| <a href="_n_e_convolution_kernel_8cpp.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) 2016-2019 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">#include "<a class="code" href="_n_e_convolution_kernel_8h.xhtml">arm_compute/core/NEON/kernels/NEConvolutionKernel.h</a>"</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> </div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#include "<a class="code" href="_coordinates_8h.xhtml">arm_compute/core/Coordinates.h</a>"</span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="preprocessor">#include "<a class="code" href="_error_8h.xhtml">arm_compute/core/Error.h</a>"</span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#include "<a class="code" href="arm__compute_2core_2_helpers_8h.xhtml">arm_compute/core/Helpers.h</a>"</span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#include "<a class="code" href="_i_tensor_8h.xhtml">arm_compute/core/ITensor.h</a>"</span></div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="preprocessor">#include "<a class="code" href="_tensor_info_8h.xhtml">arm_compute/core/TensorInfo.h</a>"</span></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#include "<a class="code" href="arm__compute_2core_2_types_8h.xhtml">arm_compute/core/Types.h</a>"</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#include "<a class="code" href="arm__compute_2core_2_utils_8h.xhtml">arm_compute/core/Utils.h</a>"</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> <span class="preprocessor">#include "<a class="code" href="_validate_8h.xhtml">arm_compute/core/Validate.h</a>"</span></div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="preprocessor">#include "<a class="code" href="_window_8h.xhtml">arm_compute/core/Window.h</a>"</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> </div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="preprocessor">#include <algorithm></span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> <span class="preprocessor">#include <arm_neon.h></span></div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="preprocessor">#include <array></span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="preprocessor">#include <cstdint></span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="preprocessor">#include <cstring></span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> <span class="preprocessor">#include <tuple></span></div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> </div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="keyword">namespace </span><a class="code" href="namespacearm__compute.xhtml">arm_compute</a></div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span> {</div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span> <span class="keyword">namespace</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="keyword">const</span> uint16x8_t max_int16 = vdupq_n_u16(INT16_MAX);</div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> </div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> int32x4_t &out, <span class="keyword">const</span> int32x4_t &out2, int16_t *output)</div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> {</div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  <span class="keyword">const</span> int16x8_t s16results = vcombine_s16(vqmovn_s32(out),</div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  vqmovn_s32(out2));</div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  vst1q_s16(output, s16results);</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> }</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span> </div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> int32x4_t &out, <span class="keyword">const</span> int32x4_t &out2, uint8_t *output)</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="keyword">const</span> uint8x8_t u8results = vqmovn_u16(vcombine_u16(vqmovun_s32(out),</div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  vqmovun_s32(out2)));</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span>  vst1_u8(output, u8results);</div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span> }</div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> </div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> uint32x4_t &out, <span class="keyword">const</span> uint32x4_t &out2, int16_t *output)</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> uint16x8_t u16results = vcombine_u16(vqmovn_u32(out), vqmovn_u32(out2));</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keyword">const</span> int16x8_t s16results = vreinterpretq_s16_u16(vminq_u16(u16results, max_int16));</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  vst1q_s16(output, s16results);</div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span> }</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span> </div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> uint32x4_t &out, <span class="keyword">const</span> uint32x4_t &out2, uint8_t *output)</div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span> {</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  <span class="keyword">const</span> uint8x8_t u8results = vqmovn_u16(vcombine_u16(vqmovn_u32(out),</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  vqmovn_u32(out2)));</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  vst1_u8(output, u8results);</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span> }</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span> </div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> int16x8_t &out, <span class="keyword">const</span> int16x8_t &out2, int16_t *output)</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> {</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  vst1q_s16(output, out);</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  vst1q_s16(output + 8, out2);</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> }</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span> </div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> int16x8_t &out, <span class="keyword">const</span> int16x8_t &out2, uint8_t *output)</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span> {</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="keyword">const</span> uint8x16_t u8results = vcombine_u8(vqmovun_s16(out),</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  vqmovun_s16(out2));</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  vst1q_u8(output, u8results);</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span> }</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> </div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> uint16x8_t &out, <span class="keyword">const</span> uint16x8_t &out2, uint8_t *output)</div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span> {</div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  <span class="keyword">const</span> uint8x16_t u8results = vcombine_u8(vqmovn_u16(out),</div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  vqmovn_u16(out2));</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  vst1q_u8(output, u8results);</div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span> }</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span> </div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(<span class="keyword">const</span> uint16x8_t &out, <span class="keyword">const</span> uint16x8_t &out2, int16_t *output)</div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span> {</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  vst1q_s16(output, vreinterpretq_s16_u16(vminq_u16(out, max_int16)));</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  vst1q_s16(output + 8, vreinterpretq_s16_u16(vminq_u16(out2, max_int16)));</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</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">inline</span> <span class="keywordtype">void</span> convolve_row3x1_unrolled(int32x4_t &out, int32x4_t &out2, <span class="keyword">const</span> uint8x16_t &row_data, <span class="keyword">const</span> int16x4_t &mat0, <span class="keyword">const</span> int16x4_t &mat1, <span class="keyword">const</span> int16x4_t &mat2)</div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span> {</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  <span class="comment">// Convert to s16 and split in blocks of 4 values:</span></div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keyword">const</span> int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data)));</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keyword">const</span> int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data)));</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>  <span class="keyword">const</span> int16x4x3_t row =</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  {</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  {</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  vget_low_s16(s16_tmp0),</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  vget_high_s16(s16_tmp0),</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  vget_low_s16(s16_tmp1)</div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  }</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  };</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span> </div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  <span class="comment">// Calculate row left value for pixels [0,3]</span></div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  out = vmlal_s16(out, row.val[0], mat0);</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  <span class="comment">// Calculate row middle value for pixels [0,3]</span></div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1);</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  <span class="comment">// Calculate row right value for pixels [0,3]</span></div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2);</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span> </div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="comment">// Calculate row left value for pixels [4,7]</span></div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  out2 = vmlal_s16(out2, row.val[1], mat0);</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  <span class="comment">// Calculate row middle value for pixels [4,7]</span></div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1);</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  <span class="comment">// Calculate row right value for pixels [4,7]</span></div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2);</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> </div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> convolve_row3x1(int32x4_t &out, int32x4_t &out2, <span class="keyword">const</span> uint8x16_t &row_data, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>)</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span> {</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  <span class="keyword">const</span> int16x4_t mat0 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>);</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="keyword">const</span> int16x4_t mat1 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 1);</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  <span class="keyword">const</span> int16x4_t mat2 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 2);</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span> </div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  convolve_row3x1_unrolled(out, out2, row_data, mat0, mat1, mat2);</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> </div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> convolve_row5x1(int32x4_t &out, int32x4_t &out2, <span class="keyword">const</span> uint8x16_t &row_data, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>)</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span> {</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="keyword">const</span> int16x4_t mat0 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>);</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="keyword">const</span> int16x4_t mat1 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 1);</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="keyword">const</span> int16x4_t mat2 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 2);</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  <span class="keyword">const</span> int16x4_t mat3 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 3);</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  <span class="keyword">const</span> int16x4_t mat4 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 4);</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span> </div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  <span class="comment">// Convert to s16 and split in blocks of 4 values:</span></div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  <span class="keyword">const</span> int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data)));</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="keyword">const</span> int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data)));</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="keyword">const</span> int16x4x3_t row =</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  {</div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  {</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  vget_low_s16(s16_tmp0),</div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  vget_high_s16(s16_tmp0),</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  vget_low_s16(s16_tmp1)</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  }</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  };</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>  <span class="comment">// Calculate row left 2 value for pixels [0,3]</span></div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  out = vmlal_s16(out, row.val[0], mat0);</div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// Calculate row left 1 value for pixels [0,3]</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1);</div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="comment">// Calculate row middle value for pixels [0,3]</span></div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2);</div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  <span class="comment">// Calculate row right +1 value for pixels [0,3]</span></div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3);</div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  <span class="comment">// Calculate row right +2 value for pixels [0,3]</span></div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  out = vmlal_s16(out, row.val[1], mat4);</div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span> </div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  <span class="comment">// Calculate row left 2 value for pixels [4,7]</span></div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  out2 = vmlal_s16(out2, row.val[1], mat0);</div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  <span class="comment">// Calculate row left 1 value for pixels [4,7]</span></div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1);</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="comment">// Calculate row middle value for pixels [4,7]</span></div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2);</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="comment">// Calculate row right +1 value for pixels [4,7]</span></div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3);</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="comment">// Calculate row right +2 value for pixels [4,7]</span></div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  out2 = vmlal_s16(out2, row.val[2], mat4);</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> </div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> convolve_row7x1(int32x4_t &out, int32x4_t &out2, <span class="keyword">const</span> uint8x16_t &row_data, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>)</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span> {</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  <span class="keyword">const</span> int16x4_t mat0 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>);</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  <span class="keyword">const</span> int16x4_t mat1 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 1);</div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  <span class="keyword">const</span> int16x4_t mat2 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 2);</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  <span class="keyword">const</span> int16x4_t mat3 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 3);</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  <span class="keyword">const</span> int16x4_t mat4 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 4);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  <span class="keyword">const</span> int16x4_t mat5 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 5);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  <span class="keyword">const</span> int16x4_t mat6 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 6);</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span> </div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <span class="comment">// Convert to s16 and split in blocks of 4 values:</span></div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  <span class="keyword">const</span> int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data)));</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  <span class="keyword">const</span> int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data)));</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span> </div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  <span class="keyword">const</span> int16x4x4_t row =</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>  {</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  vget_low_s16(s16_tmp0),</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  vget_high_s16(s16_tmp0),</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  vget_low_s16(s16_tmp1),</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  vget_high_s16(s16_tmp1)</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> </div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="comment">// Calculate row left 3 value for pixels [0,3]</span></div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  out = vmlal_s16(out, row.val[0], mat0);</div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="comment">// Calculate row left 2 value for pixels [0,3]</span></div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1);</div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="comment">// Calculate row left 1 value for pixels [0,3]</span></div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2);</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span>  <span class="comment">// Calculate row middle value for pixels [0,3]</span></div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3);</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  <span class="comment">// Calculate row right +1 value for pixels [0,3]</span></div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  out = vmlal_s16(out, row.val[1], mat4);</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <span class="comment">// Calculate row right +2 value for pixels [0,3]</span></div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 1), mat5);</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  <span class="comment">// Calculate row right +3 value for pixels [0,3]</span></div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 2), mat6);</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span> </div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span>  <span class="comment">// Calculate row left 3 value for pixels [4,7]</span></div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  out2 = vmlal_s16(out2, row.val[1], mat0);</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  <span class="comment">// Calculate row left 2 value for pixels [4,7]</span></div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1);</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  <span class="comment">// Calculate row left 1 value for pixels [4,7]</span></div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2);</div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  <span class="comment">// Calculate row middle value for pixels [4,7]</span></div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3);</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  <span class="comment">// Calculate row right +1 value for pixels [4,7]</span></div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  out2 = vmlal_s16(out2, row.val[2], mat4);</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  <span class="comment">// Calculate row right +2 value for pixels [4,7]</span></div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 1), mat5);</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="comment">// Calculate row right +3 value for pixels [4,7]</span></div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 2), mat6);</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span> }</div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span> </div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> convolve_row9x1(int32x4_t &out, int32x4_t &out2, <span class="keyword">const</span> uint8x16_t &row_data, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>)</div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span> {</div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  <span class="keyword">const</span> int16x4_t mat0 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a>);</div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  <span class="keyword">const</span> int16x4_t mat1 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 1);</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  <span class="keyword">const</span> int16x4_t mat2 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 2);</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="keyword">const</span> int16x4_t mat3 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 3);</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  <span class="keyword">const</span> int16x4_t mat4 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 4);</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  <span class="keyword">const</span> int16x4_t mat5 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 5);</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  <span class="keyword">const</span> int16x4_t mat6 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 6);</div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="keyword">const</span> int16x4_t mat7 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 7);</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  <span class="keyword">const</span> int16x4_t mat8 = vld1_dup_s16(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#ae38161fda91461cff23b750a95e47e6f">convolution</a> + 8);</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>  <span class="comment">// Convert to s16 and split in blocks of 4 values:</span></div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  <span class="keyword">const</span> int16x8_t s16_tmp0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(row_data)));</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  <span class="keyword">const</span> int16x8_t s16_tmp1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(row_data)));</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span> </div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  <span class="keyword">const</span> int16x4x4_t row =</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>  vget_low_s16(s16_tmp0),</div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  vget_high_s16(s16_tmp0),</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  vget_low_s16(s16_tmp1),</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  vget_high_s16(s16_tmp1)</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</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> </div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  <span class="comment">// Calculate row left 4 value for pixels [0,3]</span></div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  out = vmlal_s16(out, row.val[0], mat0);</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  <span class="comment">// Calculate row left 3 value for pixels [0,3]</span></div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 1), mat1);</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  <span class="comment">// Calculate row left 2 value for pixels [0,3]</span></div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 2), mat2);</div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">// Calculate row left 1 value for pixels [0,3]</span></div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  out = vmlal_s16(out, vext_s16(row.val[0], row.val[1], 3), mat3);</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  <span class="comment">// Calculate row middle value for pixels [0,3]</span></div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  out = vmlal_s16(out, row.val[1], mat4);</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span>  <span class="comment">// Calculate row right +1 value for pixels [0,3]</span></div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 1), mat5);</div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  <span class="comment">// Calculate row right +2 value for pixels [0,3]</span></div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 2), mat6);</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  <span class="comment">// Calculate row right +3 value for pixels [0,3]</span></div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  out = vmlal_s16(out, vext_s16(row.val[1], row.val[2], 3), mat7);</div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  <span class="comment">// Calculate row right +4 value for pixels [0,3]</span></div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  out = vmlal_s16(out, row.val[2], mat8);</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>  <span class="comment">// Calculate row left 4 value for pixels [0,3]</span></div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  out2 = vmlal_s16(out2, row.val[1], mat0);</div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  <span class="comment">// Calculate row left 3 value for pixels [0,3]</span></div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 1), mat1);</div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  <span class="comment">// Calculate row left 2 value for pixels [0,3]</span></div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 2), mat2);</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <span class="comment">// Calculate row left 1 value for pixels [0,3]</span></div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  out2 = vmlal_s16(out2, vext_s16(row.val[1], row.val[2], 3), mat3);</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <span class="comment">// Calculate row middle value for pixels [0,3]</span></div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  out2 = vmlal_s16(out2, row.val[2], mat4);</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  <span class="comment">// Calculate row right +1 value for pixels [0,3]</span></div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 1), mat5);</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  <span class="comment">// Calculate row right +2 value for pixels [0,3]</span></div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 2), mat6);</div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  <span class="comment">// Calculate row right +3 value for pixels [0,3]</span></div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  out2 = vmlal_s16(out2, vext_s16(row.val[2], row.val[3], 3), mat7);</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  <span class="comment">// Calculate row right +4 value for pixels [0,3]</span></div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  out2 = vmlal_s16(out2, row.val[3], mat8);</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> }</div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> } <span class="comment">// namespace</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="comment">/****************************************************************************************\</span></div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span> <span class="comment"> * Square Convolution *</span></div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span> <span class="comment">\****************************************************************************************/</span></div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span> </div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00312"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a230a4894adf8380eca9e8c2aac03e9b8"> 312</a></span> <a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a230a4894adf8380eca9e8c2aac03e9b8">NEConvolutionKernel<matrix_size>::NEConvolutionKernel</a>()</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  : <a class="code" href="classarm__compute_1_1_i_c_p_p_simple_kernel.xhtml">INESimpleKernel</a>(), _scale(0), _convolution{ {} }</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> }</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span> </div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00318"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7"> 318</a></span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a> <a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7">NEConvolutionKernel<matrix_size>::border_size</a>()<span class="keyword"> const</span></div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span> <span class="keyword"></span>{</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  <span class="keywordflow">return</span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a>{ matrix_size / 2 };</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span> }</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span> </div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00324"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a8873a8f053f3d2322e297409645c6222"> 324</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a8873a8f053f3d2322e297409645c6222">NEConvolutionKernel<matrix_size>::configure</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *input, <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *output, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>, uint32_t <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>, <span class="keywordtype">bool</span> border_undefined)</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span> {</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  <a class="code" href="_validate_8h.xhtml#a921b705e9e3e0fe928928447869e62a5">ARM_COMPUTE_ERROR_ON_NULLPTR</a>(input, output, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>);</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>  <a class="code" href="namespacearm__compute.xhtml#a9d1a839c51134b2ae171a2264c541b6f">set_shape_if_empty</a>(*output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a7c66505457d00ece3aa4b34cab80757d">tensor_shape</a>());</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>  <a class="code" href="_validate_8h.xhtml#aafce5e9928fac427dcd29d960bcf2ae3">ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES</a>(input, output);</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(input, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>);</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(output, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>);</div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span> </div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  _input = input;</div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  _output = output;</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span> </div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  std::copy_n(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>, _convolution.size(), _convolution.begin());</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span> </div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  <span class="keywordflow">if</span>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> == 0)</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  {</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  _scale = <a class="code" href="namespacearm__compute.xhtml#a0101a40c4a6acc2af3b55afa7632f16a">calculate_matrix_scale</a>(_convolution.data(), matrix_size);</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  }</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="keywordflow">else</span></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>  _scale = <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>;</div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  }</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span> </div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="comment">// Configure kernel window</span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_processed_per_iteration = 8;</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_read_per_iteration = 16;</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_written_per_iteration = 8;</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>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win = <a class="code" href="namespacearm__compute.xhtml#ab7980fa5ee693e3282a76da047a1c3b5">calculate_max_window</a>(*input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), <a class="code" href="classarm__compute_1_1_steps.xhtml">Steps</a>(num_elems_processed_per_iteration), border_undefined, border_size());</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  <a class="code" href="classarm__compute_1_1_access_window_horizontal.xhtml">AccessWindowHorizontal</a> output_access(output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), 0, num_elems_written_per_iteration);</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span> </div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  <a class="code" href="namespacearm__compute.xhtml#afc4bd8e872567d9c4c57d89eb0bb3da1">update_window_and_padding</a>(win,</div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <a class="code" href="classarm__compute_1_1_access_window_rectangle.xhtml">AccessWindowRectangle</a>(input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), -border_size().left, -border_size().top, num_elems_read_per_iteration, matrix_size),</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  output_access);</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span> </div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  output_access.set_valid_region(win, input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a74dfd07380a290c34fe7c8e065029b95">valid_region</a>(), border_undefined, border_size());</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span> </div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  INEKernel::configure(win);</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span> }</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="keyword">template</span> <></div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml">NEConvolutionKernel<3>::convolution</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &win)</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>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_input->buffer() == <span class="keyword">nullptr</span>);</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>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> input(_input, win);</div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> output(_output, win);</div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span> </div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  <span class="comment">// Load the matrix's coefficients into NEON registers:</span></div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  <span class="keyword">const</span> int16x4_t mat00 = vld1_dup_s16(_convolution.data());</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  <span class="keyword">const</span> int16x4_t mat01 = vld1_dup_s16(_convolution.data() + 1);</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  <span class="keyword">const</span> int16x4_t mat02 = vld1_dup_s16(_convolution.data() + 2);</div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  <span class="keyword">const</span> int16x4_t mat10 = vld1_dup_s16(_convolution.data() + 3);</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="keyword">const</span> int16x4_t mat11 = vld1_dup_s16(_convolution.data() + 4);</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  <span class="keyword">const</span> int16x4_t mat12 = vld1_dup_s16(_convolution.data() + 5);</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="keyword">const</span> int16x4_t mat20 = vld1_dup_s16(_convolution.data() + 6);</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  <span class="keyword">const</span> int16x4_t mat21 = vld1_dup_s16(_convolution.data() + 7);</div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  <span class="keyword">const</span> int16x4_t mat22 = vld1_dup_s16(_convolution.data() + 8);</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  <span class="keyword">const</span> float32x4_t scale_val = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span> </div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(<a class="code" href="classarm__compute_1_1_coordinates.xhtml">Coordinates</a>(-1, -1));</div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(<a class="code" href="classarm__compute_1_1_coordinates.xhtml">Coordinates</a>(-1, 0));</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(<a class="code" href="classarm__compute_1_1_coordinates.xhtml">Coordinates</a>(-1, 1));</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span> </div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_coordinates.xhtml">Coordinates</a> &)</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  {</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  int32x4_t out = vdupq_n_s32(0);</div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  int32x4_t out2 = vdupq_n_s32(0);</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>  <span class="comment">// Load 16 bytes from the top row:</span></div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  <span class="keyword">const</span> uint8x16_t top_data = vld1q_u8(input_top_ptr + input.offset());</div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  convolve_row3x1_unrolled(out, out2, top_data, mat00, mat01, mat02);</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span> </div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  <span class="comment">// Load 16 bytes from the middle row:</span></div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  <span class="keyword">const</span> uint8x16_t mid_data = vld1q_u8(input_mid_ptr + input.offset());</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  convolve_row3x1_unrolled(out, out2, mid_data, mat10, mat11, mat12);</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>  <span class="comment">// Load 16 bytes from the middle row:</span></div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  <span class="keyword">const</span> uint8x16_t low_data = vld1q_u8(input_low_ptr + input.offset());</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  convolve_row3x1_unrolled(out, out2, low_data, mat20, mat21, mat22);</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span> </div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  <span class="comment">// Apply scale</span></div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <span class="keywordflow">if</span>(_scale != 1)</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">// Convert to F32, scale and convert back to S32</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val));</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val));</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  }</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>  <span class="comment">// Clamp and store as U8 or S16:</span></div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out, out2, reinterpret_cast<OutputType *>(output.ptr()));</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>  input, output);</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"> 422</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span> <span class="keywordtype">void</span> NEConvolutionKernel<5>::convolution(<span class="keyword">const</span> Window &win)</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span> {</div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_input->buffer() == <span class="keyword">nullptr</span>);</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>  Iterator input(_input, win);</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  Iterator output(_output, win);</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>  <span class="keyword">const</span> float32x4_t scale_val = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span> </div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, -2));</div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, -1));</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 0));</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 1));</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-2, 2));</div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> </div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  {</div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  int32x4_t out = vdupq_n_s32(0);</div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  int32x4_t out2 = vdupq_n_s32(0);</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>  <span class="comment">// Load 16 bytes from the top2 row:</span></div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  <span class="keyword">const</span> uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset());</div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  convolve_row5x1(out, out2, data_t2, _convolution.data());</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>  <span class="comment">// Load 16 bytes from the top1 row:</span></div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  <span class="keyword">const</span> uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset());</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  convolve_row5x1(out, out2, data_t1, _convolution.data() + 5);</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span> </div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  <span class="comment">// Load 16 bytes from the middle row:</span></div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  <span class="keyword">const</span> uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset());</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  convolve_row5x1(out, out2, data_m, _convolution.data() + 10);</div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span> </div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  <span class="comment">// Load 16 bytes from the low1 row:</span></div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  <span class="keyword">const</span> uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset());</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  convolve_row5x1(out, out2, data_b1, _convolution.data() + 15);</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span> </div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span>  <span class="comment">// Load 16 bytes from the low2 row:</span></div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  <span class="keyword">const</span> uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset());</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  convolve_row5x1(out, out2, data_b2, _convolution.data() + 20);</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="comment">// Apply scale</span></div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span>  {</div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  <span class="comment">// Convert to F32, scale and convert back to S32</span></div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val));</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val));</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  }</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span> </div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  <span class="comment">// Clamp and store as U8 or S16:</span></div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out, out2, reinterpret_cast<OutputType *>(output.ptr()));</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  },</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  