arm_compute v17.06
diff --git a/documentation/gemm_8cl_source.xhtml b/documentation/gemm_8cl_source.xhtml
index d0cf556..a609835 100644
--- a/documentation/gemm_8cl_source.xhtml
+++ b/documentation/gemm_8cl_source.xhtml
@@ -6,7 +6,7 @@
<meta http-equiv="X-UA-Compatible" content="IE=9"/>
<meta name="generator" content="Doxygen 1.8.11"/>
<meta name="robots" content="NOINDEX, NOFOLLOW" /> <!-- Prevent indexing by search engines -->
-<title>ARM Compute Library: src/core/CL/cl_kernels/gemm.cl Source File</title>
+<title>Compute Library: src/core/CL/cl_kernels/gemm.cl Source File</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="jquery.js"></script>
<script type="text/javascript" src="dynsections.js"></script>
@@ -39,8 +39,8 @@
<tbody>
<tr style="height: 56px;">
<td style="padding-left: 0.5em;">
- <div id="projectname">ARM Compute Library
-  <span id="projectnumber">17.05</span>
+ <div id="projectname">Compute Library
+  <span id="projectnumber">v17.06</span>
</div>
</td>
</tr>
@@ -55,6 +55,7 @@
<div id="navrow1" class="tabs">
<ul class="tablist">
<li><a href="index.xhtml"><span>Main Page</span></a></li>
+ <li><a href="pages.xhtml"><span>Related Pages</span></a></li>
<li><a href="namespaces.xhtml"><span>Namespaces</span></a></li>
<li><a href="annotated.xhtml"><span>Data Structures</span></a></li>
<li class="current"><a href="files.xhtml"><span>Files</span></a></li>
@@ -116,7 +117,7 @@
<div class="title">gemm.cl</div> </div>
</div><!--header-->
<div class="contents">
-<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 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="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="l00041"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a21d7061f36aec78824e368f2ef1cafc1"> 41</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a21d7061f36aec78824e368f2ef1cafc1">gemm_transpose1x4_f32</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> {</div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> </div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> </div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> </div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  vstore4(b0, 0, (__global <span class="keywordtype">float</span> *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> }</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> </div><div class="line"><a name="l00073"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a11d87ecaf1388b4f57404e062265b374"> 73</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a11d87ecaf1388b4f57404e062265b374">gemm_transpose1x8_f16</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span> {</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> </div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> </div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</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>  half8 b0 = vload8(0, (__global half *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span> </div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  vstore8(b0, 0, (__global half *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span> }</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> </div><div class="line"><a name="l00105"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a675879eb72964ddbd5f71155da3b0cc3"> 105</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a675879eb72964ddbd5f71155da3b0cc3">gemm_transpose1x16_u8</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span> {</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span> </div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span> </div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> </div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  uchar16 b0 = vload16(0, (__global uchar *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> </div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  vstore16(b0, 0, (__global uchar *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span> }</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> </div><div class="line"><a name="l00137"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a560b2d50aa886edc0f5daf4fe729717f"> 137</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a560b2d50aa886edc0f5daf4fe729717f">gemm_interleave4x4_32bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span> {</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span> </div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  float4 a0 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  float4 a1 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  float4 a2 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  float4 a3 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span> </div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  float4 val0 = (float4)(a0.s0, a1.s0, a2.s0, a3.s0);</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span> </div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  val0 = (float4)(a0.s1, a1.s1, a2.s1, a3.s1);</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 4);</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span> </div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  val0 = (float4)(a0.s2, a1.s2, a2.s2, a3.s2);</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 8);</div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span> </div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  val0 = (float4)(a0.s3, a1.s3, a2.s3, a3.s3);</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 12);</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span> }</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span> </div><div class="line"><a name="l00178"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a2f32d740e780059f68da8aa589ed0a5b"> 178</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a2f32d740e780059f68da8aa589ed0a5b">gemm_interleave4x4_16bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span> {</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span> </div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  half8 a0 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  half8 a1 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  half8 a2 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  half8 a3 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</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>  half8 val0 = (half8)((half4)(a0.s0, a1.s0, a2.s0, a3.s0), (half4)(a0.s1, a1.s1, a2.s1, a3.s1));</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span> </div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  val0 = (half8)((half4)(a0.s2, a1.s2, a2.s2, a3.s2), (half4)(a0.s3, a1.s3, a2.s3, a3.s3));</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 8);</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span> </div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  val0 = (half8)((half4)(a0.s4, a1.s4, a2.s4, a3.s4), (half4)(a0.s5, a1.s5, a2.s5, a3.s5));</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 16);</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span> </div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  val0 = (half8)((half4)(a0.s6, a1.s6, a2.s6, a3.s6), (half4)(a0.s7, a1.s7, a2.s7, a3.s7));</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 24);</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span> }</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span> </div><div class="line"><a name="l00219"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#aa66a7b6b0420e54ec173743e6c5bfc45"> 219</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#aa66a7b6b0420e54ec173743e6c5bfc45">gemm_interleave4x4_8bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span> {</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span> </div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  uchar16 a0 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  uchar16 a1 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  uchar16 a2 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  uchar16 a3 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span> </div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  uchar16 val0 = (uchar16)((uchar4)(a0.s0, a1.s0, a2.s0, a3.s0), (uchar4)(a0.s1, a1.s1, a2.s1, a3.s1),</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  (uchar4)(a0.s2, a1.s2, a2.s2, a3.s2), (uchar4)(a0.s3, a1.s3, a2.s3, a3.s3));</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span> </div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  val0 = (uchar16)((uchar4)(a0.s4, a1.s4, a2.s4, a3.s4), (uchar4)(a0.s5, a1.s5, a2.s5, a3.s5),</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  (uchar4)(a0.s6, a1.s6, a2.s6, a3.s6), (uchar4)(a0.s7, a1.s7, a2.s7, a3.s7));</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 16);</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span> </div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  val0 = (uchar16)((uchar4)(a0.s8, a1.s8, a2.s8, a3.s8), (uchar4)(a0.s9, a1.s9, a2.s9, a3.s9),</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  (uchar4)(a0.sA, a1.sA, a2.sA, a3.sA), (uchar4)(a0.sB, a1.sB, a2.sB, a3.sB));</div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 32);</div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span> </div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  val0 = (uchar16)((uchar4)(a0.sC, a1.sC, a2.sC, a3.sC), (uchar4)(a0.sD, a1.sD, a2.sD, a3.sD),</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  (uchar4)(a0.sE, a1.sE, a2.sE, a3.sE), (uchar4)(a0.sF, a1.sF, a2.sF, a3.sF));</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 48);</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span> }</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span> </div><div class="line"><a name="l00262"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a7c619984bc9c31f5f6c19ecc88b56581"> 262</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a7c619984bc9c31f5f6c19ecc88b56581">gemm_accumulate_biases_f32</a>(</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  <a class="code" href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span> {</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <a class="code" href="struct_image.xhtml">Image</a> accum = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span> </div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  float4 accum_value = vload4(0, (__global <span class="keywordtype">float</span> *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  float4 biases_value = vload4(0, (__global <span class="keywordtype">float</span> *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  accum_value = biases_value + accum_value;</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">// Store result in the accummulate buffer</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  vstore4(accum_value, 0, (__global <span class="keywordtype">float</span> *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span> }</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span> </div><div class="line"><a name="l00290"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a76295c333dfa3c9bb20975579dce87e8"> 290</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a76295c333dfa3c9bb20975579dce87e8">gemm_accumulate_biases_f16</a>(</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <a class="code" href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span> {</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <a class="code" href="struct_image.xhtml">Image</a> accum = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span> </div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  half8 accum_value = vload8(0, (__global half *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  half8 biases_value = vload8(0, (__global half *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  accum_value = biases_value + accum_value;</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">// Store result in the accummulate buffer</span></div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  vstore8(accum_value, 0, (__global half *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span> }</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> </div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> <span class="preprocessor">#if(defined WIDTH_MATRIX_B)</span></div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> </div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span> __kernel <span class="keywordtype">void</span> gemm_mm_u8(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <span class="keywordtype">int</span> a_offset,</div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  <span class="keywordtype">int</span> b_offset,</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  <span class="keywordtype">int</span> c_offset,</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  <span class="keywordtype">int</span> c_mult_int,</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  <span class="keywordtype">int</span> shift)</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span> {</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span> </div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  (src1_stride_y));</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span> </div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span> </div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span> </div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  int16 c00 = 0.0f;</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  int16 c10 = 0.0f;</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  int16 c20 = 0.0f;</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  int16 c30 = 0.0f;</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span> </div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 32))</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  {</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  int8 a0 = (int8)a_offset + convert_int8(vload8(0, ((__global uchar *)src0_ptr) + src_addr.s0));</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));</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>  c00 += (int16)a0.s0 * b0;</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  c10 += (int16)a0.s1 * b0;</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  c20 += (int16)a0.s2 * b0;</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  c30 += (int16)a0.s3 * b0;</div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span> </div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  int16 b1 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1 + 16));</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span> </div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  c00 += (int16)a0.s4 * b1;</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  c10 += (int16)a0.s5 * b1;</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  c20 += (int16)a0.s6 * b1;</div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  c30 += (int16)a0.s7 * b1;</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  }</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span> </div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 16))</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  {</div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  int4 a0 = (int4)a_offset + convert_int4(vload4(0, ((__global uchar *)src0_ptr) + src_addr.s0));</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span> </div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  c00 += (int16)a0.s0 * b0;</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  c10 += (int16)a0.s1 * b0;</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  c20 += (int16)a0.s2 * b0;</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  c30 += (int16)a0.s3 * b0;</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  }</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span> </div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span> </div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  c00 = (((int16)c_offset + c00) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  c10 = (((int16)c_offset + c10) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  c20 = (((int16)c_offset + c20) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  c30 = (((int16)c_offset + c30) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span> </div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  <span class="comment">/* Store 4x16 block */</span></div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  vstore16(convert_uchar16_sat(c00), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  vstore16(convert_uchar16_sat(c10), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  vstore16(convert_uchar16_sat(c20), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span>  vstore16(convert_uchar16_sat(c30), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span> }</div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span> </div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span> <span class="preprocessor">#if(defined WIDTH_MATRIX_B && defined ALPHA)</span></div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span> __kernel <span class="keywordtype">void</span> gemm_mm_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> {</div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span> </div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  (src1_stride_y));</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span> </div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</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">/* Divide by 4 in order to get the src_addr in unit of float */</span></div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  src_addr = src_addr >> 2;</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span> </div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span> </div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  float4 c00 = 0.0f;</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  float4 c10 = 0.0f;</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  float4 c20 = 0.0f;</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  float4 c30 = 0.0f;</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>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 8))</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  {</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  float4 a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  float4 b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1);</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>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span> </div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0 + 4);</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1 + 4);</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span> </div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  }</div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span> </div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  <span class="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 4))</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  {</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  float4 a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  float4 b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1);</div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span> </div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  }</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>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> </div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  c00 = c00 * (float4)ALPHA;</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  c10 = c10 * (float4)ALPHA;</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  c20 = c20 * (float4)ALPHA;</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  c30 = c30 * (float4)ALPHA;</div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span> </div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  <span class="comment">/* Store 4x4 block */</span></div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  vstore4(c00, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  vstore4(c10, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  vstore4(c20, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span>  vstore4(c30, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span> }</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span> </div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span> __kernel <span class="keywordtype">void</span> gemm_mm_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span> {</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span> </div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span>  (src1_stride_y));</div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span> </div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</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>  <span class="comment">/* Divide by 2 in order to get the src_addr in unit of half */</span></div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  src_addr = src_addr >> 1;</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>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span> </div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  half8 c00 = 0.0f;</div><div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  half8 c10 = 0.0f;</div><div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  half8 c20 = 0.0f;</div><div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  half8 c30 = 0.0f;</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span> </div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 16))</div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  {</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span>  half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);</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>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00568"></a><span class="lineno"> 568</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span>  c30 += (half8)a0.s3 * b0;</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, ((__global half *)src0_ptr) + src_addr.s0 + 4);</div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1 + 8);</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>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00577"></a><span class="lineno"> 577</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span>  }</div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span> </div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span>  <span class="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 8))</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>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span> </div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  }</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>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span> </div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l00599"></a><span class="lineno"> 599</span>  c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l00600"></a><span class="lineno"> 600</span>  c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  c30 = c30 * (half8)ALPHA;</div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span> </div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  <span class="comment">/* Store 4x8 block */</span></div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  vstore8(c00, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  vstore8(c10, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00606"></a><span class="lineno"> 606</span>  vstore8(c20, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  vstore8(c30, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</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> </div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span> <span class="preprocessor">#if(defined WIDTH_VECTOR_A)</span></div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span> __kernel <span class="keywordtype">void</span> gemm_vm_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span> {</div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * 4;</div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span> </div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span>  <span class="comment">/* Compute the address for the vector A and matrix B */</span></div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span> </div><div class="line"><a name="l00646"></a><span class="lineno"> 646</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="l00647"></a><span class="lineno"> 647</span> </div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  float4 acc = 0.0f;</div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span> </div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 