Blame - documentation/gemm_8cl_source.xhtml - platform/external/ComputeLibrary

blob: d41cd5d5812b4051ba9c84b080fc79f2b959a74b [file] [log] [blame]

Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	1	<!-- HTML header for doxygen 1.8.9.1-->
				2	<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
				3	<html xmlns="http://www.w3.org/1999/xhtml">
				4	<head>
				5	<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
				6	<meta http-equiv="X-UA-Compatible" content="IE=9"/>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	7	<meta name="generator" content="Doxygen 1.8.13"/>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	8	<meta name="robots" content="NOINDEX, NOFOLLOW" /> <!-- Prevent indexing by search engines -->
Anthony Barbier	dbdab85	2017-06-23 15:42:00 +0100	[diff] [blame]	9	<title>Compute Library: src/core/CL/cl_kernels/gemm.cl Source File</title>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	10	<link href="tabs.css" rel="stylesheet" type="text/css"/>
				11	<script type="text/javascript" src="jquery.js"></script>
				12	<script type="text/javascript" src="dynsections.js"></script>
				13	<link href="navtree.css" rel="stylesheet" type="text/css"/>
				14	<script type="text/javascript" src="resize.js"></script>
Anthony Barbier	8140e1e	2017-12-14 23:48:46 +0000	[diff] [blame]	15	<script type="text/javascript" src="navtreedata.js"></script>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	16	<script type="text/javascript" src="navtree.js"></script>
				17	<script type="text/javascript">
				18	$(document).ready(initResizable);
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	19	</script>
				20	<link href="search/search.css" rel="stylesheet" type="text/css"/>
Anthony Barbier	8140e1e	2017-12-14 23:48:46 +0000	[diff] [blame]	21	<script type="text/javascript" src="search/searchdata.js"></script>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	22	<script type="text/javascript" src="search/search.js"></script>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	23	<script type="text/x-mathjax-config">
				24	MathJax.Hub.Config({
				25	extensions: ["tex2jax.js"],
				26	jax: ["input/TeX","output/HTML-CSS"],
				27	});
Anthony Barbier	8140e1e	2017-12-14 23:48:46 +0000	[diff] [blame]	28	</script><script type="text/javascript" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	29	<link href="doxygen.css" rel="stylesheet" type="text/css" />
				30	</head>
				31	<body>
				32	<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
				33	<div id="titlearea">
				34	<table cellspacing="0" cellpadding="0">
				35	<tbody>
				36	<tr style="height: 56px;">
				37	<td style="padding-left: 0.5em;">
Anthony Barbier	dbdab85	2017-06-23 15:42:00 +0100	[diff] [blame]	38	<div id="projectname">Compute Library
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	39	<span id="projectnumber">18.11</span>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	40	</div>
				41	</td>
				42	</tr>
				43	</tbody>
				44	</table>
				45	</div>
				46	<!-- end header part -->
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	47	<!-- Generated by Doxygen 1.8.13 -->
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	48	<script type="text/javascript">
				49	var searchBox = new SearchBox("searchBox", "search",false,'Search');
				50	</script>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	51	<script type="text/javascript" src="menudata.js"></script>
				52	<script type="text/javascript" src="menu.js"></script>
				53	<script type="text/javascript">
				54	$(function() {
				55	initMenu('',true,false,'search.php','Search');
				56	$(document).ready(function() { init_search(); });
				57	});
				58	</script>
				59	<div id="main-nav"></div>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	60	</div><!-- top -->
				61	<div id="side-nav" class="ui-resizable side-nav-resizable">
				62	<div id="nav-tree">
				63	<div id="nav-tree-contents">
				64	<div id="nav-sync" class="sync"></div>
				65	</div>
				66	</div>
				67	<div id="splitbar" style="-moz-user-select:none;"
				68	class="ui-resizable-handle">
				69	</div>
				70	</div>
				71	<script type="text/javascript">
				72	$(document).ready(function(){initNavTree('gemm_8cl_source.xhtml','');});
				73	</script>
				74	<div id="doc-content">
				75	<!-- window showing the filter options -->
				76	<div id="MSearchSelectWindow"
				77	onmouseover="return searchBox.OnSearchSelectShow()"
				78	onmouseout="return searchBox.OnSearchSelectHide()"
				79	onkeydown="return searchBox.OnSearchSelectKey(event)">
Anthony Barbier	8140e1e	2017-12-14 23:48:46 +0000	[diff] [blame]	80	</div>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	81
				82	<!-- iframe showing the search results (closed by default) -->
				83	<div id="MSearchResultsWindow">
				84	<iframe src="javascript:void(0)" frameborder="0"
				85	name="MSearchResults" id="MSearchResults">
				86	</iframe>
				87	</div>
				88
				89	<div class="header">
				90	<div class="headertitle">
				91	<div class="title">gemm.cl</div> </div>
				92	</div><!--header-->
				93	<div class="contents">
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	94	<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> <span class="comment">/</span></div><div class="line"><a name="l00002"></a><span class="lineno"> 2</span> <span class="comment"> Copyright (c) 2017-2018 ARM Limited.</span></div><div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> </span></div><div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> SPDX-License-Identifier: MIT</span></div><div class="line"><a name="l00005"></a><span class="lineno"> 5</span> <span class="comment"> </span></div><div class="line"><a name="l00006"></a><span class="lineno"> 6</span> <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> <span class="comment"> * of this software and associated documentation files (the "Software"), to</span></div><div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <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> <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> <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> <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> <span class="comment"> </span></div><div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <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> <span class="comment"> * copies or substantial portions of the Software.</span></div><div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="comment"> </span></div><div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="comment"> THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR</span></div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <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> <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> <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> <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> <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> <span class="comment"> * SOFTWARE.</span></div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="comment"> /</span></div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#include "<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml">helpers.h</a>"</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> </div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#if defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)</span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> </div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#if ELEMENT_SIZE == 1</span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#define DATA_TYPE uchar</span></div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="preprocessor">#elif ELEMENT_SIZE == 2</span></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#define DATA_TYPE ushort</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#elif ELEMENT_SIZE == 4</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> <span class="preprocessor">#define DATA_TYPE uint</span></div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="preprocessor">#else // ELEMENT_SIZE == 1</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="preprocessor">#error "Element size not supported"</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="preprocessor">#endif // ELEMENT_SIZE</span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> </div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> __kernel <span class="keywordtype">void</span> gemm_transpose1xW(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>),</div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>))</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> {</div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  uint z = get_global_id(2);</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span> </div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  <span class="comment">// Compute address for Matrix B - source</span></div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <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="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span> </div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  <span class="comment">// Compute address for Matrix B transposed - destination. X and Y are swapped</span></div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  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="l00072"></a><span class="lineno"> 72</span>  (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="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  dst_addr_in_bytes += z * dst_stride_z;</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span> </div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_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="l00078"></a><span class="lineno"> 78</span>  b0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(TRANSPOSE_W)(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> )src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span> </div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(TRANSPOSE_W)</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  (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="l00082"></a><span class="lineno"> 82</span> }</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span> <span class="preprocessor">#endif // defined(TRANSPOSE_W) && defined(MULT_TRANSPOSE1XW_WIDTH)</span></div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span> </div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span> <span class="preprocessor">#if defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(DATA_TYPE)</span></div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span> </div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> __kernel <span class="keywordtype">void</span> gemm_interleave4x4(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>),</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>)</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  ,</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  uint cross_plane_pad</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  )</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span> {</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  uint z = get_global_id(2);</div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span> </div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  <span class="comment">// Compute address for source tensor</span></div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <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="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span> </div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  <span class="comment">// Compute address for Matrix B transposed - destination. X and Y are swapped</span></div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  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="l00134"></a><span class="lineno"> 134</span>  (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="l00135"></a><span class="lineno"> 135</span> </div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  dst_addr_in_bytes += z * dst_stride_z;</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span> </div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  __global uchar input_ptr = src_ptr + src_offset_first_element_in_bytes + x 4 * <span class="keyword">sizeof</span>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>) + y * 4 * src_stride_y;</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span> </div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="comment">//</span></div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span> </div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  <span class="comment">// The plane (zin) is calculated dividing M (y 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(y * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span> </div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  zin = (cross_plane_pad src_stride_y);</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span> </div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// multiply src_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  input_ptr += z * src_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span> </div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  <span class="comment">// Load values from Matrix A</span></div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_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="l00170"></a><span class="lineno"> 170</span>  a0 = vload4(0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> )(input_ptr + 0 src_stride_y + zin.s0));</div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_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="l00172"></a><span class="lineno"> 172</span>  a1 = vload4(0, (__global DATA_TYPE )(input_ptr + 1 src_stride_y + zin.s1));</div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  a2 = vload4(0, (__global DATA_TYPE )(input_ptr + 2 src_stride_y + zin.s2));</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  a3 = vload4(0, (__global DATA_TYPE )(input_ptr + 3 src_stride_y + zin.s3));</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  __global uchar input_ptr = src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>;</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span> </div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="comment">// Load values from Matrix A</span></div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  a0 = vload4(0, (__global DATA_TYPE )(input_ptr + 0 * src_stride_y));</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  a1 = vload4(0, (__global DATA_TYPE )(input_ptr + 1 src_stride_y));</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  a2 = vload4(0, (__global DATA_TYPE )(input_ptr + 2 src_stride_y));</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  a3 = vload4(0, (__global DATA_TYPE )(input_ptr + 3 src_stride_y));</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span> </div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span> <span class="preprocessor">#if defined(UNROLL_BLOCK)</span></div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  vstore4(a0, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 0 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  vstore4(a1, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 4 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  vstore4(a2, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 8 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  vstore4(a3, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 12 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span> <span class="preprocessor">#else // defined(UNROLL_BLOCK)</span></div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4)</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  val0 = (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s0, a1.s0, a2.s0, a3.s0);</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  vstore4(val0, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 0 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span> </div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  val0 = (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s1, a1.s1, a2.s1, a3.s1);</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  vstore4(val0, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 4 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span> </div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  val0 = (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s2, a1.s2, a2.s2, a3.s2);</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  vstore4(val0, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 8 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span> </div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  val0 = (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 4))(a0.s3, a1.s3, a2.s3, a3.s3);</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  vstore4(val0, 0, ((__global DATA_TYPE )(dst_ptr + dst_addr_in_bytes) + 12 MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span> <span class="preprocessor">#endif // defined(UNROLL_BLOCK)</span></div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span> }</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span> <span class="preprocessor">#endif // defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(DATA_TYPE)</span></div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span> </div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span> <span class="preprocessor">#if defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT)</span></div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span> </div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span> __kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  uint src0_stride_z,</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  uint src1_stride_z,</div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  uint dst_stride_z</div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  ,</div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  uint cross_plane_pad</div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  )</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span> {</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span> </div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  <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="l00270"></a><span class="lineno"> 270</span>  <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="l00271"></a><span class="lineno"> 271</span> </div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  <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="l00275"></a><span class="lineno"> 275</span>  <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="l00276"></a><span class="lineno"> 276</span> </div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  <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="l00279"></a><span class="lineno"> 279</span>  src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span> </div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  __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="l00285"></a><span class="lineno"> 285</span>  __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="l00286"></a><span class="lineno"> 286</span> </div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  __global <span class="keywordtype">float</span> src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span> </div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span> </div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  float4 c00 = 0.0f;</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  float4 c10 = 0.0f;</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  float4 c20 = 0.0f;</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  float4 c30 = 0.0f;</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span> </div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  <span class="keywordflow">for</span>(; src_addr_b <= (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="l00300"></a><span class="lineno"> 300</span>  {</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> </div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span> </div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT);</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  b0 = vload4(0, src_addr_b + 4 * MULT_TRANSPOSE1XW_WIDTH);</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span> </div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span>  }</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span> </div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  <span class="keywordflow">for</span>(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  {</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span> </div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>  }</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span> </div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span> </div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  c00 = c00 * (float4)ALPHA;</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  c10 = c10 * (float4)ALPHA;</div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  c20 = c20 * (float4)ALPHA;</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  c30 = c30 * (float4)ALPHA;</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span> </div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span> </div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  <span class="comment">//</span></div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span> </div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span> </div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  zout = (cross_plane_pad dst_stride_y);</div><div class="line"><a name="l00368"></a><span class="lineno"> 368</span> </div><div class="line"><a name="l00369"></a><span class="lineno"> 369</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span> </div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  vstore4(c00, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  vstore4(c10, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  vstore4(c20, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  vstore4(c30, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span> </div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span> </div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  vstore4(c00, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y));</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  vstore4(c10, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y));</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  vstore4(c20, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y));</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span>  vstore4(c30, 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y));</div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span> }</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span> </div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span> __kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f32_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span>  uint src0_stride_z,</div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span>  uint src1_stride_z,</div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  uint dst_stride_z</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  ,</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  uint cross_plane_pad</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  )</div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span> {</div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span> </div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <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="l00448"></a><span class="lineno"> 448</span>  <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="l00449"></a><span class="lineno"> 449</span> </div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  <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="l00453"></a><span class="lineno"> 453</span>  <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="l00454"></a><span class="lineno"> 454</span> </div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>  <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="l00457"></a><span class="lineno"> 457</span>  src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span> </div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  __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="l00463"></a><span class="lineno"> 463</span>  __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="l00464"></a><span class="lineno"> 464</span> </div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span> </div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  <span class="keywordtype">float</span> c00 = 0.0f;</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  <span class="keywordtype">float</span> c01 = 0.0f;</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  <span class="keywordtype">float</span> c02 = 0.0f;</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  <span class="keywordtype">float</span> c03 = 0.0f;</div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  <span class="keywordtype">float</span> c10 = 0.0f;</div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  <span class="keywordtype">float</span> c11 = 0.0f;</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  <span class="keywordtype">float</span> c12 = 0.0f;</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  <span class="keywordtype">float</span> c13 = 0.0f;</div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  <span class="keywordtype">float</span> c20 = 0.0f;</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  <span class="keywordtype">float</span> c21 = 0.0f;</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  <span class="keywordtype">float</span> c22 = 0.0f;</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  <span class="keywordtype">float</span> c23 = 0.0f;</div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  <span class="keywordtype">float</span> c30 = 0.0f;</div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span>  <span class="keywordtype">float</span> c31 = 0.0f;</div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  <span class="keywordtype">float</span> c32 = 0.0f;</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  <span class="keywordtype">float</span> c33 = 0.0f;</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span> </div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span> <span class="preprocessor">#define COLS_MTX_B (COLS_B / (4 * MULT_TRANSPOSE1XW_WIDTH))</span></div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span> </div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span>  <span class="keywordflow">for</span>(; i <= (int)(COLS_MTX_B - 4); i += 4)</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  {</div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00494"></a><span class="lineno"> 494</span> </div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> </div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span> </div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span> </div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span> </div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span> </div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00519"></a><span class="lineno"> 519</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00520"></a><span class="lineno"> 520</span>  b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span> </div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00523"></a><span class="lineno"> 523</span>  src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00524"></a><span class="lineno"> 524</span> </div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span> </div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span> </div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span> </div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span> </div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span> </div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00551"></a><span class="lineno"> 551</span> </div><div class="line"><a name="l00552"></a><span class="lineno"> 552</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00556"></a><span class="lineno"> 556</span> </div><div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span> </div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00566"></a><span class="lineno"> 566</span> </div><div class="line"><a name="l00567"></a><span class="lineno"> 567</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00568"></a><span class="lineno"> 568</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00571"></a><span class="lineno"> 571</span> </div><div class="line"><a name="l00572"></a><span class="lineno"> 572</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00573"></a><span class="lineno"> 573</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00575"></a><span class="lineno"> 575</span> </div><div class="line"><a name="l00576"></a><span class="lineno"> 576</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00577"></a><span class="lineno"> 577</span>  src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span> </div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00583"></a><span class="lineno"> 583</span> </div><div class="line"><a name="l00584"></a><span class="lineno"> 584</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span> </div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span> </div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  }</div><div class="line"><a name="l00599"></a><span class="lineno"> 599</span> </div><div class="line"><a name="l00600"></a><span class="lineno"> 600</span>  <span class="keywordflow">for</span>(; i < (int)(COLS_MTX_B); ++i)</div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  {</div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span> </div><div class="line"><a name="l00606"></a><span class="lineno"> 606</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  src_addr_b += 4 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00608"></a><span class="lineno"> 608</span> </div><div class="line"><a name="l00609"></a><span class="lineno"> 609</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00611"></a><span class="lineno"> 611</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00613"></a><span class="lineno"> 613</span> </div><div class="line"><a name="l00614"></a><span class="lineno"> 614</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00615"></a><span class="lineno"> 615</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00618"></a><span class="lineno"> 618</span> </div><div class="line"><a name="l00619"></a><span class="lineno"> 619</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00621"></a><span class="lineno"> 621</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00622"></a><span class="lineno"> 622</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00623"></a><span class="lineno"> 623</span> </div><div class="line"><a name="l00624"></a><span class="lineno"> 624</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  }</div><div class="line"><a name="l00629"></a><span class="lineno"> 629</span> </div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span> </div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  c00 = c00 * ALPHA;</div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  c01 = c01 * ALPHA;</div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  c02 = c02 * ALPHA;</div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  c03 = c03 * ALPHA;</div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  c10 = c10 * ALPHA;</div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  c11 = c11 * ALPHA;</div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span>  c12 = c12 * ALPHA;</div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span>  c13 = c13 * ALPHA;</div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span>  c20 = c20 * ALPHA;</div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span>  c21 = c21 * ALPHA;</div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span>  c22 = c22 * ALPHA;</div><div class="line"><a name="l00646"></a><span class="lineno"> 646</span>  c23 = c23 * ALPHA;</div><div class="line"><a name="l00647"></a><span class="lineno"> 647</span>  c30 = c30 * ALPHA;</div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  c31 = c31 * ALPHA;</div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span>  c32 = c32 * ALPHA;</div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  c33 = c33 * ALPHA;</div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00652"></a><span class="lineno"> 652</span> </div><div class="line"><a name="l00653"></a><span class="lineno"> 653</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00654"></a><span class="lineno"> 654</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span> </div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span>  <span class="comment">//</span></div><div class="line"><a name="l00660"></a><span class="lineno"> 