input, output);</div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span> }</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span> </div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span> <span class="keywordtype">void</span> NEConvolutionKernel<7>::convolution(<span class="keyword">const</span> Window &win)</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>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_input->buffer() == <span class="keyword">nullptr</span>);</div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span> </div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  Iterator input(_input, win);</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  Iterator output(_output, win);</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="keyword">const</span> float32x4_t scale_val = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span> </div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -3));</div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -2));</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, -1));</div><div class="line"><a name="l00494"></a><span class="lineno"> 494</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 0));</div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 1));</div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 2));</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-3, 3));</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>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</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>  int32x4_t out = vdupq_n_s32(0);</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  int32x4_t out2 = vdupq_n_s32(0);</div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span> </div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  <span class="comment">// Load 16 bytes from the top3 row:</span></div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  <span class="keyword">const</span> uint8x16_t data_t3 = vld1q_u8(input_top3_ptr + input.offset());</div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  convolve_row7x1(out, out2, data_t3, _convolution.data());</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>  <span class="comment">// Load 16 bytes from the top2 row:</span></div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span>  <span class="keyword">const</span> uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset());</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  convolve_row7x1(out, out2, data_t2, _convolution.data() + 7);</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>  <span class="comment">// Load 16 bytes from the top1 row:</span></div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  <span class="keyword">const</span> uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset());</div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  convolve_row7x1(out, out2, data_t1, _convolution.data() + 14);</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>  <span class="comment">// Load 16 bytes from the middle row:</span></div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span>  <span class="keyword">const</span> uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset());</div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  convolve_row7x1(out, out2, data_m, _convolution.data() + 21);</div><div class="line"><a name="l00519"></a><span class="lineno"> 519</span> </div><div class="line"><a name="l00520"></a><span class="lineno"> 520</span>  <span class="comment">// Load 16 bytes from the low1 row:</span></div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span>  <span class="keyword">const</span> uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset());</div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  convolve_row7x1(out, out2, data_b1, _convolution.data() + 28);</div><div class="line"><a name="l00523"></a><span class="lineno"> 523</span> </div><div class="line"><a name="l00524"></a><span class="lineno"> 524</span>  <span class="comment">// Load 16 bytes from the low2 row:</span></div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span>  <span class="keyword">const</span> uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset());</div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span>  convolve_row7x1(out, out2, data_b2, _convolution.data() + 35);</div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span> </div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span>  <span class="comment">// Load 16 bytes from the low3 row:</span></div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span>  <span class="keyword">const</span> uint8x16_t data_b3 = vld1q_u8(input_low3_ptr + input.offset());</div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  convolve_row7x1(out, out2, data_b3, _convolution.data() + 42);</div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span> </div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span>  <span class="comment">// Apply scale</span></div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  {</div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span>  <span class="comment">// Convert to F32, scale and convert back to S32</span></div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val));</div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val));</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  }</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span> </div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  <span class="comment">// Clamp and store as U8 or S16:</span></div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out, out2, reinterpret_cast<OutputType *>(output.ptr()));</div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  },</div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  input, output);</div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span> }</div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span> </div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span> <span class="keywordtype">void</span> NEConvolutionKernel<9>::convolution(<span class="keyword">const</span> Window &win)</div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span> {</div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l00551"></a><span class="lineno"> 551</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_input->buffer() == <span class="keyword">nullptr</span>);</div><div class="line"><a name="l00552"></a><span class="lineno"> 552</span> </div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  Iterator input(_input, win);</div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span>  Iterator output(_output, win);</div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span> </div><div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  <span class="keyword">const</span> float32x4_t scale_val = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l00557"></a><span class="lineno"> 557</span> </div><div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top4_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -4));</div><div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -3));</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -2));</div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_top1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, -1));</div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_mid_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 0));</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low1_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 1));</div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low2_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 2));</div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low3_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 3));</div><div class="line"><a name="l00566"></a><span class="lineno"> 566</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">char</span> *input_low4_ptr = _input->buffer() + _input->info()->offset_element_in_bytes(Coordinates(-4, 4));</div><div class="line"><a name="l00567"></a><span class="lineno"> 567</span> </div><div class="line"><a name="l00568"></a><span class="lineno"> 568</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  {</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span>  int32x4_t out = vdupq_n_s32(0);</div><div class="line"><a name="l00571"></a><span class="lineno"> 571</span>  int32x4_t out2 = vdupq_n_s32(0);</div><div class="line"><a name="l00572"></a><span class="lineno"> 572</span> </div><div class="line"><a name="l00573"></a><span class="lineno"> 573</span>  <span class="comment">// Load 16 bytes from the top4 row:</span></div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  <span class="keyword">const</span> uint8x16_t data_t4 = vld1q_u8(input_top4_ptr + input.offset());</div><div class="line"><a name="l00575"></a><span class="lineno"> 575</span>  convolve_row9x1(out, out2, data_t4, _convolution.data());</div><div class="line"><a name="l00576"></a><span class="lineno"> 576</span> </div><div class="line"><a name="l00577"></a><span class="lineno"> 577</span>  <span class="comment">// Load 16 bytes from the top3 row:</span></div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span>  <span class="keyword">const</span> uint8x16_t data_t3 = vld1q_u8(input_top3_ptr + input.offset());</div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  convolve_row9x1(out, out2, data_t3, _convolution.data() + 9);</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span> </div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span>  <span class="comment">// Load 16 bytes from the top2 row:</span></div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span>  <span class="keyword">const</span> uint8x16_t data_t2 = vld1q_u8(input_top2_ptr + input.offset());</div><div class="line"><a name="l00583"></a><span class="lineno"> 583</span>  convolve_row9x1(out, out2, data_t2, _convolution.data() + 18);</div><div class="line"><a name="l00584"></a><span class="lineno"> 584</span> </div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  <span class="comment">// Load 16 bytes from the top1 row:</span></div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  <span class="keyword">const</span> uint8x16_t data_t1 = vld1q_u8(input_top1_ptr + input.offset());</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span>  convolve_row9x1(out, out2, data_t1, _convolution.data() + 27);</div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span> </div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  <span class="comment">// Load 16 bytes from the middle row:</span></div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span>  <span class="keyword">const</span> uint8x16_t data_m = vld1q_u8(input_mid_ptr + input.offset());</div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>  convolve_row9x1(out, out2, data_m, _convolution.data() + 36);</div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span> </div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span>  <span class="comment">// Load 16 bytes from the low1 row:</span></div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span>  <span class="keyword">const</span> uint8x16_t data_b1 = vld1q_u8(input_low1_ptr + input.offset());</div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  convolve_row9x1(out, out2, data_b1, _convolution.data() + 45);</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span> </div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  <span class="comment">// Load 16 bytes from the low2 row:</span></div><div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  <span class="keyword">const</span> uint8x16_t data_b2 = vld1q_u8(input_low2_ptr + input.offset());</div><div class="line"><a name="l00599"></a><span class="lineno"> 599</span>  convolve_row9x1(out, out2, data_b2, _convolution.data() + 54);</div><div class="line"><a name="l00600"></a><span class="lineno"> 600</span> </div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  <span class="comment">// Load 16 bytes from the low3 row:</span></div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  <span class="keyword">const</span> uint8x16_t data_b3 = vld1q_u8(input_low3_ptr + input.offset());</div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  convolve_row9x1(out, out2, data_b3, _convolution.data() + 63);</div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span> </div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  <span class="comment">// Load 16 bytes from the low4 row:</span></div><div class="line"><a name="l00606"></a><span class="lineno"> 606</span>  <span class="keyword">const</span> uint8x16_t data_b4 = vld1q_u8(input_low4_ptr + input.offset());</div><div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  convolve_row9x1(out, out2, data_b4, _convolution.data() + 72);</div><div class="line"><a name="l00608"></a><span class="lineno"> 608</span> </div><div class="line"><a name="l00609"></a><span class="lineno"> 609</span>  <span class="comment">// Apply scale</span></div><div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l00611"></a><span class="lineno"> 611</span>  {</div><div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  <span class="comment">// Convert to F32, scale and convert back to S32</span></div><div class="line"><a name="l00613"></a><span class="lineno"> 613</span>  out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val));</div><div class="line"><a name="l00614"></a><span class="lineno"> 614</span>  out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val));</div><div class="line"><a name="l00615"></a><span class="lineno"> 615</span>  }</div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span> </div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  <span class="comment">// Clamp and store as U8 or S16:</span></div><div class="line"><a name="l00618"></a><span class="lineno"> 618</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out, out2, reinterpret_cast<OutputType *>(output.ptr()));</div><div class="line"><a name="l00619"></a><span class="lineno"> 619</span>  },</div><div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  input, output);</div><div class="line"><a name="l00621"></a><span class="lineno"> 621</span> }</div><div class="line"><a name="l00622"></a><span class="lineno"> 622</span> </div><div class="line"><a name="l00623"></a><span class="lineno"> 623</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00624"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82"> 624</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82">NEConvolutionKernel<matrix_size>::run</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &window, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_thread_info.xhtml">ThreadInfo</a> &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>)</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span> {</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span>  <a class="code" href="_error_8h.xhtml#a6dc630a6ae9cc063b3924bcea8dee9d6">ARM_COMPUTE_UNUSED</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>);</div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  <a class="code" href="_validate_8h.xhtml#a1b35b0d258183cf9ef36adf684d0b88c">ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL</a>(<span class="keyword">this</span>);</div><div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  <a class="code" href="_validate_8h.xhtml#a6eb9ce82815fe429250189da7592ba75">ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW</a>(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">INEKernel::window</a>(), window);</div><div class="line"><a name="l00629"></a><span class="lineno"> 629</span> </div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  <span class="keywordflow">switch</span>(_output->info()->data_type())</div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>  {</div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>:</div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span>  convolution<uint8_t>(window);</div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  convolution<int16_t>(window);</div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported Data type!"</span>);</div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span>  }</div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span> }</div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span> </div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml">arm_compute::NEConvolutionKernel<3></a>;</div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml">arm_compute::NEConvolutionKernel<5></a>;</div><div class="line"><a name="l00646"></a><span class="lineno"> 646</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml">arm_compute::NEConvolutionKernel<7></a>;</div><div class="line"><a name="l00647"></a><span class="lineno"> 647</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_convolution_kernel.xhtml">arm_compute::NEConvolutionKernel<9></a>;</div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span> </div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span> <span class="comment">/****************************************************************************************\</span></div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span> <span class="comment"> * Separable Square Convolution *</span></div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span> <span class="comment">\****************************************************************************************/</span></div><div class="line"><a name="l00652"></a><span class="lineno"> 652</span> </div><div class="line"><a name="l00653"></a><span class="lineno"> 653</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00654"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a965a154580e00968b45aa6268986cd51"> 654</a></span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a965a154580e00968b45aa6268986cd51">NESeparableConvolutionHorKernel<matrix_size>::NESeparableConvolutionHorKernel</a>()</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span>  : _conv_row{ { 0 } }, _border_size(0)</div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span> {</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span> }</div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span> </div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00660"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7"> 660</a></span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7">NESeparableConvolutionHorKernel<matrix_size>::border_size</a>()<span class="keyword"> const</span></div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span> <span class="keyword"></span>{</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  <span class="keywordflow">return</span> _border_size;</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span> }</div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span> </div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00666"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a2047b01f2cdd57602126cccb6bd08625"> 666</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a2047b01f2cdd57602126cccb6bd08625">NESeparableConvolutionHorKernel<matrix_size>::configure</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *input, <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *output, <span class="keyword">const</span> int16_t *conv_row, <span class="keywordtype">bool</span> border_undefined)</div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span> {</div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  <a class="code" href="_validate_8h.xhtml#a921b705e9e3e0fe928928447869e62a5">ARM_COMPUTE_ERROR_ON_NULLPTR</a>(input, output, conv_row);</div><div class="line"><a name="l00669"></a><span class="lineno"> 669</span> </div><div class="line"><a name="l00670"></a><span class="lineno"> 670</span>  <a class="code" href="namespacearm__compute.xhtml#a9d1a839c51134b2ae171a2264c541b6f">set_shape_if_empty</a>(*output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a7c66505457d00ece3aa4b34cab80757d">tensor_shape</a>());</div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span> </div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  <a class="code" href="_validate_8h.xhtml#aafce5e9928fac427dcd29d960bcf2ae3">ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES</a>(input, output);</div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(input, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>);</div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(output, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aef9ef3ebca4d2b64b6ec83808bafa5f2">DataType::U16</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aa1e28eee0339658d39a8b4d325b56e9c">DataType::S32</a>);</div><div class="line"><a name="l00675"></a><span class="lineno"> 675</span> </div><div class="line"><a name="l00676"></a><span class="lineno"> 676</span>  _input = input;</div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>  _output = output;</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span>  std::copy_n(conv_row, _conv_row.size(), _conv_row.begin());</div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span>  _border_size = <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a>(border_undefined ? 