2 * <span class="keyword">sizeof</span>(float)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>), 2 * src1_stride_y))</div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span>  {</div><div class="line"><a name="l00652"></a><span class="lineno"> 652</span>  float2 a0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00653"></a><span class="lineno"> 653</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l00654"></a><span class="lineno"> 654</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="l00655"></a><span class="lineno"> 655</span> </div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span>  acc += b0 * (float4)a0.s0;</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span>  acc += b1 * (float4)a0.s1;</div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>  }</div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span> </div><div class="line"><a name="l00660"></a><span class="lineno"> 660</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="l00661"></a><span class="lineno"> 661</span>  {</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  <span class="keywordtype">float</span> a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span> </div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span>  acc += b0 * (float4)a0;</div><div class="line"><a name="l00666"></a><span class="lineno"> 666</span>  }</div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span> </div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00669"></a><span class="lineno"> 669</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00670"></a><span class="lineno"> 670</span> </div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span>  <span class="comment">/* Multiply by the weight of vector-matrix product */</span></div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  acc = acc * (float4)ALPHA;</div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span> </div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span>  vstore4(acc, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</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> </div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span> __kernel <span class="keywordtype">void</span> gemm_vm_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span> {</div><div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * 8;</div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span> </div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>  <span class="comment">/* Compute the address for the vector A and matrix B */</span></div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(half);</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span> </div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>  <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * <span class="keyword">sizeof</span>(half));</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span> </div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span>  half8 acc = 0.0f;</div><div class="line"><a name="l00715"></a><span class="lineno"> 715</span> </div><div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 4 * <span class="keyword">sizeof</span>(half)); src_addr += (int2)(4 * <span class="keyword">sizeof</span>(half), 4 * src1_stride_y))</div><div class="line"><a name="l00717"></a><span class="lineno"> 717</span>  {</div><div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00719"></a><span class="lineno"> 719</span>  half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));</div><div class="line"><a name="l00720"></a><span class="lineno"> 720</span>  half8 b1 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));</div><div class="line"><a name="l00721"></a><span class="lineno"> 721</span>  half8 b2 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y));</div><div class="line"><a name="l00722"></a><span class="lineno"> 722</span>  half8 b3 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y));</div><div class="line"><a name="l00723"></a><span class="lineno"> 723</span> </div><div class="line"><a name="l00724"></a><span class="lineno"> 724</span>  acc += b0 * (half8)a0.s0;</div><div class="line"><a name="l00725"></a><span class="lineno"> 725</span>  acc += b1 * (half8)a0.s1;</div><div class="line"><a name="l00726"></a><span class="lineno"> 726</span>  acc += b2 * (half8)a0.s2;</div><div class="line"><a name="l00727"></a><span class="lineno"> 727</span>  acc += b3 * (half8)a0.s3;</div><div class="line"><a name="l00728"></a><span class="lineno"> 728</span>  }</div><div class="line"><a name="l00729"></a><span class="lineno"> 729</span> </div><div class="line"><a name="l00730"></a><span class="lineno"> 730</span>  <span class="keywordflow">for</span>(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(half), src1_stride_y))</div><div class="line"><a name="l00731"></a><span class="lineno"> 731</span>  {</div><div class="line"><a name="l00732"></a><span class="lineno"> 732</span>  half a0 = *((__global half *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00733"></a><span class="lineno"> 733</span>  half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l00734"></a><span class="lineno"> 734</span> </div><div class="line"><a name="l00735"></a><span class="lineno"> 735</span>  acc += b0 * (half8)a0;</div><div class="line"><a name="l00736"></a><span class="lineno"> 736</span>  }</div><div class="line"><a name="l00737"></a><span class="lineno"> 737</span> </div><div class="line"><a name="l00738"></a><span class="lineno"> 738</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00739"></a><span class="lineno"> 739</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00740"></a><span class="lineno"> 740</span> </div><div class="line"><a name="l00741"></a><span class="lineno"> 741</span>  <span class="comment">/* Multiply by the weight of vector-matrix product */</span></div><div class="line"><a name="l00742"></a><span class="lineno"> 742</span>  acc = acc * (half8)ALPHA;</div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span> </div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span>  vstore8(acc, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span> }</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined WIDTH_VECTOR_A) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined WIDTH_MATRIX_B && defined ALPHA) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span> </div><div class="line"><a name="l00766"></a><span class="lineno"> 766</span> <span class="preprocessor">#if(defined BETA)</span></div><div class="line"><a name="l00767"></a><span class="lineno"> 767</span> __kernel <span class="keywordtype">void</span> gemm_ma_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span> {</div><div class="line"><a name="l00770"></a><span class="lineno"> 770</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span> </div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span>  <span class="comment">/* Load values from A x B */</span></div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span>  float4 alpha_ab = vload4(0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00776"></a><span class="lineno"> 776</span> </div><div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  <span class="comment">/* Load values from Matrix C */</span></div><div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  float4 c = vload4(0, (__global <span class="keywordtype">float</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00779"></a><span class="lineno"> 779</span> </div><div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  <span class="comment">/* Computes alpha * axb + beta * c */</span></div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span>  float4 out = alpha_ab + (float4)BETA * c;</div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span> </div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  <span class="comment">/* Store final result in axb matrix */</span></div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span>  vstore4(out, 0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span> }</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span> </div><div class="line"><a name="l00802"></a><span class="lineno"> 802</span> __kernel <span class="keywordtype">void</span> gemm_ma_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00803"></a><span class="lineno"> 803</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</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>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00806"></a><span class="lineno"> 806</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00807"></a><span class="lineno"> 807</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00808"></a><span class="lineno"> 808</span> </div><div class="line"><a name="l00809"></a><span class="lineno"> 809</span>  <span class="comment">/* Load values from A x B */</span></div><div class="line"><a name="l00810"></a><span class="lineno"> 810</span>  half8 alpha_ab = vload8(0, (__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00811"></a><span class="lineno"> 811</span> </div><div class="line"><a name="l00812"></a><span class="lineno"> 812</span>  <span class="comment">/* Load values from Matrix C */</span></div><div class="line"><a name="l00813"></a><span class="lineno"> 813</span>  half8 c = vload8(0, (__global half *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00814"></a><span class="lineno"> 814</span> </div><div class="line"><a name="l00815"></a><span class="lineno"> 815</span>  <span class="comment">/* Computes alpha * axb + beta * c */</span></div><div class="line"><a name="l00816"></a><span class="lineno"> 816</span>  half8 out = alpha_ab + (half8)BETA * c;</div><div class="line"><a name="l00817"></a><span class="lineno"> 817</span> </div><div class="line"><a name="l00818"></a><span class="lineno"> 818</span>  <span class="comment">/* Store final result in axb matrix */</span></div><div class="line"><a name="l00819"></a><span class="lineno"> 819</span>  vstore8(out, 0, (__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00820"></a><span class="lineno"> 820</span> }</div><div class="line"><a name="l00821"></a><span class="lineno"> 821</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined BETA) */</span><span class="preprocessor"></span></div><div class="ttc" id="struct_vector_xhtml"><div class="ttname"><a href="struct_vector.xhtml">Vector</a></div><div class="ttdoc">Structure to hold Vector information. </div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00087">helpers.h:87</a></div></div>