660</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l00666"></a><span class="lineno"> 666</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l00669"></a><span class="lineno"> 669</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00670"></a><span class="lineno"> 670</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span> </div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l00675"></a><span class="lineno"> 675</span> </div><div class="line"><a name="l00676"></a><span class="lineno"> 676</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>  zout = (cross_plane_pad dst_stride_y);</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span> </div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span> </div><div class="line"><a name="l00683"></a><span class="lineno"> 683</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00684"></a><span class="lineno"> 684</span>  vstore4((float4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span>  vstore4((float4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span>  vstore4((float4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  vstore4((float4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span> </div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00692"></a><span class="lineno"> 692</span> </div><div class="line"><a name="l00693"></a><span class="lineno"> 693</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00694"></a><span class="lineno"> 694</span>  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="l00695"></a><span class="lineno"> 695</span>  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="l00696"></a><span class="lineno"> 696</span>  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="l00697"></a><span class="lineno"> 697</span>  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="l00698"></a><span class="lineno"> 698</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00699"></a><span class="lineno"> 699</span> }</div><div class="line"><a name="l00700"></a><span class="lineno"> 700</span> </div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span> <span class="comment">// Undefine local defines</span></div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span> <span class="preprocessor">#undef COLS_MTX_B</span></div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span> </div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span> </div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span> __kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f16(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span>  uint src0_stride_z,</div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  uint src1_stride_z,</div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span>  uint dst_stride_z</div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  ,</div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span>  uint cross_plane_pad</div><div class="line"><a name="l00752"></a><span class="lineno"> 752</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00753"></a><span class="lineno"> 753</span>  )</div><div class="line"><a name="l00754"></a><span class="lineno"> 754</span> {</div><div class="line"><a name="l00755"></a><span class="lineno"> 755</span>  <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00756"></a><span class="lineno"> 756</span>  <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00757"></a><span class="lineno"> 757</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00758"></a><span class="lineno"> 758</span> </div><div class="line"><a name="l00759"></a><span class="lineno"> 759</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00760"></a><span class="lineno"> 760</span>  <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="l00761"></a><span class="lineno"> 761</span>  <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="l00762"></a><span class="lineno"> 762</span> </div><div class="line"><a name="l00763"></a><span class="lineno"> 763</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00764"></a><span class="lineno"> 764</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  <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="l00766"></a><span class="lineno"> 766</span>  <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="l00767"></a><span class="lineno"> 767</span> </div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  <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="l00770"></a><span class="lineno"> 770</span>  src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span> </div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span>  __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="l00776"></a><span class="lineno"> 776</span>  __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="l00777"></a><span class="lineno"> 777</span> </div><div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00779"></a><span class="lineno"> 779</span>  __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="l00780"></a><span class="lineno"> 780</span> </div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span> </div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span>  half8 c00 = 0.0f;</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span>  half8 c10 = 0.0f;</div><div class="line"><a name="l00787"></a><span class="lineno"> 787</span>  half8 c20 = 0.0f;</div><div class="line"><a name="l00788"></a><span class="lineno"> 788</span>  half8 c30 = 0.0f;</div><div class="line"><a name="l00789"></a><span class="lineno"> 789</span> </div><div class="line"><a name="l00790"></a><span class="lineno"> 790</span>  <span class="keywordflow">for</span>(; src_addr_b <= (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="l00791"></a><span class="lineno"> 791</span>  {</div><div class="line"><a name="l00792"></a><span class="lineno"> 792</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00793"></a><span class="lineno"> 793</span>  half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00794"></a><span class="lineno"> 794</span>  half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00795"></a><span class="lineno"> 795</span> </div><div class="line"><a name="l00796"></a><span class="lineno"> 796</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00797"></a><span class="lineno"> 797</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00798"></a><span class="lineno"> 798</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00799"></a><span class="lineno"> 799</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00800"></a><span class="lineno"> 800</span> </div><div class="line"><a name="l00801"></a><span class="lineno"> 801</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00802"></a><span class="lineno"> 802</span>  a0 = vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT);</div><div class="line"><a name="l00803"></a><span class="lineno"> 803</span>  b0 = vload8(0, src_addr_b + 8 * MULT_TRANSPOSE1XW_WIDTH);</div><div class="line"><a name="l00804"></a><span class="lineno"> 804</span> </div><div class="line"><a name="l00805"></a><span class="lineno"> 805</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00806"></a><span class="lineno"> 806</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00807"></a><span class="lineno"> 807</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00808"></a><span class="lineno"> 808</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00809"></a><span class="lineno"> 809</span>  }</div><div class="line"><a name="l00810"></a><span class="lineno"> 810</span> </div><div class="line"><a name="l00811"></a><span class="lineno"> 811</span>  <span class="keywordflow">for</span>(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00812"></a><span class="lineno"> 812</span>  {</div><div class="line"><a name="l00813"></a><span class="lineno"> 813</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00814"></a><span class="lineno"> 814</span>  half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00815"></a><span class="lineno"> 815</span>  half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l00816"></a><span class="lineno"> 816</span> </div><div class="line"><a name="l00817"></a><span class="lineno"> 817</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00818"></a><span class="lineno"> 818</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00819"></a><span class="lineno"> 819</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00820"></a><span class="lineno"> 820</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00821"></a><span class="lineno"> 821</span>  }</div><div class="line"><a name="l00822"></a><span class="lineno"> 822</span> </div><div class="line"><a name="l00823"></a><span class="lineno"> 823</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00824"></a><span class="lineno"> 824</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00825"></a><span class="lineno"> 825</span> </div><div class="line"><a name="l00826"></a><span class="lineno"> 826</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l00827"></a><span class="lineno"> 827</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00828"></a><span class="lineno"> 828</span>  c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l00829"></a><span class="lineno"> 829</span>  c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l00830"></a><span class="lineno"> 830</span>  c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l00831"></a><span class="lineno"> 831</span>  c30 = c30 * (half8)ALPHA;</div><div class="line"><a name="l00832"></a><span class="lineno"> 832</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l00833"></a><span class="lineno"> 833</span> </div><div class="line"><a name="l00834"></a><span class="lineno"> 834</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l00835"></a><span class="lineno"> 835</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l00836"></a><span class="lineno"> 836</span> </div><div class="line"><a name="l00837"></a><span class="lineno"> 837</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00838"></a><span class="lineno"> 838</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00839"></a><span class="lineno"> 839</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00840"></a><span class="lineno"> 840</span>  <span class="comment">//</span></div><div class="line"><a name="l00841"></a><span class="lineno"> 841</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00842"></a><span class="lineno"> 842</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l00843"></a><span class="lineno"> 843</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00844"></a><span class="lineno"> 844</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00845"></a><span class="lineno"> 845</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00846"></a><span class="lineno"> 846</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l00847"></a><span class="lineno"> 847</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l00848"></a><span class="lineno"> 848</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00849"></a><span class="lineno"> 849</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l00850"></a><span class="lineno"> 850</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l00851"></a><span class="lineno"> 851</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l00852"></a><span class="lineno"> 852</span> </div><div class="line"><a name="l00853"></a><span class="lineno"> 853</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00854"></a><span class="lineno"> 854</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00855"></a><span class="lineno"> 855</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l00856"></a><span class="lineno"> 856</span> </div><div class="line"><a name="l00857"></a><span class="lineno"> 857</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00858"></a><span class="lineno"> 858</span>  zout = (cross_plane_pad dst_stride_y);</div><div class="line"><a name="l00859"></a><span class="lineno"> 859</span> </div><div class="line"><a name="l00860"></a><span class="lineno"> 860</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00861"></a><span class="lineno"> 861</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00862"></a><span class="lineno"> 862</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00863"></a><span class="lineno"> 863</span> </div><div class="line"><a name="l00864"></a><span class="lineno"> 864</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l00865"></a><span class="lineno"> 865</span>  vstore8(c00, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l00866"></a><span class="lineno"> 866</span>  vstore8(c10, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l00867"></a><span class="lineno"> 867</span>  vstore8(c20, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l00868"></a><span class="lineno"> 868</span>  vstore8(c30, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l00869"></a><span class="lineno"> 869</span> </div><div class="line"><a name="l00870"></a><span class="lineno"> 870</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00871"></a><span class="lineno"> 871</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00872"></a><span class="lineno"> 872</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l00873"></a><span class="lineno"> 873</span> </div><div class="line"><a name="l00874"></a><span class="lineno"> 874</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l00875"></a><span class="lineno"> 875</span>  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="l00876"></a><span class="lineno"> 876</span>  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="l00877"></a><span class="lineno"> 877</span>  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="l00878"></a><span class="lineno"> 878</span>  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="l00879"></a><span class="lineno"> 879</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00880"></a><span class="lineno"> 880</span> }</div><div class="line"><a name="l00881"></a><span class="lineno"> 881</span> </div><div class="line"><a name="l00920"></a><span class="lineno"> 920</span> __kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f16_acc32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00921"></a><span class="lineno"> 921</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00922"></a><span class="lineno"> 922</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00923"></a><span class="lineno"> 923</span>  uint src0_stride_z,</div><div class="line"><a name="l00924"></a><span class="lineno"> 924</span>  uint src1_stride_z,</div><div class="line"><a name="l00925"></a><span class="lineno"> 925</span>  uint dst_stride_z</div><div class="line"><a name="l00926"></a><span class="lineno"> 926</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00927"></a><span class="lineno"> 927</span>  ,</div><div class="line"><a name="l00928"></a><span class="lineno"> 928</span>  uint cross_plane_pad</div><div class="line"><a name="l00929"></a><span class="lineno"> 929</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00930"></a><span class="lineno"> 930</span>  )</div><div class="line"><a name="l00931"></a><span class="lineno"> 931</span> {</div><div class="line"><a name="l00932"></a><span class="lineno"> 932</span>  <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l00933"></a><span class="lineno"> 933</span>  <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00934"></a><span class="lineno"> 934</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00935"></a><span class="lineno"> 935</span> </div><div class="line"><a name="l00936"></a><span class="lineno"> 936</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00937"></a><span class="lineno"> 937</span>  <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="l00938"></a><span class="lineno"> 938</span>  <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="l00939"></a><span class="lineno"> 939</span> </div><div class="line"><a name="l00940"></a><span class="lineno"> 940</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00941"></a><span class="lineno"> 941</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00942"></a><span class="lineno"> 942</span>  <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="l00943"></a><span class="lineno"> 943</span>  <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="l00944"></a><span class="lineno"> 944</span> </div><div class="line"><a name="l00945"></a><span class="lineno"> 945</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00946"></a><span class="lineno"> 946</span>  <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="l00947"></a><span class="lineno"> 947</span>  src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00948"></a><span class="lineno"> 948</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00949"></a><span class="lineno"> 949</span>  src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l00950"></a><span class="lineno"> 950</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00951"></a><span class="lineno"> 951</span> </div><div class="line"><a name="l00952"></a><span class="lineno"> 952</span>  __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="l00953"></a><span class="lineno"> 953</span>  __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="l00954"></a><span class="lineno"> 954</span> </div><div class="line"><a name="l00955"></a><span class="lineno"> 955</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00956"></a><span class="lineno"> 956</span>  __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="l00957"></a><span class="lineno"> 957</span> </div><div class="line"><a name="l00958"></a><span class="lineno"> 958</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00959"></a><span class="lineno"> 959</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l00960"></a><span class="lineno"> 960</span> </div><div class="line"><a name="l00961"></a><span class="lineno"> 961</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00962"></a><span class="lineno"> 962</span>  float8 c00 = 0.0f;</div><div class="line"><a name="l00963"></a><span class="lineno"> 963</span>  float8 c10 = 0.0f;</div><div class="line"><a name="l00964"></a><span class="lineno"> 964</span>  float8 c20 = 0.0f;</div><div class="line"><a name="l00965"></a><span class="lineno"> 965</span>  float8 c30 = 0.0f;</div><div class="line"><a name="l00966"></a><span class="lineno"> 966</span> </div><div class="line"><a name="l00967"></a><span class="lineno"> 967</span>  <span class="keywordflow">for</span>(; src_addr_b <= (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="l00968"></a><span class="lineno"> 968</span>  {</div><div class="line"><a name="l00969"></a><span class="lineno"> 969</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00970"></a><span class="lineno"> 970</span>  float4 a0 = convert_float4(vload4(0, src_addr_a));</div><div class="line"><a name="l00971"></a><span class="lineno"> 971</span>  float8 b0 = convert_float8(vload8(0, src_addr_b));</div><div class="line"><a name="l00972"></a><span class="lineno"> 972</span> </div><div class="line"><a name="l00973"></a><span class="lineno"> 973</span>  c00 += (float8)a0.s0 * b0;</div><div class="line"><a name="l00974"></a><span class="lineno"> 974</span>  c10 += (float8)a0.s1 * b0;</div><div class="line"><a name="l00975"></a><span class="lineno"> 975</span>  c20 += (float8)a0.s2 * b0;</div><div class="line"><a name="l00976"></a><span class="lineno"> 976</span>  c30 += (float8)a0.s3 * b0;</div><div class="line"><a name="l00977"></a><span class="lineno"> 977</span> </div><div class="line"><a name="l00978"></a><span class="lineno"> 978</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00979"></a><span class="lineno"> 979</span>  a0 = convert_float4(vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00980"></a><span class="lineno"> 980</span>  b0 = convert_float8(vload8(0, src_addr_b + 8 * MULT_TRANSPOSE1XW_WIDTH));</div><div class="line"><a name="l00981"></a><span class="lineno"> 981</span> </div><div class="line"><a name="l00982"></a><span class="lineno"> 982</span>  c00 += (float8)a0.s0 * b0;</div><div class="line"><a name="l00983"></a><span class="lineno"> 983</span>  c10 += (float8)a0.s1 * b0;</div><div class="line"><a name="l00984"></a><span class="lineno"> 984</span>  c20 += (float8)a0.s2 * b0;</div><div class="line"><a name="l00985"></a><span class="lineno"> 985</span>  c30 += (float8)a0.s3 * b0;</div><div class="line"><a name="l00986"></a><span class="lineno"> 986</span>  }</div><div class="line"><a name="l00987"></a><span class="lineno"> 987</span> </div><div class="line"><a name="l00988"></a><span class="lineno"> 988</span>  <span class="keywordflow">for</span>(; src_addr_b < src_end_addr_b; src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH)</div><div class="line"><a name="l00989"></a><span class="lineno"> 989</span>  {</div><div class="line"><a name="l00990"></a><span class="lineno"> 990</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00991"></a><span class="lineno"> 991</span>  float4 a0 = convert_float4(vload4(0, src_addr_a));</div><div class="line"><a name="l00992"></a><span class="lineno"> 992</span>  float8 b0 = convert_float8(vload8(0, src_addr_b));</div><div class="line"><a name="l00993"></a><span class="lineno"> 993</span> </div><div class="line"><a name="l00994"></a><span class="lineno"> 994</span>  c00 += (float8)a0.s0 * b0;</div><div class="line"><a name="l00995"></a><span class="lineno"> 995</span>  c10 += (float8)a0.s1 * b0;</div><div class="line"><a name="l00996"></a><span class="lineno"> 996</span>  c20 += (float8)a0.s2 * b0;</div><div class="line"><a name="l00997"></a><span class="lineno"> 997</span>  c30 += (float8)a0.s3 * b0;</div><div class="line"><a name="l00998"></a><span class="lineno"> 998</span>  }</div><div class="line"><a name="l00999"></a><span class="lineno"> 999</span> </div><div class="line"><a name="l01000"></a><span class="lineno"> 1000</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01001"></a><span class="lineno"> 1001</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01002"></a><span class="lineno"> 1002</span> </div><div class="line"><a name="l01003"></a><span class="lineno"> 1003</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01004"></a><span class="lineno"> 1004</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l01005"></a><span class="lineno"> 1005</span>  c00 = c00 * (float8)ALPHA;</div><div class="line"><a name="l01006"></a><span class="lineno"> 1006</span>  c10 = c10 * (float8)ALPHA;</div><div class="line"><a name="l01007"></a><span class="lineno"> 1007</span>  c20 = c20 * (float8)ALPHA;</div><div class="line"><a name="l01008"></a><span class="lineno"> 1008</span>  c30 = c30 * (float8)ALPHA;</div><div class="line"><a name="l01009"></a><span class="lineno"> 1009</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01010"></a><span class="lineno"> 1010</span> </div><div class="line"><a name="l01011"></a><span class="lineno"> 1011</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span> </div><div class="line"><a name="l01014"></a><span class="lineno"> 1014</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01015"></a><span class="lineno"> 1015</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01016"></a><span class="lineno"> 1016</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01017"></a><span class="lineno"> 1017</span>  <span class="comment">//</span></div><div class="line"><a name="l01018"></a><span class="lineno"> 1018</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01019"></a><span class="lineno"> 1019</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l01020"></a><span class="lineno"> 1020</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01021"></a><span class="lineno"> 1021</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01022"></a><span class="lineno"> 1022</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01023"></a><span class="lineno"> 1023</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l01024"></a><span class="lineno"> 1024</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01025"></a><span class="lineno"> 1025</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01026"></a><span class="lineno"> 1026</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l01027"></a><span class="lineno"> 1027</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01028"></a><span class="lineno"> 1028</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01029"></a><span class="lineno"> 1029</span> </div><div class="line"><a name="l01030"></a><span class="lineno"> 1030</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01031"></a><span class="lineno"> 1031</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01032"></a><span class="lineno"> 1032</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01033"></a><span class="lineno"> 1033</span> </div><div class="line"><a name="l01034"></a><span class="lineno"> 1034</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01035"></a><span class="lineno"> 1035</span>  zout = (cross_plane_pad dst_stride_y);</div><div class="line"><a name="l01036"></a><span class="lineno"> 1036</span> </div><div class="line"><a name="l01037"></a><span class="lineno"> 1037</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01038"></a><span class="lineno"> 1038</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01039"></a><span class="lineno"> 1039</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01040"></a><span class="lineno"> 1040</span> </div><div class="line"><a name="l01041"></a><span class="lineno"> 1041</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l01042"></a><span class="lineno"> 1042</span>  vstore8(convert_half8(c00), 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l01043"></a><span class="lineno"> 1043</span>  vstore8(convert_half8(c10), 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l01044"></a><span class="lineno"> 1044</span>  vstore8(convert_half8(c20), 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l01045"></a><span class="lineno"> 1045</span>  vstore8(convert_half8(c30), 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l01046"></a><span class="lineno"> 1046</span> </div><div class="line"><a name="l01047"></a><span class="lineno"> 1047</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01048"></a><span class="lineno"> 1048</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01049"></a><span class="lineno"> 1049</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01050"></a><span class="lineno"> 1050</span> </div><div class="line"><a name="l01051"></a><span class="lineno"> 1051</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l01052"></a><span class="lineno"> 1052</span>  vstore8(convert_half8(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="l01053"></a><span class="lineno"> 1053</span>  vstore8(convert_half8(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="l01054"></a><span class="lineno"> 1054</span>  vstore8(convert_half8(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="l01055"></a><span class="lineno"> 1055</span>  vstore8(convert_half8(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="l01056"></a><span class="lineno"> 1056</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01057"></a><span class="lineno"> 1057</span> }</div><div class="line"><a name="l01058"></a><span class="lineno"> 1058</span> </div><div class="line"><a name="l01094"></a><span class="lineno"> 1094</span> __kernel <span class="keywordtype">void</span> gemm_mm_interleaved_transposed_f16_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01095"></a><span class="lineno"> 1095</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01096"></a><span class="lineno"> 1096</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01097"></a><span class="lineno"> 1097</span>  uint src0_stride_z,</div><div class="line"><a name="l01098"></a><span class="lineno"> 1098</span>  uint src1_stride_z,</div><div class="line"><a name="l01099"></a><span class="lineno"> 1099</span>  uint dst_stride_z</div><div class="line"><a name="l01100"></a><span class="lineno"> 1100</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l01101"></a><span class="lineno"> 1101</span>  ,</div><div class="line"><a name="l01102"></a><span class="lineno"> 1102</span>  uint cross_plane_pad</div><div class="line"><a name="l01103"></a><span class="lineno"> 1103</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l01104"></a><span class="lineno"> 1104</span>  )</div><div class="line"><a name="l01105"></a><span class="lineno"> 1105</span> {</div><div class="line"><a name="l01106"></a><span class="lineno"> 1106</span>  <span class="keywordtype">int</span> x = get_global_id(0) / MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01107"></a><span class="lineno"> 1107</span>  <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01108"></a><span class="lineno"> 1108</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01109"></a><span class="lineno"> 1109</span> </div><div class="line"><a name="l01110"></a><span class="lineno"> 1110</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l01111"></a><span class="lineno"> 1111</span>  <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="l01112"></a><span class="lineno"> 1112</span>  <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="l01113"></a><span class="lineno"> 1113</span> </div><div class="line"><a name="l01114"></a><span class="lineno"> 1114</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l01115"></a><span class="lineno"> 1115</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l01116"></a><span class="lineno"> 1116</span>  <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="l01117"></a><span class="lineno"> 1117</span>  <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="l01118"></a><span class="lineno"> 1118</span> </div><div class="line"><a name="l01119"></a><span class="lineno"> 1119</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01120"></a><span class="lineno"> 1120</span>  <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="l01121"></a><span class="lineno"> 1121</span>  src1_addr_in_bytes += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01122"></a><span class="lineno"> 1122</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01123"></a><span class="lineno"> 1123</span>  src1_addr_in_bytes += z * src1_stride_z;</div><div class="line"><a name="l01124"></a><span class="lineno"> 1124</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01125"></a><span class="lineno"> 1125</span> </div><div class="line"><a name="l01126"></a><span class="lineno"> 1126</span>  __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="l01127"></a><span class="lineno"> 1127</span>  __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="l01128"></a><span class="lineno"> 1128</span> </div><div class="line"><a name="l01129"></a><span class="lineno"> 1129</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l01130"></a><span class="lineno"> 1130</span>  __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="l01131"></a><span class="lineno"> 1131</span> </div><div class="line"><a name="l01132"></a><span class="lineno"> 1132</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l01133"></a><span class="lineno"> 1133</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l01134"></a><span class="lineno"> 1134</span> </div><div class="line"><a name="l01135"></a><span class="lineno"> 1135</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l01136"></a><span class="lineno"> 1136</span>  half8 c00 = 0.0f;</div><div class="line"><a name="l01137"></a><span class="lineno"> 1137</span>  half8 c10 = 0.0f;</div><div class="line"><a name="l01138"></a><span class="lineno"> 1138</span>  half8 c20 = 0.0f;</div><div class="line"><a name="l01139"></a><span class="lineno"> 1139</span>  half8 c30 = 0.0f;</div><div class="line"><a name="l01140"></a><span class="lineno"> 1140</span> </div><div class="line"><a name="l01141"></a><span class="lineno"> 1141</span> <span class="preprocessor">#define COLS_MTX_B (COLS_B / (8 MULT_TRANSPOSE1XW_WIDTH))</span></div><div class="line"><a name="l01142"></a><span class="lineno"> 1142</span> </div><div class="line"><a name="l01143"></a><span class="lineno"> 1143</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01144"></a><span class="lineno"> 1144</span>  <span class="keywordflow">for</span>(; i <= (int)(COLS_MTX_B - 4); i += 4)</div><div class="line"><a name="l01145"></a><span class="lineno"> 1145</span>  {</div><div class="line"><a name="l01146"></a><span class="lineno"> 1146</span> <span class="preprocessor">#if MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l01147"></a><span class="lineno"> 1147</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01148"></a><span class="lineno"> 1148</span>  half8 a0 = vload8(0, src_addr_a);</div><div class="line"><a name="l01149"></a><span class="lineno"> 1149</span>  half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01150"></a><span class="lineno"> 1150</span> </div><div class="line"><a name="l01151"></a><span class="lineno"> 1151</span>  src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01152"></a><span class="lineno"> 1152</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01153"></a><span class="lineno"> 1153</span> </div><div class="line"><a