Jenkins | b9abeae | 2018-11-22 11:58:08 +0000 | [diff] [blame^] | 94 | <a href="gemmlowp_8cl.xhtml">Go to the documentation of this file.</a><div class="fragment"><div class="line"><a name="l00001"></a><span class="lineno"> 1</span> <span class="comment">/*</span></div><div class="line"><a name="l00002"></a><span class="lineno"> 2</span> <span class="comment"> * Copyright (c) 2017-2018 ARM Limited.</span></div><div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> *</span></div><div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> * SPDX-License-Identifier: MIT</span></div><div class="line"><a name="l00005"></a><span class="lineno"> 5</span> <span class="comment"> *</span></div><div class="line"><a name="l00006"></a><span class="lineno"> 6</span> <span class="comment"> * Permission is hereby granted, free of charge, to any person obtaining a copy</span></div><div class="line"><a name="l00007"></a><span class="lineno"> 7</span> <span class="comment"> * of this software and associated documentation files (the "Software"), to</span></div><div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <span class="comment"> * deal in the Software without restriction, including without limitation the</span></div><div class="line"><a name="l00009"></a><span class="lineno"> 9</span> <span class="comment"> * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or</span></div><div class="line"><a name="l00010"></a><span class="lineno"> 10</span> <span class="comment"> * sell copies of the Software, and to permit persons to whom the Software is</span></div><div class="line"><a name="l00011"></a><span class="lineno"> 11</span> <span class="comment"> * furnished to do so, subject to the following conditions:</span></div><div class="line"><a name="l00012"></a><span class="lineno"> 12</span> <span class="comment"> *</span></div><div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="comment"> * The above copyright notice and this permission notice shall be included in all</span></div><div class="line"><a name="l00014"></a><span class="lineno"> 14</span> <span class="comment"> * copies or substantial portions of the Software.</span></div><div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="comment"> *</span></div><div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="comment"> * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR</span></div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="comment"> * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,</span></div><div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="comment"> * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE</span></div><div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="comment"> * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER</span></div><div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="comment"> * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,</span></div><div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="comment"> * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE</span></div><div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="comment"> * SOFTWARE.</span></div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="comment"> */</span></div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#include "<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml">helpers.h</a>"</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="preprocessor">#include "<a class="code" href="helpers__asymm_8h.xhtml">helpers_asymm.h</a>"</span></div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> </div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)</span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#define ARM_DOT(x, y, val) val = arm_dot_acc((x), (y), (val));</span></div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="preprocessor">#else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)</span></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#define ARM_DOT(x, y, val) val += arm_dot((x), (y));</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8)</span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> </div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="preprocessor">#if defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP)</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> </div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_interleaved_transposed_midgard(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  uint src0_stride_z,</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  uint src1_stride_z,</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  uint dst_stride_z</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  ,</div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  uint cross_plane_pad</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  )</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span> {</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> x = get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP;</div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span> </div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;</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="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  __global uchar *src_addr_a = (__global uchar *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);</div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  __global uchar *src_addr_b = (__global uchar *)(src1_ptr + x * src1_stride_y + src1_offset_first_element_in_bytes);</div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span> </div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  src_addr_b += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  src_addr_b += z * src1_stride_z;</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span> </div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  __global uchar *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span> </div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  src_addr_b += offset_row_b;</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">// Reset accumulators</span></div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  int4 c00 = 0;</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  int4 c10 = 0;</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  int4 c20 = 0;</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  int4 c30 = 0;</div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span> </div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  <span class="keywordflow">for</span>(; src_addr_b <= (src_end_addr_b - (int)(8 * TRANSPOSE1XW_WIDTH_STEP)); src_addr_a += 8 * MULT_INTERLEAVE4X4_HEIGHT, src_addr_b += 8 * TRANSPOSE1XW_WIDTH_STEP)</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  {</div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  int4 a0 = convert_int4(vload4(0, src_addr_a));</div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  int4 b0 = convert_int4(vload4(0, src_addr_b));</div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> </div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  c00 += (int4)a0.s0 * b0;</div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  c10 += (int4)a0.s1 * b0;</div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  c20 += (int4)a0.s2 * b0;</div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  c30 += (int4)a0.s3 * b0;</div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span> </div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  a0 = convert_int4(vload4(0, src_addr_a + 4 * MULT_INTERLEAVE4X4_HEIGHT));</div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  b0 = convert_int4(vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP));</div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span> </div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  c00 += (int4)a0.s0 * b0;</div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  c10 += (int4)a0.s1 * b0;</div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  c20 += (int4)a0.s2 * b0;</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  c30 += (int4)a0.s3 * b0;</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  }</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span> </div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="keywordflow">for</span>(; src_addr_b < src_end_addr_b; src_addr_a += (4 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (4 * TRANSPOSE1XW_WIDTH_STEP))</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">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  int4 a0 = convert_int4(vload4(0, src_addr_a));</div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  int4 b0 = convert_int4(vload4(0, src_addr_b));</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span> </div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  c00 += (int4)a0.s0 * b0;</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  c10 += (int4)a0.s1 * b0;</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  c20 += (int4)a0.s2 * b0;</div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  c30 += (int4)a0.s3 * b0;</div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  }</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span> </div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  <a class="code" href="struct_image.xhtml">Image</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00150"></a><span class="lineno"> 150</span> </div><div class="line"><a name="l00151"></a><span class="lineno"> 151</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="comment">//</span></div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span> </div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  zout = min(DEPTH_GEMM3D - 1, zout);</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>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  zout *= (cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span> </div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span> </div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  vstore4(c00, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  vstore4(c10, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  vstore4(c20, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  vstore4(c30, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span> </div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z;</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span> </div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  vstore4(c00, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  vstore4(c10, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  vstore4(c20, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  vstore4(c30, 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span> }</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span> </div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_interleaved_transposed_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  uint src0_stride_z,</div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  uint src1_stride_z,</div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  uint dst_stride_z</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  ,</div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  uint cross_plane_pad</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  )</div><div class="line"><a name="l00243"></a><span class="lineno"> 243</span> {</div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> x = get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP;</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> y = get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT;</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span> </div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span> </div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  <span class="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  __global uchar *src_addr_a = (__global uchar *)(src0_ptr + z * src0_stride_z + y * src0_stride_y + src0_offset_first_element_in_bytes);</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  __global uchar *src_addr_b = (__global uchar *)(src1_ptr + x * src1_stride_y + src1_offset_first_element_in_bytes);</div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span> </div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  src_addr_b += (z % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  src_addr_b += z * src1_stride_z;</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span> </div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  <span class="comment">// Compute end row address for matrix B</span></div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  __global uchar *src_end_addr_b = src_addr_b + COLS_B;</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span> </div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  src_addr_b += offset_row_b;</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span> </div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  <span class="comment">// Reset accumulators</span></div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  uint c00 = 0;</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  uint c01 = 0;</div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  uint c02 = 0;</div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  uint c03 = 0;</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  uint c10 = 0;</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  uint c11 = 0;</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span>  uint c12 = 0;</div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  uint c13 = 0;</div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  uint c20 = 0;</div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  uint c21 = 0;</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  uint c22 = 0;</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  uint c23 = 0;</div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  uint c30 = 0;</div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  uint c31 = 0;</div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span>  uint c32 = 0;</div><div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  uint c33 = 0;</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> <span class="preprocessor">#if MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  <span class="keywordflow">for</span>(; src_addr_b <= (src_end_addr_b - (int)(32 * TRANSPOSE1XW_WIDTH_STEP)); src_addr_a += (32 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (32 * TRANSPOSE1XW_WIDTH_STEP))</div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  {</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  uchar16 a0 = vload16(0, src_addr_a);</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  uchar4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span> </div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  c00 += (ushort)a0.s0 * b0.s0;</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  c01 += (ushort)a0.s0 * b0.s1;</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  c02 += (ushort)a0.s0 * b0.s2;</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  c03 += (ushort)a0.s0 * b0.s3;</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>  c10 += (ushort)a0.s1 * b0.s0;</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  c11 += (ushort)a0.s1 * b0.s1;</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  c12 += (ushort)a0.s1 * b0.s2;</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  c13 += (ushort)a0.s1 * b0.s3;</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>  c20 += (ushort)a0.s2 * b0.s0;</div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  c21 += (ushort)a0.s2 * b0.s1;</div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  c22 += (ushort)a0.s2 * b0.s2;</div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  c23 += (ushort)a0.s2 * b0.s3;</div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span> </div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  c30 += (ushort)a0.s3 * b0.s0;</div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  c31 += (ushort)a0.s3 * b0.s1;</div><div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  c32 += (ushort)a0.s3 * b0.s2;</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  c33 += (ushort)a0.s3 * b0.s3;</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span> </div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  b0 = vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00317"></a><span class="lineno"> 317</span> </div><div class="line"><a name="l00318"></a><span class="lineno"> 318</span>  c00 += (ushort)a0.s4 * b0.s0;</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  c01 += (ushort)a0.s4 * b0.s1;</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  c02 += (ushort)a0.s4 * b0.s2;</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  c03 += (ushort)a0.s4 * b0.s3;</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span> </div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  c10 += (ushort)a0.s5 * b0.s0;</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  c11 += (ushort)a0.s5 * b0.s1;</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  c12 += (ushort)a0.s5 * b0.s2;</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span>  c13 += (ushort)a0.s5 * b0.s3;</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span> </div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  c20 += (ushort)a0.s6 * b0.s0;</div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span>  c21 += (ushort)a0.s6 * b0.s1;</div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span>  c22 += (ushort)a0.s6 * b0.s2;</div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  c23 += (ushort)a0.s6 * b0.s3;</div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span> </div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  c30 += (ushort)a0.s7 * b0.s0;</div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  c31 += (ushort)a0.s7 * b0.s1;</div><div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  c32 += (ushort)a0.s7 * b0.s2;</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  c33 += (ushort)a0.s7 * b0.s3;</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span> </div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  b0 = vload4(0, src_addr_b + 8 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span> </div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  c00 += (ushort)a0.s8 * b0.s0;</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  c01 += (ushort)a0.s8 * b0.s1;</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  c02 += (ushort)a0.s8 * b0.s2;</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  c03 += (ushort)a0.s8 * b0.s3;</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>  c10 += (ushort)a0.s9 * b0.s0;</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  c11 += (ushort)a0.s9 * b0.s1;</div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  c12 += (ushort)a0.s9 * b0.s2;</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  c13 += (ushort)a0.s9 * b0.s3;</div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span> </div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  c20 += (ushort)a0.sA * b0.s0;</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  c21 += (ushort)a0.sA * b0.s1;</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  c22 += (ushort)a0.sA * b0.s2;</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  c23 += (ushort)a0.sA * b0.s3;</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>  c30 += (ushort)a0.sB * b0.s0;</div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  c31 += (ushort)a0.sB * b0.s1;</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  c32 += (ushort)a0.sB * b0.s2;</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  c33 += (ushort)a0.sB * b0.s3;</div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span> </div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  b0 = vload4(0, src_addr_b + 12 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span> </div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  c00 += (ushort)a0.sC * b0.s0;</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  c01 += (ushort)a0.sC * b0.s1;</div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  c02 += (ushort)a0.sC * b0.s2;</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  c03 += (ushort)a0.sC * b0.s3;</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>  c10 += (ushort)a0.sD * b0.s0;</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  c11 += (ushort)a0.sD * b0.s1;</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  c12 += (ushort)a0.sD * b0.s2;</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  c13 += (ushort)a0.sD * b0.s3;</div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span> </div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  c20 += (ushort)a0.sE * b0.s0;</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  c21 += (ushort)a0.sE * b0.s1;</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  c22 += (ushort)a0.sE * b0.s2;</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  c23 += (ushort)a0.sE * b0.s3;</div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span> </div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  c30 += (ushort)a0.sF * b0.s0;</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  c31 += (ushort)a0.sF * b0.s1;</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  c32 += (ushort)a0.sF * b0.s2;</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  c33 += (ushort)a0.sF * b0.s3;</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span> </div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  a0 = vload16(0, src_addr_a + 16);</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  b0 = vload4(0, src_addr_b + 16 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span> </div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  c00 += (ushort)a0.s0 * b0.s0;</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  c01 += (ushort)a0.s0 * b0.s1;</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  c02 += (ushort)a0.s0 * b0.s2;</div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  c03 += (ushort)a0.s0 * b0.s3;</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>  c10 += (ushort)a0.s1 * b0.s0;</div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  c11 += (ushort)a0.s1 * b0.s1;</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  c12 += (ushort)a0.s1 * b0.s2;</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  c13 += (ushort)a0.s1 * b0.s3;</div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span> </div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  c20 += (ushort)a0.s2 * b0.s0;</div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span>  c21 += (ushort)a0.s2 * b0.s1;</div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  c22 += (ushort)a0.s2 * b0.s2;</div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  c23 += (ushort)a0.s2 * b0.s3;</div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span> </div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span>  c30 += (ushort)a0.s3 * b0.s0;</div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  c31 += (ushort)a0.s3 * b0.s1;</div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  c32 += (ushort)a0.s3 * b0.s2;</div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  c33 += (ushort)a0.s3 * b0.s3;</div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span> </div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  b0 = vload4(0, src_addr_b + 20 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span> </div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span>  c00 += (ushort)a0.s4 * b0.s0;</div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  