+<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 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="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="l00041"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a21d7061f36aec78824e368f2ef1cafc1"> 41</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a21d7061f36aec78824e368f2ef1cafc1">gemm_transpose1x4_f32</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> {</div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> </div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> </div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> </div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  vstore4(b0, 0, (__global <span class="keywordtype">float</span> *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> }</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> </div><div class="line"><a name="l00073"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a11d87ecaf1388b4f57404e062265b374"> 73</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a11d87ecaf1388b4f57404e062265b374">gemm_transpose1x8_f16</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span> {</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> </div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> </div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</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>  half8 b0 = vload8(0, (__global half *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span> </div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  vstore8(b0, 0, (__global half *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span> }</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> </div><div class="line"><a name="l00105"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a675879eb72964ddbd5f71155da3b0cc3"> 105</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a675879eb72964ddbd5f71155da3b0cc3">gemm_transpose1x16_u8</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span> {</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  uint x = get_global_id(0);</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span>  uint y = get_global_id(1);</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span> </div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  <span class="comment">/* Compute address for Matrix B - source */</span></div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span> </div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  <span class="comment">/* Compute address for Matrix B transposed - destination. X and Y are swapped */</span></div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  uint dst_addr_in_bytes = y * 16 + ((x * dst_stride_y + dst_offset_first_element_in_bytes));</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> </div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  uchar16 b0 = vload16(0, (__global uchar *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> </div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  vstore16(b0, 0, (__global uchar *)(dst_ptr + dst_addr_in_bytes));</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span> }</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> </div><div class="line"><a name="l00137"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a560b2d50aa886edc0f5daf4fe729717f"> 137</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a560b2d50aa886edc0f5daf4fe729717f">gemm_interleave4x4_32bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span> {</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span> </div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  float4 a0 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  float4 a1 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  float4 a2 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  float4 a3 = vload4(0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span> </div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  float4 val0 = (float4)(a0.s0, a1.s0, a2.s0, a3.s0);</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span> </div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  val0 = (float4)(a0.s1, a1.s1, a2.s1, a3.s1);</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 4);</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span> </div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  val0 = (float4)(a0.s2, a1.s2, a2.s2, a3.s2);</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 8);</div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span> </div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  val0 = (float4)(a0.s3, a1.s3, a2.s3, a3.s3);</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  vstore4(val0, 0, ((__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 12);</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span> }</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span> </div><div class="line"><a name="l00178"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a2f32d740e780059f68da8aa589ed0a5b"> 178</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a2f32d740e780059f68da8aa589ed0a5b">gemm_interleave4x4_16bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span> {</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span> </div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  half8 a0 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  half8 a1 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  half8 a2 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  half8 a3 = vload8(0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</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>  half8 val0 = (half8)((half4)(a0.s0, a1.s0, a2.s0, a3.s0), (half4)(a0.s1, a1.s1, a2.s1, a3.s1));</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span> </div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  val0 = (half8)((half4)(a0.s2, a1.s2, a2.s2, a3.s2), (half4)(a0.s3, a1.s3, a2.s3, a3.s3));</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 8);</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span> </div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  val0 = (half8)((half4)(a0.s4, a1.s4, a2.s4, a3.s4), (half4)(a0.s5, a1.s5, a2.s5, a3.s5));</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 16);</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span> </div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  val0 = (half8)((half4)(a0.s6, a1.s6, a2.s6, a3.s6), (half4)(a0.s7, a1.s7, a2.s7, a3.s7));</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  vstore8(val0, 0, ((__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 24);</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span> }</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span> </div><div class="line"><a name="l00219"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#aa66a7b6b0420e54ec173743e6c5bfc45"> 219</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#aa66a7b6b0420e54ec173743e6c5bfc45">gemm_interleave4x4_8bit</a>(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span> {</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span> </div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  <span class="comment">/* Load values from Matrix A */</span></div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  uchar16 a0 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0)));</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  uchar16 a1 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 1)));</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  uchar16 a2 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 2)));</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  uchar16 a3 = vload16(0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 3)));</div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span> </div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  uchar16 val0 = (uchar16)((uchar4)(a0.s0, a1.s0, a2.s0, a3.s0), (uchar4)(a0.s1, a1.s1, a2.s1, a3.s1),</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  (uchar4)(a0.s2, a1.s2, a2.s2, a3.s2), (uchar4)(a0.s3, a1.s3, a2.s3, a3.s3));</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 0);</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span> </div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  val0 = (uchar16)((uchar4)(a0.s4, a1.s4, a2.s4, a3.s4), (uchar4)(a0.s5, a1.s5, a2.s5, a3.s5),</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  (uchar4)(a0.s6, a1.s6, a2.s6, a3.s6), (uchar4)(a0.s7, a1.s7, a2.s7, a3.s7));</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 16);</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span> </div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  val0 = (uchar16)((uchar4)(a0.s8, a1.s8, a2.s8, a3.s8), (uchar4)(a0.s9, a1.s9, a2.s9, a3.s9),</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  (uchar4)(a0.sA, a1.sA, a2.sA, a3.sA), (uchar4)(a0.sB, a1.sB, a2.sB, a3.sB));</div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 32);</div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span> </div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  val0 = (uchar16)((uchar4)(a0.sC, a1.sC, a2.sC, a3.sC), (uchar4)(a0.sD, a1.sD, a2.sD, a3.sD),</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  (uchar4)(a0.sE, a1.sE, a2.sE, a3.sE), (uchar4)(a0.sF, a1.sF, a2.sF, a3.sF));</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  vstore16(val0, 0, ((__global uchar *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) + 48);</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span> }</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span> </div><div class="line"><a name="l00262"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a7c619984bc9c31f5f6c19ecc88b56581"> 262</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a7c619984bc9c31f5f6c19ecc88b56581">gemm_accumulate_biases_f32</a>(</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  <a class="code" href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span> {</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <a class="code" href="struct_image.xhtml">Image</a> accum = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span> </div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  float4 accum_value = vload4(0, (__global <span class="keywordtype">float</span> *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  float4 biases_value = vload4(0, (__global <span class="keywordtype">float</span> *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  accum_value = biases_value + accum_value;</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">// Store result in the accummulate buffer</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  vstore4(accum_value, 0, (__global <span class="keywordtype">float</span> *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span> }</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span> </div><div class="line"><a name="l00290"></a><span class="lineno"><a class="line" href="gemm_8cl.xhtml#a76295c333dfa3c9bb20975579dce87e8"> 290</a></span> __kernel <span class="keywordtype">void</span> <a class="code" href="gemm_8cl.xhtml#a76295c333dfa3c9bb20975579dce87e8">gemm_accumulate_biases_f16</a>(</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(accum),</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <a class="code" href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases))</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span> {</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <a class="code" href="struct_image.xhtml">Image</a> accum = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(accum);</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a>(biases);</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span> </div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  half8 accum_value = vload8(0, (__global half *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  half8 biases_value = vload8(0, (__global half *)biases.<a class="code" href="struct_vector.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  accum_value = biases_value + accum_value;</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">// Store result in the accummulate buffer</span></div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  vstore8(accum_value, 0, (__global half *)accum.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span> }</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> </div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> <span class="preprocessor">#if(defined WIDTH_MATRIX_B)</span></div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> </div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span> __kernel <span class="keywordtype">void</span> gemm_mm_u8(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <span class="keywordtype">int</span> a_offset,</div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  <span class="keywordtype">int</span> b_offset,</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  <span class="keywordtype">int</span> c_offset,</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  <span class="keywordtype">int</span> c_mult_int,</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  <span class="keywordtype">int</span> shift)</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span> {</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span> </div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  (src1_stride_y));</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span> </div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span> </div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span> </div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  int16 c00 = 0.0f;</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  int16 c10 = 0.0f;</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  int16 c20 = 0.0f;</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  int16 c30 = 0.0f;</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span> </div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 32))</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  {</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  int8 a0 = (int8)a_offset + convert_int8(vload8(0, ((__global uchar *)src0_ptr) + src_addr.s0));</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));</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>  c00 += (int16)a0.s0 * b0;</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  c10 += (int16)a0.s1 * b0;</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  c20 += (int16)a0.s2 * b0;</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  c30 += (int16)a0.s3 * b0;</div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span> </div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  int16 b1 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1 + 16));</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span> </div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  c00 += (int16)a0.s4 * b1;</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  c10 += (int16)a0.s5 * b1;</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  c20 += (int16)a0.s6 * b1;</div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  c30 += (int16)a0.s7 * b1;</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  }</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span> </div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 16))</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  {</div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  int4 a0 = (int4)a_offset + convert_int4(vload4(0, ((__global uchar *)src0_ptr) + src_addr.s0));</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  int16 b0 = (int16)b_offset + convert_int16(vload16(0, ((__global uchar *)src1_ptr) + src_addr.s1));</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span> </div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  c00 += (int16)a0.s0 * b0;</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  c10 += (int16)a0.s1 * b0;</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  c20 += (int16)a0.s2 * b0;</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  c30 += (int16)a0.s3 * b0;</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  }</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span> </div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span> </div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  c00 = (((int16)c_offset + c00) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  c10 = (((int16)c_offset + c10) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  c20 = (((int16)c_offset + c20) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  c30 = (((int16)c_offset + c30) * (int16)c_mult_int) >> shift;</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span> </div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  <span class="comment">/* Store 4x16 block */</span></div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  vstore16(convert_uchar16_sat(c00), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  vstore16(convert_uchar16_sat(c10), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  vstore16(convert_uchar16_sat(c20), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span>  vstore16(convert_uchar16_sat(c30), 0, (__global uchar *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span> }</div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span> </div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span> <span class="preprocessor">#if(defined WIDTH_MATRIX_B && defined ALPHA)</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span> </div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span> __kernel <span class="keywordtype">void</span> gemm_mm_f32_midgard(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> {</div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span> </div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  (src1_stride_y));</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span> </div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</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">/* Divide by 4 in order to get the src_addr in unit of float */</span></div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  src_addr = src_addr >> 2;</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span> </div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span> </div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  float4 c00 = 0.0f;</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  float4 c10 = 0.0f;</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  float4 c20 = 0.0f;</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  float4 c30 = 0.0f;</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>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 8))</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  {</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  float4 a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  float4 b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1);</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>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span> </div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0 + 4);</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1 + 4);</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span> </div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  }</div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span> </div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  <span class="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 4))</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  {</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  float4 a0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  float4 b0 = vload4(0, ((__global <span class="keywordtype">float</span> *)src1_ptr) + src_addr.s1);</div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span> </div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span>  c00 += (float4)a0.s0 * b0;</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  c10 += (float4)a0.s1 * b0;</div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  c20 += (float4)a0.s2 * b0;</div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  c30 += (float4)a0.s3 * b0;</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  }</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>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> </div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  c00 = c00 * (float4)ALPHA;</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  c10 = c10 * (float4)ALPHA;</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  c20 = c20 * (float4)ALPHA;</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  c30 = c30 * (float4)ALPHA;</div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span> </div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  <span class="comment">/* Store 4x4 block */</span></div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  vstore4(c00, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  vstore4(c10, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  vstore4(c20, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span>  vstore4(c30, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span> }</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span> </div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span> __kernel <span class="keywordtype">void</span> gemm_mm_f32_bifrost(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span> {</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span>  __global <span class="keywordtype">float</span> *src_addr_a = (__global <span class="keywordtype">float</span> *)(src0_ptr + get_global_id(1) * src0_stride_y + src0_offset_first_element_in_bytes);</div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  __global <span class="keywordtype">float</span> *src_addr_b = (__global <span class="keywordtype">float</span> *)(src1_ptr + get_global_id(0) * src1_stride_y + src1_offset_first_element_in_bytes);</div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span> </div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span>  __global <span class="keywordtype">float</span> *src_end_addr_b = src_addr_b + WIDTH_MATRIX_B;</div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span> </div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span>  <span class="keywordtype">float</span> c00 = 0.0f;</div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span>  <span class="keywordtype">float</span> c01 = 0.0f;</div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  <span class="keywordtype">float</span> c02 = 0.0f;</div><div class="line"><a name="l00551"></a><span class="lineno"> 551</span>  <span class="keywordtype">float</span> c03 = 0.0f;</div><div class="line"><a name="l00552"></a><span class="lineno"> 552</span>  <span class="keywordtype">float</span> c10 = 0.0f;</div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  <span class="keywordtype">float</span> c11 = 0.0f;</div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span>  <span class="keywordtype">float</span> c12 = 0.0f;</div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span>  <span class="keywordtype">float</span> c13 = 0.0f;</div><div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  <span class="keywordtype">float</span> c20 = 0.0f;</div><div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  <span class="keywordtype">float</span> c21 = 0.0f;</div><div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  <span class="keywordtype">float</span> c22 = 0.0f;</div><div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  <span class="keywordtype">float</span> c23 = 0.0f;</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  <span class="keywordtype">float</span> c30 = 0.0f;</div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span>  <span class="keywordtype">float</span> c31 = 0.0f;</div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  <span class="keywordtype">float</span> c32 = 0.0f;</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  <span class="keywordtype">float</span> c33 = 0.0f;</div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span> </div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  <span class="keywordflow">for</span>(; src_addr_b <= (src_end_addr_b - 16); src_addr_a += 16, src_addr_b += 16)</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>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00568"></a><span class="lineno"> 568</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span> </div><div class="line"><a name="l00571"></a><span class="lineno"> 571</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00572"></a><span class="lineno"> 572</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00573"></a><span class="lineno"> 573</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  c03 = fma(a0.s0, b0.s3, c03);</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>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00577"></a><span class="lineno"> 577</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span> </div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00583"></a><span class="lineno"> 583</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00584"></a><span class="lineno"> 584</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span> </div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span> </div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  a0 = vload4(0, src_addr_a + 4);</div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span>  b0 = vload4(0, src_addr_b + 4);</div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span> </div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  c03 = fma(a0.s0, b0.s3, c03);</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>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span> </div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00606"></a><span class="lineno"> 606</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00608"></a><span class="lineno"> 608</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00609"></a><span class="lineno"> 609</span> </div><div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00611"></a><span class="lineno"> 611</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00613"></a><span class="lineno"> 613</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00614"></a><span class="lineno"> 614</span> </div><div class="line"><a name="l00615"></a><span class="lineno"> 615</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span>  a0 = vload4(0, src_addr_a + 8);</div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  