name="l01154"></a><span class="lineno"> 1154</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01155"></a><span class="lineno"> 1155</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01156"></a><span class="lineno"> 1156</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01157"></a><span class="lineno"> 1157</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01158"></a><span class="lineno"> 1158</span> </div><div class="line"><a name="l01159"></a><span class="lineno"> 1159</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l01160"></a><span class="lineno"> 1160</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01161"></a><span class="lineno"> 1161</span> </div><div class="line"><a name="l01162"></a><span class="lineno"> 1162</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01163"></a><span class="lineno"> 1163</span> </div><div class="line"><a name="l01164"></a><span class="lineno"> 1164</span>  c00 = fma((half8)a0.s4, b0, c00);</div><div class="line"><a name="l01165"></a><span class="lineno"> 1165</span>  c10 = fma((half8)a0.s5, b0, c10);</div><div class="line"><a name="l01166"></a><span class="lineno"> 1166</span>  c20 = fma((half8)a0.s6, b0, c20);</div><div class="line"><a name="l01167"></a><span class="lineno"> 1167</span>  c30 = fma((half8)a0.s7, b0, c30);</div><div class="line"><a name="l01168"></a><span class="lineno"> 1168</span> </div><div class="line"><a name="l01169"></a><span class="lineno"> 1169</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01170"></a><span class="lineno"> 1170</span>  a0 = vload8(0, src_addr_a);</div><div class="line"><a name="l01171"></a><span class="lineno"> 1171</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01172"></a><span class="lineno"> 1172</span> </div><div class="line"><a name="l01173"></a><span class="lineno"> 1173</span>  src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01174"></a><span class="lineno"> 1174</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01175"></a><span class="lineno"> 1175</span> </div><div class="line"><a name="l01176"></a><span class="lineno"> 1176</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01177"></a><span class="lineno"> 1177</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01178"></a><span class="lineno"> 1178</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01179"></a><span class="lineno"> 1179</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01180"></a><span class="lineno"> 1180</span> </div><div class="line"><a name="l01181"></a><span class="lineno"> 1181</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l01182"></a><span class="lineno"> 1182</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01183"></a><span class="lineno"> 1183</span> </div><div class="line"><a name="l01184"></a><span class="lineno"> 1184</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01185"></a><span class="lineno"> 1185</span> </div><div class="line"><a name="l01186"></a><span class="lineno"> 1186</span>  c00 = fma((half8)a0.s4, b0, c00);</div><div class="line"><a name="l01187"></a><span class="lineno"> 1187</span>  c10 = fma((half8)a0.s5, b0, c10);</div><div class="line"><a name="l01188"></a><span class="lineno"> 1188</span>  c20 = fma((half8)a0.s6, b0, c20);</div><div class="line"><a name="l01189"></a><span class="lineno"> 1189</span>  c30 = fma((half8)a0.s7, b0, c30);</div><div class="line"><a name="l01190"></a><span class="lineno"> 1190</span> <span class="preprocessor">#else // MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l01191"></a><span class="lineno"> 1191</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01192"></a><span class="lineno"> 1192</span>  half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01193"></a><span class="lineno"> 1193</span>  half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01194"></a><span class="lineno"> 1194</span> </div><div class="line"><a name="l01195"></a><span class="lineno"> 1195</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01196"></a><span class="lineno"> 1196</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01197"></a><span class="lineno"> 1197</span> </div><div class="line"><a name="l01198"></a><span class="lineno"> 1198</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01199"></a><span class="lineno"> 1199</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01200"></a><span class="lineno"> 1200</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01201"></a><span class="lineno"> 1201</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01202"></a><span class="lineno"> 1202</span> </div><div class="line"><a name="l01203"></a><span class="lineno"> 1203</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01204"></a><span class="lineno"> 1204</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01205"></a><span class="lineno"> 1205</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01206"></a><span class="lineno"> 1206</span> </div><div class="line"><a name="l01207"></a><span class="lineno"> 1207</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01208"></a><span class="lineno"> 1208</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01209"></a><span class="lineno"> 1209</span> </div><div class="line"><a name="l01210"></a><span class="lineno"> 1210</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01211"></a><span class="lineno"> 1211</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01212"></a><span class="lineno"> 1212</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01213"></a><span class="lineno"> 1213</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01214"></a><span class="lineno"> 1214</span> </div><div class="line"><a name="l01215"></a><span class="lineno"> 1215</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01216"></a><span class="lineno"> 1216</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01217"></a><span class="lineno"> 1217</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01218"></a><span class="lineno"> 1218</span> </div><div class="line"><a name="l01219"></a><span class="lineno"> 1219</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01220"></a><span class="lineno"> 1220</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01221"></a><span class="lineno"> 1221</span> </div><div class="line"><a name="l01222"></a><span class="lineno"> 1222</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01223"></a><span class="lineno"> 1223</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01224"></a><span class="lineno"> 1224</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01225"></a><span class="lineno"> 1225</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01226"></a><span class="lineno"> 1226</span> </div><div class="line"><a name="l01227"></a><span class="lineno"> 1227</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01228"></a><span class="lineno"> 1228</span>  a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01229"></a><span class="lineno"> 1229</span>  b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01230"></a><span class="lineno"> 1230</span> </div><div class="line"><a name="l01231"></a><span class="lineno"> 1231</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01232"></a><span class="lineno"> 1232</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01233"></a><span class="lineno"> 1233</span> </div><div class="line"><a name="l01234"></a><span class="lineno"> 1234</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01235"></a><span class="lineno"> 1235</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01236"></a><span class="lineno"> 1236</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01237"></a><span class="lineno"> 1237</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01238"></a><span class="lineno"> 1238</span> <span class="preprocessor">#endif // MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l01239"></a><span class="lineno"> 1239</span>  }</div><div class="line"><a name="l01240"></a><span class="lineno"> 1240</span> </div><div class="line"><a name="l01241"></a><span class="lineno"> 1241</span>  <span class="keywordflow">for</span>(; i < (int)(COLS_MTX_B); ++i)</div><div class="line"><a name="l01242"></a><span class="lineno"> 1242</span>  {</div><div class="line"><a name="l01243"></a><span class="lineno"> 1243</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l01244"></a><span class="lineno"> 1244</span>  half4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l01245"></a><span class="lineno"> 1245</span>  half8 b0 = vload8(0, src_addr_b);</div><div class="line"><a name="l01246"></a><span class="lineno"> 1246</span> </div><div class="line"><a name="l01247"></a><span class="lineno"> 1247</span>  src_addr_a += 4 * MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l01248"></a><span class="lineno"> 1248</span>  src_addr_b += 8 * MULT_TRANSPOSE1XW_WIDTH;</div><div class="line"><a name="l01249"></a><span class="lineno"> 1249</span> </div><div class="line"><a name="l01250"></a><span class="lineno"> 1250</span>  c00 = fma((half8)a0.s0, b0, c00);</div><div class="line"><a name="l01251"></a><span class="lineno"> 1251</span>  c10 = fma((half8)a0.s1, b0, c10);</div><div class="line"><a name="l01252"></a><span class="lineno"> 1252</span>  c20 = fma((half8)a0.s2, b0, c20);</div><div class="line"><a name="l01253"></a><span class="lineno"> 1253</span>  c30 = fma((half8)a0.s3, b0, c30);</div><div class="line"><a name="l01254"></a><span class="lineno"> 1254</span>  }</div><div class="line"><a name="l01255"></a><span class="lineno"> 1255</span> </div><div class="line"><a name="l01256"></a><span class="lineno"> 1256</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01257"></a><span class="lineno"> 1257</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01258"></a><span class="lineno"> 1258</span> </div><div class="line"><a name="l01259"></a><span class="lineno"> 1259</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01260"></a><span class="lineno"> 1260</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l01261"></a><span class="lineno"> 1261</span>  c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l01262"></a><span class="lineno"> 1262</span>  c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l01263"></a><span class="lineno"> 1263</span>  c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l01264"></a><span class="lineno"> 1264</span>  c30 = c30 * (half8)ALPHA;</div><div class="line"><a name="l01265"></a><span class="lineno"> 1265</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01266"></a><span class="lineno"> 1266</span> </div><div class="line"><a name="l01267"></a><span class="lineno"> 1267</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01268"></a><span class="lineno"> 1268</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l01269"></a><span class="lineno"> 1269</span> </div><div class="line"><a name="l01270"></a><span class="lineno"> 1270</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01271"></a><span class="lineno"> 1271</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01272"></a><span class="lineno"> 1272</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01273"></a><span class="lineno"> 1273</span>  <span class="comment">//</span></div><div class="line"><a name="l01274"></a><span class="lineno"> 1274</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01275"></a><span class="lineno"> 1275</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l01276"></a><span class="lineno"> 1276</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01277"></a><span class="lineno"> 1277</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01278"></a><span class="lineno"> 1278</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01279"></a><span class="lineno"> 1279</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l01280"></a><span class="lineno"> 1280</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01281"></a><span class="lineno"> 1281</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01282"></a><span class="lineno"> 1282</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l01283"></a><span class="lineno"> 1283</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01284"></a><span class="lineno"> 1284</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01285"></a><span class="lineno"> 1285</span> </div><div class="line"><a name="l01286"></a><span class="lineno"> 1286</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01287"></a><span class="lineno"> 1287</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01288"></a><span class="lineno"> 1288</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01289"></a><span class="lineno"> 1289</span> </div><div class="line"><a name="l01290"></a><span class="lineno"> 1290</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01291"></a><span class="lineno"> 1291</span>  zout = (cross_plane_pad dst_stride_y);</div><div class="line"><a name="l01292"></a><span class="lineno"> 1292</span> </div><div class="line"><a name="l01293"></a><span class="lineno"> 1293</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01294"></a><span class="lineno"> 1294</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01295"></a><span class="lineno"> 1295</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01296"></a><span class="lineno"> 1296</span> </div><div class="line"><a name="l01297"></a><span class="lineno"> 1297</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l01298"></a><span class="lineno"> 1298</span>  vstore8(c00, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l01299"></a><span class="lineno"> 1299</span>  vstore8(c10, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l01300"></a><span class="lineno"> 1300</span>  vstore8(c20, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l01301"></a><span class="lineno"> 1301</span>  vstore8(c30, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l01302"></a><span class="lineno"> 1302</span> </div><div class="line"><a name="l01303"></a><span class="lineno"> 1303</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01304"></a><span class="lineno"> 1304</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01305"></a><span class="lineno"> 1305</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01306"></a><span class="lineno"> 1306</span> </div><div class="line"><a name="l01307"></a><span class="lineno"> 1307</span>  <span class="comment">// Store 4x8 block</span></div><div class="line"><a name="l01308"></a><span class="lineno"> 1308</span>  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="l01309"></a><span class="lineno"> 1309</span>  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="l01310"></a><span class="lineno"> 1310</span>  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="l01311"></a><span class="lineno"> 1311</span>  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="l01312"></a><span class="lineno"> 1312</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01313"></a><span class="lineno"> 1313</span> }</div><div class="line"><a name="l01314"></a><span class="lineno"> 1314</span> </div><div class="line"><a name="l01315"></a><span class="lineno"> 1315</span> <span class="comment">// Undefine local defines</span></div><div class="line"><a name="l01316"></a><span class="lineno"> 1316</span> <span class="preprocessor">#undef COLS_MTX_B</span></div><div class="line"><a name="l01317"></a><span class="lineno"> 1317</span> </div><div class="line"><a name="l01318"></a><span class="lineno"> 1318</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l01319"></a><span class="lineno"> 1319</span> </div><div class="line"><a name="l01320"></a><span class="lineno"> 1320</span> <span class="preprocessor">#endif // defined(COLS_B) && defined(MULT_TRANSPOSE1XW_WIDTH) && defined(MULT_INTERLEAVE4X4_HEIGHT)</span></div><div class="line"><a name="l01321"></a><span class="lineno"> 1321</span> </div><div class="line"><a name="l01322"></a><span class="lineno"> 1322</span> <span class="preprocessor">#if defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y)</span></div><div class="line"><a name="l01323"></a><span class="lineno"> 1323</span> <span class="preprocessor">#if defined(DATA_TYPE)</span></div><div class="line"><a name="l01324"></a><span class="lineno"> 1324</span> <span class="preprocessor">#define VECTOR_TYPE VEC_DATA_TYPE(DATA_TYPE, NUM_ELEMS_PROCESSED_PER_THREAD_X)</span></div><div class="line"><a name="l01325"></a><span class="lineno"> 1325</span> </div><div class="line"><a name="l01365"></a><span class="lineno"> 1365</span> __kernel <span class="keywordtype">void</span> gemm_mm_floating_point(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01366"></a><span class="lineno"> 1366</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01367"></a><span class="lineno"> 1367</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01368"></a><span class="lineno"> 1368</span>  uint src0_stride_z,</div><div class="line"><a name="l01369"></a><span class="lineno"> 1369</span>  uint src1_stride_z,</div><div class="line"><a name="l01370"></a><span class="lineno"> 1370</span>  uint dst_stride_z</div><div class="line"><a name="l01371"></a><span class="lineno"> 1371</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l01372"></a><span class="lineno"> 1372</span>  ,</div><div class="line"><a name="l01373"></a><span class="lineno"> 1373</span>  uint src_cross_plane_pad</div><div class="line"><a name="l01374"></a><span class="lineno"> 1374</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l01375"></a><span class="lineno"> 1375</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l01376"></a><span class="lineno"> 1376</span>  ,</div><div class="line"><a name="l01377"></a><span class="lineno"> 1377</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l01378"></a><span class="lineno"> 1378</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l01379"></a><span class="lineno"> 1379</span>  )</div><div class="line"><a name="l01380"></a><span class="lineno"> 1380</span> {</div><div class="line"><a name="l01381"></a><span class="lineno"> 1381</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01382"></a><span class="lineno"> 1382</span> </div><div class="line"><a name="l01383"></a><span class="lineno"> 1383</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01384"></a><span class="lineno"> 1384</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01385"></a><span class="lineno"> 1385</span> </div><div class="line"><a name="l01386"></a><span class="lineno"> 1386</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01387"></a><span class="lineno"> 1387</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01388"></a><span class="lineno"> 1388</span> </div><div class="line"><a name="l01389"></a><span class="lineno"> 1389</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01390"></a><span class="lineno"> 1390</span>  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="l01391"></a><span class="lineno"> 1391</span> </div><div class="line"><a name="l01392"></a><span class="lineno"> 1392</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01393"></a><span class="lineno"> 1393</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01394"></a><span class="lineno"> 1394</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01395"></a><span class="lineno"> 1395</span>  <span class="comment">//</span></div><div class="line"><a name="l01396"></a><span class="lineno"> 1396</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01397"></a><span class="lineno"> 1397</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l01398"></a><span class="lineno"> 1398</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01399"></a><span class="lineno"> 1399</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01400"></a><span class="lineno"> 1400</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l01401"></a><span class="lineno"> 1401</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l01402"></a><span class="lineno"> 1402</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01403"></a><span class="lineno"> 1403</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01404"></a><span class="lineno"> 1404</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l01405"></a><span class="lineno"> 1405</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01406"></a><span class="lineno"> 1406</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01407"></a><span class="lineno"> 1407</span> </div><div class="line"><a name="l01408"></a><span class="lineno"> 1408</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01409"></a><span class="lineno"> 1409</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01410"></a><span class="lineno"> 1410</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l01411"></a><span class="lineno"> 1411</span> </div><div class="line"><a name="l01412"></a><span class="lineno"> 1412</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01413"></a><span class="lineno"> 1413</span>  zin = (src_cross_plane_pad src0_stride_y);</div><div class="line"><a name="l01414"></a><span class="lineno"> 1414</span> </div><div class="line"><a name="l01415"></a><span class="lineno"> 1415</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01416"></a><span class="lineno"> 1416</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01417"></a><span class="lineno"> 1417</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01418"></a><span class="lineno"> 1418</span> </div><div class="line"><a name="l01419"></a><span class="lineno"> 1419</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01420"></a><span class="lineno"> 1420</span> </div><div class="line"><a name="l01421"></a><span class="lineno"> 1421</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01422"></a><span class="lineno"> 1422</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01423"></a><span class="lineno"> 1423</span> </div><div class="line"><a name="l01424"></a><span class="lineno"> 1424</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01425"></a><span class="lineno"> 1425</span> </div><div class="line"><a name="l01426"></a><span class="lineno"> 1426</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01427"></a><span class="lineno"> 1427</span>  <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="l01428"></a><span class="lineno"> 1428</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01429"></a><span class="lineno"> 1429</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01430"></a><span class="lineno"> 1430</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01431"></a><span class="lineno"> 1431</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01432"></a><span class="lineno"> 1432</span> </div><div class="line"><a name="l01433"></a><span class="lineno"> 1433</span>  <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="l01434"></a><span class="lineno"> 1434</span> </div><div class="line"><a name="l01435"></a><span class="lineno"> 1435</span>  VECTOR_TYPE acc0 = 0.0f;</div><div class="line"><a name="l01436"></a><span class="lineno"> 1436</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01437"></a><span class="lineno"> 1437</span>  VECTOR_TYPE acc1 = 0.0f;</div><div class="line"><a name="l01438"></a><span class="lineno"> 1438</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01439"></a><span class="lineno"> 1439</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01440"></a><span class="lineno"> 1440</span>  VECTOR_TYPE acc2 = 0.0f;</div><div class="line"><a name="l01441"></a><span class="lineno"> 1441</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01442"></a><span class="lineno"> 1442</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01443"></a><span class="lineno"> 1443</span>  VECTOR_TYPE acc3 = 0.0f;</div><div class="line"><a name="l01444"></a><span class="lineno"> 1444</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01445"></a><span class="lineno"> 1445</span> </div><div class="line"><a name="l01446"></a><span class="lineno"> 1446</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (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="l01447"></a><span class="lineno"> 1447</span>  {</div><div class="line"><a name="l01448"></a><span class="lineno"> 1448</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01449"></a><span class="lineno"> 1449</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01450"></a><span class="lineno"> 1450</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01451"></a><span class="lineno"> 1451</span>  a0 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 0 src0_stride_y + zin.s0));</div><div class="line"><a name="l01452"></a><span class="lineno"> 1452</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01453"></a><span class="lineno"> 1453</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01454"></a><span class="lineno"> 1454</span>  a1 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 1 src0_stride_y + zin.s1));</div><div class="line"><a name="l01455"></a><span class="lineno"> 1455</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01456"></a><span class="lineno"> 1456</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01457"></a><span class="lineno"> 1457</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01458"></a><span class="lineno"> 1458</span>  a2 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 2 src0_stride_y + zin.s2));</div><div class="line"><a name="l01459"></a><span class="lineno"> 1459</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01460"></a><span class="lineno"> 1460</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01461"></a><span class="lineno"> 1461</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01462"></a><span class="lineno"> 1462</span>  a3 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 3 src0_stride_y + zin.s3));</div><div class="line"><a name="l01463"></a><span class="lineno"> 1463</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01464"></a><span class="lineno"> 1464</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01465"></a><span class="lineno"> 1465</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01466"></a><span class="lineno"> 1466</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01467"></a><span class="lineno"> 1467</span>  a0 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 0 src0_stride_y));</div><div class="line"><a name="l01468"></a><span class="lineno"> 1468</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01469"></a><span class="lineno"> 1469</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01470"></a><span class="lineno"> 1470</span>  a1 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 1 src0_stride_y));</div><div class="line"><a name="l01471"></a><span class="lineno"> 1471</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01472"></a><span class="lineno"> 1472</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01473"></a><span class="lineno"> 1473</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01474"></a><span class="lineno"> 1474</span>  a2 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 2 src0_stride_y));</div><div class="line"><a name="l01475"></a><span class="lineno"> 1475</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01476"></a><span class="lineno"> 1476</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01477"></a><span class="lineno"> 1477</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(DATA_TYPE, 2)</div><div class="line"><a name="l01478"></a><span class="lineno"> 1478</span>  a3 = vload2(0, (__global DATA_TYPE )(src0_ptr + src_addr.s0 + 3 src0_stride_y));</div><div class="line"><a name="l01479"></a><span class="lineno"> 1479</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01480"></a><span class="lineno"> 1480</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01481"></a><span class="lineno"> 1481</span> </div><div class="line"><a name="l01482"></a><span class="lineno"> 1482</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01483"></a><span class="lineno"> 1483</span>  VECTOR_TYPE b0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_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="l01484"></a><span class="lineno"> 1484</span>  VECTOR_TYPE b1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_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="l01485"></a><span class="lineno"> 1485</span> </div><div class="line"><a name="l01486"></a><span class="lineno"> 1486</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01487"></a><span class="lineno"> 1487</span>  acc0 += b0 * (VECTOR_TYPE)a0.s0;</div><div class="line"><a name="l01488"></a><span class="lineno"> 1488</span>  acc0 += b1 * (VECTOR_TYPE)a0.s1;</div><div class="line"><a name="l01489"></a><span class="lineno"> 1489</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01490"></a><span class="lineno"> 1490</span>  acc1 += b0 * (VECTOR_TYPE)a1.s0;</div><div class="line"><a name="l01491"></a><span class="lineno"> 1491</span>  acc1 += b1 * (VECTOR_TYPE)a1.s1;</div><div class="line"><a name="l01492"></a><span class="lineno"> 1492</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01493"></a><span class="lineno"> 1493</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01494"></a><span class="lineno"> 1494</span>  acc2 += b0 * (VECTOR_TYPE)a2.s0;</div><div class="line"><a name="l01495"></a><span class="lineno"> 1495</span>  acc2 += b1 * (VECTOR_TYPE)a2.s1;</div><div class="line"><a name="l01496"></a><span class="lineno"> 1496</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01497"></a><span class="lineno"> 1497</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01498"></a><span class="lineno"> 1498</span>  acc3 += b0 * (VECTOR_TYPE)a3.s0;</div><div class="line"><a name="l01499"></a><span class="lineno"> 1499</span>  acc3 += b1 * (VECTOR_TYPE)a3.s1;</div><div class="line"><a name="l01500"></a><span class="lineno"> 1500</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01501"></a><span class="lineno"> 1501</span>  }</div><div class="line"><a name="l01502"></a><span class="lineno"> 1502</span> </div><div class="line"><a name="l01503"></a><span class="lineno"> 1503</span>  <span class="keywordflow">for</span>(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(DATA_TYPE), src1_stride_y))</div><div class="line"><a name="l01504"></a><span class="lineno"> 1504</span>  {</div><div class="line"><a name="l01505"></a><span class="lineno"> 1505</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01506"></a><span class="lineno"> 1506</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01507"></a><span class="lineno"> 1507</span>  DATA_TYPE a0 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));</div><div class="line"><a name="l01508"></a><span class="lineno"> 1508</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01509"></a><span class="lineno"> 1509</span>  DATA_TYPE a1 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));</div><div class="line"><a name="l01510"></a><span class="lineno"> 1510</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01511"></a><span class="lineno"> 1511</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01512"></a><span class="lineno"> 1512</span>  DATA_TYPE a2 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));</div><div class="line"><a name="l01513"></a><span class="lineno"> 1513</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01514"></a><span class="lineno"> 1514</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01515"></a><span class="lineno"> 1515</span>  DATA_TYPE a3 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));</div><div class="line"><a name="l01516"></a><span class="lineno"> 1516</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01517"></a><span class="lineno"> 1517</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01518"></a><span class="lineno"> 1518</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01519"></a><span class="lineno"> 1519</span>  DATA_TYPE a0 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01520"></a><span class="lineno"> 1520</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01521"></a><span class="lineno"> 1521</span>  DATA_TYPE a1 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01522"></a><span class="lineno"> 1522</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01523"></a><span class="lineno"> 1523</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01524"></a><span class="lineno"> 1524</span>  DATA_TYPE a2 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01525"></a><span class="lineno"> 1525</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01526"></a><span class="lineno"> 1526</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01527"></a><span class="lineno"> 1527</span>  DATA_TYPE a3 = ((__global DATA_TYPE )(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01528"></a><span class="lineno"> 1528</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01529"></a><span class="lineno"> 1529</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01530"></a><span class="lineno"> 1530</span> </div><div class="line"><a name="l01531"></a><span class="lineno"> 1531</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01532"></a><span class="lineno"> 1532</span>  VECTOR_TYPE b0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_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="l01533"></a><span