c01 += (ushort)a0.s4 * b0.s1;</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  c02 += (ushort)a0.s4 * b0.s2;</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  c03 += (ushort)a0.s4 * b0.s3;</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span> </div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span>  c10 += (ushort)a0.s5 * b0.s0;</div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  c11 += (ushort)a0.s5 * b0.s1;</div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span>  c12 += (ushort)a0.s5 * b0.s2;</div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  c13 += (ushort)a0.s5 * b0.s3;</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span> </div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span>  c20 += (ushort)a0.s6 * b0.s0;</div><div class="line"><a name="l00422"></a><span class="lineno"> 422</span>  c21 += (ushort)a0.s6 * b0.s1;</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  c22 += (ushort)a0.s6 * b0.s2;</div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  c23 += (ushort)a0.s6 * b0.s3;</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span> </div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  c30 += (ushort)a0.s7 * b0.s0;</div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  c31 += (ushort)a0.s7 * b0.s1;</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span>  c32 += (ushort)a0.s7 * b0.s2;</div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  c33 += (ushort)a0.s7 * b0.s3;</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span> </div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span>  b0 = vload4(0, src_addr_b + 24 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span> </div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span>  c00 += (ushort)a0.s8 * b0.s0;</div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  c01 += (ushort)a0.s8 * b0.s1;</div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  c02 += (ushort)a0.s8 * b0.s2;</div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  c03 += (ushort)a0.s8 * b0.s3;</div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span> </div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span>  c10 += (ushort)a0.s9 * b0.s0;</div><div class="line"><a name="l00440"></a><span class="lineno"> 440</span>  c11 += (ushort)a0.s9 * b0.s1;</div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  c12 += (ushort)a0.s9 * b0.s2;</div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  c13 += (ushort)a0.s9 * b0.s3;</div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span> </div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  c20 += (ushort)a0.sA * b0.s0;</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  c21 += (ushort)a0.sA * b0.s1;</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  c22 += (ushort)a0.sA * b0.s2;</div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  c23 += (ushort)a0.sA * b0.s3;</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span> </div><div class="line"><a name="l00449"></a><span class="lineno"> 449</span>  c30 += (ushort)a0.sB * b0.s0;</div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  c31 += (ushort)a0.sB * b0.s1;</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  c32 += (ushort)a0.sB * b0.s2;</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  c33 += (ushort)a0.sB * b0.s3;</div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span> </div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  <span class="comment">// Load values from matrix B (transposed)</span></div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  b0 = vload4(0, src_addr_b + 28 * TRANSPOSE1XW_WIDTH_STEP);</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>  c00 += (ushort)a0.sC * b0.s0;</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  c01 += (ushort)a0.sC * b0.s1;</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  c02 += (ushort)a0.sC * b0.s2;</div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  c03 += (ushort)a0.sC * b0.s3;</div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span> </div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  c10 += (ushort)a0.sD * b0.s0;</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  c11 += (ushort)a0.sD * b0.s1;</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  c12 += (ushort)a0.sD * b0.s2;</div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  c13 += (ushort)a0.sD * b0.s3;</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span> </div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span>  c20 += (ushort)a0.sE * b0.s0;</div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  c21 += (ushort)a0.sE * b0.s1;</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  c22 += (ushort)a0.sE * b0.s2;</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span>  c23 += (ushort)a0.sE * b0.s3;</div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span> </div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  c30 += (ushort)a0.sF * b0.s0;</div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  c31 += (ushort)a0.sF * b0.s1;</div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  c32 += (ushort)a0.sF * b0.s2;</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  c33 += (ushort)a0.sF * b0.s3;</div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  }</div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span> <span class="preprocessor">#endif // MULT_INTERLEAVE4X4_HEIGHT == 1</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>(; src_addr_b < src_end_addr_b; src_addr_a += (4 * MULT_INTERLEAVE4X4_HEIGHT), src_addr_b += (4 * TRANSPOSE1XW_WIDTH_STEP))</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  {</div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span>  uchar4 a0 = vload4(0, src_addr_a);</div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  uchar4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span> </div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  c00 += (ushort)a0.s0 * b0.s0;</div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  c01 += (ushort)a0.s0 * b0.s1;</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  c02 += (ushort)a0.s0 * b0.s2;</div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span>  c03 += (ushort)a0.s0 * b0.s3;</div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span> </div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span>  c10 += (ushort)a0.s1 * b0.s0;</div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  c11 += (ushort)a0.s1 * b0.s1;</div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  c12 += (ushort)a0.s1 * b0.s2;</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  c13 += (ushort)a0.s1 * b0.s3;</div><div class="line"><a name="l00494"></a><span class="lineno"> 494</span> </div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span>  c20 += (ushort)a0.s2 * b0.s0;</div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  c21 += (ushort)a0.s2 * b0.s1;</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span>  c22 += (ushort)a0.s2 * b0.s2;</div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  c23 += (ushort)a0.s2 * b0.s3;</div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span> </div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  c30 += (ushort)a0.s3 * b0.s0;</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  c31 += (ushort)a0.s3 * b0.s1;</div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  c32 += (ushort)a0.s3 * b0.s2;</div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span>  c33 += (ushort)a0.s3 * b0.s3;</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">// Compute destination address</span></div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00508"></a><span class="lineno"> 508</span> </div><div class="line"><a name="l00509"></a><span class="lineno"> 509</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span>  <span class="comment">//</span></div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l00519"></a><span class="lineno"> 519</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00520"></a><span class="lineno"> 520</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00523"></a><span class="lineno"> 523</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00524"></a><span class="lineno"> 524</span> </div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span> </div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  zout *= (cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span> </div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00535"></a><span class="lineno"> 535</span> </div><div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  vstore4((int4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  vstore4((int4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span>  vstore4((int4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  vstore4((int4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l00541"></a><span class="lineno"> 541</span> </div><div class="line"><a name="l00542"></a><span class="lineno"> 542</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z;</div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span> </div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  vstore4((int4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span>  vstore4((int4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span>  vstore4((int4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  vstore4((int4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l00551"></a><span class="lineno"> 551</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00552"></a><span class="lineno"> 552</span> }</div><div class="line"><a name="l00553"></a><span class="lineno"> 553</span> </div><div class="line"><a name="l00554"></a><span class="lineno"> 554</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l00555"></a><span class="lineno"> 555</span> </div><div class="line"><a name="l00591"></a><span class="lineno"> 591</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_interleaved_transposed_bifrost_dot8(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00593"></a><span class="lineno"> 593</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00594"></a><span class="lineno"> 594</span>  uint src0_stride_z,</div><div class="line"><a name="l00595"></a><span class="lineno"> 595</span>  uint src1_stride_z,</div><div class="line"><a name="l00596"></a><span class="lineno"> 596</span>  uint dst_stride_z</div><div class="line"><a name="l00597"></a><span class="lineno"> 597</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</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>  uint cross_plane_pad</div><div class="line"><a name="l00600"></a><span class="lineno"> 600</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  )</div><div class="line"><a name="l00602"></a><span class="lineno"> 602</span> {</div><div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  <span class="comment">// Offset</span></div><div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_a = (get_global_id(1) % MULT_INTERLEAVE4X4_HEIGHT) * 4;</div><div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> offset_row_b = (get_global_id(0) % TRANSPOSE1XW_WIDTH_STEP) * 4;</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="comment">// src_addr_a = address of matrix A</span></div><div class="line"><a name="l00608"></a><span class="lineno"> 608</span>  <span class="comment">// src_addr_b = address of matrix B</span></div><div class="line"><a name="l00609"></a><span class="lineno"> 609</span>  __global uchar *src_addr_a = (__global uchar *)(src0_ptr + (get_global_id(1) / MULT_INTERLEAVE4X4_HEIGHT) * src0_stride_y + get_global_id(2) * src0_stride_z + src0_offset_first_element_in_bytes);</div><div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  __global uchar *src_addr_b = (__global uchar *)(src1_ptr + (get_global_id(0) / TRANSPOSE1XW_WIDTH_STEP) * src1_stride_y + src1_offset_first_element_in_bytes);</div><div class="line"><a name="l00611"></a><span class="lineno"> 611</span> </div><div class="line"><a name="l00612"></a><span class="lineno"> 612</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00613"></a><span class="lineno"> 613</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00614"></a><span class="lineno"> 614</span>  src_addr_b += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00615"></a><span class="lineno"> 615</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00616"></a><span class="lineno"> 616</span>  src_addr_b += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l00617"></a><span class="lineno"> 617</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00618"></a><span class="lineno"> 618</span> </div><div class="line"><a name="l00619"></a><span class="lineno"> 619</span>  src_addr_a += offset_row_a;</div><div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  src_addr_b += offset_row_b;</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">// Reset accumulators</span></div><div class="line"><a name="l00623"></a><span class="lineno"> 623</span>  uint c00 = 0;</div><div class="line"><a name="l00624"></a><span class="lineno"> 624</span>  uint c01 = 0;</div><div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  uint c02 = 0;</div><div class="line"><a name="l00626"></a><span class="lineno"> 626</span>  uint c03 = 0;</div><div class="line"><a name="l00627"></a><span class="lineno"> 627</span> </div><div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  uint c10 = 0;</div><div class="line"><a name="l00629"></a><span class="lineno"> 629</span>  uint c11 = 0;</div><div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  uint c12 = 0;</div><div class="line"><a name="l00631"></a><span class="lineno"> 631</span>  uint c13 = 0;</div><div class="line"><a name="l00632"></a><span class="lineno"> 632</span> </div><div class="line"><a name="l00633"></a><span class="lineno"> 633</span>  uint c20 = 0;</div><div class="line"><a name="l00634"></a><span class="lineno"> 634</span>  uint c21 = 0;</div><div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  uint c22 = 0;</div><div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  uint c23 = 0;</div><div class="line"><a name="l00637"></a><span class="lineno"> 637</span> </div><div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  uint c30 = 0;</div><div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  uint c31 = 0;</div><div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  uint c32 = 0;</div><div class="line"><a name="l00641"></a><span class="lineno"> 641</span>  uint c33 = 0;</div><div class="line"><a name="l00642"></a><span class="lineno"> 642</span> </div><div class="line"><a name="l00643"></a><span class="lineno"> 643</span> <span class="preprocessor">#define COLS_MTX_B (COLS_B / (16 * MULT_TRANSPOSE1XW_WIDTH))</span></div><div class="line"><a name="l00644"></a><span class="lineno"> 644</span> </div><div class="line"><a name="l00645"></a><span class="lineno"> 645</span> <span class="preprocessor">#if MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00646"></a><span class="lineno"> 646</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l00647"></a><span class="lineno"> 647</span>  <span class="keywordflow">for</span>(; i <= (int)(COLS_MTX_B - 8); i += 8)</div><div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  {</div><div class="line"><a name="l00649"></a><span class="lineno"> 649</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  uchar16 a0 = vload16(0, src_addr_a);</div><div class="line"><a name="l00651"></a><span class="lineno"> 651</span>  uchar4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00652"></a><span class="lineno"> 652</span>  uchar4 b1 = vload4(0, src_addr_b + 4 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00653"></a><span class="lineno"> 653</span>  uchar4 b2 = vload4(0, src_addr_b + 8 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00654"></a><span class="lineno"> 654</span>  uchar4 b3 = vload4(0, src_addr_b + 12 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00655"></a><span class="lineno"> 655</span>  uchar4 b4 = vload4(0, src_addr_b + 16 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00656"></a><span class="lineno"> 656</span>  uchar4 b5 = vload4(0, src_addr_b + 20 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00657"></a><span class="lineno"> 657</span>  uchar4 b6 = vload4(0, src_addr_b + 24 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00658"></a><span class="lineno"> 658</span>  uchar4 b7 = vload4(0, src_addr_b + 28 * TRANSPOSE1XW_WIDTH_STEP);</div><div class="line"><a name="l00659"></a><span class="lineno"> 659</span> </div><div class="line"><a name="l00660"></a><span class="lineno"> 660</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00661"></a><span class="lineno"> 661</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c00);</div><div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c01);</div><div class="line"><a name="l00663"></a><span class="lineno"> 663</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c02);</div><div class="line"><a name="l00664"></a><span class="lineno"> 664</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c03);</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>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c10);</div><div class="line"><a name="l00667"></a><span class="lineno"> 667</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c11);</div><div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c12);</div><div class="line"><a name="l00669"></a><span class="lineno"> 669</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c13);</div><div class="line"><a name="l00670"></a><span class="lineno"> 670</span> </div><div class="line"><a name="l00671"></a><span class="lineno"> 671</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c20);</div><div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c21);</div><div class="line"><a name="l00673"></a><span class="lineno"> 673</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c22);</div><div class="line"><a name="l00674"></a><span class="lineno"> 674</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c23);</div><div class="line"><a name="l00675"></a><span class="lineno"> 675</span> </div><div class="line"><a name="l00676"></a><span class="lineno"> 676</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), c30);</div><div class="line"><a name="l00677"></a><span class="lineno"> 677</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), c31);</div><div class="line"><a name="l00678"></a><span class="lineno"> 678</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), c32);</div><div class="line"><a name="l00679"></a><span class="lineno"> 679</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), c33);</div><div class="line"><a name="l00680"></a><span class="lineno"> 680</span> </div><div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00682"></a><span class="lineno"> 682</span>  a0 = vload16(0, src_addr_a + 16);</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>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c00);</div><div class="line"><a name="l00685"></a><span class="lineno"> 685</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c01);</div><div class="line"><a name="l00686"></a><span class="lineno"> 686</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c02);</div><div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  ARM_DOT((uchar4)(a0.s0123), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c03);</div><div class="line"><a name="l00688"></a><span class="lineno"> 688</span> </div><div class="line"><a name="l00689"></a><span class="lineno"> 689</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c10);</div><div class="line"><a name="l00690"></a><span class="lineno"> 690</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c11);</div><div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c12);</div><div class="line"><a name="l00692"></a><span class="lineno"> 692</span>  ARM_DOT((uchar4)(a0.s4567), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c13);</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>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c20);</div><div class="line"><a name="l00695"></a><span class="lineno"> 695</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c21);</div><div class="line"><a name="l00696"></a><span class="lineno"> 696</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c22);</div><div class="line"><a name="l00697"></a><span class="lineno"> 697</span>  ARM_DOT((uchar4)(a0.s89AB), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c23);</div><div class="line"><a name="l00698"></a><span class="lineno"> 698</span> </div><div class="line"><a name="l00699"></a><span class="lineno"> 699</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s0, b5.s0, b6.s0, b7.s0), c30);</div><div class="line"><a name="l00700"></a><span class="lineno"> 700</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s1, b5.s1, b6.s1, b7.s1), c31);</div><div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s2, b5.s2, b6.s2, b7.s2), c32);</div><div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  ARM_DOT((uchar4)(a0.sCDEF), (uchar4)(b4.s3, b5.s3, b6.s3, b7.s3), c33);</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>  src_addr_a += 32;</div><div class="line"><a name="l00705"></a><span class="lineno"> 705</span>  src_addr_b += 32 * TRANSPOSE1XW_WIDTH_STEP;</div><div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  }</div><div class="line"><a name="l00707"></a><span class="lineno"> 707</span> <span class="preprocessor">#endif // MULT_INTERLEAVE4X4_HEIGHT == 1</span></div><div class="line"><a name="l00708"></a><span class="lineno"> 708</span>  <span class="keywordtype">int</span> i_left_over = 0;</div><div class="line"><a name="l00709"></a><span class="lineno"> 709</span>  <span class="keywordflow">for</span>(; i < (int)(COLS_MTX_B); ++i)</div><div class="line"><a name="l00710"></a><span class="lineno"> 710</span>  {</div><div class="line"><a name="l00711"></a><span class="lineno"> 711</span>  <span class="comment">// Load values from matrix A (interleaved) and matrix B (transposed)</span></div><div class="line"><a name="l00712"></a><span class="lineno"> 712</span>  uchar16 a0 = vload16(0, src_addr_a + (i_left_over % 4) + ((i_left_over / 4) * 16));</div><div class="line"><a name="l00713"></a><span class="lineno"> 713</span>  uchar4 b0 = vload4(0, src_addr_b);</div><div class="line"><a name="l00714"></a><span class="lineno"> 714</span> </div><div class="line"><a name="l00715"></a><span class="lineno"> 715</span>  c00 += a0.s0 * b0.s0;</div><div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  c01 += a0.s0 * b0.s1;</div><div class="line"><a name="l00717"></a><span class="lineno"> 717</span>  c02 += a0.s0 * b0.s2;</div><div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  c03 += a0.s0 * b0.s3;</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>  c10 += a0.s4 * b0.s0;</div><div class="line"><a name="l00721"></a><span class="lineno"> 721</span>  c11 += a0.s4 * b0.s1;</div><div class="line"><a name="l00722"></a><span class="lineno"> 722</span>  c12 += a0.s4 * b0.s2;</div><div class="line"><a name="l00723"></a><span class="lineno"> 723</span>  c13 += a0.s4 * b0.s3;</div><div class="line"><a name="l00724"></a><span class="lineno"> 724</span> </div><div class="line"><a name="l00725"></a><span class="lineno"> 725</span>  c20 += a0.s8 * b0.s0;</div><div class="line"><a name="l00726"></a><span class="lineno"> 726</span>  c21 += a0.s8 * b0.s1;</div><div class="line"><a name="l00727"></a><span class="lineno"> 727</span>  c22 += a0.s8 * b0.s2;</div><div class="line"><a name="l00728"></a><span class="lineno"> 728</span>  c23 += a0.s8 * b0.s3;</div><div