arm_compute v19.05
diff --git a/documentation/gemm__interleaved_8hpp_source.xhtml b/documentation/gemm__interleaved_8hpp_source.xhtml
index 88e03ca..c9b01e5 100644
--- a/documentation/gemm__interleaved_8hpp_source.xhtml
+++ b/documentation/gemm__interleaved_8hpp_source.xhtml
@@ -40,7 +40,7 @@
<img alt="Compute Library" src="https://raw.githubusercontent.com/ARM-software/ComputeLibrary/gh-pages/ACL_logo.png" style="max-width: 100%;margin-top: 15px;margin-left: 10px"/>
<td style="padding-left: 0.5em;">
<div id="projectname">
-  <span id="projectnumber">19.02</span>
+  <span id="projectnumber">19.05</span>
</div>
</td>
</tr>
@@ -100,7 +100,7 @@
<div class="title">gemm_interleaved.hpp</div> </div>
</div><!--header-->
<div class="contents">
-<a href="gemm__interleaved_8hpp.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) 2017-2019 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">#pragma once</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> </div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#include <stdio.h></span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="preprocessor">#include <assert.h></span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> </div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#include <algorithm></span></div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> </div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#include "arm_gemm.hpp"</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#include "<a class="code" href="utils_8hpp.xhtml">utils.hpp</a>"</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> </div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="preprocessor">#include "<a class="code" href="buffer__manager_8hpp.xhtml">buffer_manager.hpp</a>"</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="preprocessor">#include "<a class="code" href="mergeresults_8hpp.xhtml">mergeresults.hpp</a>"</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="preprocessor">#include "<a class="code" href="transform_8hpp.xhtml">transform.hpp</a>"</span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> </div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="preprocessor">#include "<a class="code" href="profiler_8hpp.xhtml">profiler.hpp</a>"</span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> </div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="comment">// Some macros used to decide how much working space to allocate.</span></div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="comment">// Round allocations up to the next cache line.</span></div><div class="line"><a name="l00044"></a><span class="lineno"><a class="line" href="gemm__interleaved_8hpp.xhtml#abd1ded15e3688096ead983e87f83fbd1"> 44</a></span> <span class="preprocessor">#define ALLOC_ROUND 64</span></div><div class="line"><a name="l00045"></a><span class="lineno"><a class="line" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af"> 45</a></span> <span class="preprocessor">#define ROUND_UP(x) ((((x) + ALLOC_ROUND-1) / ALLOC_ROUND) * ALLOC_ROUND)</span></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">// Implementation of the GemmCommon abstract class.</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> <span class="comment">//</span></div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="comment">// This implementation interleaves the source matrices in blocks - good for</span></div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="comment">// larger matrices.</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="keyword">namespace </span><a class="code" href="namespacearm__gemm.xhtml">arm_gemm</a> {</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> </div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span> <span class="keyword">template</span><<span class="keyword">typename</span> strategy, <span class="keyword">typename</span> To, <span class="keyword">typename</span> Tr></div><div class="line"><a name="l00054"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml"> 54</a></span> <span class="keyword">class </span><a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> : <span class="keyword">public</span> GemmCommon<To, Tr> {</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keyword">typedef</span> <span class="keyword">typename</span> strategy::operand_type Toi;</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="keyword">typedef</span> <span class="keyword">typename</span> strategy::result_type Tri;</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> </div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  <span class="comment">/* const properties set by constructor */</span></div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  <span class="keyword">const</span> CPUInfo * <span class="keyword">const</span> _ci;</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> </div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Msize;</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Nsize;</div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Ksize;</div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span> </div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _nbatches;</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _nmulti;</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span> </div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _trA;</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _trB;</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> </div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> Tr _alpha;</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  <span class="keyword">const</span> Tr _beta;</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> _maxthreads;</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  <span class="keywordtype">int</span> _nthreads;</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _pretransposed;</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> </div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="comment">/* Blocking info */</span></div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _k_block=0;</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _x_block=0;</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Mround=0;</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span> </div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  <span class="comment">/* Working space, pretransposed buffer, buffer manager */</span></div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  <span class="keyword">const</span> Toi *_B_transposed=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml">BufferManager</a> *_bm=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="keywordtype">void</span> *_working_space=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span> </div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  <span class="comment">/* We will need to walk through the blocks of B in a few contexts, so</span></div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> <span class="comment"> * factor that out. */</span></div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  <span class="keyword">class </span>blockwalker {</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="keyword">private</span>:</div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  <span class="comment">/* Size loops, etc. based on our parent's configuration */</span></div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  <span class="keyword">const</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved<strategy, To, Tr></a> &_parent;</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span> </div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="comment">/* K, X and multi parameters for current iteration. */</span></div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _k0=0, _x0=0, _multi=0;</div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> </div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _index=0;</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  <span class="keywordtype">bool</span> _done=<span class="keyword">false</span>;</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  <span class="keywordtype">bool</span> _newkblock=<span class="keyword">true</span>;</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  <span class="keywordtype">bool</span> _newmulti=<span class="keyword">true</span>;</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span> </div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  <span class="keyword">public</span>:</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  blockwalker(<span class="keyword">const</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved<strategy, To, Tr></a> &parent) : _parent(parent) { }</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span> </div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> xmax() {</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keywordflow">return</span> std::min(_x0 + _parent._x_block, _parent._Nsize);</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  }</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span> </div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> kmax() {</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  <span class="keywordflow">return</span> std::min(_k0 + _parent._k_block, _parent._Ksize);</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  }</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">/* Advance to the next block, return false at the end. */</span></div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  <span class="keywordtype">bool</span> advance(<span class="keywordtype">void</span>) {</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  <span class="keywordflow">if</span> (_done) {</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  <span class="keywordflow">return</span> <span class="keyword">false</span>;</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> </div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  _newkblock=<span class="keyword">false</span>;</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  _x0 += _parent._x_block;</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  <span class="keywordflow">if</span> (_x0 >= _parent._Nsize) {</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  _x0=0;</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  _k0 += _parent._k_block;</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="keywordflow">if</span> (_k0 >= _parent._Ksize) {</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  _k0=0;</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  _multi++;</div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="keywordflow">if</span> (_multi >= _parent._nmulti) {</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  _done=<span class="keyword">true</span>;</div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordflow">return</span> <span class="keyword">false</span>;</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  }</div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  _newmulti=<span class="keyword">true</span>;</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  }</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  _newkblock=<span class="keyword">true</span>;</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  }</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  _index++;</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span> </div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <span class="keywordflow">return</span> <span class="keyword">true</span>;</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> </div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k0(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _k0; }</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x0(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _x0; }</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> multi(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _multi; }</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> index(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _index; }</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="keywordtype">bool</span> done(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _done; }</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="keywordtype">bool</span> newkblock(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _newkblock; }</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  };</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span> </div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  <span class="comment">// A working size: One of these needed, regardless of thread count. Divided according to window.</span></div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  <span class="keywordtype">size_t</span> get_a_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Toi) * _k_block * _Mround * _nbatches);</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> </div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="comment">// B working size: 0, 1 or 3 of these needed depending on pretransposed and threading settings.</span></div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="keywordtype">size_t</span> get_b_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Toi) * _x_block * _k_block);</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  }</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>  <span class="comment">// C working size: One needed per thread.