0 : matrix_size / 2, matrix_size / 2);</div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span> </div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  <span class="comment">// Configure kernel window</span></div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_processed_per_iteration = 8;</div><div class="line"><a name="l00683"></a><span class="lineno"> 683</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_read_per_iteration = 16;</div><div class="line"><a name="l00684"></a><span class="lineno"> 684</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_written_per_iteration = 8;</div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span> </div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win = <a class="code" href="namespacearm__compute.xhtml#a9047a414d692d8814181a667a7ee68c4">calculate_max_window_horizontal</a>(*input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), <a class="code" href="classarm__compute_1_1_steps.xhtml">Steps</a>(num_elems_processed_per_iteration), border_undefined, border_size());</div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  <a class="code" href="classarm__compute_1_1_access_window_horizontal.xhtml">AccessWindowHorizontal</a> output_access(output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), 0, num_elems_written_per_iteration);</div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span> </div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span>  <a class="code" href="namespacearm__compute.xhtml#afc4bd8e872567d9c4c57d89eb0bb3da1">update_window_and_padding</a>(win,</div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span>  <a class="code" href="classarm__compute_1_1_access_window_horizontal.xhtml">AccessWindowHorizontal</a>(input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), -border_size().left, num_elems_read_per_iteration),</div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  output_access);</div><div class="line"><a name="l00692"></a><span class="lineno"> 692</span> </div><div class="line"><a name="l00693"></a><span class="lineno"> 693</span>  output_access.set_valid_region(win, input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a74dfd07380a290c34fe7c8e065029b95">valid_region</a>(), border_undefined, border_size());</div><div class="line"><a name="l00694"></a><span class="lineno"> 694</span> </div><div class="line"><a name="l00695"></a><span class="lineno"> 695</span>  INEKernel::configure(win);</div><div class="line"><a name="l00696"></a><span class="lineno"> 696</span> }</div><div class="line"><a name="l00697"></a><span class="lineno"> 697</span> </div><div class="line"><a name="l00698"></a><span class="lineno"> 698</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l00699"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82"> 699</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82">NESeparableConvolutionHorKernel<matrix_size>::run</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &window, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_thread_info.xhtml">ThreadInfo</a> &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>)</div><div class="line"><a name="l00700"></a><span class="lineno"> 700</span> {</div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  <a class="code" href="_error_8h.xhtml#a6dc630a6ae9cc063b3924bcea8dee9d6">ARM_COMPUTE_UNUSED</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>);</div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  <a class="code" href="_validate_8h.xhtml#a1b35b0d258183cf9ef36adf684d0b88c">ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL</a>(<span class="keyword">this</span>);</div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  <a class="code" href="_validate_8h.xhtml#a6eb9ce82815fe429250189da7592ba75">ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW</a>(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">INEKernel::window</a>(), window);</div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span>  <span class="keywordflow">switch</span>(_output->info()->data_type())</div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span>  {</div><div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aef9ef3ebca4d2b64b6ec83808bafa5f2">DataType::U16</a>:</div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span>  convolve<uint16_t>(window);</div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>  convolve<int16_t>(window);</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aa1e28eee0339658d39a8b4d325b56e9c">DataType::S32</a>:</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span>  convolve<int32_t>(window);</div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00715"></a><span class="lineno"> 715</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Unsupported intermediate data type!"</span>);</div><div class="line"><a name="l00717"></a><span class="lineno"> 717</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  }</div><div class="line"><a name="l00719"></a><span class="lineno"> 719</span> }</div><div class="line"><a name="l00720"></a><span class="lineno"> 720</span> </div><div class="line"><a name="l00721"></a><span class="lineno"> 721</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00722"></a><span class="lineno"> 722</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00723"></a><span class="lineno"> 723</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml">NESeparableConvolutionHorKernel<5>::convolve<uint16_t></a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &window)</div><div class="line"><a name="l00724"></a><span class="lineno"> 724</span> {</div><div class="line"><a name="l00725"></a><span class="lineno"> 725</span>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win_in(window);</div><div class="line"><a name="l00726"></a><span class="lineno"> 726</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -2);</div><div class="line"><a name="l00727"></a><span class="lineno"> 727</span> </div><div class="line"><a name="l00728"></a><span class="lineno"> 728</span>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> input(_input, win_in);</div><div class="line"><a name="l00729"></a><span class="lineno"> 729</span>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> output(_output, window);</div><div class="line"><a name="l00730"></a><span class="lineno"> 730</span> </div><div class="line"><a name="l00731"></a><span class="lineno"> 731</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_coordinates.xhtml">Coordinates</a> &)</div><div class="line"><a name="l00732"></a><span class="lineno"> 732</span>  {</div><div class="line"><a name="l00733"></a><span class="lineno"> 733</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00734"></a><span class="lineno"> 734</span> </div><div class="line"><a name="l00735"></a><span class="lineno"> 735</span>  <span class="keyword">const</span> uint16x8x2_t data_u16 =</div><div class="line"><a name="l00736"></a><span class="lineno"> 736</span>  {</div><div class="line"><a name="l00737"></a><span class="lineno"> 737</span>  {</div><div class="line"><a name="l00738"></a><span class="lineno"> 738</span>  vmovl_u8(vget_low_u8(data)),</div><div class="line"><a name="l00739"></a><span class="lineno"> 739</span>  vmovl_u8(vget_high_u8(data))</div><div class="line"><a name="l00740"></a><span class="lineno"> 740</span>  }</div><div class="line"><a name="l00741"></a><span class="lineno"> 741</span>  };</div><div class="line"><a name="l00742"></a><span class="lineno"> 742</span> </div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span>  uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]);</div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span> </div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span>  vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out);</div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  },</div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span>  input, output);</div><div class="line"><a name="l00752"></a><span class="lineno"> 752</span> }</div><div class="line"><a name="l00753"></a><span class="lineno"> 753</span> </div><div class="line"><a name="l00754"></a><span class="lineno"> 754</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00755"></a><span class="lineno"> 755</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00756"></a><span class="lineno"> 756</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<5>::convolve<int16_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00757"></a><span class="lineno"> 757</span> {</div><div class="line"><a name="l00758"></a><span class="lineno"> 758</span>  Window win_in(window);</div><div class="line"><a name="l00759"></a><span class="lineno"> 759</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -2);</div><div class="line"><a name="l00760"></a><span class="lineno"> 760</span> </div><div class="line"><a name="l00761"></a><span class="lineno"> 761</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00762"></a><span class="lineno"> 762</span>  Iterator output(_output, window);</div><div class="line"><a name="l00763"></a><span class="lineno"> 763</span> </div><div class="line"><a name="l00764"></a><span class="lineno"> 764</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  {</div><div class="line"><a name="l00766"></a><span class="lineno"> 766</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00767"></a><span class="lineno"> 767</span> </div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  {</div><div class="line"><a name="l00770"></a><span class="lineno"> 770</span>  {</div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span>  }</div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span>  };</div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span> </div><div class="line"><a name="l00776"></a><span class="lineno"> 776</span>  int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]);</div><div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l00779"></a><span class="lineno"> 779</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span> </div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out);</div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  },</div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span>  input, output);</div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span> }</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span> </div><div class="line"><a name="l00787"></a><span class="lineno"> 787</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00788"></a><span class="lineno"> 788</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00789"></a><span class="lineno"> 789</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<5>::convolve<int32_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00790"></a><span class="lineno"> 790</span> {</div><div class="line"><a name="l00791"></a><span class="lineno"> 791</span>  Window win_in(window);</div><div class="line"><a name="l00792"></a><span class="lineno"> 792</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -2);</div><div class="line"><a name="l00793"></a><span class="lineno"> 793</span> </div><div class="line"><a name="l00794"></a><span class="lineno"> 794</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00795"></a><span class="lineno"> 795</span>  Iterator output(_output, window);</div><div class="line"><a name="l00796"></a><span class="lineno"> 796</span> </div><div class="line"><a name="l00797"></a><span class="lineno"> 797</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00798"></a><span class="lineno"> 798</span>  {</div><div class="line"><a name="l00799"></a><span class="lineno"> 799</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00800"></a><span class="lineno"> 800</span> </div><div class="line"><a name="l00801"></a><span class="lineno"> 801</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l00802"></a><span class="lineno"> 802</span>  {</div><div class="line"><a name="l00803"></a><span class="lineno"> 803</span>  {</div><div class="line"><a name="l00804"></a><span class="lineno"> 804</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l00805"></a><span class="lineno"> 805</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l00806"></a><span class="lineno"> 806</span>  }</div><div class="line"><a name="l00807"></a><span class="lineno"> 807</span>  };</div><div class="line"><a name="l00808"></a><span class="lineno"> 808</span> </div><div class="line"><a name="l00809"></a><span class="lineno"> 809</span>  <span class="keyword">const</span> int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 1);</div><div class="line"><a name="l00810"></a><span class="lineno"> 810</span>  <span class="keyword">const</span> int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 2);</div><div class="line"><a name="l00811"></a><span class="lineno"> 811</span>  <span class="keyword">const</span> int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 3);</div><div class="line"><a name="l00812"></a><span class="lineno"> 812</span>  <span class="keyword">const</span> int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 4);</div><div class="line"><a name="l00813"></a><span class="lineno"> 813</span> </div><div class="line"><a name="l00814"></a><span class="lineno"> 814</span>  int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l00815"></a><span class="lineno"> 815</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[1]);</div><div class="line"><a name="l00816"></a><span class="lineno"> 816</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[2]);</div><div class="line"><a name="l00817"></a><span class="lineno"> 817</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[3]);</div><div class="line"><a name="l00818"></a><span class="lineno"> 818</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[4]);</div><div class="line"><a name="l00819"></a><span class="lineno"> 819</span> </div><div class="line"><a name="l00820"></a><span class="lineno"> 820</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low);</div><div class="line"><a name="l00821"></a><span class="lineno"> 821</span> </div><div class="line"><a name="l00822"></a><span class="lineno"> 822</span>  int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l00823"></a><span class="lineno"> 823</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[1]);</div><div class="line"><a name="l00824"></a><span class="lineno"> 824</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[2]);</div><div class="line"><a name="l00825"></a><span class="lineno"> 825</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[3]);</div><div class="line"><a name="l00826"></a><span class="lineno"> 826</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[4]);</div><div class="line"><a name="l00827"></a><span class="lineno"> 827</span> </div><div class="line"><a name="l00828"></a><span class="lineno"> 828</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high);</div><div class="line"><a name="l00829"></a><span class="lineno"> 829</span>  },</div><div class="line"><a name="l00830"></a><span class="lineno"> 830</span>  input, output);</div><div class="line"><a name="l00831"></a><span class="lineno"> 831</span> }</div><div class="line"><a name="l00832"></a><span class="lineno"> 832</span> </div><div class="line"><a name="l00833"></a><span class="lineno"> 833</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00834"></a><span class="lineno"> 834</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00835"></a><span class="lineno"> 835</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<7>::convolve<uint16_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00836"></a><span class="lineno"> 836</span> {</div><div class="line"><a name="l00837"></a><span class="lineno"> 837</span>  Window win_in(window);</div><div class="line"><a name="l00838"></a><span class="lineno"> 838</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -3);</div><div class="line"><a name="l00839"></a><span class="lineno"> 839</span> </div><div class="line"><a name="l00840"></a><span class="lineno"> 840</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00841"></a><span class="lineno"> 841</span>  Iterator output(_output, window);</div><div class="line"><a name="l00842"></a><span class="lineno"> 842</span> </div><div class="line"><a name="l00843"></a><span class="lineno"> 843</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00844"></a><span class="lineno"> 844</span>  {</div><div class="line"><a name="l00845"></a><span class="lineno"> 845</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00846"></a><span class="lineno"> 846</span> </div><div class="line"><a name="l00847"></a><span class="lineno"> 847</span>  <span class="keyword">const</span> uint16x8x2_t data_u16 =</div><div class="line"><a name="l00848"></a><span class="lineno"> 848</span>  {</div><div class="line"><a name="l00849"></a><span class="lineno"> 849</span>  {</div><div class="line"><a name="l00850"></a><span class="lineno"> 850</span>  vmovl_u8(vget_low_u8(data)),</div><div class="line"><a name="l00851"></a><span class="lineno"> 851</span>  vmovl_u8(vget_high_u8(data))</div><div class="line"><a name="l00852"></a><span class="lineno"> 852</span>  }</div><div class="line"><a name="l00853"></a><span class="lineno"> 853</span>  };</div><div class="line"><a name="l00854"></a><span class="lineno"> 854</span> </div><div class="line"><a name="l00855"></a><span class="lineno"> 855</span>  uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]);</div><div class="line"><a name="l00856"></a><span class="lineno"> 856</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l00857"></a><span class="lineno"> 857</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l00858"></a><span class="lineno"> 858</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l00859"></a><span class="lineno"> 859</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l00860"></a><span class="lineno"> 860</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 5), _conv_row[5]);</div><div class="line"><a name="l00861"></a><span class="lineno"> 861</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 6), _conv_row[6]);</div><div class="line"><a name="l00862"></a><span class="lineno"> 862</span> </div><div class="line"><a name="l00863"></a><span class="lineno"> 863</span>  vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out);</div><div class="line"><a name="l00864"></a><span class="lineno"> 864</span>  },</div><div class="line"><a name="l00865"></a><span class="lineno"> 865</span>  input, output);</div><div class="line"><a name="l00866"></a><span class="lineno"> 866</span> }</div><div class="line"><a name="l00867"></a><span class="lineno"> 867</span> </div><div class="line"><a name="l00868"></a><span class="lineno"> 868</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00869"></a><span class="lineno"> 869</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00870"></a><span class="lineno"> 870</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<7>::convolve<int16_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00871"></a><span class="lineno"> 871</span> {</div><div class="line"><a name="l00872"></a><span class="lineno"> 872</span>  Window win_in(window);</div><div class="line"><a name="l00873"></a><span class="lineno"> 873</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -3);</div><div class="line"><a name="l00874"></a><span class="lineno"> 874</span> </div><div class="line"><a name="l00875"></a><span class="lineno"> 875</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00876"></a><span class="lineno"> 876</span>  Iterator output(_output, window);</div><div class="line"><a name="l00877"></a><span class="lineno"> 877</span> </div><div class="line"><a name="l00878"></a><span class="lineno"> 878</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00879"></a><span class="lineno"> 879</span>  {</div><div class="line"><a name="l00880"></a><span class="lineno"> 880</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00881"></a><span class="lineno"> 881</span> </div><div class="line"><a name="l00882"></a><span class="lineno"> 882</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l00883"></a><span class="lineno"> 883</span>  {</div><div class="line"><a name="l00884"></a><span class="lineno"> 884</span>  {</div><div class="line"><a name="l00885"></a><span class="lineno"> 885</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l00886"></a><span class="lineno"> 886</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l00887"></a><span class="lineno"> 887</span>  }</div><div class="line"><a name="l00888"></a><span class="lineno"> 888</span>  };</div><div class="line"><a name="l00889"></a><span class="lineno"> 889</span> </div><div class="line"><a name="l00890"></a><span class="lineno"> 890</span>  int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]);</div><div class="line"><a name="l00891"></a><span class="lineno"> 891</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l00892"></a><span class="lineno"> 892</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l00893"></a><span class="lineno"> 893</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l00894"></a><span class="lineno"> 894</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l00895"></a><span class="lineno"> 895</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 5), _conv_row[5]);</div><div class="line"><a name="l00896"></a><span class="lineno"> 896</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 6), _conv_row[6]);</div><div class="line"><a name="l00897"></a><span class="lineno"> 897</span> </div><div class="line"><a name="l00898"></a><span class="lineno"> 898</span>  vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out);</div><div class="line"><a name="l00899"></a><span class="lineno"> 899</span>  },</div><div class="line"><a name="l00900"></a><span class="lineno"> 900</span>  input, output);</div><div class="line"><a name="l00901"></a><span class="lineno"> 901</span> }</div><div class="line"><a name="l00902"></a><span class="lineno"> 902</span> </div><div class="line"><a name="l00903"></a><span class="lineno"> 903</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00904"></a><span class="lineno"> 904</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00905"></a><span class="lineno"> 905</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<7>::convolve<int32_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00906"></a><span class="lineno"> 906</span> {</div><div class="line"><a name="l00907"></a><span class="lineno"> 907</span>  Window win_in(window);</div><div class="line"><a name="l00908"></a><span class="lineno"> 908</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -3);</div><div class="line"><a name="l00909"></a><span class="lineno"> 909</span> </div><div class="line"><a name="l00910"></a><span class="lineno"> 910</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00911"></a><span class="lineno"> 911</span>  Iterator output(_output, window);</div><div class="line"><a name="l00912"></a><span class="lineno"> 912</span> </div><div class="line"><a name="l00913"></a><span class="lineno"> 913</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00914"></a><span class="lineno"> 914</span>  {</div><div class="line"><a name="l00915"></a><span class="lineno"> 915</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00916"></a><span class="lineno"> 916</span> </div><div class="line"><a name="l00917"></a><span class="lineno"> 917</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l00918"></a><span class="lineno"> 918</span>  {</div><div class="line"><a name="l00919"></a><span class="lineno"> 919</span>  {</div><div class="line"><a name="l00920"></a><span class="lineno"> 920</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l00921"></a><span class="lineno"> 921</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l00922"></a><span class="lineno"> 922</span>  }</div><div class="line"><a