class="lineno"> 1533</span> </div><div class="line"><a name="l01534"></a><span class="lineno"> 1534</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01535"></a><span class="lineno"> 1535</span>  acc0 += b0 (VECTOR_TYPE)a0;</div><div class="line"><a name="l01536"></a><span class="lineno"> 1536</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01537"></a><span class="lineno"> 1537</span>  acc1 += b0 * (VECTOR_TYPE)a1;</div><div class="line"><a name="l01538"></a><span class="lineno"> 1538</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01539"></a><span class="lineno"> 1539</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01540"></a><span class="lineno"> 1540</span>  acc2 += b0 * (VECTOR_TYPE)a2;</div><div class="line"><a name="l01541"></a><span class="lineno"> 1541</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01542"></a><span class="lineno"> 1542</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01543"></a><span class="lineno"> 1543</span>  acc3 += b0 * (VECTOR_TYPE)a3;</div><div class="line"><a name="l01544"></a><span class="lineno"> 1544</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01545"></a><span class="lineno"> 1545</span>  }</div><div class="line"><a name="l01546"></a><span class="lineno"> 1546</span> </div><div class="line"><a name="l01547"></a><span class="lineno"> 1547</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01548"></a><span class="lineno"> 1548</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01549"></a><span class="lineno"> 1549</span> </div><div class="line"><a name="l01550"></a><span class="lineno"> 1550</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01551"></a><span class="lineno"> 1551</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l01552"></a><span class="lineno"> 1552</span> </div><div class="line"><a name="l01553"></a><span class="lineno"> 1553</span>  <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l01554"></a><span class="lineno"> 1554</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l01555"></a><span class="lineno"> 1555</span>  acc0 = acc0 (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01556"></a><span class="lineno"> 1556</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l01557"></a><span class="lineno"> 1557</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l01558"></a><span class="lineno"> 1558</span>  acc1 = acc1 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01559"></a><span class="lineno"> 1559</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l01560"></a><span class="lineno"> 1560</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l01561"></a><span class="lineno"> 1561</span>  acc2 = acc2 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01562"></a><span class="lineno"> 1562</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l01563"></a><span class="lineno"> 1563</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l01564"></a><span class="lineno"> 1564</span>  acc3 = acc3 * (VECTOR_TYPE)ALPHA;</div><div class="line"><a name="l01565"></a><span class="lineno"> 1565</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l01566"></a><span class="lineno"> 1566</span> </div><div class="line"><a name="l01567"></a><span class="lineno"> 1567</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01568"></a><span class="lineno"> 1568</span> </div><div class="line"><a name="l01569"></a><span class="lineno"> 1569</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01570"></a><span class="lineno"> 1570</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01571"></a><span class="lineno"> 1571</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01572"></a><span class="lineno"> 1572</span>  <span class="comment">//</span></div><div class="line"><a name="l01573"></a><span class="lineno"> 1573</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01574"></a><span class="lineno"> 1574</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l01575"></a><span class="lineno"> 1575</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01576"></a><span class="lineno"> 1576</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01577"></a><span class="lineno"> 1577</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l01578"></a><span class="lineno"> 1578</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l01579"></a><span class="lineno"> 1579</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01580"></a><span class="lineno"> 1580</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01581"></a><span class="lineno"> 1581</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l01582"></a><span class="lineno"> 1582</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01583"></a><span class="lineno"> 1583</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01584"></a><span class="lineno"> 1584</span> </div><div class="line"><a name="l01585"></a><span class="lineno"> 1585</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01586"></a><span class="lineno"> 1586</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01587"></a><span class="lineno"> 1587</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01588"></a><span class="lineno"> 1588</span> </div><div class="line"><a name="l01589"></a><span class="lineno"> 1589</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01590"></a><span class="lineno"> 1590</span>  zout = (dst_cross_plane_pad dst_stride_y);</div><div class="line"><a name="l01591"></a><span class="lineno"> 1591</span> </div><div class="line"><a name="l01592"></a><span class="lineno"> 1592</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01593"></a><span class="lineno"> 1593</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01594"></a><span class="lineno"> 1594</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01595"></a><span class="lineno"> 1595</span> </div><div class="line"><a name="l01596"></a><span class="lineno"> 1596</span>  <span class="comment">// Store output block</span></div><div class="line"><a name="l01597"></a><span class="lineno"> 1597</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01598"></a><span class="lineno"> 1598</span>  (acc0, 0, (__global DATA_TYPE )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l01599"></a><span class="lineno"> 1599</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01600"></a><span class="lineno"> 1600</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01601"></a><span class="lineno"> 1601</span>  (acc1, 0, (__global DATA_TYPE )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l01602"></a><span class="lineno"> 1602</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01603"></a><span class="lineno"> 1603</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01604"></a><span class="lineno"> 1604</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01605"></a><span class="lineno"> 1605</span>  (acc2, 0, (__global DATA_TYPE )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l01606"></a><span class="lineno"> 1606</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01607"></a><span class="lineno"> 1607</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01608"></a><span class="lineno"> 1608</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01609"></a><span class="lineno"> 1609</span>  (acc3, 0, (__global DATA_TYPE )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l01610"></a><span class="lineno"> 1610</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01611"></a><span class="lineno"> 1611</span> </div><div class="line"><a name="l01612"></a><span class="lineno"> 1612</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01613"></a><span class="lineno"> 1613</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01614"></a><span class="lineno"> 1614</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01615"></a><span class="lineno"> 1615</span> </div><div class="line"><a name="l01616"></a><span class="lineno"> 1616</span>  <span class="comment">// Store output block</span></div><div class="line"><a name="l01617"></a><span class="lineno"> 1617</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01618"></a><span class="lineno"> 1618</span>  (acc0, 0, (__global DATA_TYPE )(dst_addr + 0 dst_stride_y));</div><div class="line"><a name="l01619"></a><span class="lineno"> 1619</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01620"></a><span class="lineno"> 1620</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01621"></a><span class="lineno"> 1621</span>  (acc1, 0, (__global DATA_TYPE )(dst_addr + 1 dst_stride_y));</div><div class="line"><a name="l01622"></a><span class="lineno"> 1622</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01623"></a><span class="lineno"> 1623</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01624"></a><span class="lineno"> 1624</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01625"></a><span class="lineno"> 1625</span>  (acc2, 0, (__global DATA_TYPE )(dst_addr + 2 dst_stride_y));</div><div class="line"><a name="l01626"></a><span class="lineno"> 1626</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01627"></a><span class="lineno"> 1627</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01628"></a><span class="lineno"> 1628</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01629"></a><span class="lineno"> 1629</span>  (acc3, 0, (__global DATA_TYPE )(dst_addr + 3 dst_stride_y));</div><div class="line"><a name="l01630"></a><span class="lineno"> 1630</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01631"></a><span class="lineno"> 1631</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01632"></a><span class="lineno"> 1632</span> }</div><div class="line"><a name="l01633"></a><span class="lineno"> 1633</span> <span class="preprocessor">#endif // defined(DATA_TYPE)</span></div><div class="line"><a name="l01634"></a><span class="lineno"> 1634</span> </div><div class="line"><a name="l01676"></a><span class="lineno"> 1676</span> __kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f32_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01677"></a><span class="lineno"> 1677</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01678"></a><span class="lineno"> 1678</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01679"></a><span class="lineno"> 1679</span>  uint src0_stride_z,</div><div class="line"><a name="l01680"></a><span class="lineno"> 1680</span>  uint src1_stride_z,</div><div class="line"><a name="l01681"></a><span class="lineno"> 1681</span>  uint dst_stride_z</div><div class="line"><a name="l01682"></a><span class="lineno"> 1682</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l01683"></a><span class="lineno"> 1683</span>  ,</div><div class="line"><a name="l01684"></a><span class="lineno"> 1684</span>  uint src_cross_plane_pad</div><div class="line"><a name="l01685"></a><span class="lineno"> 1685</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l01686"></a><span class="lineno"> 1686</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l01687"></a><span class="lineno"> 1687</span>  ,</div><div class="line"><a name="l01688"></a><span class="lineno"> 1688</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l01689"></a><span class="lineno"> 1689</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l01690"></a><span class="lineno"> 1690</span>  )</div><div class="line"><a name="l01691"></a><span class="lineno"> 1691</span> {</div><div class="line"><a name="l01692"></a><span class="lineno"> 1692</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01693"></a><span class="lineno"> 1693</span> </div><div class="line"><a name="l01694"></a><span class="lineno"> 1694</span>  <span class="comment">// Compute starting address for matrix A and matrix B</span></div><div class="line"><a name="l01695"></a><span class="lineno"> 1695</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01696"></a><span class="lineno"> 1696</span> </div><div class="line"><a name="l01697"></a><span class="lineno"> 1697</span>  <span class="comment">// Update address for matrix A</span></div><div class="line"><a name="l01698"></a><span class="lineno"> 1698</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01699"></a><span class="lineno"> 1699</span> </div><div class="line"><a name="l01700"></a><span class="lineno"> 1700</span>  <span class="comment">// Update address for matrix B</span></div><div class="line"><a name="l01701"></a><span class="lineno"> 1701</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01702"></a><span class="lineno"> 1702</span> </div><div class="line"><a name="l01703"></a><span class="lineno"> 1703</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01704"></a><span class="lineno"> 1704</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01705"></a><span class="lineno"> 1705</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01706"></a><span class="lineno"> 1706</span>  <span class="comment">//</span></div><div class="line"><a name="l01707"></a><span class="lineno"> 1707</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01708"></a><span class="lineno"> 1708</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l01709"></a><span class="lineno"> 1709</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01710"></a><span class="lineno"> 1710</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01711"></a><span class="lineno"> 1711</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l01712"></a><span class="lineno"> 1712</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l01713"></a><span class="lineno"> 1713</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l01714"></a><span class="lineno"> 1714</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01715"></a><span class="lineno"> 1715</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l01716"></a><span class="lineno"> 1716</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l01717"></a><span class="lineno"> 1717</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l01718"></a><span class="lineno"> 1718</span> </div><div class="line"><a name="l01719"></a><span class="lineno"> 1719</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01720"></a><span class="lineno"> 1720</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01721"></a><span class="lineno"> 1721</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l01722"></a><span class="lineno"> 1722</span> </div><div class="line"><a name="l01723"></a><span class="lineno"> 1723</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01724"></a><span class="lineno"> 1724</span>  zin = (src_cross_plane_pad src0_stride_y);</div><div class="line"><a name="l01725"></a><span class="lineno"> 1725</span> </div><div class="line"><a name="l01726"></a><span class="lineno"> 1726</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01727"></a><span class="lineno"> 1727</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01728"></a><span class="lineno"> 1728</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01729"></a><span class="lineno"> 1729</span> </div><div class="line"><a name="l01730"></a><span class="lineno"> 1730</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01731"></a><span class="lineno"> 1731</span> </div><div class="line"><a name="l01732"></a><span class="lineno"> 1732</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01733"></a><span class="lineno"> 1733</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01734"></a><span class="lineno"> 1734</span> </div><div class="line"><a name="l01735"></a><span class="lineno"> 1735</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01736"></a><span class="lineno"> 1736</span> </div><div class="line"><a name="l01737"></a><span class="lineno"> 1737</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01738"></a><span class="lineno"> 1738</span>  <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="l01739"></a><span class="lineno"> 1739</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01740"></a><span class="lineno"> 1740</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01741"></a><span class="lineno"> 1741</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01742"></a><span class="lineno"> 1742</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01743"></a><span class="lineno"> 1743</span> </div><div class="line"><a name="l01744"></a><span class="lineno"> 1744</span>  <span class="comment">// Initialize accumulators</span></div><div class="line"><a name="l01745"></a><span class="lineno"> 1745</span>  <span class="keywordtype">float</span> acc00 = 0.0f;</div><div class="line"><a name="l01746"></a><span class="lineno"> 1746</span>  <span class="keywordtype">float</span> acc01 = 0.0f;</div><div class="line"><a name="l01747"></a><span class="lineno"> 1747</span>  <span class="keywordtype">float</span> acc02 = 0.0f;</div><div class="line"><a name="l01748"></a><span class="lineno"> 1748</span>  <span class="keywordtype">float</span> acc03 = 0.0f;</div><div class="line"><a name="l01749"></a><span class="lineno"> 1749</span> </div><div class="line"><a name="l01750"></a><span class="lineno"> 1750</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01751"></a><span class="lineno"> 1751</span>  <span class="keywordtype">float</span> acc10 = 0.0f;</div><div class="line"><a name="l01752"></a><span class="lineno"> 1752</span>  <span class="keywordtype">float</span> acc11 = 0.0f;</div><div class="line"><a name="l01753"></a><span class="lineno"> 1753</span>  <span class="keywordtype">float</span> acc12 = 0.0f;</div><div class="line"><a name="l01754"></a><span class="lineno"> 1754</span>  <span class="keywordtype">float</span> acc13 = 0.0f;</div><div class="line"><a name="l01755"></a><span class="lineno"> 1755</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01756"></a><span class="lineno"> 1756</span> </div><div class="line"><a name="l01757"></a><span class="lineno"> 1757</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01758"></a><span class="lineno"> 1758</span>  <span class="keywordtype">float</span> acc20 = 0.0f;</div><div class="line"><a name="l01759"></a><span class="lineno"> 1759</span>  <span class="keywordtype">float</span> acc21 = 0.0f;</div><div class="line"><a name="l01760"></a><span class="lineno"> 1760</span>  <span class="keywordtype">float</span> acc22 = 0.0f;</div><div class="line"><a name="l01761"></a><span class="lineno"> 1761</span>  <span class="keywordtype">float</span> acc23 = 0.0f;</div><div class="line"><a name="l01762"></a><span class="lineno"> 1762</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01763"></a><span class="lineno"> 1763</span> </div><div class="line"><a name="l01764"></a><span class="lineno"> 1764</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01765"></a><span class="lineno"> 1765</span>  <span class="keywordtype">float</span> acc30 = 0.0f;</div><div class="line"><a name="l01766"></a><span class="lineno"> 1766</span>  <span class="keywordtype">float</span> acc31 = 0.0f;</div><div class="line"><a name="l01767"></a><span class="lineno"> 1767</span>  <span class="keywordtype">float</span> acc32 = 0.0f;</div><div class="line"><a name="l01768"></a><span class="lineno"> 1768</span>  <span class="keywordtype">float</span> acc33 = 0.0f;</div><div class="line"><a name="l01769"></a><span class="lineno"> 1769</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01770"></a><span class="lineno"> 1770</span> </div><div class="line"><a name="l01771"></a><span class="lineno"> 1771</span>  <span class="comment">// A and B src indices get incremented at the same time.</span></div><div class="line"><a name="l01772"></a><span class="lineno"> 1772</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01773"></a><span class="lineno"> 1773</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 4); i += 4)</div><div class="line"><a name="l01774"></a><span class="lineno"> 1774</span>  {</div><div class="line"><a name="l01775"></a><span class="lineno"> 1775</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01776"></a><span class="lineno"> 1776</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01777"></a><span class="lineno"> 1777</span>  float4 a0 = vload4(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 0 src0_stride_y + zin.s0));</div><div class="line"><a name="l01778"></a><span class="lineno"> 1778</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01779"></a><span class="lineno"> 1779</span>  float4 a1 = vload4(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 1 src0_stride_y + zin.s1));</div><div class="line"><a name="l01780"></a><span class="lineno"> 1780</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01781"></a><span class="lineno"> 1781</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01782"></a><span class="lineno"> 1782</span>  float4 a2 = vload4(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 2 src0_stride_y + zin.s2));</div><div class="line"><a name="l01783"></a><span class="lineno"> 1783</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01784"></a><span class="lineno"> 1784</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01785"></a><span class="lineno"> 1785</span>  float4 a3 = vload4(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 3 src0_stride_y + zin.s3));</div><div class="line"><a name="l01786"></a><span class="lineno"> 1786</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01787"></a><span class="lineno"> 1787</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01788"></a><span class="lineno"> 1788</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01789"></a><span class="lineno"> 1789</span>  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="l01790"></a><span class="lineno"> 1790</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01791"></a><span class="lineno"> 1791</span>  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="l01792"></a><span class="lineno"> 1792</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01793"></a><span class="lineno"> 1793</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01794"></a><span class="lineno"> 1794</span>  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="l01795"></a><span class="lineno"> 1795</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01796"></a><span class="lineno"> 1796</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01797"></a><span class="lineno"> 1797</span>  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="l01798"></a><span class="lineno"> 1798</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01799"></a><span class="lineno"> 1799</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01800"></a><span class="lineno"> 1800</span> </div><div class="line"><a name="l01801"></a><span class="lineno"> 1801</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01802"></a><span class="lineno"> 1802</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01803"></a><span class="lineno"> 1803</span> </div><div class="line"><a name="l01804"></a><span class="lineno"> 1804</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01805"></a><span class="lineno"> 1805</span>  acc00 = fma(a0.s0, b0.s0, acc00);</div><div class="line"><a name="l01806"></a><span class="lineno"> 1806</span>  acc01 = fma(a0.s0, b0.s1, acc01);</div><div class="line"><a name="l01807"></a><span class="lineno"> 1807</span>  acc02 = fma(a0.s0, b0.s2, acc02);</div><div class="line"><a name="l01808"></a><span class="lineno"> 1808</span>  acc03 = fma(a0.s0, b0.s3, acc03);</div><div class="line"><a name="l01809"></a><span class="lineno"> 1809</span> </div><div class="line"><a name="l01810"></a><span class="lineno"> 1810</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01811"></a><span class="lineno"> 1811</span> </div><div class="line"><a name="l01812"></a><span class="lineno"> 1812</span>  acc10 = fma(a1.s0, b0.s0, acc10);</div><div class="line"><a name="l01813"></a><span class="lineno"> 1813</span>  acc11 = fma(a1.s0, b0.s1, acc11);</div><div class="line"><a name="l01814"></a><span class="lineno"> 1814</span>  acc12 = fma(a1.s0, b0.s2, acc12);</div><div class="line"><a name="l01815"></a><span class="lineno"> 1815</span>  acc13 = fma(a1.s0, b0.s3, acc13);</div><div class="line"><a name="l01816"></a><span class="lineno"> 1816</span> </div><div class="line"><a name="l01817"></a><span class="lineno"> 1817</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01818"></a><span class="lineno"> 1818</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01819"></a><span class="lineno"> 1819</span> </div><div class="line"><a name="l01820"></a><span class="lineno"> 1820</span>  acc20 = fma(a2.s0, b0.s0, acc20);</div><div class="line"><a name="l01821"></a><span class="lineno"> 1821</span>  acc21 = fma(a2.s0, b0.s1, acc21);</div><div class="line"><a name="l01822"></a><span class="lineno"> 1822</span>  acc22 = fma(a2.s0, b0.s2, acc22);</div><div class="line"><a name="l01823"></a><span class="lineno"> 1823</span>  acc23 = fma(a2.s0, b0.s3, acc23);</div><div class="line"><a name="l01824"></a><span class="lineno"> 1824</span> </div><div class="line"><a name="l01825"></a><span class="lineno"> 1825</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01826"></a><span class="lineno"> 1826</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01827"></a><span class="lineno"> 1827</span> </div><div class="line"><a name="l01828"></a><span class="lineno"> 1828</span>  acc30 = fma(a3.s0, b0.s0, acc30);</div><div class="line"><a name="l01829"></a><span class="lineno"> 1829</span>  acc31 = fma(a3.s0, b0.s1, acc31);</div><div class="line"><a name="l01830"></a><span class="lineno"> 1830</span>  acc32 = fma(a3.s0, b0.s2, acc32);</div><div class="line"><a name="l01831"></a><span class="lineno"> 1831</span>  acc33 = fma(a3.s0, b0.s3, acc33);</div><div class="line"><a name="l01832"></a><span class="lineno"> 1832</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01833"></a><span class="lineno"> 1833</span> </div><div class="line"><a name="l01834"></a><span class="lineno"> 1834</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01835"></a><span class="lineno"> 1835</span>  b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01836"></a><span class="lineno"> 1836</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01837"></a><span class="lineno"> 1837</span> </div><div class="line"><a name="l01838"></a><span class="lineno"> 1838</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01839"></a><span class="lineno"> 1839</span>  acc00 = fma(a0.s1, b0.s0, acc00);</div><div class="line"><a name="l01840"></a><span class="lineno"> 1840</span>  acc01 = fma(a0.s1, b0.s1, acc01);</div><div class="line"><a name="l01841"></a><span class="lineno"> 1841</span>  acc02 = fma(a0.s1, b0.s2, acc02);</div><div class="line"><a name="l01842"></a><span class="lineno"> 1842</span>  acc03 = fma(a0.s1, b0.s3, acc03);</div><div class="line"><a name="l01843"></a><span class="lineno"> 1843</span> </div><div class="line"><a name="l01844"></a><span class="lineno"> 1844</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01845"></a><span class="lineno"> 1845</span> </div><div class="line"><a name="l01846"></a><span class="lineno"> 1846</span>  acc10 = fma(a1.s1, b0.s0, acc10);</div><div class="line"><a name="l01847"></a><span class="lineno"> 1847</span>  acc11 = fma(a1.s1, b0.s1, acc11);</div><div class="line"><a name="l01848"></a><span class="lineno"> 1848</span>  acc12 = fma(a1.s1, b0.s2, acc12);</div><div class="line"><a name="l01849"></a><span class="lineno"> 1849</span>  acc13 = fma(a1.s1, b0.s3, acc13);</div><div class="line"><a name="l01850"></a><span class="lineno"> 1850</span> </div><div class="line"><a name="l01851"></a><span class="lineno"> 1851</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01852"></a><span class="lineno"> 1852</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01853"></a><span class="lineno"> 1853</span> </div><div class="line"><a name="l01854"></a><span class="lineno"> 1854</span>  acc20 = fma(a2.s1, b0.s0, acc20);</div><div class="line"><a name="l01855"></a><span class="lineno"> 1855</span>  acc21 = fma(a2.s1, b0.s1, acc21);</div><div class="line"><a name="l01856"></a><span class="lineno"> 1856</span>  acc22 = fma(a2.s1, b0.s2, acc22);</div><div class="line"><a name="l01857"></a><span class="lineno"> 1857</span>  acc23 = fma(a2.s1, b0.s3, acc23);</div><div class="line"><a name="l01858"></a><span class="lineno"> 1858</span> </div><div class="line"><a name="l01859"></a><span class="lineno"> 1859</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01860"></a><span class="lineno"> 1860</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01861"></a><span class="lineno"> 1861</span> </div><div class="line"><a name="l01862"></a><span class="lineno"> 1862</span>  acc30 = fma(a3.s1, b0.s0, acc30);</div><div class="line"><a name="l01863"></a><span class="lineno"> 1863</span>  acc31 = fma(a3.s1, b0.s1, acc31);</div><div class="line"><a name="l01864"></a><span class="lineno"> 1864</span>  acc32 = fma(a3.s1, b0.s2, acc32);</div><div class="line"><a name="l01865"></a><span class="lineno"> 1865</span>  acc33 = fma(a3.s1, b0.s3, acc33);</div><div class="line"><a name="l01866"></a><span class="lineno"> 1866</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01867"></a><span class="lineno"> 1867</span> </div><div class="line"><a name="l01868"></a><span class="lineno"> 1868</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01869"></a><span class="lineno"> 1869</span>  b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01870"></a><span class="lineno"> 1870</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01871"></a><span class="lineno"> 1871</span> </div><div class="line"><a name="l01872"></a><span class="lineno"> 1872</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01873"></a><span class="lineno"> 1873</span>  acc00 = fma(a0.s2, b0.s0, acc00);</div><div class="line"><a name="l01874"></a><span class="lineno"> 1874</span>  acc01 = fma(a0.s2, b0.s1, acc01);</div><div class="line"><a name="l01875"></a><span class="lineno"> 1875</span>  acc02 = fma(a0.s2, b0.s2, acc02);</div><div class="line"><a name="l01876"></a><span class="lineno"> 1876</span>  acc03 = fma(a0.s2, b0.s3, acc03);</div><div class="line"><a name="l01877"></a><span class="lineno"> 1877</span> </div><div class="line"><a name="l01878"></a><span class="lineno"> 1878</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01879"></a><span class="lineno"> 1879</span> </div><div class="line"><a name="l01880"></a><span class="lineno"> 1880</span>  acc10 = fma(a1.s2, b0.s0, acc10);</div><div class="line"><a name="l01881"></a><span class="lineno"> 1881</span>  acc11 = fma(a1.s2, b0.s1, acc11);</div><div class="line"><a name="l01882"></a><span class="lineno"> 1882</span>  acc12 = fma(a1.s2, b0.s2, acc12);</div><div class="line"><a name="l01883"></a><span class="lineno"> 1883</span>  acc13 = fma(a1.s2, b0.s3, acc13);</div><div class="line"><a name="l01884"></a><span class="lineno"> 1884</span> </div><div class="line"><a name="l01885"></a><span class="lineno"> 1885</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01886"></a><span class="lineno"> 1886</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01887"></a><span class="lineno"> 1887</span> </div><div class="line"><a name="l01888"></a><span class="lineno"> 1888</span>  acc20 = fma(a2.s2, b0.s0, acc20);</div><div class="line"><a name="l01889"></a><span class="lineno"> 1889</span>  acc21 = fma(a2.s2, b0.s1, acc21);</div><div class="line"><a name="l01890"></a><span class="lineno"> 1890</span>  acc22 = fma(a2.s2, b0.s2, acc22);</div><div class="line"><a name="l01891"></a><span class="lineno"> 1891</span>  acc23 = fma(a2.s2, b0.s3, acc23);</div><div class="line"><a name="l01892"></a><span class="lineno"> 1892</span> </div><div class="line"><a name="l01893"></a><span class="lineno"> 1893</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01894"></a><span class="lineno"> 1894</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01895"></a><span class="lineno"> 1895</span> </div><div class="line"><a name="l01896"></a><span class="lineno"> 1896</span>  acc30 = fma(a3.s2, b0.s0, acc30);</div><div class="line"><a name="l01897"></a><span class="lineno"> 1897</span>  acc31 = fma(a3.s2, b0.s1, acc31);</div><div class="line"><a name="l01898"></a><span class="lineno"> 1898</span>  acc32 = fma(a3.s2, b0.s2, acc32);</div><div class="line"><a name="l01899"></a><span class="lineno"> 1899</span>  acc33 = fma(a3.s2, b0.s3, acc33);</div><div class="line"><a name="l01900"></a><span class="lineno"> 1900</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01901"></a><span class="lineno"> 1901</span> </div><div class="line"><a name="l01902"></a><span class="lineno"> 1902</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01903"></a><span class="lineno"> 1903</span>  b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01904"></a><span class="lineno"> 1904</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01905"></a><span class="lineno"> 1905</span> </div><div class="line"><a name="l01906"></a><span class="lineno"> 1906</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01907"></a><span class="lineno"> 1907</span>  acc00 = fma(a0.s3, b0.s0, acc00);</div><div class="line"><a name="l01908"></a><span class="lineno"> 1908</span>  acc01 = fma(a0.s3, b0.s1, acc01);</div><div class="line"><a name="l01909"></a><span class="lineno"> 1909</span>  acc02 = fma(a0.s3, b0.s2, acc02);</div><div class="line"><a