span>&#160;</div><div class="line"><a name="l00220"></a><span class="lineno">  220</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l00221"></a><span class="lineno">  221</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l00222"></a><span class="lineno">  222</span>&#160;</div><div class="line"><a name="l00223"></a><span class="lineno">  223</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00224"></a><span class="lineno">  224</span>&#160;    float4 c00 = 0.0f;</div><div class="line"><a name="l00225"></a><span class="lineno">  225</span>&#160;    float4 c10 = 0.0f;</div><div class="line"><a name="l00226"></a><span class="lineno">  226</span>&#160;    float4 c20 = 0.0f;</div><div class="line"><a name="l00227"></a><span class="lineno">  227</span>&#160;    float4 c30 = 0.0f;</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="keywordflow">for</span>(; src_addr_b &lt;= (src_end_addr_b - (int)(8 * MULT_TRANSPOSE1XW_WIDTH)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00230"></a><span class="lineno">  230</span>&#160;    {</div><div class="line"><a name="l00231"></a><span class="lineno">  231</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00232"></a><span class="lineno">  232</span>&#160;        float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00233"></a><span class="lineno">  233</span>&#160;        float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00234"></a><span class="lineno">  234</span>&#160;</div><div class="line"><a name="l00235"></a><span class="lineno">  235</span>&#160;        c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00236"></a><span class="lineno">  236</span>&#160;        c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00237"></a><span class="lineno">  237</span>&#160;        c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00238"></a><span class="lineno">  238</span>&#160;        c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00239"></a><span class="lineno">  239</span>&#160;</div><div class="line"><a name="l00240"></a><span class="lineno">  240</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00241"></a><span class="lineno">  241</span>&#160;        a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT);</div><div class="line"><a name="l00242"></a><span class="lineno">  242</span>&#160;        b0 = vload4(0, src_addr_b + 4 * MULT_TRANSPOSE1XW_WIDTH);</div><div class="line"><a name="l00243"></a><span class="lineno">  243</span>&#160;</div><div class="line"><a name="l00244"></a><span class="lineno">  244</span>&#160;        c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00245"></a><span class="lineno">  245</span>&#160;        c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00246"></a><span class="lineno">  246</span>&#160;        c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00247"></a><span class="lineno">  247</span>&#160;        c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00248"></a><span class="lineno">  248</span>&#160;    }</div><div class="line"><a name="l00249"></a><span class="lineno">  249</span>&#160;</div><div class="line"><a name="l00250"></a><span class="lineno">  250</span>&#160;    <span class="keywordflow">for</span>(; src_addr_b &lt; src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00251"></a><span class="lineno">  251</span>&#160;    {</div><div class="line"><a name="l00252"></a><span class="lineno">  252</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00253"></a><span class="lineno">  253</span>&#160;        float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00254"></a><span class="lineno">  254</span>&#160;        float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00255"></a><span class="lineno">  255</span>&#160;</div><div class="line"><a name="l00256"></a><span class="lineno">  256</span>&#160;        c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00257"></a><span class="lineno">  257</span>&#160;        c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00258"></a><span class="lineno">  258</span>&#160;        c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00259"></a><span class="lineno">  259</span>&#160;        c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00260"></a><span class="lineno">  260</span>&#160;    }</div><div class="line"><a name="l00261"></a><span class="lineno">  261</span>&#160;</div><div class="line"><a name="l00262"></a><span class="lineno">  262</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00263"></a><span class="lineno">  263</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="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00264"></a><span class="lineno">  264</span>&#160;</div><div class="line"><a name="l00265"></a><span class="lineno">  265</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00266"></a><span class="lineno">  266</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00267"></a><span class="lineno">  267</span>&#160;    c00 = c00 * (float4)ALPHA;</div><div class="line"><a name="l00268"></a><span class="lineno">  268</span>&#160;    c10 = c10 * (float4)ALPHA;</div><div class="line"><a name="l00269"></a><span class="lineno">  269</span>&#160;    c20 = c20 * (float4)ALPHA;</div><div class="line"><a name="l00270"></a><span class="lineno">  270</span>&#160;    c30 = c30 * (float4)ALPHA;</div><div class="line"><a name="l00271"></a><span class="lineno">  271</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00272"></a><span class="lineno">  272</span>&#160;</div><div class="line"><a name="l00273"></a><span class="lineno">  273</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00274"></a><span class="lineno">  274</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l00275"></a><span class="lineno">  275</span>&#160;</div><div class="line"><a name="l00276"></a><span class="lineno">  276</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00277"></a><span class="lineno">  277</span>&#160;    dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00278"></a><span class="lineno">  278</span>&#160;</div><div class="line"><a name="l00279"></a><span class="lineno">  279</span>&#160;    <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00280"></a><span class="lineno">  280</span>&#160;    vstore4(c00, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l00281"></a><span class="lineno">  281</span>&#160;    vstore4(c10, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l00282"></a><span class="lineno">  282</span>&#160;    vstore4(c20, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l00283"></a><span class="lineno">  283</span>&#160;    vstore4(c30, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l00284"></a><span class="lineno">  284</span>&#160;}</div><div class="line"><a name="l00285"></a><span class="lineno">  285</span>&#160;</div><div class="line"><a name="l00315"></a><span class="lineno">  315</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f32_bifrost(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00316"></a><span class="lineno">  316</span>&#160;                                                         <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00317"></a><span class="lineno">  317</span>&#160;                                                         <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00318"></a><span class="lineno">  318</span>&#160;                                                         uint src0_stride_z,</div><div class="line"><a name="l00319"></a><span class="lineno">  319</span>&#160;                                                         uint src1_stride_z,</div><div class="line"><a name="l00320"></a><span class="lineno">  320</span>&#160;                                                         uint dst_stride_z)</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="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00323"></a><span class="lineno">  323</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00324"></a><span class="lineno">  324</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00325"></a><span class="lineno">  325</span>&#160;</div><div class="line"><a name="l00326"></a><span class="lineno">  326</span>&#160;    <span class="comment">// Offset</span></div><div class="line"><a name="l00327"></a><span class="lineno">  327</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00328"></a><span class="lineno">  328</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 4;</div><div class="line"><a name="l00329"></a><span class="lineno">  329</span>&#160;</div><div class="line"><a name="l00330"></a><span class="lineno">  330</span>&#160;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00331"></a><span class="lineno">  331</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00332"></a><span class="lineno">  332</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l00333"></a><span class="lineno">  333</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l00334"></a><span class="lineno">  334</span>&#160;</div><div class="line"><a name="l00335"></a><span class="lineno">  335</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00336"></a><span class="lineno">  336</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00337"></a><span class="lineno">  337</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00338"></a><span class="lineno">  338</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00339"></a><span class="lineno">  339</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00340"></a><span class="lineno">  340</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</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;    __global <span class="keywordtype">float</span> *src_addr_a = (__global <span class="keywordtype">float</span> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l00343"></a><span class="lineno">  343</span>&#160;    __global <span class="keywordtype">float</span> *src_addr_b = (__global <span class="keywordtype">float</span> *)(src1_ptr + src1_addr_in_bytes);</div><div class="line"><a name="l00344"></a><span class="lineno">  344</span>&#160;</div><div class="line"><a name="l00345"></a><span class="lineno">  345</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l00346"></a><span class="lineno">  346</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l00347"></a><span class="lineno">  347</span>&#160;</div><div class="line"><a name="l00348"></a><span class="lineno">  348</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00349"></a><span class="lineno">  349</span>&#160;    <span class="keywordtype">float</span> c00 = 0.0f;</div><div class="line"><a name="l00350"></a><span class="lineno">  350</span>&#160;    <span class="keywordtype">float</span> c01 = 0.0f;</div><div class="line"><a name="l00351"></a><span class="lineno">  351</span>&#160;    <span class="keywordtype">float</span> c02 = 0.0f;</div><div class="line"><a name="l00352"></a><span class="lineno">  352</span>&#160;    <span class="keywordtype">float</span> c03 = 0.0f;</div><div class="line"><a name="l00353"></a><span class="lineno">  353</span>&#160;    <span class="keywordtype">float</span> c10 = 0.0f;</div><div class="line"><a name="l00354"></a><span class="lineno">  354</span>&#160;    <span class="keywordtype">float</span> c11 = 0.0f;</div><div class="line"><a name="l00355"></a><span class="lineno">  355</span>&#160;    <span class="keywordtype">float</span> c12 = 0.0f;</div><div class="line"><a name="l00356"></a><span class="lineno">  356</span>&#160;    <span class="keywordtype">float</span> c13 = 0.0f;</div><div class="line"><a name="l00357"></a><span class="lineno">  357</span>&#160;    <span class="keywordtype">float</span> c20 = 0.0f;</div><div class="line"><a name="l00358"></a><span class="lineno">  358</span>&#160;    <span class="keywordtype">float</span> c21 = 0.0f;</div><div class="line"><a name="l00359"></a><span class="lineno">  359</span>&#160;    <span class="keywordtype">float</span> c22 = 0.0f;</div><div class="line"><a name="l00360"></a><span class="lineno">  360</span>&#160;    <span class="keywordtype">float</span> c23 = 0.0f;</div><div class="line"><a name="l00361"></a><span class="lineno">  361</span>&#160;    <span class="keywordtype">float</span> c30 = 0.0f;</div><div class="line"><a name="l00362"></a><span class="lineno">  362</span>&#160;    <span class="keywordtype">float</span> c31 = 0.0f;</div><div class="line"><a name="l00363"></a><span class="lineno">  363</span>&#160;    <span class="keywordtype">float</span> c32 = 0.0f;</div><div class="line"><a name="l00364"></a><span class="lineno">  364</span>&#160;    <span class="keywordtype">float</span> c33 = 0.0f;</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;<span class="preprocessor">#define COLS_MTX_B (COLS_B / (4 * MULT_TRANSPOSE1XW_WIDTH))</span></div><div class="line"><a name="l00367"></a><span class="lineno">  367</span>&#160;</div><div class="line"><a name="l00368"></a><span class="lineno">  368</span>&#160;    <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l00369"></a><span class="lineno">  369</span>&#160;    <span class="keywordflow">for</span>(; i &lt;= (int)(COLS_MTX_B - 4); i += 4)</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;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00372"></a><span class="lineno">  372</span>&#160;        float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00373"></a><span class="lineno">  373</span>&#160;        float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00374"></a><span class="lineno">  374</span>&#160;</div><div class="line"><a name="l00375"></a><span class="lineno">  375</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00376"></a><span class="lineno">  376</span>&#160;        src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00377"></a><span class="lineno">  377</span>&#160;</div><div class="line"><a name="l00378"></a><span class="lineno">  378</span>&#160;        c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00379"></a><span class="lineno">  379</span>&#160;        c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00380"></a><span class="lineno">  380</span>&#160;        c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00381"></a><span class="lineno">  381</span>&#160;        c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00382"></a><span class="lineno">  382</span>&#160;</div><div class="line"><a name="l00383"></a><span class="lineno">  383</span>&#160;        c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00384"></a><span class="lineno">  384</span>&#160;        c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00385"></a><span class="lineno">  385</span>&#160;        c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00386"></a><span class="lineno">  386</span>&#160;        c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00387"></a><span class="lineno">  387</span>&#160;</div><div class="line"><a name="l00388"></a><span class="lineno">  388</span>&#160;        c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00389"></a><span class="lineno">  389</span>&#160;        c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00390"></a><span class="lineno">  390</span>&#160;        c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00391"></a><span class="lineno">  391</span>&#160;        c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00392"></a><span class="lineno">  392</span>&#160;</div><div class="line"><a name="l00393"></a><span class="lineno">  393</span>&#160;        c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00394"></a><span class="lineno">  394</span>&#160;        c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00395"></a><span class="lineno">  395</span>&#160;        c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00396"></a><span class="lineno">  396</span>&#160;        c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00397"></a><span class="lineno">  397</span>&#160;</div><div class="line"><a name="l00398"></a><span class="lineno">  398</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00399"></a><span class="lineno">  399</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00400"></a><span class="lineno">  400</span>&#160;        b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00401"></a><span class="lineno">  401</span>&#160;</div><div class="line"><a name="l00402"></a><span class="lineno">  402</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00403"></a><span class="lineno">  403</span>&#160;        src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00404"></a><span class="lineno">  404</span>&#160;</div><div class="line"><a name="l00405"></a><span class="lineno">  405</span>&#160;        c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00406"></a><span class="lineno">  406</span>&#160;        c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00407"></a><span class="lineno">  407</span>&#160;        c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00408"></a><span class="lineno">  408</span>&#160;        c03 = fma(a0.s0, b0.s3, c03);</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;        c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00411"></a><span class="lineno">  411</span>&#160;        c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00412"></a><span class="lineno">  412</span>&#160;        c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00413"></a><span class="lineno">  413</span>&#160;        c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00414"></a><span class="lineno">  414</span>&#160;</div><div class="line"><a name="l00415"></a><span class="lineno">  415</span>&#160;        c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00416"></a><span class="lineno">  416</span>&#160;        c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00417"></a><span class="lineno">  417</span>&#160;        c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00418"></a><span class="lineno">  418</span>&#160;        c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00419"></a><span class="lineno">  419</span>&#160;</div><div class="line"><a name="l00420"></a><span class="lineno">  420</span>&#160;        c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00421"></a><span class="lineno">  421</span>&#160;        c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00422"></a><span class="lineno">  422</span>&#160;        c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00423"></a><span class="lineno">  423</span>&#160;        c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00424"></a><span class="lineno">  424</span>&#160;</div><div class="line"><a name="l00425"></a><span class="lineno">  425</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00426"></a><span class="lineno">  426</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00427"></a><span class="lineno">  427</span>&#160;        b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00428"></a><span class="lineno">  428</span>&#160;</div><div class="line"><a name="l00429"></a><span class="lineno">  429</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00430"></a><span class="lineno">  430</span>&#160;        src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00431"></a><span class="lineno">  431</span>&#160;</div><div class="line"><a name="l00432"></a><span class="lineno">  432</span>&#160;        c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00433"></a><span class="lineno">  433</span>&#160;        c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00434"></a><span class="lineno">  434</span>&#160;        c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00435"></a><span class="lineno">  435</span>&#160;        c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00436"></a><span class="lineno">  436</span>&#160;</div><div class="line"><a name="l00437"></a><span class="lineno">  437</span>&#160;        c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00438"></a><span class="lineno">  438</span>&#160;        c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00439"></a><span class="lineno">  439</span>&#160;        c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00440"></a><span class="lineno">  440</span>&#160;        c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00441"></a><span class="lineno">  441</span>&#160;</div><div class="line"><a name="l00442"></a><span class="lineno">  442</span>&#160;        c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00443"></a><span class="lineno">  443</span>&#160;        c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00444"></a><span class="lineno">  444</span>&#160;        c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00445"></a><span class="lineno">  445</span>&#160;        c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00446"></a><span class="lineno">  446</span>&#160;</div><div class="line"><a name="l00447"></a><span class="lineno">  447</span>&#160;        c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00448"></a><span class="lineno">  448</span>&#160;        c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00449"></a><span class="lineno">  449</span>&#160;        c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00450"></a><span class="lineno">  450</span>&#160;        c33 = fma(a0.s3, b0.s3, c33);</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="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00453"></a><span class="lineno">  453</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00454"></a><span class="lineno">  454</span>&#160;        b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00455"></a><span class="lineno">  455</span>&#160;</div><div class="line"><a name="l00456"></a><span class="lineno">  456</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00457"></a><span class="lineno">  457</span>&#160;        src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00458"></a><span class="lineno">  458</span>&#160;</div><div class="line"><a name="l00459"></a><span class="lineno">  459</span>&#160;        c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00460"></a><span class="lineno">  460</span>&#160;        c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00461"></a><span class="lineno">  461</span>&#160;        c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00462"></a><span class="lineno">  462</span>&#160;        c03 = fma(a0.s0, b0.s3, c03);</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;        c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00465"></a><span class="lineno">  465</span>&#160;        c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00466"></a><span class="lineno">  466</span>&#160;        c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00467"></a><span class="lineno">  467</span>&#160;        c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00468"></a><span class="lineno">  468</span>&#160;</div><div class="line"><a name="l00469"></a><span class="lineno">  469</span>&#160;        c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00470"></a><span class="lineno">  470</span>&#160;        c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00471"></a><span class="lineno">  471</span>&#160;        c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00472"></a><span class="lineno">  472</span>&#160;        c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00473"></a><span class="lineno">  473</span>&#160;</div><div class="line"><a name="l00474"></a><span class="lineno">  474</span>&#160;        c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00475"></a><span class="lineno">  475</span>&#160;        c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00476"></a><span class="lineno">  476</span>&#160;        c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00477"></a><span class="lineno">  477</span>&#160;        c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00478"></a><span class="lineno">  478</span>&#160;    }</div><div class="line"><a name="l00479"></a><span class="lineno">  479</span>&#160;</div><div class="line"><a name="l00480"></a><span class="lineno">  480</span>&#160;    <span class="keywordflow">for</span>(; i &lt; (int)(COLS_MTX_B); ++i)</div><div class="line"><a name="l00481"></a><span class="lineno">  481</span>&#160;    {</div><div class="line"><a name="l00482"></a><span class="lineno">  482</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00483"></a><span class="lineno">  483</span>&#160;        float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00484"></a><span class="lineno">  484</span>&#160;        float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00485"></a><span class="lineno">  485</span>&#160;</div><div class="line"><a name="l00486"></a><span class="lineno">  486</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00487"></a><span class="lineno">  487</span>&#160;        src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00488"></a><span class="lineno">  488</span>&#160;</div><div class="line"><a name="l00489"></a><span class="lineno">  489</span>&#160;        c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00490"></a><span class="lineno">  490</span>&#160;        c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00491"></a><span class="lineno">  491</span>&#160;        c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00492"></a><span class="lineno">  492</span>&#160;        c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00493"></a><span class="lineno">  493</span>&#160;</div><div class="line"><a name="l00494"></a><span class="lineno">  494</span>&#160;        c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00495"></a><span class="lineno">  495</span>&#160;        c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00496"></a><span class="lineno">  496</span>&#160;        c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00497"></a><span class="lineno">  497</span>&#160;        c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00498"></a><span class="lineno">  498</span>&#160;</div><div class="line"><a name="l00499"></a><span class="lineno">  499</span>&#160;        c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00500"></a><span class="lineno">  500</span>&#160;        c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00501"></a><span class="lineno">  501</span>&#160;        c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00502"></a><span class="lineno">  502</span>&#160;        c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00503"></a><span class="lineno">  503</span>&#160;</div><div class="line"><a name="l00504"></a><span class="lineno">  504</span>&#160;        c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00505"></a><span class="lineno">  505</span>&#160;        c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00506"></a><span class="lineno">  506</span>&#160;        c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00507"></a><span class="lineno">  507</span>&#160;        c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00508"></a><span class="lineno">  508</span>&#160;    }</div><div class="line"><a name="l00509"></a><span class="lineno">  509</span>&#160;</div><div class="line"><a name="l00510"></a><span class="lineno">  510</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00511"></a><span class="lineno">  511</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00512"></a><span class="lineno">  512</span>&#160;</div><div class="line"><a name="l00513"></a><span class="lineno">  513</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00514"></a><span class="lineno">  514</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00515"></a><span class="lineno">  515</span>&#160;    c00 = c00 * ALPHA;</div><div class="line"><a name="l00516"></a><span class="lineno">  516</span>&#160;    c01 = c01 * ALPHA;</div><div class="line"><a name="l00517"></a><span class="lineno">  517</span>&#160;    c02 = c02 * ALPHA;</div><div class="line"><a name="l00518"></a><span class="lineno">  518</span>&#160;    c03 = c03 * ALPHA;</div><div class="line"><a name="l00519"></a><span class="lineno">  519</span>&#160;    c10 = c10 * ALPHA;</div><div class="line"><a name="l00520"></a><span class="lineno">  520</span>&#160;    c11 = c11 * ALPHA;</div><div class="line"><a name="l00521"></a><span class="lineno">  521</span>&#160;    c12 = c12 * ALPHA;</div><div class="line"><a name="l00522"></a><span class="lineno">  522</span>&#160;    c13 = c13 * ALPHA;</div><div class="line"><a name="l00523"></a><span class="lineno">  523</span>&#160;    c20 = c20 * ALPHA;</div><div class="line"><a name="l00524"></a><span class="lineno">  524</span>&#160;    c21 = c21 * ALPHA;</div><div class="line"><a name="l00525"></a><span class="lineno">  525</span>&#160;    c22 = c22 * ALPHA;</div><div class="line"><a name="l00526"></a><span class="lineno">  526</span>&#160;    c23 = c23 * ALPHA;</div><div class="line"><a name="l00527"></a><span class="lineno">  527</span>&#160;    c30 = c30 * ALPHA;</div><div class="line"><a name="l00528"></a><span class="lineno">  528</span>&#160;    c31 = c31 * ALPHA;</div><div class="line"><a name="l00529"></a><span class="lineno">  529</span>&#160;    c32 = c32 * ALPHA;</div><div class="line"><a name="l00530"></a><span class="lineno">  530</span>&#160;    c33 = c33 * ALPHA;</div><div class="line"><a name="l00531"></a><span class="lineno">  531</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></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="comment">// Compute