name="l00923"></a><span class="lineno"> 923</span>  };</div><div class="line"><a name="l00924"></a><span class="lineno"> 924</span> </div><div class="line"><a name="l00925"></a><span class="lineno"> 925</span>  <span class="keyword">const</span> int16x8_t data_s16_l2 = vextq_s16(data_s16.val[0], data_s16.val[1], 1);</div><div class="line"><a name="l00926"></a><span class="lineno"> 926</span>  <span class="keyword">const</span> int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 2);</div><div class="line"><a name="l00927"></a><span class="lineno"> 927</span>  <span class="keyword">const</span> int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 3);</div><div class="line"><a name="l00928"></a><span class="lineno"> 928</span>  <span class="keyword">const</span> int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 4);</div><div class="line"><a name="l00929"></a><span class="lineno"> 929</span>  <span class="keyword">const</span> int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 5);</div><div class="line"><a name="l00930"></a><span class="lineno"> 930</span>  <span class="keyword">const</span> int16x8_t data_s16_r3 = vextq_s16(data_s16.val[0], data_s16.val[1], 6);</div><div class="line"><a name="l00931"></a><span class="lineno"> 931</span> </div><div class="line"><a name="l00932"></a><span class="lineno"> 932</span>  int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l00933"></a><span class="lineno"> 933</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l2), _conv_row[1]);</div><div class="line"><a name="l00934"></a><span class="lineno"> 934</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[2]);</div><div class="line"><a name="l00935"></a><span class="lineno"> 935</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[3]);</div><div class="line"><a name="l00936"></a><span class="lineno"> 936</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[4]);</div><div class="line"><a name="l00937"></a><span class="lineno"> 937</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[5]);</div><div class="line"><a name="l00938"></a><span class="lineno"> 938</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r3), _conv_row[6]);</div><div class="line"><a name="l00939"></a><span class="lineno"> 939</span> </div><div class="line"><a name="l00940"></a><span class="lineno"> 940</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low);</div><div class="line"><a name="l00941"></a><span class="lineno"> 941</span> </div><div class="line"><a name="l00942"></a><span class="lineno"> 942</span>  int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l00943"></a><span class="lineno"> 943</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l2), _conv_row[1]);</div><div class="line"><a name="l00944"></a><span class="lineno"> 944</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[2]);</div><div class="line"><a name="l00945"></a><span class="lineno"> 945</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[3]);</div><div class="line"><a name="l00946"></a><span class="lineno"> 946</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[4]);</div><div class="line"><a name="l00947"></a><span class="lineno"> 947</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[5]);</div><div class="line"><a name="l00948"></a><span class="lineno"> 948</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r3), _conv_row[6]);</div><div class="line"><a name="l00949"></a><span class="lineno"> 949</span> </div><div class="line"><a name="l00950"></a><span class="lineno"> 950</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high);</div><div class="line"><a name="l00951"></a><span class="lineno"> 951</span>  },</div><div class="line"><a name="l00952"></a><span class="lineno"> 952</span>  input, output);</div><div class="line"><a name="l00953"></a><span class="lineno"> 953</span> }</div><div class="line"><a name="l00954"></a><span class="lineno"> 954</span> </div><div class="line"><a name="l00955"></a><span class="lineno"> 955</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00956"></a><span class="lineno"> 956</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00957"></a><span class="lineno"> 957</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<9>::convolve<uint16_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00958"></a><span class="lineno"> 958</span> {</div><div class="line"><a name="l00959"></a><span class="lineno"> 959</span>  Window win_in(window);</div><div class="line"><a name="l00960"></a><span class="lineno"> 960</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -4);</div><div class="line"><a name="l00961"></a><span class="lineno"> 961</span> </div><div class="line"><a name="l00962"></a><span class="lineno"> 962</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l00963"></a><span class="lineno"> 963</span>  Iterator output(_output, window);</div><div class="line"><a name="l00964"></a><span class="lineno"> 964</span> </div><div class="line"><a name="l00965"></a><span class="lineno"> 965</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l00966"></a><span class="lineno"> 966</span>  {</div><div class="line"><a name="l00967"></a><span class="lineno"> 967</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l00968"></a><span class="lineno"> 968</span> </div><div class="line"><a name="l00969"></a><span class="lineno"> 969</span>  <span class="keyword">const</span> uint16x8x2_t data_u16 =</div><div class="line"><a name="l00970"></a><span class="lineno"> 970</span>  {</div><div class="line"><a name="l00971"></a><span class="lineno"> 971</span>  {</div><div class="line"><a name="l00972"></a><span class="lineno"> 972</span>  vmovl_u8(vget_low_u8(data)),</div><div class="line"><a name="l00973"></a><span class="lineno"> 973</span>  vmovl_u8(vget_high_u8(data))</div><div class="line"><a name="l00974"></a><span class="lineno"> 974</span>  }</div><div class="line"><a name="l00975"></a><span class="lineno"> 975</span>  };</div><div class="line"><a name="l00976"></a><span class="lineno"> 976</span> </div><div class="line"><a name="l00977"></a><span class="lineno"> 977</span>  uint16x8_t out = vmulq_n_u16(data_u16.val[0], _conv_row[0]);</div><div class="line"><a name="l00978"></a><span class="lineno"> 978</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l00979"></a><span class="lineno"> 979</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l00980"></a><span class="lineno"> 980</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l00981"></a><span class="lineno"> 981</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l00982"></a><span class="lineno"> 982</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 5), _conv_row[5]);</div><div class="line"><a name="l00983"></a><span class="lineno"> 983</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 6), _conv_row[6]);</div><div class="line"><a name="l00984"></a><span class="lineno"> 984</span>  out = vmlaq_n_u16(out, vextq_u16(data_u16.val[0], data_u16.val[1], 7), _conv_row[7]);</div><div class="line"><a name="l00985"></a><span class="lineno"> 985</span>  out = vmlaq_n_u16(out, data_u16.val[1], _conv_row[8]);</div><div class="line"><a name="l00986"></a><span class="lineno"> 986</span> </div><div class="line"><a name="l00987"></a><span class="lineno"> 987</span>  vst1q_u16(reinterpret_cast<uint16_t *>(output.ptr()), out);</div><div class="line"><a name="l00988"></a><span class="lineno"> 988</span>  },</div><div class="line"><a name="l00989"></a><span class="lineno"> 989</span>  input, output);</div><div class="line"><a name="l00990"></a><span class="lineno"> 990</span> }</div><div class="line"><a name="l00991"></a><span class="lineno"> 991</span> </div><div class="line"><a name="l00992"></a><span class="lineno"> 992</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00993"></a><span class="lineno"> 993</span> <span class="keyword">template</span> <></div><div class="line"><a name="l00994"></a><span class="lineno"> 994</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<9>::convolve<int16_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l00995"></a><span class="lineno"> 995</span> {</div><div class="line"><a name="l00996"></a><span class="lineno"> 996</span>  Window win_in(window);</div><div class="line"><a name="l00997"></a><span class="lineno"> 997</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -4);</div><div class="line"><a name="l00998"></a><span class="lineno"> 998</span> </div><div class="line"><a name="l00999"></a><span class="lineno"> 999</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l01000"></a><span class="lineno"> 1000</span>  Iterator output(_output, window);</div><div class="line"><a name="l01001"></a><span class="lineno"> 1001</span> </div><div class="line"><a name="l01002"></a><span class="lineno"> 1002</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01003"></a><span class="lineno"> 1003</span>  {</div><div class="line"><a name="l01004"></a><span class="lineno"> 1004</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l01005"></a><span class="lineno"> 1005</span> </div><div class="line"><a name="l01006"></a><span class="lineno"> 1006</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l01007"></a><span class="lineno"> 1007</span>  {</div><div class="line"><a name="l01008"></a><span class="lineno"> 1008</span>  {</div><div class="line"><a name="l01009"></a><span class="lineno"> 1009</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l01010"></a><span class="lineno"> 1010</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l01011"></a><span class="lineno"> 1011</span>  }</div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span>  };</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span> </div><div class="line"><a name="l01014"></a><span class="lineno"> 1014</span>  int16x8_t out = vmulq_n_s16(data_s16.val[0], _conv_row[0]);</div><div class="line"><a name="l01015"></a><span class="lineno"> 1015</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 1), _conv_row[1]);</div><div class="line"><a name="l01016"></a><span class="lineno"> 1016</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 2), _conv_row[2]);</div><div class="line"><a name="l01017"></a><span class="lineno"> 1017</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 3), _conv_row[3]);</div><div class="line"><a name="l01018"></a><span class="lineno"> 1018</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 4), _conv_row[4]);</div><div class="line"><a name="l01019"></a><span class="lineno"> 1019</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 5), _conv_row[5]);</div><div class="line"><a name="l01020"></a><span class="lineno"> 1020</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 6), _conv_row[6]);</div><div class="line"><a name="l01021"></a><span class="lineno"> 1021</span>  out = vmlaq_n_s16(out, vextq_s16(data_s16.val[0], data_s16.val[1], 7), _conv_row[7]);</div><div class="line"><a name="l01022"></a><span class="lineno"> 1022</span>  out = vmlaq_n_s16(out, data_s16.val[1], _conv_row[8]);</div><div class="line"><a name="l01023"></a><span class="lineno"> 1023</span> </div><div class="line"><a name="l01024"></a><span class="lineno"> 1024</span>  vst1q_s16(reinterpret_cast<int16_t *>(output.ptr()), out);</div><div class="line"><a name="l01025"></a><span class="lineno"> 1025</span>  },</div><div class="line"><a name="l01026"></a><span class="lineno"> 1026</span>  input, output);</div><div class="line"><a name="l01027"></a><span class="lineno"> 1027</span> }</div><div class="line"><a name="l01028"></a><span class="lineno"> 1028</span> </div><div class="line"><a name="l01029"></a><span class="lineno"> 1029</span> <span class="keyword">template</span> <></div><div class="line"><a name="l01030"></a><span class="lineno"> 1030</span> <span class="keyword">template</span> <></div><div class="line"><a name="l01031"></a><span class="lineno"> 1031</span> <span class="keywordtype">void</span> NESeparableConvolutionHorKernel<9>::convolve<int32_t>(<span class="keyword">const</span> Window &window)</div><div class="line"><a name="l01032"></a><span class="lineno"> 1032</span> {</div><div class="line"><a name="l01033"></a><span class="lineno"> 1033</span>  Window win_in(window);</div><div class="line"><a name="l01034"></a><span class="lineno"> 1034</span>  win_in.shift(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, -4);</div><div class="line"><a name="l01035"></a><span class="lineno"> 1035</span> </div><div class="line"><a name="l01036"></a><span class="lineno"> 1036</span>  Iterator input(_input, win_in);</div><div class="line"><a name="l01037"></a><span class="lineno"> 1037</span>  Iterator output(_output, window);</div><div class="line"><a name="l01038"></a><span class="lineno"> 1038</span> </div><div class="line"><a name="l01039"></a><span class="lineno"> 1039</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(window, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01040"></a><span class="lineno"> 1040</span>  {</div><div class="line"><a name="l01041"></a><span class="lineno"> 1041</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input.ptr());</div><div class="line"><a name="l01042"></a><span class="lineno"> 1042</span> </div><div class="line"><a name="l01043"></a><span class="lineno"> 1043</span>  <span class="keyword">const</span> int16x8x2_t data_s16 =</div><div class="line"><a name="l01044"></a><span class="lineno"> 1044</span>  {</div><div class="line"><a name="l01045"></a><span class="lineno"> 1045</span>  {</div><div class="line"><a name="l01046"></a><span class="lineno"> 1046</span>  vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(data))),</div><div class="line"><a name="l01047"></a><span class="lineno"> 1047</span>  vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(data)))</div><div class="line"><a name="l01048"></a><span class="lineno"> 1048</span>  }</div><div class="line"><a name="l01049"></a><span class="lineno"> 1049</span>  };</div><div class="line"><a name="l01050"></a><span class="lineno"> 1050</span> </div><div class="line"><a name="l01051"></a><span class="lineno"> 1051</span>  <span class="keyword">const</span> int16x8_t data_s16_l3 = vextq_s16(data_s16.val[0], data_s16.val[1], 1);</div><div class="line"><a name="l01052"></a><span class="lineno"> 1052</span>  <span class="keyword">const</span> int16x8_t data_s16_l2 = vextq_s16(data_s16.val[0], data_s16.val[1], 2);</div><div class="line"><a name="l01053"></a><span class="lineno"> 1053</span>  <span class="keyword">const</span> int16x8_t data_s16_l1 = vextq_s16(data_s16.val[0], data_s16.val[1], 3);</div><div class="line"><a name="l01054"></a><span class="lineno"> 1054</span>  <span class="keyword">const</span> int16x8_t data_s16_m = vextq_s16(data_s16.val[0], data_s16.val[1], 4);</div><div class="line"><a name="l01055"></a><span class="lineno"> 1055</span>  <span class="keyword">const</span> int16x8_t data_s16_r1 = vextq_s16(data_s16.val[0], data_s16.val[1], 5);</div><div class="line"><a name="l01056"></a><span class="lineno"> 1056</span>  <span class="keyword">const</span> int16x8_t data_s16_r2 = vextq_s16(data_s16.val[0], data_s16.val[1], 6);</div><div class="line"><a name="l01057"></a><span class="lineno"> 1057</span>  <span class="keyword">const</span> int16x8_t data_s16_r3 = vextq_s16(data_s16.val[0], data_s16.val[1], 7);</div><div class="line"><a name="l01058"></a><span class="lineno"> 1058</span> </div><div class="line"><a name="l01059"></a><span class="lineno"> 1059</span>  int32x4_t out_low = vmull_n_s16(vget_low_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l01060"></a><span class="lineno"> 1060</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l3), _conv_row[1]);</div><div class="line"><a name="l01061"></a><span class="lineno"> 1061</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l2), _conv_row[2]);</div><div class="line"><a name="l01062"></a><span class="lineno"> 1062</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_l1), _conv_row[3]);</div><div class="line"><a name="l01063"></a><span class="lineno"> 1063</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_m), _conv_row[4]);</div><div class="line"><a name="l01064"></a><span class="lineno"> 1064</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r1), _conv_row[5]);</div><div class="line"><a name="l01065"></a><span class="lineno"> 1065</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r2), _conv_row[6]);</div><div class="line"><a name="l01066"></a><span class="lineno"> 1066</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16_r3), _conv_row[7]);</div><div class="line"><a name="l01067"></a><span class="lineno"> 1067</span>  out_low = vmlal_n_s16(out_low, vget_low_s16(data_s16.val[1]), _conv_row[8]);</div><div class="line"><a name="l01068"></a><span class="lineno"> 1068</span> </div><div class="line"><a name="l01069"></a><span class="lineno"> 1069</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()), out_low);</div><div class="line"><a name="l01070"></a><span class="lineno"> 1070</span> </div><div class="line"><a name="l01071"></a><span class="lineno"> 1071</span>  int32x4_t out_high = vmull_n_s16(vget_high_s16(data_s16.val[0]), _conv_row[0]);</div><div class="line"><a name="l01072"></a><span class="lineno"> 1072</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l3), _conv_row[1]);</div><div class="line"><a name="l01073"></a><span class="lineno"> 1073</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l2), _conv_row[2]);</div><div class="line"><a name="l01074"></a><span class="lineno"> 1074</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_l1), _conv_row[3]);</div><div class="line"><a name="l01075"></a><span class="lineno"> 1075</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_m), _conv_row[4]);</div><div class="line"><a name="l01076"></a><span class="lineno"> 1076</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r1), _conv_row[5]);</div><div class="line"><a name="l01077"></a><span class="lineno"> 1077</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r2), _conv_row[6]);</div><div class="line"><a name="l01078"></a><span class="lineno"> 1078</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16_r3), _conv_row[7]);</div><div class="line"><a name="l01079"></a><span class="lineno"> 1079</span>  out_high = vmlal_n_s16(out_high, vget_high_s16(data_s16.val[1]), _conv_row[8]);</div><div class="line"><a name="l01080"></a><span class="lineno"> 1080</span> </div><div class="line"><a name="l01081"></a><span class="lineno"> 1081</span>  vst1q_s32(reinterpret_cast<int32_t *>(output.ptr()) + 4, out_high);</div><div class="line"><a name="l01082"></a><span class="lineno"> 1082</span>  },</div><div class="line"><a name="l01083"></a><span class="lineno"> 1083</span>  input, output);</div><div class="line"><a name="l01084"></a><span class="lineno"> 1084</span> }</div><div class="line"><a name="l01085"></a><span class="lineno"> 1085</span> </div><div class="line"><a name="l01086"></a><span class="lineno"> 1086</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml">arm_compute::NESeparableConvolutionHorKernel<5></a>;</div><div class="line"><a name="l01087"></a><span class="lineno"> 1087</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml">arm_compute::NESeparableConvolutionHorKernel<7></a>;</div><div class="line"><a name="l01088"></a><span class="lineno"> 1088</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_hor_kernel.xhtml">arm_compute::NESeparableConvolutionHorKernel<9></a>;</div><div class="line"><a name="l01089"></a><span class="lineno"> 1089</span> </div><div class="line"><a name="l01090"></a><span class="lineno"> 1090</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01091"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#ac3aaa23fc2c41fdb7a6de14841631fbd"> 1091</a></span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#ac3aaa23fc2c41fdb7a6de14841631fbd">NESeparableConvolutionVertKernel<matrix_size>::NESeparableConvolutionVertKernel</a>()</div><div class="line"><a name="l01092"></a><span class="lineno"> 1092</span>  : _conv_col{ { 0 } }, _scale(0)</div><div class="line"><a name="l01093"></a><span class="lineno"> 1093</span> {</div><div class="line"><a name="l01094"></a><span class="lineno"> 1094</span> }</div><div class="line"><a name="l01095"></a><span class="lineno"> 1095</span> </div><div class="line"><a name="l01096"></a><span class="lineno"> 1096</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01097"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7"> 1097</a></span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7">NESeparableConvolutionVertKernel<matrix_size>::border_size</a>()<span class="keyword"> const</span></div><div class="line"><a name="l01098"></a><span class="lineno"> 1098</span> <span class="keyword"></span>{</div><div class="line"><a name="l01099"></a><span class="lineno"> 1099</span>  <span class="keywordflow">return</span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a>{ matrix_size / 2, 0 };</div><div class="line"><a name="l01100"></a><span class="lineno"> 1100</span> }</div><div class="line"><a name="l01101"></a><span class="lineno"> 1101</span> </div><div class="line"><a name="l01102"></a><span class="lineno"> 1102</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01103"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#ad5b2a12ea8d073422e0615efca68ee22"> 1103</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#ad5b2a12ea8d073422e0615efca68ee22">NESeparableConvolutionVertKernel<matrix_size>::configure</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *input, <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *output, <span class="keyword">const</span> int16_t *conv_col, uint32_t <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>, <span class="keywordtype">bool</span> border_undefined)</div><div class="line"><a name="l01104"></a><span class="lineno"> 1104</span> {</div><div class="line"><a name="l01105"></a><span class="lineno"> 1105</span>  <a class="code" href="_validate_8h.xhtml#a921b705e9e3e0fe928928447869e62a5">ARM_COMPUTE_ERROR_ON_NULLPTR</a>(input, output, conv_col);</div><div class="line"><a name="l01106"></a><span class="lineno"> 1106</span> </div><div class="line"><a name="l01107"></a><span class="lineno"> 1107</span>  <a class="code" href="namespacearm__compute.xhtml#a9d1a839c51134b2ae171a2264c541b6f">set_shape_if_empty</a>(*output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a7c66505457d00ece3aa4b34cab80757d">tensor_shape</a>());</div><div class="line"><a name="l01108"></a><span class="lineno"> 1108</span> </div><div class="line"><a name="l01109"></a><span class="lineno"> 1109</span>  <a