name="l01910"></a><span class="lineno"> 1910</span>  acc03 = fma(a0.s3, b0.s3, acc03);</div><div class="line"><a name="l01911"></a><span class="lineno"> 1911</span> </div><div class="line"><a name="l01912"></a><span class="lineno"> 1912</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01913"></a><span class="lineno"> 1913</span> </div><div class="line"><a name="l01914"></a><span class="lineno"> 1914</span>  acc10 = fma(a1.s3, b0.s0, acc10);</div><div class="line"><a name="l01915"></a><span class="lineno"> 1915</span>  acc11 = fma(a1.s3, b0.s1, acc11);</div><div class="line"><a name="l01916"></a><span class="lineno"> 1916</span>  acc12 = fma(a1.s3, b0.s2, acc12);</div><div class="line"><a name="l01917"></a><span class="lineno"> 1917</span>  acc13 = fma(a1.s3, b0.s3, acc13);</div><div class="line"><a name="l01918"></a><span class="lineno"> 1918</span> </div><div class="line"><a name="l01919"></a><span class="lineno"> 1919</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01920"></a><span class="lineno"> 1920</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01921"></a><span class="lineno"> 1921</span> </div><div class="line"><a name="l01922"></a><span class="lineno"> 1922</span>  acc20 = fma(a2.s3, b0.s0, acc20);</div><div class="line"><a name="l01923"></a><span class="lineno"> 1923</span>  acc21 = fma(a2.s3, b0.s1, acc21);</div><div class="line"><a name="l01924"></a><span class="lineno"> 1924</span>  acc22 = fma(a2.s3, b0.s2, acc22);</div><div class="line"><a name="l01925"></a><span class="lineno"> 1925</span>  acc23 = fma(a2.s3, b0.s3, acc23);</div><div class="line"><a name="l01926"></a><span class="lineno"> 1926</span> </div><div class="line"><a name="l01927"></a><span class="lineno"> 1927</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01928"></a><span class="lineno"> 1928</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01929"></a><span class="lineno"> 1929</span> </div><div class="line"><a name="l01930"></a><span class="lineno"> 1930</span>  acc30 = fma(a3.s3, b0.s0, acc30);</div><div class="line"><a name="l01931"></a><span class="lineno"> 1931</span>  acc31 = fma(a3.s3, b0.s1, acc31);</div><div class="line"><a name="l01932"></a><span class="lineno"> 1932</span>  acc32 = fma(a3.s3, b0.s2, acc32);</div><div class="line"><a name="l01933"></a><span class="lineno"> 1933</span>  acc33 = fma(a3.s3, b0.s3, acc33);</div><div class="line"><a name="l01934"></a><span class="lineno"> 1934</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01935"></a><span class="lineno"> 1935</span> </div><div class="line"><a name="l01936"></a><span class="lineno"> 1936</span>  src_addr.s0 += 4 * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01937"></a><span class="lineno"> 1937</span>  }</div><div class="line"><a name="l01938"></a><span class="lineno"> 1938</span> </div><div class="line"><a name="l01939"></a><span class="lineno"> 1939</span>  <span class="keywordflow">for</span>(; i < (int)COLS_A; ++i)</div><div class="line"><a name="l01940"></a><span class="lineno"> 1940</span>  {</div><div class="line"><a name="l01941"></a><span class="lineno"> 1941</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01942"></a><span class="lineno"> 1942</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01943"></a><span class="lineno"> 1943</span>  <span class="keywordtype">float</span> a0 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));</div><div class="line"><a name="l01944"></a><span class="lineno"> 1944</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01945"></a><span class="lineno"> 1945</span>  <span class="keywordtype">float</span> a1 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));</div><div class="line"><a name="l01946"></a><span class="lineno"> 1946</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01947"></a><span class="lineno"> 1947</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01948"></a><span class="lineno"> 1948</span>  <span class="keywordtype">float</span> a2 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));</div><div class="line"><a name="l01949"></a><span class="lineno"> 1949</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01950"></a><span class="lineno"> 1950</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01951"></a><span class="lineno"> 1951</span>  <span class="keywordtype">float</span> a3 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));</div><div class="line"><a name="l01952"></a><span class="lineno"> 1952</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01953"></a><span class="lineno"> 1953</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01954"></a><span class="lineno"> 1954</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01955"></a><span class="lineno"> 1955</span>  <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="l01956"></a><span class="lineno"> 1956</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01957"></a><span class="lineno"> 1957</span>  <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="l01958"></a><span class="lineno"> 1958</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01959"></a><span class="lineno"> 1959</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01960"></a><span class="lineno"> 1960</span>  <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="l01961"></a><span class="lineno"> 1961</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01962"></a><span class="lineno"> 1962</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01963"></a><span class="lineno"> 1963</span>  <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="l01964"></a><span class="lineno"> 1964</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01965"></a><span class="lineno"> 1965</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01966"></a><span class="lineno"> 1966</span> </div><div class="line"><a name="l01967"></a><span class="lineno"> 1967</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01968"></a><span class="lineno"> 1968</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01969"></a><span class="lineno"> 1969</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01970"></a><span class="lineno"> 1970</span> </div><div class="line"><a name="l01971"></a><span class="lineno"> 1971</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l01972"></a><span class="lineno"> 1972</span>  acc00 = fma(a0, b0.s0, acc00);</div><div class="line"><a name="l01973"></a><span class="lineno"> 1973</span>  acc01 = fma(a0, b0.s1, acc01);</div><div class="line"><a name="l01974"></a><span class="lineno"> 1974</span>  acc02 = fma(a0, b0.s2, acc02);</div><div class="line"><a name="l01975"></a><span class="lineno"> 1975</span>  acc03 = fma(a0, b0.s3, acc03);</div><div class="line"><a name="l01976"></a><span class="lineno"> 1976</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01977"></a><span class="lineno"> 1977</span>  acc10 = fma(a1, b0.s0, acc10);</div><div class="line"><a name="l01978"></a><span class="lineno"> 1978</span>  acc11 = fma(a1, b0.s1, acc11);</div><div class="line"><a name="l01979"></a><span class="lineno"> 1979</span>  acc12 = fma(a1, b0.s2, acc12);</div><div class="line"><a name="l01980"></a><span class="lineno"> 1980</span>  acc13 = fma(a1, b0.s3, acc13);</div><div class="line"><a name="l01981"></a><span class="lineno"> 1981</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01982"></a><span class="lineno"> 1982</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01983"></a><span class="lineno"> 1983</span>  acc20 = fma(a2, b0.s0, acc20);</div><div class="line"><a name="l01984"></a><span class="lineno"> 1984</span>  acc21 = fma(a2, b0.s1, acc21);</div><div class="line"><a name="l01985"></a><span class="lineno"> 1985</span>  acc22 = fma(a2, b0.s2, acc22);</div><div class="line"><a name="l01986"></a><span class="lineno"> 1986</span>  acc23 = fma(a2, b0.s3, acc23);</div><div class="line"><a name="l01987"></a><span class="lineno"> 1987</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01988"></a><span class="lineno"> 1988</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01989"></a><span class="lineno"> 1989</span>  acc30 = fma(a3, b0.s0, acc30);</div><div class="line"><a name="l01990"></a><span class="lineno"> 1990</span>  acc31 = fma(a3, b0.s1, acc31);</div><div class="line"><a name="l01991"></a><span class="lineno"> 1991</span>  acc32 = fma(a3, b0.s2, acc32);</div><div class="line"><a name="l01992"></a><span class="lineno"> 1992</span>  acc33 = fma(a3, b0.s3, acc33);</div><div class="line"><a name="l01993"></a><span class="lineno"> 1993</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01994"></a><span class="lineno"> 1994</span> </div><div class="line"><a name="l01995"></a><span class="lineno"> 1995</span>  src_addr.s0 += <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01996"></a><span class="lineno"> 1996</span>  }</div><div class="line"><a name="l01997"></a><span class="lineno"> 1997</span> </div><div class="line"><a name="l01998"></a><span class="lineno"> 1998</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01999"></a><span class="lineno"> 1999</span> </div><div class="line"><a name="l02000"></a><span class="lineno"> 2000</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02001"></a><span class="lineno"> 2001</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02002"></a><span class="lineno"> 2002</span> </div><div class="line"><a name="l02003"></a><span class="lineno"> 2003</span>  <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l02004"></a><span class="lineno"> 2004</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02005"></a><span class="lineno"> 2005</span>  acc00 = acc00 ALPHA;</div><div class="line"><a name="l02006"></a><span class="lineno"> 2006</span>  acc01 = acc01 * ALPHA;</div><div class="line"><a name="l02007"></a><span class="lineno"> 2007</span>  acc02 = acc02 * ALPHA;</div><div class="line"><a name="l02008"></a><span class="lineno"> 2008</span>  acc03 = acc03 * ALPHA;</div><div class="line"><a name="l02009"></a><span class="lineno"> 2009</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02010"></a><span class="lineno"> 2010</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l02011"></a><span class="lineno"> 2011</span>  acc10 = acc10 * ALPHA;</div><div class="line"><a name="l02012"></a><span class="lineno"> 2012</span>  acc11 = acc11 * ALPHA;</div><div class="line"><a name="l02013"></a><span class="lineno"> 2013</span>  acc12 = acc12 * ALPHA;</div><div class="line"><a name="l02014"></a><span class="lineno"> 2014</span>  acc13 = acc13 * ALPHA;</div><div class="line"><a name="l02015"></a><span class="lineno"> 2015</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l02016"></a><span class="lineno"> 2016</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l02017"></a><span class="lineno"> 2017</span>  acc20 = acc20 * ALPHA;</div><div class="line"><a name="l02018"></a><span class="lineno"> 2018</span>  acc21 = acc21 * ALPHA;</div><div class="line"><a name="l02019"></a><span class="lineno"> 2019</span>  acc22 = acc22 * ALPHA;</div><div class="line"><a name="l02020"></a><span class="lineno"> 2020</span>  acc23 = acc23 * ALPHA;</div><div class="line"><a name="l02021"></a><span class="lineno"> 2021</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l02022"></a><span class="lineno"> 2022</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l02023"></a><span class="lineno"> 2023</span>  acc30 = acc30 * ALPHA;</div><div class="line"><a name="l02024"></a><span class="lineno"> 2024</span>  acc31 = acc31 * ALPHA;</div><div class="line"><a name="l02025"></a><span class="lineno"> 2025</span>  acc32 = acc32 * ALPHA;</div><div class="line"><a name="l02026"></a><span class="lineno"> 2026</span>  acc33 = acc33 * ALPHA;</div><div class="line"><a name="l02027"></a><span class="lineno"> 2027</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l02028"></a><span class="lineno"> 2028</span> </div><div class="line"><a name="l02029"></a><span class="lineno"> 2029</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02030"></a><span class="lineno"> 2030</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l02031"></a><span class="lineno"> 2031</span> </div><div class="line"><a name="l02032"></a><span class="lineno"> 2032</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02033"></a><span class="lineno"> 2033</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02034"></a><span class="lineno"> 2034</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02035"></a><span class="lineno"> 2035</span>  <span class="comment">//</span></div><div class="line"><a name="l02036"></a><span class="lineno"> 2036</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02037"></a><span class="lineno"> 2037</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02038"></a><span class="lineno"> 2038</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02039"></a><span class="lineno"> 2039</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02040"></a><span class="lineno"> 2040</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02041"></a><span class="lineno"> 2041</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02042"></a><span class="lineno"> 2042</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02043"></a><span class="lineno"> 2043</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02044"></a><span class="lineno"> 2044</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02045"></a><span class="lineno"> 2045</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02046"></a><span class="lineno"> 2046</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02047"></a><span class="lineno"> 2047</span> </div><div class="line"><a name="l02048"></a><span class="lineno"> 2048</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02049"></a><span class="lineno"> 2049</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02050"></a><span class="lineno"> 2050</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l02051"></a><span class="lineno"> 2051</span> </div><div class="line"><a name="l02052"></a><span class="lineno"> 2052</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02053"></a><span class="lineno"> 2053</span>  zout = (dst_cross_plane_pad dst_stride_y);</div><div class="line"><a name="l02054"></a><span class="lineno"> 2054</span> </div><div class="line"><a name="l02055"></a><span class="lineno"> 2055</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02056"></a><span class="lineno"> 2056</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02057"></a><span class="lineno"> 2057</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02058"></a><span class="lineno"> 2058</span> </div><div class="line"><a name="l02059"></a><span class="lineno"> 2059</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02060"></a><span class="lineno"> 2060</span>  vstore4((float4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l02061"></a><span class="lineno"> 2061</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02062"></a><span class="lineno"> 2062</span>  vstore4((float4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l02063"></a><span class="lineno"> 2063</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02064"></a><span class="lineno"> 2064</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02065"></a><span class="lineno"> 2065</span>  vstore4((float4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l02066"></a><span class="lineno"> 2066</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02067"></a><span class="lineno"> 2067</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02068"></a><span class="lineno"> 2068</span>  vstore4((float4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l02069"></a><span class="lineno"> 2069</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02070"></a><span class="lineno"> 2070</span> </div><div class="line"><a name="l02071"></a><span class="lineno"> 2071</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02072"></a><span class="lineno"> 2072</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02073"></a><span class="lineno"> 2073</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l02074"></a><span class="lineno"> 2074</span> </div><div class="line"><a name="l02075"></a><span class="lineno"> 2075</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02076"></a><span class="lineno"> 2076</span>  vstore4((float4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y));</div><div class="line"><a name="l02077"></a><span class="lineno"> 2077</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02078"></a><span class="lineno"> 2078</span>  vstore4((float4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y));</div><div class="line"><a name="l02079"></a><span class="lineno"> 2079</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02080"></a><span class="lineno"> 2080</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02081"></a><span class="lineno"> 2081</span>  vstore4((float4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y));</div><div class="line"><a name="l02082"></a><span class="lineno"> 2082</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02083"></a><span class="lineno"> 2083</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02084"></a><span class="lineno"> 2084</span>  vstore4((float4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y));</div><div class="line"><a name="l02085"></a><span class="lineno"> 2085</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02086"></a><span class="lineno"> 2086</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02087"></a><span class="lineno"> 2087</span> }</div><div class="line"><a name="l02088"></a><span class="lineno"> 2088</span> </div><div class="line"><a name="l02131"></a><span class="lineno"> 2131</span> __kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f32_bifrost_1000(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02132"></a><span class="lineno"> 2132</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l02133"></a><span class="lineno"> 2133</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l02134"></a><span class="lineno"> 2134</span>  uint src0_stride_z,</div><div class="line"><a name="l02135"></a><span class="lineno"> 2135</span>  uint src1_stride_z,</div><div class="line"><a name="l02136"></a><span class="lineno"> 2136</span>  uint dst_stride_z</div><div class="line"><a name="l02137"></a><span class="lineno"> 2137</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l02138"></a><span class="lineno"> 2138</span>  ,</div><div class="line"><a name="l02139"></a><span class="lineno"> 2139</span>  uint src_cross_plane_pad</div><div class="line"><a name="l02140"></a><span class="lineno"> 2140</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l02141"></a><span class="lineno"> 2141</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l02142"></a><span class="lineno"> 2142</span>  ,</div><div class="line"><a name="l02143"></a><span class="lineno"> 2143</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l02144"></a><span class="lineno"> 2144</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l02145"></a><span class="lineno"> 2145</span>  )</div><div class="line"><a name="l02146"></a><span class="lineno"> 2146</span> {</div><div class="line"><a name="l02147"></a><span class="lineno"> 2147</span>  <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 > 1</span></div><div class="line"><a name="l02148"></a><span class="lineno"> 2148</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l02149"></a><span class="lineno"> 2149</span> </div><div class="line"><a name="l02150"></a><span class="lineno"> 2150</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l02151"></a><span class="lineno"> 2151</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l02152"></a><span class="lineno"> 2152</span> </div><div class="line"><a name="l02153"></a><span class="lineno"> 2153</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l02154"></a><span class="lineno"> 2154</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l02155"></a><span class="lineno"> 2155</span> </div><div class="line"><a name="l02156"></a><span class="lineno"> 2156</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l02157"></a><span class="lineno"> 2157</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l02158"></a><span class="lineno"> 2158</span> </div><div class="line"><a name="l02159"></a><span class="lineno"> 2159</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02160"></a><span class="lineno"> 2160</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02161"></a><span class="lineno"> 2161</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02162"></a><span class="lineno"> 2162</span>  <span class="comment">//</span></div><div class="line"><a name="l02163"></a><span class="lineno"> 2163</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02164"></a><span class="lineno"> 2164</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02165"></a><span class="lineno"> 2165</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02166"></a><span class="lineno"> 2166</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02167"></a><span class="lineno"> 2167</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l02168"></a><span class="lineno"> 2168</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02169"></a><span class="lineno"> 2169</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02170"></a><span class="lineno"> 2170</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02171"></a><span class="lineno"> 2171</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02172"></a><span class="lineno"> 2172</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02173"></a><span class="lineno"> 2173</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02174"></a><span class="lineno"> 2174</span> </div><div class="line"><a name="l02175"></a><span class="lineno"> 2175</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02176"></a><span class="lineno"> 2176</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02177"></a><span class="lineno"> 2177</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l02178"></a><span class="lineno"> 2178</span> </div><div class="line"><a name="l02179"></a><span class="lineno"> 2179</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02180"></a><span class="lineno"> 2180</span>  zin = (src_cross_plane_pad src0_stride_y);</div><div class="line"><a name="l02181"></a><span class="lineno"> 2181</span> </div><div class="line"><a name="l02182"></a><span class="lineno"> 2182</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02183"></a><span class="lineno"> 2183</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02184"></a><span class="lineno"> 2184</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02185"></a><span class="lineno"> 2185</span> </div><div class="line"><a name="l02186"></a><span class="lineno"> 2186</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02187"></a><span class="lineno"> 2187</span> </div><div class="line"><a name="l02188"></a><span class="lineno"> 2188</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02189"></a><span class="lineno"> 2189</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l02190"></a><span class="lineno"> 2190</span> </div><div class="line"><a name="l02191"></a><span class="lineno"> 2191</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02192"></a><span class="lineno"> 2192</span> </div><div class="line"><a name="l02193"></a><span class="lineno"> 2193</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02194"></a><span class="lineno"> 2194</span>  <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="l02195"></a><span class="lineno"> 2195</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l02196"></a><span class="lineno"> 2196</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02197"></a><span class="lineno"> 2197</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l02198"></a><span class="lineno"> 2198</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02199"></a><span class="lineno"> 2199</span> </div><div class="line"><a name="l02200"></a><span class="lineno"> 2200</span>  <span class="comment">// Initialize accumulators</span></div><div class="line"><a name="l02201"></a><span class="lineno"> 2201</span>  <span class="keywordtype">float</span> acc00 = 0.0f;</div><div class="line"><a name="l02202"></a><span class="lineno"> 2202</span>  <span class="keywordtype">float</span> acc01 = 0.0f;</div><div class="line"><a name="l02203"></a><span class="lineno"> 2203</span> </div><div class="line"><a name="l02204"></a><span class="lineno"> 2204</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02205"></a><span class="lineno"> 2205</span>  <span class="keywordtype">float</span> acc10 = 0.0f;</div><div class="line"><a name="l02206"></a><span class="lineno"> 2206</span>  <span class="keywordtype">float</span> acc11 = 0.0f;</div><div class="line"><a name="l02207"></a><span class="lineno"> 2207</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02208"></a><span class="lineno"> 2208</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02209"></a><span class="lineno"> 2209</span>  <span class="keywordtype">float</span> acc20 = 0.0f;</div><div class="line"><a name="l02210"></a><span class="lineno"> 2210</span>  <span class="keywordtype">float</span> acc21 = 0.0f;</div><div class="line"><a name="l02211"></a><span class="lineno"> 2211</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02212"></a><span class="lineno"> 2212</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02213"></a><span class="lineno"> 2213</span>  <span class="keywordtype">float</span> acc30 = 0.0f;</div><div class="line"><a name="l02214"></a><span class="lineno"> 2214</span>  <span class="keywordtype">float</span> acc31 = 0.0f;</div><div class="line"><a name="l02215"></a><span class="lineno"> 2215</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02216"></a><span class="lineno"> 2216</span> </div><div class="line"><a name="l02217"></a><span class="lineno"> 2217</span>  <span class="comment">// A and B src indices get incremented at the same time.</span></div><div class="line"><a name="l02218"></a><span class="lineno"> 2218</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l02219"></a><span class="lineno"> 2219</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 8); i += 8)</div><div class="line"><a name="l02220"></a><span class="lineno"> 2220</span>  {</div><div class="line"><a name="l02221"></a><span class="lineno"> 2221</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02222"></a><span class="lineno"> 2222</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02223"></a><span class="lineno"> 2223</span>  float8 a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + zin.s0));</div><div class="line"><a name="l02224"></a><span class="lineno"> 2224</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02225"></a><span class="lineno"> 2225</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02226"></a><span class="lineno"> 2226</span>  float8 a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0));</div><div class="line"><a name="l02227"></a><span class="lineno"> 2227</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02228"></a><span class="lineno"> 2228</span> </div><div class="line"><a name="l02229"></a><span class="lineno"> 2229</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02230"></a><span class="lineno"> 2230</span>  float2 b0 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02231"></a><span class="lineno"> 2231</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02232"></a><span class="lineno"> 2232</span>  float2 b1 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02233"></a><span class="lineno"> 2233</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02234"></a><span class="lineno"> 2234</span>  float2 b2 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02235"></a><span class="lineno"> 2235</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02236"></a><span class="lineno"> 2236</span>  float2 b3 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02237"></a><span class="lineno"> 2237</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02238"></a><span class="lineno"> 2238</span>  float2 b4 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02239"></a><span class="lineno"> 2239</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02240"></a><span class="lineno"> 2240</span>  float2 b5 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02241"></a><span class="lineno"> 2241</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02242"></a><span class="lineno"> 2242</span>  float2 b6 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02243"></a><span class="lineno"> 2243</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02244"></a><span class="lineno"> 2244</span>  float2 b7 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02245"></a><span class="lineno"> 2245</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02246"></a><span class="lineno"> 2246</span> </div><div class="line"><a name="l02247"></a><span class="lineno"> 2247</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l02248"></a><span class="lineno"> 2248</span>  acc00 = fma(a0.s0, b0.s0, acc00);</div><div class="line"><a name="l02249"></a><span class="lineno"> 2249</span>  acc00 = fma(a0.s1, b1.s0, acc00);</div><div class="line"><a name="l02250"></a><span class="lineno"> 2250</span>  acc00 = fma(a0.s2, b2.s0, acc00);</div><div class="line"><a name="l02251"></a><span class="lineno"> 2251</span>  acc00 = fma(a0.s3, b3.s0, acc00);</div><div class="line"><a name="l02252"></a><span class="lineno"> 2252</span>  acc00 = fma(a0.s4, b4.s0, acc00);</div><div class="line"><a name="l02253"></a><span class="lineno"> 2253</span>  acc00 = fma(a0.s5, b5.s0, acc00);</div><div class="line"><a name="l02254"></a><span class="lineno"> 2254</span>  acc00 = fma(a0.s6, b6.s0, acc00);</div><div class="line"><a name="l02255"></a><span class="lineno"> 2255</span>  acc00 = fma(a0.s7, b7.s0, acc00);</div><div class="line"><a name="l02256"></a><span class="lineno"> 2256</span> </div><div class="line"><a name="l02257"></a><span class="lineno"> 2257</span>  acc01 = fma(a0.s0, b0.s1, acc01);</div><div class="line"><a name="l02258"></a><span class="lineno"> 2258</span>  acc01 = fma(a0.s1, b1.s1, acc01);</div><div class="line"><a name="l02259"></a><span class="lineno"> 2259</span>  acc01 = fma(a0.s2, b2.s1, acc01);</div><div class="line"><a name="l02260"></a><span class="lineno"> 2260</span>  acc01 = fma(a0.s3, b3.s1, acc01);</div><div class="line"><a name="l02261"></a><span class="lineno"> 2261</span>  acc01 = fma(a0.s4, b4.s1, acc01);</div><div class="line"><a name="l02262"></a><span class="lineno"> 2262</span>  acc01 = fma(a0.s5, b5.s1, acc01);</div><div class="line"><a name="l02263"></a><span class="lineno"> 2263</span>  acc01 = fma(a0.s6, b6.s1, acc01);</div><div class="line"><a name="l02264"></a><span class="lineno"> 2264</span>  acc01 = fma(a0.s7, b7.s1, acc01);</div><div class="line"><a name="l02265"></a><span class="lineno"> 2265</span> </div><div class="line"><a name="l02266"></a><span class="lineno"> 2266</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02267"></a><span class="lineno"> 2267</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02268"></a><span class="lineno"> 2268</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 1 src0_stride_y + zin.s1));</div><div class="line"><a name="l02269"></a><span class="lineno"> 2269</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02270"></a><span class="lineno"> 2270</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 1 src0_stride_y));</div><div class="line"><a name="l02271"></a><span class="lineno"> 2271</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02272"></a><span class="lineno"> 2272</span>  acc10 = fma(a0.s0, b0.s0, acc10);</div><div class="line"><a name="l02273"></a><span class="lineno"> 2273</span>  acc10 = fma(a0.s1, b1.s0, acc10);</div><div class="line"><a name="l02274"></a><span class="lineno"> 2274</span>  acc10 = fma(a0.s2, b2.s0, acc10);</div><div class="line"><a name="l02275"></a><span class="lineno"> 2275</span>  acc10 = fma(a0.s3, b3.s0, acc10);</div><div class="line"><a name="l02276"></a><span class="lineno"> 2276</span>  acc10 = fma(a0.s4, b4.s0, acc10);</div><div class="line"><a name="l02277"></a><span class="lineno"> 2277</span>  acc10 = fma(a0.s5, b5.s0, acc10);</div><div class="line"><a name="l02278"></a><span class="lineno"> 2278</span>  acc10 = fma(a0.s6, b6.s0, acc10);</div><div class="line"><a name="l02279"></a><span class="lineno"> 2279</span>  acc10 = fma(a0.s7, b7.s0, acc10);</div><div class="line"><a name="l02280"></a><span class="lineno"> 2280</span> </div><div class="line"><a name="l02281"></a><span class="lineno"> 2281</span>  acc11 = fma(a0.s0, b0.s1, acc11);</div><div class="line"><a name="l02282"></a><span class="lineno"> 2282</span>  acc11 = fma(a0.s1, b1.s1, acc11);</div><div class="line"><a name="l02283"></a><span class="lineno"> 2283</span>  acc11 = fma(a0.s2, b2.s1, acc11);</div><div class="line"><a name="l02284"></a><span class="lineno"> 2284</span>  acc11 = fma(a0.s3, b3.s1, acc11);</div><div class="line"><a name="l02285"></a><span class="lineno"> 2285</span>  acc11 = fma(a0.s4, b4.s1, acc11);</div><div class="line"><a name="l02286"></a><span class="lineno"> 