| <a href="quantized_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) 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">#ifdef __aarch64__</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 "arm_gemm.hpp"</span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> </div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#include <arm_neon.h></span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> </div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="keyword">namespace </span><a class="code" href="namespacearm__gemm.xhtml">arm_gemm</a> {</div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> </div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="keyword">namespace </span>{</div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> </div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="comment">/* Requantize a block of data, using the requantize parameters in 'qp'.</span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="comment"> *</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="comment"> * row_bias and col_bias are assumed to be precomputed values which include</span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> <span class="comment"> * any externally supplied bias, plus the row/column contibution sums, plus</span></div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="comment"> * the overall constant offset (A_offset * B_offset * depth).</span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="comment"> *</span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="comment"> * Note that this function works equally well for uint8_t output: just set</span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> <span class="comment"> * minval/maxval appropriately and cast the output pointer. It is caller's</span></div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="comment"> * responsibility to ensure that minval/maxval are representable in the</span></div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="comment"> * target type - the downcast to (u)int8_t is done by simply extracting the</span></div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span> <span class="comment"> * LSB.</span></div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span> <span class="comment"> *</span></div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> <span class="comment"> * The 'do_shift_correction' template parameter turns on the correction</span></div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span> <span class="comment"> * applied to negative values being shifted right to make sure they round</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> <span class="comment"> * properly - if negative values are never output (e.g. fused ReLU) this is</span></div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="comment"> * unnecessary.</span></div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="comment"> */</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="keyword">template</span><<span class="keywordtype">bool</span> do_shift_correction></div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> <span class="keywordtype">void</span> requantize_block_32_int(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height,</div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  <span class="keyword">const</span> int32_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int8_t *output, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> out_stride,</div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  <span class="keyword">const</span> int32_t *row_bias, <span class="keyword">const</span> int32_t *col_bias) {</div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keyword">const</span> int32x4_t v_mul = vdupq_n_s32(qp.requant_mul);</div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="keyword">const</span> int32x4_t v_shift = vdupq_n_s32(qp.requant_shift);</div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  <span class="keyword">const</span> int32x4_t v_minval = vdupq_n_s32(qp.minval);</div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  <span class="keyword">const</span> int32x4_t v_maxval = vdupq_n_s32(qp.maxval);</div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  <span class="keyword">const</span> int32x4_t v_c_offset = vdupq_n_s32(qp.c_offset);</div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> </div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  <span class="comment">/* To make sure we have plenty of accumulators, compute two rows at a</span></div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> <span class="comment"> * time. If the number of rows is odd, compute the bottom row twice to</span></div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span> <span class="comment"> * avoid needing a duplicate codepath. */</span></div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> row=0; row<height; row+=2) {</div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="comment">/* Prefer to do 4 vectors (16 values) at once as this collapses</span></div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span> <span class="comment"> * neatly to a single vector of output, failing that a vector at a</span></div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span> <span class="comment"> * time and then the odd ones out at the end. */</span></div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> blocks=(width / 16);</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> regs=(width % 16) / 4;</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> odds=(width % 4);</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> int32_t *colptr = col_bias;</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keyword">const</span> int32_t *in_ptr = input + (row * in_stride);</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  int8_t *out_ptr = output + (row * out_stride);</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  int32_t row_sum = row_bias[row];</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> </div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="keyword">const</span> int32_t *in_ptr1;</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  int8_t *out_ptr1;</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  int32_t row_sum1;</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> </div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="keywordflow">if</span> (row == height-1) {</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  in_ptr1 = in_ptr;</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  out_ptr1 = out_ptr;</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  row_sum1 = row_sum;</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  in_ptr1 = in_ptr + in_stride;</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  out_ptr1 = out_ptr + out_stride;</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  row_sum1 = row_bias[row+1];</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  }</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> int32x4_t v_row_sum = vdupq_n_s32(row_sum);</div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  <span class="keyword">const</span> int32x4_t v_row_sum1 = vdupq_n_s32(row_sum1);</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span> </div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="keywordflow">while</span> (blocks--) {</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  <span class="comment">// Load column pointers</span></div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  int32x4_t v_col0 = vld1q_s32(colptr);</div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  int32x4_t v_col1 = vld1q_s32(colptr + 4);</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  int32x4_t v_col2 = vld1q_s32(colptr + 8);</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  int32x4_t v_col3 = vld1q_s32(colptr + 12);</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  colptr += 16;</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="comment">// Load input data (row 0);</span></div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  int32x4_t v_in00 = vld1q_s32(in_ptr);</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  int32x4_t v_in01 = vld1q_s32(in_ptr + 4);</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  int32x4_t v_in02 = vld1q_s32(in_ptr + 8);</div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  int32x4_t v_in03 = vld1q_s32(in_ptr + 12);</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  in_ptr += 16;</div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span> </div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  <span class="comment">// Load input data (row 1);</span></div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  int32x4_t v_in10 = vld1q_s32(in_ptr1);</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  int32x4_t v_in11 = vld1q_s32(in_ptr1 + 4);</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  int32x4_t v_in12 = vld1q_s32(in_ptr1 + 8);</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  int32x4_t v_in13 = vld1q_s32(in_ptr1 + 12);</div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  in_ptr1 += 16;</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>  <span class="comment">// Add on row bias and column bias</span></div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  v_in00 = vaddq_s32(v_in00, v_row_sum);</div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  v_in01 = vaddq_s32(v_in01, v_row_sum);</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  v_in02 = vaddq_s32(v_in02, v_row_sum);</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  v_in03 = vaddq_s32(v_in03, v_row_sum);</div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span> </div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  v_in10 = vaddq_s32(v_in10, v_row_sum1);</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  