class="line"><a name="l00729"></a><span class="lineno"> 729</span> </div><div class="line"><a name="l00730"></a><span class="lineno"> 730</span>  c30 += a0.sC * b0.s0;</div><div class="line"><a name="l00731"></a><span class="lineno"> 731</span>  c31 += a0.sC * b0.s1;</div><div class="line"><a name="l00732"></a><span class="lineno"> 732</span>  c32 += a0.sC * b0.s2;</div><div class="line"><a name="l00733"></a><span class="lineno"> 733</span>  c33 += a0.sC * b0.s3;</div><div class="line"><a name="l00734"></a><span class="lineno"> 734</span> </div><div class="line"><a name="l00735"></a><span class="lineno"> 735</span>  i_left_over++;</div><div class="line"><a name="l00736"></a><span class="lineno"> 736</span>  src_addr_b += 4 * TRANSPOSE1XW_WIDTH_STEP;</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>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00740"></a><span class="lineno"> 740</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00741"></a><span class="lineno"> 741</span> </div><div class="line"><a name="l00742"></a><span class="lineno"> 742</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00743"></a><span class="lineno"> 743</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00744"></a><span class="lineno"> 744</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00745"></a><span class="lineno"> 745</span>  <span class="comment">//</span></div><div class="line"><a name="l00746"></a><span class="lineno"> 746</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l00748"></a><span class="lineno"> 748</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00749"></a><span class="lineno"> 749</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00751"></a><span class="lineno"> 751</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l00752"></a><span class="lineno"> 752</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00753"></a><span class="lineno"> 753</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00754"></a><span class="lineno"> 754</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l00755"></a><span class="lineno"> 755</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00756"></a><span class="lineno"> 756</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00757"></a><span class="lineno"> 757</span> </div><div class="line"><a name="l00758"></a><span class="lineno"> 758</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * 4) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00759"></a><span class="lineno"> 759</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * 4)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00760"></a><span class="lineno"> 760</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l00761"></a><span class="lineno"> 761</span> </div><div class="line"><a name="l00762"></a><span class="lineno"> 762</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00763"></a><span class="lineno"> 763</span>  zout *= (cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l00764"></a><span class="lineno"> 764</span> </div><div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00766"></a><span class="lineno"> 766</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00767"></a><span class="lineno"> 767</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += get_global_id(2) * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00768"></a><span class="lineno"> 768</span> </div><div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00770"></a><span class="lineno"> 770</span>  vstore4((int4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  vstore4((int4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  vstore4((int4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l00773"></a><span class="lineno"> 773</span>  vstore4((int4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l00774"></a><span class="lineno"> 774</span> </div><div class="line"><a name="l00775"></a><span class="lineno"> 775</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00776"></a><span class="lineno"> 776</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += get_global_id(2) * dst_stride_z;</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>  <span class="comment">// Store 4x4 block</span></div><div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  vstore4((int4)(c00, c01, c02, c03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l00781"></a><span class="lineno"> 781</span>  vstore4((int4)(c10, c11, c12, c13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  vstore4((int4)(c20, c21, c22, c23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  vstore4((int4)(c30, c31, c32, c33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l00784"></a><span class="lineno"> 784</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00785"></a><span class="lineno"> 785</span> }</div><div class="line"><a name="l00786"></a><span class="lineno"> 786</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></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> <span class="preprocessor">#endif // defined(COLS_B) && defined(MULT_INTERLEAVE4X4_HEIGHT) && defined(TRANSPOSE1XW_WIDTH_STEP)</span></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="preprocessor">#if defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A)</span></div><div class="line"><a name="l00791"></a><span class="lineno"> 791</span> <span class="preprocessor">#define VECTOR_UCHAR VEC_DATA_TYPE(uchar, NUM_ELEMS_PROCESSED_PER_THREAD_X)</span></div><div class="line"><a name="l00792"></a><span class="lineno"> 792</span> <span class="preprocessor">#define VECTOR_UINT VEC_DATA_TYPE(uint, NUM_ELEMS_PROCESSED_PER_THREAD_X)</span></div><div class="line"><a name="l00793"></a><span class="lineno"> 793</span> <span class="preprocessor">#define VECTOR_INT VEC_DATA_TYPE(int, NUM_ELEMS_PROCESSED_PER_THREAD_X)</span></div><div class="line"><a name="l00794"></a><span class="lineno"> 794</span> </div><div class="line"><a name="l00829"></a><span class="lineno"> 829</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_midgard(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l00830"></a><span class="lineno"> 830</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l00831"></a><span class="lineno"> 831</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l00832"></a><span class="lineno"> 832</span>  uint src0_stride_z,</div><div class="line"><a name="l00833"></a><span class="lineno"> 833</span>  uint src1_stride_z,</div><div class="line"><a name="l00834"></a><span class="lineno"> 834</span>  uint dst_stride_z</div><div class="line"><a name="l00835"></a><span class="lineno"> 835</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l00836"></a><span class="lineno"> 836</span>  ,</div><div class="line"><a name="l00837"></a><span class="lineno"> 837</span>  uint src_cross_plane_pad</div><div class="line"><a name="l00838"></a><span class="lineno"> 838</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l00839"></a><span class="lineno"> 839</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l00840"></a><span class="lineno"> 840</span>  ,</div><div class="line"><a name="l00841"></a><span class="lineno"> 841</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l00842"></a><span class="lineno"> 842</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l00843"></a><span class="lineno"> 843</span>  )</div><div class="line"><a name="l00844"></a><span class="lineno"> 844</span> {</div><div class="line"><a name="l00845"></a><span class="lineno"> 845</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l00846"></a><span class="lineno"> 846</span> </div><div class="line"><a name="l00847"></a><span class="lineno"> 847</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l00848"></a><span class="lineno"> 848</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l00849"></a><span class="lineno"> 849</span> </div><div class="line"><a name="l00850"></a><span class="lineno"> 850</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l00851"></a><span class="lineno"> 851</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l00852"></a><span class="lineno"> 852</span> </div><div class="line"><a name="l00853"></a><span class="lineno"> 853</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l00854"></a><span class="lineno"> 854</span>  src_addr.s1 += idx;</div><div class="line"><a name="l00855"></a><span class="lineno"> 855</span> </div><div class="line"><a name="l00856"></a><span class="lineno"> 856</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00857"></a><span class="lineno"> 857</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00858"></a><span class="lineno"> 858</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00859"></a><span class="lineno"> 859</span>  <span class="comment">//</span></div><div class="line"><a name="l00860"></a><span class="lineno"> 860</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00861"></a><span class="lineno"> 861</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l00862"></a><span class="lineno"> 862</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00863"></a><span class="lineno"> 863</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00864"></a><span class="lineno"> 864</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00865"></a><span class="lineno"> 865</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l00866"></a><span class="lineno"> 866</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l00867"></a><span class="lineno"> 867</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00868"></a><span class="lineno"> 868</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l00869"></a><span class="lineno"> 869</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00870"></a><span class="lineno"> 870</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l00871"></a><span class="lineno"> 871</span> </div><div class="line"><a name="l00872"></a><span class="lineno"> 872</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l00873"></a><span class="lineno"> 873</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l00874"></a><span class="lineno"> 874</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l00875"></a><span class="lineno"> 875</span> </div><div class="line"><a name="l00876"></a><span class="lineno"> 876</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l00877"></a><span class="lineno"> 877</span>  zin *= (src_cross_plane_pad * src0_stride_y);</div><div class="line"><a name="l00878"></a><span class="lineno"> 878</span> </div><div class="line"><a name="l00879"></a><span class="lineno"> 879</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l00880"></a><span class="lineno"> 880</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l00881"></a><span class="lineno"> 881</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l00882"></a><span class="lineno"> 882</span> </div><div class="line"><a name="l00883"></a><span class="lineno"> 883</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00884"></a><span class="lineno"> 884</span> </div><div class="line"><a name="l00885"></a><span class="lineno"> 885</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l00886"></a><span class="lineno"> 886</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l00887"></a><span class="lineno"> 887</span> </div><div class="line"><a name="l00888"></a><span class="lineno"> 888</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l00889"></a><span class="lineno"> 889</span> </div><div class="line"><a name="l00890"></a><span class="lineno"> 890</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00891"></a><span class="lineno"> 891</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l00892"></a><span class="lineno"> 892</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l00893"></a><span class="lineno"> 893</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00894"></a><span class="lineno"> 894</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l00895"></a><span class="lineno"> 895</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l00896"></a><span class="lineno"> 896</span> </div><div class="line"><a name="l00897"></a><span class="lineno"> 897</span>  <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + COLS_A;</div><div class="line"><a name="l00898"></a><span class="lineno"> 898</span> </div><div class="line"><a name="l00899"></a><span class="lineno"> 899</span>  VECTOR_UINT acc0 = 0;</div><div class="line"><a name="l00900"></a><span class="lineno"> 900</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00901"></a><span class="lineno"> 901</span>  VECTOR_UINT acc1 = 0;</div><div class="line"><a name="l00902"></a><span class="lineno"> 902</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00903"></a><span class="lineno"> 903</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00904"></a><span class="lineno"> 904</span>  VECTOR_UINT acc2 = 0;</div><div class="line"><a name="l00905"></a><span class="lineno"> 905</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00906"></a><span class="lineno"> 906</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00907"></a><span class="lineno"> 907</span>  VECTOR_UINT acc3 = 0;</div><div class="line"><a name="l00908"></a><span class="lineno"> 908</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00909"></a><span class="lineno"> 909</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00910"></a><span class="lineno"> 910</span>  VECTOR_UINT acc4 = 0;</div><div class="line"><a name="l00911"></a><span class="lineno"> 911</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00912"></a><span class="lineno"> 912</span> </div><div class="line"><a name="l00913"></a><span class="lineno"> 913</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 2); src_addr += (int2)(2, 2 * src1_stride_y))</div><div class="line"><a name="l00914"></a><span class="lineno"> 914</span>  {</div><div class="line"><a name="l00915"></a><span class="lineno"> 915</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l00916"></a><span class="lineno"> 916</span>  uchar2 a0 = vload2(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y);</div><div class="line"><a name="l00917"></a><span class="lineno"> 917</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00918"></a><span class="lineno"> 918</span>  uchar2 a1 = vload2(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y);</div><div class="line"><a name="l00919"></a><span class="lineno"> 919</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00920"></a><span class="lineno"> 920</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00921"></a><span class="lineno"> 921</span>  uchar2 a2 = vload2(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y);</div><div class="line"><a name="l00922"></a><span class="lineno"> 922</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00923"></a><span class="lineno"> 923</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00924"></a><span class="lineno"> 924</span>  uchar2 a3 = vload2(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y);</div><div class="line"><a name="l00925"></a><span class="lineno"> 925</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00926"></a><span class="lineno"> 926</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00927"></a><span class="lineno"> 927</span>  uchar2 a4 = vload2(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y);</div><div class="line"><a name="l00928"></a><span class="lineno"> 928</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00929"></a><span class="lineno"> 929</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l00930"></a><span class="lineno"> 930</span>  VECTOR_UCHAR b0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1);</div><div class="line"><a name="l00931"></a><span class="lineno"> 931</span>  VECTOR_UCHAR b1 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1 + src1_stride_y);</div><div class="line"><a name="l00932"></a><span class="lineno"> 932</span> </div><div class="line"><a name="l00933"></a><span class="lineno"> 933</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00934"></a><span class="lineno"> 934</span>  acc0 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a0.s0;</div><div class="line"><a name="l00935"></a><span class="lineno"> 935</span>  acc0 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b1, VECTOR_UINT) * (VECTOR_UINT)a0.s1;</div><div class="line"><a name="l00936"></a><span class="lineno"> 936</span> #<span class="keywordflow">if</span> NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</div><div class="line"><a name="l00937"></a><span class="lineno"> 937</span>  acc1 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a1.s0;</div><div class="line"><a name="l00938"></a><span class="lineno"> 938</span>  acc1 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b1, VECTOR_UINT) * (VECTOR_UINT)a1.s1;</div><div class="line"><a name="l00939"></a><span class="lineno"> 939</span> #endif <span class="comment">// NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00940"></a><span class="lineno"> 940</span> #<span class="keywordflow">if</span> NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</div><div class="line"><a name="l00941"></a><span class="lineno"> 941</span>  acc2 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a2.s0;</div><div class="line"><a name="l00942"></a><span class="lineno"> 942</span>  acc2 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b1, VECTOR_UINT) * (VECTOR_UINT)a2.s1;</div><div class="line"><a name="l00943"></a><span class="lineno"> 943</span> #endif <span class="comment">// NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00944"></a><span class="lineno"> 944</span> #<span class="keywordflow">if</span> NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</div><div class="line"><a name="l00945"></a><span class="lineno"> 945</span>  acc3 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a3.s0;</div><div class="line"><a name="l00946"></a><span class="lineno"> 946</span>  acc3 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b1, VECTOR_UINT) * (VECTOR_UINT)a3.s1;</div><div class="line"><a name="l00947"></a><span class="lineno"> 947</span> #endif <span class="comment">// NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00948"></a><span class="lineno"> 948</span> #<span class="keywordflow">if</span> NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</div><div class="line"><a name="l00949"></a><span class="lineno"> 949</span>  acc4 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a4.s0;</div><div class="line"><a name="l00950"></a><span class="lineno"> 950</span>  acc4 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b1, VECTOR_UINT) * (VECTOR_UINT)a4.s1;</div><div class="line"><a name="l00951"></a><span class="lineno"> 951</span> #endif <span class="comment">// NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00952"></a><span class="lineno"> 952</span>  }</div><div class="line"><a name="l00953"></a><span class="lineno"> 953</span> </div><div class="line"><a name="l00954"></a><span class="lineno"> 954</span>  <span class="keywordflow">for</span>(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y))</div><div class="line"><a name="l00955"></a><span class="lineno"> 955</span>  {</div><div class="line"><a name="l00956"></a><span class="lineno"> 956</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l00957"></a><span class="lineno"> 957</span>  uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y);</div><div class="line"><a name="l00958"></a><span class="lineno"> 958</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00959"></a><span class="lineno"> 959</span>  uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y);</div><div class="line"><a name="l00960"></a><span class="lineno"> 960</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00961"></a><span class="lineno"> 961</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00962"></a><span class="lineno"> 962</span>  uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y);</div><div class="line"><a name="l00963"></a><span class="lineno"> 963</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00964"></a><span class="lineno"> 964</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00965"></a><span class="lineno"> 965</span>  uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y);</div><div class="line"><a name="l00966"></a><span class="lineno"> 966</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00967"></a><span class="lineno"> 967</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00968"></a><span class="lineno"> 968</span>  uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y);</div><div class="line"><a name="l00969"></a><span class="lineno"> 969</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00970"></a><span class="lineno"> 970</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l00971"></a><span class="lineno"> 971</span>  VECTOR_UCHAR b0 = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a287e2fc366c312b468382c95bb90f91f">VLOAD</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)(0, src1_ptr + src_addr.s1);</div><div class="line"><a name="l00972"></a><span class="lineno"> 972</span> </div><div class="line"><a name="l00973"></a><span class="lineno"> 973</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00974"></a><span class="lineno"> 974</span>  acc0 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a0;</div><div class="line"><a name="l00975"></a><span class="lineno"> 975</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00976"></a><span class="lineno"> 976</span>  acc1 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a1;</div><div class="line"><a name="l00977"></a><span class="lineno"> 977</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l00978"></a><span class="lineno"> 978</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00979"></a><span class="lineno"> 979</span>  acc2 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a2;</div><div class="line"><a name="l00980"></a><span class="lineno"> 980</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l00981"></a><span class="lineno"> 981</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00982"></a><span class="lineno"> 982</span>  acc3 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a3;</div><div class="line"><a name="l00983"></a><span class="lineno"> 983</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l00984"></a><span class="lineno"> 984</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00985"></a><span class="lineno"> 985</span>  acc4 += <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(b0, VECTOR_UINT) * (VECTOR_UINT)a4;</div><div class="line"><a name="l00986"></a><span class="lineno"> 986</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l00987"></a><span class="lineno"> 987</span>  }</div><div class="line"><a name="l00988"></a><span class="lineno"> 988</span> </div><div class="line"><a name="l00989"></a><span class="lineno"> 989</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l00990"></a><span class="lineno"> 990</span> </div><div class="line"><a name="l00991"></a><span class="lineno"> 991</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l00992"></a><span class="lineno"> 992</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l00993"></a><span class="lineno"> 993</span> </div><div class="line"><a name="l00994"></a><span class="lineno"> 994</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l00995"></a><span class="lineno"> 995</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l00996"></a><span class="lineno"> 996</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l00997"></a><span class="lineno"> 997</span>  <span class="comment">//</span></div><div class="line"><a name="l00998"></a><span class="lineno"> 998</span>  <span class="comment">// | |</span></div><div