2286</span>  acc11 = fma(a0.s5, b5.s1, acc11);</div><div class="line"><a name="l02287"></a><span class="lineno"> 2287</span>  acc11 = fma(a0.s6, b6.s1, acc11);</div><div class="line"><a name="l02288"></a><span class="lineno"> 2288</span>  acc11 = fma(a0.s7, b7.s1, acc11);</div><div class="line"><a name="l02289"></a><span class="lineno"> 2289</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02290"></a><span class="lineno"> 2290</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02291"></a><span class="lineno"> 2291</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02292"></a><span class="lineno"> 2292</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 2 src0_stride_y + zin.s2));</div><div class="line"><a name="l02293"></a><span class="lineno"> 2293</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02294"></a><span class="lineno"> 2294</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 2 src0_stride_y));</div><div class="line"><a name="l02295"></a><span class="lineno"> 2295</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02296"></a><span class="lineno"> 2296</span>  acc20 = fma(a0.s0, b0.s0, acc20);</div><div class="line"><a name="l02297"></a><span class="lineno"> 2297</span>  acc20 = fma(a0.s1, b1.s0, acc20);</div><div class="line"><a name="l02298"></a><span class="lineno"> 2298</span>  acc20 = fma(a0.s2, b2.s0, acc20);</div><div class="line"><a name="l02299"></a><span class="lineno"> 2299</span>  acc20 = fma(a0.s3, b3.s0, acc20);</div><div class="line"><a name="l02300"></a><span class="lineno"> 2300</span>  acc20 = fma(a0.s4, b4.s0, acc20);</div><div class="line"><a name="l02301"></a><span class="lineno"> 2301</span>  acc20 = fma(a0.s5, b5.s0, acc20);</div><div class="line"><a name="l02302"></a><span class="lineno"> 2302</span>  acc20 = fma(a0.s6, b6.s0, acc20);</div><div class="line"><a name="l02303"></a><span class="lineno"> 2303</span>  acc20 = fma(a0.s7, b7.s0, acc20);</div><div class="line"><a name="l02304"></a><span class="lineno"> 2304</span> </div><div class="line"><a name="l02305"></a><span class="lineno"> 2305</span>  acc21 = fma(a0.s0, b0.s1, acc21);</div><div class="line"><a name="l02306"></a><span class="lineno"> 2306</span>  acc21 = fma(a0.s1, b1.s1, acc21);</div><div class="line"><a name="l02307"></a><span class="lineno"> 2307</span>  acc21 = fma(a0.s2, b2.s1, acc21);</div><div class="line"><a name="l02308"></a><span class="lineno"> 2308</span>  acc21 = fma(a0.s3, b3.s1, acc21);</div><div class="line"><a name="l02309"></a><span class="lineno"> 2309</span>  acc21 = fma(a0.s4, b4.s1, acc21);</div><div class="line"><a name="l02310"></a><span class="lineno"> 2310</span>  acc21 = fma(a0.s5, b5.s1, acc21);</div><div class="line"><a name="l02311"></a><span class="lineno"> 2311</span>  acc21 = fma(a0.s6, b6.s1, acc21);</div><div class="line"><a name="l02312"></a><span class="lineno"> 2312</span>  acc21 = fma(a0.s7, b7.s1, acc21);</div><div class="line"><a name="l02313"></a><span class="lineno"> 2313</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02314"></a><span class="lineno"> 2314</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02315"></a><span class="lineno"> 2315</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02316"></a><span class="lineno"> 2316</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 3 src0_stride_y + zin.s3));</div><div class="line"><a name="l02317"></a><span class="lineno"> 2317</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02318"></a><span class="lineno"> 2318</span>  a0 = vload8(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 3 src0_stride_y));</div><div class="line"><a name="l02319"></a><span class="lineno"> 2319</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02320"></a><span class="lineno"> 2320</span>  acc30 = fma(a0.s0, b0.s0, acc30);</div><div class="line"><a name="l02321"></a><span class="lineno"> 2321</span>  acc30 = fma(a0.s1, b1.s0, acc30);</div><div class="line"><a name="l02322"></a><span class="lineno"> 2322</span>  acc30 = fma(a0.s2, b2.s0, acc30);</div><div class="line"><a name="l02323"></a><span class="lineno"> 2323</span>  acc30 = fma(a0.s3, b3.s0, acc30);</div><div class="line"><a name="l02324"></a><span class="lineno"> 2324</span>  acc30 = fma(a0.s4, b4.s0, acc30);</div><div class="line"><a name="l02325"></a><span class="lineno"> 2325</span>  acc30 = fma(a0.s5, b5.s0, acc30);</div><div class="line"><a name="l02326"></a><span class="lineno"> 2326</span>  acc30 = fma(a0.s6, b6.s0, acc30);</div><div class="line"><a name="l02327"></a><span class="lineno"> 2327</span>  acc30 = fma(a0.s7, b7.s0, acc30);</div><div class="line"><a name="l02328"></a><span class="lineno"> 2328</span> </div><div class="line"><a name="l02329"></a><span class="lineno"> 2329</span>  acc31 = fma(a0.s0, b0.s1, acc31);</div><div class="line"><a name="l02330"></a><span class="lineno"> 2330</span>  acc31 = fma(a0.s1, b1.s1, acc31);</div><div class="line"><a name="l02331"></a><span class="lineno"> 2331</span>  acc31 = fma(a0.s2, b2.s1, acc31);</div><div class="line"><a name="l02332"></a><span class="lineno"> 2332</span>  acc31 = fma(a0.s3, b3.s1, acc31);</div><div class="line"><a name="l02333"></a><span class="lineno"> 2333</span>  acc31 = fma(a0.s4, b4.s1, acc31);</div><div class="line"><a name="l02334"></a><span class="lineno"> 2334</span>  acc31 = fma(a0.s5, b5.s1, acc31);</div><div class="line"><a name="l02335"></a><span class="lineno"> 2335</span>  acc31 = fma(a0.s6, b6.s1, acc31);</div><div class="line"><a name="l02336"></a><span class="lineno"> 2336</span>  acc31 = fma(a0.s7, b7.s1, acc31);</div><div class="line"><a name="l02337"></a><span class="lineno"> 2337</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02338"></a><span class="lineno"> 2338</span> </div><div class="line"><a name="l02339"></a><span class="lineno"> 2339</span>  src_addr.s0 += <span class="keyword">sizeof</span>(float) * 8;</div><div class="line"><a name="l02340"></a><span class="lineno"> 2340</span>  }</div><div class="line"><a name="l02341"></a><span class="lineno"> 2341</span>  <span class="comment">// float size increment</span></div><div class="line"><a name="l02342"></a><span class="lineno"> 2342</span>  <span class="keywordflow">for</span>(; i < (int)COLS_A; ++i)</div><div class="line"><a name="l02343"></a><span class="lineno"> 2343</span>  {</div><div class="line"><a name="l02344"></a><span class="lineno"> 2344</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02345"></a><span class="lineno"> 2345</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02346"></a><span class="lineno"> 2346</span>  <span class="keywordtype">float</span> a0 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 0 * src0_stride_y + zin.s0));</div><div class="line"><a name="l02347"></a><span class="lineno"> 2347</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02348"></a><span class="lineno"> 2348</span>  <span class="keywordtype">float</span> a1 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 1 * src0_stride_y + zin.s1));</div><div class="line"><a name="l02349"></a><span class="lineno"> 2349</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02350"></a><span class="lineno"> 2350</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02351"></a><span class="lineno"> 2351</span>  <span class="keywordtype">float</span> a2 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 2 * src0_stride_y + zin.s2));</div><div class="line"><a name="l02352"></a><span class="lineno"> 2352</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02353"></a><span class="lineno"> 2353</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02354"></a><span class="lineno"> 2354</span>  <span class="keywordtype">float</span> a3 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0 + 3 * src0_stride_y + zin.s3));</div><div class="line"><a name="l02355"></a><span class="lineno"> 2355</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02356"></a><span class="lineno"> 2356</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02357"></a><span class="lineno"> 2357</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02358"></a><span class="lineno"> 2358</span>  <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="l02359"></a><span class="lineno"> 2359</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02360"></a><span class="lineno"> 2360</span>  <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="l02361"></a><span class="lineno"> 2361</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02362"></a><span class="lineno"> 2362</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02363"></a><span class="lineno"> 2363</span>  <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="l02364"></a><span class="lineno"> 2364</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02365"></a><span class="lineno"> 2365</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02366"></a><span class="lineno"> 2366</span>  <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="l02367"></a><span class="lineno"> 2367</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02368"></a><span class="lineno"> 2368</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02369"></a><span class="lineno"> 2369</span> </div><div class="line"><a name="l02370"></a><span class="lineno"> 2370</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02371"></a><span class="lineno"> 2371</span>  float2 b0 = vload2(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l02372"></a><span class="lineno"> 2372</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02373"></a><span class="lineno"> 2373</span> </div><div class="line"><a name="l02374"></a><span class="lineno"> 2374</span>  <span class="comment">// Multiply and accumulate</span></div><div class="line"><a name="l02375"></a><span class="lineno"> 2375</span>  acc00 = fma(a0, b0.s0, acc00);</div><div class="line"><a name="l02376"></a><span class="lineno"> 2376</span>  acc01 = fma(a0, b0.s1, acc01);</div><div class="line"><a name="l02377"></a><span class="lineno"> 2377</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02378"></a><span class="lineno"> 2378</span>  acc10 = fma(a1, b0.s0, acc10);</div><div class="line"><a name="l02379"></a><span class="lineno"> 2379</span>  acc11 = fma(a1, b0.s1, acc11);</div><div class="line"><a name="l02380"></a><span class="lineno"> 2380</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02381"></a><span class="lineno"> 2381</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02382"></a><span class="lineno"> 2382</span>  acc20 = fma(a2, b0.s0, acc20);</div><div class="line"><a name="l02383"></a><span class="lineno"> 2383</span>  acc21 = fma(a2, b0.s1, acc21);</div><div class="line"><a name="l02384"></a><span class="lineno"> 2384</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02385"></a><span class="lineno"> 2385</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02386"></a><span class="lineno"> 2386</span>  acc30 = fma(a3, b0.s0, acc30);</div><div class="line"><a name="l02387"></a><span class="lineno"> 2387</span>  acc31 = fma(a3, b0.s1, acc31);</div><div class="line"><a name="l02388"></a><span class="lineno"> 2388</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02389"></a><span class="lineno"> 2389</span> </div><div class="line"><a name="l02390"></a><span class="lineno"> 2390</span>  src_addr.s0 += <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l02391"></a><span class="lineno"> 2391</span>  }</div><div class="line"><a name="l02392"></a><span class="lineno"> 2392</span> </div><div class="line"><a name="l02393"></a><span class="lineno"> 2393</span>  <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l02394"></a><span class="lineno"> 2394</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02395"></a><span class="lineno"> 2395</span>  acc00 = acc00 ALPHA;</div><div class="line"><a name="l02396"></a><span class="lineno"> 2396</span>  acc01 = acc01 * ALPHA;</div><div class="line"><a name="l02397"></a><span class="lineno"> 2397</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02398"></a><span class="lineno"> 2398</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l02399"></a><span class="lineno"> 2399</span>  acc10 = acc10 * ALPHA;</div><div class="line"><a name="l02400"></a><span class="lineno"> 2400</span>  acc11 = acc11 * ALPHA;</div><div class="line"><a name="l02401"></a><span class="lineno"> 2401</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l02402"></a><span class="lineno"> 2402</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l02403"></a><span class="lineno"> 2403</span>  acc20 = acc20 * ALPHA;</div><div class="line"><a name="l02404"></a><span class="lineno"> 2404</span>  acc21 = acc21 * ALPHA;</div><div class="line"><a name="l02405"></a><span class="lineno"> 2405</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l02406"></a><span class="lineno"> 2406</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l02407"></a><span class="lineno"> 2407</span>  acc30 = acc30 * ALPHA;</div><div class="line"><a name="l02408"></a><span class="lineno"> 2408</span>  acc31 = acc31 * ALPHA;</div><div class="line"><a name="l02409"></a><span class="lineno"> 2409</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l02410"></a><span class="lineno"> 2410</span> </div><div class="line"><a name="l02411"></a><span class="lineno"> 2411</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02412"></a><span class="lineno"> 2412</span> </div><div class="line"><a name="l02413"></a><span class="lineno"> 2413</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02414"></a><span class="lineno"> 2414</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02415"></a><span class="lineno"> 2415</span> </div><div class="line"><a name="l02416"></a><span class="lineno"> 2416</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02417"></a><span class="lineno"> 2417</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l02418"></a><span class="lineno"> 2418</span> </div><div class="line"><a name="l02419"></a><span class="lineno"> 2419</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02420"></a><span class="lineno"> 2420</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02421"></a><span class="lineno"> 2421</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02422"></a><span class="lineno"> 2422</span>  <span class="comment">//</span></div><div class="line"><a name="l02423"></a><span class="lineno"> 2423</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02424"></a><span class="lineno"> 2424</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02425"></a><span class="lineno"> 2425</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02426"></a><span class="lineno"> 2426</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02427"></a><span class="lineno"> 2427</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02428"></a><span class="lineno"> 2428</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02429"></a><span class="lineno"> 2429</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02430"></a><span class="lineno"> 2430</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02431"></a><span class="lineno"> 2431</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02432"></a><span class="lineno"> 2432</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02433"></a><span class="lineno"> 2433</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02434"></a><span class="lineno"> 2434</span> </div><div class="line"><a name="l02435"></a><span class="lineno"> 2435</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02436"></a><span class="lineno"> 2436</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02437"></a><span class="lineno"> 2437</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l02438"></a><span class="lineno"> 2438</span> </div><div class="line"><a name="l02439"></a><span class="lineno"> 2439</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02440"></a><span class="lineno"> 2440</span>  zout = (dst_cross_plane_pad dst_stride_y);</div><div class="line"><a name="l02441"></a><span class="lineno"> 2441</span> </div><div class="line"><a name="l02442"></a><span class="lineno"> 2442</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02443"></a><span class="lineno"> 2443</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02444"></a><span class="lineno"> 2444</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02445"></a><span class="lineno"> 2445</span> </div><div class="line"><a name="l02446"></a><span class="lineno"> 2446</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02447"></a><span class="lineno"> 2447</span>  vstore2((float2)(acc00, acc01), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l02448"></a><span class="lineno"> 2448</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02449"></a><span class="lineno"> 2449</span>  vstore2((float2)(acc10, acc11), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l02450"></a><span class="lineno"> 2450</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02451"></a><span class="lineno"> 2451</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02452"></a><span class="lineno"> 2452</span>  vstore2((float2)(acc20, acc21), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l02453"></a><span class="lineno"> 2453</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02454"></a><span class="lineno"> 2454</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02455"></a><span class="lineno"> 2455</span>  vstore2((float2)(acc30, acc31), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l02456"></a><span class="lineno"> 2456</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02457"></a><span class="lineno"> 2457</span> </div><div class="line"><a name="l02458"></a><span class="lineno"> 2458</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02459"></a><span class="lineno"> 2459</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02460"></a><span class="lineno"> 2460</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l02461"></a><span class="lineno"> 2461</span> </div><div class="line"><a name="l02462"></a><span class="lineno"> 2462</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02463"></a><span class="lineno"> 2463</span>  vstore2((float2)(acc00, acc01), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 0 dst_stride_y));</div><div class="line"><a name="l02464"></a><span class="lineno"> 2464</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02465"></a><span class="lineno"> 2465</span>  vstore2((float2)(acc10, acc11), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 1 dst_stride_y));</div><div class="line"><a name="l02466"></a><span class="lineno"> 2466</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02467"></a><span class="lineno"> 2467</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02468"></a><span class="lineno"> 2468</span>  vstore2((float2)(acc20, acc21), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 2 dst_stride_y));</div><div class="line"><a name="l02469"></a><span class="lineno"> 2469</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02470"></a><span class="lineno"> 2470</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02471"></a><span class="lineno"> 2471</span>  vstore2((float2)(acc30, acc31), 0, (__global <span class="keywordtype">float</span> )(dst_addr + 3 dst_stride_y));</div><div class="line"><a name="l02472"></a><span class="lineno"> 2472</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02473"></a><span class="lineno"> 2473</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02474"></a><span class="lineno"> 2474</span> }</div><div class="line"><a name="l02475"></a><span class="lineno"> 2475</span> </div><div class="line"><a name="l02476"></a><span class="lineno"> 2476</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l02477"></a><span class="lineno"> 2477</span> </div><div class="line"><a name="l02518"></a><span class="lineno"> 2518</span> __kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f16_bifrost_acc32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02519"></a><span class="lineno"> 2519</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l02520"></a><span class="lineno"> 2520</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l02521"></a><span class="lineno"> 2521</span>  uint src0_stride_z,</div><div class="line"><a name="l02522"></a><span class="lineno"> 2522</span>  uint src1_stride_z,</div><div class="line"><a name="l02523"></a><span class="lineno"> 2523</span>  uint dst_stride_z</div><div class="line"><a name="l02524"></a><span class="lineno"> 2524</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l02525"></a><span class="lineno"> 2525</span>  ,</div><div class="line"><a name="l02526"></a><span class="lineno"> 2526</span>  uint src_cross_plane_pad</div><div class="line"><a name="l02527"></a><span class="lineno"> 2527</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l02528"></a><span class="lineno"> 2528</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l02529"></a><span class="lineno"> 2529</span>  ,</div><div class="line"><a name="l02530"></a><span class="lineno"> 2530</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l02531"></a><span class="lineno"> 2531</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l02532"></a><span class="lineno"> 2532</span>  )</div><div class="line"><a name="l02533"></a><span class="lineno"> 2533</span> {</div><div class="line"><a name="l02534"></a><span class="lineno"> 2534</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l02535"></a><span class="lineno"> 2535</span> </div><div class="line"><a name="l02536"></a><span class="lineno"> 2536</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l02537"></a><span class="lineno"> 2537</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l02538"></a><span class="lineno"> 2538</span> </div><div class="line"><a name="l02539"></a><span class="lineno"> 2539</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l02540"></a><span class="lineno"> 2540</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l02541"></a><span class="lineno"> 2541</span> </div><div class="line"><a name="l02542"></a><span class="lineno"> 2542</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l02543"></a><span class="lineno"> 2543</span>  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="l02544"></a><span class="lineno"> 2544</span> </div><div class="line"><a name="l02545"></a><span class="lineno"> 2545</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02546"></a><span class="lineno"> 2546</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02547"></a><span class="lineno"> 2547</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02548"></a><span class="lineno"> 2548</span>  <span class="comment">//</span></div><div class="line"><a name="l02549"></a><span class="lineno"> 2549</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02550"></a><span class="lineno"> 2550</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02551"></a><span class="lineno"> 2551</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02552"></a><span class="lineno"> 2552</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02553"></a><span class="lineno"> 2553</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l02554"></a><span class="lineno"> 2554</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02555"></a><span class="lineno"> 2555</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02556"></a><span class="lineno"> 2556</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02557"></a><span class="lineno"> 2557</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02558"></a><span class="lineno"> 2558</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02559"></a><span class="lineno"> 2559</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02560"></a><span class="lineno"> 2560</span> </div><div class="line"><a name="l02561"></a><span class="lineno"> 2561</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02562"></a><span class="lineno"> 2562</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02563"></a><span class="lineno"> 2563</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l02564"></a><span class="lineno"> 2564</span> </div><div class="line"><a name="l02565"></a><span class="lineno"> 2565</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02566"></a><span class="lineno"> 2566</span>  zin = (src_cross_plane_pad src0_stride_y);</div><div class="line"><a name="l02567"></a><span class="lineno"> 2567</span> </div><div class="line"><a name="l02568"></a><span class="lineno"> 2568</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02569"></a><span class="lineno"> 2569</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02570"></a><span class="lineno"> 2570</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02571"></a><span class="lineno"> 2571</span> </div><div class="line"><a name="l02572"></a><span class="lineno"> 2572</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02573"></a><span class="lineno"> 2573</span> </div><div class="line"><a name="l02574"></a><span class="lineno"> 2574</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02575"></a><span class="lineno"> 2575</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l02576"></a><span class="lineno"> 2576</span> </div><div class="line"><a name="l02577"></a><span class="lineno"> 2577</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02578"></a><span class="lineno"> 2578</span> </div><div class="line"><a name="l02579"></a><span class="lineno"> 2579</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02580"></a><span class="lineno"> 2580</span>  <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="l02581"></a><span class="lineno"> 2581</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l02582"></a><span class="lineno"> 2582</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02583"></a><span class="lineno"> 2583</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l02584"></a><span class="lineno"> 2584</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02585"></a><span class="lineno"> 2585</span> </div><div class="line"><a name="l02586"></a><span class="lineno"> 2586</span>  float8 acc0 = 0.0h;</div><div class="line"><a name="l02587"></a><span class="lineno"> 2587</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02588"></a><span class="lineno"> 2588</span>  float8 acc1 = 0.0h;</div><div class="line"><a name="l02589"></a><span class="lineno"> 2589</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02590"></a><span class="lineno"> 2590</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02591"></a><span class="lineno"> 2591</span>  float8 acc2 = 0.0h;</div><div class="line"><a name="l02592"></a><span class="lineno"> 2592</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02593"></a><span class="lineno"> 2593</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02594"></a><span class="lineno"> 2594</span>  float8 acc3 = 0.0h;</div><div class="line"><a name="l02595"></a><span class="lineno"> 2595</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02596"></a><span class="lineno"> 2596</span> </div><div class="line"><a name="l02597"></a><span class="lineno"> 2597</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l02598"></a><span class="lineno"> 2598</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 4); i += 4)</div><div class="line"><a name="l02599"></a><span class="lineno"> 2599</span>  {</div><div class="line"><a name="l02600"></a><span class="lineno"> 2600</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02601"></a><span class="lineno"> 2601</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02602"></a><span class="lineno"> 2602</span>  half4 a0 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 0 src0_stride_y + zin.s0));</div><div class="line"><a name="l02603"></a><span class="lineno"> 2603</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02604"></a><span class="lineno"> 2604</span>  half4 a1 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 1 src0_stride_y + zin.s1));</div><div class="line"><a name="l02605"></a><span class="lineno"> 2605</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02606"></a><span class="lineno"> 2606</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02607"></a><span class="lineno"> 2607</span>  half4 a2 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 2 src0_stride_y + zin.s2));</div><div class="line"><a name="l02608"></a><span class="lineno"> 2608</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02609"></a><span class="lineno"> 2609</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02610"></a><span class="lineno"> 2610</span>  half4 a3 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 3 src0_stride_y + zin.s3));</div><div class="line"><a name="l02611"></a><span class="lineno"> 2611</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02612"></a><span class="lineno"> 2612</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02613"></a><span class="lineno"> 2613</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02614"></a><span class="lineno"> 2614</span>  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="l02615"></a><span class="lineno"> 2615</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02616"></a><span class="lineno"> 2616</span>  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="l02617"></a><span class="lineno"> 2617</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02618"></a><span class="lineno"> 2618</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02619"></a><span class="lineno"> 2619</span>  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="l02620"></a><span class="lineno"> 2620</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02621"></a><span class="lineno"> 2621</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02622"></a><span class="lineno"> 2622</span>  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="l02623"></a><span class="lineno"> 2623</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02624"></a><span class="lineno"> 2624</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02625"></a><span class="lineno"> 2625</span> </div><div class="line"><a name="l02626"></a><span class="lineno"> 2626</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02627"></a><span class="lineno"> 2627</span>  float8 b0 = convert_float8(vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src1_ptr + src_addr.s1)));</div><div class="line"><a name="l02628"></a><span class="lineno"> 2628</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02629"></a><span class="lineno"> 2629</span> </div><div class="line"><a name="l02630"></a><span class="lineno"> 2630</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l02631"></a><span class="lineno"> 2631</span>  acc0 = fma(b0, (float8)a0.s0, acc0);</div><div class="line"><a name="l02632"></a><span class="lineno"> 2632</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02633"></a><span class="lineno"> 2633</span>  acc1 = fma(b0, (float8)a1.s0, acc1);</div><div class="line"><a name="l02634"></a><span class="lineno"> 2634</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02635"></a><span class="lineno"> 2635</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02636"></a><span class="lineno"> 2636</span>  acc2 = fma(b0, (float8)a2.s0, acc2);</div><div class="line"><a name="l02637"></a><span class="lineno"> 2637</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02638"></a><span class="lineno"> 2638</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02639"></a><span class="lineno"> 2639</span>  acc3 = fma(b0, (float8)a3.s0, acc3);</div><div class="line"><a name="l02640"></a><span class="lineno"> 2640</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02641"></a><span class="lineno"> 2641</span> </div><div class="line"><a name="l02642"></a><span class="lineno"> 2642</span>  b0 = convert_float8(vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src1_ptr + src_addr.s1)));</div><div class="line"><a name="l02643"></a><span class="lineno"> 2643</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02644"></a><span class="lineno"> 2644</span>  acc0 = fma(b0, (float8)a0.s1, acc0);</div><div class="line"><a name="l02645"></a><span class="lineno"> 2645</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02646"></a><span class="lineno"> 2646</span>  acc1 = fma(b0, (float8)a1.s1, acc1);</div><div class="line"><a name="l02647"></a><span class="lineno"> 2647</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02648"></a><span class="lineno"> 2648</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02649"></a><span class="lineno"> 2649</span>  acc2 = fma(b0, (float8)a2.s1, acc2);</div><div class="line"><a name="l02650"></a><span class="lineno"> 2650</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02651"></a><span class="lineno"> 2651</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02652"></a><span class="lineno"> 2652</span>  acc3 = fma(b0, (float8)a3.s1, acc3);</div><div class="line"><a name="l02653"></a><span class="lineno"> 2653</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02654"></a><span class="lineno"> 2654</span> </div><div class="line"><a name="l02655"></a><span class="lineno"> 2655</span>  b0 = convert_float8(vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src1_ptr + src_addr.s1)));</div><div class="line"><a name="l02656"></a><span class="lineno"> 