v_in11 = vaddq_s32(v_in11, v_row_sum1);</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  v_in12 = vaddq_s32(v_in12, v_row_sum1);</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  v_in13 = vaddq_s32(v_in13, v_row_sum1);</div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span> </div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  v_in00 = vaddq_s32(v_in00, v_col0);</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  v_in01 = vaddq_s32(v_in01, v_col1);</div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  v_in02 = vaddq_s32(v_in02, v_col2);</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  v_in03 = vaddq_s32(v_in03, v_col3);</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>  v_in10 = vaddq_s32(v_in10, v_col0);</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  v_in11 = vaddq_s32(v_in11, v_col1);</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  v_in12 = vaddq_s32(v_in12, v_col2);</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  v_in13 = vaddq_s32(v_in13, v_col3);</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span> </div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <span class="comment">// Quantize - start with multiply</span></div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  v_in00 = vqrdmulhq_s32(v_in00, v_mul);</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  v_in01 = vqrdmulhq_s32(v_in01, v_mul);</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  v_in02 = vqrdmulhq_s32(v_in02, v_mul);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  v_in03 = vqrdmulhq_s32(v_in03, v_mul);</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>  v_in10 = vqrdmulhq_s32(v_in10, v_mul);</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  v_in11 = vqrdmulhq_s32(v_in11, v_mul);</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  v_in12 = vqrdmulhq_s32(v_in12, v_mul);</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  v_in13 = vqrdmulhq_s32(v_in13, v_mul);</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span> </div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  <span class="comment">// Compute and add on corrective offset</span></div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  <span class="keywordflow">if</span> (do_shift_correction) {</div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  int32x4_t v_temp00 = vandq_s32(v_in00, v_shift);</div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  int32x4_t v_temp01 = vandq_s32(v_in01, v_shift);</div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  int32x4_t v_temp02 = vandq_s32(v_in02, v_shift);</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  int32x4_t v_temp03 = vandq_s32(v_in03, v_shift);</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>  int32x4_t v_temp10 = vandq_s32(v_in10, v_shift);</div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  int32x4_t v_temp11 = vandq_s32(v_in11, v_shift);</div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  int32x4_t v_temp12 = vandq_s32(v_in12, v_shift);</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  int32x4_t v_temp13 = vandq_s32(v_in13, v_shift);</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>  v_temp00 = vshrq_n_s32(v_temp00, 31);</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  v_temp01 = vshrq_n_s32(v_temp01, 31);</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  v_temp02 = vshrq_n_s32(v_temp02, 31);</div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  v_temp03 = vshrq_n_s32(v_temp03, 31);</div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span> </div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  v_temp10 = vshrq_n_s32(v_temp10, 31);</div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  v_temp11 = vshrq_n_s32(v_temp11, 31);</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  v_temp12 = vshrq_n_s32(v_temp12, 31);</div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  v_temp13 = vshrq_n_s32(v_temp13, 31);</div><div class="line"><a name="l00170"></a><span class="lineno"> 170</span> </div><div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  v_in00 = vqaddq_s32(v_in00, v_temp00);</div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  v_in01 = vqaddq_s32(v_in01, v_temp01);</div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  v_in02 = vqaddq_s32(v_in02, v_temp02);</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  v_in03 = vqaddq_s32(v_in03, v_temp03);</div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span> </div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  v_in10 = vqaddq_s32(v_in10, v_temp10);</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  v_in11 = vqaddq_s32(v_in11, v_temp11);</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  v_in12 = vqaddq_s32(v_in12, v_temp12);</div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  v_in13 = vqaddq_s32(v_in13, v_temp13);</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  }</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span> </div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  v_in00 = vrshlq_s32(v_in00, v_shift);</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  v_in01 = vrshlq_s32(v_in01, v_shift);</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  v_in02 = vrshlq_s32(v_in02, v_shift);</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  v_in03 = vrshlq_s32(v_in03, v_shift);</div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span> </div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  v_in10 = vrshlq_s32(v_in10, v_shift);</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  v_in11 = vrshlq_s32(v_in11, v_shift);</div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  v_in12 = vrshlq_s32(v_in12, v_shift);</div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  v_in13 = vrshlq_s32(v_in13, v_shift);</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span> </div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  v_in00 = vaddq_s32(v_in00, v_c_offset);</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  v_in01 = vaddq_s32(v_in01, v_c_offset);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  v_in02 = vaddq_s32(v_in02, v_c_offset);</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  v_in03 = vaddq_s32(v_in03, v_c_offset);</div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span> </div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  v_in10 = vaddq_s32(v_in10, v_c_offset);</div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  v_in11 = vaddq_s32(v_in11, v_c_offset);</div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  v_in12 = vaddq_s32(v_in12, v_c_offset);</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  v_in13 = vaddq_s32(v_in13, v_c_offset);</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>  v_in00 = vmaxq_s32(v_in00, v_minval);</div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  v_in01 = vmaxq_s32(v_in01, v_minval);</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  v_in02 = vmaxq_s32(v_in02, v_minval);</div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  v_in03 = vmaxq_s32(v_in03, v_minval);</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span> </div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  v_in10 = vmaxq_s32(v_in10, v_minval);</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  v_in11 = vmaxq_s32(v_in11, v_minval);</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  v_in12 = vmaxq_s32(v_in12, v_minval);</div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  v_in13 = vmaxq_s32(v_in13, v_minval);</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span> </div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  v_in00 = vminq_s32(v_in00, v_maxval);</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  v_in01 = vminq_s32(v_in01, v_maxval);</div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  v_in02 = vminq_s32(v_in02, v_maxval);</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  v_in03 = vminq_s32(v_in03, v_maxval);</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span> </div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  v_in10 = vminq_s32(v_in10, v_maxval);</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  v_in11 = vminq_s32(v_in11, v_maxval);</div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  v_in12 = vminq_s32(v_in12, v_maxval);</div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  v_in13 = vminq_s32(v_in13, v_maxval);</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span> </div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  int16x8_t v_uz00 = vuzp1q_s16(vreinterpretq_s16_s32(v_in00), vreinterpretq_s16_s32(v_in01));</div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  int16x8_t v_uz01 = vuzp1q_s16(vreinterpretq_s16_s32(v_in02), vreinterpretq_s16_s32(v_in03));</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>  int16x8_t v_uz10 = vuzp1q_s16(vreinterpretq_s16_s32(v_in10), vreinterpretq_s16_s32(v_in11));</div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  int16x8_t v_uz11 = vuzp1q_s16(vreinterpretq_s16_s32(v_in12), vreinterpretq_s16_s32(v_in13));</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span> </div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  int8x16_t v_uz0 = vuzp1q_s8(vreinterpretq_s8_s16(v_uz00), vreinterpretq_s8_s16(v_uz01));</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  int8x16_t v_uz1 = vuzp1q_s8(vreinterpretq_s8_s16(v_uz10), vreinterpretq_s8_s16(v_uz11));</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span> </div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  vst1q_s8(out_ptr, v_uz0);</div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  out_ptr += 16;</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  vst1q_s8(out_ptr1, v_uz1);</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  out_ptr1 += 