dst address</span></div><div class="line"><a name="l00534"></a><span class="lineno">  534</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l00535"></a><span class="lineno">  535</span>&#160;</div><div class="line"><a name="l00536"></a><span class="lineno">  536</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00537"></a><span class="lineno">  537</span>&#160;    dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00538"></a><span class="lineno">  538</span>&#160;</div><div class="line"><a name="l00539"></a><span class="lineno">  539</span>&#160;    <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00540"></a><span class="lineno">  540</span>&#160;    vstore4((float4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l00541"></a><span class="lineno">  541</span>&#160;    vstore4((float4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l00542"></a><span class="lineno">  542</span>&#160;    vstore4((float4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l00543"></a><span class="lineno">  543</span>&#160;    vstore4((float4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l00544"></a><span class="lineno">  544</span>&#160;}</div><div class="line"><a name="l00545"></a><span class="lineno">  545</span>&#160;</div><div class="line"><a name="l00546"></a><span class="lineno">  546</span>&#160;<span class="comment">// Undefine local defines</span></div><div class="line"><a name="l00547"></a><span class="lineno">  547</span>&#160;<span class="preprocessor">#undef COLS_MTX_B</span></div><div class="line"><a name="l00548"></a><span class="lineno">  548</span>&#160;</div><div class="line"><a name="l00549"></a><span class="lineno">  549</span>&#160;<span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l00550"></a><span class="lineno">  550</span>&#160;</div><div class="line"><a name="l00578"></a><span class="lineno">  578</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00579"></a><span class="lineno">  579</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00580"></a><span class="lineno">  580</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00581"></a><span class="lineno">  581</span>&#160;                                                 uint src0_stride_z,</div><div class="line"><a name="l00582"></a><span class="lineno">  582</span>&#160;                                                 uint src1_stride_z,</div><div class="line"><a name="l00583"></a><span class="lineno">  583</span>&#160;                                                 uint dst_stride_z)</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;    <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00586"></a><span class="lineno">  586</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00587"></a><span class="lineno">  587</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</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="comment">// Offset</span></div><div class="line"><a name="l00590"></a><span class="lineno">  590</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00591"></a><span class="lineno">  591</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8;</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;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00594"></a><span class="lineno">  594</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00595"></a><span class="lineno">  595</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l00596"></a><span class="lineno">  596</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l00597"></a><span class="lineno">  597</span>&#160;</div><div class="line"><a name="l00598"></a><span class="lineno">  598</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00599"></a><span class="lineno">  599</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00600"></a><span class="lineno">  600</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00601"></a><span class="lineno">  601</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00602"></a><span class="lineno">  602</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00603"></a><span class="lineno">  603</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00604"></a><span class="lineno">  604</span>&#160;</div><div class="line"><a name="l00605"></a><span class="lineno">  605</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_addr_a = (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l00606"></a><span class="lineno">  606</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_addr_b = (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src1_addr_in_bytes);</div><div class="line"><a name="l00607"></a><span class="lineno">  607</span>&#160;</div><div class="line"><a name="l00608"></a><span class="lineno">  608</span>&#160;    <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00609"></a><span class="lineno">  609</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00610"></a><span class="lineno">  610</span>&#160;</div><div class="line"><a name="l00611"></a><span class="lineno">  611</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l00612"></a><span class="lineno">  612</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l00613"></a><span class="lineno">  613</span>&#160;</div><div class="line"><a name="l00614"></a><span class="lineno">  614</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00615"></a><span class="lineno">  615</span>&#160;    half8 c00 = 0.0f;</div><div class="line"><a name="l00616"></a><span class="lineno">  616</span>&#160;    half8 c10 = 0.0f;</div><div class="line"><a name="l00617"></a><span class="lineno">  617</span>&#160;    half8 c20 = 0.0f;</div><div class="line"><a name="l00618"></a><span class="lineno">  618</span>&#160;    half8 c30 = 0.0f;</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;    <span class="keywordflow">for</span>(; src_addr_b &lt;= (src_end_addr_b - (int)(16 * MULT_TRANSPOSE1XW_WIDTH)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH)</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="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00623"></a><span class="lineno">  623</span>&#160;        half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00624"></a><span class="lineno">  624</span>&#160;        half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00625"></a><span class="lineno">  625</span>&#160;</div><div class="line"><a name="l00626"></a><span class="lineno">  626</span>&#160;        c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00627"></a><span class="lineno">  627</span>&#160;        c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00628"></a><span class="lineno">  628</span>&#160;        c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00629"></a><span class="lineno">  629</span>&#160;        c30 += (half8)a0.s3 * b0;</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="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00632"></a><span class="lineno">  632</span>&#160;        a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT);</div><div class="line"><a name="l00633"></a><span class="lineno">  633</span>&#160;        b0 = vload8(0, src_addr_b + 8 * MULT_TRANSPOSE1XW_WIDTH);</div><div class="line"><a name="l00634"></a><span class="lineno">  634</span>&#160;</div><div class="line"><a name="l00635"></a><span class="lineno">  635</span>&#160;        c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00636"></a><span class="lineno">  636</span>&#160;        c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00637"></a><span class="lineno">  637</span>&#160;        c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00638"></a><span class="lineno">  638</span>&#160;        c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00639"></a><span class="lineno">  639</span>&#160;    }</div><div class="line"><a name="l00640"></a><span class="lineno">  640</span>&#160;</div><div class="line"><a name="l00641"></a><span class="lineno">  641</span>&#160;    <span class="keywordflow">for</span>(; src_addr_b &lt; src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00642"></a><span class="lineno">  642</span>&#160;    {</div><div class="line"><a name="l00643"></a><span class="lineno">  643</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00644"></a><span class="lineno">  644</span>&#160;        half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00645"></a><span class="lineno">  645</span>&#160;        half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00646"></a><span class="lineno">  646</span>&#160;</div><div class="line"><a name="l00647"></a><span class="lineno">  647</span>&#160;        c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00648"></a><span class="lineno">  648</span>&#160;        c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00649"></a><span class="lineno">  649</span>&#160;        c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00650"></a><span class="lineno">  650</span>&#160;        c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00651"></a><span class="lineno">  651</span>&#160;    }</div><div class="line"><a name="l00652"></a><span class="lineno">  652</span>&#160;</div><div class="line"><a name="l00653"></a><span class="lineno">  653</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00654"></a><span class="lineno">  654</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00655"></a><span class="lineno">  655</span>&#160;</div><div class="line"><a name="l00656"></a><span class="lineno">  656</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00657"></a><span class="lineno">  657</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00658"></a><span class="lineno">  658</span>&#160;    c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l00659"></a><span class="lineno">  659</span>&#160;    c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l00660"></a><span class="lineno">  660</span>&#160;    c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l00661"></a><span class="lineno">  661</span>&#160;    c30 = c30 * (half8)ALPHA;</div><div class="line"><a name="l00662"></a><span class="lineno">  662</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00663"></a><span class="lineno">  663</span>&#160;</div><div class="line"><a name="l00664"></a><span class="lineno">  664</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00665"></a><span class="lineno">  665</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</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;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00668"></a><span class="lineno">  668</span>&#160;    dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00669"></a><span class="lineno">  669</span>&#160;</div><div class="line"><a name="l00670"></a><span class="lineno">  670</span>&#160;    <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l00671"></a><span class="lineno">  671</span>&#160;    vstore8(c00, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l00672"></a><span class="lineno">  672</span>&#160;    vstore8(c10, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l00673"></a><span class="lineno">  673</span>&#160;    vstore8(c20, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l00674"></a><span class="lineno">  674</span>&#160;    vstore8(c30, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l00675"></a><span class="lineno">  675</span>&#160;}</div><div class="line"><a name="l00676"></a><span class="lineno">  676</span>&#160;</div><div class="line"><a name="l00705"></a><span class="lineno">  705</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f16_bifrost(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00706"></a><span class="lineno">  706</span>&#160;                                                         <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00707"></a><span class="lineno">  707</span>&#160;                                                         <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00708"></a><span class="lineno">  708</span>&#160;                                                         uint src0_stride_z,</div><div class="line"><a name="l00709"></a><span class="lineno">  709</span>&#160;                                                         uint src1_stride_z,</div><div class="line"><a name="l00710"></a><span class="lineno">  710</span>&#160;                                                         uint dst_stride_z)</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="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00713"></a><span class="lineno">  713</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00714"></a><span class="lineno">  714</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00715"></a><span class="lineno">  715</span>&#160;</div><div class="line"><a name="l00716"></a><span class="lineno">  716</span>&#160;    <span class="comment">// Offset</span></div><div class="line"><a name="l00717"></a><span class="lineno">  717</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00718"></a><span class="lineno">  718</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8;</div><div class="line"><a name="l00719"></a><span class="lineno">  719</span>&#160;</div><div class="line"><a name="l00720"></a><span class="lineno">  720</span>&#160;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00721"></a><span class="lineno">  721</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00722"></a><span class="lineno">  722</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l00723"></a><span class="lineno">  723</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l00724"></a><span class="lineno">  724</span>&#160;</div><div class="line"><a name="l00725"></a><span class="lineno">  725</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00726"></a><span class="lineno">  726</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00727"></a><span class="lineno">  727</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00728"></a><span class="lineno">  728</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00729"></a><span class="lineno">  729</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00730"></a><span class="lineno">  730</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00731"></a><span class="lineno">  731</span>&#160;</div><div class="line"><a name="l00732"></a><span class="lineno">  732</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_addr_a = (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l00733"></a><span class="lineno">  733</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_addr_b = (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src1_addr_in_bytes);</div><div class="line"><a name="l00734"></a><span class="lineno">  734</span>&#160;</div><div class="line"><a name="l00735"></a><span class="lineno">  735</span>&#160;    <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00736"></a><span class="lineno">  736</span>&#160;    __global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00737"></a><span class="lineno">  737</span>&#160;</div><div class="line"><a name="l00738"></a><span class="lineno">  738</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l00739"></a><span class="lineno">  739</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l00740"></a><span class="lineno">  740</span>&#160;</div><div class="line"><a name="l00741"></a><span class="lineno">  741</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00742"></a><span class="lineno">  742</span>&#160;    half8 c00 = 0.0f;</div><div class="line"><a name="l00743"></a><span class="lineno">  743</span>&#160;    half8 c10 = 0.0f;</div><div class="line"><a name="l00744"></a><span class="lineno">  744</span>&#160;    half8 c20 = 0.0f;</div><div class="line"><a name="l00745"></a><span class="lineno">  745</span>&#160;    half8 c30 = 0.0f;</div><div class="line"><a name="l00746"></a><span class="lineno">  746</span>&#160;</div><div class="line"><a name="l00747"></a><span class="lineno">  747</span>&#160;<span class="preprocessor">#define COLS_MTX_B (COLS_B / (8 * MULT_TRANSPOSE1XW_WIDTH))</span></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="keywordtype">int</span> i = 0;</div><div class="line"><a name="l00750"></a><span class="lineno">  750</span>&#160;    <span class="keywordflow">for</span>(; i &lt;= (int)(COLS_MTX_B - 4); i += 4)</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 MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00753"></a><span class="lineno">  753</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00754"></a><span class="lineno">  754</span>&#160;        half8 a0 = vload8(0, src_addr_a);</div><div class="line"><a name="l00755"></a><span class="lineno">  755</span>&#160;        half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00756"></a><span class="lineno">  756</span>&#160;</div><div class="line"><a name="l00757"></a><span class="lineno">  757</span>&#160;        src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00758"></a><span class="lineno">  758</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00759"></a><span class="lineno">  759</span>&#160;</div><div class="line"><a name="l00760"></a><span class="lineno">  760</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00761"></a><span class="lineno">  761</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00762"></a><span class="lineno">  762</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00763"></a><span class="lineno">  763</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l00764"></a><span class="lineno">  764</span>&#160;</div><div class="line"><a name="l00765"></a><span class="lineno">  765</span>&#160;        <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00766"></a><span class="lineno">  766</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00767"></a><span class="lineno">  767</span>&#160;</div><div class="line"><a name="l00768"></a><span class="lineno">  768</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00769"></a><span class="lineno">  769</span>&#160;</div><div class="line"><a name="l00770"></a><span class="lineno">  770</span>&#160;        c00 = fma((half8)a0.s4, b0, c00);</div><div class="line"><a name="l00771"></a><span class="lineno">  771</span>&#160;        c10 = fma((half8)a0.s5, b0, c10);</div><div class="line"><a name="l00772"></a><span class="lineno">  772</span>&#160;        c20 = fma((half8)a0.s6, b0, c20);</div><div class="line"><a name="l00773"></a><span class="lineno">  773</span>&#160;        c30 = fma((half8)a0.s7, b0, c30);</div><div class="line"><a name="l00774"></a><span class="lineno">  774</span>&#160;</div><div class="line"><a name="l00775"></a><span class="lineno">  775</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00776"></a><span class="lineno">  776</span>&#160;        a0 = vload8(0, src_addr_a);</div><div class="line"><a name="l00777"></a><span class="lineno">  777</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00778"></a><span class="lineno">  778</span>&#160;</div><div class="line"><a name="l00779"></a><span class="lineno">  779</span>&#160;        src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00780"></a><span class="lineno">  780</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00781"></a><span class="lineno">  781</span>&#160;</div><div class="line"><a name="l00782"></a><span class="lineno">  782</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00783"></a><span class="lineno">  783</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00784"></a><span class="lineno">  784</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00785"></a><span class="lineno">  785</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l00786"></a><span class="lineno">  786</span>&#160;</div><div class="line"><a name="l00787"></a><span class="lineno">  787</span>&#160;        <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00788"></a><span class="lineno">  788</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00789"></a><span class="lineno">  789</span>&#160;</div><div class="line"><a name="l00790"></a><span class="lineno">  790</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00791"></a><span class="lineno">  791</span>&#160;</div><div class="line"><a name="l00792"></a><span class="lineno">  792</span>&#160;        c00 = fma((half8)a0.s4, b0, c00);</div><div class="line"><a name="l00793"></a><span class="lineno">  793</span>&#160;        c10 = fma((half8)a0.s5, b0, c10);</div><div class="line"><a name="l00794"></a><span class="lineno">  794</span>&#160;        c20 = fma((half8)a0.s6, b0, c20);</div><div class="line"><a name="l00795"></a><span class="lineno">  795</span>&#160;        c30 = fma((half8)a0.s7, b0, c30);</div><div class="line"><a name="l00796"></a><span class="lineno">  796</span>&#160;<span class="preprocessor">#else  // MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00797"></a><span class="lineno">  797</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00798"></a><span class="lineno">  798</span>&#160;        half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00799"></a><span class="lineno">  799</span>&#160;        half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00800"></a><span class="lineno">  800</span>&#160;</div><div class="line"><a name="l00801"></a><span class="lineno">  801</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00802"></a><span class="lineno">  802</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00803"></a><span class="lineno">  803</span>&#160;</div><div class="line"><a name="l00804"></a><span class="lineno">  804</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00805"></a><span class="lineno">  805</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00806"></a><span class="lineno">  806</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00807"></a><span class="lineno">  807</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</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;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00810"></a><span class="lineno">  810</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00811"></a><span class="lineno">  811</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00812"></a><span class="lineno">  812</span>&#160;</div><div class="line"><a name="l00813"></a><span class="lineno">  813</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00814"></a><span class="lineno">  814</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00815"></a><span class="lineno">  815</span>&#160;</div><div class="line"><a name="l00816"></a><span class="lineno">  816</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00817"></a><span class="lineno">  817</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00818"></a><span class="lineno">  818</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00819"></a><span class="lineno">  819</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</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;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00822"></a><span class="lineno">  822</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00823"></a><span class="lineno">  823</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00824"></a><span class="lineno">  824</span>&#160;</div><div class="line"><a name="l00825"></a><span class="lineno">  825</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00826"></a><span class="lineno">  826</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00827"></a><span class="lineno">  827</span>&#160;</div><div class="line"><a name="l00828"></a><span class="lineno">  828</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00829"></a><span class="lineno">  829</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00830"></a><span class="lineno">  830</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00831"></a><span class="lineno">  831</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l00832"></a><span class="lineno">  832</span>&#160;</div><div class="line"><a name="l00833"></a><span class="lineno">  833</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00834"></a><span class="lineno">  834</span>&#160;        a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00835"></a><span class="lineno">  835</span>&#160;        b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00836"></a><span class="lineno">  836</span>&#160;</div><div class="line"><a name="l00837"></a><span class="lineno">  837</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00838"></a><span class="lineno">  838</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00839"></a><span class="lineno">  839</span>&#160;</div><div class="line"><a name="l00840"></a><span class="lineno">  840</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00841"></a><span class="lineno">  841</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00842"></a><span class="lineno">  842</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00843"></a><span class="lineno">  843</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l00844"></a><span class="lineno">  844</span>&#160;<span class="preprocessor">#endif // MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00845"></a><span class="lineno">  845</span>&#160;    }</div><div class="line"><a name="l00846"></a><span class="lineno">  846</span>&#160;</div><div class="line"><a name="l00847"></a><span class="lineno">  847</span>&#160;    <span class="keywordflow">for</span>(; i &lt; (int)(COLS_MTX_B); ++i)</div><div class="line"><a name="l00848"></a><span class="lineno">  848</span>&#160;    {</div><div class="line"><a name="l00849"></a><span class="lineno">  849</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00850"></a><span class="lineno">  850</span>&#160;        half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00851"></a><span class="lineno">  851</span>&#160;        half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00852"></a><span class="lineno">  852</span>&#160;</div><div class="line"><a name="l00853"></a><span class="lineno">  853</span>&#160;        src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00854"></a><span class="lineno">  854</span>&#160;        src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00855"></a><span class="lineno">  855</span>&#160;</div><div class="line"><a name="l00856"></a><span class="lineno">  856</span>&#160;        c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l00857"></a><span class="lineno">  857</span>&#160;        c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l00858"></a><span class="lineno">  858</span>&#160;        c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l00859"></a><span class="lineno">  859</span>&#160;        c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l00860"></a><span class="lineno">  860</span>&#160;    }</div><div class="line"><a name="l00861"></a><span class="lineno">  861</span>&#160;</div><div class="line"><a name="l00862"></a><span class="lineno">  862</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00863"></a><span class="lineno">  863</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00864"></a><span class="lineno">  864</span>&#160;</div><div class="line"><a name="l00865"></a><span class="lineno">  865</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00866"></a><span class="lineno">  866</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00867"></a><span class="lineno">  867</span>&#160;    c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l00868"></a><span class="lineno">  868</span>&#160;    c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l00869"></a><span class="lineno">  869</span>&#160;    c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l00870"></a><span class="lineno">  870</span>&#160;    c30 = c30 * (half8)ALPHA;</div><div class="line"><a name="l00871"></a><span class="lineno">  871</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00872"></a><span class="lineno">  872</span>&#160;</div><div class="line"><a name="l00873"></a><span class="lineno">  873</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00874"></a><span class="lineno">  874</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</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="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00877"></a><span class="lineno">  877</span>&#160;    dst_addr += z * dst_stride_z;</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;    <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l00880"></a><span class="lineno">  880</span>&#160;    vstore8(c00, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l00881"></a><span class="lineno">  881</span>&#160;    vstore8(c10, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l00882"></a><span class="lineno">  882</span>&#160;    vstore8(c20, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l00883"></a><span class="lineno">  883</span>&#160;    vstore8(c30, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l00884"></a><span class="lineno">  884</span>&#160;}</div><div class="line"><a name="l00885"></a><span class="lineno">  885</span>&#160;</div><div class="line"><a name="l00886"></a><span class="lineno">  886</span>&#160;<span class="comment">// Undefine local defines</span></div><div class="line"><a name="l00887"></a><span class="lineno">  887</span>&#160;<span class="preprocessor">#undef COLS_MTX_B</span></div><div class="line"><a name="l00888"></a><span class="lineno">  888</span>&#160;</div><div class="line"><a name="l00889"></a><span class="lineno">  889</span>&#160;<span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></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="preprocessor">#if defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l00892"></a><span class="lineno">  892</span>&#160;</div><div class="line"><a name="l00921"></a><span class="lineno">  921</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_qs8(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00922"></a><span class="lineno">  922</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00923"></a><span class="lineno">  923</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00924"></a><span class="lineno">  924</span>&#160;                                                 uint src0_stride_z,</div><div class="line"><a name="l00925"></a><span class="lineno">  925</span>&#160;                                                 uint src1_stride_z,</div><div class="line"><a name="l00926"></a><span class="lineno">  926</span>&#160;                                                 uint dst_stride_z)</div><div class="line"><a name="l00927"></a><span class="lineno">  927</span>&#160;{</div><div class="line"><a name="l00928"></a><span class="lineno">  928</span>&#160;    <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00929"></a><span class="lineno">  929</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00930"></a><span class="lineno">  930</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00931"></a><span class="lineno">  931</span>&#160;</div><div class="line"><a name="l00932"></a><span class="lineno">  932</span>&#160;    <span class="comment">// Offset</span></div><div class="line"><a name="l00933"></a><span class="lineno">  933</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00934"></a><span class="lineno">  934</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 16;</div><div class="line"><a name="l00935"></a><span class="lineno">  935</span>&#160;</div><div class="line"><a name="l00936"></a><span class="lineno">  936</span>&#160;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00937"></a><span class="lineno">  937</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00938"></a><span class="lineno">  938</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l00939"></a><span class="lineno">  939</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l00940"></a><span class="lineno">  940</span>&#160;</div><div class="line"><a name="l00941"></a><span class="lineno">  941</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00942"></a><span class="lineno">  942</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00943"></a><span class="lineno">  943</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00944"></a><span class="lineno">  944</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00945"></a><span class="lineno">  945</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00946"></a><span class="lineno">  946</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00947"></a><span class="lineno">  947</span>&#160;</div><div class="line"><a name="l00948"></a><span class="lineno">  948</span>&#160;    __global <span class="keywordtype">char</span> *src_addr_a = (__global <span class="keywordtype">char</span> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l00949"></a><span class="lineno">  949</span>&#160;    __global <span class="keywordtype">char</span> *src_addr_b = (__global <span class="keywordtype">char</span> *)(src1_ptr + src1_addr_in_bytes);</div><div class="line"><a name="l00950"></a><span class="lineno">  950</span>&#160;</div><div class="line"><a name="l00951"></a><span class="lineno">  951</span>&#160;    <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00952"></a><span class="lineno">  952</span>&#160;    __global <span class="keywordtype">char</span> *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00953"></a><span class="lineno">  953</span>&#160;</div><div class="line"><a name="l00954"></a><span class="lineno">  954</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l00955"></a><span class="lineno">  955</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l00956"></a><span class="lineno">  956</span>&#160;</div><div class="line"><a name="l00957"></a><span class="lineno">  957</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00958"></a><span class="lineno">  958</span>&#160;    short8 c00 = 0.0f;</div><div class="line"><a name="l00959"></a><span class="lineno">  959</span>&#160;    short8 c10 = 0.0f;</div><div class="line"><a name="l00960"></a><span class="lineno">  960</span>&#160;    short8 c20 = 0.0f;</div><div class="line"><a name="l00961"></a><span class="lineno">  961</span>&#160;    short8 c30 = 0.0f;</div><div class="line"><a name="l00962"></a><span class="lineno">  962</span>&#160;    short8 c01 = 0.0f;</div><div class="line"><a name="l00963"></a><span class="lineno">  963</span>&#160;    short8 c11 = 0.0f;</div><div class="line"><a name="l00964"></a><span class="lineno">  964</span>&#160;    short8 c21 = 0.0f;</div><div class="line"><a name="l00965"></a><span class="lineno">  965</span>&#160;    short8 c31 = 0.0f;</div><div class="line"><a name="l00966"></a><span class="lineno">  966</span>&#160;</div><div class="line"><a name="l00967"></a><span class="lineno">  967</span>&#160;    <span class="comment">// This for loop performs 1 accumulation for each iteration</span></div><div class="line"><a name="l00968"></a><span class="lineno">  968</span>&#160;    <span class="keywordflow">for</span>(; src_addr_b &lt; src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 16 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00969"></a><span class="lineno">  969</span>&#160;    {</div><div class="line"><a name="l00970"></a><span class="lineno">  970</span>&#160;        <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00971"></a><span class="lineno">  971</span>&#160;        char4  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00972"></a><span class="lineno">  972</span>&#160;        char16 b0 = vload16(0, src_addr_b);</div><div class="line"><a name="l00973"></a><span class="lineno">  973</span>&#160;</div><div class="line"><a name="l00974"></a><span class="lineno">  974</span>&#160;        c00 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l00975"></a><span class="lineno">  975</span>&#160;        c10 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c10, (char8)a0.s1, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l00976"></a><span class="lineno">  976</span>&#160;        c20 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c20, (char8)a0.s2, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l00977"></a><span class="lineno">  977</span>&#160;        c30 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c30, (char8)a0.s3, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l00978"></a><span class="lineno">  978</span>&#160;</div><div class="line"><a name="l00979"></a><span class="lineno">  979</span>&#160;        c01 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l00980"></a><span class="lineno">  980</span>&#160;        c11 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c11, (char8)a0.s1, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l00981"></a><span class="lineno">  981</span>&#160;        c21 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c21, (char8)a0.s2, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l00982"></a><span class="lineno">  982</span>&#160;        c31 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(c31, (char8)a0.s3, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l00983"></a><span class="lineno">  983</span>&#160;    }</div><div class="line"><a name="l00984"></a><span class="lineno">  984</span>&#160;</div><div class="line"><a name="l00985"></a><span class="lineno">  985</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00986"></a><span class="lineno">  986</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00987"></a><span class="lineno">  987</span>&#160;</div><div class="line"><a name="l00988"></a><span class="lineno">  988</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00989"></a><span class="lineno">  989</span>&#160;    char16 c00_qs8 = convert_char16_sat((short16)(c00, c01));</div><div class="line"><a name="l00990"></a><span class="lineno">  990</span>&#160;    char16 c10_qs8 = convert_char16_sat((short16)(c10, c11));</div><div class="line"><a name="l00991"></a><span class="lineno">  991</span>&#160;    char16 c20_qs8 = convert_char16_sat((short16)(c20, c21));</div><div class="line"><a name="l00992"></a><span class="lineno">  992</span>&#160;    char16 c30_qs8 = convert_char16_sat((short16)(c30, c31));</div><div class="line"><a name="l00993"></a><span class="lineno">  993</span>&#160;</div><div class="line"><a name="l00994"></a><span class="lineno">  994</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00995"></a><span class="lineno">  995</span>&#160;    c00_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(c00_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l00996"></a><span class="lineno">  996</span>&#160;    c10_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(c10_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l00997"></a><span class="lineno">  997</span>&#160;    c20_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(c20_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l00998"></a><span class="lineno">  998</span>&#160;    c30_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(c30_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l00999"></a><span class="lineno">  999</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01000"></a><span class="lineno"> 1000</span>&#160;</div><div class="line"><a name="l01001"></a><span class="lineno"> 1001</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01002"></a><span class="lineno"> 1002</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l01003"></a><span class="lineno"> 1003</span>&#160;</div><div class="line"><a name="l01004"></a><span class="lineno"> 1004</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01005"></a><span class="lineno"> 1005</span>&#160;    dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01006"></a><span class="lineno"> 1006</span>&#160;</div><div class="line"><a name="l01007"></a><span class="lineno"> 1007</span>&#160;    <span class="comment">// Store 16x4 block</span></div><div class="line"><a name="l01008"></a><span class="lineno"> 1008</span>&#160;    vstore16(c00_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01009"></a><span class="lineno"> 1009</span>&#160;    vstore16(c10_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01010"></a><span class="lineno"> 1010</span>&#160;    vstore16(c20_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01011"></a><span class="lineno"> 1011</span>&#160;    vstore16(c30_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span>&#160;}</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span>&#160;</div><div class="line"><a name="l01043"></a><span class="lineno"> 1043</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_qs16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01044"></a><span class="lineno"> 1044</span>&#160;                                                  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01045"></a><span class="lineno"> 1045</span>&#160;                                                  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01046"></a><span class="lineno"> 1046</span>&#160;                                                  uint src0_stride_z,</div><div class="line"><a name="l01047"></a><span class="lineno"> 1047</span>&#160;                                                  uint src1_stride_z,</div><div class="line"><a name="l01048"></a><span class="lineno"> 1048</span>&#160;                                                  uint dst_stride_z)</div><div class="line"><a name="l01049"></a><span class="lineno"> 1049</span>&#160;{</div><div class="line"><a name="l01050"></a><span class="lineno"> 1050</span>&#160;    <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01051"></a><span class="lineno"> 1051</span>&#160;    <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01052"></a><span class="lineno"> 1052</span>&#160;    <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01053"></a><span class="lineno"> 1053</span>&#160;</div><div class="line"><a name="l01054"></a><span class="lineno"> 1054</span>&#160;    <span class="comment">// Offset</span></div><div class="line"><a name="l01055"></a><span class="lineno"> 1055</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l01056"></a><span class="lineno"> 1056</span>&#160;    <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % MULT_TRANSPOSE1XW_WIDTH) * 8;</div><div class="line"><a name="l01057"></a><span class="lineno"> 1057</span>&#160;</div><div class="line"><a name="l01058"></a><span class="lineno"> 1058</span>&#160;    <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l01059"></a><span class="lineno"> 1059</span>&#160;    <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l01060"></a><span class="lineno"> 1060</span>&#160;    <span class="keywordtype">int</span> src0_addr_in_bytes = z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes;</div><div class="line"><a name="l01061"></a><span class="lineno"> 1061</span>&#160;    <span class="keywordtype">int</span> src1_addr_in_bytes = x * src1_stride_y + src1_offset_first_element_in_bytes;</div><div class="line"><a name="l01062"></a><span class="lineno"> 1062</span>&#160;</div><div class="line"><a name="l01063"></a><span class="lineno"> 1063</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01064"></a><span class="lineno"> 1064</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01065"></a><span class="lineno"> 1065</span>&#160;    src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01066"></a><span class="lineno"> 1066</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01067"></a><span class="lineno"> 1067</span>&#160;    src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l01068"></a><span class="lineno"> 1068</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01069"></a><span class="lineno"> 1069</span>&#160;</div><div class="line"><a name="l01070"></a><span class="lineno"> 1070</span>&#160;    __global <span class="keywordtype">short</span> *src_addr_a = (__global <span class="keywordtype">short</span> *)(src0_ptr + src0_addr_in_bytes);</div><div class="line"><a name="l01071"></a><span class="lineno"> 1071</span>&#160;    __global <span class="keywordtype">short</span> *src_addr_b = (__global <span class="keywordtype">short</span> *)(src1_ptr + src1_addr_in_bytes);</div><div class="line"><a name="l01072"></a><span class="lineno"> 1072</span>&#160;</div><div class="line"><a name="l01073"></a><span class="lineno"> 1073</span>&#160;    <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l01074"></a><span class="lineno"> 1074</span>&#160;    __global <span class="keywordtype">short</span> *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l01075"></a><span class="lineno"> 1075</span>&#160;</div><div class="line"><a name="l01076"></a><span class="lineno"> 1076</span>&#160;    src_addr_a += offset_row_a;</div><div class="line"><a name="l01077"></a><span class="lineno"> 1077</span>&#160;    src_addr_b += offset_row_b;</div><div class="line"><a name="l01078"></a><span class="lineno"> 1078</span>&#160;</div><div class="line"><a name="l01079"></a><span class="lineno"> 1079</span>&#160;    <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l01080"></a><span class="lineno"> 1080</span>&#160;    int8 c00 = 0.0f;</div><div class="line"><a name="l01081"></a><span class="lineno"> 1081</span>&#160;    int8 c10 = 0.0f;</div><div class="line"><a name="l01082"></a><span class="lineno"> 1082</span>&#160;    int8 c20 = 0.0f;</div><div class="line"><a name="l01083"></a><span class="lineno"> 1083</span>&#160;    int8 c30 = 0.0f;</div><div class="line"><a name="l01084"></a><span class="lineno"> 1084</span>&#160;</div><div class="line"><a name="l01085"></a><span class="lineno"> 1085</span>&#160;    <span class="comment">// This for loop performs 1 accumulation for each iteration</span></div><div class="line"><a name="l01086"></a><span class="lineno"> 1086</span>&#160;    <span class="keywordflow">for</span>(; src_addr_b &lt; src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l01087"></a><span class="lineno"> 1087</span>&#160;    {</div><div class="line"><a name="l01088"></a><span class="lineno"> 1088</span>&#160;        <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l01089"></a><span class="lineno"> 1089</span>&#160;        short4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01090"></a><span class="lineno"> 1090</span>&#160;        short8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01091"></a><span class="lineno"> 1091</span>&#160;</div><div class="line"><a name="l01092"></a><span class="lineno"> 1092</span>&#160;        c00 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(c00, (short8)a0.s0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l01093"></a><span class="lineno"> 1093</span>&#160;        c10 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(c10, (short8)a0.s1, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l01094"></a><span class="lineno"> 1094</span>&#160;        c20 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(c20, (short8)a0.s2, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l01095"></a><span class="lineno"> 1095</span>&#160;        c30 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(c30, (short8)a0.s3, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l01096"></a><span class="lineno"> 1096</span>&#160;    }</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;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01099"></a><span class="lineno"> 1099</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01100"></a><span class="lineno"> 1100</span>&#160;</div><div class="line"><a name="l01101"></a><span class="lineno"> 1101</span>&#160;    <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l01102"></a><span class="lineno"> 1102</span>&#160;    short8 c00_qs16 = convert_short8_sat(c00);</div><div class="line"><a name="l01103"></a><span class="lineno"> 1103</span>&#160;    short8 c10_qs16 = convert_short8_sat(c10);</div><div class="line"><a name="l01104"></a><span class="lineno"> 1104</span>&#160;    short8 c20_qs16 = convert_short8_sat(c20);</div><div class="line"><a name="l01105"></a><span class="lineno"> 1105</span>&#160;    short8 c30_qs16 = convert_short8_sat(c30);</div><div class="line"><a name="l01106"></a><span class="lineno"> 1106</span>&#160;</div><div class="line"><a name="l01107"></a><span class="lineno"> 1107</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01108"></a><span class="lineno"> 1108</span>&#160;    c00_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(c00_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l01109"></a><span class="lineno"> 1109</span>&#160;    c10_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(c10_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l01110"></a><span class="lineno"> 1110</span>&#160;    c20_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(c20_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l01111"></a><span class="lineno"> 1111</span>&#160;    c30_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(c30_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l01112"></a><span class="lineno"> 1112</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01113"></a><span class="lineno"> 1113</span>&#160;</div><div class="line"><a name="l01114"></a><span class="lineno"> 1114</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01115"></a><span class="lineno"> 1115</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l01116"></a><span class="lineno"> 1116</span>&#160;</div><div class="line"><a name="l01117"></a><span class="lineno"> 1117</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01118"></a><span class="lineno"> 1118</span>&#160;    dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01119"></a><span class="lineno"> 1119</span>&#160;</div><div class="line"><a name="l01120"></a><span class="lineno"> 1120</span>&#160;    <span class="comment">// Store 8x4 block</span></div><div class="line"><a name="l01121"></a><span class="lineno"> 1121</span>&#160;    vstore8(c00_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01122"></a><span class="lineno"> 1122</span>&#160;    vstore8(c10_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01123"></a><span class="lineno"> 1123</span>&#160;    vstore8(c20_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01124"></a><span class="lineno"> 1124</span>&#160;    vstore8(c30_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01125"></a><span class="lineno"> 1125</span>&#160;}</div><div class="line"><a name="l01126"></a><span class="lineno"> 1126</span>&#160;<span class="preprocessor">#endif // defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l01127"></a><span class="lineno"> 1127</span>&#160;<span class="preprocessor">#endif // defined(COLS_B) &amp;&amp; defined(MULT_TRANSPOSE1XW_WIDTH) &amp;&amp; defined(MULT_INTERLEAVE4X4_HEIGHT)</span></div><div class="line"><a name="l01128"></a><span class="lineno"> 1128</span>&#160;</div><div class="line"><a name="l01129"></a><span class="lineno"> 1129</span>&#160;<span class="preprocessor">#if defined(COLS_A) &amp;&amp; defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) &amp;&amp; (NUM_ELEMS_PROCESSED_PER_THREAD_Y)</span></div><div class="line"><a name="l01130"></a><span class="lineno"> 1130</span>&#160;<span class="preprocessor">#if defined(DATA_TYPE)</span></div><div class="line"><a name="l01131"></a><span class="lineno"> 1131</span>&#160;<span class="preprocessor">#define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, NUM_ELEMS_PROCESSED_PER_THREAD_X)</span></div><div class="line"><a name="l01132"></a><span class="lineno"> 1132</span>&#160;</div><div class="line"><a name="l01160"></a><span class="lineno"> 1160</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_floating_point(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01161"></a><span class="lineno"> 1161</span>&#160;                                     <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01162"></a><span class="lineno"> 1162</span>&#160;                                     <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01163"></a><span class="lineno"> 1163</span>&#160;                                     uint src0_stride_z,</div><div class="line"><a name="l01164"></a><span class="lineno"> 1164</span>&#160;                                     uint src1_stride_z,</div><div class="line"><a name="l01165"></a><span class="lineno"> 1165</span>&#160;                                     uint dst_stride_z)</div><div class="line"><a name="l01166"></a><span class="lineno"> 1166</span>&#160;{</div><div class="line"><a name="l01167"></a><span class="lineno"> 1167</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01168"></a><span class="lineno"> 1168</span>&#160;</div><div class="line"><a name="l01169"></a><span class="lineno"> 1169</span>&#160;    <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01170"></a><span class="lineno"> 1170</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01171"></a><span class="lineno"> 1171</span>&#160;</div><div class="line"><a name="l01172"></a><span class="lineno"> 1172</span>&#160;    <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01173"></a><span class="lineno"> 1173</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01174"></a><span class="lineno"> 1174</span>&#160;</div><div class="line"><a name="l01175"></a><span class="lineno"> 1175</span>&#160;    <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01176"></a><span class="lineno"> 1176</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>);</div><div class="line"><a name="l01177"></a><span class="lineno"> 1177</span>&#160;</div><div class="line"><a name="l01178"></a><span class="lineno"> 1178</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01179"></a><span class="lineno"> 1179</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01180"></a><span class="lineno"> 1180</span>&#160;</div><div class="line"><a name="l01181"></a><span class="lineno"> 1181</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01182"></a><span class="lineno"> 1182</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01183"></a><span class="lineno"> 1183</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01184"></a><span class="lineno"> 1184</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01185"></a><span class="lineno"> 1185</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01186"></a><span class="lineno"> 1186</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01187"></a><span class="lineno"> 1187</span>&#160;</div><div class="line"><a name="l01188"></a><span class="lineno"> 1188</span>&#160;    <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (COLS_A * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>));</div><div class="line"><a name="l01189"></a><span class="lineno"> 1189</span>&#160;</div><div class="line"><a name="l01190"></a><span class="lineno"> 1190</span>&#160;    VECTOR_TYPE acc0 = 0.0f;</div><div class="line"><a name="l01191"></a><span class="lineno"> 1191</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01192"></a><span class="lineno"> 1192</span>&#160;    VECTOR_TYPE acc1 = 0.0f;</div><div class="line"><a name="l01193"></a><span class="lineno"> 1193</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01194"></a><span class="lineno"> 1194</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01195"></a><span class="lineno"> 1195</span>&#160;    VECTOR_TYPE acc2 = 0.0f;</div><div class="line"><a name="l01196"></a><span class="lineno"> 1196</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01197"></a><span class="lineno"> 1197</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01198"></a><span class="lineno"> 1198</span>&#160;    VECTOR_TYPE acc3 = 0.0f;</div><div class="line"><a name="l01199"></a><span class="lineno"> 1199</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01200"></a><span class="lineno"> 1200</span>&#160;</div><div class="line"><a name="l01201"></a><span class="lineno"> 1201</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt;= (end_row_vec_a - 2 * (int)<span class="keyword">sizeof</span>(DATA_TYPE)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(DATA_TYPE), 2 * src1_stride_y))</div><div class="line"><a name="l01202"></a><span class="lineno"> 1202</span>&#160;    {</div><div class="line"><a name="l01203"></a><span class="lineno"> 1203</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01204"></a><span class="lineno"> 1204</span>&#160;        <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01205"></a><span class="lineno"> 1205</span>&#160;        a0 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01206"></a><span class="lineno"> 1206</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01207"></a><span class="lineno"> 1207</span>&#160;        <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01208"></a><span class="lineno"> 1208</span>&#160;        a1 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01209"></a><span class="lineno"> 1209</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01210"></a><span class="lineno"> 1210</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01211"></a><span class="lineno"> 1211</span>&#160;        <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01212"></a><span class="lineno"> 1212</span>&#160;        a2 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01213"></a><span class="lineno"> 1213</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01214"></a><span class="lineno"> 1214</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01215"></a><span class="lineno"> 1215</span>&#160;        <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01216"></a><span class="lineno"> 1216</span>&#160;        a3 = vload2(0, (__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01217"></a><span class="lineno"> 1217</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01218"></a><span class="lineno"> 1218</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01219"></a><span class="lineno"> 1219</span>&#160;        VECTOR_TYPE b0 = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01220"></a><span class="lineno"> 1220</span>&#160;        VECTOR_TYPE b1 = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1 + src1_stride_y));</div><div class="line"><a name="l01221"></a><span class="lineno"> 1221</span>&#160;</div><div class="line"><a name="l01222"></a><span class="lineno"> 1222</span>&#160;        <span class="comment">// Accumulate</span></div><div class="line"><a name="l01223"></a><span class="lineno"> 1223</span>&#160;        acc0 += b0 * (VECTOR_TYPE)a0.s0;</div><div class="line"><a name="l01224"></a><span class="lineno"> 1224</span>&#160;        acc0 += b1 * (VECTOR_TYPE)a0.s1;</div><div class="line"><a name="l01225"></a><span class="lineno"> 1225</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01226"></a><span class="lineno"> 1226</span>&#160;        acc1 += b0 * (VECTOR_TYPE)a1.s0;</div><div class="line"><a name="l01227"></a><span class="lineno"> 1227</span>&#160;        acc1 += b1 * (VECTOR_TYPE)a1.s1;</div><div class="line"><a name="l01228"></a><span class="lineno"> 1228</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01229"></a><span class="lineno"> 1229</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01230"></a><span class="lineno"> 1230</span>&#160;        