</span></div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  <span class="keywordtype">size_t</span> get_c_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Tri) * _x_block * strategy::out_height());</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  }</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span> </div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  <span class="comment">// Internal execute function.</span></div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// This supports both the "pretransposed" and "standard" interfaces via the template parameter.</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="keyword">template</span><<span class="keywordtype">bool</span> pretransposed></div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="keywordtype">void</span> execute_internal(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> start, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> end, <span class="keywordtype">int</span> threadid) {</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  profiler prof;</div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  strategy strat(_ci);</div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span> </div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  blockwalker next=current;</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span> </div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  <span class="comment">/* Translate 'start' and 'end' into a position within the batches and rows. */</span></div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> window_per_batch = _Mround / strategy::out_height();</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch_0 = start / window_per_batch;</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch_end = end / window_per_batch;</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 the M values to operate on */</span></div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> m_0 = (start - (batch_0 * window_per_batch)) * strategy::out_height();</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> m_max = (end - (batch_end * window_per_batch)) * strategy::out_height();</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">/* Make sure we've been set up correctly. */</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  assert(_B_transposed);</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  assert(_bm);</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> </div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  assert(_working_space);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  int8_t *working_space_bytes = reinterpret_cast<int8_t *>(_working_space);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span> </div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  <span class="comment">// Private buffers. Treat working_space as an array of C buffers</span></div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <span class="comment">// (one per thread) first, followed by the (window-divided) A</span></div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  <span class="comment">// buffer.</span></div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  <span class="comment">// Set a_panel to the base of the A buffers - compute offsets into it based on M/batches later.</span></div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  Toi * <span class="keyword">const</span> a_panel = reinterpret_cast<Toi *>(working_space_bytes + (_maxthreads * get_c_working_size()));</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  Tri * <span class="keyword">const</span> c_panel = reinterpret_cast<Tri *>(working_space_bytes + (threadid * get_c_working_size()));</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span> </div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  <span class="comment">// Shared buffers - these come either from BufferManager or _B_transposed.</span></div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  <span class="keyword">const</span> Toi *b_panel;</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span> </div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  b_panel = _B_transposed;</div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  }</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span> </div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="comment">//printf("Starting GEMM loop, x_block=%d, k_block=%d\n", _x_block, _k_block);</span></div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span> </div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  <span class="comment">// newkblock() is always true on the first iteration, so this will be set properly on the first loop.</span></div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="keywordtype">int</span> kern_k = 0;</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span> </div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="keywordflow">for</span> (;!current.done();current.advance()) {</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  <span class="keywordflow">if</span> (current.newkblock()) {</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#ac19bf2c5c388b9fb21cd0668475b2cf7">PROFILE_PREPA</a>, (end - start) * strategy::out_height() * (current.kmax()-current.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = batch_0; batch <= batch_end; batch++) {</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_m = (batch == batch_0) ? m_0 : 0;</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> last_m = (batch == batch_end) ? m_max : _Msize;</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span> </div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <span class="keywordflow">if</span> (first_m >= last_m)</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  <span class="keywordflow">continue</span>;</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>  strat.transforms.PrepareA(a_panel + ((batch * _Mround + first_m) * _k_block),</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  this->_Aptr + (batch * this->_A_batch_stride) + (current.multi() * this->_A_multi_stride),</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  this->_lda, first_m, last_m, current.k0(), current.kmax(), _trA);</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  }</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span> </div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  <span class="comment">// Figure out how many "K" the kernel will actually process.</span></div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  kern_k = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(current.kmax() - current.k0(), strategy::k_unroll());</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  kern_k *= strat.k_unroll();</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  }</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>  <span class="keywordtype">int</span> bblocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(current.xmax() - current.x0(), strategy::out_width());</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span> </div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  <span class="keywordflow">if</span> (!pretransposed) {</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  <span class="comment">/* Look ahead to the next block and populate it if necessary.</span></div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span> <span class="comment"> * This avoids the populate operation becoming a bottleneck, and</span></div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> <span class="comment"> * helps keep the threads synchronized (the first thread to get</span></div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span> <span class="comment"> * here will populate while the rest will advance).</span></div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span> <span class="comment"> *</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span> <span class="comment"> * If we are running single threaded, bm->try_populate() will do</span></div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span> <span class="comment"> * nothing.</span></div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span> <span class="comment"> */</span></div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  <span class="keywordflow">if</span> (next.advance()) {</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a97814ef98fcd29d1fb5df1e877f8eabc">try_populate</a>(next.index(), [&](<span class="keywordtype">void</span> *buffer) {</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#af0477dfaa86fc08825cb3f31212f423c">PROFILE_PREPB</a>, (next.xmax()-next.x0()) * (next.kmax()-next.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span> </div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  Toi *b_panel = reinterpret_cast<Toi *>(buffer);</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span> </div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  strat.transforms.PrepareB(b_panel, this->_Bptr + (next.multi() * this->_B_multi_stride), this->_ldb,</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  next.x0(), next.xmax(), next.k0(), next.kmax(), _trB);</div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  });</div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  }</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span> </div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="comment">/* Get the buffer for this iteration from the BufferManager. */</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  b_panel = reinterpret_cast<Toi *>(_bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a427705bf76071f12aa29992bd5cffa73">get</a>(current.index(), [&](<span class="keywordtype">void</span> *bpv) {</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#af0477dfaa86fc08825cb3f31212f423c">PROFILE_PREPB</a>, (current.xmax()-current.x0()) * (current.kmax()-current.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span> <span class="preprocessor">#endif</span></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>  Toi *b_panel = reinterpret_cast<Toi *>(bpv);</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span> </div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  strat.transforms.PrepareB(b_panel, this->_Bptr + (current.multi() * this->_B_multi_stride), this->_ldb,</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  current.x0(), current.xmax(), current.k0(), current.kmax(), _trB);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  }));</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  }</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">/* Do the actual work. */</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = batch_0; batch <= batch_end; batch++) {</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_m = (batch == batch_0) ? m_0 : 0;</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> last_m = (batch == batch_end) ? m_max : _Msize;</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span> </div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  <span class="keyword">const</span> Toi *a_ptr = a_panel + (batch * _Mround + first_m) * _k_block;</div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span> </div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  <span class="keywordflow">if</span> (first_m >= last_m)</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  <span class="keywordflow">continue</span>;</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span> </div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> y=first_m; y<last_m; y+=strategy::out_height()) {</div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> ymax = std::min(_Msize, y + strategy::out_height());</div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span> </div><div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  {</div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#a97f4e398174b5e094f5262e5d785ac20">PROFILE_KERNEL</a>, (strategy::out_height() * bblocks * strategy::out_width() * kern_k));</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span> </div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  strat.kernel(a_ptr, b_panel, c_panel, 1, bblocks, kern_k);</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span> </div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  a_ptr += (strategy::out_height() * kern_k);</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  }</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span> </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> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#a2f7c3e65a9ccab509d7776832f562dd0">PROFILE_MERGE</a>, (strategy::out_height() * bblocks * strategy::out_width() * <span class="keyword">sizeof</span>(Tr)));</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  strat.transforms.Merge(this->_Cptr + (batch * this->_C_batch_stride) + (current.multi() * this->_C_multi_stride),</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  c_panel, this->_ldc, y, ymax, current.x0(), current.xmax(),</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  _alpha, (current.k0()==0 ? _beta : static_cast<Tr>(1)));</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>  }</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> </div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  b_panel += (bblocks * strat.out_width() * kern_k);</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a29a2894a4403bc4cc19aacc35cb91d54">release</a>(current.index());</div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  }</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  }</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  }</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span> </div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span> <span class="keyword">public</span>:</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#add13963d432a1937370721fa61755b9c">GemmInterleaved</a>(<a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> &) = <span class="keyword">delete</span>;</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> & <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#af03b8e4289fa0e112d315dd7e25ec696">operator= </a>(<a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> &) = <span class="keyword">delete</span>;</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span> </div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  <span class="comment">/* Constructor */</span></div><div class="line"><a name="l00320"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aec95ac91ca276b1ad8baf91520c0306f"> 320</a></span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aec95ac91ca276b1ad8baf91520c0306f">GemmInterleaved</a>(<span class="keyword">const</span> GemmArgs<Tr> &args)</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  : _ci(args._ci), _Msize(args._Msize), _Nsize(args._Nsize), _Ksize(args._Ksize),</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  _nbatches(args._nbatches), _nmulti(args._nmulti), _trA(args._trA), _trB(args._trB),</div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  _alpha(args._alpha), _beta(args._beta), _maxthreads(args._maxthreads), _nthreads(args._maxthreads),</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  _pretransposed(args._pretransposed_hint) {</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> L1_size = _ci->get_L1_cache_size();</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> L2_size = _ci->get_L2_cache_size();</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span> </div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  assert(_maxthreads > 0);</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span> </div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>  <span class="comment">// Work out blocking parameters, or override from provided GemmConfig</span></div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  <span class="keywordflow">if</span> (args._cfg && args._cfg->inner_block_size) {</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  _k_block = args._cfg->inner_block_size;</div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  <span class="comment">// k_block: Find out how much of the larger array can be loaded into half the cache.</span></div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  <span class="comment">// This should account for associative caches.</span></div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  _k_block = (L1_size / 2) / (<span class="keyword">sizeof</span>(Toi) * (std::max(strategy::out_width(), strategy::out_height())));</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span> </div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <span class="comment">// Needs to be (at least a single) multiple of the K unroll level.</span></div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  _k_block /= strategy::k_unroll();</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  _k_block = std::max(_k_block, 1U) * strategy::k_unroll();</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span> </div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  <span class="comment">// Now tune to presented problem size; this is how many blocks we need.</span></div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_k_blocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Ksize, _k_block);</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span> </div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <span class="comment">// So divide the space equally into that many blocks.</span></div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  _k_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Ksize, num_k_blocks);</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span> </div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="comment">// And round UP to the K unroll level required.</span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  _k_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_k_block, strategy::k_unroll());</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  _k_block *= strategy::k_unroll();</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  }</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span> </div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <span class="keywordflow">if</span> (args._cfg && args._cfg->outer_block_size) {</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  _x_block = args._cfg->outer_block_size;</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  <span class="comment">// x_block: Work out how many rows (of length k_block) will fit in the L2</span></div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="comment">// Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents.</span></div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  _x_block = (((L2_size * 9) / 10) - (_k_block * <span class="keyword">sizeof</span>(Toi) * (strategy::out_width() + strategy::out_height()))) /</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  (<span class="keyword">sizeof</span>(Toi) * _k_block);</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span> </div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  <span class="comment">// Needs to be (at least a single) multiple of the kernel output width.</span></div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  _x_block /= strategy::out_width();</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  _x_block = std::max(_x_block, 1U) * strategy::out_width();</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">// And tune to the presented problem size.</span></div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_x_blocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Nsize, _x_block);</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  _x_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Nsize, num_x_blocks);</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>  _x_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_x_block, strategy::out_width());</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  _x_block *= strategy::out_width();</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  }</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span> </div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  <span class="comment">// Work out the rounded size of M - needed for some buffers.</span></div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  _Mround = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Msize, strategy::out_height());</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  _Mround *= strategy::out_height();</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  }</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span> </div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  <span class="comment">// Interface implementation - Compulsory functions</span></div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span> </div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="comment">// Window size: Only the last thread should do a ragged block, so dole</span></div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  <span class="comment">// out work in units of out_height. Factor batches into the window, but</span></div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="comment">// not multi for now (as this would cause problems with the buffer</span></div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  <span class="comment">// manager).</span></div><div class="line"><a name="l00384"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a9fe814165d032363ea4f9cfdd36dd5e2"> 384</a></span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a9fe814165d032363ea4f9cfdd36dd5e2">get_window_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  <span class="comment">// _Mround is a multiple of out_height by definition.</span></div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  <span class="keywordflow">return</span> (_Mround / strategy::out_height()) * _nbatches;</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> </div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  <span class="comment">// set_nthreads: pass on to buffer manager to avoid it waiting for non-existant threads.</span></div><div class="line"><a name="l00390"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a468e9c50d4decc2ba86f9bd393ba27d6"> 390</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a468e9c50d4decc2ba86f9bd393ba27d6">set_nthreads</a>(<span class="keywordtype">int</span> nthreads)<span class="keyword"> override </span>{</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  _nthreads = std::min(nthreads, _maxthreads);</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  <span class="keywordflow">if</span> (_bm) {</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#abe8297b7d4cf647e06eff027982cbb8e">set_nthreads</a>(_nthreads);</div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  }</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  }</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">// Execute</span></div><div class="line"><a name="l00398"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0b21ce5febc3f7eed9c7b37c0b9a3560"> 398</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0b21ce5febc3f7eed9c7b37c0b9a3560">execute</a>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> start, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> end, <span class="keywordtype">int</span> threadid)<span class="keyword"> override </span>{</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  <span class="keywordflow">if</span> (_pretransposed) {</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  execute_internal<true>(start, end, threadid);</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  execute_internal<false>(start, end, threadid);</div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  }</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  }</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span> </div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  <span class="comment">// Interface implementation - working space</span></div><div class="line"><a name="l00407"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#abf5a58f6feffeae31f48b750cba9303d"> 407</a></span>  <span class="keywordtype">size_t</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#abf5a58f6feffeae31f48b750cba9303d">get_working_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  <span class="comment">// In all cases, we need one A buffer plus a C buffer per thread.</span></div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <span class="keywordtype">size_t</span> size = get_a_working_size() + (get_c_working_size() * _maxthreads);</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="comment">// For pretransposed case, there is no working space needed for B.</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="comment">// Otherwise, we need a BufferManager.</span></div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  <span class="keywordflow">if</span> (!_pretransposed) {</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  size += <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#ad94ec2bb5758590521fe20ae781ca0b5">BufferManager::get_storage_requirement</a>(_maxthreads, get_b_working_size());</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span>  }</div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span> </div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  size += 64; <span class="comment">// Add on a cache line extra for alignment.</span></div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span> </div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  <span class="keywordflow">return</span> size;</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span>  }</div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span> </div><div class="line"><a name="l00422"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aee006b8384820aa25720e84b91b32a27"> 422</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aee006b8384820aa25720e84b91b32a27">set_working_space</a>(<span class="keywordtype">void</span> *working_space)<span class="keyword"> override </span>{</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  <span class="comment">// Make sure everything ends up cache line aligned</span></div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  int8_t *working_space_bytes = reinterpret_cast<int8_t *>(working_space);</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span>  intptr_t working_space_int = reinterpret_cast<intptr_t>(working_space);</div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span> </div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  <span class="keywordtype">size_t</span> diff=0;</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span> </div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  <span class="keywordflow">if</span> (working_space_int & 0x3F) {</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  diff = 0x40 - (working_space_int & 0x3F);</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  }</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span> </div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span>  working_space_bytes += diff;</div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span> </div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  <span class="keywordflow">if</span> (_pretransposed) {</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  <span class="comment">// Pretransposed case: just set internal pointer to parameter value.