2656</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02657"></a><span class="lineno"> 2657</span>  acc0 = fma(b0, (float8)a0.s2, acc0);</div><div class="line"><a name="l02658"></a><span class="lineno"> 2658</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02659"></a><span class="lineno"> 2659</span>  acc1 = fma(b0, (float8)a1.s2, acc1);</div><div class="line"><a name="l02660"></a><span class="lineno"> 2660</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02661"></a><span class="lineno"> 2661</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02662"></a><span class="lineno"> 2662</span>  acc2 = fma(b0, (float8)a2.s2, acc2);</div><div class="line"><a name="l02663"></a><span class="lineno"> 2663</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02664"></a><span class="lineno"> 2664</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02665"></a><span class="lineno"> 2665</span>  acc3 = fma(b0, (float8)a3.s2, acc3);</div><div class="line"><a name="l02666"></a><span class="lineno"> 2666</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02667"></a><span class="lineno"> 2667</span> </div><div class="line"><a name="l02668"></a><span class="lineno"> 2668</span>  b0 = convert_float8(vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src1_ptr + src_addr.s1)));</div><div class="line"><a name="l02669"></a><span class="lineno"> 2669</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02670"></a><span class="lineno"> 2670</span>  acc0 = fma(b0, (float8)a0.s3, acc0);</div><div class="line"><a name="l02671"></a><span class="lineno"> 2671</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02672"></a><span class="lineno"> 2672</span>  acc1 = fma(b0, (float8)a1.s3, acc1);</div><div class="line"><a name="l02673"></a><span class="lineno"> 2673</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02674"></a><span class="lineno"> 2674</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02675"></a><span class="lineno"> 2675</span>  acc2 = fma(b0, (float8)a2.s3, acc2);</div><div class="line"><a name="l02676"></a><span class="lineno"> 2676</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02677"></a><span class="lineno"> 2677</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02678"></a><span class="lineno"> 2678</span>  acc3 = fma(b0, (float8)a3.s3, acc3);</div><div class="line"><a name="l02679"></a><span class="lineno"> 2679</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02680"></a><span class="lineno"> 2680</span> </div><div class="line"><a name="l02681"></a><span class="lineno"> 2681</span>  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="l02682"></a><span class="lineno"> 2682</span>  }</div><div class="line"><a name="l02683"></a><span class="lineno"> 2683</span> </div><div class="line"><a name="l02684"></a><span class="lineno"> 2684</span>  <span class="keywordflow">for</span>(; i < (int)COLS_A; ++i)</div><div class="line"><a name="l02685"></a><span class="lineno"> 2685</span>  {</div><div class="line"><a name="l02686"></a><span class="lineno"> 2686</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02687"></a><span class="lineno"> 2687</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02688"></a><span class="lineno"> 2688</span>  <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 + zin.s0));</div><div class="line"><a name="l02689"></a><span class="lineno"> 2689</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02690"></a><span class="lineno"> 2690</span>  <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 + zin.s1));</div><div class="line"><a name="l02691"></a><span class="lineno"> 2691</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02692"></a><span class="lineno"> 2692</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02693"></a><span class="lineno"> 2693</span>  <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 + zin.s2));</div><div class="line"><a name="l02694"></a><span class="lineno"> 2694</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02695"></a><span class="lineno"> 2695</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02696"></a><span class="lineno"> 2696</span>  <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 + zin.s3));</div><div class="line"><a name="l02697"></a><span class="lineno"> 2697</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02698"></a><span class="lineno"> 2698</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02699"></a><span class="lineno"> 2699</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02700"></a><span class="lineno"> 2700</span>  <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="l02701"></a><span class="lineno"> 2701</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02702"></a><span class="lineno"> 2702</span>  <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="l02703"></a><span class="lineno"> 2703</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02704"></a><span class="lineno"> 2704</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02705"></a><span class="lineno"> 2705</span>  <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="l02706"></a><span class="lineno"> 2706</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02707"></a><span class="lineno"> 2707</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02708"></a><span class="lineno"> 2708</span>  <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="l02709"></a><span class="lineno"> 2709</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02710"></a><span class="lineno"> 2710</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02711"></a><span class="lineno"> 2711</span> </div><div class="line"><a name="l02712"></a><span class="lineno"> 2712</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02713"></a><span class="lineno"> 2713</span>  float8 b0 = convert_float8(vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src1_ptr + src_addr.s1)));</div><div class="line"><a name="l02714"></a><span class="lineno"> 2714</span> </div><div class="line"><a name="l02715"></a><span class="lineno"> 2715</span>  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="l02716"></a><span class="lineno"> 2716</span> </div><div class="line"><a name="l02717"></a><span class="lineno"> 2717</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l02718"></a><span class="lineno"> 2718</span>  acc0 = fma(b0, (float8)a0, acc0); <span class="comment">// b0 (half8)a0;</span></div><div class="line"><a name="l02719"></a><span class="lineno"> 2719</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02720"></a><span class="lineno"> 2720</span>  acc1 = fma(b0, (float8)a1, acc1); <span class="comment">// b0 * (half8)a1;</span></div><div class="line"><a name="l02721"></a><span class="lineno"> 2721</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02722"></a><span class="lineno"> 2722</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02723"></a><span class="lineno"> 2723</span>  acc2 = fma(b0, (float8)a2, acc2); <span class="comment">// b0 * (half8)a2;</span></div><div class="line"><a name="l02724"></a><span class="lineno"> 2724</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02725"></a><span class="lineno"> 2725</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02726"></a><span class="lineno"> 2726</span>  acc3 = fma(b0, (float8)a3, acc3); <span class="comment">// b0 * (half8)a3;</span></div><div class="line"><a name="l02727"></a><span class="lineno"> 2727</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02728"></a><span class="lineno"> 2728</span>  }</div><div class="line"><a name="l02729"></a><span class="lineno"> 2729</span> </div><div class="line"><a name="l02730"></a><span class="lineno"> 2730</span>  <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l02731"></a><span class="lineno"> 2731</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02732"></a><span class="lineno"> 2732</span>  half8 hacc0 = convert_half8(acc0) * (half8)ALPHA;</div><div class="line"><a name="l02733"></a><span class="lineno"> 2733</span> <span class="preprocessor">#else //defined(ALPHA)</span></div><div class="line"><a name="l02734"></a><span class="lineno"> 2734</span>  half8 hacc0 = convert_half8(acc0);</div><div class="line"><a name="l02735"></a><span class="lineno"> 2735</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02736"></a><span class="lineno"> 2736</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02737"></a><span class="lineno"> 2737</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02738"></a><span class="lineno"> 2738</span>  half8 hacc1 = convert_half8(acc1) * (half8)ALPHA;</div><div class="line"><a name="l02739"></a><span class="lineno"> 2739</span> <span class="preprocessor">#else //defined(ALPHA)</span></div><div class="line"><a name="l02740"></a><span class="lineno"> 2740</span>  half8 hacc1 = convert_half8(acc1);</div><div class="line"><a name="l02741"></a><span class="lineno"> 2741</span> <span class="preprocessor">#endif //defined(ALPHA)</span></div><div class="line"><a name="l02742"></a><span class="lineno"> 2742</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y</span></div><div class="line"><a name="l02743"></a><span class="lineno"> 2743</span> </div><div class="line"><a name="l02744"></a><span class="lineno"> 2744</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02745"></a><span class="lineno"> 2745</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02746"></a><span class="lineno"> 2746</span>  half8 hacc2 = convert_half8(acc2) * (half8)ALPHA;</div><div class="line"><a name="l02747"></a><span class="lineno"> 2747</span> <span class="preprocessor">#else //defined(ALPHA)</span></div><div class="line"><a name="l02748"></a><span class="lineno"> 2748</span>  half8 hacc2 = convert_half8(acc2);</div><div class="line"><a name="l02749"></a><span class="lineno"> 2749</span> <span class="preprocessor">#endif //defined(ALPHA)</span></div><div class="line"><a name="l02750"></a><span class="lineno"> 2750</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02751"></a><span class="lineno"> 2751</span> </div><div class="line"><a name="l02752"></a><span class="lineno"> 2752</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02753"></a><span class="lineno"> 2753</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l02754"></a><span class="lineno"> 2754</span>  half8 hacc3 = convert_half8(acc3) * (half8)ALPHA;</div><div class="line"><a name="l02755"></a><span class="lineno"> 2755</span> <span class="preprocessor">#else //defined(ALPHA)</span></div><div class="line"><a name="l02756"></a><span class="lineno"> 2756</span>  half8 hacc3 = convert_half8(acc3);</div><div class="line"><a name="l02757"></a><span class="lineno"> 2757</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l02758"></a><span class="lineno"> 2758</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02759"></a><span class="lineno"> 2759</span> </div><div class="line"><a name="l02760"></a><span class="lineno"> 2760</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02761"></a><span class="lineno"> 2761</span> </div><div class="line"><a name="l02762"></a><span class="lineno"> 2762</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l02763"></a><span class="lineno"> 2763</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02764"></a><span class="lineno"> 2764</span> </div><div class="line"><a name="l02765"></a><span class="lineno"> 2765</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l02766"></a><span class="lineno"> 2766</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l02767"></a><span class="lineno"> 2767</span> </div><div class="line"><a name="l02768"></a><span class="lineno"> 2768</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02769"></a><span class="lineno"> 2769</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02770"></a><span class="lineno"> 2770</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02771"></a><span class="lineno"> 2771</span>  <span class="comment">//</span></div><div class="line"><a name="l02772"></a><span class="lineno"> 2772</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02773"></a><span class="lineno"> 2773</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02774"></a><span class="lineno"> 2774</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02775"></a><span class="lineno"> 2775</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02776"></a><span class="lineno"> 2776</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02777"></a><span class="lineno"> 2777</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02778"></a><span class="lineno"> 2778</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02779"></a><span class="lineno"> 2779</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02780"></a><span class="lineno"> 2780</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02781"></a><span class="lineno"> 2781</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02782"></a><span class="lineno"> 2782</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02783"></a><span class="lineno"> 2783</span> </div><div class="line"><a name="l02784"></a><span class="lineno"> 2784</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02785"></a><span class="lineno"> 2785</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02786"></a><span class="lineno"> 2786</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l02787"></a><span class="lineno"> 2787</span> </div><div class="line"><a name="l02788"></a><span class="lineno"> 2788</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02789"></a><span class="lineno"> 2789</span>  zout = (dst_cross_plane_pad dst_stride_y);</div><div class="line"><a name="l02790"></a><span class="lineno"> 2790</span> </div><div class="line"><a name="l02791"></a><span class="lineno"> 2791</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02792"></a><span class="lineno"> 2792</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02793"></a><span class="lineno"> 2793</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02794"></a><span class="lineno"> 2794</span> </div><div class="line"><a name="l02795"></a><span class="lineno"> 2795</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02796"></a><span class="lineno"> 2796</span>  vstore8(hacc0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l02797"></a><span class="lineno"> 2797</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02798"></a><span class="lineno"> 2798</span>  vstore8(hacc1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l02799"></a><span class="lineno"> 2799</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02800"></a><span class="lineno"> 2800</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02801"></a><span class="lineno"> 2801</span>  vstore8(hacc2, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l02802"></a><span class="lineno"> 2802</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02803"></a><span class="lineno"> 2803</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02804"></a><span class="lineno"> 2804</span>  vstore8(hacc3, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l02805"></a><span class="lineno"> 2805</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02806"></a><span class="lineno"> 2806</span> </div><div class="line"><a name="l02807"></a><span class="lineno"> 2807</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l02808"></a><span class="lineno"> 2808</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02809"></a><span class="lineno"> 2809</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l02810"></a><span class="lineno"> 2810</span> </div><div class="line"><a name="l02811"></a><span class="lineno"> 2811</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l02812"></a><span class="lineno"> 2812</span>  vstore8(hacc0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y));</div><div class="line"><a name="l02813"></a><span class="lineno"> 2813</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02814"></a><span class="lineno"> 2814</span>  vstore8(hacc1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y));</div><div class="line"><a name="l02815"></a><span class="lineno"> 2815</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02816"></a><span class="lineno"> 2816</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02817"></a><span class="lineno"> 2817</span>  vstore8(hacc2, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y));</div><div class="line"><a name="l02818"></a><span class="lineno"> 2818</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02819"></a><span class="lineno"> 2819</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02820"></a><span class="lineno"> 2820</span>  vstore8(hacc3, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y));</div><div class="line"><a name="l02821"></a><span class="lineno"> 2821</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02822"></a><span class="lineno"> 2822</span> <span class="preprocessor">#endif // REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l02823"></a><span class="lineno"> 2823</span> }</div><div class="line"><a name="l02824"></a><span class="lineno"> 2824</span> </div><div class="line"><a name="l02866"></a><span class="lineno"> 2866</span> __kernel <span class="keywordtype">void</span> gemm_mm_floating_point_f16_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l02867"></a><span class="lineno"> 2867</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l02868"></a><span class="lineno"> 2868</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l02869"></a><span class="lineno"> 2869</span>  uint src0_stride_z,</div><div class="line"><a name="l02870"></a><span class="lineno"> 2870</span>  uint src1_stride_z,</div><div class="line"><a name="l02871"></a><span class="lineno"> 2871</span>  uint dst_stride_z</div><div class="line"><a name="l02872"></a><span class="lineno"> 2872</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l02873"></a><span class="lineno"> 2873</span>  ,</div><div class="line"><a name="l02874"></a><span class="lineno"> 2874</span>  uint src_cross_plane_pad</div><div class="line"><a name="l02875"></a><span class="lineno"> 2875</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l02876"></a><span class="lineno"> 2876</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l02877"></a><span class="lineno"> 2877</span>  ,</div><div class="line"><a name="l02878"></a><span class="lineno"> 2878</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l02879"></a><span class="lineno"> 2879</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l02880"></a><span class="lineno"> 2880</span>  )</div><div class="line"><a name="l02881"></a><span class="lineno"> 2881</span> {</div><div class="line"><a name="l02882"></a><span class="lineno"> 2882</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l02883"></a><span class="lineno"> 2883</span> </div><div class="line"><a name="l02884"></a><span class="lineno"> 2884</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l02885"></a><span class="lineno"> 2885</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l02886"></a><span class="lineno"> 2886</span> </div><div class="line"><a name="l02887"></a><span class="lineno"> 2887</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l02888"></a><span class="lineno"> 2888</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l02889"></a><span class="lineno"> 2889</span> </div><div class="line"><a name="l02890"></a><span class="lineno"> 2890</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l02891"></a><span class="lineno"> 2891</span>  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="l02892"></a><span class="lineno"> 2892</span> </div><div class="line"><a name="l02893"></a><span class="lineno"> 2893</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02894"></a><span class="lineno"> 2894</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l02895"></a><span class="lineno"> 2895</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l02896"></a><span class="lineno"> 2896</span>  <span class="comment">//</span></div><div class="line"><a name="l02897"></a><span class="lineno"> 2897</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02898"></a><span class="lineno"> 2898</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l02899"></a><span class="lineno"> 2899</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02900"></a><span class="lineno"> 2900</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02901"></a><span class="lineno"> 2901</span>  <span class="comment">// \|****************\|</span></div><div class="line"><a name="l02902"></a><span class="lineno"> 2902</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l02903"></a><span class="lineno"> 2903</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l02904"></a><span class="lineno"> 2904</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02905"></a><span class="lineno"> 2905</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l02906"></a><span class="lineno"> 2906</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l02907"></a><span class="lineno"> 2907</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l02908"></a><span class="lineno"> 2908</span> </div><div class="line"><a name="l02909"></a><span class="lineno"> 2909</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l02910"></a><span class="lineno"> 2910</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l02911"></a><span class="lineno"> 2911</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l02912"></a><span class="lineno"> 2912</span> </div><div class="line"><a name="l02913"></a><span class="lineno"> 2913</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l02914"></a><span class="lineno"> 2914</span>  zin = (src_cross_plane_pad src0_stride_y);</div><div class="line"><a name="l02915"></a><span class="lineno"> 2915</span> </div><div class="line"><a name="l02916"></a><span class="lineno"> 2916</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l02917"></a><span class="lineno"> 2917</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l02918"></a><span class="lineno"> 2918</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l02919"></a><span class="lineno"> 2919</span> </div><div class="line"><a name="l02920"></a><span class="lineno"> 2920</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02921"></a><span class="lineno"> 2921</span> </div><div class="line"><a name="l02922"></a><span class="lineno"> 2922</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l02923"></a><span class="lineno"> 2923</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l02924"></a><span class="lineno"> 2924</span> </div><div class="line"><a name="l02925"></a><span class="lineno"> 2925</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02926"></a><span class="lineno"> 2926</span> </div><div class="line"><a name="l02927"></a><span class="lineno"> 2927</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02928"></a><span class="lineno"> 2928</span>  <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="l02929"></a><span class="lineno"> 2929</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l02930"></a><span class="lineno"> 2930</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02931"></a><span class="lineno"> 2931</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l02932"></a><span class="lineno"> 2932</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l02933"></a><span class="lineno"> 2933</span> </div><div class="line"><a name="l02934"></a><span class="lineno"> 2934</span>  half8 acc0 = 0.0h;</div><div class="line"><a name="l02935"></a><span class="lineno"> 2935</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02936"></a><span class="lineno"> 2936</span>  half8 acc1 = 0.0h;</div><div class="line"><a name="l02937"></a><span class="lineno"> 2937</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02938"></a><span class="lineno"> 2938</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02939"></a><span class="lineno"> 2939</span>  half8 acc2 = 0.0h;</div><div class="line"><a name="l02940"></a><span class="lineno"> 2940</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02941"></a><span class="lineno"> 2941</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02942"></a><span class="lineno"> 2942</span>  half8 acc3 = 0.0h;</div><div class="line"><a name="l02943"></a><span class="lineno"> 2943</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02944"></a><span class="lineno"> 2944</span> </div><div class="line"><a name="l02945"></a><span class="lineno"> 2945</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l02946"></a><span class="lineno"> 2946</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 4); i += 4)</div><div class="line"><a name="l02947"></a><span class="lineno"> 2947</span>  {</div><div class="line"><a name="l02948"></a><span class="lineno"> 2948</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02949"></a><span class="lineno"> 2949</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02950"></a><span class="lineno"> 2950</span>  half4 a0 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 0 src0_stride_y + zin.s0));</div><div class="line"><a name="l02951"></a><span class="lineno"> 2951</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02952"></a><span class="lineno"> 2952</span>  half4 a1 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 1 src0_stride_y + zin.s1));</div><div class="line"><a name="l02953"></a><span class="lineno"> 2953</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02954"></a><span class="lineno"> 2954</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02955"></a><span class="lineno"> 2955</span>  half4 a2 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 2 src0_stride_y + zin.s2));</div><div class="line"><a name="l02956"></a><span class="lineno"> 2956</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02957"></a><span class="lineno"> 2957</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02958"></a><span class="lineno"> 2958</span>  half4 a3 = vload4(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(src0_ptr + src_addr.s0 + 3 src0_stride_y + zin.s3));</div><div class="line"><a name="l02959"></a><span class="lineno"> 2959</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02960"></a><span class="lineno"> 2960</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02961"></a><span class="lineno"> 2961</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l02962"></a><span class="lineno"> 2962</span>  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="l02963"></a><span class="lineno"> 2963</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02964"></a><span class="lineno"> 2964</span>  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="l02965"></a><span class="lineno"> 2965</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02966"></a><span class="lineno"> 2966</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02967"></a><span class="lineno"> 2967</span>  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="l02968"></a><span class="lineno"> 2968</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02969"></a><span class="lineno"> 2969</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02970"></a><span class="lineno"> 2970</span>  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="l02971"></a><span class="lineno"> 2971</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02972"></a><span class="lineno"> 2972</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l02973"></a><span class="lineno"> 2973</span> </div><div class="line"><a name="l02974"></a><span class="lineno"> 2974</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l02975"></a><span class="lineno"> 2975</span>  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="l02976"></a><span class="lineno"> 2976</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02977"></a><span class="lineno"> 2977</span> </div><div class="line"><a name="l02978"></a><span class="lineno"> 2978</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l02979"></a><span class="lineno"> 2979</span>  acc0 = fma(b0, (half8)a0.s0, acc0);</div><div class="line"><a name="l02980"></a><span class="lineno"> 2980</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02981"></a><span class="lineno"> 2981</span>  acc1 = fma(b0, (half8)a1.s0, acc1);</div><div class="line"><a name="l02982"></a><span class="lineno"> 2982</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02983"></a><span class="lineno"> 2983</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02984"></a><span class="lineno"> 2984</span>  acc2 = fma(b0, (half8)a2.s0, acc2);</div><div class="line"><a name="l02985"></a><span class="lineno"> 2985</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02986"></a><span class="lineno"> 2986</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02987"></a><span class="lineno"> 2987</span>  acc3 = fma(b0, (half8)a3.s0, acc3);</div><div class="line"><a name="l02988"></a><span class="lineno"> 2988</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l02989"></a><span class="lineno"> 2989</span> </div><div class="line"><a name="l02990"></a><span class="lineno"> 2990</span>  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="l02991"></a><span class="lineno"> 2991</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l02992"></a><span class="lineno"> 2992</span>  acc0 = fma(b0, (half8)a0.s1, acc0);</div><div class="line"><a name="l02993"></a><span class="lineno"> 2993</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02994"></a><span class="lineno"> 2994</span>  acc1 = fma(b0, (half8)a1.s1, acc1);</div><div class="line"><a name="l02995"></a><span class="lineno"> 2995</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l02996"></a><span class="lineno"> 2996</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02997"></a><span class="lineno"> 2997</span>  acc2 = fma(b0, (half8)a2.s1, acc2);</div><div class="line"><a name="l02998"></a><span class="lineno"> 2998</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l02999"></a><span class="lineno"> 2999</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03000"></a><span class="lineno"> 3000</span>  acc3 = fma(b0, (half8)a3.s1, acc3);</div><div class="line"><a name="l03001"></a><span class="lineno"> 3001</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03002"></a><span class="lineno"> 3002</span> </div><div class="line"><a name="l03003"></a><span class="lineno"> 3003</span>  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="l03004"></a><span class="lineno"> 3004</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l03005"></a><span class="lineno"> 3005</span>  acc0 = fma(b0, (half8)a0.s2, acc0);</div><div class="line"><a name="l03006"></a><span class="lineno"> 3006</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03007"></a><span class="lineno"> 3007</span>  acc1 = fma(b0, (half8)a1.s2, acc1);</div><div class="line"><a name="l03008"></a><span class="lineno"> 3008</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03009"></a><span class="lineno"> 3009</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03010"></a><span class="lineno"> 3010</span>  acc2 = fma(b0, (half8)a2.s2, acc2);</div><div class="line"><a name="l03011"></a><span class="lineno"> 3011</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03012"></a><span class="lineno"> 3012</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03013"></a><span class="lineno"> 3013</span>  acc3 = fma(b0, (half8)a3.s2, acc3);</div><div class="line"><a name="l03014"></a><span class="lineno"> 3014</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03015"></a><span class="lineno"> 3015</span> </div><div class="line"><a name="l03016"></a><span class="lineno"> 3016</span>  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="l03017"></a><span class="lineno"> 3017</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l03018"></a><span class="lineno"> 3018</span>  acc0 = fma(b0, (half8)a0.s3, acc0);</div><div class="line"><a name="l03019"></a><span class="lineno"> 3019</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03020"></a><span class="lineno"> 3020</span>  acc1 = fma(b0, (half8)a1.s3, acc1);</div><div class="line"><a name="l03021"></a><span class="lineno"> 3021</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03022"></a><span class="lineno"> 3022</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03023"></a><span class="lineno"> 3023</span>  acc2 = fma(b0, (half8)a2.s3, acc2);</div><div class="line"><a name="l03024"></a><span class="lineno"> 3024</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03025"></a><span class="lineno"> 3025</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03026"></a><span class="lineno"> 3026</span>  acc3 = fma(b0, (half8)a3.s3, acc3);</div><div class="line"><a name="l03027"></a><span class="lineno"> 3027</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03028"></a><span class="lineno"> 3028</span> </div><div class="line"><a name="l03029"></a><span class="lineno"> 3029</span>  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="l03030"></a><span class="lineno"> 3030</span>  }</div><div class="line"><a name="l03031"></a><span class="lineno"> 3031</span> </div><div class="line"><a name="l03032"></a><span class="lineno"> 3032</span>  <span class="keywordflow">for</span>(; i < (int)COLS_A; ++i)</div><div class="line"><a name="l03033"></a><span