acc2 += b0 * (VECTOR_TYPE)a2.s0;</div><div class="line"><a name="l01231"></a><span class="lineno"> 1231</span>&#160;        acc2 += b1 * (VECTOR_TYPE)a2.s1;</div><div class="line"><a name="l01232"></a><span class="lineno"> 1232</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01233"></a><span class="lineno"> 1233</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01234"></a><span class="lineno"> 1234</span>&#160;        acc3 += b0 * (VECTOR_TYPE)a3.s0;</div><div class="line"><a name="l01235"></a><span class="lineno"> 1235</span>&#160;        acc3 += b1 * (VECTOR_TYPE)a3.s1;</div><div class="line"><a name="l01236"></a><span class="lineno"> 1236</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></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;</div><div class="line"><a name="l01239"></a><span class="lineno"> 1239</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt; end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(DATA_TYPE), src1_stride_y))</div><div class="line"><a name="l01240"></a><span class="lineno"> 1240</span>&#160;    {</div><div class="line"><a name="l01241"></a><span class="lineno"> 1241</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01242"></a><span class="lineno"> 1242</span>&#160;        DATA_TYPE a0 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01243"></a><span class="lineno"> 1243</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01244"></a><span class="lineno"> 1244</span>&#160;        DATA_TYPE a1 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01245"></a><span class="lineno"> 1245</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01246"></a><span class="lineno"> 1246</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01247"></a><span class="lineno"> 1247</span>&#160;        DATA_TYPE a2 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01248"></a><span class="lineno"> 1248</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01249"></a><span class="lineno"> 1249</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01250"></a><span class="lineno"> 1250</span>&#160;        DATA_TYPE a3 = *((__global DATA_TYPE *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01251"></a><span class="lineno"> 1251</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01252"></a><span class="lineno"> 1252</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01253"></a><span class="lineno"> 1253</span>&#160;        VECTOR_TYPE b0 = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, (__global DATA_TYPE *)(src1_ptr + src_addr.s1));</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="comment">// Accumulate</span></div><div class="line"><a name="l01256"></a><span class="lineno"> 1256</span>&#160;        acc0 += b0 * (VECTOR_TYPE)a0;</div><div class="line"><a name="l01257"></a><span class="lineno"> 1257</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01258"></a><span class="lineno"> 1258</span>&#160;        acc1 += b0 * (VECTOR_TYPE)a1;</div><div class="line"><a name="l01259"></a><span class="lineno"> 1259</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01260"></a><span class="lineno"> 1260</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01261"></a><span class="lineno"> 1261</span>&#160;        acc2 += b0 * (VECTOR_TYPE)a2;</div><div class="line"><a name="l01262"></a><span class="lineno"> 1262</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01263"></a><span class="lineno"> 1263</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01264"></a><span class="lineno"> 1264</span>&#160;        acc3 += b0 * (VECTOR_TYPE)a3;</div><div class="line"><a name="l01265"></a><span class="lineno"> 1265</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01266"></a><span class="lineno"> 1266</span>&#160;    }</div><div class="line"><a name="l01267"></a><span class="lineno"> 1267</span>&#160;</div><div class="line"><a name="l01268"></a><span class="lineno"> 1268</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01269"></a><span class="lineno"> 1269</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01270"></a><span class="lineno"> 1270</span>&#160;</div><div class="line"><a name="l01271"></a><span class="lineno"> 1271</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01272"></a><span class="lineno"> 1272</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l01273"></a><span class="lineno"> 1273</span>&#160;</div><div class="line"><a name="l01274"></a><span class="lineno"> 1274</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01275"></a><span class="lineno"> 1275</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l01276"></a><span class="lineno"> 1276</span>&#160;</div><div class="line"><a name="l01277"></a><span class="lineno"> 1277</span>&#160;    <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l01278"></a><span class="lineno"> 1278</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01279"></a><span class="lineno"> 1279</span>&#160;    acc0 = acc0 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01280"></a><span class="lineno"> 1280</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01281"></a><span class="lineno"> 1281</span>&#160;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01282"></a><span class="lineno"> 1282</span>&#160;    (acc0, 0, (__global DATA_TYPE *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01283"></a><span class="lineno"> 1283</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01284"></a><span class="lineno"> 1284</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01285"></a><span class="lineno"> 1285</span>&#160;    acc1 = acc1 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01286"></a><span class="lineno"> 1286</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01287"></a><span class="lineno"> 1287</span>&#160;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01288"></a><span class="lineno"> 1288</span>&#160;    (acc1, 0, (__global DATA_TYPE *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01289"></a><span class="lineno"> 1289</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01290"></a><span class="lineno"> 1290</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01291"></a><span class="lineno"> 1291</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01292"></a><span class="lineno"> 1292</span>&#160;    acc2 = acc2 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01293"></a><span class="lineno"> 1293</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01294"></a><span class="lineno"> 1294</span>&#160;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01295"></a><span class="lineno"> 1295</span>&#160;    (acc2, 0, (__global DATA_TYPE *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01296"></a><span class="lineno"> 1296</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01297"></a><span class="lineno"> 1297</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01298"></a><span class="lineno"> 1298</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01299"></a><span class="lineno"> 1299</span>&#160;    acc3 = acc3 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01300"></a><span class="lineno"> 1300</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01301"></a><span class="lineno"> 1301</span>&#160;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01302"></a><span class="lineno"> 1302</span>&#160;    (acc3, 0, (__global DATA_TYPE *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01303"></a><span class="lineno"> 1303</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01304"></a><span class="lineno"> 1304</span>&#160;}</div><div class="line"><a name="l01305"></a><span class="lineno"> 1305</span>&#160;<span class="preprocessor">#endif // defined(DATA_TYPE)</span></div><div class="line"><a name="l01306"></a><span class="lineno"> 1306</span>&#160;</div><div class="line"><a name="l01336"></a><span class="lineno"> 1336</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f32_bifrost(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01337"></a><span class="lineno"> 1337</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01338"></a><span class="lineno"> 1338</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01339"></a><span class="lineno"> 1339</span>&#160;                                                 uint src0_stride_z,</div><div class="line"><a name="l01340"></a><span class="lineno"> 1340</span>&#160;                                                 uint src1_stride_z,</div><div class="line"><a name="l01341"></a><span class="lineno"> 1341</span>&#160;                                                 uint dst_stride_z)</div><div class="line"><a name="l01342"></a><span class="lineno"> 1342</span>&#160;{</div><div class="line"><a name="l01343"></a><span class="lineno"> 1343</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01344"></a><span class="lineno"> 1344</span>&#160;</div><div class="line"><a name="l01345"></a><span class="lineno"> 1345</span>&#160;    <span class="comment">// Compute starting address for matrix A and matrix B</span></div><div class="line"><a name="l01346"></a><span class="lineno"> 1346</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01347"></a><span class="lineno"> 1347</span>&#160;</div><div class="line"><a name="l01348"></a><span class="lineno"> 1348</span>&#160;    <span class="comment">// Update address for matrix A</span></div><div class="line"><a name="l01349"></a><span class="lineno"> 1349</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01350"></a><span class="lineno"> 1350</span>&#160;</div><div class="line"><a name="l01351"></a><span class="lineno"> 1351</span>&#160;    <span class="comment">// Update address for matrix B</span></div><div class="line"><a name="l01352"></a><span class="lineno"> 1352</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01353"></a><span class="lineno"> 1353</span>&#160;</div><div class="line"><a name="l01354"></a><span class="lineno"> 1354</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01355"></a><span class="lineno"> 1355</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01356"></a><span class="lineno"> 1356</span>&#160;</div><div class="line"><a name="l01357"></a><span class="lineno"> 1357</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01358"></a><span class="lineno"> 1358</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01359"></a><span class="lineno"> 1359</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01360"></a><span class="lineno"> 1360</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01361"></a><span class="lineno"> 1361</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01362"></a><span class="lineno"> 1362</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01363"></a><span class="lineno"> 1363</span>&#160;</div><div class="line"><a name="l01364"></a><span class="lineno"> 1364</span>&#160;    <span class="comment">// Initialize accumulators</span></div><div class="line"><a name="l01365"></a><span class="lineno"> 1365</span>&#160;    <span class="keywordtype">float</span> acc00 = 0.0f;</div><div class="line"><a name="l01366"></a><span class="lineno"> 1366</span>&#160;    <span class="keywordtype">float</span> acc01 = 0.0f;</div><div class="line"><a name="l01367"></a><span class="lineno"> 1367</span>&#160;    <span class="keywordtype">float</span> acc02 = 0.0f;</div><div class="line"><a name="l01368"></a><span class="lineno"> 1368</span>&#160;    <span class="keywordtype">float</span> acc03 = 0.0f;</div><div class="line"><a name="l01369"></a><span class="lineno"> 1369</span>&#160;</div><div class="line"><a name="l01370"></a><span class="lineno"> 1370</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01371"></a><span class="lineno"> 1371</span>&#160;    <span class="keywordtype">float</span> acc10 = 0.0f;</div><div class="line"><a name="l01372"></a><span class="lineno"> 1372</span>&#160;    <span class="keywordtype">float</span> acc11 = 0.0f;</div><div class="line"><a name="l01373"></a><span class="lineno"> 1373</span>&#160;    <span class="keywordtype">float</span> acc12 = 0.0f;</div><div class="line"><a name="l01374"></a><span class="lineno"> 1374</span>&#160;    <span class="keywordtype">float</span> acc13 = 0.0f;</div><div class="line"><a name="l01375"></a><span class="lineno"> 1375</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01376"></a><span class="lineno"> 1376</span>&#160;</div><div class="line"><a name="l01377"></a><span class="lineno"> 1377</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01378"></a><span class="lineno"> 1378</span>&#160;    <span class="keywordtype">float</span> acc20 = 0.0f;</div><div class="line"><a name="l01379"></a><span class="lineno"> 1379</span>&#160;    <span class="keywordtype">float</span> acc21 = 0.0f;</div><div class="line"><a name="l01380"></a><span class="lineno"> 1380</span>&#160;    <span class="keywordtype">float</span> acc22 = 0.0f;</div><div class="line"><a name="l01381"></a><span class="lineno"> 1381</span>&#160;    <span class="keywordtype">float</span> acc23 = 0.0f;</div><div class="line"><a name="l01382"></a><span class="lineno"> 1382</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01383"></a><span class="lineno"> 1383</span>&#160;</div><div class="line"><a name="l01384"></a><span class="lineno"> 1384</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01385"></a><span class="lineno"> 1385</span>&#160;    <span class="keywordtype">float</span> acc30 = 0.0f;</div><div class="line"><a name="l01386"></a><span class="lineno"> 1386</span>&#160;    <span class="keywordtype">float</span> acc31 = 0.0f;</div><div class="line"><a name="l01387"></a><span class="lineno"> 1387</span>&#160;    <span class="keywordtype">float</span> acc32 = 0.0f;</div><div class="line"><a name="l01388"></a><span class="lineno"> 1388</span>&#160;    <span class="keywordtype">float</span> acc33 = 0.0f;</div><div class="line"><a name="l01389"></a><span class="lineno"> 1389</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01390"></a><span class="lineno"> 1390</span>&#160;</div><div class="line"><a name="l01391"></a><span class="lineno"> 1391</span>&#160;    <span class="comment">// A and B src indices get incremented at the same time.</span></div><div class="line"><a name="l01392"></a><span class="lineno"> 1392</span>&#160;    <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01393"></a><span class="lineno"> 1393</span>&#160;    <span class="keywordflow">for</span>(; i &lt;= ((int)COLS_A - 4); i += 4)</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;        <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01396"></a><span class="lineno"> 1396</span>&#160;        float4 a0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01397"></a><span class="lineno"> 1397</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01398"></a><span class="lineno"> 1398</span>&#160;        float4 a1 = vload4(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01399"></a><span class="lineno"> 1399</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01400"></a><span class="lineno"> 1400</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01401"></a><span class="lineno"> 1401</span>&#160;        float4 a2 = vload4(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01402"></a><span class="lineno"> 1402</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01403"></a><span class="lineno"> 1403</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01404"></a><span class="lineno"> 1404</span>&#160;        float4 a3 = vload4(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01405"></a><span class="lineno"> 1405</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01406"></a><span class="lineno"> 1406</span>&#160;        float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01407"></a><span class="lineno"> 1407</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01408"></a><span class="lineno"> 1408</span>&#160;</div><div class="line"><a name="l01409"></a><span class="lineno"> 1409</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01410"></a><span class="lineno"> 1410</span>&#160;        acc00 = fma(a0.s0, b0.s0, acc00);</div><div class="line"><a name="l01411"></a><span class="lineno"> 1411</span>&#160;        acc01 = fma(a0.s0, b0.s1, acc01);</div><div class="line"><a name="l01412"></a><span class="lineno"> 1412</span>&#160;        acc02 = fma(a0.s0, b0.s2, acc02);</div><div class="line"><a name="l01413"></a><span class="lineno"> 1413</span>&#160;        acc03 = fma(a0.s0, b0.s3, acc03);</div><div class="line"><a name="l01414"></a><span class="lineno"> 1414</span>&#160;</div><div class="line"><a name="l01415"></a><span class="lineno"> 1415</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01416"></a><span class="lineno"> 1416</span>&#160;</div><div class="line"><a name="l01417"></a><span class="lineno"> 1417</span>&#160;        acc10 = fma(a1.s0, b0.s0, acc10);</div><div class="line"><a name="l01418"></a><span class="lineno"> 1418</span>&#160;        acc11 = fma(a1.s0, b0.s1, acc11);</div><div class="line"><a name="l01419"></a><span class="lineno"> 1419</span>&#160;        acc12 = fma(a1.s0, b0.s2, acc12);</div><div class="line"><a name="l01420"></a><span class="lineno"> 1420</span>&#160;        acc13 = fma(a1.s0, b0.s3, acc13);</div><div class="line"><a name="l01421"></a><span class="lineno"> 1421</span>&#160;</div><div class="line"><a name="l01422"></a><span class="lineno"> 1422</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01423"></a><span class="lineno"> 1423</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01424"></a><span class="lineno"> 1424</span>&#160;</div><div class="line"><a name="l01425"></a><span class="lineno"> 1425</span>&#160;        acc20 = fma(a2.s0, b0.s0, acc20);</div><div class="line"><a name="l01426"></a><span class="lineno"> 1426</span>&#160;        acc21 = fma(a2.s0, b0.s1, acc21);</div><div class="line"><a name="l01427"></a><span class="lineno"> 1427</span>&#160;        acc22 = fma(a2.s0, b0.s2, acc22);</div><div class="line"><a name="l01428"></a><span class="lineno"> 1428</span>&#160;        acc23 = fma(a2.s0, b0.s3, acc23);</div><div class="line"><a name="l01429"></a><span class="lineno"> 1429</span>&#160;</div><div class="line"><a name="l01430"></a><span class="lineno"> 1430</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01431"></a><span class="lineno"> 1431</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01432"></a><span class="lineno"> 1432</span>&#160;</div><div class="line"><a name="l01433"></a><span class="lineno"> 1433</span>&#160;        acc30 = fma(a3.s0, b0.s0, acc30);</div><div class="line"><a name="l01434"></a><span class="lineno"> 1434</span>&#160;        acc31 = fma(a3.s0, b0.s1, acc31);</div><div class="line"><a name="l01435"></a><span class="lineno"> 1435</span>&#160;        acc32 = fma(a3.s0, b0.s2, acc32);</div><div class="line"><a name="l01436"></a><span class="lineno"> 1436</span>&#160;        acc33 = fma(a3.s0, b0.s3, acc33);</div><div class="line"><a name="l01437"></a><span class="lineno"> 1437</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01438"></a><span class="lineno"> 1438</span>&#160;</div><div class="line"><a name="l01439"></a><span class="lineno"> 1439</span>&#160;        <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01440"></a><span class="lineno"> 1440</span>&#160;        b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01441"></a><span class="lineno"> 1441</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01442"></a><span class="lineno"> 1442</span>&#160;</div><div class="line"><a name="l01443"></a><span class="lineno"> 1443</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01444"></a><span class="lineno"> 1444</span>&#160;        acc00 = fma(a0.s1, b0.s0, acc00);</div><div class="line"><a name="l01445"></a><span class="lineno"> 1445</span>&#160;        acc01 = fma(a0.s1, b0.s1, acc01);</div><div class="line"><a name="l01446"></a><span class="lineno"> 1446</span>&#160;        acc02 = fma(a0.s1, b0.s2, acc02);</div><div class="line"><a name="l01447"></a><span class="lineno"> 1447</span>&#160;        acc03 = fma(a0.s1, b0.s3, acc03);</div><div class="line"><a name="l01448"></a><span class="lineno"> 1448</span>&#160;</div><div class="line"><a name="l01449"></a><span class="lineno"> 1449</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01450"></a><span class="lineno"> 1450</span>&#160;</div><div class="line"><a name="l01451"></a><span class="lineno"> 1451</span>&#160;        acc10 = fma(a1.s1, b0.s0, acc10);</div><div class="line"><a name="l01452"></a><span class="lineno"> 1452</span>&#160;        acc11 = fma(a1.s1, b0.s1, acc11);</div><div class="line"><a name="l01453"></a><span class="lineno"> 1453</span>&#160;        acc12 = fma(a1.s1, b0.s2, acc12);</div><div class="line"><a name="l01454"></a><span class="lineno"> 1454</span>&#160;        acc13 = fma(a1.s1, b0.s3, acc13);</div><div class="line"><a name="l01455"></a><span class="lineno"> 1455</span>&#160;</div><div class="line"><a name="l01456"></a><span class="lineno"> 1456</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01457"></a><span class="lineno"> 1457</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01458"></a><span class="lineno"> 1458</span>&#160;</div><div class="line"><a name="l01459"></a><span class="lineno"> 1459</span>&#160;        acc20 = fma(a2.s1, b0.s0, acc20);</div><div class="line"><a name="l01460"></a><span class="lineno"> 1460</span>&#160;        acc21 = fma(a2.s1, b0.s1, acc21);</div><div class="line"><a name="l01461"></a><span class="lineno"> 1461</span>&#160;        acc22 = fma(a2.s1, b0.s2, acc22);</div><div class="line"><a name="l01462"></a><span class="lineno"> 1462</span>&#160;        acc23 = fma(a2.s1, b0.s3, acc23);</div><div class="line"><a name="l01463"></a><span class="lineno"> 1463</span>&#160;</div><div class="line"><a name="l01464"></a><span class="lineno"> 1464</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01465"></a><span class="lineno"> 1465</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01466"></a><span class="lineno"> 1466</span>&#160;</div><div class="line"><a name="l01467"></a><span class="lineno"> 1467</span>&#160;        acc30 = fma(a3.s1, b0.s0, acc30);</div><div class="line"><a name="l01468"></a><span class="lineno"> 1468</span>&#160;        acc31 = fma(a3.s1, b0.s1, acc31);</div><div class="line"><a name="l01469"></a><span class="lineno"> 1469</span>&#160;        acc32 = fma(a3.s1, b0.s2, acc32);</div><div class="line"><a name="l01470"></a><span class="lineno"> 1470</span>&#160;        acc33 = fma(a3.s1, b0.s3, acc33);</div><div class="line"><a name="l01471"></a><span class="lineno"> 1471</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01472"></a><span class="lineno"> 1472</span>&#160;</div><div class="line"><a name="l01473"></a><span class="lineno"> 1473</span>&#160;        <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01474"></a><span class="lineno"> 1474</span>&#160;        b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01475"></a><span class="lineno"> 1475</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01476"></a><span class="lineno"> 1476</span>&#160;</div><div class="line"><a name="l01477"></a><span class="lineno"> 1477</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01478"></a><span class="lineno"> 1478</span>&#160;        acc00 = fma(a0.s2, b0.s0, acc00);</div><div class="line"><a name="l01479"></a><span class="lineno"> 1479</span>&#160;        acc01 = fma(a0.s2, b0.s1, acc01);</div><div class="line"><a name="l01480"></a><span class="lineno"> 1480</span>&#160;        acc02 = fma(a0.s2, b0.s2, acc02);</div><div class="line"><a name="l01481"></a><span class="lineno"> 1481</span>&#160;        acc03 = fma(a0.s2, b0.s3, acc03);</div><div class="line"><a name="l01482"></a><span class="lineno"> 1482</span>&#160;</div><div class="line"><a name="l01483"></a><span class="lineno"> 1483</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01484"></a><span class="lineno"> 1484</span>&#160;</div><div class="line"><a name="l01485"></a><span class="lineno"> 1485</span>&#160;        acc10 = fma(a1.s2, b0.s0, acc10);</div><div class="line"><a name="l01486"></a><span class="lineno"> 1486</span>&#160;        acc11 = fma(a1.s2, b0.s1, acc11);</div><div class="line"><a name="l01487"></a><span class="lineno"> 1487</span>&#160;        acc12 = fma(a1.s2, b0.s2, acc12);</div><div class="line"><a name="l01488"></a><span class="lineno"> 1488</span>&#160;        acc13 = fma(a1.s2, b0.s3, acc13);</div><div class="line"><a name="l01489"></a><span class="lineno"> 1489</span>&#160;</div><div class="line"><a name="l01490"></a><span class="lineno"> 1490</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01491"></a><span class="lineno"> 1491</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01492"></a><span class="lineno"> 1492</span>&#160;</div><div class="line"><a name="l01493"></a><span class="lineno"> 1493</span>&#160;        acc20 = fma(a2.s2, b0.s0, acc20);</div><div class="line"><a name="l01494"></a><span class="lineno"> 1494</span>&#160;        acc21 = fma(a2.s2, b0.s1, acc21);</div><div class="line"><a name="l01495"></a><span class="lineno"> 1495</span>&#160;        acc22 = fma(a2.s2, b0.s2, acc22);</div><div class="line"><a name="l01496"></a><span class="lineno"> 1496</span>&#160;        acc23 = fma(a2.s2, b0.s3, acc23);</div><div class="line"><a name="l01497"></a><span class="lineno"> 1497</span>&#160;</div><div class="line"><a name="l01498"></a><span class="lineno"> 1498</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01499"></a><span class="lineno"> 1499</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01500"></a><span class="lineno"> 1500</span>&#160;</div><div class="line"><a name="l01501"></a><span class="lineno"> 1501</span>&#160;        acc30 = fma(a3.s2, b0.s0, acc30);</div><div class="line"><a name="l01502"></a><span class="lineno"> 1502</span>&#160;        acc31 = fma(a3.s2, b0.s1, acc31);</div><div class="line"><a name="l01503"></a><span class="lineno"> 1503</span>&#160;        acc32 = fma(a3.s2, b0.s2, acc32);</div><div class="line"><a name="l01504"></a><span class="lineno"> 1504</span>&#160;        acc33 = fma(a3.s2, b0.s3, acc33);</div><div class="line"><a name="l01505"></a><span class="lineno"> 1505</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01506"></a><span class="lineno"> 1506</span>&#160;</div><div class="line"><a name="l01507"></a><span class="lineno"> 1507</span>&#160;        <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01508"></a><span class="lineno"> 1508</span>&#160;        b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01509"></a><span class="lineno"> 1509</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01510"></a><span class="lineno"> 1510</span>&#160;</div><div class="line"><a name="l01511"></a><span class="lineno"> 1511</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01512"></a><span class="lineno"> 1512</span>&#160;        acc00 = fma(a0.s3, b0.s0, acc00);</div><div class="line"><a name="l01513"></a><span class="lineno"> 1513</span>&#160;        acc01 = fma(a0.s3, b0.s1, acc01);</div><div class="line"><a name="l01514"></a><span class="lineno"> 1514</span>&#160;        acc02 = fma(a0.s3, b0.s2, acc02);</div><div class="line"><a name="l01515"></a><span class="lineno"> 1515</span>&#160;        acc03 = fma(a0.s3, b0.s3, acc03);</div><div class="line"><a name="l01516"></a><span class="lineno"> 1516</span>&#160;</div><div class="line"><a name="l01517"></a><span class="lineno"> 1517</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01518"></a><span class="lineno"> 1518</span>&#160;</div><div class="line"><a name="l01519"></a><span class="lineno"> 1519</span>&#160;        acc10 = fma(a1.s3, b0.s0, acc10);</div><div class="line"><a name="l01520"></a><span class="lineno"> 1520</span>&#160;        acc11 = fma(a1.s3, b0.s1, acc11);</div><div class="line"><a name="l01521"></a><span class="lineno"> 1521</span>&#160;        acc12 = fma(a1.s3, b0.s2, acc12);</div><div class="line"><a name="l01522"></a><span class="lineno"> 1522</span>&#160;        acc13 = fma(a1.s3, b0.s3, acc13);</div><div class="line"><a name="l01523"></a><span class="lineno"> 1523</span>&#160;</div><div class="line"><a name="l01524"></a><span class="lineno"> 1524</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01525"></a><span class="lineno"> 1525</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01526"></a><span class="lineno"> 1526</span>&#160;</div><div class="line"><a name="l01527"></a><span class="lineno"> 1527</span>&#160;        acc20 = fma(a2.s3, b0.s0, acc20);</div><div class="line"><a name="l01528"></a><span class="lineno"> 1528</span>&#160;        acc21 = fma(a2.s3, b0.s1, acc21);</div><div class="line"><a name="l01529"></a><span class="lineno"> 1529</span>&#160;        acc22 = fma(a2.s3, b0.s2, acc22);</div><div class="line"><a name="l01530"></a><span class="lineno"> 1530</span>&#160;        acc23 = fma(a2.s3, b0.s3, acc23);</div><div class="line"><a name="l01531"></a><span class="lineno"> 1531</span>&#160;</div><div class="line"><a name="l01532"></a><span class="lineno"> 1532</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01533"></a><span class="lineno"> 1533</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01534"></a><span class="lineno"> 1534</span>&#160;</div><div class="line"><a name="l01535"></a><span class="lineno"> 1535</span>&#160;        acc30 = fma(a3.s3, b0.s0, acc30);</div><div class="line"><a name="l01536"></a><span class="lineno"> 1536</span>&#160;        acc31 = fma(a3.s3, b0.s1, acc31);</div><div class="line"><a name="l01537"></a><span class="lineno"> 1537</span>&#160;        