</span></div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  _working_space = reinterpret_cast<void *>(working_space_bytes);</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span>  <span class="comment">// Otherwise, use the first part of the working space for the buffer manager.</span></div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  <span class="comment">// It's legal to call this again so don't leak a buffer manager if it already existed.</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  <span class="keyword">delete</span> _bm;</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>  _bm = <span class="keyword">new</span> <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml">BufferManager</a>(_nthreads, get_b_working_size(), reinterpret_cast<void *>(working_space_bytes));</div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span> </div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  working_space_bytes += <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#ad94ec2bb5758590521fe20ae781ca0b5">BufferManager::get_storage_requirement</a>(_maxthreads, get_b_working_size());</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>  _working_space = reinterpret_cast<void *>(working_space_bytes);</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  }</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> </div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="comment">// Interface implementation - pretransposed</span></div><div class="line"><a name="l00452"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ac5dec5063a75604fdc60555f15577b94"> 452</a></span>  <span class="keywordtype">bool</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ac5dec5063a75604fdc60555f15577b94">B_is_pretransposed</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="keywordflow">return</span> _pretransposed;</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  }</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span> </div><div class="line"><a name="l00456"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aa7cfff39cbb2be65cd40042a75e5ae1c"> 456</a></span>  <span class="keywordtype">bool</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aa7cfff39cbb2be65cd40042a75e5ae1c">B_pretranspose_required</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  <span class="keywordflow">return</span> _pretransposed && (_B_transposed==<span class="keyword">nullptr</span>);</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  }</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span> </div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  <span class="comment">// TODO: this could almost certainly be considerably simpler.</span></div><div class="line"><a name="l00461"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae9fc7380e533c43a1d91e48a7578e412"> 461</a></span>  <span class="keywordtype">size_t</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae9fc7380e533c43a1d91e48a7578e412">get_B_pretransposed_array_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  <span class="keywordtype">size_t</span> total=0;</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span> </div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  <span class="keywordflow">do</span> {</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  <span class="comment">/* Figure out the size of each block. */</span></div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x_size = (current.xmax() - current.x0());</div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k_size = (current.kmax() - current.k0());</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span> </div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  <span class="comment">/* Round sizes up as needed. */</span></div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  x_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(x_size, strategy::out_width());</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  x_size *= strategy::out_width();</div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span> </div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  k_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(k_size, strategy::k_unroll());</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  k_size *= strategy::k_unroll();</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>  total += x_size * k_size * <span class="keyword">sizeof</span>(Toi);</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  } <span class="keywordflow">while</span> (current.advance());</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span> </div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  <span class="keywordflow">return</span> total;</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="keyword">using</span> GemmCommon<To, Tr>::pretranspose_B_array;</div><div class="line"><a name="l00484"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110"> 484</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110">pretranspose_B_array</a>(<span class="keywordtype">void</span> *in_buffer, <span class="keyword">const</span> To *B, <span class="keyword">const</span> <span class="keywordtype">int</span> ldb, <span class="keyword">const</span> <span class="keywordtype">int</span> B_multi_stride)<span class="keyword"> override </span>{</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  Toi *buffer = reinterpret_cast<Toi *>(in_buffer);</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  _B_transposed = buffer;</div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span>  strategy strat(_ci);</div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span> </div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  <span class="keywordflow">do</span> {</div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="comment">/* Figure out the size of each block. */</span></div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x_size = (current.xmax() - current.x0());</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k_size = (current.kmax() - current.k0());</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">/* Round sizes up as needed. */</span></div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  x_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(x_size, strategy::out_width());</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span>  x_size *= strategy::out_width();</div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span> </div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  k_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(k_size, strategy::k_unroll());</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  k_size *= strategy::k_unroll();</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span> </div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  strat.transforms.PrepareB(buffer, B + (current.multi() * B_multi_stride), ldb,</div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span>  current.x0(), current.xmax(), current.k0(), current.kmax(), _trB);</div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span> </div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  buffer += (x_size * k_size);</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  } <span class="keywordflow">while</span> (current.advance());</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  }</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span> </div><div class="line"><a name="l00509"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac"> 509</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac">set_pretransposed_B_data</a>(<span class="keywordtype">void</span> *in_buffer)<span class="keyword"> override </span>{</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  _B_transposed = reinterpret_cast<Toi *>(in_buffer);</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  }</div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span> </div><div class="line"><a name="l00513"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5"> 513</a></span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5">~GemmInterleaved</a>()<span class="keyword"> override </span>{</div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  <span class="keyword">delete</span> _bm;</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>  }</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span> };</div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span> </div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span> } <span class="comment">// namespace arm_gemm</span></div><div class="ttc" id="utils_8hpp_xhtml_aa4508679e1d089c2bdcf000f72357683"><div class="ttname"><a href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a></div><div class="ttdeci">T iceildiv(const T a, const T b)</div><div class="ttdef"><b>Definition:</b> <a href="utils_8hpp_source.xhtml#l00036">utils.hpp:36</a></div></div>
+<a href="gemm__interleaved_8hpp.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) 2017-2019 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">#pragma once</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> </div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#include <stdio.h></span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="preprocessor">#include <assert.h></span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> </div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#include <algorithm></span></div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> </div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#include "arm_gemm.hpp"</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#include "<a class="code" href="utils_8hpp.xhtml">utils.hpp</a>"</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> </div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="preprocessor">#include "<a class="code" href="buffer__manager_8hpp.xhtml">buffer_manager.hpp</a>"</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="preprocessor">#include "<a class="code" href="mergeresults_8hpp.xhtml">mergeresults.hpp</a>"</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="preprocessor">#include "<a class="code" href="transform_8hpp.xhtml">transform.hpp</a>"</span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> </div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="preprocessor">#include "<a class="code" href="profiler_8hpp.xhtml">profiler.hpp</a>"</span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> </div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="comment">// Some macros used to decide how much working space to allocate.</span></div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="comment">// Round allocations up to the next cache line.</span></div><div class="line"><a name="l00044"></a><span class="lineno"><a class="line" href="gemm__interleaved_8hpp.xhtml#abd1ded15e3688096ead983e87f83fbd1"> 44</a></span> <span class="preprocessor">#define ALLOC_ROUND 64</span></div><div class="line"><a name="l00045"></a><span class="lineno"><a class="line" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af"> 45</a></span> <span class="preprocessor">#define ROUND_UP(x) ((((x) + ALLOC_ROUND-1) / ALLOC_ROUND) * ALLOC_ROUND)</span></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">// Implementation of the GemmCommon abstract class.</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> <span class="comment">//</span></div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="comment">// This implementation interleaves the source matrices in blocks - good for</span></div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="comment">// larger matrices.