b0 = vload4(0, src_addr_b + 8);</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>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00621"></a><span class="lineno"> 621</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00622"></a><span class="lineno"> 622</span>  c03 = fma(a0.s0, b0.s3, c03);</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>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  c13 = fma(a0.s1, b0.s3, c13);</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>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span> </div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span> </div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  a0 = vload4(0, src_addr_a + 12);</div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span>  b0 = vload4(0, src_addr_b + 12);</div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span> </div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00646"></a><span class="lineno"> 646</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00647"></a><span class="lineno"> 647</span> </div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span>  c13 = fma(a0.s1, b0.s3, c13);</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>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00654"></a><span class="lineno"> 654</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span> </div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00660"></a><span class="lineno"> 660</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  }</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span> </div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span>  <span class="keywordflow">for</span>(; src_addr_b < src_end_addr_b; src_addr_a += 4, src_addr_b += 4)</div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span>  {</div><div class="line"><a name="l00666"></a><span class="lineno"> 666</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span>  float4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  float4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00669"></a><span class="lineno"> 669</span> </div><div class="line"><a name="l00670"></a><span class="lineno"> 670</span>  c00 = fma(a0.s0, b0.s0, c00);</div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span>  c01 = fma(a0.s0, b0.s1, c01);</div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  c02 = fma(a0.s0, b0.s2, c02);</div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span>  c03 = fma(a0.s0, b0.s3, c03);</div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span> </div><div class="line"><a name="l00675"></a><span class="lineno"> 675</span>  c10 = fma(a0.s1, b0.s0, c10);</div><div class="line"><a name="l00676"></a><span class="lineno"> 676</span>  c11 = fma(a0.s1, b0.s1, c11);</div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>  c12 = fma(a0.s1, b0.s2, c12);</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span>  c13 = fma(a0.s1, b0.s3, c13);</div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span> </div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span>  c20 = fma(a0.s2, b0.s0, c20);</div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  c21 = fma(a0.s2, b0.s1, c21);</div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span>  c22 = fma(a0.s2, b0.s2, c22);</div><div class="line"><a name="l00683"></a><span class="lineno"> 683</span>  c23 = fma(a0.s2, b0.s3, c23);</div><div class="line"><a name="l00684"></a><span class="lineno"> 684</span> </div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span>  c30 = fma(a0.s3, b0.s0, c30);</div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span>  c31 = fma(a0.s3, b0.s1, c31);</div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  c32 = fma(a0.s3, b0.s2, c32);</div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span>  c33 = fma(a0.s3, b0.s3, c33);</div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span>  }</div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span> </div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00692"></a><span class="lineno"> 692</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00693"></a><span class="lineno"> 693</span> </div><div class="line"><a name="l00694"></a><span class="lineno"> 694</span>  <span class="comment">// Multiply by the weight of matrix product</span></div><div class="line"><a name="l00695"></a><span class="lineno"> 695</span>  c00 = c00 * ALPHA;</div><div class="line"><a name="l00696"></a><span class="lineno"> 696</span>  c01 = c01 * ALPHA;</div><div class="line"><a name="l00697"></a><span class="lineno"> 697</span>  c02 = c02 * ALPHA;</div><div class="line"><a name="l00698"></a><span class="lineno"> 698</span>  c03 = c03 * ALPHA;</div><div class="line"><a name="l00699"></a><span class="lineno"> 699</span>  c10 = c10 * ALPHA;</div><div class="line"><a name="l00700"></a><span class="lineno"> 700</span>  c11 = c11 * ALPHA;</div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  c12 = c12 * ALPHA;</div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  c13 = c13 * ALPHA;</div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  c20 = c20 * ALPHA;</div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span>  c21 = c21 * ALPHA;</div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span>  c22 = c22 * ALPHA;</div><div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  c23 = c23 * ALPHA;</div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span>  c30 = c30 * ALPHA;</div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>  c31 = c31 * ALPHA;</div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span>  c32 = c32 * ALPHA;</div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>  c33 = c33 * ALPHA;</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span> </div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>  barrier(CLK_GLOBAL_MEM_FENCE);</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span> </div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00715"></a><span class="lineno"> 715</span>  vstore4((float4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  vstore4((float4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00717"></a><span class="lineno"> 717</span>  vstore4((float4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  vstore4((float4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00719"></a><span class="lineno"> 719</span> }</div><div class="line"><a name="l00720"></a><span class="lineno"> 720</span> </div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span> __kernel <span class="keywordtype">void</span> gemm_mm_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span> {</div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span>  <span class="comment">/* src_addr.s0 = address of matrix A */</span></div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  <span class="comment">/* src_addr.s1 = address of matrix B */</span></div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span> </div><div class="line"><a name="l00752"></a><span class="lineno"> 752</span>  <span class="comment">/* Compute address for matrix A and B */</span></div><div class="line"><a name="l00753"></a><span class="lineno"> 753</span>  int2 src_addr = (int2)(get_global_id(1), get_global_id(0)) * (int2)((src0_stride_y),</div><div class="line"><a name="l00754"></a><span class="lineno"> 754</span>  (src1_stride_y));</div><div class="line"><a name="l00755"></a><span class="lineno"> 755</span> </div><div class="line"><a name="l00756"></a><span class="lineno"> 756</span>  <span class="comment">/* Add offset_first_element_in_bytes */</span></div><div class="line"><a name="l00757"></a><span class="lineno"> 757</span>  src_addr = src_addr + ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</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">/* Divide by 2 in order to get the src_addr in unit of half */</span></div><div class="line"><a name="l00760"></a><span class="lineno"> 760</span>  src_addr = src_addr >> 1;</div><div class="line"><a name="l00761"></a><span class="lineno"> 761</span> </div><div class="line"><a name="l00762"></a><span class="lineno"> 762</span>  <span class="comment">/* Compute end row address for matrix B */</span></div><div class="line"><a name="l00763"></a><span class="lineno"> 763</span>  <span class="keywordtype">int</span> end_row_mtx_b = src_addr.s1 + WIDTH_MATRIX_B;</div><div class="line"><a name="l00764"></a><span class="lineno"> 764</span> </div><div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  <span class="comment">/* Reset accumulators */</span></div><div class="line"><a name="l00766"></a><span class="lineno"> 766</span>  half8 c00 = 0.0f;</div><div class="line"><a name="l00767"></a><span class="lineno"> 767</span>  half8 c10 = 0.0f;</div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span>  half8 c20 = 0.0f;</div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  half8 c30 = 0.0f;</div><div class="line"><a name="l00770"></a><span class="lineno"> 770</span> </div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  <span class="keywordflow">for</span>(; src_addr.s1 <= (end_row_mtx_b - 8); src_addr += (int2)(8, 16))</div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  {</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span>  half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span>  half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);</div><div class="line"><a name="l00776"></a><span class="lineno"> 776</span> </div><div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00779"></a><span class="lineno"> 779</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span> </div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0 + 4);</div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span>  b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1 + 8);</div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span> </div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00787"></a><span class="lineno"> 787</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00788"></a><span class="lineno"> 788</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00789"></a><span class="lineno"> 789</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00790"></a><span class="lineno"> 790</span>  }</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="keywordflow">for</span>(; src_addr.s1 < end_row_mtx_b; src_addr += (int2)(4, 8))</div><div class="line"><a name="l00793"></a><span class="lineno"> 793</span>  {</div><div class="line"><a name="l00794"></a><span class="lineno"> 794</span>  <span class="comment">/* Load values from matrix A (interleaved) and matrix B (transposed) */</span></div><div class="line"><a name="l00795"></a><span class="lineno"> 795</span>  half4 a0 = vload4(0, ((__global half *)src0_ptr) + src_addr.s0);</div><div class="line"><a name="l00796"></a><span class="lineno"> 796</span>  half8 b0 = vload8(0, ((__global half *)src1_ptr) + src_addr.s1);</div><div class="line"><a name="l00797"></a><span class="lineno"> 797</span> </div><div class="line"><a name="l00798"></a><span class="lineno"> 798</span>  c00 += (half8)a0.s0 * b0;</div><div class="line"><a name="l00799"></a><span class="lineno"> 799</span>  c10 += (half8)a0.s1 * b0;</div><div class="line"><a name="l00800"></a><span class="lineno"> 800</span>  c20 += (half8)a0.s2 * b0;</div><div class="line"><a name="l00801"></a><span class="lineno"> 801</span>  c30 += (half8)a0.s3 * b0;</div><div class="line"><a name="l00802"></a><span class="lineno"> 802</span>  }</div><div class="line"><a name="l00803"></a><span class="lineno"> 803</span> </div><div class="line"><a name="l00804"></a><span class="lineno"> 804</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00805"></a><span class="lineno"> 805</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00806"></a><span class="lineno"> 806</span> </div><div class="line"><a name="l00807"></a><span class="lineno"> 807</span>  <span class="comment">/* Multiply by the weight of matrix product */</span></div><div class="line"><a name="l00808"></a><span class="lineno"> 808</span>  c00 = c00 * (half8)ALPHA;</div><div class="line"><a name="l00809"></a><span class="lineno"> 809</span>  c10 = c10 * (half8)ALPHA;</div><div class="line"><a name="l00810"></a><span class="lineno"> 810</span>  c20 = c20 * (half8)ALPHA;</div><div class="line"><a name="l00811"></a><span class="lineno"> 