acc32 = fma(a3.s3, b0.s2, acc32);</div><div class="line"><a name="l01538"></a><span class="lineno"> 1538</span>&#160;        acc33 = fma(a3.s3, b0.s3, acc33);</div><div class="line"><a name="l01539"></a><span class="lineno"> 1539</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01540"></a><span class="lineno"> 1540</span>&#160;</div><div class="line"><a name="l01541"></a><span class="lineno"> 1541</span>&#160;        src_addr.s0 += 4 * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01542"></a><span class="lineno"> 1542</span>&#160;    }</div><div class="line"><a name="l01543"></a><span class="lineno"> 1543</span>&#160;</div><div class="line"><a name="l01544"></a><span class="lineno"> 1544</span>&#160;    <span class="keywordflow">for</span>(; i &lt; (int)COLS_A; ++i)</div><div class="line"><a name="l01545"></a><span class="lineno"> 1545</span>&#160;    {</div><div class="line"><a name="l01546"></a><span class="lineno"> 1546</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01547"></a><span class="lineno"> 1547</span>&#160;        <span class="keywordtype">float</span> a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l01548"></a><span class="lineno"> 1548</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01549"></a><span class="lineno"> 1549</span>&#160;        <span class="keywordtype">float</span> a1 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01550"></a><span class="lineno"> 1550</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01551"></a><span class="lineno"> 1551</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01552"></a><span class="lineno"> 1552</span>&#160;        <span class="keywordtype">float</span> a2 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01553"></a><span class="lineno"> 1553</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01554"></a><span class="lineno"> 1554</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01555"></a><span class="lineno"> 1555</span>&#160;        <span class="keywordtype">float</span> a3 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01556"></a><span class="lineno"> 1556</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01557"></a><span class="lineno"> 1557</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01558"></a><span class="lineno"> 1558</span>&#160;        float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01559"></a><span class="lineno"> 1559</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01560"></a><span class="lineno"> 1560</span>&#160;</div><div class="line"><a name="l01561"></a><span class="lineno"> 1561</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01562"></a><span class="lineno"> 1562</span>&#160;        acc00 = fma(a0, b0.s0, acc00);</div><div class="line"><a name="l01563"></a><span class="lineno"> 1563</span>&#160;        acc01 = fma(a0, b0.s1, acc01);</div><div class="line"><a name="l01564"></a><span class="lineno"> 1564</span>&#160;        acc02 = fma(a0, b0.s2, acc02);</div><div class="line"><a name="l01565"></a><span class="lineno"> 1565</span>&#160;        acc03 = fma(a0, b0.s3, acc03);</div><div class="line"><a name="l01566"></a><span class="lineno"> 1566</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01567"></a><span class="lineno"> 1567</span>&#160;        acc10 = fma(a1, b0.s0, acc10);</div><div class="line"><a name="l01568"></a><span class="lineno"> 1568</span>&#160;        acc11 = fma(a1, b0.s1, acc11);</div><div class="line"><a name="l01569"></a><span class="lineno"> 1569</span>&#160;        acc12 = fma(a1, b0.s2, acc12);</div><div class="line"><a name="l01570"></a><span class="lineno"> 1570</span>&#160;        acc13 = fma(a1, b0.s3, acc13);</div><div class="line"><a name="l01571"></a><span class="lineno"> 1571</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01572"></a><span class="lineno"> 1572</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01573"></a><span class="lineno"> 1573</span>&#160;        acc20 = fma(a2, b0.s0, acc20);</div><div class="line"><a name="l01574"></a><span class="lineno"> 1574</span>&#160;        acc21 = fma(a2, b0.s1, acc21);</div><div class="line"><a name="l01575"></a><span class="lineno"> 1575</span>&#160;        acc22 = fma(a2, b0.s2, acc22);</div><div class="line"><a name="l01576"></a><span class="lineno"> 1576</span>&#160;        acc23 = fma(a2, b0.s3, acc23);</div><div class="line"><a name="l01577"></a><span class="lineno"> 1577</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01578"></a><span class="lineno"> 1578</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01579"></a><span class="lineno"> 1579</span>&#160;        acc30 = fma(a3, b0.s0, acc30);</div><div class="line"><a name="l01580"></a><span class="lineno"> 1580</span>&#160;        acc31 = fma(a3, b0.s1, acc31);</div><div class="line"><a name="l01581"></a><span class="lineno"> 1581</span>&#160;        acc32 = fma(a3, b0.s2, acc32);</div><div class="line"><a name="l01582"></a><span class="lineno"> 1582</span>&#160;        acc33 = fma(a3, b0.s3, acc33);</div><div class="line"><a name="l01583"></a><span class="lineno"> 1583</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01584"></a><span class="lineno"> 1584</span>&#160;</div><div class="line"><a name="l01585"></a><span class="lineno"> 1585</span>&#160;        src_addr.s0 += <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01586"></a><span class="lineno"> 1586</span>&#160;    }</div><div class="line"><a name="l01587"></a><span class="lineno"> 1587</span>&#160;</div><div class="line"><a name="l01588"></a><span class="lineno"> 1588</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01589"></a><span class="lineno"> 1589</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01590"></a><span class="lineno"> 1590</span>&#160;</div><div class="line"><a name="l01591"></a><span class="lineno"> 1591</span>&#160;    <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l01592"></a><span class="lineno"> 1592</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01593"></a><span class="lineno"> 1593</span>&#160;    acc00 = acc00 * ALPHA;</div><div class="line"><a name="l01594"></a><span class="lineno"> 1594</span>&#160;    acc01 = acc01 * ALPHA;</div><div class="line"><a name="l01595"></a><span class="lineno"> 1595</span>&#160;    acc02 = acc02 * ALPHA;</div><div class="line"><a name="l01596"></a><span class="lineno"> 1596</span>&#160;    acc03 = acc03 * ALPHA;</div><div class="line"><a name="l01597"></a><span class="lineno"> 1597</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01598"></a><span class="lineno"> 1598</span>&#160;</div><div class="line"><a name="l01599"></a><span class="lineno"> 1599</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01600"></a><span class="lineno"> 1600</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l01601"></a><span class="lineno"> 1601</span>&#160;</div><div class="line"><a name="l01602"></a><span class="lineno"> 1602</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01603"></a><span class="lineno"> 1603</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l01604"></a><span class="lineno"> 1604</span>&#160;</div><div class="line"><a name="l01605"></a><span class="lineno"> 1605</span>&#160;    float4 acc0 = ((float4)(acc00, acc01, acc02, acc03));</div><div class="line"><a name="l01606"></a><span class="lineno"> 1606</span>&#160;    vstore4(acc0, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01607"></a><span class="lineno"> 1607</span>&#160;</div><div class="line"><a name="l01608"></a><span class="lineno"> 1608</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01609"></a><span class="lineno"> 1609</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01610"></a><span class="lineno"> 1610</span>&#160;    acc10 = acc10 * ALPHA;</div><div class="line"><a name="l01611"></a><span class="lineno"> 1611</span>&#160;    acc11 = acc11 * ALPHA;</div><div class="line"><a name="l01612"></a><span class="lineno"> 1612</span>&#160;    acc12 = acc12 * ALPHA;</div><div class="line"><a name="l01613"></a><span class="lineno"> 1613</span>&#160;    acc13 = acc13 * ALPHA;</div><div class="line"><a name="l01614"></a><span class="lineno"> 1614</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01615"></a><span class="lineno"> 1615</span>&#160;    float4 acc1 = ((float4)(acc10, acc11, acc12, acc13));</div><div class="line"><a name="l01616"></a><span class="lineno"> 1616</span>&#160;    vstore4(acc1, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01617"></a><span class="lineno"> 1617</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01618"></a><span class="lineno"> 1618</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01619"></a><span class="lineno"> 1619</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01620"></a><span class="lineno"> 1620</span>&#160;    acc20 = acc20 * ALPHA;</div><div class="line"><a name="l01621"></a><span class="lineno"> 1621</span>&#160;    acc21 = acc21 * ALPHA;</div><div class="line"><a name="l01622"></a><span class="lineno"> 1622</span>&#160;    acc22 = acc22 * ALPHA;</div><div class="line"><a name="l01623"></a><span class="lineno"> 1623</span>&#160;    acc23 = acc23 * ALPHA;</div><div class="line"><a name="l01624"></a><span class="lineno"> 1624</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01625"></a><span class="lineno"> 1625</span>&#160;    float4 acc2 = ((float4)(acc20, acc21, acc22, acc23));</div><div class="line"><a name="l01626"></a><span class="lineno"> 1626</span>&#160;    vstore4(acc2, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01627"></a><span class="lineno"> 1627</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01628"></a><span class="lineno"> 1628</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01629"></a><span class="lineno"> 1629</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01630"></a><span class="lineno"> 1630</span>&#160;    acc30 = acc30 * ALPHA;</div><div class="line"><a name="l01631"></a><span class="lineno"> 1631</span>&#160;    acc31 = acc31 * ALPHA;</div><div class="line"><a name="l01632"></a><span class="lineno"> 1632</span>&#160;    acc32 = acc32 * ALPHA;</div><div class="line"><a name="l01633"></a><span class="lineno"> 1633</span>&#160;    acc33 = acc33 * ALPHA;</div><div class="line"><a name="l01634"></a><span class="lineno"> 1634</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01635"></a><span class="lineno"> 1635</span>&#160;    float4 acc3 = ((float4)(acc30, acc31, acc32, acc33));</div><div class="line"><a name="l01636"></a><span class="lineno"> 1636</span>&#160;    vstore4(acc3, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01637"></a><span class="lineno"> 1637</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01638"></a><span class="lineno"> 1638</span>&#160;}</div><div class="line"><a name="l01639"></a><span class="lineno"> 1639</span>&#160;</div><div class="line"><a name="l01670"></a><span class="lineno"> 1670</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f32_bifrost_1000(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01671"></a><span class="lineno"> 1671</span>&#160;                                                      <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01672"></a><span class="lineno"> 1672</span>&#160;                                                      <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01673"></a><span class="lineno"> 1673</span>&#160;                                                      uint src0_stride_z,</div><div class="line"><a name="l01674"></a><span class="lineno"> 1674</span>&#160;                                                      uint src1_stride_z,</div><div class="line"><a name="l01675"></a><span class="lineno"> 1675</span>&#160;                                                      uint dst_stride_z)</div><div class="line"><a name="l01676"></a><span class="lineno"> 1676</span>&#160;{</div><div class="line"><a name="l01677"></a><span class="lineno"> 1677</span>&#160;    <span class="comment">// Requires 2 NUM_ELEMS_PROCESSED_PER_THREAD_X, C vect2, A vect4, B (2 vload2) // to fix for NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01678"></a><span class="lineno"> 1678</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01679"></a><span class="lineno"> 1679</span>&#160;</div><div class="line"><a name="l01680"></a><span class="lineno"> 1680</span>&#160;    <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01681"></a><span class="lineno"> 1681</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01682"></a><span class="lineno"> 1682</span>&#160;</div><div class="line"><a name="l01683"></a><span class="lineno"> 1683</span>&#160;    <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01684"></a><span class="lineno"> 1684</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01685"></a><span class="lineno"> 1685</span>&#160;</div><div class="line"><a name="l01686"></a><span class="lineno"> 1686</span>&#160;    <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01687"></a><span class="lineno"> 1687</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01688"></a><span class="lineno"> 1688</span>&#160;</div><div class="line"><a name="l01689"></a><span class="lineno"> 1689</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01690"></a><span class="lineno"> 1690</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01691"></a><span class="lineno"> 1691</span>&#160;</div><div class="line"><a name="l01692"></a><span class="lineno"> 1692</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01693"></a><span class="lineno"> 1693</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01694"></a><span class="lineno"> 1694</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01695"></a><span class="lineno"> 1695</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01696"></a><span class="lineno"> 1696</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01697"></a><span class="lineno"> 1697</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01698"></a><span class="lineno"> 1698</span>&#160;</div><div class="line"><a name="l01699"></a><span class="lineno"> 1699</span>&#160;    <span class="comment">// Initialize accumulators</span></div><div class="line"><a name="l01700"></a><span class="lineno"> 1700</span>&#160;    <span class="keywordtype">float</span> acc00 = 0.0f;</div><div class="line"><a name="l01701"></a><span class="lineno"> 1701</span>&#160;    <span class="keywordtype">float</span> acc01 = 0.0f;</div><div class="line"><a name="l01702"></a><span class="lineno"> 1702</span>&#160;</div><div class="line"><a name="l01703"></a><span class="lineno"> 1703</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01704"></a><span class="lineno"> 1704</span>&#160;    <span class="keywordtype">float</span> acc10 = 0.0f;</div><div class="line"><a name="l01705"></a><span class="lineno"> 1705</span>&#160;    <span class="keywordtype">float</span> acc11 = 0.0f;</div><div class="line"><a name="l01706"></a><span class="lineno"> 1706</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01707"></a><span class="lineno"> 1707</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01708"></a><span class="lineno"> 1708</span>&#160;    <span class="keywordtype">float</span> acc20 = 0.0f;</div><div class="line"><a name="l01709"></a><span class="lineno"> 1709</span>&#160;    <span class="keywordtype">float</span> acc21 = 0.0f;</div><div class="line"><a name="l01710"></a><span class="lineno"> 1710</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01711"></a><span class="lineno"> 1711</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01712"></a><span class="lineno"> 1712</span>&#160;    <span class="keywordtype">float</span> acc30 = 0.0f;</div><div class="line"><a name="l01713"></a><span class="lineno"> 1713</span>&#160;    <span class="keywordtype">float</span> acc31 = 0.0f;</div><div class="line"><a name="l01714"></a><span class="lineno"> 1714</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01715"></a><span class="lineno"> 1715</span>&#160;</div><div class="line"><a name="l01716"></a><span class="lineno"> 1716</span>&#160;    <span class="comment">// A and B src indices get incremented at the same time.</span></div><div class="line"><a name="l01717"></a><span class="lineno"> 1717</span>&#160;    <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01718"></a><span class="lineno"> 1718</span>&#160;    <span class="keywordflow">for</span>(; i &lt;= ((int)COLS_A - 8); i += 8)</div><div class="line"><a name="l01719"></a><span class="lineno"> 1719</span>&#160;    {</div><div class="line"><a name="l01720"></a><span class="lineno"> 1720</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01721"></a><span class="lineno"> 1721</span>&#160;        float8 a0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l01722"></a><span class="lineno"> 1722</span>&#160;</div><div class="line"><a name="l01723"></a><span class="lineno"> 1723</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01724"></a><span class="lineno"> 1724</span>&#160;        float2 b0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01725"></a><span class="lineno"> 1725</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01726"></a><span class="lineno"> 1726</span>&#160;        float2 b1 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01727"></a><span class="lineno"> 1727</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01728"></a><span class="lineno"> 1728</span>&#160;        float2 b2 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01729"></a><span class="lineno"> 1729</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01730"></a><span class="lineno"> 1730</span>&#160;        float2 b3 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01731"></a><span class="lineno"> 1731</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01732"></a><span class="lineno"> 1732</span>&#160;        float2 b4 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01733"></a><span class="lineno"> 1733</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01734"></a><span class="lineno"> 1734</span>&#160;        float2 b5 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01735"></a><span class="lineno"> 1735</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01736"></a><span class="lineno"> 1736</span>&#160;        float2 b6 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01737"></a><span class="lineno"> 1737</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01738"></a><span class="lineno"> 1738</span>&#160;        float2 b7 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01739"></a><span class="lineno"> 1739</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01740"></a><span class="lineno"> 1740</span>&#160;</div><div class="line"><a name="l01741"></a><span class="lineno"> 1741</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01742"></a><span class="lineno"> 1742</span>&#160;        acc00 = fma(a0.s0, b0.s0, acc00);</div><div class="line"><a name="l01743"></a><span class="lineno"> 1743</span>&#160;        acc00 = fma(a0.s1, b1.s0, acc00);</div><div class="line"><a name="l01744"></a><span class="lineno"> 1744</span>&#160;        acc00 = fma(a0.s2, b2.s0, acc00);</div><div class="line"><a name="l01745"></a><span class="lineno"> 1745</span>&#160;        acc00 = fma(a0.s3, b3.s0, acc00);</div><div class="line"><a name="l01746"></a><span class="lineno"> 1746</span>&#160;        acc00 = fma(a0.s4, b4.s0, acc00);</div><div class="line"><a name="l01747"></a><span class="lineno"> 1747</span>&#160;        acc00 = fma(a0.s5, b5.s0, acc00);</div><div class="line"><a name="l01748"></a><span class="lineno"> 1748</span>&#160;        acc00 = fma(a0.s6, b6.s0, acc00);</div><div class="line"><a name="l01749"></a><span class="lineno"> 1749</span>&#160;        acc00 = fma(a0.s7, b7.s0, acc00);</div><div class="line"><a name="l01750"></a><span class="lineno"> 1750</span>&#160;</div><div class="line"><a name="l01751"></a><span class="lineno"> 1751</span>&#160;        acc01 = fma(a0.s0, b0.s1, acc01);</div><div class="line"><a name="l01752"></a><span class="lineno"> 1752</span>&#160;        acc01 = fma(a0.s1, b1.s1, acc01);</div><div class="line"><a name="l01753"></a><span class="lineno"> 1753</span>&#160;        acc01 = fma(a0.s2, b2.s1, acc01);</div><div class="line"><a name="l01754"></a><span class="lineno"> 1754</span>&#160;        acc01 = fma(a0.s3, b3.s1, acc01);</div><div class="line"><a name="l01755"></a><span class="lineno"> 1755</span>&#160;        acc01 = fma(a0.s4, b4.s1, acc01);</div><div class="line"><a name="l01756"></a><span class="lineno"> 1756</span>&#160;        acc01 = fma(a0.s5, b5.s1, acc01);</div><div class="line"><a name="l01757"></a><span class="lineno"> 1757</span>&#160;        acc01 = fma(a0.s6, b6.s1, acc01);</div><div class="line"><a name="l01758"></a><span class="lineno"> 1758</span>&#160;        acc01 = fma(a0.s7, b7.s1, acc01);</div><div class="line"><a name="l01759"></a><span class="lineno"> 1759</span>&#160;</div><div class="line"><a name="l01760"></a><span class="lineno"> 1760</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01761"></a><span class="lineno"> 1761</span>&#160;        a0    = vload8(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01762"></a><span class="lineno"> 1762</span>&#160;        acc10 = fma(a0.s0, b0.s0, acc10);</div><div class="line"><a name="l01763"></a><span class="lineno"> 1763</span>&#160;        acc10 = fma(a0.s1, b1.s0, acc10);</div><div class="line"><a name="l01764"></a><span class="lineno"> 1764</span>&#160;        acc10 = fma(a0.s2, b2.s0, acc10);</div><div class="line"><a name="l01765"></a><span class="lineno"> 1765</span>&#160;        acc10 = fma(a0.s3, b3.s0, acc10);</div><div class="line"><a name="l01766"></a><span class="lineno"> 1766</span>&#160;        acc10 = fma(a0.s4, b4.s0, acc10);</div><div class="line"><a name="l01767"></a><span class="lineno"> 1767</span>&#160;        acc10 = fma(a0.s5, b5.s0, acc10);</div><div class="line"><a name="l01768"></a><span class="lineno"> 1768</span>&#160;        acc10 = fma(a0.s6, b6.s0, acc10);</div><div class="line"><a name="l01769"></a><span class="lineno"> 1769</span>&#160;        acc10 = fma(a0.s7, b7.s0, acc10);</div><div class="line"><a name="l01770"></a><span class="lineno"> 1770</span>&#160;</div><div class="line"><a name="l01771"></a><span class="lineno"> 1771</span>&#160;        acc11 = fma(a0.s0, b0.s1, acc11);</div><div class="line"><a name="l01772"></a><span class="lineno"> 1772</span>&#160;        acc11 = fma(a0.s1, b1.s1, acc11);</div><div class="line"><a name="l01773"></a><span class="lineno"> 1773</span>&#160;        acc11 = fma(a0.s2, b2.s1, acc11);</div><div class="line"><a name="l01774"></a><span class="lineno"> 1774</span>&#160;        acc11 = fma(a0.s3, b3.s1, acc11);</div><div class="line"><a name="l01775"></a><span class="lineno"> 1775</span>&#160;        acc11 = fma(a0.s4, b4.s1, acc11);</div><div class="line"><a name="l01776"></a><span class="lineno"> 1776</span>&#160;        acc11 = fma(a0.s5, b5.s1, acc11);</div><div class="line"><a name="l01777"></a><span class="lineno"> 1777</span>&#160;        acc11 = fma(a0.s6, b6.s1, acc11);</div><div class="line"><a name="l01778"></a><span class="lineno"> 1778</span>&#160;        acc11 = fma(a0.s7, b7.s1, acc11);</div><div class="line"><a name="l01779"></a><span class="lineno"> 1779</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01780"></a><span class="lineno"> 1780</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01781"></a><span class="lineno"> 1781</span>&#160;        a0    = vload8(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01782"></a><span class="lineno"> 1782</span>&#160;        acc20 = fma(a0.s0, b0.s0, acc20);</div><div class="line"><a name="l01783"></a><span class="lineno"> 1783</span>&#160;        acc20 = fma(a0.s1, b1.s0, acc20);</div><div class="line"><a name="l01784"></a><span class="lineno"> 1784</span>&#160;        acc20 = fma(a0.s2, b2.s0, acc20);</div><div class="line"><a name="l01785"></a><span class="lineno"> 1785</span>&#160;        acc20 = fma(a0.s3, b3.s0, acc20);</div><div class="line"><a name="l01786"></a><span class="lineno"> 1786</span>&#160;        acc20 = fma(a0.s4, b4.s0, acc20);</div><div class="line"><a name="l01787"></a><span class="lineno"> 1787</span>&#160;        acc20 = fma(a0.s5, b5.s0, acc20);</div><div class="line"><a name="l01788"></a><span class="lineno"> 1788</span>&#160;        acc20 = fma(a0.s6, b6.s0, acc20);</div><div class="line"><a name="l01789"></a><span class="lineno"> 1789</span>&#160;        acc20 = fma(a0.s7, b7.s0, acc20);</div><div class="line"><a name="l01790"></a><span class="lineno"> 1790</span>&#160;</div><div class="line"><a name="l01791"></a><span class="lineno"> 1791</span>&#160;        acc21 = fma(a0.s0, b0.s1, acc21);</div><div class="line"><a name="l01792"></a><span class="lineno"> 1792</span>&#160;        acc21 = fma(a0.s1, b1.s1, acc21);</div><div class="line"><a name="l01793"></a><span class="lineno"> 1793</span>&#160;        acc21 = fma(a0.s2, b2.s1, acc21);</div><div class="line"><a name="l01794"></a><span class="lineno"> 1794</span>&#160;        acc21 = fma(a0.s3, b3.s1, acc21);</div><div class="line"><a name="l01795"></a><span class="lineno"> 1795</span>&#160;        acc21 = fma(a0.s4, b4.s1, acc21);</div><div class="line"><a name="l01796"></a><span class="lineno"> 1796</span>&#160;        acc21 = fma(a0.s5, b5.s1, acc21);</div><div class="line"><a name="l01797"></a><span class="lineno"> 1797</span>&#160;        acc21 = fma(a0.s6, b6.s1, acc21);</div><div class="line"><a name="l01798"></a><span class="lineno"> 1798</span>&#160;        acc21 = fma(a0.s7, b7.s1, acc21);</div><div class="line"><a name="l01799"></a><span class="lineno"> 1799</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01800"></a><span class="lineno"> 1800</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01801"></a><span class="lineno"> 1801</span>&#160;        a0    = vload8(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01802"></a><span class="lineno"> 1802</span>&#160;        acc30 = fma(a0.s0, b0.s0, acc30);</div><div class="line"><a name="l01803"></a><span class="lineno"> 1803</span>&#160;        acc30 = fma(a0.s1, b1.s0, acc30);</div><div class="line"><a name="l01804"></a><span class="lineno"> 1804</span>&#160;        acc30 = fma(a0.s2, b2.s0, acc30);</div><div class="line"><a name="l01805"></a><span class="lineno"> 1805</span>&#160;        acc30 = fma(a0.s3, b3.s0, acc30);</div><div class="line"><a name="l01806"></a><span class="lineno"> 1806</span>&#160;        acc30 = fma(a0.s4, b4.s0, acc30);</div><div class="line"><a name="l01807"></a><span class="lineno"> 1807</span>&#160;        acc30 = fma(a0.s5, b5.s0, acc30);</div><div class="line"><a name="l01808"></a><span class="lineno"> 1808</span>&#160;        acc30 = fma(a0.s6, b6.s0, acc30);</div><div class="line"><a name="l01809"></a><span class="lineno"> 1809</span>&#160;        acc30 = fma(a0.s7, b7.s0, acc30);</div><div class="line"><a name="l01810"></a><span class="lineno"> 1810</span>&#160;</div><div class="line"><a name="l01811"></a><span class="lineno"> 1811</span>&#160;        acc31 = fma(a0.s0, b0.s1, acc31);</div><div class="line"><a name="l01812"></a><span class="lineno"> 1812</span>&#160;        acc31 = fma(a0.s1, b1.s1, acc31);</div><div class="line"><a name="l01813"></a><span class="lineno"> 1813</span>&#160;        acc31 = fma(a0.s2, b2.s1, acc31);</div><div class="line"><a name="l01814"></a><span class="lineno"> 1814</span>&#160;        acc31 = fma(a0.s3, b3.s1, acc31);</div><div class="line"><a name="l01815"></a><span class="lineno"> 1815</span>&#160;        acc31 = fma(a0.s4, b4.s1, acc31);</div><div class="line"><a name="l01816"></a><span class="lineno"> 1816</span>&#160;        acc31 = fma(a0.s5, b5.s1, acc31);</div><div class="line"><a name="l01817"></a><span class="lineno"> 1817</span>&#160;        acc31 = fma(a0.s6, b6.s1, acc31);</div><div class="line"><a name="l01818"></a><span class="lineno"> 1818</span>&#160;        acc31 = fma(a0.s7, b7.s1, acc31);</div><div class="line"><a name="l01819"></a><span class="lineno"> 1819</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01820"></a><span class="lineno"> 1820</span>&#160;</div><div class="line"><a name="l01821"></a><span class="lineno"> 1821</span>&#160;        src_addr.s0 += <span class="keyword">sizeof</span>(float) * 8;</div><div class="line"><a name="l01822"></a><span class="lineno"> 1822</span>&#160;    }</div><div class="line"><a name="l01823"></a><span class="lineno"> 1823</span>&#160;    <span class="comment">// float size increment</span></div><div class="line"><a name="l01824"></a><span class="lineno"> 1824</span>&#160;    <span class="keywordflow">for</span>(; i &lt; (int)COLS_A; ++i)</div><div class="line"><a name="l01825"></a><span class="lineno"> 1825</span>&#160;    {</div><div class="line"><a name="l01826"></a><span class="lineno"> 1826</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01827"></a><span class="lineno"> 1827</span>&#160;        <span class="keywordtype">float</span> a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01828"></a><span class="lineno"> 1828</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01829"></a><span class="lineno"> 1829</span>&#160;        <span class="keywordtype">float</span> a1 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01830"></a><span class="lineno"> 1830</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01831"></a><span class="lineno"> 1831</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01832"></a><span class="lineno"> 1832</span>&#160;        <span class="keywordtype">float</span> a2 