</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="keyword">namespace </span><a class="code" href="namespacearm__gemm.xhtml">arm_gemm</a> {</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> </div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span> <span class="keyword">template</span><<span class="keyword">typename</span> strategy, <span class="keyword">typename</span> To, <span class="keyword">typename</span> Tr></div><div class="line"><a name="l00054"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml"> 54</a></span> <span class="keyword">class </span><a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> : <span class="keyword">public</span> GemmCommon<To, Tr> {</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keyword">typedef</span> <span class="keyword">typename</span> strategy::operand_type Toi;</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="keyword">typedef</span> <span class="keyword">typename</span> strategy::result_type Tri;</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> </div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  <span class="comment">/* const properties set by constructor */</span></div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  <span class="keyword">const</span> CPUInfo * <span class="keyword">const</span> _ci;</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> </div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Msize;</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Nsize;</div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Ksize;</div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span> </div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _nbatches;</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _nmulti;</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span> </div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _trA;</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _trB;</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> </div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> Tr _alpha;</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  <span class="keyword">const</span> Tr _beta;</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> _maxthreads;</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  <span class="keywordtype">int</span> _nthreads;</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  <span class="keyword">const</span> <span class="keywordtype">bool</span> _pretransposed;</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> </div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="comment">/* Blocking info */</span></div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _k_block=0;</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _x_block=0;</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _Mround=0;</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span> </div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  <span class="comment">/* Working space, pretransposed buffer, buffer manager */</span></div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  <span class="keyword">const</span> Toi *_B_transposed=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml">BufferManager</a> *_bm=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="keywordtype">void</span> *_working_space=<span class="keyword">nullptr</span>;</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span> </div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  <span class="comment">/* We will need to walk through the blocks of B in a few contexts, so</span></div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> <span class="comment"> * factor that out. */</span></div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  <span class="keyword">class </span>blockwalker {</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="keyword">private</span>:</div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  <span class="comment">/* Size loops, etc. based on our parent's configuration */</span></div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  <span class="keyword">const</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved<strategy, To, Tr></a> &_parent;</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span> </div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="comment">/* K, X and multi parameters for current iteration. */</span></div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _k0=0, _x0=0, _multi=0;</div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> </div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> _index=0;</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  <span class="keywordtype">bool</span> _done=<span class="keyword">false</span>;</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  <span class="keywordtype">bool</span> _newkblock=<span class="keyword">true</span>;</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  <span class="keywordtype">bool</span> _newmulti=<span class="keyword">true</span>;</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span> </div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  <span class="keyword">public</span>:</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  blockwalker(<span class="keyword">const</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved<strategy, To, Tr></a> &parent) : _parent(parent) { }</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span> </div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> xmax() {</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keywordflow">return</span> std::min(_x0 + _parent._x_block, _parent._Nsize);</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  }</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span> </div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> kmax() {</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  <span class="keywordflow">return</span> std::min(_k0 + _parent._k_block, _parent._Ksize);</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  }</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">/* Advance to the next block, return false at the end. */</span></div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  <span class="keywordtype">bool</span> advance(<span class="keywordtype">void</span>) {</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  <span class="keywordflow">if</span> (_done) {</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  <span class="keywordflow">return</span> <span class="keyword">false</span>;</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> </div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  _newkblock=<span class="keyword">false</span>;</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  _x0 += _parent._x_block;</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  <span class="keywordflow">if</span> (_x0 >= _parent._Nsize) {</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  _x0=0;</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  _k0 += _parent._k_block;</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="keywordflow">if</span> (_k0 >= _parent._Ksize) {</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  _k0=0;</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  _multi++;</div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="keywordflow">if</span> (_multi >= _parent._nmulti) {</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  _done=<span class="keyword">true</span>;</div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordflow">return</span> <span class="keyword">false</span>;</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  }</div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  _newmulti=<span class="keyword">true</span>;</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  }</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  _newkblock=<span class="keyword">true</span>;</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  }</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  _index++;</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span> </div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <span class="keywordflow">return</span> <span class="keyword">true</span>;</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> </div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k0(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _k0; }</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x0(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _x0; }</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> multi(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _multi; }</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> index(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _index; }</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="keywordtype">bool</span> done(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _done; }</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="keywordtype">bool</span> newkblock(<span class="keywordtype">void</span>) { <span class="keywordflow">return</span> _newkblock; }</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  };</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span> </div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  <span class="comment">// A working size: One of these needed, regardless of thread count. Divided according to window.</span></div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  <span class="keywordtype">size_t</span> get_a_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Toi) * _k_block * _Mround * _nbatches);</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> </div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="comment">// B working size: 0, 1 or 3 of these needed depending on pretransposed and threading settings.</span></div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="keywordtype">size_t</span> get_b_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Toi) * _x_block * _k_block);</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  }</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>  <span class="comment">// C working size: One needed per thread.</span></div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  <span class="keywordtype">size_t</span> get_c_working_size()<span class="keyword"> const </span>{</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="keywordflow">return</span> <a class="code" href="gemm__interleaved_8hpp.xhtml#a5560b3c73e2e50a0b1bda1fa70ec58af">ROUND_UP</a>(<span class="keyword">sizeof</span>(Tri) * _x_block * strategy::out_height());</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  }</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span> </div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  <span class="comment">// Internal execute function.</span></div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// This supports both the "pretransposed" and "standard" interfaces via the template parameter.</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="keyword">template</span><<span class="keywordtype">bool</span> pretransposed></div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="keywordtype">void</span> execute_internal(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> start, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> end, <span class="keywordtype">int</span> threadid) {</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  profiler prof;</div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  strategy strat(_ci);</div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span> </div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  blockwalker next=current;</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span> </div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  <span class="comment">/* Translate 'start' and 'end' into a position within the batches and rows. */</span></div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> window_per_batch = _Mround / strategy::out_height();</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch_0 = start / window_per_batch;</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch_end = end / window_per_batch;</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 the M values to operate on */</span></div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> m_0 = (start - (batch_0 * window_per_batch)) * strategy::out_height();</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> m_max = (end - (batch_end * window_per_batch)) * strategy::out_height();</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">/* Make sure we've been set up correctly. */</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  assert(_B_transposed);</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  assert(_bm);</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> </div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  assert(_working_space);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  int8_t *working_space_bytes = reinterpret_cast<int8_t *>(_working_space);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span> </div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  <span class="comment">// Private buffers. Treat working_space as an array of C buffers</span></div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <span class="comment">// (one per thread) first, followed by the (window-divided) A</span></div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  <span class="comment">// buffer.</span></div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  <span class="comment">// Set a_panel to the base of the A buffers - compute offsets into it based on M/batches later.</span></div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  Toi * <span class="keyword">const</span> a_panel = reinterpret_cast<Toi *>(working_space_bytes + (_maxthreads * get_c_working_size()));</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  Tri * <span class="keyword">const</span> c_panel = reinterpret_cast<Tri *>(working_space_bytes + (threadid * get_c_working_size()));</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span> </div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  <span class="comment">// Shared buffers - these come either from BufferManager or _B_transposed.</span></div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  <span class="keyword">const</span> Toi *b_panel;</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span> </div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  b_panel = _B_transposed;</div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  }</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span> </div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="comment">//printf("Starting GEMM loop, x_block=%d, k_block=%d\n", _x_block, _k_block);</span></div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span> </div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  <span class="comment">// newkblock() is always true on the first iteration, so this will be set properly on the first loop.