16;</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  }</div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span> </div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="keywordflow">while</span> (regs--) {</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  <span class="comment">// Load column pointers</span></div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  int32x4_t v_col0 = vld1q_s32(colptr);</div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  colptr += 4;</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> </div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  <span class="comment">// Load input data (row 0);</span></div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  int32x4_t v_in00 = vld1q_s32(in_ptr);</div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  in_ptr += 4;</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span> </div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="comment">// Load input data (row 1);</span></div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  int32x4_t v_in10 = vld1q_s32(in_ptr1);</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  in_ptr1 += 4;</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span> </div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="comment">// Add on row sum and bias constant</span></div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  v_in00 = vaddq_s32(v_in00, v_row_sum);</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>  v_in10 = vaddq_s32(v_in10, v_row_sum1);</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span> </div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  <span class="comment">// Subtract col sum * a_offset</span></div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  v_in00 = vaddq_s32(v_in00, v_col0);</div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span> </div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  v_in10 = vaddq_s32(v_in10, v_col0);</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span> </div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="comment">// Quantize - start with multiply</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  v_in00 = vqrdmulhq_s32(v_in00, v_mul);</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span> </div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  v_in10 = vqrdmulhq_s32(v_in10, v_mul);</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>  <span class="comment">// Compute and add on corrective offset</span></div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <span class="keywordflow">if</span> (do_shift_correction) {</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  int32x4_t v_temp00 = vandq_s32(v_in00, v_shift);</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span> </div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  int32x4_t v_temp10 = vandq_s32(v_in10, v_shift);</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span> </div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  v_temp00 = vshrq_n_s32(v_temp00, 31);</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  v_temp10 = vshrq_n_s32(v_temp10, 31);</div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span> </div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  v_in00 = vqaddq_s32(v_in00, v_temp00);</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span> </div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span>  v_in10 = vqaddq_s32(v_in10, v_temp10);</div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  }</div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span> </div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  v_in00 = vrshlq_s32(v_in00, v_shift);</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span> </div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  v_in10 = vrshlq_s32(v_in10, v_shift);</div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span> </div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  v_in00 = vaddq_s32(v_in00, v_c_offset);</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>  v_in10 = vaddq_s32(v_in10, v_c_offset);</div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span> </div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  v_in00 = vmaxq_s32(v_in00, v_minval);</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span> </div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  v_in10 = vmaxq_s32(v_in10, v_minval);</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span> </div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  v_in00 = vminq_s32(v_in00, v_maxval);</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span> </div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  v_in10 = vminq_s32(v_in10, v_maxval);</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span> </div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  int16x8_t v_uz00 = vuzp1q_s16(vreinterpretq_s16_s32(v_in00), vreinterpretq_s16_s32(v_in10));</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span> </div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  int8x16_t v_uz0 = vuzp1q_s8(vreinterpretq_s8_s16(v_uz00), vreinterpretq_s8_s16(v_uz00));</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span> </div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  vst1q_lane_s32(reinterpret_cast<int32_t *>(out_ptr), vreinterpretq_s32_s8(v_uz0), 0);</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  out_ptr += 4;</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  vst1q_lane_s32(reinterpret_cast<int32_t *>(out_ptr1), vreinterpretq_s32_s8(v_uz0), 1);</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  out_ptr1 += 4;</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span>  }</div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> </div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  <span class="keywordflow">if</span> (odds) {</div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  int32x4_t v_col0 = vdupq_n_s32(0);</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  int32x4_t v_in00 = vdupq_n_s32(0);</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  int32x4_t v_in10 = vdupq_n_s32(0);</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="keywordflow">do</span> {</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  v_col0 = vld1q_lane_s32(colptr, v_col0, 0);</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  v_in00 = vld1q_lane_s32(in_ptr, v_in00, 0);</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  v_in10 = vld1q_lane_s32(in_ptr1, v_in10, 0);</div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  <span class="keywordflow">if</span> (odds == 1) { <span class="keywordflow">break</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>  v_col0 = vld1q_lane_s32(colptr + 1, v_col0, 1);</div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span>  v_in00 = vld1q_lane_s32(in_ptr + 1, v_in00, 1);</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  v_in10 = vld1q_lane_s32(in_ptr1 + 1, v_in10, 1);</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  <span class="keywordflow">if</span> (odds == 2) { <span class="keywordflow">break</span>; }</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>  v_col0 = vld1q_lane_s32(colptr + 2, v_col0, 2);</div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  v_in00 = vld1q_lane_s32(in_ptr + 2, v_in00, 2);</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  v_in10 = vld1q_lane_s32(in_ptr1 + 2, v_in10, 2);</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  } <span class="keywordflow">while</span> (0);</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span> </div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span>  <span class="comment">// Add on row sum and bias constant</span></div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  v_in00 = vaddq_s32(v_in00, v_row_sum);</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>  v_in10 = vaddq_s32(v_in10, v_row_sum1);</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span> </div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  <span class="comment">// Subtract col sum * a_offset</span></div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  v_in00 = vaddq_s32(v_in00, v_col0);</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span> </div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  v_in10 = vaddq_s32(v_in10, v_col0);</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>  <span class="comment">// Quantize - start with multiply</span></div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  v_in00 = vqrdmulhq_s32(v_in00, v_mul);</div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span> </div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  v_in10 = vqrdmulhq_s32(v_in10, v_mul);</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span> </div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  <span class="comment">// Compute and add on corrective offset</span></div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="keywordflow">if</span> (do_shift_correction) {</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  int32x4_t v_temp00 = vandq_s32(v_in00, v_shift);</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span> </div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  int32x4_t v_temp10 = vandq_s32(v_in10, v_shift);</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>  v_temp00 = vshrq_n_s32(v_temp00, 31);</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span> </div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  v_temp10 = vshrq_n_s32(v_temp10, 31);</div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span> </div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  v_in00 = vqaddq_s32(v_in00, v_temp00);</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span> </div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  v_in10 = vqaddq_s32(v_in10, v_temp10);</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> </div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  v_in00 = vrshlq_s32(v_in00, v_shift);</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span> </div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  v_in10 = vrshlq_s32(v_in10, v_shift);</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span> </div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  v_in00 = vaddq_s32(v_in00, v_c_offset);</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span> </div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  v_in10 = vaddq_s32(v_in10, v_c_offset);</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>  v_in00 = vmaxq_s32(v_in00, v_minval);</div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span> </div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  v_in10 = vmaxq_s32(v_in10, v_minval);</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>  v_in00 = vminq_s32(v_in00, v_maxval);</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span> </div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  v_in10 = vminq_s32(v_in10, v_maxval);</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span> </div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  <span class="keywordflow">do</span> {</div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  vst1q_lane_s8(out_ptr, vreinterpretq_s8_s32(v_in00), 0);</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  vst1q_lane_s8(out_ptr1, vreinterpretq_s8_s32(v_in10), 0);</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span> </div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  <span class="keywordflow">if</span> (odds==1) { <span class="keywordflow">break</span>; }</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span> </div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  vst1q_lane_s8(out_ptr + 1, vreinterpretq_s8_s32(v_in00), 4);</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  vst1q_lane_s8(out_ptr1 + 1, vreinterpretq_s8_s32(v_in10), 4);</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span> </div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="keywordflow">if</span> (odds==2) { <span class="keywordflow">break</span>; }</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span> </div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  vst1q_lane_s8(out_ptr + 2, vreinterpretq_s8_s32(v_in00), 8);</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  vst1q_lane_s8(out_ptr1 + 2, vreinterpretq_s8_s32(v_in10), 8);</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  } <span class="keywordflow">while</span>(0);</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span>  }</div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  }</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span> }</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> } <span class="comment">// anonymous namespace</span></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> <span class="keyword">template</span><<span class="keyword">typename</span> Tin, <span class="keyword">typename</span> Tout></div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a5367d66fa8e4648db3f690e9cfa7d498">requantize_block_32</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height,</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  <span class="keyword">const</span> Tin *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, Tout *output, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> out_stride,</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  <span class="keyword">const</span> int32_t *row_bias, <span class="keyword">const</span> int32_t *col_bias) {</div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  <span class="keywordflow">if</span> (qp.minval >= qp.c_offset) {</div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  requantize_block_32_int<false>(qp, width, height, reinterpret_cast<const int32_t *>(input), in_stride,</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  reinterpret_cast<int8_t *>(output), out_stride, row_bias, col_bias);</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  requantize_block_32_int<true>(qp, width, height, reinterpret_cast<const int32_t *>(input), in_stride,</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  reinterpret_cast<int8_t *>(output), out_stride, row_bias, col_bias);</div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  }</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span> }</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span> </div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a5367d66fa8e4648db3f690e9cfa7d498">requantize_block_32</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height,</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span>  <span class="keyword">const</span> int32_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int8_t *output, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> out_stride,</div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  <span class="keyword">const</span> int32_t *row_bias, <span class="keyword">const</span> int32_t *col_bias);</div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span> </div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a5367d66fa8e4648db3f690e9cfa7d498">requantize_block_32</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height,</div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span>  <span class="keyword">const</span> uint32_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, uint8_t *output, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> out_stride,</div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="keyword">const</span> int32_t *row_bias, <span class="keyword">const</span> int32_t *col_bias);</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span> </div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span> <span class="comment">/*</span></div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span> <span class="comment"> * Routine (and helpers) to compute row sums needed for offset correction.</span></div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span> <span class="comment"> *</span></div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span> <span class="comment"> * This is often needed for a lot of short rows (e.g. Syrax 5 - 6400 rows</span></div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span> <span class="comment"> * of length 27), therefore it's important not to sacrifice performance on</span></div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span> <span class="comment"> * odd length rows.</span></div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span> <span class="comment"> *</span></div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span> <span class="comment"> * To minimize performance loss in these cases, this routine will overread</span></div><div class="line"><a name="l00422"></a><span class="lineno"> 422</span> <span class="comment"> * by up to 7 bytes.</span></div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span> <span class="comment"> *</span></div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span> <span class="comment"> * This is handled via "mask" and "mask mode" parameters to the inner</span></div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span> <span class="comment"> * routines; mask mode == 1 indicates that are between 1 and 8 bytes</span></div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span> <span class="comment"> * (inclusive) needed at the end; in these cases we always read 8 bytes.</span></div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span> <span class="comment"> * mask mode == 2 indicates that there are between 9 and 15 bytes needed at</span></div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span> <span class="comment"> * the end, and in this case we always read 16 bytes. In both cases the</span></div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span> <span class="comment"> * 'mask' vector is set up so that the read value can be masked off to clear</span></div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span> <span class="comment"> * the overread lanes. This is handled by 'accumulate_masked_8' and</span></div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span> <span class="comment"> * 'accumulate_masked_16' above.</span></div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span> <span class="comment"> *</span></div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span> <span class="comment"> * This routine is templated on the type to be accumulated, because the</span></div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span> <span class="comment"> * innermost instruction used needs to be of the correct signedness.</span></div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span> <span class="comment"> * However, beyond this point we always use signed values in both cases.</span></div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span> <span class="comment"> * The instructions that need to be different are therefore wrapped in</span></div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span> <span class="comment"> * helper functions below.</span></div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span> <span class="comment"> *</span></div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> <span class="comment"> * The general strategy used is to load vectors of 16 bytes and accumulate</span></div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span> <span class="comment"> * (using uadalp/sadalp or AArch32 equivalents) into 8x16-bit accumulators.