class="code" href="_validate_8h.xhtml#aafce5e9928fac427dcd29d960bcf2ae3">ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES</a>(input, output);</div><div class="line"><a name="l01110"></a><span class="lineno"> 1110</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(input, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aef9ef3ebca4d2b64b6ec83808bafa5f2">DataType::U16</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aa1e28eee0339658d39a8b4d325b56e9c">DataType::S32</a>);</div><div class="line"><a name="l01111"></a><span class="lineno"> 1111</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(output, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>);</div><div class="line"><a name="l01112"></a><span class="lineno"> 1112</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> == 0);</div><div class="line"><a name="l01113"></a><span class="lineno"> 1113</span> </div><div class="line"><a name="l01114"></a><span class="lineno"> 1114</span>  _input = input;</div><div class="line"><a name="l01115"></a><span class="lineno"> 1115</span>  _output = output;</div><div class="line"><a name="l01116"></a><span class="lineno"> 1116</span>  std::copy_n(conv_col, _conv_col.size(), _conv_col.begin());</div><div class="line"><a name="l01117"></a><span class="lineno"> 1117</span>  _scale = <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>;</div><div class="line"><a name="l01118"></a><span class="lineno"> 1118</span> </div><div class="line"><a name="l01119"></a><span class="lineno"> 1119</span>  <span class="comment">// Configure kernel window</span></div><div class="line"><a name="l01120"></a><span class="lineno"> 1120</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_processed_per_iteration = 16;</div><div class="line"><a name="l01121"></a><span class="lineno"> 1121</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_read_per_iteration = 16;</div><div class="line"><a name="l01122"></a><span class="lineno"> 1122</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_written_per_iteration = 16;</div><div class="line"><a name="l01123"></a><span class="lineno"> 1123</span> </div><div class="line"><a name="l01124"></a><span class="lineno"> 1124</span>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win = <a class="code" href="namespacearm__compute.xhtml#ab7980fa5ee693e3282a76da047a1c3b5">calculate_max_window</a>(*input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), <a class="code" href="classarm__compute_1_1_steps.xhtml">Steps</a>(num_elems_processed_per_iteration), border_undefined, border_size());</div><div class="line"><a name="l01125"></a><span class="lineno"> 1125</span>  <a class="code" href="classarm__compute_1_1_access_window_horizontal.xhtml">AccessWindowHorizontal</a> output_access(output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), 0, num_elems_written_per_iteration);</div><div class="line"><a name="l01126"></a><span class="lineno"> 1126</span> </div><div class="line"><a name="l01127"></a><span class="lineno"> 1127</span>  <a class="code" href="namespacearm__compute.xhtml#afc4bd8e872567d9c4c57d89eb0bb3da1">update_window_and_padding</a>(win,</div><div class="line"><a name="l01128"></a><span class="lineno"> 1128</span>  <a class="code" href="classarm__compute_1_1_access_window_rectangle.xhtml">AccessWindowRectangle</a>(input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), 0, -border_size().top, num_elems_read_per_iteration, matrix_size),</div><div class="line"><a name="l01129"></a><span class="lineno"> 1129</span>  output_access);</div><div class="line"><a name="l01130"></a><span class="lineno"> 1130</span> </div><div class="line"><a name="l01131"></a><span class="lineno"> 1131</span>  output_access.set_valid_region(win, input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a74dfd07380a290c34fe7c8e065029b95">valid_region</a>(), border_undefined, border_size());</div><div class="line"><a name="l01132"></a><span class="lineno"> 1132</span> </div><div class="line"><a name="l01133"></a><span class="lineno"> 1133</span>  INEKernel::configure(win);</div><div class="line"><a name="l01134"></a><span class="lineno"> 1134</span> }</div><div class="line"><a name="l01135"></a><span class="lineno"> 1135</span> </div><div class="line"><a name="l01136"></a><span class="lineno"> 1136</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01137"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82"> 1137</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82">NESeparableConvolutionVertKernel<matrix_size>::run</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &window, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_thread_info.xhtml">ThreadInfo</a> &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>)</div><div class="line"><a name="l01138"></a><span class="lineno"> 1138</span> {</div><div class="line"><a name="l01139"></a><span class="lineno"> 1139</span>  <a class="code" href="_error_8h.xhtml#a6dc630a6ae9cc063b3924bcea8dee9d6">ARM_COMPUTE_UNUSED</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>);</div><div class="line"><a name="l01140"></a><span class="lineno"> 1140</span>  <a class="code" href="_validate_8h.xhtml#a1b35b0d258183cf9ef36adf684d0b88c">ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL</a>(<span class="keyword">this</span>);</div><div class="line"><a name="l01141"></a><span class="lineno"> 1141</span>  <a class="code" href="_validate_8h.xhtml#a6eb9ce82815fe429250189da7592ba75">ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW</a>(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">INEKernel::window</a>(), window);</div><div class="line"><a name="l01142"></a><span class="lineno"> 1142</span> </div><div class="line"><a name="l01143"></a><span class="lineno"> 1143</span>  <span class="keywordflow">switch</span>(_input->info()->data_type())</div><div class="line"><a name="l01144"></a><span class="lineno"> 1144</span>  {</div><div class="line"><a name="l01145"></a><span class="lineno"> 1145</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aef9ef3ebca4d2b64b6ec83808bafa5f2">DataType::U16</a>:</div><div class="line"><a name="l01146"></a><span class="lineno"> 1146</span>  <span class="keywordflow">switch</span>(_output->info()->data_type())</div><div class="line"><a name="l01147"></a><span class="lineno"> 1147</span>  {</div><div class="line"><a name="l01148"></a><span class="lineno"> 1148</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>:</div><div class="line"><a name="l01149"></a><span class="lineno"> 1149</span>  convolution_u16<uint8_t>(window);</div><div class="line"><a name="l01150"></a><span class="lineno"> 1150</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01151"></a><span class="lineno"> 1151</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l01152"></a><span class="lineno"> 1152</span>  convolution_u16<int16_t>(window);</div><div class="line"><a name="l01153"></a><span class="lineno"> 1153</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01154"></a><span class="lineno"> 1154</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01155"></a><span class="lineno"> 1155</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported"</span>);</div><div class="line"><a name="l01156"></a><span class="lineno"> 1156</span>  }</div><div class="line"><a name="l01157"></a><span class="lineno"> 1157</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01158"></a><span class="lineno"> 1158</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l01159"></a><span class="lineno"> 1159</span>  <span class="keywordflow">switch</span>(_output->info()->data_type())</div><div class="line"><a name="l01160"></a><span class="lineno"> 1160</span>  {</div><div class="line"><a name="l01161"></a><span class="lineno"> 1161</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>:</div><div class="line"><a name="l01162"></a><span class="lineno"> 1162</span>  convolution_s16<uint8_t>(window);</div><div class="line"><a name="l01163"></a><span class="lineno"> 1163</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01164"></a><span class="lineno"> 1164</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l01165"></a><span class="lineno"> 1165</span>  convolution_s16<int16_t>(window);</div><div class="line"><a name="l01166"></a><span class="lineno"> 1166</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01167"></a><span class="lineno"> 1167</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01168"></a><span class="lineno"> 1168</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported"</span>);</div><div class="line"><a name="l01169"></a><span class="lineno"> 1169</span>  }</div><div class="line"><a name="l01170"></a><span class="lineno"> 1170</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01171"></a><span class="lineno"> 1171</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58aa1e28eee0339658d39a8b4d325b56e9c">DataType::S32</a>:</div><div class="line"><a name="l01172"></a><span class="lineno"> 1172</span>  <span class="keywordflow">switch</span>(_output->info()->data_type())</div><div class="line"><a name="l01173"></a><span class="lineno"> 1173</span>  {</div><div class="line"><a name="l01174"></a><span class="lineno"> 1174</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>:</div><div class="line"><a name="l01175"></a><span class="lineno"> 1175</span>  convolution_s32<uint8_t>(window);</div><div class="line"><a name="l01176"></a><span class="lineno"> 1176</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01177"></a><span class="lineno"> 1177</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l01178"></a><span class="lineno"> 1178</span>  convolution_s32<int16_t>(window);</div><div class="line"><a name="l01179"></a><span class="lineno"> 1179</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01180"></a><span class="lineno"> 1180</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01181"></a><span class="lineno"> 1181</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported"</span>);</div><div class="line"><a name="l01182"></a><span class="lineno"> 1182</span>  }</div><div class="line"><a name="l01183"></a><span class="lineno"> 1183</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01184"></a><span class="lineno"> 1184</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01185"></a><span class="lineno"> 1185</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Unsupported intermediate data type!"</span>);</div><div class="line"><a name="l01186"></a><span class="lineno"> 1186</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01187"></a><span class="lineno"> 1187</span>  }</div><div class="line"><a name="l01188"></a><span class="lineno"> 1188</span> }</div><div class="line"><a name="l01189"></a><span class="lineno"> 1189</span> </div><div class="line"><a name="l01190"></a><span class="lineno"> 1190</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01191"></a><span class="lineno"> 1191</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l01192"></a><span class="lineno"> 1192</span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml">NESeparableConvolutionVertKernel<matrix_size>::convolution_u16</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &win)</div><div class="line"><a name="l01193"></a><span class="lineno"> 1193</span> {</div><div class="line"><a name="l01194"></a><span class="lineno"> 1194</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l01195"></a><span class="lineno"> 1195</span> </div><div class="line"><a name="l01196"></a><span class="lineno"> 1196</span>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win_in(win);</div><div class="line"><a name="l01197"></a><span class="lineno"> 1197</span>  win_in.set_dimension_step(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, 8);</div><div class="line"><a name="l01198"></a><span class="lineno"> 1198</span> </div><div class="line"><a name="l01199"></a><span class="lineno"> 1199</span>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> in(_input, win_in);</div><div class="line"><a name="l01200"></a><span class="lineno"> 1200</span>  <a class="code" href="classarm__compute_1_1_iterator.xhtml">Iterator</a> out(_output, win);</div><div class="line"><a name="l01201"></a><span class="lineno"> 1201</span> </div><div class="line"><a name="l01202"></a><span class="lineno"> 1202</span>  std::array<unsigned char *, matrix_size> input_ptrs{ {} };</div><div class="line"><a name="l01203"></a><span class="lineno"> 1203</span>  <span class="keyword">const</span> float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l01204"></a><span class="lineno"> 1204</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> k_half = matrix_size / 2;</div><div class="line"><a name="l01205"></a><span class="lineno"> 1205</span> </div><div class="line"><a name="l01206"></a><span class="lineno"> 1206</span>  <span class="comment">// Set row pointers</span></div><div class="line"><a name="l01207"></a><span class="lineno"> 1207</span>  <span class="keywordflow">for</span>(<span class="keywordtype">int</span> i = -k_half; i <= k_half; ++i)</div><div class="line"><a name="l01208"></a><span class="lineno"> 1208</span>  {</div><div class="line"><a name="l01209"></a><span class="lineno"> 1209</span>  input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i));</div><div class="line"><a name="l01210"></a><span class="lineno"> 1210</span>  }</div><div class="line"><a name="l01211"></a><span class="lineno"> 1211</span> </div><div class="line"><a name="l01212"></a><span class="lineno"> 1212</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01213"></a><span class="lineno"> 1213</span>  {</div><div class="line"><a name="l01214"></a><span class="lineno"> 1214</span>  uint16x8_t out0 = vdupq_n_u16(0);</div><div class="line"><a name="l01215"></a><span class="lineno"> 1215</span>  uint16x8_t out1 = vdupq_n_u16(0);</div><div class="line"><a name="l01216"></a><span class="lineno"> 1216</span> </div><div class="line"><a name="l01217"></a><span class="lineno"> 1217</span>  <span class="comment">// First half</span></div><div class="line"><a name="l01218"></a><span class="lineno"> 1218</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01219"></a><span class="lineno"> 1219</span>  {</div><div class="line"><a name="l01220"></a><span class="lineno"> 1220</span>  <span class="keyword">const</span> uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01221"></a><span class="lineno"> 1221</span>  out0 = vmlaq_n_u16(out0, data, _conv_col[r]);</div><div class="line"><a name="l01222"></a><span class="lineno"> 1222</span>  }</div><div class="line"><a name="l01223"></a><span class="lineno"> 1223</span> </div><div class="line"><a name="l01224"></a><span class="lineno"> 1224</span>  in.increment(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>);</div><div class="line"><a name="l01225"></a><span class="lineno"> 1225</span> </div><div class="line"><a name="l01226"></a><span class="lineno"> 1226</span>  <span class="comment">// Second half</span></div><div class="line"><a name="l01227"></a><span class="lineno"> 1227</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01228"></a><span class="lineno"> 1228</span>  {</div><div class="line"><a name="l01229"></a><span class="lineno"> 1229</span>  <span class="keyword">const</span> uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01230"></a><span class="lineno"> 1230</span>  out1 = vmlaq_n_u16(out1, data, _conv_col[r]);</div><div class="line"><a name="l01231"></a><span class="lineno"> 1231</span>  }</div><div class="line"><a name="l01232"></a><span class="lineno"> 1232</span> </div><div class="line"><a name="l01233"></a><span class="lineno"> 1233</span>  <span class="comment">//scale the result if needed</span></div><div class="line"><a name="l01234"></a><span class="lineno"> 1234</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l01235"></a><span class="lineno"> 1235</span>  {</div><div class="line"><a name="l01236"></a><span class="lineno"> 1236</span>  float32x4_t out0_f32_high = vcvtq_f32_u32(vmovl_u16(vget_high_u16(out0)));</div><div class="line"><a name="l01237"></a><span class="lineno"> 1237</span>  float32x4_t out0_f32_low = vcvtq_f32_u32(vmovl_u16(vget_low_u16(out0)));</div><div class="line"><a name="l01238"></a><span class="lineno"> 1238</span>  out0_f32_high = vmulq_f32(out0_f32_high, oneoverscale);</div><div class="line"><a name="l01239"></a><span class="lineno"> 1239</span>  out0_f32_low = vmulq_f32(out0_f32_low, oneoverscale);</div><div class="line"><a name="l01240"></a><span class="lineno"> 1240</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(vcvtq_u32_f32(out0_f32_low), vcvtq_u32_f32(out0_f32_high), reinterpret_cast<OutputType *>(out.ptr()));</div><div class="line"><a name="l01241"></a><span class="lineno"> 1241</span> </div><div class="line"><a name="l01242"></a><span class="lineno"> 1242</span>  float32x4_t out1_f32_high = vcvtq_f32_u32(vmovl_u16(vget_high_u16(out1)));</div><div class="line"><a name="l01243"></a><span class="lineno"> 1243</span>  float32x4_t out1_f32_low = vcvtq_f32_u32(vmovl_u16(vget_low_u16(out1)));</div><div class="line"><a name="l01244"></a><span class="lineno"> 1244</span>  out1_f32_high = vmulq_f32(out1_f32_high, oneoverscale);</div><div class="line"><a name="l01245"></a><span class="lineno"> 1245</span>  out1_f32_low = vmulq_f32(out1_f32_low, oneoverscale);</div><div class="line"><a name="l01246"></a><span class="lineno"> 1246</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(vcvtq_u32_f32(out1_f32_low), vcvtq_u32_f32(out1_f32_high), reinterpret_cast<OutputType *>(out.ptr()) + 8);</div><div class="line"><a name="l01247"></a><span class="lineno"> 1247</span>  }</div><div class="line"><a name="l01248"></a><span class="lineno"> 1248</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l01249"></a><span class="lineno"> 1249</span>  {</div><div class="line"><a name="l01250"></a><span class="lineno"> 1250</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out0, out1, reinterpret_cast<OutputType *>(out.ptr()));</div><div class="line"><a name="l01251"></a><span class="lineno"> 1251</span>  }</div><div class="line"><a name="l01252"></a><span class="lineno"> 1252</span>  },</div><div class="line"><a name="l01253"></a><span class="lineno"> 1253</span>  in, out);</div><div class="line"><a name="l01254"></a><span class="lineno"> 1254</span> }</div><div class="line"><a name="l01255"></a><span class="lineno"> 1255</span> </div><div class="line"><a name="l01256"></a><span class="lineno"> 1256</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01257"></a><span class="lineno"> 1257</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l01258"></a><span class="lineno"> 1258</span> <span class="keywordtype">void</span> NESeparableConvolutionVertKernel<matrix_size>::convolution_s16(<span class="keyword">const</span> Window &win)</div><div class="line"><a name="l01259"></a><span class="lineno"> 1259</span> {</div><div class="line"><a name="l01260"></a><span class="lineno"> 1260</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l01261"></a><span class="lineno"> 1261</span> </div><div class="line"><a name="l01262"></a><span class="lineno"> 1262</span>  Window win_in(win);</div><div class="line"><a name="l01263"></a><span class="lineno"> 1263</span>  win_in.set_dimension_step(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, 8);</div><div class="line"><a name="l01264"></a><span class="lineno"> 1264</span> </div><div class="line"><a name="l01265"></a><span class="lineno"> 1265</span>  Iterator in(_input, win_in);</div><div class="line"><a name="l01266"></a><span class="lineno"> 1266</span>  Iterator out(_output, win);</div><div class="line"><a name="l01267"></a><span class="lineno"> 1267</span> </div><div class="line"><a name="l01268"></a><span class="lineno"> 1268</span>  std::array<unsigned char *, matrix_size> input_ptrs{ {} };</div><div class="line"><a name="l01269"></a><span class="lineno"> 1269</span>  <span class="keyword">const</span> float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l01270"></a><span class="lineno"> 1270</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> k_half = matrix_size / 2;</div><div class="line"><a name="l01271"></a><span class="lineno"> 1271</span> </div><div class="line"><a name="l01272"></a><span class="lineno"> 1272</span>  <span class="comment">// Set row pointers</span></div><div class="line"><a name="l01273"></a><span class="lineno"> 1273</span>  <span class="keywordflow">for</span>(<span class="keywordtype">int</span> i = -k_half; i <= k_half; ++i)</div><div class="line"><a name="l01274"></a><span class="lineno"> 1274</span>  {</div><div class="line"><a name="l01275"></a><span class="lineno"> 1275</span>  input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i));</div><div class="line"><a name="l01276"></a><span class="lineno"> 1276</span>  }</div><div class="line"><a name="l01277"></a><span class="lineno"> 1277</span> </div><div class="line"><a name="l01278"></a><span class="lineno"> 1278</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01279"></a><span class="lineno"> 1279</span>  {</div><div class="line"><a name="l01280"></a><span class="lineno"> 1280</span>  int16x8_t out0 = vdupq_n_s16(0);</div><div class="line"><a name="l01281"></a><span class="lineno"> 1281</span>  int16x8_t out1 = vdupq_n_s16(0);</div><div class="line"><a name="l01282"></a><span class="lineno"> 1282</span> </div><div class="line"><a name="l01283"></a><span class="lineno"> 1283</span>  <span class="comment">// First half</span></div><div class="line"><a name="l01284"></a><span class="lineno"> 1284</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01285"></a><span class="lineno"> 1285</span>  {</div><div class="line"><a name="l01286"></a><span class="lineno"> 1286</span>  <span class="keyword">const</span> int16x8_t data = vld1q_s16(reinterpret_cast<const int16_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01287"></a><span class="lineno"> 1287</span>  out0 = vmlaq_n_s16(out0, data, _conv_col[r]);</div><div class="line"><a name="l01288"></a><span class="lineno"> 1288</span>  }</div><div class="line"><a name="l01289"></a><span class="lineno"> 1289</span> </div><div class="line"><a name="l01290"></a><span class="lineno"> 1290</span>  in.increment(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>);</div><div class="line"><a name="l01291"></a><span class="lineno"> 1291</span> </div><div class="line"><a name="l01292"></a><span class="lineno"> 1292</span>  <span class="comment">// Second half</span></div><div class="line"><a name="l01293"></a><span class="lineno"> 1293</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01294"></a><span class="lineno"> 