class="line"><a name="l00999"></a><span class="lineno"> 999</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l01000"></a><span class="lineno"> 1000</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01001"></a><span class="lineno"> 1001</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01002"></a><span class="lineno"> 1002</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01003"></a><span class="lineno"> 1003</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l01004"></a><span class="lineno"> 1004</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01005"></a><span class="lineno"> 1005</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01006"></a><span class="lineno"> 1006</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l01007"></a><span class="lineno"> 1007</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01008"></a><span class="lineno"> 1008</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01009"></a><span class="lineno"> 1009</span> </div><div class="line"><a name="l01010"></a><span class="lineno"> 1010</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01011"></a><span class="lineno"> 1011</span>  uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D;</div><div class="line"><a name="l01012"></a><span class="lineno"> 1012</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01013"></a><span class="lineno"> 1013</span> </div><div class="line"><a name="l01014"></a><span class="lineno"> 1014</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01015"></a><span class="lineno"> 1015</span>  zout *= (dst_cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l01016"></a><span class="lineno"> 1016</span> </div><div class="line"><a name="l01017"></a><span class="lineno"> 1017</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01018"></a><span class="lineno"> 1018</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01019"></a><span class="lineno"> 1019</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01020"></a><span class="lineno"> 1020</span> </div><div class="line"><a name="l01021"></a><span class="lineno"> 1021</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01022"></a><span class="lineno"> 1022</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01023"></a><span class="lineno"> 1023</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc0, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l01024"></a><span class="lineno"> 1024</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01025"></a><span class="lineno"> 1025</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01026"></a><span class="lineno"> 1026</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc1, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l01027"></a><span class="lineno"> 1027</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01028"></a><span class="lineno"> 1028</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01029"></a><span class="lineno"> 1029</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01030"></a><span class="lineno"> 1030</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc2, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l01031"></a><span class="lineno"> 1031</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01032"></a><span class="lineno"> 1032</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01033"></a><span class="lineno"> 1033</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01034"></a><span class="lineno"> 1034</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc3, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l01035"></a><span class="lineno"> 1035</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01036"></a><span class="lineno"> 1036</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01037"></a><span class="lineno"> 1037</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01038"></a><span class="lineno"> 1038</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc4, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 4 * dst_stride_y + zout.s4));</div><div class="line"><a name="l01039"></a><span class="lineno"> 1039</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01040"></a><span class="lineno"> 1040</span> </div><div class="line"><a name="l01041"></a><span class="lineno"> 1041</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01042"></a><span class="lineno"> 1042</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01043"></a><span class="lineno"> 1043</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z;</div><div class="line"><a name="l01044"></a><span class="lineno"> 1044</span> </div><div class="line"><a name="l01045"></a><span class="lineno"> 1045</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01046"></a><span class="lineno"> 1046</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01047"></a><span class="lineno"> 1047</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc0, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l01048"></a><span class="lineno"> 1048</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01049"></a><span class="lineno"> 1049</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01050"></a><span class="lineno"> 1050</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc1, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l01051"></a><span class="lineno"> 1051</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01052"></a><span class="lineno"> 1052</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01053"></a><span class="lineno"> 1053</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01054"></a><span class="lineno"> 1054</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc2, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l01055"></a><span class="lineno"> 1055</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01056"></a><span class="lineno"> 1056</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01057"></a><span class="lineno"> 1057</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01058"></a><span class="lineno"> 1058</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc3, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l01059"></a><span class="lineno"> 1059</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01060"></a><span class="lineno"> 1060</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01061"></a><span class="lineno"> 1061</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#acb282042d1edeeaa3cc979a206f78b54">VSTORE</a>(NUM_ELEMS_PROCESSED_PER_THREAD_X)</div><div class="line"><a name="l01062"></a><span class="lineno"> 1062</span>  (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a>(acc4, VECTOR_INT), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 4 * dst_stride_y));</div><div class="line"><a name="l01063"></a><span class="lineno"> 1063</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01064"></a><span class="lineno"> 1064</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01065"></a><span class="lineno"> 1065</span> }</div><div class="line"><a name="l01066"></a><span class="lineno"> 1066</span> </div><div class="line"><a name="l01102"></a><span class="lineno"> 1102</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_bifrost(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01103"></a><span class="lineno"> 1103</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01104"></a><span class="lineno"> 1104</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01105"></a><span class="lineno"> 1105</span>  uint src0_stride_z,</div><div class="line"><a name="l01106"></a><span class="lineno"> 1106</span>  uint src1_stride_z,</div><div class="line"><a name="l01107"></a><span class="lineno"> 1107</span>  uint dst_stride_z</div><div class="line"><a name="l01108"></a><span class="lineno"> 1108</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l01109"></a><span class="lineno"> 1109</span>  ,</div><div class="line"><a name="l01110"></a><span class="lineno"> 1110</span>  uint src_cross_plane_pad</div><div class="line"><a name="l01111"></a><span class="lineno"> 1111</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l01112"></a><span class="lineno"> 1112</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l01113"></a><span class="lineno"> 1113</span>  ,</div><div class="line"><a name="l01114"></a><span class="lineno"> 1114</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l01115"></a><span class="lineno"> 1115</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D</span></div><div class="line"><a name="l01116"></a><span class="lineno"> 1116</span>  )</div><div class="line"><a name="l01117"></a><span class="lineno"> 1117</span> {</div><div class="line"><a name="l01118"></a><span class="lineno"> 1118</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01119"></a><span class="lineno"> 1119</span> </div><div class="line"><a name="l01120"></a><span class="lineno"> 1120</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01121"></a><span class="lineno"> 1121</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01122"></a><span class="lineno"> 1122</span> </div><div class="line"><a name="l01123"></a><span class="lineno"> 1123</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01124"></a><span class="lineno"> 1124</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01125"></a><span class="lineno"> 1125</span> </div><div class="line"><a name="l01126"></a><span class="lineno"> 1126</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01127"></a><span class="lineno"> 1127</span>  src_addr.s1 += idx;</div><div class="line"><a name="l01128"></a><span class="lineno"> 1128</span> </div><div class="line"><a name="l01129"></a><span class="lineno"> 1129</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01130"></a><span class="lineno"> 1130</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01131"></a><span class="lineno"> 1131</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01132"></a><span class="lineno"> 1132</span>  <span class="comment">//</span></div><div class="line"><a name="l01133"></a><span class="lineno"> 1133</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01134"></a><span class="lineno"> 1134</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l01135"></a><span class="lineno"> 1135</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01136"></a><span class="lineno"> 1136</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01137"></a><span class="lineno"> 1137</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01138"></a><span class="lineno"> 1138</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l01139"></a><span class="lineno"> 1139</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01140"></a><span class="lineno"> 1140</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01141"></a><span class="lineno"> 1141</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l01142"></a><span class="lineno"> 1142</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01143"></a><span class="lineno"> 1143</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01144"></a><span class="lineno"> 1144</span> </div><div class="line"><a name="l01145"></a><span class="lineno"> 1145</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01146"></a><span class="lineno"> 1146</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01147"></a><span class="lineno"> 1147</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l01148"></a><span class="lineno"> 1148</span> </div><div class="line"><a name="l01149"></a><span class="lineno"> 1149</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01150"></a><span class="lineno"> 1150</span>  zin *= (src_cross_plane_pad * src0_stride_y);</div><div class="line"><a name="l01151"></a><span class="lineno"> 1151</span> </div><div class="line"><a name="l01152"></a><span class="lineno"> 1152</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01153"></a><span class="lineno"> 1153</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01154"></a><span class="lineno"> 1154</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01155"></a><span class="lineno"> 1155</span> </div><div class="line"><a name="l01156"></a><span class="lineno"> 1156</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01157"></a><span class="lineno"> 1157</span> </div><div class="line"><a name="l01158"></a><span class="lineno"> 1158</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01159"></a><span class="lineno"> 1159</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01160"></a><span class="lineno"> 1160</span> </div><div class="line"><a name="l01161"></a><span class="lineno"> 1161</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01162"></a><span class="lineno"> 1162</span> </div><div class="line"><a name="l01163"></a><span class="lineno"> 1163</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01164"></a><span class="lineno"> 1164</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01165"></a><span class="lineno"> 1165</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01166"></a><span class="lineno"> 1166</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01167"></a><span class="lineno"> 1167</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01168"></a><span class="lineno"> 1168</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01169"></a><span class="lineno"> 1169</span> </div><div class="line"><a name="l01170"></a><span class="lineno"> 1170</span>  <span class="keywordtype">int</span> end_row_vec_a = src_addr.s0 + COLS_A;</div><div class="line"><a name="l01171"></a><span class="lineno"> 1171</span> </div><div class="line"><a name="l01172"></a><span class="lineno"> 1172</span>  uint acc00 = 0;</div><div class="line"><a name="l01173"></a><span class="lineno"> 1173</span>  uint acc01 = 0;</div><div class="line"><a name="l01174"></a><span class="lineno"> 1174</span>  uint acc02 = 0;</div><div class="line"><a name="l01175"></a><span class="lineno"> 1175</span>  uint acc03 = 0;</div><div class="line"><a name="l01176"></a><span class="lineno"> 1176</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01177"></a><span class="lineno"> 1177</span>  uint acc10 = 0;</div><div class="line"><a name="l01178"></a><span class="lineno"> 1178</span>  uint acc11 = 0;</div><div class="line"><a name="l01179"></a><span class="lineno"> 1179</span>  uint acc12 = 0;</div><div class="line"><a name="l01180"></a><span class="lineno"> 1180</span>  uint acc13 = 0;</div><div class="line"><a name="l01181"></a><span class="lineno"> 1181</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01182"></a><span class="lineno"> 1182</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01183"></a><span class="lineno"> 1183</span>  uint acc20 = 0;</div><div class="line"><a name="l01184"></a><span class="lineno"> 1184</span>  uint acc21 = 0;</div><div class="line"><a name="l01185"></a><span class="lineno"> 1185</span>  uint acc22 = 0;</div><div class="line"><a name="l01186"></a><span class="lineno"> 1186</span>  uint acc23 = 0;</div><div class="line"><a name="l01187"></a><span class="lineno"> 1187</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01188"></a><span class="lineno"> 1188</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01189"></a><span class="lineno"> 1189</span>  uint acc30 = 0;</div><div class="line"><a name="l01190"></a><span class="lineno"> 1190</span>  uint acc31 = 0;</div><div class="line"><a name="l01191"></a><span class="lineno"> 1191</span>  uint acc32 = 0;</div><div class="line"><a name="l01192"></a><span class="lineno"> 1192</span>  uint acc33 = 0;</div><div class="line"><a name="l01193"></a><span class="lineno"> 1193</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01194"></a><span class="lineno"> 1194</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01195"></a><span class="lineno"> 1195</span>  uint acc40 = 0;</div><div class="line"><a name="l01196"></a><span class="lineno"> 1196</span>  uint acc41 = 0;</div><div class="line"><a name="l01197"></a><span class="lineno"> 1197</span>  uint acc42 = 0;</div><div class="line"><a name="l01198"></a><span class="lineno"> 1198</span>  uint acc43 = 0;</div><div class="line"><a name="l01199"></a><span class="lineno"> 1199</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01200"></a><span class="lineno"> 1200</span> </div><div class="line"><a name="l01201"></a><span class="lineno"> 1201</span>  <span class="keywordflow">for</span>(; src_addr.s0 <= (end_row_vec_a - 4); src_addr += (int2)(4, 4 * src1_stride_y))</div><div class="line"><a name="l01202"></a><span class="lineno"> 1202</span>  {</div><div class="line"><a name="l01203"></a><span class="lineno"> 1203</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01204"></a><span class="lineno"> 1204</span>  uchar4 a0 = vload4(0, src0_ptr + src_addr.s0 + 0 * src0_stride_y);</div><div class="line"><a name="l01205"></a><span class="lineno"> 1205</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01206"></a><span class="lineno"> 1206</span>  uchar4 a1 = vload4(0, src0_ptr + src_addr.s0 + 1 * src0_stride_y);</div><div class="line"><a name="l01207"></a><span class="lineno"> 1207</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01208"></a><span class="lineno"> 1208</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01209"></a><span class="lineno"> 1209</span>  uchar4 a2 = vload4(0, src0_ptr + src_addr.s0 + 2 * src0_stride_y);</div><div class="line"><a name="l01210"></a><span class="lineno"> 1210</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01211"></a><span class="lineno"> 1211</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01212"></a><span class="lineno"> 1212</span>  uchar4 a3 = vload4(0, src0_ptr + src_addr.s0 + 3 * src0_stride_y);</div><div class="line"><a name="l01213"></a><span class="lineno"> 1213</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01214"></a><span class="lineno"> 1214</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01215"></a><span class="lineno"> 1215</span>  uchar4 a4 = vload4(0, src0_ptr + src_addr.s0 + 4 * src0_stride_y);</div><div class="line"><a name="l01216"></a><span class="lineno"> 1216</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01217"></a><span class="lineno"> 1217</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01218"></a><span class="lineno"> 1218</span>  uchar4 b0 = vload4(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);</div><div class="line"><a name="l01219"></a><span class="lineno"> 1219</span>  uchar4 b1 = vload4(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);</div><div class="line"><a name="l01220"></a><span class="lineno"> 1220</span>  uchar4 b2 = vload4(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);</div><div class="line"><a name="l01221"></a><span class="lineno"> 1221</span>  uchar4 b3 = vload4(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);</div><div class="line"><a name="l01222"></a><span class="lineno"> 1222</span> </div><div class="line"><a name="l01223"></a><span class="lineno"> 1223</span>  {</div><div class="line"><a name="l01224"></a><span class="lineno"> 1224</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01225"></a><span class="lineno"> 1225</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a0.s0;</div><div class="line"><a name="l01226"></a><span class="lineno"> 1226</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a0.s0;</div><div class="line"><a name="l01227"></a><span class="lineno"> 1227</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a0.s0;</div><div class="line"><a name="l01228"></a><span class="lineno"> 1228</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a0.s0;</div><div class="line"><a name="l01229"></a><span class="lineno"> 1229</span> </div><div class="line"><a name="l01230"></a><span class="lineno"> 1230</span>  ushort tmp4 = (ushort)b1.s0 * (ushort)a0.s1;</div><div class="line"><a name="l01231"></a><span class="lineno"> 1231</span>  ushort tmp5 = (ushort)b1.s1 * (ushort)a0.s1;</div><div class="line"><a name="l01232"></a><span class="lineno"> 1232</span>  ushort tmp6 = (ushort)b1.s2 * (ushort)a0.s1;</div><div class="line"><a name="l01233"></a><span class="lineno"> 1233</span>  ushort tmp7 = (ushort)b1.s3 * (ushort)a0.s1;</div><div class="line"><a name="l01234"></a><span class="lineno"> 1234</span> </div><div class="line"><a name="l01235"></a><span class="lineno"> 1235</span>  ushort tmp8 = (ushort)b2.s0 * (ushort)a0.s2;</div><div class="line"><a name="l01236"></a><span class="lineno"> 1236</span>  ushort tmp9 = (ushort)b2.s1 * (ushort)a0.s2;</div><div class="line"><a name="l01237"></a><span class="lineno"> 1237</span>  ushort tmpA = (ushort)b2.s2 * (ushort)a0.s2;</div><div class="line"><a name="l01238"></a><span class="lineno"> 1238</span>  ushort tmpB = (ushort)b2.s3 * (ushort)a0.s2;</div><div class="line"><a name="l01239"></a><span class="lineno"> 1239</span> </div><div class="line"><a name="l01240"></a><span class="lineno"> 1240</span>  ushort tmpC = (ushort)b3.s0 * (ushort)a0.s3;</div><div class="line"><a name="l01241"></a><span class="lineno"> 1241</span>  ushort tmpD = (ushort)b3.s1 * (ushort)a0.s3;</div><div class="line"><a name="l01242"></a><span class="lineno"> 1242</span>  ushort tmpE = (ushort)b3.s2 * (ushort)a0.s3;</div><div class="line"><a name="l01243"></a><span class="lineno"> 1243</span>  ushort tmpF = (ushort)b3.s3 * (ushort)a0.s3;</div><div class="line"><a name="l01244"></a><span class="lineno"> 1244</span> </div><div class="line"><a name="l01245"></a><span class="lineno"> 1245</span>  acc00 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);</div><div class="line"><a name="l01246"></a><span class="lineno"> 1246</span>  acc01 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);</div><div class="line"><a name="l01247"></a><span class="lineno"> 1247</span>  acc02 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);</div><div class="line"><a name="l01248"></a><span class="lineno"> 1248</span>  acc03 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);</div><div class="line"><a name="l01249"></a><span class="lineno"> 1249</span>  }</div><div class="line"><a name="l01250"></a><span class="lineno"> 1250</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01251"></a><span class="lineno"> 1251</span>  {</div><div class="line"><a name="l01252"></a><span class="lineno"> 1252</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01253"></a><span class="lineno"> 1253</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a1.s0;</div><div class="line"><a name="l01254"></a><span class="lineno"> 1254</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a1.s0;</div><div class="line"><a name="l01255"></a><span class="lineno"> 1255</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a1.s0;</div><div class="line"><a name="l01256"></a><span class="lineno"> 1256</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a1.s0;</div><div class="line"><a name="l01257"></a><span class="lineno"> 1257</span> </div><div class="line"><a name="l01258"></a><span class="lineno"> 1258</span>  ushort tmp4 = (ushort)b1.s0 * (ushort)a1.s1;</div><div class="line"><a name="l01259"></a><span class="lineno"> 1259</span>  ushort tmp5 = (ushort)b1.s1 * (ushort)a1.s1;</div><div class="line"><a name="l01260"></a><span class="lineno"> 1260</span>  ushort tmp6 = (ushort)b1.s2 * (ushort)a1.s1;</div><div class="line"><a name="l01261"></a><span class="lineno"> 1261</span>  ushort tmp7 = (ushort)b1.s3 * (ushort)a1.s1;</div><div class="line"><a name="l01262"></a><span class="lineno"> 1262</span> </div><div class="line"><a name="l01263"></a><span class="lineno"> 1263</span>  ushort tmp8 = (ushort)b2.s0 * (ushort)a1.s2;</div><div class="line"><a name="l01264"></a><span class="lineno"> 1264</span>  ushort tmp9 = (ushort)b2.s1 * (ushort)a1.s2;</div><div class="line"><a name="l01265"></a><span