</span></div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="keywordtype">int</span> kern_k = 0;</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span> </div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="keywordflow">for</span> (;!current.done();current.advance()) {</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  <span class="keywordflow">if</span> (current.newkblock()) {</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#ac19bf2c5c388b9fb21cd0668475b2cf7">PROFILE_PREPA</a>, (end - start) * strategy::out_height() * (current.kmax()-current.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = batch_0; batch <= batch_end; batch++) {</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_m = (batch == batch_0) ? m_0 : 0;</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> last_m = (batch == batch_end) ? m_max : _Msize;</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span> </div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <span class="keywordflow">if</span> (first_m >= last_m)</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  <span class="keywordflow">continue</span>;</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>  strat.transforms.PrepareA(a_panel + ((batch * _Mround + first_m) * _k_block),</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  this->_Aptr + (batch * this->_A_batch_stride) + (current.multi() * this->_A_multi_stride),</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  this->_lda, first_m, last_m, current.k0(), current.kmax(), _trA);</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  }</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span> </div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  <span class="comment">// Figure out how many "K" the kernel will actually process.</span></div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  kern_k = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(current.kmax() - current.k0(), strategy::k_unroll());</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  kern_k *= strat.k_unroll();</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  }</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>  <span class="keywordtype">int</span> bblocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(current.xmax() - current.x0(), strategy::out_width());</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span> </div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  <span class="keywordflow">if</span> (!pretransposed) {</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  <span class="comment">/* Look ahead to the next block and populate it if necessary.</span></div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span> <span class="comment"> * This avoids the populate operation becoming a bottleneck, and</span></div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> <span class="comment"> * helps keep the threads synchronized (the first thread to get</span></div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span> <span class="comment"> * here will populate while the rest will advance).</span></div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span> <span class="comment"> *</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span> <span class="comment"> * If we are running single threaded, bm->try_populate() will do</span></div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span> <span class="comment"> * nothing.</span></div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span> <span class="comment"> */</span></div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  <span class="keywordflow">if</span> (next.advance()) {</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a97814ef98fcd29d1fb5df1e877f8eabc">try_populate</a>(next.index(), [&](<span class="keywordtype">void</span> *buffer) {</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#af0477dfaa86fc08825cb3f31212f423c">PROFILE_PREPB</a>, (next.xmax()-next.x0()) * (next.kmax()-next.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span> </div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  Toi *b_panel = reinterpret_cast<Toi *>(buffer);</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span> </div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  strat.transforms.PrepareB(b_panel, this->_Bptr + (next.multi() * this->_B_multi_stride), this->_ldb,</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  next.x0(), next.xmax(), next.k0(), next.kmax(), _trB);</div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  });</div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  }</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span> </div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="comment">/* Get the buffer for this iteration from the BufferManager. */</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  b_panel = reinterpret_cast<Toi *>(_bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a427705bf76071f12aa29992bd5cffa73">get</a>(current.index(), [&](<span class="keywordtype">void</span> *bpv) {</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#af0477dfaa86fc08825cb3f31212f423c">PROFILE_PREPB</a>, (current.xmax()-current.x0()) * (current.kmax()-current.k0()) * <span class="keyword">sizeof</span>(Toi));</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span> <span class="preprocessor">#endif</span></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>  Toi *b_panel = reinterpret_cast<Toi *>(bpv);</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span> </div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  strat.transforms.PrepareB(b_panel, this->_Bptr + (current.multi() * this->_B_multi_stride), this->_ldb,</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  current.x0(), current.xmax(), current.k0(), current.kmax(), _trB);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  }));</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  }</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">/* Do the actual work. */</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> batch = batch_0; batch <= batch_end; batch++) {</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_m = (batch == batch_0) ? m_0 : 0;</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> last_m = (batch == batch_end) ? m_max : _Msize;</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span> </div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  <span class="keyword">const</span> Toi *a_ptr = a_panel + (batch * _Mround + first_m) * _k_block;</div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span> </div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  <span class="keywordflow">if</span> (first_m >= last_m)</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  <span class="keywordflow">continue</span>;</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span> </div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> y=first_m; y<last_m; y+=strategy::out_height()) {</div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> ymax = std::min(_Msize, y + strategy::out_height());</div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span> </div><div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  {</div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#a97f4e398174b5e094f5262e5d785ac20">PROFILE_KERNEL</a>, (strategy::out_height() * bblocks * strategy::out_width() * kern_k));</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span> </div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  strat.kernel(a_ptr, b_panel, c_panel, 1, bblocks, kern_k);</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span> </div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  a_ptr += (strategy::out_height() * kern_k);</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  }</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span> </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> <span class="preprocessor">#ifdef CYCLE_PROFILING</span></div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  <span class="keyword">auto</span> p=prof.ScopedProfiler(<a class="code" href="profiler_8hpp.xhtml#a2f7c3e65a9ccab509d7776832f562dd0">PROFILE_MERGE</a>, (strategy::out_height() * bblocks * strategy::out_width() * <span class="keyword">sizeof</span>(Tr)));</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span> <span class="preprocessor">#endif</span></div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  strat.transforms.Merge(this->_Cptr + (batch * this->_C_batch_stride) + (current.multi() * this->_C_multi_stride),</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  c_panel, this->_ldc, y, ymax, current.x0(), current.xmax(),</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  _alpha, (current.k0()==0 ? _beta : static_cast<Tr>(1)));</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>  }</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> </div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  <span class="keywordflow">if</span> (pretransposed) {</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  b_panel += (bblocks * strat.out_width() * kern_k);</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#a29a2894a4403bc4cc19aacc35cb91d54">release</a>(current.index());</div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  }</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  }</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  }</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span> </div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span> <span class="keyword">public</span>:</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#add13963d432a1937370721fa61755b9c">GemmInterleaved</a>(<a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> &) = <span class="keyword">delete</span>;</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> & <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#af03b8e4289fa0e112d315dd7e25ec696">operator= </a>(<a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml">GemmInterleaved</a> &) = <span class="keyword">delete</span>;</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span> </div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  <span class="comment">/* Constructor */</span></div><div class="line"><a name="l00320"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aec95ac91ca276b1ad8baf91520c0306f"> 320</a></span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aec95ac91ca276b1ad8baf91520c0306f">GemmInterleaved</a>(<span class="keyword">const</span> GemmArgs<Tr> &args)</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  : _ci(args._ci), _Msize(args._Msize), _Nsize(args._Nsize), _Ksize(args._Ksize),</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  _nbatches(args._nbatches), _nmulti(args._nmulti), _trA(args._trA), _trB(args._trB),</div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  _alpha(args._alpha), _beta(args._beta), _maxthreads(args._maxthreads), _nthreads(args._maxthreads),</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  _pretransposed(args._pretransposed_hint) {</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> L1_size = _ci->get_L1_cache_size();</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> L2_size = _ci->get_L2_cache_size();</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span> </div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  assert(_maxthreads > 0);</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span> </div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>  <span class="comment">// Work out blocking parameters, or override from provided GemmConfig</span></div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  <span class="keywordflow">if</span> (args._cfg && args._cfg->inner_block_size) {</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  _k_block = args._cfg->inner_block_size;</div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  <span class="comment">// k_block: Find out how much of the larger array can be loaded into half the cache.</span></div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  <span class="comment">// This should account for associative caches.</span></div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  _k_block = (L1_size / 2) / (<span class="keyword">sizeof</span>(Toi) * (std::max(strategy::out_width(), strategy::out_height())));</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span> </div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <span class="comment">// Needs to be (at least a single) multiple of the K unroll level.</span></div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  _k_block /= strategy::k_unroll();</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  _k_block = std::max(_k_block, 1U) * strategy::k_unroll();</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span> </div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  <span class="comment">// Now tune to presented problem size; this is how many blocks we need.</span></div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_k_blocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Ksize, _k_block);</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span> </div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <span class="comment">// So divide the space equally into that many blocks.