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01833"></a><span class="lineno"> 1833</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01834"></a><span class="lineno"> 1834</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01835"></a><span class="lineno"> 1835</span>&#160;        <span class="keywordtype">float</span> a3 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01836"></a><span class="lineno"> 1836</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01837"></a><span class="lineno"> 1837</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01838"></a><span class="lineno"> 1838</span>&#160;        float2 b0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01839"></a><span class="lineno"> 1839</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01840"></a><span class="lineno"> 1840</span>&#160;</div><div class="line"><a name="l01841"></a><span class="lineno"> 1841</span>&#160;        <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01842"></a><span class="lineno"> 1842</span>&#160;        acc00 = fma(a0, b0.s0, acc00);</div><div class="line"><a name="l01843"></a><span class="lineno"> 1843</span>&#160;        acc01 = fma(a0, b0.s1, acc01);</div><div class="line"><a name="l01844"></a><span class="lineno"> 1844</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01845"></a><span class="lineno"> 1845</span>&#160;        acc10 = fma(a1, b0.s0, acc10);</div><div class="line"><a name="l01846"></a><span class="lineno"> 1846</span>&#160;        acc11 = fma(a1, b0.s1, acc11);</div><div class="line"><a name="l01847"></a><span class="lineno"> 1847</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01848"></a><span class="lineno"> 1848</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01849"></a><span class="lineno"> 1849</span>&#160;        acc20 = fma(a2, b0.s0, acc20);</div><div class="line"><a name="l01850"></a><span class="lineno"> 1850</span>&#160;        acc21 = fma(a2, b0.s1, acc21);</div><div class="line"><a name="l01851"></a><span class="lineno"> 1851</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01852"></a><span class="lineno"> 1852</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01853"></a><span class="lineno"> 1853</span>&#160;        acc30 = fma(a3, b0.s0, acc30);</div><div class="line"><a name="l01854"></a><span class="lineno"> 1854</span>&#160;        acc31 = fma(a3, b0.s1, acc31);</div><div class="line"><a name="l01855"></a><span class="lineno"> 1855</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01856"></a><span class="lineno"> 1856</span>&#160;</div><div class="line"><a name="l01857"></a><span class="lineno"> 1857</span>&#160;        src_addr.s0 += <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01858"></a><span class="lineno"> 1858</span>&#160;    }</div><div class="line"><a name="l01859"></a><span class="lineno"> 1859</span>&#160;</div><div class="line"><a name="l01860"></a><span class="lineno"> 1860</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01861"></a><span class="lineno"> 1861</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01862"></a><span class="lineno"> 1862</span>&#160;</div><div class="line"><a name="l01863"></a><span class="lineno"> 1863</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01864"></a><span class="lineno"> 1864</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l01865"></a><span class="lineno"> 1865</span>&#160;</div><div class="line"><a name="l01866"></a><span class="lineno"> 1866</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01867"></a><span class="lineno"> 1867</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l01868"></a><span class="lineno"> 1868</span>&#160;</div><div class="line"><a name="l01869"></a><span class="lineno"> 1869</span>&#160;    <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l01870"></a><span class="lineno"> 1870</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01871"></a><span class="lineno"> 1871</span>&#160;    acc00 = acc00 * ALPHA;</div><div class="line"><a name="l01872"></a><span class="lineno"> 1872</span>&#160;    acc01 = acc01 * ALPHA;</div><div class="line"><a name="l01873"></a><span class="lineno"> 1873</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01874"></a><span class="lineno"> 1874</span>&#160;    float2 acc0 = ((float2)(acc00, acc01));</div><div class="line"><a name="l01875"></a><span class="lineno"> 1875</span>&#160;    vstore2(acc0, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01876"></a><span class="lineno"> 1876</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01877"></a><span class="lineno"> 1877</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01878"></a><span class="lineno"> 1878</span>&#160;    acc10 = acc10 * ALPHA;</div><div class="line"><a name="l01879"></a><span class="lineno"> 1879</span>&#160;    acc11 = acc11 * ALPHA;</div><div class="line"><a name="l01880"></a><span class="lineno"> 1880</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01881"></a><span class="lineno"> 1881</span>&#160;    float2 acc1 = ((float2)(acc10, acc11));</div><div class="line"><a name="l01882"></a><span class="lineno"> 1882</span>&#160;    vstore2(acc1, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01883"></a><span class="lineno"> 1883</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01884"></a><span class="lineno"> 1884</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01885"></a><span class="lineno"> 1885</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01886"></a><span class="lineno"> 1886</span>&#160;    acc20 = acc20 * ALPHA;</div><div class="line"><a name="l01887"></a><span class="lineno"> 1887</span>&#160;    acc21 = acc21 * ALPHA;</div><div class="line"><a name="l01888"></a><span class="lineno"> 1888</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01889"></a><span class="lineno"> 1889</span>&#160;    float2 acc2 = ((float2)(acc20, acc21));</div><div class="line"><a name="l01890"></a><span class="lineno"> 1890</span>&#160;    vstore2(acc2, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01891"></a><span class="lineno"> 1891</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01892"></a><span class="lineno"> 1892</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01893"></a><span class="lineno"> 1893</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01894"></a><span class="lineno"> 1894</span>&#160;    acc30 = acc30 * ALPHA;</div><div class="line"><a name="l01895"></a><span class="lineno"> 1895</span>&#160;    acc31 = acc31 * ALPHA;</div><div class="line"><a name="l01896"></a><span class="lineno"> 1896</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01897"></a><span class="lineno"> 1897</span>&#160;    float2 acc3 = (float2)(acc30, acc31);</div><div class="line"><a name="l01898"></a><span class="lineno"> 1898</span>&#160;    vstore2(acc3, 0, (__global <span class="keywordtype">float</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01899"></a><span class="lineno"> 1899</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01900"></a><span class="lineno"> 1900</span>&#160;}</div><div class="line"><a name="l01901"></a><span class="lineno"> 1901</span>&#160;</div><div class="line"><a name="l01931"></a><span class="lineno"> 1931</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f16_bifrost(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01932"></a><span class="lineno"> 1932</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01933"></a><span class="lineno"> 1933</span>&#160;                                                 <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01934"></a><span class="lineno"> 1934</span>&#160;                                                 uint src0_stride_z,</div><div class="line"><a name="l01935"></a><span class="lineno"> 1935</span>&#160;                                                 uint src1_stride_z,</div><div class="line"><a name="l01936"></a><span class="lineno"> 1936</span>&#160;                                                 uint dst_stride_z)</div><div class="line"><a name="l01937"></a><span class="lineno"> 1937</span>&#160;{</div><div class="line"><a name="l01938"></a><span class="lineno"> 1938</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01939"></a><span class="lineno"> 1939</span>&#160;</div><div class="line"><a name="l01940"></a><span class="lineno"> 1940</span>&#160;    <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01941"></a><span class="lineno"> 1941</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01942"></a><span class="lineno"> 1942</span>&#160;</div><div class="line"><a name="l01943"></a><span class="lineno"> 1943</span>&#160;    <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01944"></a><span class="lineno"> 1944</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01945"></a><span class="lineno"> 1945</span>&#160;</div><div class="line"><a name="l01946"></a><span class="lineno"> 1946</span>&#160;    <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01947"></a><span class="lineno"> 1947</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(<a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a>);</div><div class="line"><a name="l01948"></a><span class="lineno"> 1948</span>&#160;</div><div class="line"><a name="l01949"></a><span class="lineno"> 1949</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01950"></a><span class="lineno"> 1950</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01951"></a><span class="lineno"> 1951</span>&#160;</div><div class="line"><a name="l01952"></a><span class="lineno"> 1952</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01953"></a><span class="lineno"> 1953</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01954"></a><span class="lineno"> 1954</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01955"></a><span class="lineno"> 1955</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01956"></a><span class="lineno"> 1956</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01957"></a><span class="lineno"> 1957</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01958"></a><span class="lineno"> 1958</span>&#160;</div><div class="line"><a name="l01959"></a><span class="lineno"> 1959</span>&#160;    half8 acc0 = 0.0h;</div><div class="line"><a name="l01960"></a><span class="lineno"> 1960</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01961"></a><span class="lineno"> 1961</span>&#160;    half8 acc1 = 0.0h;</div><div class="line"><a name="l01962"></a><span class="lineno"> 1962</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01963"></a><span class="lineno"> 1963</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01964"></a><span class="lineno"> 1964</span>&#160;    half8 acc2 = 0.0h;</div><div class="line"><a name="l01965"></a><span class="lineno"> 1965</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01966"></a><span class="lineno"> 1966</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01967"></a><span class="lineno"> 1967</span>&#160;    half8 acc3 = 0.0h;</div><div class="line"><a name="l01968"></a><span class="lineno"> 1968</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01969"></a><span class="lineno"> 1969</span>&#160;</div><div class="line"><a name="l01970"></a><span class="lineno"> 1970</span>&#160;    <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01971"></a><span class="lineno"> 1971</span>&#160;    <span class="keywordflow">for</span>(; i &lt;= ((int)COLS_A - 4); i += 4)</div><div class="line"><a name="l01972"></a><span class="lineno"> 1972</span>&#160;    {</div><div class="line"><a name="l01973"></a><span class="lineno"> 1973</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01974"></a><span class="lineno"> 1974</span>&#160;        half4 a0 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01975"></a><span class="lineno"> 1975</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01976"></a><span class="lineno"> 1976</span>&#160;        half4 a1 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01977"></a><span class="lineno"> 1977</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01978"></a><span class="lineno"> 1978</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01979"></a><span class="lineno"> 1979</span>&#160;        half4 a2 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01980"></a><span class="lineno"> 1980</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01981"></a><span class="lineno"> 1981</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01982"></a><span class="lineno"> 1982</span>&#160;        half4 a3 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01983"></a><span class="lineno"> 1983</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01984"></a><span class="lineno"> 1984</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01985"></a><span class="lineno"> 1985</span>&#160;        half8 b0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01986"></a><span class="lineno"> 1986</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01987"></a><span class="lineno"> 1987</span>&#160;</div><div class="line"><a name="l01988"></a><span class="lineno"> 1988</span>&#160;        <span class="comment">// Accumulate</span></div><div class="line"><a name="l01989"></a><span class="lineno"> 1989</span>&#160;        acc0 = fma(b0, (half8)a0.s0, acc0);</div><div class="line"><a name="l01990"></a><span class="lineno"> 1990</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01991"></a><span class="lineno"> 1991</span>&#160;        acc1 = fma(b0, (half8)a1.s0, acc1);</div><div class="line"><a name="l01992"></a><span class="lineno"> 1992</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l01993"></a><span class="lineno"> 1993</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01994"></a><span class="lineno"> 1994</span>&#160;        acc2 = fma(b0, (half8)a2.s0, acc2);</div><div class="line"><a name="l01995"></a><span class="lineno"> 1995</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l01996"></a><span class="lineno"> 1996</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01997"></a><span class="lineno"> 1997</span>&#160;        acc3 = fma(b0, (half8)a3.s0, acc3);</div><div class="line"><a name="l01998"></a><span class="lineno"> 1998</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l01999"></a><span class="lineno"> 1999</span>&#160;</div><div class="line"><a name="l02000"></a><span class="lineno"> 2000</span>&#160;        b0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02001"></a><span class="lineno"> 2001</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02002"></a><span class="lineno"> 2002</span>&#160;        acc0 = fma(b0, (half8)a0.s1, acc0);</div><div class="line"><a name="l02003"></a><span class="lineno"> 2003</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02004"></a><span class="lineno"> 2004</span>&#160;        acc1 = fma(b0, (half8)a1.s1, acc1);</div><div class="line"><a name="l02005"></a><span class="lineno"> 2005</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02006"></a><span class="lineno"> 2006</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02007"></a><span class="lineno"> 2007</span>&#160;        acc2 = fma(b0, (half8)a2.s1, acc2);</div><div class="line"><a name="l02008"></a><span class="lineno"> 2008</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02009"></a><span class="lineno"> 2009</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02010"></a><span class="lineno"> 2010</span>&#160;        acc3 = fma(b0, (half8)a3.s1, acc3);</div><div class="line"><a name="l02011"></a><span class="lineno"> 2011</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02012"></a><span class="lineno"> 2012</span>&#160;</div><div class="line"><a name="l02013"></a><span class="lineno"> 2013</span>&#160;        b0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02014"></a><span class="lineno"> 2014</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02015"></a><span class="lineno"> 2015</span>&#160;        acc0 = fma(b0, (half8)a0.s2, acc0);</div><div class="line"><a name="l02016"></a><span class="lineno"> 2016</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02017"></a><span class="lineno"> 2017</span>&#160;        acc1 = fma(b0, (half8)a1.s2, acc1);</div><div class="line"><a name="l02018"></a><span class="lineno"> 2018</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02019"></a><span class="lineno"> 2019</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02020"></a><span class="lineno"> 2020</span>&#160;        acc2 = fma(b0, (half8)a2.s2, acc2);</div><div class="line"><a name="l02021"></a><span class="lineno"> 2021</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02022"></a><span class="lineno"> 2022</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02023"></a><span class="lineno"> 2023</span>&#160;        acc3 = fma(b0, (half8)a3.s2, acc3);</div><div class="line"><a name="l02024"></a><span class="lineno"> 2024</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02025"></a><span class="lineno"> 2025</span>&#160;</div><div class="line"><a name="l02026"></a><span class="lineno"> 2026</span>&#160;        b0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02027"></a><span class="lineno"> 2027</span>&#160;        src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02028"></a><span class="lineno"> 2028</span>&#160;        acc0 = fma(b0, (half8)a0.s3, acc0);</div><div class="line"><a name="l02029"></a><span class="lineno"> 2029</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02030"></a><span class="lineno"> 2030</span>&#160;        acc1 = fma(b0, (half8)a1.s3, acc1);</div><div class="line"><a name="l02031"></a><span class="lineno"> 2031</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02032"></a><span class="lineno"> 2032</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02033"></a><span class="lineno"> 2033</span>&#160;        acc2 = fma(b0, (half8)a2.s3, acc2);</div><div class="line"><a name="l02034"></a><span class="lineno"> 2034</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02035"></a><span class="lineno"> 2035</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02036"></a><span class="lineno"> 2036</span>&#160;        acc3 = fma(b0, (half8)a3.s3, acc3);</div><div class="line"><a name="l02037"></a><span class="lineno"> 2037</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02038"></a><span class="lineno"> 2038</span>&#160;</div><div class="line"><a name="l02039"></a><span class="lineno"> 2039</span>&#160;        src_addr.s0 += 4 * <span class="keyword">sizeof</span>(<a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a>);</div><div class="line"><a name="l02040"></a><span class="lineno"> 2040</span>&#160;    }</div><div class="line"><a name="l02041"></a><span class="lineno"> 2041</span>&#160;</div><div class="line"><a name="l02042"></a><span class="lineno"> 2042</span>&#160;    <span class="keywordflow">for</span>(; i &lt; (int)COLS_A; ++i)</div><div class="line"><a name="l02043"></a><span class="lineno"> 2043</span>&#160;    {</div><div class="line"><a name="l02044"></a><span class="lineno"> 2044</span>&#160;        <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02045"></a><span class="lineno"> 2045</span>&#160;        <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> a0 = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l02046"></a><span class="lineno"> 2046</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02047"></a><span class="lineno"> 2047</span>&#160;        <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> a1 = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l02048"></a><span class="lineno"> 2048</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02049"></a><span class="lineno"> 2049</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02050"></a><span class="lineno"> 2050</span>&#160;        <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> a2 = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l02051"></a><span class="lineno"> 2051</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02052"></a><span class="lineno"> 2052</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02053"></a><span class="lineno"> 2053</span>&#160;        <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> a3 = *((__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l02054"></a><span class="lineno"> 2054</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02055"></a><span class="lineno"> 2055</span>&#160;        <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02056"></a><span class="lineno"> 2056</span>&#160;        half8 b0 = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02057"></a><span class="lineno"> 2057</span>&#160;</div><div class="line"><a name="l02058"></a><span class="lineno"> 2058</span>&#160;        src_addr += (int2)(<span class="keyword">sizeof</span>(<a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a>), src1_stride_y);</div><div class="line"><a name="l02059"></a><span class="lineno"> 2059</span>&#160;</div><div class="line"><a name="l02060"></a><span class="lineno"> 2060</span>&#160;        <span class="comment">// Accumulate</span></div><div class="line"><a name="l02061"></a><span class="lineno"> 2061</span>&#160;        acc0 = fma(b0, (half8)a0, acc0); <span class="comment">// b0 * (half8)a0;</span></div><div class="line"><a name="l02062"></a><span class="lineno"> 2062</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02063"></a><span class="lineno"> 2063</span>&#160;        acc1 = fma(b0, (half8)a1, acc1); <span class="comment">// b0 * (half8)a1;</span></div><div class="line"><a name="l02064"></a><span class="lineno"> 2064</span>&#160;<span class="preprocessor">#endif                                   // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02065"></a><span class="lineno"> 2065</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02066"></a><span class="lineno"> 2066</span>&#160;        acc2 = fma(b0, (half8)a2, acc2); <span class="comment">// b0 * (half8)a2;</span></div><div class="line"><a name="l02067"></a><span class="lineno"> 2067</span>&#160;<span class="preprocessor">#endif                                   // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02068"></a><span class="lineno"> 2068</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02069"></a><span class="lineno"> 2069</span>&#160;        acc3 = fma(b0, (half8)a3, acc3); <span class="comment">// b0 * (half8)a3;</span></div><div class="line"><a name="l02070"></a><span class="lineno"> 2070</span>&#160;<span class="preprocessor">#endif                                   // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02071"></a><span class="lineno"> 2071</span>&#160;    }</div><div class="line"><a name="l02072"></a><span class="lineno"> 2072</span>&#160;</div><div class="line"><a name="l02073"></a><span class="lineno"> 2073</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02074"></a><span class="lineno"> 2074</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02075"></a><span class="lineno"> 2075</span>&#160;</div><div class="line"><a name="l02076"></a><span class="lineno"> 2076</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02077"></a><span class="lineno"> 2077</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l02078"></a><span class="lineno"> 2078</span>&#160;</div><div class="line"><a name="l02079"></a><span class="lineno"> 2079</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02080"></a><span class="lineno"> 2080</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l02081"></a><span class="lineno"> 2081</span>&#160;</div><div class="line"><a name="l02082"></a><span class="lineno"> 2082</span>&#160;    <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l02083"></a><span class="lineno"> 2083</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02084"></a><span class="lineno"> 2084</span>&#160;    acc0 = acc0 * (half8)ALPHA;</div><div class="line"><a name="l02085"></a><span class="lineno"> 2085</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02086"></a><span class="lineno"> 2086</span>&#160;    vstore8(acc0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l02087"></a><span class="lineno"> 2087</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02088"></a><span class="lineno"> 2088</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02089"></a><span class="lineno"> 2089</span>&#160;    acc1 = acc1 * (half8)ALPHA;</div><div class="line"><a name="l02090"></a><span class="lineno"> 2090</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02091"></a><span class="lineno"> 2091</span>&#160;    vstore8(acc1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l02092"></a><span class="lineno"> 2092</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02093"></a><span class="lineno"> 2093</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02094"></a><span class="lineno"> 2094</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02095"></a><span class="lineno"> 2095</span>&#160;    acc2 = acc2 * (half8)ALPHA;</div><div class="line"><a name="l02096"></a><span class="lineno"> 2096</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02097"></a><span class="lineno"> 2097</span>&#160;    vstore8(acc2, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l02098"></a><span class="lineno"> 2098</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02099"></a><span class="lineno"> 2099</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02100"></a><span class="lineno"> 2100</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02101"></a><span class="lineno"> 2101</span>&#160;    acc3 = acc3 * (half8)ALPHA;</div><div class="line"><a name="l02102"></a><span class="lineno"> 2102</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02103"></a><span class="lineno"> 2103</span>&#160;    vstore8(acc3, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l02104"></a><span class="lineno"> 2104</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02105"></a><span class="lineno"> 2105</span>&#160;}</div><div class="line"><a name="l02106"></a><span class="lineno"> 2106</span>&#160;</div><div class="line"><a name="l02107"></a><span class="lineno"> 2107</span>&#160;<span class="preprocessor">#if defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l02108"></a><span class="lineno"> 2108</span>&#160;</div><div class="line"><a name="l02137"></a><span class="lineno"> 2137</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_qs8(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02138"></a><span class="lineno"> 2138</span>&#160;                          <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l02139"></a><span class="lineno"> 2139</span>&#160;                          <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l02140"></a><span class="lineno"> 2140</span>&#160;                          uint src0_stride_z,</div><div class="line"><a name="l02141"></a><span class="lineno"> 2141</span>&#160;                          uint src1_stride_z,</div><div class="line"><a name="l02142"></a><span class="lineno"> 2142</span>&#160;                          uint dst_stride_z)</div><div class="line"><a name="l02143"></a><span class="lineno"> 2143</span>&#160;{</div><div class="line"><a name="l02144"></a><span class="lineno"> 2144</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l02145"></a><span class="lineno"> 2145</span>&#160;</div><div class="line"><a name="l02146"></a><span class="lineno"> 2146</span>&#160;    <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l02147"></a><span class="lineno"> 2147</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l02148"></a><span class="lineno"> 2148</span>&#160;</div><div class="line"><a name="l02149"></a><span class="lineno"> 2149</span>&#160;    <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l02150"></a><span class="lineno"> 2150</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l02151"></a><span class="lineno"> 2151</span>&#160;</div><div class="line"><a name="l02152"></a><span class="lineno"> 2152</span>&#160;    <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l02153"></a><span class="lineno"> 2153</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(char);</div><div class="line"><a name="l02154"></a><span class="lineno"> 2154</span>&#160;</div><div class="line"><a name="l02155"></a><span class="lineno"> 2155</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02156"></a><span class="lineno"> 2156</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l02157"></a><span class="lineno"> 2157</span>&#160;</div><div class="line"><a name="l02158"></a><span class="lineno"> 2158</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02159"></a><span class="lineno"> 2159</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l02160"></a><span class="lineno"> 2160</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l02161"></a><span class="lineno"> 2161</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02162"></a><span class="lineno"> 2162</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l02163"></a><span class="lineno"> 2163</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02164"></a><span class="lineno"> 2164</span>&#160;</div><div class="line"><a name="l02165"></a><span class="lineno"> 2165</span>&#160;    <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (COLS_A * <span class="keyword">sizeof</span>(char));</div><div class="line"><a name="l02166"></a><span class="lineno"> 2166</span>&#160;</div><div class="line"><a name="l02167"></a><span class="lineno"> 2167</span>&#160;    short8 acc00 = 0;</div><div class="line"><a name="l02168"></a><span class="lineno"> 2168</span>&#160;    short8 acc01 = 0;</div><div class="line"><a name="l02169"></a><span class="lineno"> 2169</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02170"></a><span class="lineno"> 2170</span>&#160;    short8 acc10 = 0;</div><div class="line"><a name="l02171"></a><span class="lineno"> 2171</span>&#160;    short8 acc11 = 0;</div><div class="line"><a name="l02172"></a><span class="lineno"> 2172</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02173"></a><span class="lineno"> 2173</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02174"></a><span class="lineno"> 2174</span>&#160;    short8 acc20 = 0;</div><div class="line"><a name="l02175"></a><span class="lineno"> 2175</span>&#160;    short8 acc21 = 0;</div><div class="line"><a name="l02176"></a><span class="lineno"> 2176</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02177"></a><span class="lineno"> 2177</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02178"></a><span class="lineno"> 2178</span>&#160;    short8 acc30 = 0;</div><div class="line"><a name="l02179"></a><span class="lineno"> 2179</span>&#160;    short8 acc31 = 0;</div><div class="line"><a name="l02180"></a><span class="lineno"> 2180</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02181"></a><span class="lineno"> 2181</span>&#160;</div><div class="line"><a name="l02182"></a><span class="lineno"> 2182</span>&#160;    <span class="comment">// This for loop performs 4 accumulations per iteration</span></div><div class="line"><a name="l02183"></a><span class="lineno"> 2183</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt;= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y))</div><div class="line"><a name="l02184"></a><span class="lineno"> 2184</span>&#160;    {</div><div class="line"><a name="l02185"></a><span class="lineno"> 2185</span>&#160;        char2 a0 = vload2(0, (__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l02186"></a><span class="lineno"> 2186</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02187"></a><span class="lineno"> 2187</span>&#160;        char2 a1 = vload2(0, (__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l02188"></a><span class="lineno"> 2188</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02189"></a><span class="lineno"> 2189</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02190"></a><span class="lineno"> 2190</span>&#160;        char2 a2 = vload2(0, (__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l02191"></a><span class="lineno"> 2191</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02192"></a><span class="lineno"> 2192</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02193"></a><span class="lineno"> 2193</span>&#160;        char2 a3 = vload2(0, (__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l02194"></a><span class="lineno"> 2194</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02195"></a><span class="lineno"> 2195</span>&#160;        char16 b0 = vload16(0, (__global <span class="keywordtype">char</span> *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));</div><div class="line"><a name="l02196"></a><span class="lineno"> 2196</span>&#160;        char16 b1 = vload16(0, (__global <span class="keywordtype">char</span> *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));</div><div class="line"><a name="l02197"></a><span class="lineno"> 2197</span>&#160;</div><div class="line"><a name="l02198"></a><span class="lineno"> 2198</span>&#160;        acc00 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc00, (char8)a0.s0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02199"></a><span class="lineno"> 2199</span>&#160;        acc00 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc00, (char8)a0.s1, b1.