class="lineno"> 1265</span>  ushort tmpA = (ushort)b2.s2 * (ushort)a1.s2;</div><div class="line"><a name="l01266"></a><span class="lineno"> 1266</span>  ushort tmpB = (ushort)b2.s3 * (ushort)a1.s2;</div><div class="line"><a name="l01267"></a><span class="lineno"> 1267</span> </div><div class="line"><a name="l01268"></a><span class="lineno"> 1268</span>  ushort tmpC = (ushort)b3.s0 * (ushort)a1.s3;</div><div class="line"><a name="l01269"></a><span class="lineno"> 1269</span>  ushort tmpD = (ushort)b3.s1 * (ushort)a1.s3;</div><div class="line"><a name="l01270"></a><span class="lineno"> 1270</span>  ushort tmpE = (ushort)b3.s2 * (ushort)a1.s3;</div><div class="line"><a name="l01271"></a><span class="lineno"> 1271</span>  ushort tmpF = (ushort)b3.s3 * (ushort)a1.s3;</div><div class="line"><a name="l01272"></a><span class="lineno"> 1272</span> </div><div class="line"><a name="l01273"></a><span class="lineno"> 1273</span>  acc10 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);</div><div class="line"><a name="l01274"></a><span class="lineno"> 1274</span>  acc11 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);</div><div class="line"><a name="l01275"></a><span class="lineno"> 1275</span>  acc12 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);</div><div class="line"><a name="l01276"></a><span class="lineno"> 1276</span>  acc13 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);</div><div class="line"><a name="l01277"></a><span class="lineno"> 1277</span>  }</div><div class="line"><a name="l01278"></a><span class="lineno"> 1278</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01279"></a><span class="lineno"> 1279</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01280"></a><span class="lineno"> 1280</span>  {</div><div class="line"><a name="l01281"></a><span class="lineno"> 1281</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01282"></a><span class="lineno"> 1282</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a2.s0;</div><div class="line"><a name="l01283"></a><span class="lineno"> 1283</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a2.s0;</div><div class="line"><a name="l01284"></a><span class="lineno"> 1284</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a2.s0;</div><div class="line"><a name="l01285"></a><span class="lineno"> 1285</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a2.s0;</div><div class="line"><a name="l01286"></a><span class="lineno"> 1286</span> </div><div class="line"><a name="l01287"></a><span class="lineno"> 1287</span>  ushort tmp4 = (ushort)b1.s0 * (ushort)a2.s1;</div><div class="line"><a name="l01288"></a><span class="lineno"> 1288</span>  ushort tmp5 = (ushort)b1.s1 * (ushort)a2.s1;</div><div class="line"><a name="l01289"></a><span class="lineno"> 1289</span>  ushort tmp6 = (ushort)b1.s2 * (ushort)a2.s1;</div><div class="line"><a name="l01290"></a><span class="lineno"> 1290</span>  ushort tmp7 = (ushort)b1.s3 * (ushort)a2.s1;</div><div class="line"><a name="l01291"></a><span class="lineno"> 1291</span> </div><div class="line"><a name="l01292"></a><span class="lineno"> 1292</span>  ushort tmp8 = (ushort)b2.s0 * (ushort)a2.s2;</div><div class="line"><a name="l01293"></a><span class="lineno"> 1293</span>  ushort tmp9 = (ushort)b2.s1 * (ushort)a2.s2;</div><div class="line"><a name="l01294"></a><span class="lineno"> 1294</span>  ushort tmpA = (ushort)b2.s2 * (ushort)a2.s2;</div><div class="line"><a name="l01295"></a><span class="lineno"> 1295</span>  ushort tmpB = (ushort)b2.s3 * (ushort)a2.s2;</div><div class="line"><a name="l01296"></a><span class="lineno"> 1296</span> </div><div class="line"><a name="l01297"></a><span class="lineno"> 1297</span>  ushort tmpC = (ushort)b3.s0 * (ushort)a2.s3;</div><div class="line"><a name="l01298"></a><span class="lineno"> 1298</span>  ushort tmpD = (ushort)b3.s1 * (ushort)a2.s3;</div><div class="line"><a name="l01299"></a><span class="lineno"> 1299</span>  ushort tmpE = (ushort)b3.s2 * (ushort)a2.s3;</div><div class="line"><a name="l01300"></a><span class="lineno"> 1300</span>  ushort tmpF = (ushort)b3.s3 * (ushort)a2.s3;</div><div class="line"><a name="l01301"></a><span class="lineno"> 1301</span> </div><div class="line"><a name="l01302"></a><span class="lineno"> 1302</span>  acc20 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);</div><div class="line"><a name="l01303"></a><span class="lineno"> 1303</span>  acc21 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);</div><div class="line"><a name="l01304"></a><span class="lineno"> 1304</span>  acc22 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);</div><div class="line"><a name="l01305"></a><span class="lineno"> 1305</span>  acc23 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);</div><div class="line"><a name="l01306"></a><span class="lineno"> 1306</span>  }</div><div class="line"><a name="l01307"></a><span class="lineno"> 1307</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01308"></a><span class="lineno"> 1308</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01309"></a><span class="lineno"> 1309</span>  {</div><div class="line"><a name="l01310"></a><span class="lineno"> 1310</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01311"></a><span class="lineno"> 1311</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a3.s0;</div><div class="line"><a name="l01312"></a><span class="lineno"> 1312</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a3.s0;</div><div class="line"><a name="l01313"></a><span class="lineno"> 1313</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a3.s0;</div><div class="line"><a name="l01314"></a><span class="lineno"> 1314</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a3.s0;</div><div class="line"><a name="l01315"></a><span class="lineno"> 1315</span> </div><div class="line"><a name="l01316"></a><span class="lineno"> 1316</span>  ushort tmp4 = (ushort)b1.s0 * (ushort)a3.s1;</div><div class="line"><a name="l01317"></a><span class="lineno"> 1317</span>  ushort tmp5 = (ushort)b1.s1 * (ushort)a3.s1;</div><div class="line"><a name="l01318"></a><span class="lineno"> 1318</span>  ushort tmp6 = (ushort)b1.s2 * (ushort)a3.s1;</div><div class="line"><a name="l01319"></a><span class="lineno"> 1319</span>  ushort tmp7 = (ushort)b1.s3 * (ushort)a3.s1;</div><div class="line"><a name="l01320"></a><span class="lineno"> 1320</span> </div><div class="line"><a name="l01321"></a><span class="lineno"> 1321</span>  ushort tmp8 = (ushort)b2.s0 * (ushort)a3.s2;</div><div class="line"><a name="l01322"></a><span class="lineno"> 1322</span>  ushort tmp9 = (ushort)b2.s1 * (ushort)a3.s2;</div><div class="line"><a name="l01323"></a><span class="lineno"> 1323</span>  ushort tmpA = (ushort)b2.s2 * (ushort)a3.s2;</div><div class="line"><a name="l01324"></a><span class="lineno"> 1324</span>  ushort tmpB = (ushort)b2.s3 * (ushort)a3.s2;</div><div class="line"><a name="l01325"></a><span class="lineno"> 1325</span> </div><div class="line"><a name="l01326"></a><span class="lineno"> 1326</span>  ushort tmpC = (ushort)b3.s0 * (ushort)a3.s3;</div><div class="line"><a name="l01327"></a><span class="lineno"> 1327</span>  ushort tmpD = (ushort)b3.s1 * (ushort)a3.s3;</div><div class="line"><a name="l01328"></a><span class="lineno"> 1328</span>  ushort tmpE = (ushort)b3.s2 * (ushort)a3.s3;</div><div class="line"><a name="l01329"></a><span class="lineno"> 1329</span>  ushort tmpF = (ushort)b3.s3 * (ushort)a3.s3;</div><div class="line"><a name="l01330"></a><span class="lineno"> 1330</span> </div><div class="line"><a name="l01331"></a><span class="lineno"> 1331</span>  acc30 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);</div><div class="line"><a name="l01332"></a><span class="lineno"> 1332</span>  acc31 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);</div><div class="line"><a name="l01333"></a><span class="lineno"> 1333</span>  acc32 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);</div><div class="line"><a name="l01334"></a><span class="lineno"> 1334</span>  acc33 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);</div><div class="line"><a name="l01335"></a><span class="lineno"> 1335</span>  }</div><div class="line"><a name="l01336"></a><span class="lineno"> 1336</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01337"></a><span class="lineno"> 1337</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01338"></a><span class="lineno"> 1338</span>  {</div><div class="line"><a name="l01339"></a><span class="lineno"> 1339</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01340"></a><span class="lineno"> 1340</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a4.s0;</div><div class="line"><a name="l01341"></a><span class="lineno"> 1341</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a4.s0;</div><div class="line"><a name="l01342"></a><span class="lineno"> 1342</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a4.s0;</div><div class="line"><a name="l01343"></a><span class="lineno"> 1343</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a4.s0;</div><div class="line"><a name="l01344"></a><span class="lineno"> 1344</span> </div><div class="line"><a name="l01345"></a><span class="lineno"> 1345</span>  ushort tmp4 = (ushort)b1.s0 * (ushort)a4.s1;</div><div class="line"><a name="l01346"></a><span class="lineno"> 1346</span>  ushort tmp5 = (ushort)b1.s1 * (ushort)a4.s1;</div><div class="line"><a name="l01347"></a><span class="lineno"> 1347</span>  ushort tmp6 = (ushort)b1.s2 * (ushort)a4.s1;</div><div class="line"><a name="l01348"></a><span class="lineno"> 1348</span>  ushort tmp7 = (ushort)b1.s3 * (ushort)a4.s1;</div><div class="line"><a name="l01349"></a><span class="lineno"> 1349</span> </div><div class="line"><a name="l01350"></a><span class="lineno"> 1350</span>  ushort tmp8 = (ushort)b2.s0 * (ushort)a4.s2;</div><div class="line"><a name="l01351"></a><span class="lineno"> 1351</span>  ushort tmp9 = (ushort)b2.s1 * (ushort)a4.s2;</div><div class="line"><a name="l01352"></a><span class="lineno"> 1352</span>  ushort tmpA = (ushort)b2.s2 * (ushort)a4.s2;</div><div class="line"><a name="l01353"></a><span class="lineno"> 1353</span>  ushort tmpB = (ushort)b2.s3 * (ushort)a4.s2;</div><div class="line"><a name="l01354"></a><span class="lineno"> 1354</span> </div><div class="line"><a name="l01355"></a><span class="lineno"> 1355</span>  ushort tmpC = (ushort)b3.s0 * (ushort)a4.s3;</div><div class="line"><a name="l01356"></a><span class="lineno"> 1356</span>  ushort tmpD = (ushort)b3.s1 * (ushort)a4.s3;</div><div class="line"><a name="l01357"></a><span class="lineno"> 1357</span>  ushort tmpE = (ushort)b3.s2 * (ushort)a4.s3;</div><div class="line"><a name="l01358"></a><span class="lineno"> 1358</span>  ushort tmpF = (ushort)b3.s3 * (ushort)a4.s3;</div><div class="line"><a name="l01359"></a><span class="lineno"> 1359</span> </div><div class="line"><a name="l01360"></a><span class="lineno"> 1360</span>  acc40 += ((uint)tmp0 + (uint)tmp4 + (uint)tmp8 + (uint)tmpC);</div><div class="line"><a name="l01361"></a><span class="lineno"> 1361</span>  acc41 += ((uint)tmp1 + (uint)tmp5 + (uint)tmp9 + (uint)tmpD);</div><div class="line"><a name="l01362"></a><span class="lineno"> 1362</span>  acc42 += ((uint)tmp2 + (uint)tmp6 + (uint)tmpA + (uint)tmpE);</div><div class="line"><a name="l01363"></a><span class="lineno"> 1363</span>  acc43 += ((uint)tmp3 + (uint)tmp7 + (uint)tmpB + (uint)tmpF);</div><div class="line"><a name="l01364"></a><span class="lineno"> 1364</span>  }</div><div class="line"><a name="l01365"></a><span class="lineno"> 1365</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01366"></a><span class="lineno"> 1366</span>  }</div><div class="line"><a name="l01367"></a><span class="lineno"> 1367</span> </div><div class="line"><a name="l01368"></a><span class="lineno"> 1368</span>  <span class="keywordflow">for</span>(; src_addr.s0 < end_row_vec_a; src_addr += (int2)(1, src1_stride_y))</div><div class="line"><a name="l01369"></a><span class="lineno"> 1369</span>  {</div><div class="line"><a name="l01370"></a><span class="lineno"> 1370</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01371"></a><span class="lineno"> 1371</span>  uchar a0 = *(src0_ptr + src_addr.s0 + 0 * src0_stride_y);</div><div class="line"><a name="l01372"></a><span class="lineno"> 1372</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01373"></a><span class="lineno"> 1373</span>  uchar a1 = *(src0_ptr + src_addr.s0 + 1 * src0_stride_y);</div><div class="line"><a name="l01374"></a><span class="lineno"> 1374</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01375"></a><span class="lineno"> 1375</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01376"></a><span class="lineno"> 1376</span>  uchar a2 = *(src0_ptr + src_addr.s0 + 2 * src0_stride_y);</div><div class="line"><a name="l01377"></a><span class="lineno"> 1377</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01378"></a><span class="lineno"> 1378</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01379"></a><span class="lineno"> 1379</span>  uchar a3 = *(src0_ptr + src_addr.s0 + 3 * src0_stride_y);</div><div class="line"><a name="l01380"></a><span class="lineno"> 1380</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01381"></a><span class="lineno"> 1381</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01382"></a><span class="lineno"> 1382</span>  uchar a4 = *(src0_ptr + src_addr.s0 + 4 * src0_stride_y);</div><div class="line"><a name="l01383"></a><span class="lineno"> 1383</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01384"></a><span class="lineno"> 1384</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01385"></a><span class="lineno"> 1385</span>  uchar4 b0 = vload4(0, src1_ptr + src_addr.s1);</div><div class="line"><a name="l01386"></a><span class="lineno"> 1386</span> </div><div class="line"><a name="l01387"></a><span class="lineno"> 1387</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01388"></a><span class="lineno"> 1388</span>  {</div><div class="line"><a name="l01389"></a><span class="lineno"> 1389</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01390"></a><span class="lineno"> 1390</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a0;</div><div class="line"><a name="l01391"></a><span class="lineno"> 1391</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a0;</div><div class="line"><a name="l01392"></a><span class="lineno"> 1392</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a0;</div><div class="line"><a name="l01393"></a><span class="lineno"> 1393</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a0;</div><div class="line"><a name="l01394"></a><span class="lineno"> 1394</span> </div><div class="line"><a name="l01395"></a><span class="lineno"> 1395</span>  acc00 += ((uint)tmp0);</div><div class="line"><a name="l01396"></a><span class="lineno"> 1396</span>  acc01 += ((uint)tmp1);</div><div class="line"><a name="l01397"></a><span class="lineno"> 1397</span>  acc02 += ((uint)tmp2);</div><div class="line"><a name="l01398"></a><span class="lineno"> 1398</span>  acc03 += ((uint)tmp3);</div><div class="line"><a name="l01399"></a><span class="lineno"> 1399</span>  }</div><div class="line"><a name="l01400"></a><span class="lineno"> 1400</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01401"></a><span class="lineno"> 1401</span>  {</div><div class="line"><a name="l01402"></a><span class="lineno"> 1402</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01403"></a><span class="lineno"> 1403</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a1;</div><div class="line"><a name="l01404"></a><span class="lineno"> 1404</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a1;</div><div class="line"><a name="l01405"></a><span class="lineno"> 1405</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a1;</div><div class="line"><a name="l01406"></a><span class="lineno"> 1406</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a1;</div><div class="line"><a name="l01407"></a><span class="lineno"> 1407</span> </div><div class="line"><a name="l01408"></a><span class="lineno"> 1408</span>  acc10 += ((uint)tmp0);</div><div class="line"><a name="l01409"></a><span class="lineno"> 1409</span>  acc11 += ((uint)tmp1);</div><div class="line"><a name="l01410"></a><span class="lineno"> 1410</span>  acc12 += ((uint)tmp2);</div><div class="line"><a name="l01411"></a><span class="lineno"> 1411</span>  acc13 += ((uint)tmp3);</div><div class="line"><a name="l01412"></a><span class="lineno"> 1412</span>  }</div><div class="line"><a name="l01413"></a><span class="lineno"> 1413</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01414"></a><span class="lineno"> 1414</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01415"></a><span class="lineno"> 1415</span>  {</div><div class="line"><a name="l01416"></a><span class="lineno"> 1416</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01417"></a><span class="lineno"> 1417</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a2;</div><div class="line"><a name="l01418"></a><span class="lineno"> 1418</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a2;</div><div class="line"><a name="l01419"></a><span class="lineno"> 1419</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a2;</div><div class="line"><a name="l01420"></a><span class="lineno"> 1420</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a2;</div><div class="line"><a name="l01421"></a><span class="lineno"> 1421</span> </div><div class="line"><a name="l01422"></a><span class="lineno"> 1422</span>  acc20 += ((uint)tmp0);</div><div class="line"><a name="l01423"></a><span class="lineno"> 1423</span>  acc21 += ((uint)tmp1);</div><div class="line"><a name="l01424"></a><span class="lineno"> 1424</span>  acc22 += ((uint)tmp2);</div><div class="line"><a name="l01425"></a><span class="lineno"> 1425</span>  acc23 += ((uint)tmp3);</div><div class="line"><a name="l01426"></a><span class="lineno"> 1426</span>  }</div><div class="line"><a name="l01427"></a><span class="lineno"> 1427</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01428"></a><span class="lineno"> 1428</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01429"></a><span class="lineno"> 1429</span>  {</div><div class="line"><a name="l01430"></a><span class="lineno"> 1430</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01431"></a><span class="lineno"> 1431</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a3;</div><div class="line"><a name="l01432"></a><span class="lineno"> 1432</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a3;</div><div class="line"><a name="l01433"></a><span class="lineno"> 1433</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a3;</div><div class="line"><a name="l01434"></a><span class="lineno"> 1434</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a3;</div><div class="line"><a name="l01435"></a><span class="lineno"> 1435</span> </div><div class="line"><a name="l01436"></a><span class="lineno"> 1436</span>  acc30 += ((uint)tmp0);</div><div class="line"><a name="l01437"></a><span class="lineno"> 1437</span>  acc31 += ((uint)tmp1);</div><div class="line"><a name="l01438"></a><span class="lineno"> 1438</span>  acc32 += ((uint)tmp2);</div><div class="line"><a name="l01439"></a><span class="lineno"> 1439</span>  acc33 += ((uint)tmp3);</div><div class="line"><a name="l01440"></a><span class="lineno"> 1440</span>  }</div><div class="line"><a name="l01441"></a><span class="lineno"> 1441</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01442"></a><span class="lineno"> 1442</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01443"></a><span class="lineno"> 1443</span>  {</div><div class="line"><a name="l01444"></a><span class="lineno"> 1444</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l01445"></a><span class="lineno"> 1445</span>  ushort tmp0 = (ushort)b0.s0 * (ushort)a4;</div><div class="line"><a name="l01446"></a><span class="lineno"> 1446</span>  ushort tmp1 = (ushort)b0.s1 * (ushort)a4;</div><div class="line"><a name="l01447"></a><span class="lineno"> 1447</span>  ushort tmp2 = (ushort)b0.s2 * (ushort)a4;</div><div class="line"><a name="l01448"></a><span class="lineno"> 1448</span>  ushort tmp3 = (ushort)b0.s3 * (ushort)a4;</div><div class="line"><a name="l01449"></a><span class="lineno"> 1449</span> </div><div class="line"><a name="l01450"></a><span class="lineno"> 1450</span>  acc40 += ((uint)tmp0);</div><div class="line"><a name="l01451"></a><span class="lineno"> 1451</span>  acc41 += ((uint)tmp1);</div><div class="line"><a name="l01452"></a><span class="lineno"> 1452</span>  acc42 += ((uint)tmp2);</div><div class="line"><a name="l01453"></a><span class="lineno"> 1453</span>  acc43 += ((uint)tmp3);</div><div class="line"><a name="l01454"></a><span class="lineno"> 1454</span>  }</div><div class="line"><a name="l01455"></a><span class="lineno"> 1455</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01456"></a><span class="lineno"> 1456</span>  }</div><div class="line"><a name="l01457"></a><span class="lineno"> 1457</span> </div><div class="line"><a name="l01458"></a><span class="lineno"> 1458</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01459"></a><span class="lineno"> 1459</span> </div><div class="line"><a name="l01460"></a><span class="lineno"> 1460</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01461"></a><span class="lineno"> 1461</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01462"></a><span class="lineno"> 1462</span> </div><div class="line"><a name="l01463"></a><span class="lineno"> 1463</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01464"></a><span class="lineno"> 1464</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01465"></a><span class="lineno"> 1465</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01466"></a><span class="lineno"> 1466</span>  <span class="comment">//</span></div><div class="line"><a name="l01467"></a><span class="lineno"> 1467</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01468"></a><span class="lineno"> 1468</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l01469"></a><span class="lineno"> 1469</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01470"></a><span class="lineno"> 1470</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01471"></a><span class="lineno"> 1471</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01472"></a><span class="lineno"> 1472</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l01473"></a><span class="lineno"> 1473</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01474"></a><span class="lineno"> 1474</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01475"></a><span class="lineno"> 1475</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l01476"></a><span class="lineno"> 1476</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01477"></a><span class="lineno"> 1477</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01478"></a><span class="lineno"> 1478</span> </div><div class="line"><a name="l01479"></a><span class="lineno"> 1479</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01480"></a><span class="lineno"> 1480</span>  uint8 zout = ((uint8)(0, 1, 2, 3, 4, 5, 6, 7) + (uint8)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint8)HEIGHT_GEMM3D;</div><div class="line"><a name="l01481"></a><span class="lineno"> 1481</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01482"></a><span class="lineno"> 1482</span> </div><div class="line"><a name="l01483"></a><span class="lineno"> 1483</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01484"></a><span class="lineno"> 1484</span>  zout *= (dst_cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l01485"></a><span class="lineno"> 1485</span> </div><div class="line"><a name="l01486"></a><span