</span></div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  _k_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Ksize, num_k_blocks);</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span> </div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="comment">// And round UP to the K unroll level required.</span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  _k_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_k_block, strategy::k_unroll());</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  _k_block *= strategy::k_unroll();</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  }</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span> </div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <span class="keywordflow">if</span> (args._cfg && args._cfg->outer_block_size) {</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  _x_block = args._cfg->outer_block_size;</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  <span class="comment">// x_block: Work out how many rows (of length k_block) will fit in the L2</span></div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="comment">// Don't allocate more than 90% of the L2 to allow for overheads, and subtract off the L1 contents.</span></div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  _x_block = (((L2_size * 9) / 10) - (_k_block * <span class="keyword">sizeof</span>(Toi) * (strategy::out_width() + strategy::out_height()))) /</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  (<span class="keyword">sizeof</span>(Toi) * _k_block);</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span> </div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  <span class="comment">// Needs to be (at least a single) multiple of the kernel output width.</span></div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  _x_block /= strategy::out_width();</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  _x_block = std::max(_x_block, 1U) * strategy::out_width();</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">// And tune to the presented problem size.</span></div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_x_blocks = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Nsize, _x_block);</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  _x_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Nsize, num_x_blocks);</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>  _x_block = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_x_block, strategy::out_width());</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  _x_block *= strategy::out_width();</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  }</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span> </div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  <span class="comment">// Work out the rounded size of M - needed for some buffers.</span></div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  _Mround = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(_Msize, strategy::out_height());</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  _Mround *= strategy::out_height();</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  }</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span> </div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  <span class="comment">// Interface implementation - Compulsory functions</span></div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span> </div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="comment">// Window size: Only the last thread should do a ragged block, so dole</span></div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  <span class="comment">// out work in units of out_height. Factor batches into the window, but</span></div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="comment">// not multi for now (as this would cause problems with the buffer</span></div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  <span class="comment">// manager).</span></div><div class="line"><a name="l00384"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a9fe814165d032363ea4f9cfdd36dd5e2"> 384</a></span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a9fe814165d032363ea4f9cfdd36dd5e2">get_window_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  <span class="comment">// _Mround is a multiple of out_height by definition.</span></div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  <span class="keywordflow">return</span> (_Mround / strategy::out_height()) * _nbatches;</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> </div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  <span class="comment">// set_nthreads: pass on to buffer manager to avoid it waiting for non-existant threads.</span></div><div class="line"><a name="l00390"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a468e9c50d4decc2ba86f9bd393ba27d6"> 390</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a468e9c50d4decc2ba86f9bd393ba27d6">set_nthreads</a>(<span class="keywordtype">int</span> nthreads)<span class="keyword"> override </span>{</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  _nthreads = std::min(nthreads, _maxthreads);</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  <span class="keywordflow">if</span> (_bm) {</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  _bm-><a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#abe8297b7d4cf647e06eff027982cbb8e">set_nthreads</a>(_nthreads);</div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  }</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  }</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">// Execute</span></div><div class="line"><a name="l00398"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0b21ce5febc3f7eed9c7b37c0b9a3560"> 398</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0b21ce5febc3f7eed9c7b37c0b9a3560">execute</a>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> start, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> end, <span class="keywordtype">int</span> threadid)<span class="keyword"> override </span>{</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  <span class="keywordflow">if</span> (_pretransposed) {</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  execute_internal<true>(start, end, threadid);</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  execute_internal<false>(start, end, threadid);</div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  }</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  }</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span> </div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  <span class="comment">// Interface implementation - working space</span></div><div class="line"><a name="l00407"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#abf5a58f6feffeae31f48b750cba9303d"> 407</a></span>  <span class="keywordtype">size_t</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#abf5a58f6feffeae31f48b750cba9303d">get_working_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  <span class="comment">// In all cases, we need one A buffer plus a C buffer per thread.</span></div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <span class="keywordtype">size_t</span> size = get_a_working_size() + (get_c_working_size() * _maxthreads);</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="comment">// For pretransposed case, there is no working space needed for B.</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="comment">// Otherwise, we need a BufferManager.</span></div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  <span class="keywordflow">if</span> (!_pretransposed) {</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  size += <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#ad94ec2bb5758590521fe20ae781ca0b5">BufferManager::get_storage_requirement</a>(_maxthreads, get_b_working_size());</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span>  }</div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span> </div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  size += 64; <span class="comment">// Add on a cache line extra for alignment.</span></div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span> </div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  <span class="keywordflow">return</span> size;</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span>  }</div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span> </div><div class="line"><a name="l00422"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aee006b8384820aa25720e84b91b32a27"> 422</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aee006b8384820aa25720e84b91b32a27">set_working_space</a>(<span class="keywordtype">void</span> *working_space)<span class="keyword"> override </span>{</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  <span class="comment">// Make sure everything ends up cache line aligned</span></div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  int8_t *working_space_bytes = reinterpret_cast<int8_t *>(working_space);</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span>  intptr_t working_space_int = reinterpret_cast<intptr_t>(working_space);</div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span> </div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  <span class="keywordtype">size_t</span> diff=0;</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span> </div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  <span class="keywordflow">if</span> (working_space_int & 0x3F) {</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  diff = 0x40 - (working_space_int & 0x3F);</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  }</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span> </div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span>  working_space_bytes += diff;</div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span> </div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  <span class="keywordflow">if</span> (_pretransposed) {</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  <span class="comment">// Pretransposed case: just set internal pointer to parameter value.</span></div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  _working_space = reinterpret_cast<void *>(working_space_bytes);</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span>  <span class="comment">// Otherwise, use the first part of the working space for the buffer manager.</span></div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  <span class="comment">// It's legal to call this again so don't leak a buffer manager if it already existed.</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  <span class="keyword">delete</span> _bm;</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>  _bm = <span class="keyword">new</span> <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml">BufferManager</a>(_nthreads, get_b_working_size(), reinterpret_cast<void *>(working_space_bytes));</div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span> </div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  working_space_bytes += <a class="code" href="classarm__gemm_1_1_buffer_manager.xhtml#ad94ec2bb5758590521fe20ae781ca0b5">BufferManager::get_storage_requirement</a>(_maxthreads, get_b_working_size());</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>  _working_space = reinterpret_cast<void *>(working_space_bytes);</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  }</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> </div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="comment">// Interface implementation - pretransposed</span></div><div class="line"><a name="l00452"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ac5dec5063a75604fdc60555f15577b94"> 452</a></span>  <span class="keywordtype">bool</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ac5dec5063a75604fdc60555f15577b94">B_is_pretransposed</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="keywordflow">return</span> _pretransposed;</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  }</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span> </div><div class="line"><a name="l00456"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aa7cfff39cbb2be65cd40042a75e5ae1c"> 456</a></span>  <span class="keywordtype">bool</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#aa7cfff39cbb2be65cd40042a75e5ae1c">B_pretranspose_required</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  <span class="keywordflow">return</span> _pretransposed && (_B_transposed==<span class="keyword">nullptr</span>);</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  }</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span> </div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  <span class="comment">// TODO: this could almost certainly be considerably simpler.