811</span>  c30 = c30 * (half8)ALPHA;</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">/* Store 4x8 block */</span></div><div class="line"><a name="l00814"></a><span class="lineno"> 814</span>  vstore8(c00, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00815"></a><span class="lineno"> 815</span>  vstore8(c10, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 1)));</div><div class="line"><a name="l00816"></a><span class="lineno"> 816</span>  vstore8(c20, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 2)));</div><div class="line"><a name="l00817"></a><span class="lineno"> 817</span>  vstore8(c30, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 3)));</div><div class="line"><a name="l00818"></a><span class="lineno"> 818</span> }</div><div class="line"><a name="l00819"></a><span class="lineno"> 819</span> </div><div class="line"><a name="l00820"></a><span class="lineno"> 820</span> <span class="preprocessor">#if(defined WIDTH_VECTOR_A)</span></div><div class="line"><a name="l00821"></a><span class="lineno"> 821</span> </div><div class="line"><a name="l00846"></a><span class="lineno"> 846</span> __kernel <span class="keywordtype">void</span> gemm_vm_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00847"></a><span class="lineno"> 847</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00848"></a><span class="lineno"> 848</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00849"></a><span class="lineno"> 849</span> {</div><div class="line"><a name="l00850"></a><span class="lineno"> 850</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * 4;</div><div class="line"><a name="l00851"></a><span class="lineno"> 851</span> </div><div class="line"><a name="l00852"></a><span class="lineno"> 852</span>  <span class="comment">/* Compute the address for the vector A and matrix B */</span></div><div class="line"><a name="l00853"></a><span class="lineno"> 853</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00854"></a><span class="lineno"> 854</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l00855"></a><span class="lineno"> 855</span> </div><div class="line"><a name="l00856"></a><span class="lineno"> 856</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="l00857"></a><span class="lineno"> 857</span> </div><div class="line"><a name="l00858"></a><span class="lineno"> 858</span>  float4 acc = 0.0f;</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="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 2 * <span class="keyword">sizeof</span>(float)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>), 2 * src1_stride_y))</div><div class="line"><a name="l00861"></a><span class="lineno"> 861</span>  {</div><div class="line"><a name="l00862"></a><span class="lineno"> 862</span>  float2 a0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00863"></a><span class="lineno"> 863</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l00864"></a><span class="lineno"> 864</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="l00865"></a><span class="lineno"> 865</span> </div><div class="line"><a name="l00866"></a><span class="lineno"> 866</span>  acc += b0 * (float4)a0.s0;</div><div class="line"><a name="l00867"></a><span class="lineno"> 867</span>  acc += b1 * (float4)a0.s1;</div><div class="line"><a name="l00868"></a><span class="lineno"> 868</span>  }</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="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="l00871"></a><span class="lineno"> 871</span>  {</div><div class="line"><a name="l00872"></a><span class="lineno"> 872</span>  <span class="keywordtype">float</span> a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00873"></a><span class="lineno"> 873</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l00874"></a><span class="lineno"> 874</span> </div><div class="line"><a name="l00875"></a><span class="lineno"> 875</span>  acc += b0 * (float4)a0;</div><div class="line"><a name="l00876"></a><span class="lineno"> 876</span>  }</div><div class="line"><a name="l00877"></a><span class="lineno"> 877</span> </div><div class="line"><a name="l00878"></a><span class="lineno"> 878</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00879"></a><span class="lineno"> 879</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00880"></a><span class="lineno"> 880</span> </div><div class="line"><a name="l00881"></a><span class="lineno"> 881</span>  <span class="comment">/* Multiply by the weight of vector-matrix product */</span></div><div class="line"><a name="l00882"></a><span class="lineno"> 882</span>  acc = acc * (float4)ALPHA;</div><div class="line"><a name="l00883"></a><span class="lineno"> 883</span> </div><div class="line"><a name="l00884"></a><span class="lineno"> 884</span>  vstore4(acc, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00885"></a><span class="lineno"> 885</span> }</div><div class="line"><a name="l00886"></a><span class="lineno"> 886</span> </div><div class="line"><a name="l00912"></a><span class="lineno"> 912</span> __kernel <span class="keywordtype">void</span> gemm_vm_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00913"></a><span class="lineno"> 913</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00914"></a><span class="lineno"> 914</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00915"></a><span class="lineno"> 915</span> {</div><div class="line"><a name="l00916"></a><span class="lineno"> 916</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * 8;</div><div class="line"><a name="l00917"></a><span class="lineno"> 917</span> </div><div class="line"><a name="l00918"></a><span class="lineno"> 918</span>  <span class="comment">/* Compute the address for the vector A and matrix B */</span></div><div class="line"><a name="l00919"></a><span class="lineno"> 919</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00920"></a><span class="lineno"> 920</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(half);</div><div class="line"><a name="l00921"></a><span class="lineno"> 921</span> </div><div class="line"><a name="l00922"></a><span class="lineno"> 922</span>  <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + (WIDTH_VECTOR_A * <span class="keyword">sizeof</span>(half));</div><div class="line"><a name="l00923"></a><span class="lineno"> 923</span> </div><div class="line"><a name="l00924"></a><span class="lineno"> 924</span>  half8 acc = 0.0f;</div><div class="line"><a name="l00925"></a><span class="lineno"> 925</span> </div><div class="line"><a name="l00926"></a><span class="lineno"> 926</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 4 * <span class="keyword">sizeof</span>(half)); src_addr += (int2)(4 * <span class="keyword">sizeof</span>(half), 4 * src1_stride_y))</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>  half4 a0 = vload4(0, (__global half *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00929"></a><span class="lineno"> 929</span>  half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 0 * src1_stride_y));</div><div class="line"><a name="l00930"></a><span class="lineno"> 930</span>  half8 b1 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 1 * src1_stride_y));</div><div class="line"><a name="l00931"></a><span class="lineno"> 931</span>  half8 b2 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 2 * src1_stride_y));</div><div class="line"><a name="l00932"></a><span class="lineno"> 932</span>  half8 b3 = vload8(0, (__global half *)(src1_ptr + src_addr.s1 + 3 * src1_stride_y));</div><div class="line"><a name="l00933"></a><span class="lineno"> 933</span> </div><div class="line"><a name="l00934"></a><span class="lineno"> 934</span>  acc += b0 * (half8)a0.s0;</div><div class="line"><a name="l00935"></a><span class="lineno"> 935</span>  acc += b1 * (half8)a0.s1;</div><div class="line"><a name="l00936"></a><span class="lineno"> 936</span>  acc += b2 * (half8)a0.s2;</div><div class="line"><a name="l00937"></a><span class="lineno"> 937</span>  acc += b3 * (half8)a0.s3;</div><div class="line"><a name="l00938"></a><span class="lineno"> 938</span>  }</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="keywordflow">for</span>(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(<span class="keyword">sizeof</span>(half), src1_stride_y))</div><div class="line"><a name="l00941"></a><span class="lineno"> 941</span>  {</div><div class="line"><a name="l00942"></a><span class="lineno"> 942</span>  half a0 = *((__global half *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l00943"></a><span class="lineno"> 943</span>  half8 b0 = vload8(0, (__global half *)(src1_ptr + src_addr.s1));</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>  acc += b0 * (half8)a0;</div><div class="line"><a name="l00946"></a><span class="lineno"> 946</span>  }</div><div class="line"><a name="l00947"></a><span class="lineno"> 947</span> </div><div class="line"><a name="l00948"></a><span class="lineno"> 948</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l00949"></a><span class="lineno"> 949</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00950"></a><span class="lineno"> 950</span> </div><div class="line"><a name="l00951"></a><span class="lineno"> 951</span>  <span class="comment">/* Multiply by the weight of vector-matrix product */</span></div><div class="line"><a name="l00952"></a><span class="lineno"> 952</span>  acc = acc * (half8)ALPHA;</div><div class="line"><a name="l00953"></a><span class="lineno"> 953</span> </div><div class="line"><a name="l00954"></a><span class="lineno"> 954</span>  vstore8(acc, 0, (__global half *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l00955"></a><span class="lineno"> 955</span> }</div><div class="line"><a name="l00956"></a><span class="lineno"> 956</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined WIDTH_VECTOR_A) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00957"></a><span class="lineno"> 957</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined WIDTH_MATRIX_B && defined ALPHA) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00958"></a><span class="lineno"> 958</span> </div><div class="line"><a name="l00959"></a><span class="lineno"> 959</span> <span class="preprocessor">#if(defined BETA)</span></div><div class="line"><a name="l00960"></a><span class="lineno"> 960</span> </div><div class="line"><a name="l00977"></a><span class="lineno"> 977</span> __kernel <span class="keywordtype">void</span> gemm_ma_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l00978"></a><span class="lineno"> 978</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l00979"></a><span class="lineno"> 979</span> {</div><div class="line"><a name="l00980"></a><span class="lineno"> 980</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l00981"></a><span class="lineno"> 981</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00982"></a><span class="lineno"> 982</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00983"></a><span class="lineno"> 983</span> </div><div class="line"><a name="l00984"></a><span class="lineno"> 984</span>  <span class="comment">/* Load values from A x B */</span></div><div class="line"><a name="l00985"></a><span class="lineno"> 985</span>  float4 alpha_ab = vload4(0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00986"></a><span class="lineno"> 986</span> </div><div class="line"><a name="l00987"></a><span class="lineno"> 987</span>  <span class="comment">/* Load values from Matrix C */</span></div><div class="line"><a name="l00988"></a><span class="lineno"> 988</span>  float4 c = vload4(0, (__global <span class="keywordtype">float</span> *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00989"></a><span class="lineno"> 989</span> </div><div class="line"><a name="l00990"></a><span class="lineno"> 990</span>  <span class="comment">/* Computes alpha * axb + beta * c */</span></div><div class="line"><a name="l00991"></a><span class="lineno"> 991</span>  float4 out = alpha_ab + (float4)BETA * c;</div><div class="line"><a name="l00992"></a><span class="lineno"> 992</span> </div><div class="line"><a name="l00993"></a><span class="lineno"> 993</span>  <span class="comment">/* Store final result in axb matrix */</span></div><div class="line"><a name="l00994"></a><span class="lineno"> 994</span>  vstore4(out, 0, (__global <span class="keywordtype">float</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00995"></a><span class="lineno"> 995</span> }</div><div class="line"><a name="l00996"></a><span class="lineno"> 996</span> </div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span> __kernel <span class="keywordtype">void</span> gemm_ma_f16(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src),</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l01014"></a><span class="lineno"> 1014</span> {</div><div class="line"><a name="l01015"></a><span class="lineno"> 1015</span>  <span class="comment">/* Compute source and destination addresses */</span></div><div class="line"><a name="l01016"></a><span class="lineno"> 1016</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l01017"></a><span class="lineno"> 1017</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01018"></a><span class="lineno"> 1018</span> </div><div class="line"><a name="l01019"></a><span class="lineno"> 1019</span>  <span class="comment">/* Load values from A x B */</span></div><div class="line"><a name="l01020"></a><span class="lineno"> 1020</span>  half8 alpha_ab = vload8(0, (__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l01021"></a><span class="lineno"> 1021</span> </div><div class="line"><a name="l01022"></a><span class="lineno"> 1022</span>  <span class="comment">/* Load values from Matrix C */</span></div><div class="line"><a name="l01023"></a><span class="lineno"> 1023</span>  half8 c = vload8(0, (__global half *)src.