</span></div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span> <span class="comment"> * These are then reduced (using uadalp/sadalp again) into 4x32-bit</span></div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span> <span class="comment"> * accumulators. The 4 accumulators for up to 4 rows being processed are</span></div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span> <span class="comment"> * then added together into a single output vector using pairwise adds.</span></div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span> <span class="comment"> *</span></div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span> <span class="comment"> * This reduction from the 8x16-bit into the 4x32-bit accumulators needs to</span></div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span> <span class="comment"> * occur before the 16-bit accumulators can overflow - which is every 32</span></div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span> <span class="comment"> * iterations (512 total bytes processed). This is explained more below.</span></div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span> <span class="comment"> */</span></div><div class="line"><a name="l00449"></a><span class="lineno"> 449</span> <span class="keyword">namespace </span>{</div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  <span class="keyword">struct </span>row_sum_helpers {</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="keyword">const</span> ARequantizeLayer32 &qp;</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 a full 16 byte vector, pairwise accumulate into 'sum' with uadalp or sadalp */</span></div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  <span class="keyword">template</span><<span class="keyword">typename</span> T></div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  <span class="keyword">inline</span> int16x8_t accumulate_16(<span class="keyword">const</span> T *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>);</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 a full 16 byte vector, but mask before accumulation (see above). */</span></div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  <span class="keyword">template</span><<span class="keyword">typename</span> T></div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  <span class="keyword">inline</span> int16x8_t accumulate_masked_16(<span class="keyword">const</span> T *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask);</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 8 bytes and mask before accumulation. */</span></div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  <span class="keyword">template</span><<span class="keyword">typename</span> T></div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  <span class="keyword">inline</span> int16x8_t accumulate_masked_8(<span class="keyword">const</span> T *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask);</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">/* This function does the actual work for up to 4 rows at a time.</span></div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span> <span class="comment"> * It's pulled out so we can template on the row count to generate</span></div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span> <span class="comment"> * the 4 different cases. 4 rows are computed at a time as this</span></div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span> <span class="comment"> * reduces to a single vector write. */</span></div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  <span class="keyword">template</span><<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> rows, <span class="keyword">typename</span> T></div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  <span class="keywordtype">void</span> compute_some_rows(<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> blocks, <span class="keyword">const</span> T *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *row_bias, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> mask_mode, uint64x2_t mask, int32x4_t offset_mul) {</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  int16x8_t sums[rows];</div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  int32x4_t finalsums[rows];</div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span> </div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<rows; i++) {</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  sums[i] = vdupq_n_s16(0);</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  finalsums[i] = vdupq_n_s32(0);</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="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<blocks; i++) {</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</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="l00481"></a><span class="lineno"> 481</span>  <span class="comment">/* If we add too many blocks together, we run the risk</span></div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span> <span class="comment"> * of overflowing the intermediate 16-bit accumulators,</span></div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span> <span class="comment"> * especially in the unsigned case where we later treat</span></div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span> <span class="comment"> * the accumulator as signed.</span></div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span> <span class="comment"> *</span></div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span> <span class="comment"> * In that case, the maximum (signed) value is 16383,</span></div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span> <span class="comment"> * which is safe for 64 (unsigned) accumulations (255*64</span></div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span> <span class="comment"> * = 16,320).</span></div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span> <span class="comment"> *</span></div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span> <span class="comment"> * Each invocation of pairwise add adds 2 values to the</span></div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span> <span class="comment"> * accumulator - so in the unsigned case we can do 32</span></div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span> <span class="comment"> * adds before we need to reset the 16-bit accumulator</span></div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span> <span class="comment"> * by adding into the 32-bit 'finalsums'.</span></div><div class="line"><a name="l00494"></a><span class="lineno"> 494</span> <span class="comment"> *</span></div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span> <span class="comment"> * We could do 64 adds in the signed case, but that</span></div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span> <span class="comment"> * optimization is not worth the complexity.</span></div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> <span class="comment"> */</span></div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="keywordflow">if</span> (i > 0 && ((i & 31) == 0)) {</div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  finalsums[r] = vpadalq_s16(finalsums[r], sums[r]);</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  sums[r] = vdupq_n_s16(0);</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  }</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  sums[r] = accumulate_16(input + (r * in_stride) + (i * 16), sums[r]);</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>  }</div><div class="line"><a name="l00505"></a><span class="lineno"> 505</span> </div><div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  <span class="comment">/* Handle the final masked read if needed. */</span></div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  <span class="keywordflow">if</span> (mask_mode > 0) {</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</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="l00509"></a><span class="lineno"> 509</span>  <span class="keywordflow">if</span> (mask_mode == 1) {</div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  sums[r] = accumulate_masked_8(input + (r * in_stride) + (blocks * 16), sums[r], mask);</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span>  sums[r] = accumulate_masked_16(input + (r * in_stride) + (blocks * 16), sums[r], mask);</div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  }</div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  }</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>  }</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span> </div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<rows; i++) {</div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  finalsums[i] = vpadalq_s16(finalsums[i], sums[i]);</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> </div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span>  int32x4_t t0, t1;</div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  int32x2_t t2;</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">/* Result writeback - need to write back one value per row</span></div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span> <span class="comment"> * processed. Multiply all the final totals by -b_offset so</span></div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span> <span class="comment"> * that the terms can simply be added in the requantize code.