</span></div><div class="line"><a name="l00461"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae9fc7380e533c43a1d91e48a7578e412"> 461</a></span>  <span class="keywordtype">size_t</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae9fc7380e533c43a1d91e48a7578e412">get_B_pretransposed_array_size</a>()<span class="keyword"> const override </span>{</div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  <span class="keywordtype">size_t</span> total=0;</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span> </div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  <span class="keywordflow">do</span> {</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  <span class="comment">/* Figure out the size of each block. */</span></div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x_size = (current.xmax() - current.x0());</div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k_size = (current.kmax() - current.k0());</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span> </div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  <span class="comment">/* Round sizes up as needed. */</span></div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  x_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(x_size, strategy::out_width());</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  x_size *= strategy::out_width();</div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span> </div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  k_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(k_size, strategy::k_unroll());</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  k_size *= strategy::k_unroll();</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>  total += x_size * k_size * <span class="keyword">sizeof</span>(Toi);</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  } <span class="keywordflow">while</span> (current.advance());</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span> </div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  <span class="keywordflow">return</span> total;</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"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110"> 483</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110">pretranspose_B_array</a>(<span class="keywordtype">void</span> *in_buffer, <span class="keyword">const</span> To *B, <span class="keyword">const</span> <span class="keywordtype">int</span> ldb, <span class="keyword">const</span> <span class="keywordtype">int</span> B_multi_stride)<span class="keyword"> override </span>{</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  blockwalker current(*<span class="keyword">this</span>);</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  Toi *buffer = reinterpret_cast<Toi *>(in_buffer);</div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  _B_transposed = buffer;</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  strategy strat(_ci);</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>  <span class="keywordflow">do</span> {</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  <span class="comment">/* Figure out the size of each block. */</span></div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> x_size = (current.xmax() - current.x0());</div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> k_size = (current.kmax() - current.k0());</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>  <span class="comment">/* Round sizes up as needed. */</span></div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span>  x_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(x_size, strategy::out_width());</div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  x_size *= strategy::out_width();</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> </div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  k_size = <a class="code" href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a>(k_size, strategy::k_unroll());</div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  k_size *= strategy::k_unroll();</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span> </div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  strat.transforms.PrepareB(buffer, B + (current.multi() * B_multi_stride), ldb,</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  current.x0(), current.xmax(), current.k0(), current.kmax(), _trB);</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>  buffer += (x_size * k_size);</div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  } <span class="keywordflow">while</span> (current.advance());</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  }</div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span> </div><div class="line"><a name="l00508"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac"> 508</a></span>  <span class="keywordtype">void</span> <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac">set_pretransposed_B_data</a>(<span class="keywordtype">void</span> *in_buffer)<span class="keyword"> override </span>{</div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span>  _B_transposed = reinterpret_cast<Toi *>(in_buffer);</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  }</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span> </div><div class="line"><a name="l00512"></a><span class="lineno"><a class="line" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5"> 512</a></span>  <a class="code" href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5">~GemmInterleaved</a>()<span class="keyword"> override </span>{</div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  <span class="keyword">delete</span> _bm;</div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  }</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span> };</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span> </div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span> } <span class="comment">// namespace arm_gemm</span></div><div class="ttc" id="utils_8hpp_xhtml_aa4508679e1d089c2bdcf000f72357683"><div class="ttname"><a href="utils_8hpp.xhtml#aa4508679e1d089c2bdcf000f72357683">iceildiv</a></div><div class="ttdeci">T iceildiv(const T a, const T b)</div><div class="ttdef"><b>Definition:</b> <a href="utils_8hpp_source.xhtml#l00036">utils.hpp:36</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_buffer_manager_xhtml_ad94ec2bb5758590521fe20ae781ca0b5"><div class="ttname"><a href="classarm__gemm_1_1_buffer_manager.xhtml#ad94ec2bb5758590521fe20ae781ca0b5">arm_gemm::BufferManager::get_storage_requirement</a></div><div class="ttdeci">static size_t get_storage_requirement(const int maxthreads, const size_t buffersize)</div><div class="ttdef"><b>Definition:</b> <a href="buffer__manager_8hpp_source.xhtml#l00218">buffer_manager.hpp:218</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml">arm_gemm::GemmInterleaved</a></div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00054">gemm_interleaved.hpp:54</a></div></div>
<div class="ttc" id="profiler_8hpp_xhtml_af0477dfaa86fc08825cb3f31212f423c"><div class="ttname"><a href="profiler_8hpp.xhtml#af0477dfaa86fc08825cb3f31212f423c">PROFILE_PREPB</a></div><div class="ttdeci">#define PROFILE_PREPB</div><div class="ttdef"><b>Definition:</b> <a href="profiler_8hpp_source.xhtml#l00135">profiler.hpp:135</a></div></div>
@@ -114,8 +114,8 @@
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_ae9fc7380e533c43a1d91e48a7578e412"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae9fc7380e533c43a1d91e48a7578e412">arm_gemm::GemmInterleaved::get_B_pretransposed_array_size</a></div><div class="ttdeci">size_t get_B_pretransposed_array_size() const override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00461">gemm_interleaved.hpp:461</a></div></div>
<div class="ttc" id="transform_8hpp_xhtml"><div class="ttname"><a href="transform_8hpp.xhtml">transform.hpp</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_abf5a58f6feffeae31f48b750cba9303d"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#abf5a58f6feffeae31f48b750cba9303d">arm_gemm::GemmInterleaved::get_working_size</a></div><div class="ttdeci">size_t get_working_size() const override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00407">gemm_interleaved.hpp:407</a></div></div>
-<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_ae385dce1a98e03f0a3ca311f031e5110"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110">arm_gemm::GemmInterleaved::pretranspose_B_array</a></div><div class="ttdeci">void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00484">gemm_interleaved.hpp:484</a></div></div>
-<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_a6e76d099f51fc0ef7f48dbd7d00214ac"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac">arm_gemm::GemmInterleaved::set_pretransposed_B_data</a></div><div class="ttdeci">void set_pretransposed_B_data(void *in_buffer) override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00509">gemm_interleaved.hpp:509</a></div></div>
+<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_ae385dce1a98e03f0a3ca311f031e5110"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#ae385dce1a98e03f0a3ca311f031e5110">arm_gemm::GemmInterleaved::pretranspose_B_array</a></div><div class="ttdeci">void pretranspose_B_array(void *in_buffer, const To *B, const int ldb, const int B_multi_stride) override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00483">gemm_interleaved.hpp:483</a></div></div>
+<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_a6e76d099f51fc0ef7f48dbd7d00214ac"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#a6e76d099f51fc0ef7f48dbd7d00214ac">arm_gemm::GemmInterleaved::set_pretransposed_B_data</a></div><div class="ttdeci">void set_pretransposed_B_data(void *in_buffer) override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00508">gemm_interleaved.hpp:508</a></div></div>
<div class="ttc" id="profiler_8hpp_xhtml_a97f4e398174b5e094f5262e5d785ac20"><div class="ttname"><a href="profiler_8hpp.xhtml#a97f4e398174b5e094f5262e5d785ac20">PROFILE_KERNEL</a></div><div class="ttdeci">#define PROFILE_KERNEL</div><div class="ttdef"><b>Definition:</b> <a href="profiler_8hpp_source.xhtml#l00136">profiler.hpp:136</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_buffer_manager_xhtml_a97814ef98fcd29d1fb5df1e877f8eabc"><div class="ttname"><a href="classarm__gemm_1_1_buffer_manager.xhtml#a97814ef98fcd29d1fb5df1e877f8eabc">arm_gemm::BufferManager::try_populate</a></div><div class="ttdeci">void try_populate(const int index, T func)</div><div class="ttdef"><b>Definition:</b> <a href="buffer__manager_8hpp_source.xhtml#l00260">buffer_manager.hpp:260</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_a9fe814165d032363ea4f9cfdd36dd5e2"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#a9fe814165d032363ea4f9cfdd36dd5e2">arm_gemm::GemmInterleaved::get_window_size</a></div><div class="ttdeci">unsigned int get_window_size() const override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00384">gemm_interleaved.hpp:384</a></div></div>
@@ -130,7 +130,7 @@
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_aee006b8384820aa25720e84b91b32a27"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#aee006b8384820aa25720e84b91b32a27">arm_gemm::GemmInterleaved::set_working_space</a></div><div class="ttdeci">void set_working_space(void *working_space) override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00422">gemm_interleaved.hpp:422</a></div></div>
<div class="ttc" id="profiler_8hpp_xhtml_a2f7c3e65a9ccab509d7776832f562dd0"><div class="ttname"><a href="profiler_8hpp.xhtml#a2f7c3e65a9ccab509d7776832f562dd0">PROFILE_MERGE</a></div><div class="ttdeci">#define PROFILE_MERGE</div><div class="ttdef"><b>Definition:</b> <a href="profiler_8hpp_source.xhtml#l00137">profiler.hpp:137</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_buffer_manager_xhtml"><div class="ttname"><a href="classarm__gemm_1_1_buffer_manager.xhtml">arm_gemm::BufferManager</a></div><div class="ttdef"><b>Definition:</b> <a href="buffer__manager_8hpp_source.xhtml#l00205">buffer_manager.hpp:205</a></div></div>
-<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_a0ca149c067af4f39661ecce72734e0c5"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5">arm_gemm::GemmInterleaved::~GemmInterleaved</a></div><div class="ttdeci">~GemmInterleaved() override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00513">gemm_interleaved.hpp:513</a></div></div>
+<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_a0ca149c067af4f39661ecce72734e0c5"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#a0ca149c067af4f39661ecce72734e0c5">arm_gemm::GemmInterleaved::~GemmInterleaved</a></div><div class="ttdeci">~GemmInterleaved() override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00512">gemm_interleaved.hpp:512</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_ac5dec5063a75604fdc60555f15577b94"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#ac5dec5063a75604fdc60555f15577b94">arm_gemm::GemmInterleaved::B_is_pretransposed</a></div><div class="ttdeci">bool B_is_pretransposed() const override</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00452">gemm_interleaved.hpp:452</a></div></div>
<div class="ttc" id="classarm__gemm_1_1_gemm_interleaved_xhtml_aec95ac91ca276b1ad8baf91520c0306f"><div class="ttname"><a href="classarm__gemm_1_1_gemm_interleaved.xhtml#aec95ac91ca276b1ad8baf91520c0306f">arm_gemm::GemmInterleaved::GemmInterleaved</a></div><div class="ttdeci">GemmInterleaved(const GemmArgs< Tr > &args)</div><div class="ttdef"><b>Definition:</b> <a href="gemm__interleaved_8hpp_source.xhtml#l00320">gemm_interleaved.hpp:320</a></div></div>
</div><!-- fragment --></div><!-- contents -->
@@ -139,7 +139,7 @@
<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_d91e1966cc57df4f49d1e07be7697713.xhtml">NEON</a></li><li class="navelem"><a class="el" href="dir_1ce938f580a6f304b99685e2c1dd19db.xhtml">kernels</a></li><li class="navelem"><a class="el" href="dir_9847dd9cb7a28e46201717fa7336cf62.xhtml">arm_gemm</a></li><li class="navelem"><a class="el" href="gemm__interleaved_8hpp.xhtml">gemm_interleaved.hpp</a></li>
- <li class="footer">Generated on Thu Feb 28 2019 12:24:50 for Compute Library by
+ <li class="footer">Generated on Thu May 23 2019 17:11:20 for Compute Library by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.15 </li>
</ul>