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02200"></a><span class="lineno"> 2200</span>&#160;        acc01 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc01, (char8)a0.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02201"></a><span class="lineno"> 2201</span>&#160;        acc01 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc01, (char8)a0.s1, b1.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02202"></a><span class="lineno"> 2202</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02203"></a><span class="lineno"> 2203</span>&#160;        acc10 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc10, (char8)a1.s0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02204"></a><span class="lineno"> 2204</span>&#160;        acc10 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc10, (char8)a1.s1, b1.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02205"></a><span class="lineno"> 2205</span>&#160;        acc11 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc11, (char8)a1.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02206"></a><span class="lineno"> 2206</span>&#160;        acc11 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc11, (char8)a1.s1, b1.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02207"></a><span class="lineno"> 2207</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02208"></a><span class="lineno"> 2208</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02209"></a><span class="lineno"> 2209</span>&#160;        acc20 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc20, (char8)a2.s0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02210"></a><span class="lineno"> 2210</span>&#160;        acc20 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc20, (char8)a2.s1, b1.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02211"></a><span class="lineno"> 2211</span>&#160;        acc21 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc21, (char8)a2.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02212"></a><span class="lineno"> 2212</span>&#160;        acc21 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc21, (char8)a2.s1, b1.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02213"></a><span class="lineno"> 2213</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02214"></a><span class="lineno"> 2214</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02215"></a><span class="lineno"> 2215</span>&#160;        acc30 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc30, (char8)a3.s0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02216"></a><span class="lineno"> 2216</span>&#160;        acc30 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc30, (char8)a3.s1, b1.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02217"></a><span class="lineno"> 2217</span>&#160;        acc31 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc31, (char8)a3.s0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02218"></a><span class="lineno"> 2218</span>&#160;        acc31 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc31, (char8)a3.s1, b1.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02219"></a><span class="lineno"> 2219</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02220"></a><span class="lineno"> 2220</span>&#160;    }</div><div class="line"><a name="l02221"></a><span class="lineno"> 2221</span>&#160;</div><div class="line"><a name="l02222"></a><span class="lineno"> 2222</span>&#160;    <span class="comment">// Left-over accumulations</span></div><div class="line"><a name="l02223"></a><span class="lineno"> 2223</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt; end_row_vec_a; src_addr += (int2)(1, src1_stride_y))</div><div class="line"><a name="l02224"></a><span class="lineno"> 2224</span>&#160;    {</div><div class="line"><a name="l02225"></a><span class="lineno"> 2225</span>&#160;        <span class="keywordtype">char</span> a0 = *((__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l02226"></a><span class="lineno"> 2226</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02227"></a><span class="lineno"> 2227</span>&#160;        <span class="keywordtype">char</span> a1 = *((__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l02228"></a><span class="lineno"> 2228</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02229"></a><span class="lineno"> 2229</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02230"></a><span class="lineno"> 2230</span>&#160;        <span class="keywordtype">char</span> a2 = *((__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l02231"></a><span class="lineno"> 2231</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02232"></a><span class="lineno"> 2232</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02233"></a><span class="lineno"> 2233</span>&#160;        <span class="keywordtype">char</span> a3 = *((__global <span class="keywordtype">char</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l02234"></a><span class="lineno"> 2234</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02235"></a><span class="lineno"> 2235</span>&#160;        char16 b0 = vload16(0, (__global <span class="keywordtype">char</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02236"></a><span class="lineno"> 2236</span>&#160;</div><div class="line"><a name="l02237"></a><span class="lineno"> 2237</span>&#160;        acc00 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc00, (char8)a0, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02238"></a><span class="lineno"> 2238</span>&#160;        acc01 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc01, (char8)a0, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02239"></a><span class="lineno"> 2239</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02240"></a><span class="lineno"> 2240</span>&#160;        acc10 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc10, (char8)a1, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02241"></a><span class="lineno"> 2241</span>&#160;        acc11 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc11, (char8)a1, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02242"></a><span class="lineno"> 2242</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02243"></a><span class="lineno"> 2243</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02244"></a><span class="lineno"> 2244</span>&#160;        acc20 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc20, (char8)a2, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02245"></a><span class="lineno"> 2245</span>&#160;        acc21 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc21, (char8)a2, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02246"></a><span class="lineno"> 2246</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02247"></a><span class="lineno"> 2247</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02248"></a><span class="lineno"> 2248</span>&#160;        acc30 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc30, (char8)a3, b0.s01234567, FIXED_POINT_POSITION);</div><div class="line"><a name="l02249"></a><span class="lineno"> 2249</span>&#160;        acc31 = <a class="code" href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a>(acc31, (char8)a3, b0.s89ABCDEF, FIXED_POINT_POSITION);</div><div class="line"><a name="l02250"></a><span class="lineno"> 2250</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02251"></a><span class="lineno"> 2251</span>&#160;    }</div><div class="line"><a name="l02252"></a><span class="lineno"> 2252</span>&#160;</div><div class="line"><a name="l02253"></a><span class="lineno"> 2253</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02254"></a><span class="lineno"> 2254</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02255"></a><span class="lineno"> 2255</span>&#160;</div><div class="line"><a name="l02256"></a><span class="lineno"> 2256</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02257"></a><span class="lineno"> 2257</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l02258"></a><span class="lineno"> 2258</span>&#160;</div><div class="line"><a name="l02259"></a><span class="lineno"> 2259</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02260"></a><span class="lineno"> 2260</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l02261"></a><span class="lineno"> 2261</span>&#160;</div><div class="line"><a name="l02262"></a><span class="lineno"> 2262</span>&#160;    <span class="comment">// Multiply by the weight of matrix product and store the result</span></div><div class="line"><a name="l02263"></a><span class="lineno"> 2263</span>&#160;    char16 acc_qs8;</div><div class="line"><a name="l02264"></a><span class="lineno"> 2264</span>&#160;    acc_qs8 = convert_char16_sat((short16)(acc00, acc01));</div><div class="line"><a name="l02265"></a><span class="lineno"> 2265</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02266"></a><span class="lineno"> 2266</span>&#160;    acc_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02267"></a><span class="lineno"> 2267</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02268"></a><span class="lineno"> 2268</span>&#160;    vstore16(acc_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l02269"></a><span class="lineno"> 2269</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02270"></a><span class="lineno"> 2270</span>&#160;    acc_qs8 = convert_char16_sat((short16)(acc10, acc11));</div><div class="line"><a name="l02271"></a><span class="lineno"> 2271</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02272"></a><span class="lineno"> 2272</span>&#160;    acc_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02273"></a><span class="lineno"> 2273</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02274"></a><span class="lineno"> 2274</span>&#160;    vstore16(acc_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l02275"></a><span class="lineno"> 2275</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02276"></a><span class="lineno"> 2276</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02277"></a><span class="lineno"> 2277</span>&#160;    acc_qs8 = convert_char16_sat((short16)(acc20, acc21));</div><div class="line"><a name="l02278"></a><span class="lineno"> 2278</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02279"></a><span class="lineno"> 2279</span>&#160;    acc_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02280"></a><span class="lineno"> 2280</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02281"></a><span class="lineno"> 2281</span>&#160;    vstore16(acc_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l02282"></a><span class="lineno"> 2282</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02283"></a><span class="lineno"> 2283</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02284"></a><span class="lineno"> 2284</span>&#160;    acc_qs8 = convert_char16_sat((short16)(acc30, acc31));</div><div class="line"><a name="l02285"></a><span class="lineno"> 2285</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02286"></a><span class="lineno"> 2286</span>&#160;    acc_qs8 = <a class="code" href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a>(acc_qs8, (char16)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02287"></a><span class="lineno"> 2287</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02288"></a><span class="lineno"> 2288</span>&#160;    vstore16(acc_qs8, 0, (__global <span class="keywordtype">char</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l02289"></a><span class="lineno"> 2289</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02290"></a><span class="lineno"> 2290</span>&#160;}</div><div class="line"><a name="l02291"></a><span class="lineno"> 2291</span>&#160;</div><div class="line"><a name="l02321"></a><span class="lineno"> 2321</span>&#160;__kernel <span class="keywordtype">void</span> gemm_mm_qs16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02322"></a><span class="lineno"> 2322</span>&#160;                           <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l02323"></a><span class="lineno"> 2323</span>&#160;                           <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l02324"></a><span class="lineno"> 2324</span>&#160;                           uint src0_stride_z,</div><div class="line"><a name="l02325"></a><span class="lineno"> 2325</span>&#160;                           uint src1_stride_z,</div><div class="line"><a name="l02326"></a><span class="lineno"> 2326</span>&#160;                           uint dst_stride_z)</div><div class="line"><a name="l02327"></a><span class="lineno"> 2327</span>&#160;{</div><div class="line"><a name="l02328"></a><span class="lineno"> 2328</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l02329"></a><span class="lineno"> 2329</span>&#160;</div><div class="line"><a name="l02330"></a><span class="lineno"> 2330</span>&#160;    <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l02331"></a><span class="lineno"> 2331</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l02332"></a><span class="lineno"> 2332</span>&#160;</div><div class="line"><a name="l02333"></a><span class="lineno"> 2333</span>&#160;    <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l02334"></a><span class="lineno"> 2334</span>&#160;    src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l02335"></a><span class="lineno"> 2335</span>&#160;</div><div class="line"><a name="l02336"></a><span class="lineno"> 2336</span>&#160;    <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l02337"></a><span class="lineno"> 2337</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(short);</div><div class="line"><a name="l02338"></a><span class="lineno"> 2338</span>&#160;</div><div class="line"><a name="l02339"></a><span class="lineno"> 2339</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02340"></a><span class="lineno"> 2340</span>&#160;    src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l02341"></a><span class="lineno"> 2341</span>&#160;</div><div class="line"><a name="l02342"></a><span class="lineno"> 2342</span>&#160;<span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02343"></a><span class="lineno"> 2343</span>&#160;    <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l02344"></a><span class="lineno"> 2344</span>&#160;    src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l02345"></a><span class="lineno"> 2345</span>&#160;<span class="preprocessor">#else  // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02346"></a><span class="lineno"> 2346</span>&#160;    src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l02347"></a><span class="lineno"> 2347</span>&#160;<span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02348"></a><span class="lineno"> 2348</span>&#160;</div><div class="line"><a name="l02349"></a><span class="lineno"> 2349</span>&#160;    <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (COLS_A * <span class="keyword">sizeof</span>(short));</div><div class="line"><a name="l02350"></a><span class="lineno"> 2350</span>&#160;</div><div class="line"><a name="l02351"></a><span class="lineno"> 2351</span>&#160;    int8 acc0 = 0;</div><div class="line"><a name="l02352"></a><span class="lineno"> 2352</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02353"></a><span class="lineno"> 2353</span>&#160;    int8 acc1 = 0;</div><div class="line"><a name="l02354"></a><span class="lineno"> 2354</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02355"></a><span class="lineno"> 2355</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02356"></a><span class="lineno"> 2356</span>&#160;    int8 acc2 = 0;</div><div class="line"><a name="l02357"></a><span class="lineno"> 2357</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02358"></a><span class="lineno"> 2358</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02359"></a><span class="lineno"> 2359</span>&#160;    int8 acc3 = 0;</div><div class="line"><a name="l02360"></a><span class="lineno"> 2360</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02361"></a><span class="lineno"> 2361</span>&#160;</div><div class="line"><a name="l02362"></a><span class="lineno"> 2362</span>&#160;    <span class="comment">// This for loop performs 4 accumulations per iteration</span></div><div class="line"><a name="l02363"></a><span class="lineno"> 2363</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt;= (end_row_vec_a - 2 * (int)<span class="keyword">sizeof</span>(<span class="keywordtype">short</span>)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(<span class="keywordtype">short</span>), 2 * src1_stride_y))</div><div class="line"><a name="l02364"></a><span class="lineno"> 2364</span>&#160;    {</div><div class="line"><a name="l02365"></a><span class="lineno"> 2365</span>&#160;        short2 a0 = vload2(0, (__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l02366"></a><span class="lineno"> 2366</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02367"></a><span class="lineno"> 2367</span>&#160;        short2 a1 = vload2(0, (__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l02368"></a><span class="lineno"> 2368</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02369"></a><span class="lineno"> 2369</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02370"></a><span class="lineno"> 2370</span>&#160;        short2 a2 = vload2(0, (__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l02371"></a><span class="lineno"> 2371</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02372"></a><span class="lineno"> 2372</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02373"></a><span class="lineno"> 2373</span>&#160;        short2 a3 = vload2(0, (__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l02374"></a><span class="lineno"> 2374</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02375"></a><span class="lineno"> 2375</span>&#160;        short8 b0 = vload8(0, (__global <span class="keywordtype">short</span> *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));</div><div class="line"><a name="l02376"></a><span class="lineno"> 2376</span>&#160;        short8 b1 = vload8(0, (__global <span class="keywordtype">short</span> *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));</div><div class="line"><a name="l02377"></a><span class="lineno"> 2377</span>&#160;</div><div class="line"><a name="l02378"></a><span class="lineno"> 2378</span>&#160;        acc0 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc0, (short8)a0.s0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02379"></a><span class="lineno"> 2379</span>&#160;        acc0 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc0, (short8)a0.s1, b1, FIXED_POINT_POSITION);</div><div class="line"><a name="l02380"></a><span class="lineno"> 2380</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02381"></a><span class="lineno"> 2381</span>&#160;        acc1 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc1, (short8)a1.s0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02382"></a><span class="lineno"> 2382</span>&#160;        acc1 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc1, (short8)a1.s1, b1, FIXED_POINT_POSITION);</div><div class="line"><a name="l02383"></a><span class="lineno"> 2383</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02384"></a><span class="lineno"> 2384</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02385"></a><span class="lineno"> 2385</span>&#160;        acc2 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc2, (short8)a2.s0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02386"></a><span class="lineno"> 2386</span>&#160;        acc2 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc2, (short8)a2.s1, b1, FIXED_POINT_POSITION);</div><div class="line"><a name="l02387"></a><span class="lineno"> 2387</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02388"></a><span class="lineno"> 2388</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02389"></a><span class="lineno"> 2389</span>&#160;        acc3 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc3, (short8)a3.s0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02390"></a><span class="lineno"> 2390</span>&#160;        acc3 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc3, (short8)a3.s1, b1, FIXED_POINT_POSITION);</div><div class="line"><a name="l02391"></a><span class="lineno"> 2391</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02392"></a><span class="lineno"> 2392</span>&#160;    }</div><div class="line"><a name="l02393"></a><span class="lineno"> 2393</span>&#160;</div><div class="line"><a name="l02394"></a><span class="lineno"> 2394</span>&#160;    <span class="comment">// Left-over accumulations</span></div><div class="line"><a name="l02395"></a><span class="lineno"> 2395</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt; end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(<span class="keywordtype">short</span>), src1_stride_y))</div><div class="line"><a name="l02396"></a><span class="lineno"> 2396</span>&#160;    {</div><div class="line"><a name="l02397"></a><span class="lineno"> 2397</span>&#160;        <span class="keywordtype">short</span> a0 = *((__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l02398"></a><span class="lineno"> 2398</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02399"></a><span class="lineno"> 2399</span>&#160;        <span class="keywordtype">short</span> a1 = *((__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l02400"></a><span class="lineno"> 2400</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02401"></a><span class="lineno"> 2401</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02402"></a><span class="lineno"> 2402</span>&#160;        <span class="keywordtype">short</span> a2 = *((__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l02403"></a><span class="lineno"> 2403</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02404"></a><span class="lineno"> 2404</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02405"></a><span class="lineno"> 2405</span>&#160;        <span class="keywordtype">short</span> a3 = *((__global <span class="keywordtype">short</span> *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l02406"></a><span class="lineno"> 2406</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02407"></a><span class="lineno"> 2407</span>&#160;        short8 b0 = vload8(0, (__global <span class="keywordtype">short</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02408"></a><span class="lineno"> 2408</span>&#160;</div><div class="line"><a name="l02409"></a><span class="lineno"> 2409</span>&#160;        acc0 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc0, (short8)a0, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02410"></a><span class="lineno"> 2410</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02411"></a><span class="lineno"> 2411</span>&#160;        acc1 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc1, (short8)a1, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02412"></a><span class="lineno"> 2412</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02413"></a><span class="lineno"> 2413</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02414"></a><span class="lineno"> 2414</span>&#160;        acc2 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc2, (short8)a2, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02415"></a><span class="lineno"> 2415</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02416"></a><span class="lineno"> 2416</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02417"></a><span class="lineno"> 2417</span>&#160;        acc3 = <a class="code" href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a>(acc3, (short8)a3, b0, FIXED_POINT_POSITION);</div><div class="line"><a name="l02418"></a><span class="lineno"> 2418</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02419"></a><span class="lineno"> 2419</span>&#160;    }</div><div class="line"><a name="l02420"></a><span class="lineno"> 2420</span>&#160;</div><div class="line"><a name="l02421"></a><span class="lineno"> 2421</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02422"></a><span class="lineno"> 2422</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02423"></a><span class="lineno"> 2423</span>&#160;</div><div class="line"><a name="l02424"></a><span class="lineno"> 2424</span>&#160;    <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02425"></a><span class="lineno"> 2425</span>&#160;    __global uchar *dst_addr = <a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0);</div><div class="line"><a name="l02426"></a><span class="lineno"> 2426</span>&#160;</div><div class="line"><a name="l02427"></a><span class="lineno"> 2427</span>&#160;    <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02428"></a><span class="lineno"> 2428</span>&#160;    dst_addr += get_global_id(2) * dst_stride_z;</div><div class="line"><a name="l02429"></a><span class="lineno"> 2429</span>&#160;</div><div class="line"><a name="l02430"></a><span class="lineno"> 2430</span>&#160;    <span class="comment">// Multiply by the weight of matrix product and store the result</span></div><div class="line"><a name="l02431"></a><span class="lineno"> 2431</span>&#160;    short8 acc_qs16;</div><div class="line"><a name="l02432"></a><span class="lineno"> 2432</span>&#160;    acc_qs16 = convert_short8_sat(acc0);</div><div class="line"><a name="l02433"></a><span class="lineno"> 2433</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02434"></a><span class="lineno"> 2434</span>&#160;    acc_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02435"></a><span class="lineno"> 2435</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02436"></a><span class="lineno"> 2436</span>&#160;    vstore8(acc_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l02437"></a><span class="lineno"> 2437</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02438"></a><span class="lineno"> 2438</span>&#160;    acc_qs16 = convert_short8_sat(acc1);</div><div class="line"><a name="l02439"></a><span class="lineno"> 2439</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02440"></a><span class="lineno"> 2440</span>&#160;    acc_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02441"></a><span class="lineno"> 2441</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02442"></a><span class="lineno"> 2442</span>&#160;    vstore8(acc_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l02443"></a><span class="lineno"> 2443</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 1</span></div><div class="line"><a name="l02444"></a><span class="lineno"> 2444</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02445"></a><span class="lineno"> 2445</span>&#160;    acc_qs16 = convert_short8_sat(acc2);</div><div class="line"><a name="l02446"></a><span class="lineno"> 2446</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02447"></a><span class="lineno"> 2447</span>&#160;    acc_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02448"></a><span class="lineno"> 2448</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02449"></a><span class="lineno"> 2449</span>&#160;    vstore8(acc_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l02450"></a><span class="lineno"> 2450</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 2</span></div><div class="line"><a name="l02451"></a><span class="lineno"> 2451</span>&#160;<span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02452"></a><span class="lineno"> 2452</span>&#160;    acc_qs16 = convert_short8_sat(acc3);</div><div class="line"><a name="l02453"></a><span class="lineno"> 2453</span>&#160;<span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02454"></a><span class="lineno"> 2454</span>&#160;    acc_qs16 = <a class="code" href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a>(acc_qs16, (short8)ALPHA, FIXED_POINT_POSITION);</div><div class="line"><a name="l02455"></a><span class="lineno"> 2455</span>&#160;<span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02456"></a><span class="lineno"> 2456</span>&#160;    vstore8(acc_qs16, 0, (__global <span class="keywordtype">short</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l02457"></a><span class="lineno"> 2457</span>&#160;<span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y &gt; 3</span></div><div class="line"><a name="l02458"></a><span class="lineno"> 2458</span>&#160;}</div><div class="line"><a name="l02459"></a><span class="lineno"> 2459</span>&#160;<span class="preprocessor">#endif // defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l02460"></a><span class="lineno"> 2460</span>&#160;<span class="preprocessor">#endif // defined(COLS_A) &amp;&amp; defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) &amp;&amp; (NUM_ELEMS_PROCESSED_PER_THREAD_Y)</span></div><div class="line"><a name="l02461"></a><span class="lineno"> 2461</span>&#160;</div><div class="line"><a name="l02462"></a><span class="lineno"> 2462</span>&#160;<span class="preprocessor">#if defined(BETA)</span></div><div class="line"><a name="l02463"></a><span class="lineno"> 2463</span>&#160;</div><div class="line"><a name="l02480"></a><span class="lineno"> 2480</span>&#160;__kernel <span class="keywordtype">void</span> gemm_ma_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l02481"></a><span class="lineno"> 2481</span>&#160;                          <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02482"></a><span class="lineno"> 2482</span>&#160;{</div><div class="line"><a name="l02483"></a><span class="lineno"> 2483</span>&#160;    <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02484"></a><span class="lineno"> 2484</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l02485"></a><span class="lineno"> 2485</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02486"></a><span class="lineno"> 2486</span>&#160;</div><div class="line"><a name="l02487"></a><span class="lineno"> 2487</span>&#160;    <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l02488"></a><span class="lineno"> 2488</span>&#160;    float4 alpha_ab = vload4(0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02489"></a><span class="lineno"> 2489</span>&#160;</div><div class="line"><a name="l02490"></a><span class="lineno"> 2490</span>&#160;    <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l02491"></a><span class="lineno"> 2491</span>&#160;    float4 <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a497347573faa3778743ddf277f861094">c</a> = vload4(0, (__global <span class="keywordtype">float</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02492"></a><span class="lineno"> 2492</span>&#160;</div><div class="line"><a name="l02493"></a><span class="lineno"> 2493</span>&#160;    <span class="comment">// Computes alpha * axb + beta * c</span></div><div class="line"><a name="l02494"></a><span class="lineno"> 2494</span>&#160;    float4 out = alpha_ab + (float4)BETA * c;</div><div class="line"><a name="l02495"></a><span class="lineno"> 2495</span>&#160;</div><div class="line"><a name="l02496"></a><span class="lineno"> 2496</span>&#160;    <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l02497"></a><span class="lineno"> 2497</span>&#160;    vstore4(out, 0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02498"></a><span class="lineno"> 2498</span>&#160;}</div><div class="line"><a name="l02499"></a><span class="lineno"> 2499</span>&#160;</div><div class="line"><a name="l02517"></a><span class="lineno"> 2517</span>&#160;__kernel <span class="keywordtype">void</span> gemm_ma_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l02518"></a><span class="lineno"> 2518</span>&#160;                          <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02519"></a><span class="lineno"> 2519</span>&#160;{</div><div class="line"><a name="l02520"></a><span class="lineno"> 2520</span>&#160;    <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02521"></a><span class="lineno"> 2521</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l02522"></a><span class="lineno"> 2522</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02523"></a><span class="lineno"> 2523</span>&#160;</div><div class="line"><a name="l02524"></a><span class="lineno"> 2524</span>&#160;    <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l02525"></a><span class="lineno"> 2525</span>&#160;    half8 alpha_ab = vload8(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="l02526"></a><span class="lineno"> 2526</span>&#160;</div><div class="line"><a name="l02527"></a><span class="lineno"> 2527</span>&#160;    <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l02528"></a><span class="lineno"> 2528</span>&#160;    half8 c = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02529"></a><span class="lineno"> 2529</span>&#160;</div><div class="line"><a name="l02530"></a><span class="lineno"> 2530</span>&#160;    <span class="comment">// Computes alpha * axb + beta * c</span></div><div class="line"><a name="l02531"></a><span class="lineno"> 2531</span>&#160;    half8 out = alpha_ab + (half8)BETA * c;</div><div class="line"><a name="l02532"></a><span class="lineno"> 2532</span>&#160;</div><div class="line"><a name="l02533"></a><span class="lineno"> 2533</span>&#160;    <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l02534"></a><span class="lineno"> 2534</span>&#160;    vstore8(out, 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="l02535"></a><span class="lineno"> 2535</span>&#160;}</div><div class="line"><a name="l02536"></a><span class="lineno"> 2536</span>&#160;</div><div class="line"><a name="l02537"></a><span class="lineno"> 2537</span>&#160;<span class="preprocessor">#if defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l02538"></a><span class="lineno"> 2538</span>&#160;</div><div class="line"><a name="l02557"></a><span class="lineno"> 2557</span>&#160;__kernel <span class="keywordtype">void</span> gemm_ma_qs8(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l02558"></a><span class="lineno"> 2558</span>&#160;                          <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02559"></a><span class="lineno"> 2559</span>&#160;{</div><div class="line"><a name="l02560"></a><span class="lineno"> 2560</span>&#160;    <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02561"></a><span class="lineno"> 2561</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l02562"></a><span class="lineno"> 2562</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02563"></a><span class="lineno"> 2563</span>&#160;</div><div class="line"><a name="l02564"></a><span class="lineno"> 2564</span>&#160;    <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l02565"></a><span class="lineno"> 2565</span>&#160;    char16 alpha_ab = vload16(0, (__global <span class="keywordtype">char</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02566"></a><span class="lineno"> 2566</span>&#160;</div><div class="line"><a name="l02567"></a><span class="lineno"> 2567</span>&#160;    <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l02568"></a><span class="lineno"> 2568</span>&#160;    char16 c = vload16(0, (__global <span class="keywordtype">char</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02569"></a><span class="lineno"> 2569</span>&#160;</div><div class="line"><a name="l02570"></a><span class="lineno"> 2570</span>&#160;    <span class="comment">// Computes alpha * axb + beta * c</span></div><div class="line"><a name="l02571"></a><span class="lineno"> 2571</span>&#160;    char16 out = <a class="code" href="fixed__point_8h.xhtml#ad59f32bf4f61044a698ff4da02080e5a">mla_sat_qs8x16</a>(alpha_ab, (char16)BETA, c, FIXED_POINT_POSITION);</div><div class="line"><a name="l02572"></a><span class="lineno"> 2572</span>&#160;</div><div class="line"><a name="l02573"></a><span class="lineno"> 2573</span>&#160;    <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l02574"></a><span class="lineno"> 2574</span>&#160;    vstore16(out, 0, (__global <span class="keywordtype">char</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02575"></a><span class="lineno"> 2575</span>&#160;}</div><div class="line"><a name="l02576"></a><span class="lineno"> 2576</span>&#160;</div><div class="line"><a name="l02596"></a><span class="lineno"> 2596</span>&#160;__kernel <span class="keywordtype">void</span> gemm_ma_qs16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l02597"></a><span class="lineno"> 2597</span>&#160;                           <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02598"></a><span class="lineno"> 2598</span>&#160;{</div><div class="line"><a name="l02599"></a><span class="lineno"> 2599</span>&#160;    <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02600"></a><span class="lineno"> 2600</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l02601"></a><span class="lineno"> 2601</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02602"></a><span class="lineno"> 2602</span>&#160;</div><div class="line"><a name="l02603"></a><span class="lineno"> 2603</span>&#160;    <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l02604"></a><span class="lineno"> 2604</span>&#160;    short8 alpha_ab = vload8(0, (__global <span class="keywordtype">short</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02605"></a><span class="lineno"> 2605</span>&#160;</div><div class="line"><a name="l02606"></a><span class="lineno"> 2606</span>&#160;    <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l02607"></a><span class="lineno"> 2607</span>&#160;    short8 c = vload8(0, (__global <span class="keywordtype">short</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02608"></a><span class="lineno"> 2608</span>&#160;</div><div class="line"><a name="l02609"></a><span class="lineno"> 2609</span>&#160;    <span class="comment">// Computes alpha * axb + beta * c</span></div><div class="line"><a name="l02610"></a><span class="lineno"> 2610</span>&#160;    short8 out = <a class="code" href="fixed__point_8h.xhtml#a7a3c295a894e0529251bd4529b36a574">mla_sat_qs16x8</a>(alpha_ab, (short8)BETA, c, FIXED_POINT_POSITION);</div><div class="line"><a name="l02611"></a><span class="lineno"> 2611</span>&#160;</div><div class="line"><a name="l02612"></a><span class="lineno"> 2612</span>&#160;    <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l02613"></a><span class="lineno"> 2613</span>&#160;    vstore8(out, 0, (__global <span class="keywordtype">short</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02614"></a><span class="lineno"> 2614</span>&#160;}</div><div class="line"><a name="l02615"></a><span class="lineno"> 2615</span>&#160;<span class="preprocessor">#endif // defined(FIXED_POINT_POSITION)</span></div><div class="line"><a name="l02616"></a><span class="lineno"> 2616</span>&#160;<span class="preprocessor">#endif // defined(BETA)</span></div><div class="line"><a name="l02617"></a><span class="lineno"> 2617</span>&#160;</div><div class="line"><a name="l02618"></a><span class="lineno"> 2618</span>&#160;<span class="preprocessor">#if defined(WIDTH_VECTOR_A)</span></div><div class="line"><a name="l02619"></a><span class="lineno"> 2619</span>&#160;</div><div class="line"><a name="l02646"></a><span class="lineno"> 2646</span>&#160;__kernel <span class="keywordtype">void</span> gemm_lc_vm_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02647"></a><span class="lineno"> 2647</span>&#160;                             <a class="code" href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src1),</div><div class="line"><a name="l02648"></a><span class="lineno"> 2648</span>&#160;                             <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02649"></a><span class="lineno"> 2649</span>&#160;{</div><div class="line"><a name="l02650"></a><span class="lineno"> 2650</span>&#160;    <span class="keywordtype">int</span> idx = get_global_id(0) * 4;</div><div class="line"><a name="l02651"></a><span class="lineno"> 2651</span>&#160;    <span class="keywordtype">int</span> idy = get_global_id(1);</div><div class="line"><a name="l02652"></a><span class="lineno"> 2652</span>&#160;</div><div class="line"><a name="l02653"></a><span class="lineno"> 2653</span>&#160;    <span class="comment">// Compute the address for the vector A and matrix B</span></div><div class="line"><a name="l02654"></a><span class="lineno"> 2654</span>&#160;    int2 src_addr = ((int2)(src0_offset_first_element_in_bytes + src0_stride_y * idy, src1_offset_first_element_in_bytes + src1_stride_z * idy));</div><div class="line"><a name="l02655"></a><span class="lineno"> 2655</span>&#160;    src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l02656"></a><span class="lineno"> 2656</span>&#160;</div><div class="line"><a name="l02657"></a><span class="lineno"> 2657</span>&#160;    <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * <span class="keyword">sizeof</span>(float));</div><div class="line"><a name="l02658"></a><span class="lineno"> 2658</span>&#160;</div><div class="line"><a name="l02659"></a><span class="lineno"> 2659</span>&#160;    float4 acc = 0.0f;</div><div class="line"><a name="l02660"></a><span class="lineno"> 2660</span>&#160;</div><div class="line"><a name="l02661"></a><span class="lineno"> 2661</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt;= (end_row_vec_a - 2 * (int)<span class="keyword">sizeof</span>(<span class="keywordtype">float</span>)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>), 2 * src1_stride_y))</div><div class="line"><a name="l02662"></a><span class="lineno"> 2662</span>&#160;    {</div><div class="line"><a name="l02663"></a><span class="lineno"> 2663</span>&#160;        float2 a0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l02664"></a><span class="lineno"> 2664</span>&#160;        float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02665"></a><span class="lineno"> 2665</span>&#160;        float4 b1 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1 + src1_stride_y));</div><div class="line"><a name="l02666"></a><span class="lineno"> 2666</span>&#160;</div><div class="line"><a name="l02667"></a><span class="lineno"> 2667</span>&#160;        acc += b0 * (float4)a0.s0;</div><div class="line"><a name="l02668"></a><span class="lineno"> 2668</span>&#160;        acc += b1 * (float4)a0.s1;</div><div class="line"><a name="l02669"></a><span class="lineno"> 2669</span>&#160;    }</div><div class="line"><a name="l02670"></a><span class="lineno"> 2670</span>&#160;</div><div class="line"><a name="l02671"></a><span class="lineno"> 2671</span>&#160;    <span class="keywordflow">for</span>(; src_addr.s0 &lt; end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(<span class="keywordtype">float</span>), src1_stride_y))</div><div class="line"><a name="l02672"></a><span class="lineno"> 2672</span>&#160;    {</div><div class="line"><a name="l02673"></a><span class="lineno"> 2673</span>&#160;        <span class="keywordtype">float</span>  a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l02674"></a><span class="lineno"> 2674</span>&#160;        float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02675"></a><span class="lineno"> 2675</span>&#160;</div><div class="line"><a name="l02676"></a><span class="lineno"> 2676</span>&#160;        acc += b0 * (float4)a0;</div><div class="line"><a name="l02677"></a><span class="lineno"> 2677</span>&#160;    }</div><div class="line"><a name="l02678"></a><span class="lineno"> 2678</span>&#160;</div><div class="line"><a name="l02679"></a><span class="lineno"> 2679</span>&#160;    <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02680"></a><span class="lineno"> 2680</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02681"></a><span class="lineno"> 2681</span>&#160;</div><div class="line"><a name="l02682"></a><span class="lineno"> 2682</span>&#160;    vstore4(acc, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&amp;dst, 0, 0)));</div><div class="line"><a name="l02683"></a><span class="lineno"> 2683</span>&#160;}</div><div class="line"><a name="l02684"></a><span class="lineno"> 2684</span>&#160;<span class="preprocessor">#endif // defined(WIDTH_VECTOR_A)</span></div><div class="line"><a name="l02685"></a><span class="lineno"> 2685</span>&#160;</div><div class="line"><a name="l02702"></a><span class="lineno"> 2702</span>&#160;<span class="preprocessor">#if defined(DATA_TYPE) &amp;&amp; defined(VECTOR_SIZE)</span></div><div class="line"><a name="l02703"></a><span class="lineno"> 2703</span>&#160;__kernel <span class="keywordtype">void</span> gemm_accumulate_biases(</div><div class="line"><a name="l02704"></a><span class="lineno"> 2704</span>&#160;    <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l02705"></a><span class="lineno"> 2705</span>&#160;    <a class="code" href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l02706"></a><span class="lineno"> 2706</span>&#160;{</div><div class="line"><a name="l02707"></a><span class="lineno"> 2707</span>&#160;    <a class="code" href="struct_image.xhtml">Image</a>  accum  = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l02708"></a><span class="lineno"> 2708</span>&#160;    <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l02709"></a><span class="lineno"> 2709</span>&#160;</div><div class="line"><a name="l02710"></a><span class="lineno"> 2710</span>&#160;    <span class="comment">// Vector size, i.e. number of vector elements.</span></div><div class="line"><a name="l02711"></a><span class="lineno"> 2711</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, <a class="code" href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a>)</div><div class="line"><a name="l02712"></a><span class="lineno"> 2712</span>&#160;    accum_value = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(<a class="code" href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a>)(0, (__global DATA_TYPE *)accum.ptr);</div><div class="line"><a name="l02713"></a><span class="lineno"> 2713</span>&#160;    <a class="code" href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, VECTOR_SIZE)</div><div class="line"><a name="l02714"></a><span class="lineno"> 2714</span>&#160;    biases_value = <a class="code" href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(VECTOR_SIZE)(0, (__global DATA_TYPE *)biases.ptr);</div><div class="line"><a name="l02715"></a><span class="lineno"> 2715</span>&#160;<span class="preprocessor">#ifdef FIXED_POINT_POSITION</span></div><div class="line"><a name="l02716"></a><span class="lineno"> 2716</span>&#160;    accum_value = <a class="code" href="fixed__point_8h.xhtml#a6b1acbaff6cb3bad4edda9a93dac1f9e">ADD_SAT_OP_EXPAND</a>(biases_value, accum_value, DATA_TYPE, VECTOR_SIZE);</div><div class="line"><a name="l02717"></a><span class="lineno"> 2717</span>&#160;<span class="preprocessor">#else  // FIXED_POINT_POSITION</span></div><div class="line"><a name="l02718"></a><span class="lineno"> 2718</span>&#160;    accum_value = biases_value + accum_value;</div><div class="line"><a name="l02719"></a><span class="lineno"> 2719</span>&#160;<span class="preprocessor">#endif // FIXED_POINT_POSITION</span></div><div class="line"><a name="l02720"></a><span class="lineno"> 2720</span>&#160;    <span class="comment">// Store result in the accumulate buffer</span></div><div class="line"><a name="l02721"></a><span class="lineno"> 2721</span>&#160;    <a class="code" href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(VECTOR_SIZE)</div><div class="line"><a name="l02722"></a><span class="lineno"> 2722</span>&#160;    (accum_value, 0, (__global DATA_TYPE *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02723"></a><span class="lineno"> 2723</span>&#160;}</div><div class="line"><a name="l02724"></a><span class="lineno"> 2724</span>&#160;<span class="preprocessor">#endif // defined(DATA_TYPE) &amp;&amp; defined(VECTOR_SIZE)</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="helpers_8h_source.xhtml#l00134">helpers.h:134</a></div></div>
 121 <div class="ttc" id="fixed__point_8h_xhtml_a286a2ecb07602fd7268d31c6e88e62a6"><div class="ttname"><a href="fixed__point_8h.xhtml#a286a2ecb07602fd7268d31c6e88e62a6">mul_sat_qs8x16</a></div><div class="ttdeci">qs8x16 mul_sat_qs8x16(qs8x16 VopA, qs8x16 VopB, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00249">fixed_point.h:249</a></div></div>
 122 <div class="ttc" id="fixed__point_8h_xhtml_a93fefd4972b6b9440c8371509b46d19e"><div class="ttname"><a href="fixed__point_8h.xhtml#a93fefd4972b6b9440c8371509b46d19e">mul_sat_qs16x8</a></div><div class="ttdeci">qs16x8 mul_sat_qs16x8(qs16x8 VopA, qs16x8 VopB, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00254">fixed_point.h:254</a></div></div>
 123 <div class="ttc" id="helpers_8h_xhtml_a287e2fc366c312b468382c95bb90f91f"><div class="ttname"><a href="helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a></div><div class="ttdeci">#define VLOAD(size)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00042">helpers.h:42</a></div></div>
 124 <div class="ttc" id="helpers_8h_xhtml_a31c8c760f08fb1a331b16b7c204321dc"><div class="ttname"><a href="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="helpers_8h_source.xhtml#l00119">helpers.h:119</a></div></div>
 125 <div class="ttc" id="helpers_8h_xhtml_a527bfdf5eeb306f1cf01c4a8e29f38e0"><div class="ttname"><a href="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="helpers_8h_source.xhtml#l00098">helpers.h:98</a></div></div>
 126 <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>
 127 <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>
 128 <div class="ttc" id="helpers_8h_xhtml_a22f42fcf2077d951271df83b55c1a71a"><div class="ttname"><a href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a></div><div class="ttdeci">#define IMAGE_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00068">helpers.h:68</a></div></div>
 129 <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="helpers_8h_source.xhtml#l00151">helpers.h:151</a></div></div>
 130 <div class="ttc" id="fixed__point_8h_xhtml_ad59f32bf4f61044a698ff4da02080e5a"><div class="ttname"><a href="fixed__point_8h.xhtml#ad59f32bf4f61044a698ff4da02080e5a">mla_sat_qs8x16</a></div><div class="ttdeci">qs8x16 mla_sat_qs8x16(qs8x16 VopA, qs8x16 VopB, qs8x16 VopC, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00275">fixed_point.h:275</a></div></div>
 131 <div class="ttc" id="helpers_8h_xhtml_a009469e4d9b8fce3b6d5e97d2077827d"><div class="ttname"><a href="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="helpers_8h_source.xhtml#l00303">helpers.h:303</a></div></div>
 132 <div class="ttc" id="helpers_8h_xhtml_acb282042d1edeeaa3cc979a206f78b54"><div class="ttname"><a href="helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a></div><div class="ttdeci">#define VSTORE(size)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00045">helpers.h:45</a></div></div>
 133 <div class="ttc" id="fixed__point_8h_xhtml_a7a3c295a894e0529251bd4529b36a574"><div class="ttname"><a href="fixed__point_8h.xhtml#a7a3c295a894e0529251bd4529b36a574">mla_sat_qs16x8</a></div><div class="ttdeci">qs16x8 mla_sat_qs16x8(qs16x8 VopA, qs16x8 VopB, qs16x8 VopC, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00276">fixed_point.h:276</a></div></div>
 134 <div class="ttc" id="helpers_8h_xhtml"><div class="ttname"><a href="helpers_8h.xhtml">helpers.h</a></div></div>
 135 <div class="ttc" id="helpers_8h_xhtml_aebe814363556c244be043b13e7969197"><div class="ttname"><a href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_IMAGE_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00104">helpers.h:104</a></div></div>
 136 <div class="ttc" id="softmax__layer_8cl_xhtml_a7c78836761fa3b5b124efea237dac70f"><div class="ttname"><a href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a></div><div class="ttdeci">#define VECTOR_SIZE</div><div class="ttdef"><b>Definition:</b> <a href="softmax__layer_8cl_source.xhtml#l00079">softmax_layer.cl:79</a></div></div>
 137 <div class="ttc" id="fixed__point_8h_xhtml_a50b6bfd3a44a0ac605e88135e230d1e5"><div class="ttname"><a href="fixed__point_8h.xhtml#a50b6bfd3a44a0ac605e88135e230d1e5">mlal_sat_qs16x8</a></div><div class="ttdeci">qs32x8 mlal_sat_qs16x8(qs32x8 VopA, qs16x8 VopB, qs16x8 VopC, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00296">fixed_point.h:296</a></div></div>
 138 <div class="ttc" id="fixed__point_8h_xhtml"><div class="ttname"><a href="fixed__point_8h.xhtml">fixed_point.h</a></div></div>
 139 <div class="ttc" id="helpers_8h_xhtml_a40a6eb9f2a7712f08d6bb8ff6c9e6ca7"><div class="ttname"><a href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a></div><div class="ttdeci">#define VECTOR_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00062">helpers.h:62</a></div></div>
 140 <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_convolution_8cpp_source.xhtml#l00137">Convolution.cpp:137</a></div></div>
 141 <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="helpers_8h_source.xhtml#l00142">helpers.h:142</a></div></div>
 142 <div class="ttc" id="helpers_8h_xhtml_a6b83038822d1ae7ab619b684ed3b7fc0"><div class="ttname"><a href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a></div><div class="ttdeci">#define TENSOR3D_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00076">helpers.h:76</a></div></div>
 143 <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="helpers_8h_source.xhtml#l00144">helpers.h:144</a></div></div>
 144 <div class="ttc" id="fixed__point_8h_xhtml_a36f754c05b6fddf6df0d8d0a74f8159f"><div class="ttname"><a href="fixed__point_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a></div><div class="ttdeci">#define VEC_DATA_TYPE(type, size)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00093">fixed_point.h:93</a></div></div>
 145 <div class="ttc" id="fixed__point_8h_xhtml_a09f5d764b62b47a1894a4cb71a0a62f6"><div class="ttname"><a href="fixed__point_8h.xhtml#a09f5d764b62b47a1894a4cb71a0a62f6">mlal_sat_qs8x8</a></div><div class="ttdeci">qs16x8 mlal_sat_qs8x8(qs16x8 VopA, qs8x8 VopB, qs8x8 VopC, int fixed_point_position)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00295">fixed_point.h:295</a></div></div>
 146 <div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_a497347573faa3778743ddf277f861094"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#a497347573faa3778743ddf277f861094">arm_compute::test::validation::c</a></div><div class="ttdeci">CLTensor c</div><div class="ttdef"><b>Definition:</b> <a href="validation_2_c_l_2_g_e_m_m_8cpp_source.xhtml#l00123">GEMM.cpp:123</a></div></div>
 147 <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="helpers_8h_source.xhtml#l00153">helpers.h:153</a></div></div>
 148 <div class="ttc" id="fixed__point_8h_xhtml_a6b1acbaff6cb3bad4edda9a93dac1f9e"><div class="ttname"><a href="fixed__point_8h.xhtml#a6b1acbaff6cb3bad4edda9a93dac1f9e">ADD_SAT_OP_EXPAND</a></div><div class="ttdeci">#define ADD_SAT_OP_EXPAND(a, b, type, size)</div><div class="ttdef"><b>Definition:</b> <a href="fixed__point_8h_source.xhtml#l00178">fixed_point.h:178</a></div></div>
 149 <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="_c_l_2_convolution_8cpp_source.xhtml#l00147">Convolution.cpp:147</a></div></div>
 150 </div><!-- fragment --></div><!-- contents -->
 151 </div><!-- doc-content -->
 152 <!-- start footer part -->
 153 <div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
 154   <ul>
 155     <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="gemm_8cl.xhtml">gemm.cl</a></li>
 156     <li class="footer">Generated on Wed May 23 2018 11:36:36 for Compute Library by
 157     <a href="http://www.doxygen.org/index.html">
 158     <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.11 </li>
 159   </ul>
 160 </div>
 161 </body>
 162 </html>