</span></div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span> <span class="comment"> * */</span></div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span>  <span class="keywordflow">switch</span> (rows) {</div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span>  <span class="keywordflow">case</span> 1:</div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  <span class="comment">/* If we only have one output, just use ADDV. Multiply</span></div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span> <span class="comment"> * the offset into all four components separately so it</span></div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span> <span class="comment"> * can stay in the SIMD register file. */</span></div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  t0 = vmulq_s32(finalsums[0], offset_mul);</div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  *row_bias = vaddvq_s32(t0);</div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span> </div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  <span class="keywordflow">case</span> 2:</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  <span class="comment">/* For two outputs, two rounds of pairwise adds will</span></div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span> <span class="comment"> * generate the result in a 2-vector we can store in one</span></div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span> <span class="comment"> * go. */</span></div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span>  t0 = vpaddq_s32(finalsums[0], finalsums[1]);</div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  t0 = vpaddq_s32(t0, t0);</div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  t2 = vmul_s32(vget_low_s32(t0), vget_low_s32(offset_mul));</div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span>  vst1_s32(row_bias, t2);</div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span> </div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  <span class="keywordflow">case</span> 3:</div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span>  <span class="comment">/* Three rows - need to store the low two words plus the odd value from lane 2 */</span></div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span>  t0 = vpaddq_s32(finalsums[0], finalsums[1]);</div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  t1 = vpaddq_s32(finalsums[2], finalsums[2]);</div><div class="line"><a name="l00551"></a><span class="lineno"> 551</span> </div><div class="line"><a name="l00552"></a><span class="lineno"> 552</span>  t0 = vpaddq_s32(t0, t1);</div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  t0 = vmulq_s32(t0, offset_mul);</div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span> </div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span>  vst1_s32(row_bias, vget_low_s32(t0));</div><div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  row_bias[2] = vgetq_lane_s32(t0, 2);</div><div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00558"></a><span class="lineno"> 558</span> </div><div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  <span class="keywordflow">case</span> 4:</div><div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  <span class="comment">/* Four rows (most common case) - reduce to a single</span></div><div class="line"><a name="l00561"></a><span class="lineno"> 561</span> <span class="comment"> * vector with pairwise adds. */</span></div><div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  t0 = vpaddq_s32(finalsums[0], finalsums[1]);</div><div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  t1 = vpaddq_s32(finalsums[2], finalsums[3]);</div><div class="line"><a name="l00564"></a><span class="lineno"> 564</span> </div><div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  t0 = vpaddq_s32(t0, t1);</div><div class="line"><a name="l00566"></a><span class="lineno"> 566</span>  t0 = vmulq_s32(t0, offset_mul);</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>  vst1q_s32(row_bias, t0);</div><div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00570"></a><span class="lineno"> 570</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l00571"></a><span class="lineno"> 571</span>  <span class="keywordflow">break</span>;</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>  }</div><div class="line"><a name="l00574"></a><span class="lineno"> 574</span> </div><div class="line"><a name="l00575"></a><span class="lineno"> 575</span>  row_sum_helpers(<span class="keyword">const</span> ARequantizeLayer32 &qp) : qp(qp) { }</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> </div><div class="line"><a name="l00578"></a><span class="lineno"> 578</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  int16x8_t row_sum_helpers::accumulate_16(<span class="keyword">const</span> uint8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>) {</div><div class="line"><a name="l00580"></a><span class="lineno"> 580</span>  <span class="keywordflow">return</span> vreinterpretq_s16_u16(vpadalq_u8(vreinterpretq_u16_s16(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>), vld1q_u8(ptr)));</div><div class="line"><a name="l00581"></a><span class="lineno"> 581</span>  }</div><div class="line"><a name="l00582"></a><span class="lineno"> 582</span> </div><div class="line"><a name="l00583"></a><span class="lineno"> 583</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00584"></a><span class="lineno"> 584</span>  int16x8_t row_sum_helpers::accumulate_16(<span class="keyword">const</span> int8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>) {</div><div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  <span class="keywordflow">return</span> vpadalq_s8(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, vld1q_s8(ptr));</div><div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  }</div><div class="line"><a name="l00587"></a><span class="lineno"> 587</span> </div><div class="line"><a name="l00588"></a><span class="lineno"> 588</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  int16x8_t row_sum_helpers::accumulate_masked_16(<span class="keyword">const</span> int8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask) {</div><div class="line"><a name="l00590"></a><span class="lineno"> 590</span>  int8x16_t v = vandq_s8(vld1q_s8(ptr), vreinterpretq_s8_u64(mask));</div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span>  <span class="keywordflow">return</span> vpadalq_s8(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, v);</div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  }</div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span> </div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  int16x8_t row_sum_helpers::accumulate_masked_16(<span class="keyword">const</span> uint8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask) {</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span>  uint8x16_t v = vandq_u8(vld1q_u8(ptr), vreinterpretq_u8_u64(mask));</div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  <span class="keywordflow">return</span> vreinterpretq_s16_u16(vpadalq_u8(vreinterpretq_u16_s16(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>), v));</div><div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  }</div><div class="line"><a name="l00599"></a><span class="lineno"> 599</span> </div><div class="line"><a name="l00600"></a><span class="lineno"> 600</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  int16x8_t row_sum_helpers::accumulate_masked_8(<span class="keyword">const</span> int8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask) {</div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  int8x16_t v = vcombine_s8(vld1_s8(ptr), vdup_n_s8(0));</div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  v = vreinterpretq_s8_u64(vandq_u64(mask, vreinterpretq_u64_s8(v)));</div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  <span class="keywordflow">return</span> vpadalq_s8(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, v);</div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  }</div><div class="line"><a name="l00606"></a><span class="lineno"> 606</span> </div><div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  <span class="keyword">template</span><></div><div class="line"><a name="l00608"></a><span class="lineno"> 608</span>  int16x8_t row_sum_helpers::accumulate_masked_8(<span class="keyword">const</span> uint8_t *ptr, int16x8_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>, uint64x2_t mask) {</div><div class="line"><a name="l00609"></a><span class="lineno"> 609</span>  uint8x16_t v = vcombine_u8(vld1_u8(ptr), vdup_n_u8(0));</div><div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  v = vreinterpretq_u8_u64(vandq_u64(mask, vreinterpretq_u64_u8(v)));</div><div class="line"><a name="l00611"></a><span class="lineno"> 611</span>  <span class="keywordflow">return</span> vreinterpretq_s16_u16(vpadalq_u8(vreinterpretq_u16_s16(<a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>), v));</div><div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  }</div><div class="line"><a name="l00613"></a><span class="lineno"> 613</span> }</div><div class="line"><a name="l00614"></a><span class="lineno"> 614</span> </div><div class="line"><a name="l00615"></a><span class="lineno"> 615</span> <span class="keyword">template</span><<span class="keyword">typename</span> T></div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a79d7971bf26ac0d4df7662a47d65d068">compute_row_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height,</div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  <span class="keyword">const</span> T *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *row_bias) {</div><div class="line"><a name="l00618"></a><span class="lineno"> 618</span>  row_sum_helpers thehelpers(qp);</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>  <span class="keyword">const</span> int32x4_t offset_mul = vdupq_n_s32(-qp.b_offset);</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>  <span class="comment">/* Work out how many full vectors of 16 bytes we will read, and how many</span></div><div class="line"><a name="l00623"></a><span class="lineno"> 623</span> <span class="comment"> * odd bytes at the end */</span></div><div class="line"><a name="l00624"></a><span class="lineno"> 624</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> blocks = (width / 16);</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  <span