1294</span>  {</div><div class="line"><a name="l01295"></a><span class="lineno"> 1295</span>  <span class="keyword">const</span> int16x8_t data = vld1q_s16(reinterpret_cast<const int16_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01296"></a><span class="lineno"> 1296</span>  out1 = vmlaq_n_s16(out1, data, _conv_col[r]);</div><div class="line"><a name="l01297"></a><span class="lineno"> 1297</span>  }</div><div class="line"><a name="l01298"></a><span class="lineno"> 1298</span> </div><div class="line"><a name="l01299"></a><span class="lineno"> 1299</span>  <span class="comment">//scale the result if needed</span></div><div class="line"><a name="l01300"></a><span class="lineno"> 1300</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l01301"></a><span class="lineno"> 1301</span>  {</div><div class="line"><a name="l01302"></a><span class="lineno"> 1302</span>  float32x4_t out0_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out0)));</div><div class="line"><a name="l01303"></a><span class="lineno"> 1303</span>  float32x4_t out0_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out0)));</div><div class="line"><a name="l01304"></a><span class="lineno"> 1304</span>  out0_f32_high = vmulq_f32(out0_f32_high, oneoverscale);</div><div class="line"><a name="l01305"></a><span class="lineno"> 1305</span>  out0_f32_low = vmulq_f32(out0_f32_low, oneoverscale);</div><div class="line"><a name="l01306"></a><span class="lineno"> 1306</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(vcvtq_s32_f32(out0_f32_low), vcvtq_s32_f32(out0_f32_high), reinterpret_cast<OutputType *>(out.ptr()));</div><div class="line"><a name="l01307"></a><span class="lineno"> 1307</span> </div><div class="line"><a name="l01308"></a><span class="lineno"> 1308</span>  float32x4_t out1_f32_high = vcvtq_f32_s32(vmovl_s16(vget_high_s16(out1)));</div><div class="line"><a name="l01309"></a><span class="lineno"> 1309</span>  float32x4_t out1_f32_low = vcvtq_f32_s32(vmovl_s16(vget_low_s16(out1)));</div><div class="line"><a name="l01310"></a><span class="lineno"> 1310</span>  out1_f32_high = vmulq_f32(out1_f32_high, oneoverscale);</div><div class="line"><a name="l01311"></a><span class="lineno"> 1311</span>  out1_f32_low = vmulq_f32(out1_f32_low, oneoverscale);</div><div class="line"><a name="l01312"></a><span class="lineno"> 1312</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(vcvtq_s32_f32(out1_f32_low), vcvtq_s32_f32(out1_f32_high), reinterpret_cast<OutputType *>(out.ptr()) + 8);</div><div class="line"><a name="l01313"></a><span class="lineno"> 1313</span>  }</div><div class="line"><a name="l01314"></a><span class="lineno"> 1314</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l01315"></a><span class="lineno"> 1315</span>  {</div><div class="line"><a name="l01316"></a><span class="lineno"> 1316</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out0, out1, reinterpret_cast<OutputType *>(out.ptr()));</div><div class="line"><a name="l01317"></a><span class="lineno"> 1317</span>  }</div><div class="line"><a name="l01318"></a><span class="lineno"> 1318</span>  },</div><div class="line"><a name="l01319"></a><span class="lineno"> 1319</span>  in, out);</div><div class="line"><a name="l01320"></a><span class="lineno"> 1320</span> }</div><div class="line"><a name="l01321"></a><span class="lineno"> 1321</span> </div><div class="line"><a name="l01322"></a><span class="lineno"> 1322</span> <span class="keyword">template</span> <<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> matrix_size></div><div class="line"><a name="l01323"></a><span class="lineno"> 1323</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType></div><div class="line"><a name="l01324"></a><span class="lineno"> 1324</span> <span class="keywordtype">void</span> NESeparableConvolutionVertKernel<matrix_size>::convolution_s32(<span class="keyword">const</span> Window &win)</div><div class="line"><a name="l01325"></a><span class="lineno"> 1325</span> {</div><div class="line"><a name="l01326"></a><span class="lineno"> 1326</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l01327"></a><span class="lineno"> 1327</span> </div><div class="line"><a name="l01328"></a><span class="lineno"> 1328</span>  Window win_in(win);</div><div class="line"><a name="l01329"></a><span class="lineno"> 1329</span>  win_in.set_dimension_step(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>, 8);</div><div class="line"><a name="l01330"></a><span class="lineno"> 1330</span> </div><div class="line"><a name="l01331"></a><span class="lineno"> 1331</span>  Iterator in(_input, win_in);</div><div class="line"><a name="l01332"></a><span class="lineno"> 1332</span>  Iterator out(_output, win);</div><div class="line"><a name="l01333"></a><span class="lineno"> 1333</span> </div><div class="line"><a name="l01334"></a><span class="lineno"> 1334</span>  std::array<unsigned char *, matrix_size> input_ptrs{ {} };</div><div class="line"><a name="l01335"></a><span class="lineno"> 1335</span>  <span class="keyword">const</span> float32x4_t oneoverscale = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l01336"></a><span class="lineno"> 1336</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> k_half = matrix_size / 2;</div><div class="line"><a name="l01337"></a><span class="lineno"> 1337</span> </div><div class="line"><a name="l01338"></a><span class="lineno"> 1338</span>  <span class="comment">// Set row pointers</span></div><div class="line"><a name="l01339"></a><span class="lineno"> 1339</span>  <span class="keywordflow">for</span>(<span class="keywordtype">int</span> i = -k_half; i <= k_half; ++i)</div><div class="line"><a name="l01340"></a><span class="lineno"> 1340</span>  {</div><div class="line"><a name="l01341"></a><span class="lineno"> 1341</span>  input_ptrs[k_half + i] = _input->ptr_to_element(Coordinates(0, i));</div><div class="line"><a name="l01342"></a><span class="lineno"> 1342</span>  }</div><div class="line"><a name="l01343"></a><span class="lineno"> 1343</span> </div><div class="line"><a name="l01344"></a><span class="lineno"> 1344</span>  <span class="keyword">const</span> int32x4_t zero = vdupq_n_s32(0);</div><div class="line"><a name="l01345"></a><span class="lineno"> 1345</span> </div><div class="line"><a name="l01346"></a><span class="lineno"> 1346</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01347"></a><span class="lineno"> 1347</span>  {</div><div class="line"><a name="l01348"></a><span class="lineno"> 1348</span>  int32x4x2_t out0 =</div><div class="line"><a name="l01349"></a><span class="lineno"> 1349</span>  {</div><div class="line"><a name="l01350"></a><span class="lineno"> 1350</span>  {</div><div class="line"><a name="l01351"></a><span class="lineno"> 1351</span>  zero,</div><div class="line"><a name="l01352"></a><span class="lineno"> 1352</span>  zero</div><div class="line"><a name="l01353"></a><span class="lineno"> 1353</span>  }</div><div class="line"><a name="l01354"></a><span class="lineno"> 1354</span>  };</div><div class="line"><a name="l01355"></a><span class="lineno"> 1355</span> </div><div class="line"><a name="l01356"></a><span class="lineno"> 1356</span>  int32x4x2_t out1 =</div><div class="line"><a name="l01357"></a><span class="lineno"> 1357</span>  {</div><div class="line"><a name="l01358"></a><span class="lineno"> 1358</span>  {</div><div class="line"><a name="l01359"></a><span class="lineno"> 1359</span>  zero,</div><div class="line"><a name="l01360"></a><span class="lineno"> 1360</span>  zero</div><div class="line"><a name="l01361"></a><span class="lineno"> 1361</span>  }</div><div class="line"><a name="l01362"></a><span class="lineno"> 1362</span>  };</div><div class="line"><a name="l01363"></a><span class="lineno"> 1363</span> </div><div class="line"><a name="l01364"></a><span class="lineno"> 1364</span>  <span class="comment">// First half</span></div><div class="line"><a name="l01365"></a><span class="lineno"> 1365</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01366"></a><span class="lineno"> 1366</span>  {</div><div class="line"><a name="l01367"></a><span class="lineno"> 1367</span>  <span class="keyword">const</span> int32x4x2_t data = vld2q_s32(reinterpret_cast<const int32_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01368"></a><span class="lineno"> 1368</span>  out0.val[0] = vmlaq_n_s32(out0.val[0], data.val[0], _conv_col[r]);</div><div class="line"><a name="l01369"></a><span class="lineno"> 1369</span>  out0.val[1] = vmlaq_n_s32(out0.val[1], data.val[1], _conv_col[r]);</div><div class="line"><a name="l01370"></a><span class="lineno"> 1370</span>  }</div><div class="line"><a name="l01371"></a><span class="lineno"> 1371</span> </div><div class="line"><a name="l01372"></a><span class="lineno"> 1372</span>  in.increment(<a class="code" href="classarm__compute_1_1_window.xhtml#aa96e81276ee4f87ab386cd05a5539a7d">Window::DimX</a>);</div><div class="line"><a name="l01373"></a><span class="lineno"> 1373</span> </div><div class="line"><a name="l01374"></a><span class="lineno"> 1374</span>  <span class="comment">// Second half</span></div><div class="line"><a name="l01375"></a><span class="lineno"> 1375</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < matrix_size; ++r)</div><div class="line"><a name="l01376"></a><span class="lineno"> 1376</span>  {</div><div class="line"><a name="l01377"></a><span class="lineno"> 1377</span>  <span class="keyword">const</span> int32x4x2_t data = vld2q_s32(reinterpret_cast<const int32_t *>(input_ptrs[r] + in.offset()));</div><div class="line"><a name="l01378"></a><span class="lineno"> 1378</span>  out1.val[0] = vmlaq_n_s32(out1.val[0], data.val[0], _conv_col[r]);</div><div class="line"><a name="l01379"></a><span class="lineno"> 1379</span>  out1.val[1] = vmlaq_n_s32(out1.val[1], data.val[1], _conv_col[r]);</div><div class="line"><a name="l01380"></a><span class="lineno"> 1380</span>  }</div><div class="line"><a name="l01381"></a><span class="lineno"> 1381</span> </div><div class="line"><a name="l01382"></a><span class="lineno"> 1382</span>  <span class="comment">//scale the result if needed</span></div><div class="line"><a name="l01383"></a><span class="lineno"> 1383</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l01384"></a><span class="lineno"> 1384</span>  {</div><div class="line"><a name="l01385"></a><span class="lineno"> 1385</span>  float32x4_t out0_f32_odd = vcvtq_f32_s32(out0.val[0]);</div><div class="line"><a name="l01386"></a><span class="lineno"> 1386</span>  float32x4_t out0_f32_even = vcvtq_f32_s32(out0.val[1]);</div><div class="line"><a name="l01387"></a><span class="lineno"> 1387</span>  out0_f32_odd = vmulq_f32(out0_f32_odd, oneoverscale);</div><div class="line"><a name="l01388"></a><span class="lineno"> 1388</span>  out0_f32_even = vmulq_f32(out0_f32_even, oneoverscale);</div><div class="line"><a name="l01389"></a><span class="lineno"> 1389</span>  out0.val[0] = vcvtq_s32_f32(out0_f32_odd);</div><div class="line"><a name="l01390"></a><span class="lineno"> 1390</span>  out0.val[1] = vcvtq_s32_f32(out0_f32_even);</div><div class="line"><a name="l01391"></a><span class="lineno"> 1391</span> </div><div class="line"><a name="l01392"></a><span class="lineno"> 1392</span>  float32x4_t out1_f32_odd = vcvtq_f32_s32(out1.val[0]);</div><div class="line"><a name="l01393"></a><span class="lineno"> 1393</span>  float32x4_t out1_f32_even = vcvtq_f32_s32(out1.val[1]);</div><div class="line"><a name="l01394"></a><span class="lineno"> 1394</span>  out1_f32_odd = vmulq_f32(out1_f32_odd, oneoverscale);</div><div class="line"><a name="l01395"></a><span class="lineno"> 1395</span>  out1_f32_even = vmulq_f32(out1_f32_even, oneoverscale);</div><div class="line"><a name="l01396"></a><span class="lineno"> 1396</span>  out1.val[0] = vcvtq_s32_f32(out1_f32_odd);</div><div class="line"><a name="l01397"></a><span class="lineno"> 1397</span>  out1.val[1] = vcvtq_s32_f32(out1_f32_even);</div><div class="line"><a name="l01398"></a><span class="lineno"> 1398</span>  }</div><div class="line"><a name="l01399"></a><span class="lineno"> 1399</span> </div><div class="line"><a name="l01400"></a><span class="lineno"> 1400</span>  <span class="keyword">const</span> int32x4x2_t out0_s32 = vzipq_s32(out0.val[0], out0.val[1]);</div><div class="line"><a name="l01401"></a><span class="lineno"> 1401</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out0_s32.val[0], out0_s32.val[1], reinterpret_cast<OutputType *>(out.ptr()));</div><div class="line"><a name="l01402"></a><span class="lineno"> 1402</span> </div><div class="line"><a name="l01403"></a><span class="lineno"> 1403</span>  <span class="keyword">const</span> int32x4x2_t out1_s32 = vzipq_s32(out1.val[0], out1.val[1]);</div><div class="line"><a name="l01404"></a><span class="lineno"> 1404</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out1_s32.val[0], out1_s32.val[1], reinterpret_cast<OutputType *>(out.ptr()) + 8);</div><div class="line"><a name="l01405"></a><span class="lineno"> 1405</span>  },</div><div class="line"><a name="l01406"></a><span class="lineno"> 1406</span>  in, out);</div><div class="line"><a name="l01407"></a><span class="lineno"> 1407</span> }</div><div class="line"><a name="l01408"></a><span class="lineno"> 1408</span> </div><div class="line"><a name="l01409"></a><span class="lineno"> 1409</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml">arm_compute::NESeparableConvolutionVertKernel<5></a>;</div><div class="line"><a name="l01410"></a><span class="lineno"> 1410</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml">arm_compute::NESeparableConvolutionVertKernel<7></a>;</div><div class="line"><a name="l01411"></a><span class="lineno"> 1411</span> <span class="keyword">template</span> <span class="keyword">class </span><a class="code" href="classarm__compute_1_1_n_e_separable_convolution_vert_kernel.xhtml">arm_compute::NESeparableConvolutionVertKernel<9></a>;</div><div class="line"><a name="l01412"></a><span class="lineno"> 1412</span> </div><div class="line"><a name="l01413"></a><span class="lineno"> 1413</span> <span class="comment">/****************************************************************************************\</span></div><div class="line"><a name="l01414"></a><span class="lineno"> 1414</span> <span class="comment"> * Rectangle Convolution *</span></div><div class="line"><a name="l01415"></a><span class="lineno"> 1415</span> <span class="comment">\****************************************************************************************/</span></div><div class="line"><a name="l01416"></a><span class="lineno"> 1416</span> </div><div class="line"><a name="l01417"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a437bfa972bdb089215368c344cce05d3"> 1417</a></span> <a class="code" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a437bfa972bdb089215368c344cce05d3">NEConvolutionRectangleKernel::NEConvolutionRectangleKernel</a>()</div><div class="line"><a name="l01418"></a><span class="lineno"> 1418</span>  : _input(nullptr), _output(nullptr), _scale(0), _convolution(), _border_size(), _func_idx(0)</div><div class="line"><a name="l01419"></a><span class="lineno"> 1419</span> {</div><div class="line"><a name="l01420"></a><span class="lineno"> 1420</span> }</div><div class="line"><a name="l01421"></a><span class="lineno"> 1421</span> </div><div class="line"><a name="l01422"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7"> 1422</a></span> <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a> <a class="code" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a423f9a45a52983b4de5e2b347f4369c7">NEConvolutionRectangleKernel::border_size</a>()<span class="keyword"> const</span></div><div class="line"><a name="l01423"></a><span class="lineno"> 1423</span> <span class="keyword"></span>{</div><div class="line"><a name="l01424"></a><span class="lineno"> 1424</span>  <span class="keywordflow">return</span> _border_size;</div><div class="line"><a name="l01425"></a><span class="lineno"> 1425</span> }</div><div class="line"><a name="l01426"></a><span class="lineno"> 1426</span> </div><div class="line"><a name="l01427"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a3bb75d0640105ec72239550ec8a086b9"> 1427</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a3bb75d0640105ec72239550ec8a086b9">NEConvolutionRectangleKernel::configure</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *input, <a class="code" href="classarm__compute_1_1_i_tensor.xhtml">ITensor</a> *output, <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>, uint32_t width, uint32_t height, uint32_t <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>, <span class="keywordtype">bool</span> border_undefined)</div><div class="line"><a name="l01428"></a><span class="lineno"> 1428</span> {</div><div class="line"><a name="l01429"></a><span class="lineno"> 1429</span>  <a class="code" href="_validate_8h.xhtml#a921b705e9e3e0fe928928447869e62a5">ARM_COMPUTE_ERROR_ON_NULLPTR</a>(input, output, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>);</div><div class="line"><a name="l01430"></a><span class="lineno"> 1430</span> </div><div class="line"><a name="l01431"></a><span class="lineno"> 1431</span>  <a class="code" href="namespacearm__compute.xhtml#a9d1a839c51134b2ae171a2264c541b6f">set_shape_if_empty</a>(*output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a7c66505457d00ece3aa4b34cab80757d">tensor_shape</a>());</div><div class="line"><a name="l01432"></a><span class="lineno"> 1432</span> </div><div class="line"><a name="l01433"></a><span class="lineno"> 1433</span>  <a class="code" href="_validate_8h.xhtml#aafce5e9928fac427dcd29d960bcf2ae3">ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES</a>(input, output);</div><div class="line"><a name="l01434"></a><span class="lineno"> 1434</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(input, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>);</div><div class="line"><a name="l01435"></a><span class="lineno"> 1435</span>  <a class="code" href="_validate_8h.xhtml#aadf5c9cff86327b96d88d04649d9715e">ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN</a>(output, 1, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>, <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>);</div><div class="line"><a name="l01436"></a><span class="lineno"> 1436</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(width != 3 && width != 5 && width != 7 && width != 9);</div><div class="line"><a name="l01437"></a><span class="lineno"> 1437</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(height != 3 && height != 5 && height != 7 && height != 9);</div><div class="line"><a name="l01438"></a><span class="lineno"> 1438</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(0 == <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l01439"></a><span class="lineno"> 1439</span> </div><div class="line"><a name="l01440"></a><span class="lineno"> 1440</span>  _input = input;</div><div class="line"><a name="l01441"></a><span class="lineno"> 1441</span>  _output = output;</div><div class="line"><a name="l01442"></a><span class="lineno"> 1442</span>  _scale = <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>;</div><div class="line"><a name="l01443"></a><span class="lineno"> 1443</span>  _border_size = <a class="code" href="structarm__compute_1_1_border_size.xhtml">BorderSize</a>(height / 2, width / 2);</div><div class="line"><a name="l01444"></a><span class="lineno"> 1444</span> </div><div class="line"><a name="l01445"></a><span class="lineno"> 1445</span>  <span class="comment">// Setup the convolution matrix</span></div><div class="line"><a name="l01446"></a><span class="lineno"> 1446</span>  <span class="keyword">const</span> uint32_t nr_elements = width * height;</div><div class="line"><a name="l01447"></a><span class="lineno"> 1447</span>  _convolution.resize(nr_elements);</div><div class="line"><a name="l01448"></a><span class="lineno"> 1448</span>  std::copy_n(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a>, nr_elements, _convolution.begin());</div><div class="line"><a name="l01449"></a><span class="lineno"> 1449</span> </div><div class="line"><a name="l01450"></a><span class="lineno"> 1450</span>  <span class="comment">// Set function index to help choose appropriate function in run()</span></div><div class="line"><a name="l01451"></a><span class="lineno"> 1451</span>  _func_idx = get_index(height) * 4 + get_index(width);</div><div class="line"><a name="l01452"></a><span class="lineno"> 1452</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_func_idx > (_nr_supported_sizes * _nr_supported_sizes));</div><div class="line"><a name="l01453"></a><span class="lineno"> 1453</span> </div><div class="line"><a name="l01454"></a><span class="lineno"> 1454</span>  <span class="comment">// Configure kernel window</span></div><div class="line"><a name="l01455"></a><span class="lineno"> 1455</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_processed_per_iteration = 8;</div><div class="line"><a name="l01456"></a><span class="lineno"> 1456</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_read_per_iteration = 16;</div><div class="line"><a name="l01457"></a><span class="lineno"> 1457</span>  constexpr <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> num_elems_written_per_iteration = 8;</div><div class="line"><a name="l01458"></a><span class="lineno"> 1458</span> </div><div class="line"><a name="l01459"></a><span class="lineno"> 1459</span>  <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> win = <a class="code" href="namespacearm__compute.xhtml#ab7980fa5ee693e3282a76da047a1c3b5">calculate_max_window</a>(*input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), <a class="code" href="classarm__compute_1_1_steps.xhtml">Steps</a>(num_elems_processed_per_iteration), border_undefined, _border_size);</div><div class="line"><a name="l01460"></a><span class="lineno"> 1460</span>  <a class="code" href="classarm__compute_1_1_access_window_horizontal.xhtml">AccessWindowHorizontal</a> output_access(output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), 0, num_elems_written_per_iteration);</div><div class="line"><a name="l01461"></a><span class="lineno"> 1461</span> </div><div class="line"><a name="l01462"></a><span class="lineno"> 1462</span>  <a class="code" href="namespacearm__compute.xhtml#afc4bd8e872567d9c4c57d89eb0bb3da1">update_window_and_padding</a>(win,</div><div class="line"><a name="l01463"></a><span class="lineno"> 1463</span>  <a class="code" href="classarm__compute_1_1_access_window_rectangle.xhtml">AccessWindowRectangle</a>(input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>(), -_border_size.