class="lineno"> 3033</span>  {</div><div class="line"><a name="l03034"></a><span class="lineno"> 3034</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l03035"></a><span class="lineno"> 3035</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l03036"></a><span class="lineno"> 3036</span>  <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 + zin.s0));</div><div class="line"><a name="l03037"></a><span class="lineno"> 3037</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03038"></a><span class="lineno"> 3038</span>  <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 + zin.s1));</div><div class="line"><a name="l03039"></a><span class="lineno"> 3039</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03040"></a><span class="lineno"> 3040</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03041"></a><span class="lineno"> 3041</span>  <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 + zin.s2));</div><div class="line"><a name="l03042"></a><span class="lineno"> 3042</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03043"></a><span class="lineno"> 3043</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03044"></a><span class="lineno"> 3044</span>  <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 + zin.s3));</div><div class="line"><a name="l03045"></a><span class="lineno"> 3045</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03046"></a><span class="lineno"> 3046</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l03047"></a><span class="lineno"> 3047</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l03048"></a><span class="lineno"> 3048</span>  <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="l03049"></a><span class="lineno"> 3049</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03050"></a><span class="lineno"> 3050</span>  <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="l03051"></a><span class="lineno"> 3051</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03052"></a><span class="lineno"> 3052</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03053"></a><span class="lineno"> 3053</span>  <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="l03054"></a><span class="lineno"> 3054</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03055"></a><span class="lineno"> 3055</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03056"></a><span class="lineno"> 3056</span>  <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="l03057"></a><span class="lineno"> 3057</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03058"></a><span class="lineno"> 3058</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l03059"></a><span class="lineno"> 3059</span> </div><div class="line"><a name="l03060"></a><span class="lineno"> 3060</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l03061"></a><span class="lineno"> 3061</span>  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="l03062"></a><span class="lineno"> 3062</span> </div><div class="line"><a name="l03063"></a><span class="lineno"> 3063</span>  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="l03064"></a><span class="lineno"> 3064</span> </div><div class="line"><a name="l03065"></a><span class="lineno"> 3065</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l03066"></a><span class="lineno"> 3066</span>  acc0 = fma(b0, (half8)a0, acc0); <span class="comment">// b0 (half8)a0;</span></div><div class="line"><a name="l03067"></a><span class="lineno"> 3067</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03068"></a><span class="lineno"> 3068</span>  acc1 = fma(b0, (half8)a1, acc1); <span class="comment">// b0 * (half8)a1;</span></div><div class="line"><a name="l03069"></a><span class="lineno"> 3069</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03070"></a><span class="lineno"> 3070</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03071"></a><span class="lineno"> 3071</span>  acc2 = fma(b0, (half8)a2, acc2); <span class="comment">// b0 * (half8)a2;</span></div><div class="line"><a name="l03072"></a><span class="lineno"> 3072</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03073"></a><span class="lineno"> 3073</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03074"></a><span class="lineno"> 3074</span>  acc3 = fma(b0, (half8)a3, acc3); <span class="comment">// b0 * (half8)a3;</span></div><div class="line"><a name="l03075"></a><span class="lineno"> 3075</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03076"></a><span class="lineno"> 3076</span>  }</div><div class="line"><a name="l03077"></a><span class="lineno"> 3077</span> </div><div class="line"><a name="l03078"></a><span class="lineno"> 3078</span>  <span class="comment">// Multiply by the weight of matrix-matrix product and store the result</span></div><div class="line"><a name="l03079"></a><span class="lineno"> 3079</span> <span class="preprocessor">#if defined(ALPHA)</span></div><div class="line"><a name="l03080"></a><span class="lineno"> 3080</span>  acc0 = acc0 * (half8)ALPHA;</div><div class="line"><a name="l03081"></a><span class="lineno"> 3081</span> <span class="preprocessor">#endif // defined(ALPHA)</span></div><div class="line"><a name="l03082"></a><span class="lineno"> 3082</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l03083"></a><span class="lineno"> 3083</span>  acc1 = acc1 * (half8)ALPHA;</div><div class="line"><a name="l03084"></a><span class="lineno"> 3084</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1 && defined(ALPHA)</span></div><div class="line"><a name="l03085"></a><span class="lineno"> 3085</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l03086"></a><span class="lineno"> 3086</span>  acc2 = acc2 * (half8)ALPHA;</div><div class="line"><a name="l03087"></a><span class="lineno"> 3087</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2 && defined(ALPHA)</span></div><div class="line"><a name="l03088"></a><span class="lineno"> 3088</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l03089"></a><span class="lineno"> 3089</span>  acc3 = acc3 * (half8)ALPHA;</div><div class="line"><a name="l03090"></a><span class="lineno"> 3090</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3 && defined(ALPHA)</span></div><div class="line"><a name="l03091"></a><span class="lineno"> 3091</span> </div><div class="line"><a name="l03092"></a><span class="lineno"> 3092</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l03093"></a><span class="lineno"> 3093</span> </div><div class="line"><a name="l03094"></a><span class="lineno"> 3094</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l03095"></a><span class="lineno"> 3095</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l03096"></a><span class="lineno"> 3096</span> </div><div class="line"><a name="l03097"></a><span class="lineno"> 3097</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l03098"></a><span class="lineno"> 3098</span>  __global uchar dst_addr = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0);</div><div class="line"><a name="l03099"></a><span class="lineno"> 3099</span> </div><div class="line"><a name="l03100"></a><span class="lineno"> 3100</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l03101"></a><span class="lineno"> 3101</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l03102"></a><span class="lineno"> 3102</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l03103"></a><span class="lineno"> 3103</span>  <span class="comment">//</span></div><div class="line"><a name="l03104"></a><span class="lineno"> 3104</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l03105"></a><span class="lineno"> 3105</span>  <span class="comment">// \| plane0 \|</span></div><div class="line"><a name="l03106"></a><span class="lineno"> 3106</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l03107"></a><span class="lineno"> 3107</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l03108"></a><span class="lineno"> 3108</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l03109"></a><span class="lineno"> 3109</span>  <span class="comment">// \| cross_plane_pad \|</span></div><div class="line"><a name="l03110"></a><span class="lineno"> 3110</span>  <span class="comment">// \|***************\|</span></div><div class="line"><a name="l03111"></a><span class="lineno"> 3111</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l03112"></a><span class="lineno"> 3112</span>  <span class="comment">// \| plane1 \|</span></div><div class="line"><a name="l03113"></a><span class="lineno"> 3113</span>  <span class="comment">// \| \|</span></div><div class="line"><a name="l03114"></a><span class="lineno"> 3114</span>  <span class="comment">// \|__________________\|</span></div><div class="line"><a name="l03115"></a><span class="lineno"> 3115</span> </div><div class="line"><a name="l03116"></a><span class="lineno"> 3116</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l03117"></a><span class="lineno"> 3117</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l03118"></a><span class="lineno"> 3118</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l03119"></a><span class="lineno"> 3119</span> </div><div class="line"><a name="l03120"></a><span class="lineno"> 3120</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l03121"></a><span class="lineno"> 3121</span>  zout = (dst_cross_plane_pad dst_stride_y);</div><div class="line"><a name="l03122"></a><span class="lineno"> 3122</span> </div><div class="line"><a name="l03123"></a><span class="lineno"> 3123</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l03124"></a><span class="lineno"> 3124</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l03125"></a><span class="lineno"> 3125</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l03126"></a><span class="lineno"> 3126</span> </div><div class="line"><a name="l03127"></a><span class="lineno"> 3127</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l03128"></a><span class="lineno"> 3128</span>  vstore8(acc0, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 0 dst_stride_y + zout.s0));</div><div class="line"><a name="l03129"></a><span class="lineno"> 3129</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03130"></a><span class="lineno"> 3130</span>  vstore8(acc1, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 1 dst_stride_y + zout.s1));</div><div class="line"><a name="l03131"></a><span class="lineno"> 3131</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03132"></a><span class="lineno"> 3132</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03133"></a><span class="lineno"> 3133</span>  vstore8(acc2, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 2 dst_stride_y + zout.s2));</div><div class="line"><a name="l03134"></a><span class="lineno"> 3134</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03135"></a><span class="lineno"> 3135</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03136"></a><span class="lineno"> 3136</span>  vstore8(acc3, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )(dst_addr + 3 dst_stride_y + zout.s3));</div><div class="line"><a name="l03137"></a><span class="lineno"> 3137</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03138"></a><span class="lineno"> 3138</span> </div><div class="line"><a name="l03139"></a><span class="lineno"> 3139</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l03140"></a><span class="lineno"> 3140</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l03141"></a><span class="lineno"> 3141</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l03142"></a><span class="lineno"> 3142</span> </div><div class="line"><a name="l03143"></a><span class="lineno"> 3143</span>  <span class="comment">// Store the output block</span></div><div class="line"><a name="l03144"></a><span class="lineno"> 3144</span>  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="l03145"></a><span class="lineno"> 3145</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03146"></a><span class="lineno"> 3146</span>  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="l03147"></a><span class="lineno"> 3147</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l03148"></a><span class="lineno"> 3148</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03149"></a><span class="lineno"> 3149</span>  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="l03150"></a><span class="lineno"> 3150</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l03151"></a><span class="lineno"> 3151</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03152"></a><span class="lineno"> 3152</span>  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="l03153"></a><span class="lineno"> 3153</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l03154"></a><span class="lineno"> 3154</span> <span class="preprocessor">#endif // REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l03155"></a><span class="lineno"> 3155</span> }</div><div class="line"><a name="l03156"></a><span class="lineno"> 3156</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l03157"></a><span class="lineno"> 3157</span> </div><div class="line"><a name="l03158"></a><span class="lineno"> 3158</span> <span class="preprocessor">#endif // defined(COLS_A) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && (NUM_ELEMS_PROCESSED_PER_THREAD_Y)</span></div><div class="line"><a name="l03159"></a><span class="lineno"> 3159</span> </div><div class="line"><a name="l03160"></a><span class="lineno"> 3160</span> <span class="preprocessor">#if defined(BETA)</span></div><div class="line"><a name="l03161"></a><span class="lineno"> 3161</span> </div><div class="line"><a name="l03182"></a><span class="lineno"> 3182</span> __kernel <span class="keywordtype">void</span> gemm_ma_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l03183"></a><span class="lineno"> 3183</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l03184"></a><span class="lineno"> 3184</span> {</div><div class="line"><a name="l03185"></a><span class="lineno"> 3185</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l03186"></a><span class="lineno"> 3186</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l03187"></a><span class="lineno"> 3187</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l03188"></a><span class="lineno"> 3188</span> </div><div class="line"><a name="l03189"></a><span class="lineno"> 3189</span>  <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l03190"></a><span class="lineno"> 3190</span>  float4 alpha_ab = vload4(0, (__global <span class="keywordtype">float</span> )dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03191"></a><span class="lineno"> 3191</span> </div><div class="line"><a name="l03192"></a><span class="lineno"> 3192</span>  <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l03193"></a><span class="lineno"> 3193</span>  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_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03194"></a><span class="lineno"> 3194</span> </div><div class="line"><a name="l03195"></a><span class="lineno"> 3195</span>  <span class="comment">// Computes alpha * axb + beta * c</span></div><div class="line"><a name="l03196"></a><span class="lineno"> 3196</span>  float4 <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#af8a8a0625e7981212a0af48deb9d2a09">out</a> = alpha_ab + (float4)BETA * c;</div><div class="line"><a name="l03197"></a><span class="lineno"> 3197</span> </div><div class="line"><a name="l03198"></a><span class="lineno"> 3198</span>  <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l03199"></a><span class="lineno"> 3199</span>  vstore4(out, 0, (__global <span class="keywordtype">float</span> )dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03200"></a><span class="lineno"> 3200</span> }</div><div class="line"><a name="l03201"></a><span class="lineno"> 3201</span> </div><div class="line"><a name="l03202"></a><span class="lineno"> 3202</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l03203"></a><span class="lineno"> 3203</span> </div><div class="line"><a name="l03224"></a><span class="lineno"> 3224</span> __kernel <span class="keywordtype">void</span> gemm_ma_f16(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l03225"></a><span class="lineno"> 3225</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l03226"></a><span class="lineno"> 3226</span> {</div><div class="line"><a name="l03227"></a><span class="lineno"> 3227</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l03228"></a><span class="lineno"> 3228</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l03229"></a><span class="lineno"> 3229</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l03230"></a><span class="lineno"> 3230</span> </div><div class="line"><a name="l03231"></a><span class="lineno"> 3231</span>  <span class="comment">// Load values from A x B</span></div><div class="line"><a name="l03232"></a><span class="lineno"> 3232</span>  half8 alpha_ab = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03233"></a><span class="lineno"> 3233</span> </div><div class="line"><a name="l03234"></a><span class="lineno"> 3234</span>  <span class="comment">// Load values from Matrix C</span></div><div class="line"><a name="l03235"></a><span class="lineno"> 3235</span>  half8 c = vload8(0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03236"></a><span class="lineno"> 3236</span> </div><div class="line"><a name="l03237"></a><span class="lineno"> 3237</span>  <span class="comment">// Computes alpha axb + beta * c</span></div><div class="line"><a name="l03238"></a><span class="lineno"> 3238</span>  half8 out = alpha_ab + (half8)BETA * c;</div><div class="line"><a name="l03239"></a><span class="lineno"> 3239</span> </div><div class="line"><a name="l03240"></a><span class="lineno"> 3240</span>  <span class="comment">// Store final result in axb matrix</span></div><div class="line"><a name="l03241"></a><span class="lineno"> 3241</span>  vstore8(out, 0, (__global <a class="code" href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">half</a> )dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03242"></a><span class="lineno"> 3242</span> }</div><div class="line"><a name="l03243"></a><span class="lineno"> 3243</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_FP16_ENABLED)</span></div><div class="line"><a name="l03244"></a><span class="lineno"> 3244</span> <span class="preprocessor">#endif // defined(BETA)</span></div><div class="line"><a name="l03245"></a><span class="lineno"> 3245</span> </div><div class="line"><a name="l03246"></a><span class="lineno"> 3246</span> <span class="preprocessor">#if defined(WIDTH_VECTOR_A)</span></div><div class="line"><a name="l03247"></a><span class="lineno"> 3247</span> </div><div class="line"><a name="l03274"></a><span class="lineno"> 3274</span> __kernel <span class="keywordtype">void</span> gemm_lc_vm_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l03275"></a><span class="lineno"> 3275</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src1),</div><div class="line"><a name="l03276"></a><span class="lineno"> 3276</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l03277"></a><span class="lineno"> 3277</span> {</div><div class="line"><a name="l03278"></a><span class="lineno"> 3278</span>  <span class="keywordtype">int</span> idx = get_global_id(0) 4;</div><div class="line"><a name="l03279"></a><span class="lineno"> 3279</span>  <span class="keywordtype">int</span> idy = get_global_id(1);</div><div class="line"><a name="l03280"></a><span class="lineno"> 3280</span> </div><div class="line"><a name="l03281"></a><span class="lineno"> 3281</span>  <span class="comment">// Compute the address for the vector A and matrix B</span></div><div class="line"><a name="l03282"></a><span class="lineno"> 3282</span>  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="l03283"></a><span class="lineno"> 3283</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l03284"></a><span class="lineno"> 3284</span> </div><div class="line"><a name="l03285"></a><span class="lineno"> 3285</span>  <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="l03286"></a><span class="lineno"> 3286</span> </div><div class="line"><a name="l03287"></a><span class="lineno"> 3287</span>  float4 acc = 0.0f;</div><div class="line"><a name="l03288"></a><span class="lineno"> 3288</span> </div><div class="line"><a name="l03289"></a><span class="lineno"> 3289</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (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="l03290"></a><span class="lineno"> 3290</span>  {</div><div class="line"><a name="l03291"></a><span class="lineno"> 3291</span>  float2 a0 = vload2(0, (__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0));</div><div class="line"><a name="l03292"></a><span class="lineno"> 3292</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l03293"></a><span class="lineno"> 3293</span>  float4 b1 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1 + src1_stride_y));</div><div class="line"><a name="l03294"></a><span class="lineno"> 3294</span> </div><div class="line"><a name="l03295"></a><span class="lineno"> 3295</span>  acc += b0 (float4)a0.s0;</div><div class="line"><a name="l03296"></a><span class="lineno"> 3296</span>  acc += b1 * (float4)a0.s1;</div><div class="line"><a name="l03297"></a><span class="lineno"> 3297</span>  }</div><div class="line"><a name="l03298"></a><span class="lineno"> 3298</span> </div><div class="line"><a name="l03299"></a><span class="lineno"> 3299</span>  <span class="keywordflow">for</span>(; src_addr.s0 < 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="l03300"></a><span class="lineno"> 3300</span>  {</div><div class="line"><a name="l03301"></a><span class="lineno"> 3301</span>  <span class="keywordtype">float</span> a0 = ((__global <span class="keywordtype">float</span> )(src0_ptr + src_addr.s0));</div><div class="line"><a name="l03302"></a><span class="lineno"> 3302</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> )(src1_ptr + src_addr.s1));</div><div class="line"><a name="l03303"></a><span class="lineno"> 3303</span> </div><div class="line"><a name="l03304"></a><span class="lineno"> 3304</span>  acc += b0 (float4)a0;</div><div class="line"><a name="l03305"></a><span class="lineno"> 3305</span>  }</div><div class="line"><a name="l03306"></a><span class="lineno"> 3306</span> </div><div class="line"><a name="l03307"></a><span class="lineno"> 3307</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l03308"></a><span class="lineno"> 3308</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l03309"></a><span class="lineno"> 3309</span> </div><div class="line"><a name="l03310"></a><span class="lineno"> 3310</span>  vstore4(acc, 0, (__global <span class="keywordtype">float</span> )(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l03311"></a><span class="lineno"> 3311</span> }</div><div class="line"><a name="l03312"></a><span class="lineno"> 3312</span> <span class="preprocessor">#endif // defined(WIDTH_VECTOR_A)</span></div><div class="line"><a name="l03313"></a><span class="lineno"> 3313</span> </div><div class="line"><a name="l03330"></a><span class="lineno"> 3330</span> <span class="preprocessor">#if defined(DATA_TYPE) && defined(VECTOR_SIZE)</span></div><div class="line"><a name="l03331"></a><span class="lineno"> 3331</span> __kernel <span class="keywordtype">void</span> gemm_accumulate_biases(</div><div class="line"><a name="l03332"></a><span class="lineno"> 3332</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l03333"></a><span class="lineno"> 3333</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l03334"></a><span class="lineno"> 3334</span> {</div><div class="line"><a name="l03335"></a><span class="lineno"> 3335</span>  <a class="code" href="struct_image.xhtml">Image</a> accum = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l03336"></a><span class="lineno"> 3336</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l03337"></a><span class="lineno"> 3337</span> </div><div class="line"><a name="l03338"></a><span class="lineno"> 3338</span>  <span class="comment">// Vector size, i.e. number of vector elements.</span></div><div class="line"><a name="l03339"></a><span class="lineno"> 3339</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_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="l03340"></a><span class="lineno"> 3340</span>  accum_value = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(<a class="code" href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a>)(0, (__global DATA_TYPE )accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03341"></a><span class="lineno"> 3341</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_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="l03342"></a><span class="lineno"> 3342</span>  biases_value = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(<a class="code" href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a>)(0, (__global DATA_TYPE )biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03343"></a><span class="lineno"> 3343</span>  accum_value = biases_value + accum_value;</div><div class="line"><a name="l03344"></a><span class="lineno"> 3344</span>  <span class="comment">// Store result in the accumulate buffer</span></div><div class="line"><a name="l03345"></a><span class="lineno"> 3345</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(<a class="code" href="softmax__layer_8cl.xhtml#a7c78836761fa3b5b124efea237dac70f">VECTOR_SIZE</a>)</div><div class="line"><a name="l03346"></a><span class="lineno"> 3346</span>  (accum_value, 0, (__global DATA_TYPE )accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l03347"></a><span class="lineno"> 3347</span> }</div><div class="line"><a name="l03348"></a><span class="lineno"> 3348</span> <span class="preprocessor">#endif // defined(DATA_TYPE) && 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="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00140">helpers.h:140</a></div></div>
				95	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a009469e4d9b8fce3b6d5e97d2077827d"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a></div><div class="ttdeci">__global uchar * offset(const Image *img, int x, int y)</div><div class="ttdoc">Get the pointer position of a Image. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00309">helpers.h:309</a></div></div>
				96	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_aebe814363556c244be043b13e7969197"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_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="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00110">helpers.h:110</a></div></div>
Kaizen	8938bd3	2017-09-28 14:38:23 +0100	[diff] [blame]	97	<div class="ttc" id="convolution3x3_8cl_xhtml_afb8c72ce35c4a1f4a2588d6573e54aa1"><div class="ttname"><a href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a></div><div class="ttdeci">#define DATA_TYPE</div><div class="ttdef"><b>Definition:</b> <a href="convolution3x3_8cl_source.xhtml#l00027">convolution3x3.cl:27</a></div></div>
Anthony Barbier	06ea048	2018-02-22 15:45:35 +0000	[diff] [blame]	98	<div class="ttc" id="namespacearm__compute_xhtml_a73e2825fd61d349c5ca2f5313e3c8ea1"><div class="ttname"><a href="namespacearm__compute.xhtml#a73e2825fd61d349c5ca2f5313e3c8ea1">arm_compute::half</a></div><div class="ttdeci">half_float::half half</div><div class="ttdoc">16-bit floating point type </div><div class="ttdef"><b>Definition:</b> <a href="arm__compute_2core_2_types_8h_source.xhtml#l00044">Types.h:44</a></div></div>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	99	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a22f42fcf2077d951271df83b55c1a71a"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a></div><div class="ttdeci">#define IMAGE_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00074">helpers.h:74</a></div></div>
				100	<div class="ttc" id="struct_tensor3_d_xhtml"><div class="ttname"><a href="struct_tensor3_d.xhtml">Tensor3D</a></div><div class="ttdoc">Structure to hold 3D tensor information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00157">helpers.h:157</a></div></div>
				101	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a527bfdf5eeb306f1cf01c4a8e29f38e0"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_VECTOR_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00104">helpers.h:104</a></div></div>
Jenkins	52ba29e	2018-08-29 15:32:11 +0000	[diff] [blame]	102	<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#l00060">softmax_layer.cl:60</a></div></div>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	103	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a40a6eb9f2a7712f08d6bb8ff6c9e6ca7"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a></div><div class="ttdeci">#define VECTOR_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00068">helpers.h:68</a></div></div>
				104	<div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_af8a8a0625e7981212a0af48deb9d2a09"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#af8a8a0625e7981212a0af48deb9d2a09">arm_compute::test::validation::out</a></div><div class="ttdeci">CLTensor out</div><div class="ttdef"><b>Definition:</b> <a href="_c_l_2_winograd_8cpp_source.xhtml#l00266">Winograd.cpp:266</a></div></div>
				105	<div class="ttc" id="struct_vector_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">Vector::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00142">helpers.h:142</a></div></div>
Jenkins	52ba29e	2018-08-29 15:32:11 +0000	[diff] [blame]	106	<div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_adbf67dcee294e673cf796f1ed8aeb6a4"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">arm_compute::test::validation::dst</a></div><div class="ttdeci">CLTensor dst</div><div class="ttdef"><b>Definition:</b> <a href="_c_l_2_arithmetic_division_8cpp_source.xhtml#l00100">ArithmeticDivision.cpp:100</a></div></div>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	107	<div class="ttc" id="struct_image_xhtml"><div class="ttname"><a href="struct_image.xhtml">Image</a></div><div class="ttdoc">Structure to hold Image information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00148">helpers.h:148</a></div></div>
				108	<div class="ttc" id="struct_image_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">Image::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00150">helpers.h:150</a></div></div>
				109	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a31c8c760f08fb1a331b16b7c204321dc"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_TENSOR3D_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00125">helpers.h:125</a></div></div>
				110	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml">helpers.h</a></div></div>
				111	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_acb282042d1edeeaa3cc979a206f78b54"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a></div><div class="ttdeci">#define VSTORE(size)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00051">helpers.h:51</a></div></div>
Jenkins	52ba29e	2018-08-29 15:32:11 +0000	[diff] [blame]	112	<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#l00092">GEMM.cpp:92</a></div></div>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	113	<div class="ttc" id="struct_tensor3_d_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">Tensor3D::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00159">helpers.h:159</a></div></div>
				114	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a287e2fc366c312b468382c95bb90f91f"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a></div><div class="ttdeci">#define VLOAD(size)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00048">helpers.h:48</a></div></div>
				115	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a6b83038822d1ae7ab619b684ed3b7fc0"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a></div><div class="ttdeci">#define TENSOR3D_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00082">helpers.h:82</a></div></div>
Jenkins	52ba29e	2018-08-29 15:32:11 +0000	[diff] [blame]	116	<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 & src</div><div class="ttdef"><b>Definition:</b> <a href="validation_2_c_l_2_convolution_8cpp_source.xhtml#l00147">Convolution.cpp:147</a></div></div>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	117	<div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a36f754c05b6fddf6df0d8d0a74f8159f"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_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="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00054">helpers.h:54</a></div></div>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	118	</div><!-- fragment --></div><!-- contents -->
				119	</div><!-- doc-content -->
				120	<!-- start footer part -->
				121	<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
				122	<ul>
Anthony Barbier	8140e1e	2017-12-14 23:48:46 +0000	[diff] [blame]	123	<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>
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	124	<li class="footer">Generated on Thu Nov 22 2018 11:57:38 for Compute Library by
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	125	<a href="http://www.doxygen.org/index.html">
Jenkins	b9abeae	2018-11-22 11:58:08 +0000	[diff] [blame]	126	<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.13 </li>
Anthony Barbier	871448e	2017-03-24 14:54:29 +0000	[diff] [blame]	127	</ul>
				128	</div>
				129	</body>
				130	</html>