class="keyword">const</span> <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> odds = width % 16;</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span> </div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  <span class="comment">/* Generate a mask to use on the last iteration, if necessary. */</span></div><div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  uint64x2_t mask;</div><div class="line"><a name="l00629"></a><span class="lineno"> 629</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> mask_mode = 0;</div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span> </div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>  <span class="keywordflow">if</span> (odds > 0 && odds <= 8) {</div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span>  <span class="comment">/* 1-8 odds: mask in the low lane, 0 in the top */</span></div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span>  uint64_t maskval = (~0ULL) >> (8 * (8-odds));</div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span> </div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  mask = vsetq_lane_u64(maskval, vdupq_n_u64(0), 0);</div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span> </div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  mask_mode = 1;</div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (odds > 8) {</div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  <span class="comment">/* 9-15 odds: mask in the top lane, all 1s in the bottom. */</span></div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  uint64_t maskval = (~0ULL) >> (8 * (16-odds));</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>  mask = vsetq_lane_u64(maskval, vdupq_n_u64(~0ULL), 1);</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>  mask_mode = 2;</div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span>  }</div><div class="line"><a name="l00646"></a><span class="lineno"> 646</span> </div><div class="line"><a name="l00647"></a><span class="lineno"> 647</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> row=0; row<height; row+=4) {</div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  <span class="keywordflow">switch</span>(height-row) {</div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  <span class="keywordflow">case</span> 4:</div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span>  thehelpers.compute_some_rows<4>(blocks, input + (row * in_stride), in_stride, row_bias + row, mask_mode, mask, offset_mul);</div><div class="line"><a name="l00652"></a><span class="lineno"> 652</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00653"></a><span class="lineno"> 653</span>  <span class="keywordflow">case</span> 3:</div><div class="line"><a name="l00654"></a><span class="lineno"> 654</span>  thehelpers.compute_some_rows<3>(blocks, input + (row * in_stride), in_stride, row_bias + row, mask_mode, mask, offset_mul);</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span>  <span class="keywordflow">case</span> 2:</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span>  thehelpers.compute_some_rows<2>(blocks, input + (row * in_stride), in_stride, row_bias + row, mask_mode, mask, offset_mul);</div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span>  <span class="keywordflow">case</span> 1:</div><div class="line"><a name="l00660"></a><span class="lineno"> 660</span>  thehelpers.compute_some_rows<1>(blocks, input + (row * in_stride), in_stride, row_bias + row, mask_mode, mask, offset_mul);</div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  }</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span>  }</div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span> }</div><div class="line"><a name="l00665"></a><span class="lineno"> 665</span> </div><div class="line"><a name="l00666"></a><span class="lineno"> 666</span> <span class="comment">/* Instantiate the two versions for uint8_t and int8_t. */</span></div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a79d7971bf26ac0d4df7662a47d65d068">compute_row_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, <span class="keyword">const</span> int8_t *, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, int32_t *);</div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#a79d7971bf26ac0d4df7662a47d65d068">compute_row_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, <span class="keyword">const</span> uint8_t *, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>, int32_t *);</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> <span class="keyword">template</span><<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> active_rows, <span class="keyword">typename</span> T></div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> add_block(<span class="keyword">const</span> T *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *output);</div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span> </div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span> <span class="keyword">template</span><<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> active_rows></div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> add_block(<span class="keyword">const</span> uint8_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *output) {</div><div class="line"><a name="l00675"></a><span class="lineno"> 675</span>  uint8x16_t inputs[4];</div><div class="line"><a name="l00676"></a><span class="lineno"> 676</span> </div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<4; i++) {</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span>  <span class="keywordflow">if</span> (i < active_rows) {</div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span>  inputs[i] = vld1q_u8(input + i * in_stride);</div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  inputs[i] = vdupq_n_u8(0);</div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span>  }</div><div class="line"><a name="l00683"></a><span class="lineno"> 683</span>  }</div><div class="line"><a name="l00684"></a><span class="lineno"> 684</span> </div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span>  int16x8_t sums_16b[4];</div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span> </div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  <span class="comment">// Two adds for the low pairs</span></div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span>  sums_16b[0]=vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(inputs[0]), vget_low_u8(inputs[1])));</div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span>  sums_16b[1]=vreinterpretq_s16_u16(vaddl_u8(vget_low_u8(inputs[2]), vget_low_u8(inputs[3])));</div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span>  <span class="comment">// Two adds for the high pairs</span></div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  sums_16b[2]=vreinterpretq_s16_u16(vaddl_high_u8(inputs[0], inputs[1]));</div><div class="line"><a name="l00692"></a><span class="lineno"> 692</span>  sums_16b[3]=vreinterpretq_s16_u16(vaddl_high_u8(inputs[2], inputs[3]));</div><div class="line"><a name="l00693"></a><span class="lineno"> 693</span> </div><div class="line"><a name="l00694"></a><span class="lineno"> 694</span>  int32x4_t sums_32b[4];</div><div class="line"><a name="l00695"></a><span class="lineno"> 695</span> </div><div class="line"><a name="l00696"></a><span class="lineno"> 696</span>  sums_32b[0]=vaddl_s16(vget_low_s16(sums_16b[0]), vget_low_s16(sums_16b[1]));</div><div class="line"><a name="l00697"></a><span class="lineno"> 697</span>  sums_32b[1]=vaddl_high_s16(sums_16b[0], sums_16b[1]);</div><div class="line"><a name="l00698"></a><span class="lineno"> 698</span>  sums_32b[2]=vaddl_s16(vget_low_s16(sums_16b[2]), vget_low_s16(sums_16b[3]));</div><div class="line"><a name="l00699"></a><span class="lineno"> 699</span>  sums_32b[3]=vaddl_high_s16(sums_16b[2], sums_16b[3]);</div><div class="line"><a name="l00700"></a><span class="lineno"> 700</span> </div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<4; i++) {</div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  vst1q_s32(output + 4*i, vaddq_s32(sums_32b[i], vld1q_s32(output + 4*i)));</div><div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  }</div><div class="line"><a name="l00704"></a><span class="lineno"> 704</span> }</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="keyword">template</span><<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> active_rows></div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span> <span class="keyword">inline</span> <span class="keywordtype">void</span> add_block(<span class="keyword">const</span> int8_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *output) {</div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>  int8x16_t inputs[4];</div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span> </div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<4; i++) {</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span>  <span class="keywordflow">if</span> (i < active_rows) {</div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>  inputs[i] = vld1q_s8(input + i * in_stride);</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span>  inputs[i] = vdupq_n_s8(0);</div><div class="line"><a name="l00715"></a><span class="lineno"> 715</span>  }</div><div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  }</div><div class="line"><a name="l00717"></a><span class="lineno"> 717</span> </div><div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  int16x8_t sums_16b[4];</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>  <span class="comment">// Two adds for the low pairs</span></div><div class="line"><a name="l00721"></a><span class="lineno"> 721</span>  sums_16b[0]=vaddl_s8(vget_low_s8(inputs[0]), vget_low_s8(inputs[1]));</div><div class="line"><a name="l00722"></a><span class="lineno"> 722</span>  sums_16b[1]=vaddl_s8(vget_low_s8(inputs[2]), vget_low_s8(inputs[3]));</div><div