class="lineno"> 1486</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01487"></a><span class="lineno"> 1487</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01488"></a><span class="lineno"> 1488</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01489"></a><span class="lineno"> 1489</span> </div><div class="line"><a name="l01490"></a><span class="lineno"> 1490</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01491"></a><span class="lineno"> 1491</span>  vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l01492"></a><span class="lineno"> 1492</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01493"></a><span class="lineno"> 1493</span>  vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l01494"></a><span class="lineno"> 1494</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01495"></a><span class="lineno"> 1495</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01496"></a><span class="lineno"> 1496</span>  vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l01497"></a><span class="lineno"> 1497</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01498"></a><span class="lineno"> 1498</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01499"></a><span class="lineno"> 1499</span>  vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l01500"></a><span class="lineno"> 1500</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01501"></a><span class="lineno"> 1501</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01502"></a><span class="lineno"> 1502</span>  vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 4 * dst_stride_y + zout.s4));</div><div class="line"><a name="l01503"></a><span class="lineno"> 1503</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01504"></a><span class="lineno"> 1504</span> </div><div class="line"><a name="l01505"></a><span class="lineno"> 1505</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01506"></a><span class="lineno"> 1506</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01507"></a><span class="lineno"> 1507</span>  dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> += z * dst_stride_z;</div><div class="line"><a name="l01508"></a><span class="lineno"> 1508</span> </div><div class="line"><a name="l01509"></a><span class="lineno"> 1509</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01510"></a><span class="lineno"> 1510</span>  vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l01511"></a><span class="lineno"> 1511</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01512"></a><span class="lineno"> 1512</span>  vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l01513"></a><span class="lineno"> 1513</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01514"></a><span class="lineno"> 1514</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01515"></a><span class="lineno"> 1515</span>  vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y));</div><div class="line"><a name="l01516"></a><span class="lineno"> 1516</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01517"></a><span class="lineno"> 1517</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01518"></a><span class="lineno"> 1518</span>  vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y));</div><div class="line"><a name="l01519"></a><span class="lineno"> 1519</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01520"></a><span class="lineno"> 1520</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01521"></a><span class="lineno"> 1521</span>  vstore4((int4)(acc40, acc41, acc42, acc43), 0, (__global <span class="keywordtype">int</span> *)(dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 4 * dst_stride_y));</div><div class="line"><a name="l01522"></a><span class="lineno"> 1522</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 4</span></div><div class="line"><a name="l01523"></a><span class="lineno"> 1523</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01524"></a><span class="lineno"> 1524</span> }</div><div class="line"><a name="l01525"></a><span class="lineno"> 1525</span> </div><div class="line"><a name="l01526"></a><span class="lineno"> 1526</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l01527"></a><span class="lineno"> 1527</span> </div><div class="line"><a name="l01562"></a><span class="lineno"> 1562</span> __kernel <span class="keywordtype">void</span> gemmlowp_mm_bifrost_dot8(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src0),</div><div class="line"><a name="l01563"></a><span class="lineno"> 1563</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(src1),</div><div class="line"><a name="l01564"></a><span class="lineno"> 1564</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst),</div><div class="line"><a name="l01565"></a><span class="lineno"> 1565</span>  uint src0_stride_z,</div><div class="line"><a name="l01566"></a><span class="lineno"> 1566</span>  uint src1_stride_z,</div><div class="line"><a name="l01567"></a><span class="lineno"> 1567</span>  uint dst_stride_z</div><div class="line"><a name="l01568"></a><span class="lineno"> 1568</span> #<span class="keywordflow">if</span> defined(REINTERPRET_INPUT_AS_3D)</div><div class="line"><a name="l01569"></a><span class="lineno"> 1569</span>  ,</div><div class="line"><a name="l01570"></a><span class="lineno"> 1570</span>  uint src_cross_plane_pad</div><div class="line"><a name="l01571"></a><span class="lineno"> 1571</span> #endif <span class="comment">// REINTERPRET_INPUT_AS_3D</span></div><div class="line"><a name="l01572"></a><span class="lineno"> 1572</span> #<span class="keywordflow">if</span> defined(REINTERPRET_OUTPUT_AS_3D)</div><div class="line"><a name="l01573"></a><span class="lineno"> 1573</span>  ,</div><div class="line"><a name="l01574"></a><span class="lineno"> 1574</span>  uint dst_cross_plane_pad</div><div class="line"><a name="l01575"></a><span class="lineno"> 1575</span> #endif <span class="comment">// REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01576"></a><span class="lineno"> 1576</span>  )</div><div class="line"><a name="l01577"></a><span class="lineno"> 1577</span> {</div><div class="line"><a name="l01578"></a><span class="lineno"> 1578</span>  <span class="keywordtype">int</span> idx = get_global_id(0) * NUM_ELEMS_PROCESSED_PER_THREAD_X;</div><div class="line"><a name="l01579"></a><span class="lineno"> 1579</span> </div><div class="line"><a name="l01580"></a><span class="lineno"> 1580</span>  <span class="comment">// Compute starting address for matrix A and Matrix B</span></div><div class="line"><a name="l01581"></a><span class="lineno"> 1581</span>  int2 src_addr = ((int2)(src0_offset_first_element_in_bytes, src1_offset_first_element_in_bytes));</div><div class="line"><a name="l01582"></a><span class="lineno"> 1582</span> </div><div class="line"><a name="l01583"></a><span class="lineno"> 1583</span>  <span class="comment">// Update address for the matrix A</span></div><div class="line"><a name="l01584"></a><span class="lineno"> 1584</span>  src_addr.s0 += get_global_id(1) * src0_stride_y * NUM_ELEMS_PROCESSED_PER_THREAD_Y;</div><div class="line"><a name="l01585"></a><span class="lineno"> 1585</span> </div><div class="line"><a name="l01586"></a><span class="lineno"> 1586</span>  <span class="comment">// Update address for the matrix B</span></div><div class="line"><a name="l01587"></a><span class="lineno"> 1587</span>  src_addr.s1 += idx;</div><div class="line"><a name="l01588"></a><span class="lineno"> 1588</span> </div><div class="line"><a name="l01589"></a><span class="lineno"> 1589</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01590"></a><span class="lineno"> 1590</span>  <span class="comment">// Since we load a 2D input tile from a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01591"></a><span class="lineno"> 1591</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01592"></a><span class="lineno"> 1592</span>  <span class="comment">//</span></div><div class="line"><a name="l01593"></a><span class="lineno"> 1593</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01594"></a><span class="lineno"> 1594</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l01595"></a><span class="lineno"> 1595</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01596"></a><span class="lineno"> 1596</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01597"></a><span class="lineno"> 1597</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01598"></a><span class="lineno"> 1598</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l01599"></a><span class="lineno"> 1599</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01600"></a><span class="lineno"> 1600</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01601"></a><span class="lineno"> 1601</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l01602"></a><span class="lineno"> 1602</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01603"></a><span class="lineno"> 1603</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01604"></a><span class="lineno"> 1604</span> </div><div class="line"><a name="l01605"></a><span class="lineno"> 1605</span>  <span class="comment">// The plane (zin) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01606"></a><span class="lineno"> 1606</span>  uint4 zin = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01607"></a><span class="lineno"> 1607</span>  zin = min(DEPTH_GEMM3D - 1, zin);</div><div class="line"><a name="l01608"></a><span class="lineno"> 1608</span> </div><div class="line"><a name="l01609"></a><span class="lineno"> 1609</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01610"></a><span class="lineno"> 1610</span>  zin *= (src_cross_plane_pad * src0_stride_y);</div><div class="line"><a name="l01611"></a><span class="lineno"> 1611</span> </div><div class="line"><a name="l01612"></a><span class="lineno"> 1612</span>  zin += ((uint4)(0, 1, 2, 3)) * src0_stride_y;</div><div class="line"><a name="l01613"></a><span class="lineno"> 1613</span> </div><div class="line"><a name="l01614"></a><span class="lineno"> 1614</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01615"></a><span class="lineno"> 1615</span>  <span class="comment">// multiply src0_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01616"></a><span class="lineno"> 1616</span>  src_addr.s0 += get_global_id(2) * src0_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01617"></a><span class="lineno"> 1617</span> </div><div class="line"><a name="l01618"></a><span class="lineno"> 1618</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01619"></a><span class="lineno"> 1619</span> </div><div class="line"><a name="l01620"></a><span class="lineno"> 1620</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01621"></a><span class="lineno"> 1621</span>  src_addr.s0 += get_global_id(2) * src0_stride_z;</div><div class="line"><a name="l01622"></a><span class="lineno"> 1622</span> </div><div class="line"><a name="l01623"></a><span class="lineno"> 1623</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01624"></a><span class="lineno"> 1624</span> </div><div class="line"><a name="l01625"></a><span class="lineno"> 1625</span> <span class="preprocessor">#if defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01626"></a><span class="lineno"> 1626</span>  <span class="comment">// Do not slide matrix B if the matrix B has 3 dimensions and matrix A more than 3</span></div><div class="line"><a name="l01627"></a><span class="lineno"> 1627</span>  src_addr.s1 += (get_global_id(2) % MATRIX_B_DEPTH) * src1_stride_z;</div><div class="line"><a name="l01628"></a><span class="lineno"> 1628</span> <span class="preprocessor">#else // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01629"></a><span class="lineno"> 1629</span>  src_addr.s1 += get_global_id(2) * src1_stride_z;</div><div class="line"><a name="l01630"></a><span class="lineno"> 1630</span> <span class="preprocessor">#endif // defined(MATRIX_B_DEPTH)</span></div><div class="line"><a name="l01631"></a><span class="lineno"> 1631</span> </div><div class="line"><a name="l01632"></a><span class="lineno"> 1632</span>  uint acc00 = 0;</div><div class="line"><a name="l01633"></a><span class="lineno"> 1633</span>  uint acc01 = 0;</div><div class="line"><a name="l01634"></a><span class="lineno"> 1634</span>  uint acc02 = 0;</div><div class="line"><a name="l01635"></a><span class="lineno"> 1635</span>  uint acc03 = 0;</div><div class="line"><a name="l01636"></a><span class="lineno"> 1636</span>  uint acc04 = 0;</div><div class="line"><a name="l01637"></a><span class="lineno"> 1637</span>  uint acc05 = 0;</div><div class="line"><a name="l01638"></a><span class="lineno"> 1638</span>  uint acc06 = 0;</div><div class="line"><a name="l01639"></a><span class="lineno"> 1639</span>  uint acc07 = 0;</div><div class="line"><a name="l01640"></a><span class="lineno"> 1640</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01641"></a><span class="lineno"> 1641</span>  uint acc10 = 0;</div><div class="line"><a name="l01642"></a><span class="lineno"> 1642</span>  uint acc11 = 0;</div><div class="line"><a name="l01643"></a><span class="lineno"> 1643</span>  uint acc12 = 0;</div><div class="line"><a name="l01644"></a><span class="lineno"> 1644</span>  uint acc13 = 0;</div><div class="line"><a name="l01645"></a><span class="lineno"> 1645</span>  uint acc14 = 0;</div><div class="line"><a name="l01646"></a><span class="lineno"> 1646</span>  uint acc15 = 0;</div><div class="line"><a name="l01647"></a><span class="lineno"> 1647</span>  uint acc16 = 0;</div><div class="line"><a name="l01648"></a><span class="lineno"> 1648</span>  uint acc17 = 0;</div><div class="line"><a name="l01649"></a><span class="lineno"> 1649</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01650"></a><span class="lineno"> 1650</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01651"></a><span class="lineno"> 1651</span>  uint acc20 = 0;</div><div class="line"><a name="l01652"></a><span class="lineno"> 1652</span>  uint acc21 = 0;</div><div class="line"><a name="l01653"></a><span class="lineno"> 1653</span>  uint acc22 = 0;</div><div class="line"><a name="l01654"></a><span class="lineno"> 1654</span>  uint acc23 = 0;</div><div class="line"><a name="l01655"></a><span class="lineno"> 1655</span>  uint acc24 = 0;</div><div class="line"><a name="l01656"></a><span class="lineno"> 1656</span>  uint acc25 = 0;</div><div class="line"><a name="l01657"></a><span class="lineno"> 1657</span>  uint acc26 = 0;</div><div class="line"><a name="l01658"></a><span class="lineno"> 1658</span>  uint acc27 = 0;</div><div class="line"><a name="l01659"></a><span class="lineno"> 1659</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01660"></a><span class="lineno"> 1660</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01661"></a><span class="lineno"> 1661</span>  uint acc30 = 0;</div><div class="line"><a name="l01662"></a><span class="lineno"> 1662</span>  uint acc31 = 0;</div><div class="line"><a name="l01663"></a><span class="lineno"> 1663</span>  uint acc32 = 0;</div><div class="line"><a name="l01664"></a><span class="lineno"> 1664</span>  uint acc33 = 0;</div><div class="line"><a name="l01665"></a><span class="lineno"> 1665</span>  uint acc34 = 0;</div><div class="line"><a name="l01666"></a><span class="lineno"> 1666</span>  uint acc35 = 0;</div><div class="line"><a name="l01667"></a><span class="lineno"> 1667</span>  uint acc36 = 0;</div><div class="line"><a name="l01668"></a><span class="lineno"> 1668</span>  uint acc37 = 0;</div><div class="line"><a name="l01669"></a><span class="lineno"> 1669</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01670"></a><span class="lineno"> 1670</span> </div><div class="line"><a name="l01671"></a><span class="lineno"> 1671</span>  <span class="comment">// A and B src indices get incremented at the same time.</span></div><div class="line"><a name="l01672"></a><span class="lineno"> 1672</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01673"></a><span class="lineno"> 1673</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 8); i += 8)</div><div class="line"><a name="l01674"></a><span class="lineno"> 1674</span>  {</div><div class="line"><a name="l01675"></a><span class="lineno"> 1675</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01676"></a><span class="lineno"> 1676</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01677"></a><span class="lineno"> 1677</span>  uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s0));</div><div class="line"><a name="l01678"></a><span class="lineno"> 1678</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01679"></a><span class="lineno"> 1679</span>  uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s1));</div><div class="line"><a name="l01680"></a><span class="lineno"> 1680</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01681"></a><span class="lineno"> 1681</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01682"></a><span class="lineno"> 1682</span>  uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s2));</div><div class="line"><a name="l01683"></a><span class="lineno"> 1683</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01684"></a><span class="lineno"> 1684</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01685"></a><span class="lineno"> 1685</span>  uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + zin.s3));</div><div class="line"><a name="l01686"></a><span class="lineno"> 1686</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01687"></a><span class="lineno"> 1687</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01688"></a><span class="lineno"> 1688</span>  <span class="comment">// Load values from matrix A and matrix B</span></div><div class="line"><a name="l01689"></a><span class="lineno"> 1689</span>  uchar8 a0 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01690"></a><span class="lineno"> 1690</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01691"></a><span class="lineno"> 1691</span>  uchar8 a1 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01692"></a><span class="lineno"> 1692</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01693"></a><span class="lineno"> 1693</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01694"></a><span class="lineno"> 1694</span>  uchar8 a2 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01695"></a><span class="lineno"> 1695</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01696"></a><span class="lineno"> 1696</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01697"></a><span class="lineno"> 1697</span>  uchar8 a3 = vload8(0, (__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01698"></a><span class="lineno"> 1698</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01699"></a><span class="lineno"> 1699</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01700"></a><span class="lineno"> 1700</span> </div><div class="line"><a name="l01701"></a><span class="lineno"> 1701</span>  uchar8 b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);</div><div class="line"><a name="l01702"></a><span class="lineno"> 1702</span>  uchar8 b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);</div><div class="line"><a name="l01703"></a><span class="lineno"> 1703</span>  uchar8 b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);</div><div class="line"><a name="l01704"></a><span class="lineno"> 1704</span>  uchar8 b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);</div><div class="line"><a name="l01705"></a><span class="lineno"> 1705</span>  src_addr.s1 += 4 * src1_stride_y;</div><div class="line"><a name="l01706"></a><span class="lineno"> 1706</span> </div><div class="line"><a name="l01707"></a><span class="lineno"> 1707</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00);</div><div class="line"><a name="l01708"></a><span class="lineno"> 1708</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01);</div><div class="line"><a name="l01709"></a><span class="lineno"> 1709</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02);</div><div class="line"><a name="l01710"></a><span class="lineno"> 1710</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03);</div><div class="line"><a name="l01711"></a><span class="lineno"> 1711</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04);</div><div class="line"><a name="l01712"></a><span class="lineno"> 1712</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05);</div><div class="line"><a name="l01713"></a><span class="lineno"> 1713</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06);</div><div class="line"><a name="l01714"></a><span class="lineno"> 1714</span>  ARM_DOT(a0.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07);</div><div class="line"><a name="l01715"></a><span class="lineno"> 1715</span> </div><div class="line"><a name="l01716"></a><span class="lineno"> 1716</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01717"></a><span class="lineno"> 1717</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10);</div><div class="line"><a name="l01718"></a><span class="lineno"> 1718</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11);</div><div class="line"><a name="l01719"></a><span class="lineno"> 1719</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12);</div><div class="line"><a name="l01720"></a><span class="lineno"> 1720</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13);</div><div class="line"><a name="l01721"></a><span class="lineno"> 1721</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14);</div><div class="line"><a name="l01722"></a><span class="lineno"> 1722</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15);</div><div class="line"><a name="l01723"></a><span class="lineno"> 1723</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16);</div><div class="line"><a name="l01724"></a><span class="lineno"> 1724</span>  ARM_DOT(a1.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17);</div><div class="line"><a name="l01725"></a><span class="lineno"> 1725</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01726"></a><span class="lineno"> 1726</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01727"></a><span class="lineno"> 1727</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20);</div><div class="line"><a name="l01728"></a><span class="lineno"> 1728</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21);</div><div class="line"><a name="l01729"></a><span class="lineno"> 1729</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22);</div><div class="line"><a name="l01730"></a><span class="lineno"> 1730</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23);</div><div class="line"><a name="l01731"></a><span class="lineno"> 1731</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24);</div><div class="line"><a name="l01732"></a><span class="lineno"> 1732</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25);</div><div class="line"><a name="l01733"></a><span class="lineno"> 1733</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26);</div><div class="line"><a name="l01734"></a><span class="lineno"> 1734</span>  ARM_DOT(a2.