<a class="code" href="structarm__compute_1_1_border_size.xhtml#a05374b750b0fc472c34ee61e6f028bba">left</a>, -_border_size.<a class="code" href="structarm__compute_1_1_border_size.xhtml#a3fdd42ea34070a54e696b3adc28c4be3">top</a>, num_elems_read_per_iteration, height),</div><div class="line"><a name="l01464"></a><span class="lineno"> 1464</span>  output_access);</div><div class="line"><a name="l01465"></a><span class="lineno"> 1465</span> </div><div class="line"><a name="l01466"></a><span class="lineno"> 1466</span>  output_access.set_valid_region(win, input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a74dfd07380a290c34fe7c8e065029b95">valid_region</a>(), border_undefined, _border_size);</div><div class="line"><a name="l01467"></a><span class="lineno"> 1467</span> </div><div class="line"><a name="l01468"></a><span class="lineno"> 1468</span>  INEKernel::configure(win);</div><div class="line"><a name="l01469"></a><span class="lineno"> 1469</span> }</div><div class="line"><a name="l01470"></a><span class="lineno"> 1470</span> </div><div class="line"><a name="l01471"></a><span class="lineno"><a class="line" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82"> 1471</a></span> <span class="keywordtype">void</span> <a class="code" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml#a112b35dd205c62ea6ed1447ef226da82">NEConvolutionRectangleKernel::run</a>(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> &window, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_thread_info.xhtml">ThreadInfo</a> &<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>)</div><div class="line"><a name="l01472"></a><span class="lineno"> 1472</span> {</div><div class="line"><a name="l01473"></a><span class="lineno"> 1473</span>  <a class="code" href="_error_8h.xhtml#a6dc630a6ae9cc063b3924bcea8dee9d6">ARM_COMPUTE_UNUSED</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a4f4125dba5283887b34f889b1c615c0c">info</a>);</div><div class="line"><a name="l01474"></a><span class="lineno"> 1474</span>  <a class="code" href="_validate_8h.xhtml#a1b35b0d258183cf9ef36adf684d0b88c">ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL</a>(<span class="keyword">this</span>);</div><div class="line"><a name="l01475"></a><span class="lineno"> 1475</span>  <a class="code" href="_validate_8h.xhtml#a6eb9ce82815fe429250189da7592ba75">ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW</a>(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">INEKernel::window</a>(), <a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">window</a>);</div><div class="line"><a name="l01476"></a><span class="lineno"> 1476</span> </div><div class="line"><a name="l01477"></a><span class="lineno"> 1477</span>  <span class="keyword">using</span> ConvolutionRectangleFunction = void (<a class="code" href="classarm__compute_1_1_n_e_convolution_rectangle_kernel.xhtml">NEConvolutionRectangleKernel</a>::*)(<span class="keyword">const</span> <a class="code" href="classarm__compute_1_1_window.xhtml">Window</a> & <a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">window</a>);</div><div class="line"><a name="l01478"></a><span class="lineno"> 1478</span> </div><div class="line"><a name="l01479"></a><span class="lineno"> 1479</span>  <span class="comment">// uint8_t function table</span></div><div class="line"><a name="l01480"></a><span class="lineno"> 1480</span>  <span class="keyword">static</span> <span class="keyword">const</span> std::array<ConvolutionRectangleFunction, 16> func_table_u8 =</div><div class="line"><a name="l01481"></a><span class="lineno"> 1481</span>  {</div><div class="line"><a name="l01482"></a><span class="lineno"> 1482</span>  {</div><div class="line"><a name="l01483"></a><span class="lineno"> 1483</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 3>,</div><div class="line"><a name="l01484"></a><span class="lineno"> 1484</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 5>,</div><div class="line"><a name="l01485"></a><span class="lineno"> 1485</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 7>,</div><div class="line"><a name="l01486"></a><span class="lineno"> 1486</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 3, 9>,</div><div class="line"><a name="l01487"></a><span class="lineno"> 1487</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 3>,</div><div class="line"><a name="l01488"></a><span class="lineno"> 1488</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 5>,</div><div class="line"><a name="l01489"></a><span class="lineno"> 1489</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 7>,</div><div class="line"><a name="l01490"></a><span class="lineno"> 1490</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 5, 9>,</div><div class="line"><a name="l01491"></a><span class="lineno"> 1491</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 3>,</div><div class="line"><a name="l01492"></a><span class="lineno"> 1492</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 5>,</div><div class="line"><a name="l01493"></a><span class="lineno"> 1493</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 7>,</div><div class="line"><a name="l01494"></a><span class="lineno"> 1494</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 7, 9>,</div><div class="line"><a name="l01495"></a><span class="lineno"> 1495</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 3>,</div><div class="line"><a name="l01496"></a><span class="lineno"> 1496</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 5>,</div><div class="line"><a name="l01497"></a><span class="lineno"> 1497</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 7>,</div><div class="line"><a name="l01498"></a><span class="lineno"> 1498</span>  &NEConvolutionRectangleKernel::convolution<uint8_t, 9, 9></div><div class="line"><a name="l01499"></a><span class="lineno"> 1499</span>  }</div><div class="line"><a name="l01500"></a><span class="lineno"> 1500</span>  };</div><div class="line"><a name="l01501"></a><span class="lineno"> 1501</span>  <span class="comment">// int16_t function table</span></div><div class="line"><a name="l01502"></a><span class="lineno"> 1502</span>  <span class="keyword">static</span> <span class="keyword">const</span> std::array<ConvolutionRectangleFunction, 16> func_table_s16 =</div><div class="line"><a name="l01503"></a><span class="lineno"> 1503</span>  {</div><div class="line"><a name="l01504"></a><span class="lineno"> 1504</span>  {</div><div class="line"><a name="l01505"></a><span class="lineno"> 1505</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 3, 3>,</div><div class="line"><a name="l01506"></a><span class="lineno"> 1506</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 3, 5>,</div><div class="line"><a name="l01507"></a><span class="lineno"> 1507</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 3, 7>,</div><div class="line"><a name="l01508"></a><span class="lineno"> 1508</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 3, 9>,</div><div class="line"><a name="l01509"></a><span class="lineno"> 1509</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 5, 3>,</div><div class="line"><a name="l01510"></a><span class="lineno"> 1510</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 5, 5>,</div><div class="line"><a name="l01511"></a><span class="lineno"> 1511</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 5, 7>,</div><div class="line"><a name="l01512"></a><span class="lineno"> 1512</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 5, 9>,</div><div class="line"><a name="l01513"></a><span class="lineno"> 1513</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 7, 3>,</div><div class="line"><a name="l01514"></a><span class="lineno"> 1514</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 7, 5>,</div><div class="line"><a name="l01515"></a><span class="lineno"> 1515</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 7, 7>,</div><div class="line"><a name="l01516"></a><span class="lineno"> 1516</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 7, 9>,</div><div class="line"><a name="l01517"></a><span class="lineno"> 1517</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 9, 3>,</div><div class="line"><a name="l01518"></a><span class="lineno"> 1518</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 9, 5>,</div><div class="line"><a name="l01519"></a><span class="lineno"> 1519</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 9, 7>,</div><div class="line"><a name="l01520"></a><span class="lineno"> 1520</span>  &NEConvolutionRectangleKernel::convolution<int16_t, 9, 9></div><div class="line"><a name="l01521"></a><span class="lineno"> 1521</span>  }</div><div class="line"><a name="l01522"></a><span class="lineno"> 1522</span>  };</div><div class="line"><a name="l01523"></a><span class="lineno"> 1523</span> </div><div class="line"><a name="l01524"></a><span class="lineno"> 1524</span>  <span class="comment">// Run appropriate function</span></div><div class="line"><a name="l01525"></a><span class="lineno"> 1525</span>  <span class="keywordflow">switch</span>(_output-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a7cfb31af63202568efef5214acfbf3ba">data_type</a>())</div><div class="line"><a name="l01526"></a><span class="lineno"> 1526</span>  {</div><div class="line"><a name="l01527"></a><span class="lineno"> 1527</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6669348b484e3008dca2bfa8e85e40b5">DataType::U8</a>:</div><div class="line"><a name="l01528"></a><span class="lineno"> 1528</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_func_idx >= func_table_u8.size());</div><div class="line"><a name="l01529"></a><span class="lineno"> 1529</span>  (this->*func_table_u8[_func_idx])(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">window</a>);</div><div class="line"><a name="l01530"></a><span class="lineno"> 1530</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01531"></a><span class="lineno"> 1531</span>  <span class="keywordflow">case</span> <a class="code" href="namespacearm__compute.xhtml#ab4e88c89b3b7ea1735996cc4def22d58a6e0b0886efb94aec797f6b830329b72c">DataType::S16</a>:</div><div class="line"><a name="l01532"></a><span class="lineno"> 1532</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_func_idx >= func_table_s16.size());</div><div class="line"><a name="l01533"></a><span class="lineno"> 1533</span>  (this->*func_table_s16[_func_idx])(<a class="code" href="classarm__compute_1_1_i_kernel.xhtml#ad34a46f53686c12a5c5e717cc9617fb6">window</a>);</div><div class="line"><a name="l01534"></a><span class="lineno"> 1534</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l01535"></a><span class="lineno"> 1535</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01536"></a><span class="lineno"> 1536</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported"</span>);</div><div class="line"><a name="l01537"></a><span class="lineno"> 1537</span>  }</div><div class="line"><a name="l01538"></a><span class="lineno"> 1538</span> }</div><div class="line"><a name="l01539"></a><span class="lineno"> 1539</span> </div><div class="line"><a name="l01540"></a><span class="lineno"> 1540</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> NEConvolutionRectangleKernel::get_index(uint32_t val)</div><div class="line"><a name="l01541"></a><span class="lineno"> 1541</span> {</div><div class="line"><a name="l01542"></a><span class="lineno"> 1542</span>  <span class="keywordflow">switch</span>(val)</div><div class="line"><a name="l01543"></a><span class="lineno"> 1543</span>  {</div><div class="line"><a name="l01544"></a><span class="lineno"> 1544</span>  <span class="keywordflow">case</span> 3:</div><div class="line"><a name="l01545"></a><span class="lineno"> 1545</span>  <span class="keywordflow">return</span> 0;</div><div class="line"><a name="l01546"></a><span class="lineno"> 1546</span>  <span class="keywordflow">case</span> 5:</div><div class="line"><a name="l01547"></a><span class="lineno"> 1547</span>  <span class="keywordflow">return</span> 1;</div><div class="line"><a name="l01548"></a><span class="lineno"> 1548</span>  <span class="keywordflow">case</span> 7:</div><div class="line"><a name="l01549"></a><span class="lineno"> 1549</span>  <span class="keywordflow">return</span> 2;</div><div class="line"><a name="l01550"></a><span class="lineno"> 1550</span>  <span class="keywordflow">case</span> 9:</div><div class="line"><a name="l01551"></a><span class="lineno"> 1551</span>  <span class="keywordflow">return</span> 3;</div><div class="line"><a name="l01552"></a><span class="lineno"> 1552</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l01553"></a><span class="lineno"> 1553</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Not supported dimension size"</span>);</div><div class="line"><a name="l01554"></a><span class="lineno"> 1554</span>  <span class="keywordflow">return</span> 0;</div><div class="line"><a name="l01555"></a><span class="lineno"> 1555</span>  }</div><div class="line"><a name="l01556"></a><span class="lineno"> 1556</span> }</div><div class="line"><a name="l01557"></a><span class="lineno"> 1557</span> </div><div class="line"><a name="l01558"></a><span class="lineno"> 1558</span> <span class="keyword">template</span> <<span class="keyword">typename</span> OutputType, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> rows, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> cols></div><div class="line"><a name="l01559"></a><span class="lineno"> 1559</span> <span class="keywordtype">void</span> NEConvolutionRectangleKernel::convolution(<span class="keyword">const</span> Window &win)</div><div class="line"><a name="l01560"></a><span class="lineno"> 1560</span> {</div><div class="line"><a name="l01561"></a><span class="lineno"> 1561</span>  static_assert(<span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(uint8_t) || <span class="keyword">sizeof</span>(OutputType) == <span class="keyword">sizeof</span>(int16_t), <span class="stringliteral">"The output buffer can only be u8 or s16"</span>);</div><div class="line"><a name="l01562"></a><span class="lineno"> 1562</span>  <a class="code" href="_error_8h.xhtml#a54a6080c9f4df1f908e57a9bbb46f5da">ARM_COMPUTE_ERROR_ON</a>(_input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#ab988210662dbd3bf32fd563c7dd1bdbf">buffer</a>() == <span class="keyword">nullptr</span>);</div><div class="line"><a name="l01563"></a><span class="lineno"> 1563</span> </div><div class="line"><a name="l01564"></a><span class="lineno"> 1564</span>  Iterator input(_input, win);</div><div class="line"><a name="l01565"></a><span class="lineno"> 1565</span>  Iterator output(_output, win);</div><div class="line"><a name="l01566"></a><span class="lineno"> 1566</span> </div><div class="line"><a name="l01567"></a><span class="lineno"> 1567</span>  std::array<unsigned char *, rows> input_ptrs{ {} };</div><div class="line"><a name="l01568"></a><span class="lineno"> 1568</span>  <span class="keyword">const</span> int16_t *<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a> = _convolution.data();</div><div class="line"><a name="l01569"></a><span class="lineno"> 1569</span>  <span class="keyword">const</span> float32x4_t scale_val = vdupq_n_f32(1.0f / _scale);</div><div class="line"><a name="l01570"></a><span class="lineno"> 1570</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> k_row_half = rows / 2;</div><div class="line"><a name="l01571"></a><span class="lineno"> 1571</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> k_col_half = cols / 2;</div><div class="line"><a name="l01572"></a><span class="lineno"> 1572</span> </div><div class="line"><a name="l01573"></a><span class="lineno"> 1573</span>  <span class="comment">// Set row pointers</span></div><div class="line"><a name="l01574"></a><span class="lineno"> 1574</span>  <span class="keywordflow">for</span>(<span class="keywordtype">int</span> i = -k_row_half; i <= k_row_half; ++i)</div><div class="line"><a name="l01575"></a><span class="lineno"> 1575</span>  {</div><div class="line"><a name="l01576"></a><span class="lineno"> 1576</span>  input_ptrs[k_row_half + i] = _input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#ab988210662dbd3bf32fd563c7dd1bdbf">buffer</a>() + _input-><a class="code" href="classarm__compute_1_1_i_tensor.xhtml#a0e95dc1e53c361348314873b168ae237">info</a>()-><a class="code" href="classarm__compute_1_1_i_tensor_info.xhtml#a8a5312c9a822d27a3b38dc238c1c7de3">offset_element_in_bytes</a>(Coordinates(-k_col_half, i));</div><div class="line"><a name="l01577"></a><span class="lineno"> 1577</span>  }</div><div class="line"><a name="l01578"></a><span class="lineno"> 1578</span> </div><div class="line"><a name="l01579"></a><span class="lineno"> 1579</span>  <a class="code" href="namespacearm__compute.xhtml#a5002bf7ec46d52971f9526e94172cfee">execute_window_loop</a>(win, [&](<span class="keyword">const</span> Coordinates &)</div><div class="line"><a name="l01580"></a><span class="lineno"> 1580</span>  {</div><div class="line"><a name="l01581"></a><span class="lineno"> 1581</span>  int32x4_t out = vdupq_n_s32(0);</div><div class="line"><a name="l01582"></a><span class="lineno"> 1582</span>  int32x4_t out2 = vdupq_n_s32(0);</div><div class="line"><a name="l01583"></a><span class="lineno"> 1583</span> </div><div class="line"><a name="l01584"></a><span class="lineno"> 1584</span>  <span class="comment">// Perform appropriate convolution</span></div><div class="line"><a name="l01585"></a><span class="lineno"> 1585</span>  <span class="keywordflow">for</span>(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r = 0; r < rows; ++r)</div><div class="line"><a name="l01586"></a><span class="lineno"> 1586</span>  {</div><div class="line"><a name="l01587"></a><span class="lineno"> 1587</span>  <span class="keyword">const</span> uint8x16_t data = vld1q_u8(input_ptrs[r] + input.offset());</div><div class="line"><a name="l01588"></a><span class="lineno"> 1588</span>  <span class="keywordflow">if</span>(3 == cols)</div><div class="line"><a name="l01589"></a><span class="lineno"> 1589</span>  {</div><div class="line"><a name="l01590"></a><span class="lineno"> 1590</span>  convolve_row3x1(out, out2, data, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a> + r * cols);</div><div class="line"><a name="l01591"></a><span class="lineno"> 1591</span>  }</div><div class="line"><a name="l01592"></a><span class="lineno"> 1592</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(5 == cols)</div><div class="line"><a name="l01593"></a><span class="lineno"> 1593</span>  {</div><div class="line"><a name="l01594"></a><span class="lineno"> 1594</span>  convolve_row5x1(out, out2, data, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a> + r * cols);</div><div class="line"><a name="l01595"></a><span class="lineno"> 1595</span>  }</div><div class="line"><a name="l01596"></a><span class="lineno"> 1596</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(7 == cols)</div><div class="line"><a name="l01597"></a><span class="lineno"> 1597</span>  {</div><div class="line"><a name="l01598"></a><span class="lineno"> 1598</span>  convolve_row7x1(out, out2, data, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a> + r * cols);</div><div class="line"><a name="l01599"></a><span class="lineno"> 1599</span>  }</div><div class="line"><a name="l01600"></a><span class="lineno"> 1600</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(9 == cols)</div><div class="line"><a name="l01601"></a><span class="lineno"> 1601</span>  {</div><div class="line"><a name="l01602"></a><span class="lineno"> 1602</span>  convolve_row9x1(out, out2, data, <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a006546051719c5fb4b20c966a26b9c76">conv</a> + r * cols);</div><div class="line"><a name="l01603"></a><span class="lineno"> 1603</span>  }</div><div class="line"><a name="l01604"></a><span class="lineno"> 1604</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l01605"></a><span class="lineno"> 1605</span>  {</div><div class="line"><a name="l01606"></a><span class="lineno"> 1606</span>  <a class="code" href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a>(<span class="stringliteral">"Unsupported number of columns"</span>);</div><div class="line"><a name="l01607"></a><span class="lineno"> 1607</span>  }</div><div class="line"><a name="l01608"></a><span class="lineno"> 1608</span>  }</div><div class="line"><a name="l01609"></a><span class="lineno"> 1609</span> </div><div class="line"><a name="l01610"></a><span class="lineno"> 1610</span>  <span class="comment">// Apply scale</span></div><div class="line"><a name="l01611"></a><span class="lineno"> 1611</span>  <span class="keywordflow">if</span>(_scale != 1)</div><div class="line"><a name="l01612"></a><span class="lineno"> 1612</span>  {</div><div class="line"><a name="l01613"></a><span class="lineno"> 1613</span>  <span class="comment">// Convert to F32, scale and convert back to S32</span></div><div class="line"><a name="l01614"></a><span class="lineno"> 1614</span>  out = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out), scale_val));</div><div class="line"><a name="l01615"></a><span class="lineno"> 1615</span>  out2 = vcvtq_s32_f32(vmulq_f32(vcvtq_f32_s32(out2), scale_val));</div><div class="line"><a name="l01616"></a><span class="lineno"> 1616</span>  }</div><div class="line"><a name="l01617"></a><span class="lineno"> 1617</span> </div><div class="line"><a name="l01618"></a><span class="lineno"> 1618</span>  <span class="comment">// Clamp and store as U8 or S16:</span></div><div class="line"><a name="l01619"></a><span class="lineno"> 1619</span>  <a class="code" href="namespacearm__compute_1_1detail.xhtml#a1ac67082e6f68cb7ed1939240fb6bdb5">store_results</a>(out, out2, reinterpret_cast<OutputType *>(output.ptr()));</div><div class="line"><a name="l01620"></a><span class="lineno"> 1620</span>  },</div><div class="line"><a name="l01621"></a><span class="lineno"> 1621</span>  input, output);</div><div class="line"><a name="l01622"></a><span class="lineno"> 1622</span> }</div><div class="line"><a name="l01623"></a><span class="lineno"> 1623</span> } <span class="comment">// namespace arm_compute</span></div><div class="ttc" id="_error_8h_xhtml_a05b19c75afe9c24200a62b9724734bbd"><div class="ttname"><a href="_error_8h.xhtml#a05b19c75afe9c24200a62b9724734bbd">ARM_COMPUTE_ERROR</a></div><div class="ttdeci">#define ARM_COMPUTE_ERROR(...)</div><div class="ttdoc">Print the given message then throw an std::runtime_error.</div><div class="ttdef"><b>Definition:</b> <a href="_error_8h_source.xhtml#l00261">Error.h:261</a></div></div> |