class="line"><a name="l00723"></a><span class="lineno"> 723</span>  <span class="comment">// Two adds for the high pairs</span></div><div class="line"><a name="l00724"></a><span class="lineno"> 724</span>  sums_16b[2]=vaddl_high_s8(inputs[0], inputs[1]);</div><div class="line"><a name="l00725"></a><span class="lineno"> 725</span>  sums_16b[3]=vaddl_high_s8(inputs[2], inputs[3]);</div><div class="line"><a name="l00726"></a><span class="lineno"> 726</span> </div><div class="line"><a name="l00727"></a><span class="lineno"> 727</span>  int32x4_t sums_32b[4];</div><div class="line"><a name="l00728"></a><span class="lineno"> 728</span> </div><div class="line"><a name="l00729"></a><span class="lineno"> 729</span>  sums_32b[0]=vaddl_s16(vget_low_s16(sums_16b[0]), vget_low_s16(sums_16b[1]));</div><div class="line"><a name="l00730"></a><span class="lineno"> 730</span>  sums_32b[1]=vaddl_high_s16(sums_16b[0], sums_16b[1]);</div><div class="line"><a name="l00731"></a><span class="lineno"> 731</span>  sums_32b[2]=vaddl_s16(vget_low_s16(sums_16b[2]), vget_low_s16(sums_16b[3]));</div><div class="line"><a name="l00732"></a><span class="lineno"> 732</span>  sums_32b[3]=vaddl_high_s16(sums_16b[2], sums_16b[3]);</div><div class="line"><a name="l00733"></a><span class="lineno"> 733</span> </div><div class="line"><a name="l00734"></a><span class="lineno"> 734</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> i=0; i<4; i++) {</div><div class="line"><a name="l00735"></a><span class="lineno"> 735</span>  vst1q_s32(output + 4*i, vaddq_s32(sums_32b[i], vld1q_s32(output + 4*i)));</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> </div><div class="line"><a name="l00739"></a><span class="lineno"> 739</span> </div><div class="line"><a name="l00740"></a><span class="lineno"> 740</span> <span class="comment">/* "first_col" parameter is used to offset the read into the qp.bias array,</span></div><div class="line"><a name="l00741"></a><span class="lineno"> 741</span> <span class="comment"> * in cases where we are not computing the first columns of the output (i.e.</span></div><div class="line"><a name="l00742"></a><span class="lineno"> 742</span> <span class="comment"> * in multithreaded cases where we divide columns across threads) */</span></div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span> <span class="keyword">template</span><<span class="keyword">typename</span> T></div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#ac6ac0fe47fd694fe6d6dbba1ce6e9aac">compute_col_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height, <span class="keyword">const</span> T *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *col_bias, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> depth, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_col) {</div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span>  memset(reinterpret_cast<void *>(col_bias), 0, width * <span class="keyword">sizeof</span>(int32_t));</div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span> </div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> row=0; row<height; row+=4) {</div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> numrows=std::min(height-row, 4u);</div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span> </div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> col=0; col<width; col+=16) {</div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> numcols=std::min(width-col, 16u);</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>  <span class="keywordflow">if</span> (numcols==16) {</div><div class="line"><a name="l00754"></a><span class="lineno"> 754</span>  <span class="keywordflow">switch</span>(numrows) {</div><div class="line"><a name="l00755"></a><span class="lineno"> 755</span>  <span class="keywordflow">case</span> 1:</div><div class="line"><a name="l00756"></a><span class="lineno"> 756</span>  add_block<1>(input + row * in_stride + col, in_stride, col_bias + col);</div><div class="line"><a name="l00757"></a><span class="lineno"> 757</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00758"></a><span class="lineno"> 758</span> </div><div class="line"><a name="l00759"></a><span class="lineno"> 759</span>  <span class="keywordflow">case</span> 2:</div><div class="line"><a name="l00760"></a><span class="lineno"> 760</span>  add_block<2>(input + row * in_stride + col, in_stride, col_bias + col);</div><div class="line"><a name="l00761"></a><span class="lineno"> 761</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00762"></a><span class="lineno"> 762</span> </div><div class="line"><a name="l00763"></a><span class="lineno"> 763</span>  <span class="keywordflow">case</span> 3:</div><div class="line"><a name="l00764"></a><span class="lineno"> 764</span>  add_block<3>(input + row * in_stride + col, in_stride, col_bias + col);</div><div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00766"></a><span class="lineno"> 766</span> </div><div class="line"><a name="l00767"></a><span class="lineno"> 767</span>  <span class="keywordflow">case</span> 4:</div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span>  add_block<4>(input + row * in_stride + col, in_stride, col_bias + col);</div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00770"></a><span class="lineno"> 770</span>  <span class="keywordflow">default</span>:</div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  <span class="keywordflow">break</span>;</div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  }</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span>  } <span class="keywordflow">else</span> {</div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span>  <span class="keywordflow">for</span> (; col<width; col++) {</div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span>  int32_t <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>=0;</div><div class="line"><a name="l00776"></a><span class="lineno"> 776</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> r=0; r<numrows; r++) {</div><div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a> += input[(row + r)*in_stride + col];</div><div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  }</div><div class="line"><a name="l00779"></a><span class="lineno"> 779</span>  col_bias[col] += <a class="code" href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a>;</div><div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  }</div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span>  }</div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  }</div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  }</div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span> </div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span>  <span class="keywordflow">for</span> (<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> col=0; col<width; col++) {</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span>  int32_t result = col_bias[col];</div><div class="line"><a name="l00787"></a><span class="lineno"> 787</span> </div><div class="line"><a name="l00788"></a><span class="lineno"> 788</span>  result = (qp.a_offset * qp.b_offset * depth) - (result * qp.a_offset);</div><div class="line"><a name="l00789"></a><span class="lineno"> 789</span> </div><div class="line"><a name="l00790"></a><span class="lineno"> 790</span>  <span class="keywordflow">if</span> (qp.bias != <span class="keyword">nullptr</span>) {</div><div class="line"><a name="l00791"></a><span class="lineno"> 791</span>  result += qp.bias[col + first_col];</div><div class="line"><a name="l00792"></a><span class="lineno"> 792</span>  }</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>  col_bias[col] = result;</div><div class="line"><a name="l00795"></a><span class="lineno"> 795</span>  }</div><div class="line"><a name="l00796"></a><span class="lineno"> 796</span> }</div><div class="line"><a name="l00797"></a><span class="lineno"> 797</span> </div><div class="line"><a name="l00798"></a><span class="lineno"> 798</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#ac6ac0fe47fd694fe6d6dbba1ce6e9aac">compute_col_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height, <span class="keyword">const</span> int8_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *col_bias, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> depth, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_col);</div><div class="line"><a name="l00799"></a><span class="lineno"> 799</span> <span class="keyword">template</span> <span class="keywordtype">void</span> <a class="code" href="namespacearm__gemm.xhtml#ac6ac0fe47fd694fe6d6dbba1ce6e9aac">compute_col_sums</a>(<span class="keyword">const</span> ARequantizeLayer32 &qp, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> width, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> height, <span class="keyword">const</span> uint8_t *input, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> in_stride, int32_t *col_bias, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> depth, <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> first_col);</div><div class="line"><a name="l00800"></a><span class="lineno"> 800</span> </div><div class="line"><a name="l00801"></a><span class="lineno"> 801</span> } <span class="comment">// namespace arm_gemm</span></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> <span class="preprocessor">#endif // __aarch64__</span></div><div class="ttc" id="reduction__operation_8cl_xhtml_ab0df00f5333da51860deb93deb44a782"><div class="ttname"><a href="reduction__operation_8cl.xhtml#ab0df00f5333da51860deb93deb44a782">sum</a></div><div class="ttdeci">DATA_TYPE sum(__global const DATA_TYPE *input)</div><div class="ttdoc">Calculate sum of a vector.</div><div class="ttdef"><b>Definition:</b> <a href="reduction__operation_8cl_source.xhtml#l00065">reduction_operation.cl:65</a></div></div> |