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27);</div><div class="line"><a name="l01735"></a><span class="lineno"> 1735</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01736"></a><span class="lineno"> 1736</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01737"></a><span class="lineno"> 1737</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30);</div><div class="line"><a name="l01738"></a><span class="lineno"> 1738</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31);</div><div class="line"><a name="l01739"></a><span class="lineno"> 1739</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32);</div><div class="line"><a name="l01740"></a><span class="lineno"> 1740</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33);</div><div class="line"><a name="l01741"></a><span class="lineno"> 1741</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34);</div><div class="line"><a name="l01742"></a><span class="lineno"> 1742</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35);</div><div class="line"><a name="l01743"></a><span class="lineno"> 1743</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36);</div><div class="line"><a name="l01744"></a><span class="lineno"> 1744</span>  ARM_DOT(a3.s0123, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37);</div><div class="line"><a name="l01745"></a><span class="lineno"> 1745</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01746"></a><span class="lineno"> 1746</span> </div><div class="line"><a name="l01747"></a><span class="lineno"> 1747</span>  b0 = vload8(0, src1_ptr + src_addr.s1 + 0 * src1_stride_y);</div><div class="line"><a name="l01748"></a><span class="lineno"> 1748</span>  b1 = vload8(0, src1_ptr + src_addr.s1 + 1 * src1_stride_y);</div><div class="line"><a name="l01749"></a><span class="lineno"> 1749</span>  b2 = vload8(0, src1_ptr + src_addr.s1 + 2 * src1_stride_y);</div><div class="line"><a name="l01750"></a><span class="lineno"> 1750</span>  b3 = vload8(0, src1_ptr + src_addr.s1 + 3 * src1_stride_y);</div><div class="line"><a name="l01751"></a><span class="lineno"> 1751</span>  src_addr.s1 += 4 * src1_stride_y;</div><div class="line"><a name="l01752"></a><span class="lineno"> 1752</span> </div><div class="line"><a name="l01753"></a><span class="lineno"> 1753</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc00);</div><div class="line"><a name="l01754"></a><span class="lineno"> 1754</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc01);</div><div class="line"><a name="l01755"></a><span class="lineno"> 1755</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc02);</div><div class="line"><a name="l01756"></a><span class="lineno"> 1756</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc03);</div><div class="line"><a name="l01757"></a><span class="lineno"> 1757</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc04);</div><div class="line"><a name="l01758"></a><span class="lineno"> 1758</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc05);</div><div class="line"><a name="l01759"></a><span class="lineno"> 1759</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc06);</div><div class="line"><a name="l01760"></a><span class="lineno"> 1760</span>  ARM_DOT(a0.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc07);</div><div class="line"><a name="l01761"></a><span class="lineno"> 1761</span> </div><div class="line"><a name="l01762"></a><span class="lineno"> 1762</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01763"></a><span class="lineno"> 1763</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc10);</div><div class="line"><a name="l01764"></a><span class="lineno"> 1764</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc11);</div><div class="line"><a name="l01765"></a><span class="lineno"> 1765</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc12);</div><div class="line"><a name="l01766"></a><span class="lineno"> 1766</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc13);</div><div class="line"><a name="l01767"></a><span class="lineno"> 1767</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc14);</div><div class="line"><a name="l01768"></a><span class="lineno"> 1768</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc15);</div><div class="line"><a name="l01769"></a><span class="lineno"> 1769</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc16);</div><div class="line"><a name="l01770"></a><span class="lineno"> 1770</span>  ARM_DOT(a1.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc17);</div><div class="line"><a name="l01771"></a><span class="lineno"> 1771</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01772"></a><span class="lineno"> 1772</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01773"></a><span class="lineno"> 1773</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc20);</div><div class="line"><a name="l01774"></a><span class="lineno"> 1774</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc21);</div><div class="line"><a name="l01775"></a><span class="lineno"> 1775</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc22);</div><div class="line"><a name="l01776"></a><span class="lineno"> 1776</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc23);</div><div class="line"><a name="l01777"></a><span class="lineno"> 1777</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc24);</div><div class="line"><a name="l01778"></a><span class="lineno"> 1778</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc25);</div><div class="line"><a name="l01779"></a><span class="lineno"> 1779</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc26);</div><div class="line"><a name="l01780"></a><span class="lineno"> 1780</span>  ARM_DOT(a2.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc27);</div><div class="line"><a name="l01781"></a><span class="lineno"> 1781</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01782"></a><span class="lineno"> 1782</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01783"></a><span class="lineno"> 1783</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s0, b1.s0, b2.s0, b3.s0), acc30);</div><div class="line"><a name="l01784"></a><span class="lineno"> 1784</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s1, b1.s1, b2.s1, b3.s1), acc31);</div><div class="line"><a name="l01785"></a><span class="lineno"> 1785</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s2, b1.s2, b2.s2, b3.s2), acc32);</div><div class="line"><a name="l01786"></a><span class="lineno"> 1786</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s3, b1.s3, b2.s3, b3.s3), acc33);</div><div class="line"><a name="l01787"></a><span class="lineno"> 1787</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s4, b1.s4, b2.s4, b3.s4), acc34);</div><div class="line"><a name="l01788"></a><span class="lineno"> 1788</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s5, b1.s5, b2.s5, b3.s5), acc35);</div><div class="line"><a name="l01789"></a><span class="lineno"> 1789</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s6, b1.s6, b2.s6, b3.s6), acc36);</div><div class="line"><a name="l01790"></a><span class="lineno"> 1790</span>  ARM_DOT(a3.s4567, (uchar4)(b0.s7, b1.s7, b2.s7, b3.s7), acc37);</div><div class="line"><a name="l01791"></a><span class="lineno"> 1791</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01792"></a><span class="lineno"> 1792</span> </div><div class="line"><a name="l01793"></a><span class="lineno"> 1793</span>  src_addr.s0 += 8;</div><div class="line"><a name="l01794"></a><span class="lineno"> 1794</span>  }</div><div class="line"><a name="l01795"></a><span class="lineno"> 1795</span> </div><div class="line"><a name="l01796"></a><span class="lineno"> 1796</span>  <span class="keywordflow">for</span>(; i < (int)COLS_A; ++i)</div><div class="line"><a name="l01797"></a><span class="lineno"> 1797</span>  {</div><div class="line"><a name="l01798"></a><span class="lineno"> 1798</span> <span class="preprocessor">#if defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01799"></a><span class="lineno"> 1799</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01800"></a><span class="lineno"> 1800</span>  uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s0));</div><div class="line"><a name="l01801"></a><span class="lineno"> 1801</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01802"></a><span class="lineno"> 1802</span>  uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s1));</div><div class="line"><a name="l01803"></a><span class="lineno"> 1803</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01804"></a><span class="lineno"> 1804</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01805"></a><span class="lineno"> 1805</span>  uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s2));</div><div class="line"><a name="l01806"></a><span class="lineno"> 1806</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01807"></a><span class="lineno"> 1807</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01808"></a><span class="lineno"> 1808</span>  uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + zin.s3));</div><div class="line"><a name="l01809"></a><span class="lineno"> 1809</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01810"></a><span class="lineno"> 1810</span> <span class="preprocessor">#else // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01811"></a><span class="lineno"> 1811</span>  <span class="comment">// Load values from matrix A</span></div><div class="line"><a name="l01812"></a><span class="lineno"> 1812</span>  uchar a0 = *((__global uchar *)(src0_ptr + src_addr.s0 + 0 * src0_stride_y));</div><div class="line"><a name="l01813"></a><span class="lineno"> 1813</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01814"></a><span class="lineno"> 1814</span>  uchar a1 = *((__global uchar *)(src0_ptr + src_addr.s0 + 1 * src0_stride_y));</div><div class="line"><a name="l01815"></a><span class="lineno"> 1815</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01816"></a><span class="lineno"> 1816</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01817"></a><span class="lineno"> 1817</span>  uchar a2 = *((__global uchar *)(src0_ptr + src_addr.s0 + 2 * src0_stride_y));</div><div class="line"><a name="l01818"></a><span class="lineno"> 1818</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01819"></a><span class="lineno"> 1819</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01820"></a><span class="lineno"> 1820</span>  uchar a3 = *((__global uchar *)(src0_ptr + src_addr.s0 + 3 * src0_stride_y));</div><div class="line"><a name="l01821"></a><span class="lineno"> 1821</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01822"></a><span class="lineno"> 1822</span> <span class="preprocessor">#endif // defined(REINTERPRET_INPUT_AS_3D)</span></div><div class="line"><a name="l01823"></a><span class="lineno"> 1823</span> </div><div class="line"><a name="l01824"></a><span class="lineno"> 1824</span>  <span class="comment">// Load values from matrix B</span></div><div class="line"><a name="l01825"></a><span class="lineno"> 1825</span>  uchar8 b0 = vload8(0, src1_ptr + src_addr.s1);</div><div class="line"><a name="l01826"></a><span class="lineno"> 1826</span>  src_addr.s1 += src1_stride_y;</div><div class="line"><a name="l01827"></a><span class="lineno"> 1827</span> </div><div class="line"><a name="l01828"></a><span class="lineno"> 1828</span>  acc00 += (uint)a0 * b0.s0;</div><div class="line"><a name="l01829"></a><span class="lineno"> 1829</span>  acc01 += (uint)a0 * b0.s1;</div><div class="line"><a name="l01830"></a><span class="lineno"> 1830</span>  acc02 += (uint)a0 * b0.s2;</div><div class="line"><a name="l01831"></a><span class="lineno"> 1831</span>  acc03 += (uint)a0 * b0.s3;</div><div class="line"><a name="l01832"></a><span class="lineno"> 1832</span>  acc04 += (uint)a0 * b0.s4;</div><div class="line"><a name="l01833"></a><span class="lineno"> 1833</span>  acc05 += (uint)a0 * b0.s5;</div><div class="line"><a name="l01834"></a><span class="lineno"> 1834</span>  acc06 += (uint)a0 * b0.s6;</div><div class="line"><a name="l01835"></a><span class="lineno"> 1835</span>  acc07 += (uint)a0 * b0.s7;</div><div class="line"><a name="l01836"></a><span class="lineno"> 1836</span> </div><div class="line"><a name="l01837"></a><span class="lineno"> 1837</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01838"></a><span class="lineno"> 1838</span>  acc10 += (uint)a1 * b0.s0;</div><div class="line"><a name="l01839"></a><span class="lineno"> 1839</span>  acc11 += (uint)a1 * b0.s1;</div><div class="line"><a name="l01840"></a><span class="lineno"> 1840</span>  acc12 += (uint)a1 * b0.s2;</div><div class="line"><a name="l01841"></a><span class="lineno"> 1841</span>  acc13 += (uint)a1 * b0.s3;</div><div class="line"><a name="l01842"></a><span class="lineno"> 1842</span>  acc14 += (uint)a1 * b0.s4;</div><div class="line"><a name="l01843"></a><span class="lineno"> 1843</span>  acc15 += (uint)a1 * b0.s5;</div><div class="line"><a name="l01844"></a><span class="lineno"> 1844</span>  acc16 += (uint)a1 * b0.s6;</div><div class="line"><a name="l01845"></a><span class="lineno"> 1845</span>  acc17 += (uint)a1 * b0.s7;</div><div class="line"><a name="l01846"></a><span class="lineno"> 1846</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01847"></a><span class="lineno"> 1847</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01848"></a><span class="lineno"> 1848</span>  acc20 += (uint)a2 * b0.s0;</div><div class="line"><a name="l01849"></a><span class="lineno"> 1849</span>  acc21 += (uint)a2 * b0.s1;</div><div class="line"><a name="l01850"></a><span class="lineno"> 1850</span>  acc22 += (uint)a2 * b0.s2;</div><div class="line"><a name="l01851"></a><span class="lineno"> 1851</span>  acc23 += (uint)a2 * b0.s3;</div><div class="line"><a name="l01852"></a><span class="lineno"> 1852</span>  acc24 += (uint)a2 * b0.s4;</div><div class="line"><a name="l01853"></a><span class="lineno"> 1853</span>  acc25 += (uint)a2 * b0.s5;</div><div class="line"><a name="l01854"></a><span class="lineno"> 1854</span>  acc26 += (uint)a2 * b0.s6;</div><div class="line"><a name="l01855"></a><span class="lineno"> 1855</span>  acc27 += (uint)a2 * b0.s7;</div><div class="line"><a name="l01856"></a><span class="lineno"> 1856</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01857"></a><span class="lineno"> 1857</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01858"></a><span class="lineno"> 1858</span>  acc30 += (uint)a3 * b0.s0;</div><div class="line"><a name="l01859"></a><span class="lineno"> 1859</span>  acc31 += (uint)a3 * b0.s1;</div><div class="line"><a name="l01860"></a><span class="lineno"> 1860</span>  acc32 += (uint)a3 * b0.s2;</div><div class="line"><a name="l01861"></a><span class="lineno"> 1861</span>  acc33 += (uint)a3 * b0.s3;</div><div class="line"><a name="l01862"></a><span class="lineno"> 1862</span>  acc34 += (uint)a3 * b0.s4;</div><div class="line"><a name="l01863"></a><span class="lineno"> 1863</span>  acc35 += (uint)a3 * b0.s5;</div><div class="line"><a name="l01864"></a><span class="lineno"> 1864</span>  acc36 += (uint)a3 * b0.s6;</div><div class="line"><a name="l01865"></a><span class="lineno"> 1865</span>  acc37 += (uint)a3 * b0.s7;</div><div class="line"><a name="l01866"></a><span class="lineno"> 1866</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01867"></a><span class="lineno"> 1867</span> </div><div class="line"><a name="l01868"></a><span class="lineno"> 1868</span>  src_addr.s0 += 1;</div><div class="line"><a name="l01869"></a><span class="lineno"> 1869</span>  }</div><div class="line"><a name="l01870"></a><span class="lineno"> 1870</span> </div><div class="line"><a name="l01871"></a><span class="lineno"> 1871</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l01872"></a><span class="lineno"> 1872</span> </div><div class="line"><a name="l01873"></a><span class="lineno"> 1873</span>  <span class="comment">// Compute destination address</span></div><div class="line"><a name="l01874"></a><span class="lineno"> 1874</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01875"></a><span class="lineno"> 1875</span> </div><div class="line"><a name="l01876"></a><span class="lineno"> 1876</span>  <span class="comment">// Compute dst address</span></div><div class="line"><a name="l01877"></a><span class="lineno"> 1877</span>  __global uchar *dst_addr = dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>;</div><div class="line"><a name="l01878"></a><span class="lineno"> 1878</span> </div><div class="line"><a name="l01879"></a><span class="lineno"> 1879</span> <span class="preprocessor">#if defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01880"></a><span class="lineno"> 1880</span>  <span class="comment">// Since we store a 2D output tile in a 3D tensor, we need to check when the plane changes across the z dimension</span></div><div class="line"><a name="l01881"></a><span class="lineno"> 1881</span>  <span class="comment">// in order to take into account the presence of possible cross plane paddings</span></div><div class="line"><a name="l01882"></a><span class="lineno"> 1882</span>  <span class="comment">//</span></div><div class="line"><a name="l01883"></a><span class="lineno"> 1883</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01884"></a><span class="lineno"> 1884</span>  <span class="comment">// | plane0 |</span></div><div class="line"><a name="l01885"></a><span class="lineno"> 1885</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01886"></a><span class="lineno"> 1886</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01887"></a><span class="lineno"> 1887</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01888"></a><span class="lineno"> 1888</span>  <span class="comment">// | cross_plane_pad |</span></div><div class="line"><a name="l01889"></a><span class="lineno"> 1889</span>  <span class="comment">// |******************|</span></div><div class="line"><a name="l01890"></a><span class="lineno"> 1890</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01891"></a><span class="lineno"> 1891</span>  <span class="comment">// | plane1 |</span></div><div class="line"><a name="l01892"></a><span class="lineno"> 1892</span>  <span class="comment">// | |</span></div><div class="line"><a name="l01893"></a><span class="lineno"> 1893</span>  <span class="comment">// |__________________|</span></div><div class="line"><a name="l01894"></a><span class="lineno"> 1894</span> </div><div class="line"><a name="l01895"></a><span class="lineno"> 1895</span>  <span class="comment">// The plane (zout) is calculated dividing M (get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y) by HEIGHT_GEMM3D</span></div><div class="line"><a name="l01896"></a><span class="lineno"> 1896</span>  uint4 zout = ((uint4)(0, 1, 2, 3) + (uint4)(get_global_id(1) * NUM_ELEMS_PROCESSED_PER_THREAD_Y)) / (uint4)HEIGHT_GEMM3D;</div><div class="line"><a name="l01897"></a><span class="lineno"> 1897</span>  zout = min(DEPTH_GEMM3D - 1, zout);</div><div class="line"><a name="l01898"></a><span class="lineno"> 1898</span> </div><div class="line"><a name="l01899"></a><span class="lineno"> 1899</span>  <span class="comment">// Add offset due to the cross plane paddings</span></div><div class="line"><a name="l01900"></a><span class="lineno"> 1900</span>  zout *= (dst_cross_plane_pad * dst_stride_y);</div><div class="line"><a name="l01901"></a><span class="lineno"> 1901</span> </div><div class="line"><a name="l01902"></a><span class="lineno"> 1902</span>  <span class="comment">// Add offset for batched GEMM. The batches will be in the fourth dimension and for this reason we</span></div><div class="line"><a name="l01903"></a><span class="lineno"> 1903</span>  <span class="comment">// multiply dst_stride_z by DEPTH_GEMM3D</span></div><div class="line"><a name="l01904"></a><span class="lineno"> 1904</span>  dst_addr += z * dst_stride_z * DEPTH_GEMM3D;</div><div class="line"><a name="l01905"></a><span class="lineno"> 1905</span> </div><div class="line"><a name="l01906"></a><span class="lineno"> 1906</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01907"></a><span class="lineno"> 1907</span>  vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l01908"></a><span class="lineno"> 1908</span>  vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 0 * dst_stride_y + zout.s0));</div><div class="line"><a name="l01909"></a><span class="lineno"> 1909</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01910"></a><span class="lineno"> 1910</span>  vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l01911"></a><span class="lineno"> 1911</span>  vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 1 * dst_stride_y + zout.s1));</div><div class="line"><a name="l01912"></a><span class="lineno"> 1912</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01913"></a><span class="lineno"> 1913</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01914"></a><span class="lineno"> 1914</span>  vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l01915"></a><span class="lineno"> 1915</span>  vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 2 * dst_stride_y + zout.s2));</div><div class="line"><a name="l01916"></a><span class="lineno"> 1916</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01917"></a><span class="lineno"> 1917</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01918"></a><span class="lineno"> 1918</span>  vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l01919"></a><span class="lineno"> 1919</span>  vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 3 * dst_stride_y + zout.s3));</div><div class="line"><a name="l01920"></a><span class="lineno"> 1920</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01921"></a><span class="lineno"> 1921</span> </div><div class="line"><a name="l01922"></a><span class="lineno"> 1922</span> <span class="preprocessor">#else // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01923"></a><span class="lineno"> 1923</span>  <span class="comment">// Add offset for batched GEMM</span></div><div class="line"><a name="l01924"></a><span class="lineno"> 1924</span>  dst_addr += z * dst_stride_z;</div><div class="line"><a name="l01925"></a><span class="lineno"> 1925</span> </div><div class="line"><a name="l01926"></a><span class="lineno"> 1926</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l01927"></a><span class="lineno"> 1927</span>  vstore4((int4)(acc00, acc01, acc02, acc03), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01928"></a><span class="lineno"> 1928</span>  vstore4((int4)(acc04, acc05, acc06, acc07), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 0 * dst_stride_y));</div><div class="line"><a name="l01929"></a><span class="lineno"> 1929</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01930"></a><span class="lineno"> 1930</span>  vstore4((int4)(acc10, acc11, acc12, acc13), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01931"></a><span class="lineno"> 1931</span>  vstore4((int4)(acc14, acc15, acc16, acc17), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 1 * dst_stride_y));</div><div class="line"><a name="l01932"></a><span class="lineno"> 1932</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1</span></div><div class="line"><a name="l01933"></a><span class="lineno"> 1933</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01934"></a><span class="lineno"> 1934</span>  vstore4((int4)(acc20, acc21, acc22, acc23), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01935"></a><span class="lineno"> 1935</span>  vstore4((int4)(acc24, acc25, acc26, acc27), 1, (__global <span class="keywordtype">int</span> *)(dst_addr + 2 * dst_stride_y));</div><div class="line"><a name="l01936"></a><span class="lineno"> 1936</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2</span></div><div class="line"><a name="l01937"></a><span class="lineno"> 1937</span> <span class="preprocessor">#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01938"></a><span class="lineno"> 1938</span>  vstore4((int4)(acc30, acc31, acc32, acc33), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01939"></a><span class="lineno"> 1939</span>  vstore4((int4)(acc34, acc35, acc36, acc37), 0, (__global <span class="keywordtype">int</span> *)(dst_addr + 3 * dst_stride_y));</div><div class="line"><a name="l01940"></a><span class="lineno"> 1940</span> <span class="preprocessor">#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3</span></div><div