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l01024"></a><span class="lineno"> 1024</span> </div><div class="line"><a name="l01025"></a><span class="lineno"> 1025</span>  <span class="comment">/* Computes alpha * axb + beta * c */</span></div><div class="line"><a name="l01026"></a><span class="lineno"> 1026</span>  half8 out = alpha_ab + (half8)BETA * c;</div><div class="line"><a name="l01027"></a><span class="lineno"> 1027</span> </div><div class="line"><a name="l01028"></a><span class="lineno"> 1028</span>  <span class="comment">/* Store final result in axb matrix */</span></div><div class="line"><a name="l01029"></a><span class="lineno"> 1029</span>  vstore8(out, 0, (__global half *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l01030"></a><span class="lineno"> 1030</span> }</div><div class="line"><a name="l01031"></a><span class="lineno"> 1031</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined BETA) */</span><span class="preprocessor"></span></div><div class="line"><a name="l01032"></a><span class="lineno"> 1032</span> </div><div class="line"><a name="l01033"></a><span class="lineno"> 1033</span> <span class="preprocessor">#if(defined WIDTH_VECTOR_A)</span></div><div class="line"><a name="l01034"></a><span class="lineno"> 1034</span> </div><div class="line"><a name="l01061"></a><span class="lineno"> 1061</span> __kernel <span class="keywordtype">void</span> gemm_lc_vm_f32(<a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01062"></a><span class="lineno"> 1062</span>  <a class="code" href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src1),</div><div class="line"><a name="l01063"></a><span class="lineno"> 1063</span>  <a class="code" href="helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l01064"></a><span class="lineno"> 1064</span> {</div><div class="line"><a name="l01065"></a><span class="lineno"> 1065</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * 4;</div><div class="line"><a name="l01066"></a><span class="lineno"> 1066</span>  <span class="keywordtype">int</span> idy = get_global_id(1);</div><div class="line"><a name="l01067"></a><span class="lineno"> 1067</span> </div><div class="line"><a name="l01068"></a><span class="lineno"> 1068</span>  <span class="comment">/* Compute the address for the vector A and matrix B */</span></div><div class="line"><a name="l01069"></a><span class="lineno"> 1069</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="l01070"></a><span class="lineno"> 1070</span>  src_addr.s1 += idx * <span class="keyword">sizeof</span>(float);</div><div class="line"><a name="l01071"></a><span class="lineno"> 1071</span> </div><div class="line"><a name="l01072"></a><span class="lineno"> 1072</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="l01073"></a><span class="lineno"> 1073</span> </div><div class="line"><a name="l01074"></a><span class="lineno"> 1074</span>  float4 acc = 0.0f;</div><div class="line"><a name="l01075"></a><span class="lineno"> 1075</span> </div><div class="line"><a name="l01076"></a><span class="lineno"> 1076</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 2 * <span class="keyword">sizeof</span>(float)); src_addr += (int2)(2 * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>), 2 * src1_stride_y))</div><div class="line"><a name="l01077"></a><span class="lineno"> 1077</span>  {</div><div class="line"><a name="l01078"></a><span class="lineno"> 1078</span>  float2 a0 = vload2(0, (__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l01079"></a><span class="lineno"> 1079</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01080"></a><span class="lineno"> 1080</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="l01081"></a><span class="lineno"> 1081</span> </div><div class="line"><a name="l01082"></a><span class="lineno"> 1082</span>  acc += b0 * (float4)a0.s0;</div><div class="line"><a name="l01083"></a><span class="lineno"> 1083</span>  acc += b1 * (float4)a0.s1;</div><div class="line"><a name="l01084"></a><span class="lineno"> 1084</span>  }</div><div class="line"><a name="l01085"></a><span class="lineno"> 1085</span> </div><div class="line"><a name="l01086"></a><span class="lineno"> 1086</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="l01087"></a><span class="lineno"> 1087</span>  {</div><div class="line"><a name="l01088"></a><span class="lineno"> 1088</span>  <span class="keywordtype">float</span> a0 = *((__global <span class="keywordtype">float</span> *)(src0_ptr + src_addr.s0));</div><div class="line"><a name="l01089"></a><span class="lineno"> 1089</span>  float4 b0 = vload4(0, (__global <span class="keywordtype">float</span> *)(src1_ptr + src_addr.s1));</div><div class="line"><a name="l01090"></a><span class="lineno"> 1090</span> </div><div class="line"><a name="l01091"></a><span class="lineno"> 1091</span>  acc += b0 * (float4)a0;</div><div class="line"><a name="l01092"></a><span class="lineno"> 1092</span>  }</div><div class="line"><a name="l01093"></a><span class="lineno"> 1093</span> </div><div class="line"><a name="l01094"></a><span class="lineno"> 1094</span>  <span class="comment">/* Compute destination address */</span></div><div class="line"><a name="l01095"></a><span class="lineno"> 1095</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01096"></a><span class="lineno"> 1096</span> </div><div class="line"><a name="l01097"></a><span class="lineno"> 1097</span>  vstore4(acc, 0, (__global <span class="keywordtype">float</span> *)(<a class="code" href="helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&dst, 0, 0)));</div><div class="line"><a name="l01098"></a><span class="lineno"> 1098</span> }</div><div class="line"><a name="l01099"></a><span class="lineno"> 1099</span> <span class="preprocessor">#endif </span><span class="comment">/* (defined WIDTH_VECTOR_A) */</span><span class="preprocessor"></span></div><div class="ttc" id="struct_vector_xhtml"><div class="ttname"><a href="struct_vector.xhtml">Vector</a></div><div class="ttdoc">Structure to hold Vector information. </div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00087">helpers.h:87</a></div></div>
<div class="ttc" id="helpers_8h_xhtml_a527bfdf5eeb306f1cf01c4a8e29f38e0"><div class="ttname"><a href="helpers_8h.xhtml#a527bfdf5eeb306f1cf01c4a8e29f38e0">CONVERT_TO_VECTOR_STRUCT</a></div><div class="ttdeci">#define CONVERT_TO_VECTOR_STRUCT(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00067">helpers.h:67</a></div></div>
<div class="ttc" id="gemm_8cl_xhtml_aa66a7b6b0420e54ec173743e6c5bfc45"><div class="ttname"><a href="gemm_8cl.xhtml#aa66a7b6b0420e54ec173743e6c5bfc45">gemm_interleave4x4_8bit</a></div><div class="ttdeci">__kernel void gemm_interleave4x4_8bit(__global uchar *src_ptr, uint src_stride_x, uint src_step_x, uint src_stride_y, uint src_step_y, uint src_offset_first_element_in_bytes, __global uchar *dst_ptr, uint dst_stride_x, uint dst_step_x, uint dst_stride_y, uint dst_step_y, uint dst_offset_first_element_in_bytes)</div><div class="ttdoc">This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values...</div><div class="ttdef"><b>Definition:</b> <a href="gemm_8cl_source.xhtml#l00219">gemm.cl:219</a></div></div>
<div class="ttc" id="gemm_8cl_xhtml_a76295c333dfa3c9bb20975579dce87e8"><div class="ttname"><a href="gemm_8cl.xhtml#a76295c333dfa3c9bb20975579dce87e8">gemm_accumulate_biases_f16</a></div><div class="ttdeci">__kernel void gemm_accumulate_biases_f16(__global uchar *accum_ptr, uint accum_stride_x, uint accum_step_x, uint accum_stride_y, uint accum_step_y, uint accum_offset_first_element_in_bytes, __global uchar *biases_ptr, uint biases_stride_x, uint biases_step_x, uint biases_offset_first_element_in_bytes)</div><div class="ttdoc">This kernel accumulates each row with the biases vector. </div><div class="ttdef"><b>Definition:</b> <a href="gemm_8cl_source.xhtml#l00290">gemm.cl:290</a></div></div>
@@ -127,6 +128,7 @@
<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="helpers_8h_source.xhtml#l00089">helpers.h:89</a></div></div>
<div class="ttc" id="helpers_8h_xhtml_a40a6eb9f2a7712f08d6bb8ff6c9e6ca7"><div class="ttname"><a href="helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a></div><div class="ttdeci">#define VECTOR_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00043">helpers.h:43</a></div></div>
<div class="ttc" id="struct_image_xhtml"><div class="ttname"><a href="struct_image.xhtml">Image</a></div><div class="ttdoc">Structure to hold Image information. </div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00095">helpers.h:95</a></div></div>
+<div class="ttc" id="helpers_8h_xhtml_a6b83038822d1ae7ab619b684ed3b7fc0"><div class="ttname"><a href="helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a></div><div class="ttdeci">#define TENSOR3D_DECLARATION(name)</div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00057">helpers.h:57</a></div></div>
<div class="ttc" id="gemm_8cl_xhtml_a11d87ecaf1388b4f57404e062265b374"><div class="ttname"><a href="gemm_8cl.xhtml#a11d87ecaf1388b4f57404e062265b374">gemm_transpose1x8_f16</a></div><div class="ttdeci">__kernel void gemm_transpose1x8_f16(__global uchar *src_ptr, uint src_stride_x, uint src_step_x, uint src_stride_y, uint src_step_y, uint src_offset_first_element_in_bytes, __global uchar *dst_ptr, uint dst_stride_x, uint dst_step_x, uint dst_stride_y, uint dst_step_y, uint dst_offset_first_element_in_bytes)</div><div class="ttdoc">This OpenCL kernel computes the "vector" 1x8 transposition of input matrix. </div><div class="ttdef"><b>Definition:</b> <a href="gemm_8cl_source.xhtml#l00073">gemm.cl:73</a></div></div>
<div class="ttc" id="struct_image_xhtml_acf52c23cbd7424606c10a606524e3e32"><div class="ttname"><a href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">Image::ptr</a></div><div class="ttdeci">__global uchar * ptr</div><div class="ttdoc">Pointer to the starting postion of the buffer. </div><div class="ttdef"><b>Definition:</b> <a href="helpers_8h_source.xhtml#l00097">helpers.h:97</a></div></div>
<div class="ttc" id="gemm_8cl_xhtml_a560b2d50aa886edc0f5daf4fe729717f"><div class="ttname"><a href="gemm_8cl.xhtml#a560b2d50aa886edc0f5daf4fe729717f">gemm_interleave4x4_32bit</a></div><div class="ttdeci">__kernel void gemm_interleave4x4_32bit(__global uchar *src_ptr, uint src_stride_x, uint src_step_x, uint src_stride_y, uint src_step_y, uint src_offset_first_element_in_bytes, __global uchar *dst_ptr, uint dst_stride_x, uint dst_step_x, uint dst_stride_y, uint dst_step_y, uint dst_offset_first_element_in_bytes)</div><div class="ttdoc">This OpenCL kernel reshapes the input matrix transposing each 4x4 block and interleaving the values...</div><div class="ttdef"><b>Definition:</b> <a href="gemm_8cl_source.xhtml#l00137">gemm.cl:137</a></div></div>
@@ -140,7 +142,7 @@
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<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>
- <li class="footer">Generated on Wed May 3 2017 17:20:04 for ARM Compute Library by
+ <li class="footer">Generated on Fri Jun 23 2017 15:44:33 for Compute Library by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.11 </li>
</ul>