class="line"><a name="l01941"></a><span class="lineno"> 1941</span> <span class="preprocessor">#endif // defined(REINTERPRET_OUTPUT_AS_3D)</span></div><div class="line"><a name="l01942"></a><span class="lineno"> 1942</span> }</div><div class="line"><a name="l01943"></a><span class="lineno"> 1943</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l01944"></a><span class="lineno"> 1944</span> <span class="preprocessor">#endif // defined(NUM_ELEMS_PROCESSED_PER_THREAD_X) && defined(NUM_ELEMS_PROCESSED_PER_THREAD_Y) && defined(COLS_A)</span></div><div class="line"><a name="l01945"></a><span class="lineno"> 1945</span> </div><div class="line"><a name="l01946"></a><span class="lineno"> 1946</span> <span class="preprocessor">#if defined(COLS_A)</span></div><div class="line"><a name="l01947"></a><span class="lineno"> 1947</span> </div><div class="line"><a name="l01969"></a><span class="lineno"> 1969</span> __kernel <span class="keywordtype">void</span> gemmlowp_matrix_a_reduction(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a>),</div><div class="line"><a name="l01970"></a><span class="lineno"> 1970</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l01971"></a><span class="lineno"> 1971</span> {</div><div class="line"><a name="l01972"></a><span class="lineno"> 1972</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l01973"></a><span class="lineno"> 1973</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#a6743f0a130e8311e6f5b1a23df102472">src</a> = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l01974"></a><span class="lineno"> 1974</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l01975"></a><span class="lineno"> 1975</span> </div><div class="line"><a name="l01976"></a><span class="lineno"> 1976</span>  uint4 sum_row_u32 = (uint4)0;</div><div class="line"><a name="l01977"></a><span class="lineno"> 1977</span>  uint sum_row = 0;</div><div class="line"><a name="l01978"></a><span class="lineno"> 1978</span> </div><div class="line"><a name="l01979"></a><span class="lineno"> 1979</span>  __global <span class="keyword">const</span> uchar *matrix_a = (__global <span class="keyword">const</span> uchar *)(src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z);</div><div class="line"><a name="l01980"></a><span class="lineno"> 1980</span> </div><div class="line"><a name="l01981"></a><span class="lineno"> 1981</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l01982"></a><span class="lineno"> 1982</span> </div><div class="line"><a name="l01983"></a><span class="lineno"> 1983</span>  <span class="comment">// This for loop performs 16 accumulations</span></div><div class="line"><a name="l01984"></a><span class="lineno"> 1984</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 16); i += 16)</div><div class="line"><a name="l01985"></a><span class="lineno"> 1985</span>  {</div><div class="line"><a name="l01986"></a><span class="lineno"> 1986</span>  <span class="keyword">const</span> uchar16 a0_u8 = vload16(0, matrix_a + i);</div><div class="line"><a name="l01987"></a><span class="lineno"> 1987</span> </div><div class="line"><a name="l01988"></a><span class="lineno"> 1988</span>  sum_row_u32 += convert_uint4(a0_u8.s0123) + convert_uint4(a0_u8.s4567) + convert_uint4(a0_u8.s89AB) + convert_uint4(a0_u8.sCDEF);</div><div class="line"><a name="l01989"></a><span class="lineno"> 1989</span>  }</div><div class="line"><a name="l01990"></a><span class="lineno"> 1990</span> </div><div class="line"><a name="l01991"></a><span class="lineno"> 1991</span>  <span class="comment">// This for loop performs the leftover accumulations</span></div><div class="line"><a name="l01992"></a><span class="lineno"> 1992</span>  <span class="keywordflow">for</span>(; i < COLS_A; ++i)</div><div class="line"><a name="l01993"></a><span class="lineno"> 1993</span>  {</div><div class="line"><a name="l01994"></a><span class="lineno"> 1994</span>  sum_row += matrix_a[i];</div><div class="line"><a name="l01995"></a><span class="lineno"> 1995</span>  }</div><div class="line"><a name="l01996"></a><span class="lineno"> 1996</span> </div><div class="line"><a name="l01997"></a><span class="lineno"> 1997</span>  sum_row += sum_row_u32.s0 + sum_row_u32.s1 + sum_row_u32.s2 + sum_row_u32.s3;</div><div class="line"><a name="l01998"></a><span class="lineno"> 1998</span> </div><div class="line"><a name="l01999"></a><span class="lineno"> 1999</span>  *((__global <span class="keywordtype">int</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) = (int)sum_row;</div><div class="line"><a name="l02000"></a><span class="lineno"> 2000</span> }</div><div class="line"><a name="l02001"></a><span class="lineno"> 2001</span> </div><div class="line"><a name="l02002"></a><span class="lineno"> 2002</span> <span class="preprocessor">#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l02003"></a><span class="lineno"> 2003</span> </div><div class="line"><a name="l02025"></a><span class="lineno"> 2025</span> __kernel <span class="keywordtype">void</span> gemmlowp_matrix_a_reduction_dot8(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l02026"></a><span class="lineno"> 2026</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02027"></a><span class="lineno"> 2027</span> {</div><div class="line"><a name="l02028"></a><span class="lineno"> 2028</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02029"></a><span class="lineno"> 2029</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l02030"></a><span class="lineno"> 2030</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02031"></a><span class="lineno"> 2031</span> </div><div class="line"><a name="l02032"></a><span class="lineno"> 2032</span>  uint sum_row = 0;</div><div class="line"><a name="l02033"></a><span class="lineno"> 2033</span> </div><div class="line"><a name="l02034"></a><span class="lineno"> 2034</span>  __global <span class="keyword">const</span> uchar *matrix_a = (__global <span class="keyword">const</span> uchar *)(src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(0) * src_stride_y + get_global_id(1) * src_stride_z);</div><div class="line"><a name="l02035"></a><span class="lineno"> 2035</span> </div><div class="line"><a name="l02036"></a><span class="lineno"> 2036</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l02037"></a><span class="lineno"> 2037</span> </div><div class="line"><a name="l02038"></a><span class="lineno"> 2038</span>  <span class="comment">// This for loop performs 16 accumulations</span></div><div class="line"><a name="l02039"></a><span class="lineno"> 2039</span>  <span class="keywordflow">for</span>(; i <= ((int)COLS_A - 32); i += 32)</div><div class="line"><a name="l02040"></a><span class="lineno"> 2040</span>  {</div><div class="line"><a name="l02041"></a><span class="lineno"> 2041</span>  uchar16 a0_u8 = vload16(0, matrix_a + i);</div><div class="line"><a name="l02042"></a><span class="lineno"> 2042</span> </div><div class="line"><a name="l02043"></a><span class="lineno"> 2043</span>  sum_row += arm_dot(a0_u8.s0123, (uchar4)(1));</div><div class="line"><a name="l02044"></a><span class="lineno"> 2044</span>  sum_row += arm_dot(a0_u8.s4567, (uchar4)(1));</div><div class="line"><a name="l02045"></a><span class="lineno"> 2045</span>  sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1));</div><div class="line"><a name="l02046"></a><span class="lineno"> 2046</span>  sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1));</div><div class="line"><a name="l02047"></a><span class="lineno"> 2047</span> </div><div class="line"><a name="l02048"></a><span class="lineno"> 2048</span>  a0_u8 = vload16(1, matrix_a + i);</div><div class="line"><a name="l02049"></a><span class="lineno"> 2049</span> </div><div class="line"><a name="l02050"></a><span class="lineno"> 2050</span>  sum_row += arm_dot(a0_u8.s0123, (uchar4)(1));</div><div class="line"><a name="l02051"></a><span class="lineno"> 2051</span>  sum_row += arm_dot(a0_u8.s4567, (uchar4)(1));</div><div class="line"><a name="l02052"></a><span class="lineno"> 2052</span>  sum_row += arm_dot(a0_u8.s89AB, (uchar4)(1));</div><div class="line"><a name="l02053"></a><span class="lineno"> 2053</span>  sum_row += arm_dot(a0_u8.sCDEF, (uchar4)(1));</div><div class="line"><a name="l02054"></a><span class="lineno"> 2054</span>  }</div><div class="line"><a name="l02055"></a><span class="lineno"> 2055</span> </div><div class="line"><a name="l02056"></a><span class="lineno"> 2056</span>  <span class="comment">// This for loop performs the leftover accumulations</span></div><div class="line"><a name="l02057"></a><span class="lineno"> 2057</span>  <span class="keywordflow">for</span>(; i < COLS_A; ++i)</div><div class="line"><a name="l02058"></a><span class="lineno"> 2058</span>  {</div><div class="line"><a name="l02059"></a><span class="lineno"> 2059</span>  sum_row += matrix_a[i];</div><div class="line"><a name="l02060"></a><span class="lineno"> 2060</span>  }</div><div class="line"><a name="l02061"></a><span class="lineno"> 2061</span> </div><div class="line"><a name="l02062"></a><span class="lineno"> 2062</span>  *((__global <span class="keywordtype">int</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>) = (int)sum_row;</div><div class="line"><a name="l02063"></a><span class="lineno"> 2063</span> }</div><div class="line"><a name="l02064"></a><span class="lineno"> 2064</span> <span class="preprocessor">#endif // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8)</span></div><div class="line"><a name="l02065"></a><span class="lineno"> 2065</span> <span class="preprocessor">#endif // defined(COLS_A)</span></div><div class="line"><a name="l02066"></a><span class="lineno"> 2066</span> </div><div class="line"><a name="l02067"></a><span class="lineno"> 2067</span> <span class="preprocessor">#if defined(COLS_B) && defined(ROWS_B)</span></div><div class="line"><a name="l02068"></a><span class="lineno"> 2068</span> </div><div class="line"><a name="l02090"></a><span class="lineno"> 2090</span> __kernel <span class="keywordtype">void</span> gemmlowp_matrix_b_reduction(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l02091"></a><span class="lineno"> 2091</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(dst))</div><div class="line"><a name="l02092"></a><span class="lineno"> 2092</span> {</div><div class="line"><a name="l02093"></a><span class="lineno"> 2093</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02094"></a><span class="lineno"> 2094</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(src);</div><div class="line"><a name="l02095"></a><span class="lineno"> 2095</span>  <a class="code" href="struct_image.xhtml">Image</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(dst);</div><div class="line"><a name="l02096"></a><span class="lineno"> 2096</span> </div><div class="line"><a name="l02097"></a><span class="lineno"> 2097</span>  uint16 sum_col_u32 = (uint16)0;</div><div class="line"><a name="l02098"></a><span class="lineno"> 2098</span> </div><div class="line"><a name="l02099"></a><span class="lineno"> 2099</span>  __global <span class="keyword">const</span> uchar *matrix_b = (__global <span class="keyword">const</span> uchar *)(src.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + get_global_id(1) * src_stride_z);</div><div class="line"><a name="l02100"></a><span class="lineno"> 2100</span> </div><div class="line"><a name="l02101"></a><span class="lineno"> 2101</span>  <span class="keywordtype">int</span> i = 0;</div><div class="line"><a name="l02102"></a><span class="lineno"> 2102</span>  <span class="comment">// This for loop performs 4 accumulations</span></div><div class="line"><a name="l02103"></a><span class="lineno"> 2103</span>  <span class="keywordflow">for</span>(; i <= ((int)ROWS_B - 4); i += 4)</div><div class="line"><a name="l02104"></a><span class="lineno"> 2104</span>  {</div><div class="line"><a name="l02105"></a><span class="lineno"> 2105</span>  <span class="keyword">const</span> uchar16 b0_u8 = vload16(0, matrix_b + 0 * src_stride_y);</div><div class="line"><a name="l02106"></a><span class="lineno"> 2106</span>  <span class="keyword">const</span> uchar16 b1_u8 = vload16(0, matrix_b + 1 * src_stride_y);</div><div class="line"><a name="l02107"></a><span class="lineno"> 2107</span>  <span class="keyword">const</span> uchar16 b2_u8 = vload16(0, matrix_b + 2 * src_stride_y);</div><div class="line"><a name="l02108"></a><span class="lineno"> 2108</span>  <span class="keyword">const</span> uchar16 b3_u8 = vload16(0, matrix_b + 3 * src_stride_y);</div><div class="line"><a name="l02109"></a><span class="lineno"> 2109</span> </div><div class="line"><a name="l02110"></a><span class="lineno"> 2110</span>  sum_col_u32 += convert_uint16(b0_u8) + convert_uint16(b1_u8) + convert_uint16(b2_u8) + convert_uint16(b3_u8);</div><div class="line"><a name="l02111"></a><span class="lineno"> 2111</span> </div><div class="line"><a name="l02112"></a><span class="lineno"> 2112</span>  matrix_b += 4 * src_stride_y;</div><div class="line"><a name="l02113"></a><span class="lineno"> 2113</span>  }</div><div class="line"><a name="l02114"></a><span class="lineno"> 2114</span> </div><div class="line"><a name="l02115"></a><span class="lineno"> 2115</span>  <span class="comment">// This for loop perfoms the leftover accumulations</span></div><div class="line"><a name="l02116"></a><span class="lineno"> 2116</span>  <span class="keywordflow">for</span>(; i < (int)ROWS_B; ++i)</div><div class="line"><a name="l02117"></a><span class="lineno"> 2117</span>  {</div><div class="line"><a name="l02118"></a><span class="lineno"> 2118</span>  <span class="keyword">const</span> uchar16 b0_u8 = vload16(0, matrix_b);</div><div class="line"><a name="l02119"></a><span class="lineno"> 2119</span> </div><div class="line"><a name="l02120"></a><span class="lineno"> 2120</span>  sum_col_u32 += convert_uint16(b0_u8);</div><div class="line"><a name="l02121"></a><span class="lineno"> 2121</span> </div><div class="line"><a name="l02122"></a><span class="lineno"> 2122</span>  matrix_b += src_stride_y;</div><div class="line"><a name="l02123"></a><span class="lineno"> 2123</span>  }</div><div class="line"><a name="l02124"></a><span class="lineno"> 2124</span> </div><div class="line"><a name="l02125"></a><span class="lineno"> 2125</span>  vstore16(convert_int16(sum_col_u32), 0, (__global <span class="keywordtype">int</span> *)dst.<a class="code" href="struct_image.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l02126"></a><span class="lineno"> 2126</span> }</div><div class="line"><a name="l02127"></a><span class="lineno"> 2127</span> <span class="preprocessor">#endif // defined(COLS_B) && defined(ROWS_B)</span></div><div class="line"><a name="l02128"></a><span class="lineno"> 2128</span> </div><div class="line"><a name="l02129"></a><span class="lineno"> 2129</span> <span class="preprocessor">#if defined(K_OFFSET)</span></div><div class="line"><a name="l02130"></a><span class="lineno"> 2130</span> </div><div class="line"><a name="l02131"></a><span class="lineno"> 2131</span> <span class="comment">/* Helper function used to calculate the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel.</span></div><div class="line"><a name="l02132"></a><span class="lineno"> 2132</span> <span class="comment"> *</span></div><div class="line"><a name="l02133"></a><span class="lineno"> 2133</span> <span class="comment"> * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel),</span></div><div class="line"><a name="l02134"></a><span class="lineno"> 2134</span> <span class="comment"> * and calculates the offset contribution of matrix A and matrix B.</span></div><div class="line"><a name="l02135"></a><span class="lineno"> 2135</span> <span class="comment"> *</span></div><div class="line"><a name="l02136"></a><span class="lineno"> 2136</span> <span class="comment"> * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)</span></div><div class="line"><a name="l02137"></a><span class="lineno"> 2137</span> <span class="comment"> * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)</span></div><div class="line"><a name="l02138"></a><span class="lineno"> 2138</span> <span class="comment"> * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)</span></div><div class="line"><a name="l02139"></a><span class="lineno"> 2139</span> <span class="comment"> * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches</span></div><div class="line"><a name="l02140"></a><span class="lineno"> 2140</span> <span class="comment"> *</span></div><div class="line"><a name="l02141"></a><span class="lineno"> 2141</span> <span class="comment"> * @param[in] x get_global_id(0) * 4</span></div><div class="line"><a name="l02142"></a><span class="lineno"> 2142</span> <span class="comment"> * @param[in] y get_global_id(1)</span></div><div class="line"><a name="l02143"></a><span class="lineno"> 2143</span> <span class="comment"> * @param[in] z get_global_id(2)</span></div><div class="line"><a name="l02144"></a><span class="lineno"> 2144</span> <span class="comment"> * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02145"></a><span class="lineno"> 2145</span> <span class="comment"> * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02146"></a><span class="lineno"> 2146</span> <span class="comment"> * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02147"></a><span class="lineno"> 2147</span> <span class="comment"> * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02148"></a><span class="lineno"> 2148</span> <span class="comment"> * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02149"></a><span class="lineno"> 2149</span> <span class="comment"> * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02150"></a><span class="lineno"> 2150</span> <span class="comment"> * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02151"></a><span class="lineno"> 2151</span> <span class="comment"> * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02152"></a><span class="lineno"> 2152</span> <span class="comment"> * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02153"></a><span class="lineno"> 2153</span> <span class="comment"> * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02154"></a><span class="lineno"> 2154</span> <span class="comment"> * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02155"></a><span class="lineno"> 2155</span> <span class="comment"> * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02156"></a><span class="lineno"> 2156</span> <span class="comment"> * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr</span></div><div class="line"><a name="l02157"></a><span class="lineno"> 2157</span> <span class="comment"> * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02158"></a><span class="lineno"> 2158</span> <span class="comment"> * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02159"></a><span class="lineno"> 2159</span> <span class="comment"> * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor</span></div><div class="line"><a name="l02160"></a><span class="lineno"> 2160</span> <span class="comment"> */</span></div><div class="line"><a name="l02161"></a><span class="lineno"> 2161</span> <span class="keyword">inline</span> int4 offset_contribution(</div><div class="line"><a name="l02162"></a><span class="lineno"> 2162</span>  <span class="keywordtype">int</span> x,</div><div class="line"><a name="l02163"></a><span class="lineno"> 2163</span>  <span class="keywordtype">int</span> y,</div><div class="line"><a name="l02164"></a><span class="lineno"> 2164</span>  <span class="keywordtype">int</span> z</div><div class="line"><a name="l02165"></a><span class="lineno"> 2165</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02166"></a><span class="lineno"> 2166</span>  ,</div><div class="line"><a name="l02167"></a><span class="lineno"> 2167</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_col)</div><div class="line"><a name="l02168"></a><span class="lineno"> 2168</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02169"></a><span class="lineno"> 2169</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02170"></a><span class="lineno"> 2170</span>  ,</div><div class="line"><a name="l02171"></a><span class="lineno"> 2171</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_row)</div><div class="line"><a name="l02172"></a><span class="lineno"> 2172</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02173"></a><span class="lineno"> 2173</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02174"></a><span class="lineno"> 2174</span>  ,</div><div class="line"><a name="l02175"></a><span class="lineno"> 2175</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l02176"></a><span class="lineno"> 2176</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02177"></a><span class="lineno"> 2177</span> )</div><div class="line"><a name="l02178"></a><span class="lineno"> 2178</span> {</div><div class="line"><a name="l02179"></a><span class="lineno"> 2179</span>  int4 a_offset_s32 = (int4)0;</div><div class="line"><a name="l02180"></a><span class="lineno"> 2180</span>  int4 b_offset_s32 = (int4)0;</div><div class="line"><a name="l02181"></a><span class="lineno"> 2181</span> </div><div class="line"><a name="l02182"></a><span class="lineno"> 2182</span>  <span class="keywordtype">int</span> batch_id = z;</div><div class="line"><a name="l02183"></a><span class="lineno"> 2183</span> <span class="preprocessor">#if defined(DEPTH_INPUT3D)</span></div><div class="line"><a name="l02184"></a><span class="lineno"> 2184</span>  batch_id /= (int)DEPTH_INPUT3D;</div><div class="line"><a name="l02185"></a><span class="lineno"> 2185</span> <span class="preprocessor">#endif // defined(DEPTH_INPUT3D)</span></div><div class="line"><a name="l02186"></a><span class="lineno"> 2186</span> </div><div class="line"><a name="l02187"></a><span class="lineno"> 2187</span> <span class="preprocessor">#if defined(A_OFFSET)</span></div><div class="line"><a name="l02188"></a><span class="lineno"> 2188</span>  <span class="comment">// Compute the offset contribution due to A_OFFSET</span></div><div class="line"><a name="l02189"></a><span class="lineno"> 2189</span>  __global uchar *sum_col_addr = sum_col_ptr + sum_col_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02190"></a><span class="lineno"> 2190</span> </div><div class="line"><a name="l02191"></a><span class="lineno"> 2191</span>  <span class="comment">// Compute the offset contribution due to A_OFFSET</span></div><div class="line"><a name="l02192"></a><span class="lineno"> 2192</span> <span class="preprocessor">#if defined(SUM_COL_HAS_BATCHES)</span></div><div class="line"><a name="l02193"></a><span class="lineno"> 2193</span>  a_offset_s32 = vload4(0, (__global <span class="keywordtype">int</span> *)(sum_col_addr + batch_id * sum_col_stride_y));</div><div class="line"><a name="l02194"></a><span class="lineno"> 2194</span> <span class="preprocessor">#else // defined(SUM_COL_HAS_BATCHES)</span></div><div class="line"><a name="l02195"></a><span class="lineno"> 2195</span>  a_offset_s32 = vload4(0, (__global <span class="keywordtype">int</span> *)sum_col_addr);</div><div class="line"><a name="l02196"></a><span class="lineno"> 2196</span> <span class="preprocessor">#endif // defined(SUM_COL_HAS_BATCHES)</span></div><div class="line"><a name="l02197"></a><span class="lineno"> 2197</span> </div><div class="line"><a name="l02198"></a><span class="lineno"> 2198</span>  a_offset_s32 *= (int4)A_OFFSET;</div><div class="line"><a name="l02199"></a><span class="lineno"> 2199</span> <span class="preprocessor">#endif // defined(A_OFFSET)</span></div><div class="line"><a name="l02200"></a><span class="lineno"> 2200</span> </div><div class="line"><a name="l02201"></a><span class="lineno"> 2201</span> <span class="preprocessor">#if defined(B_OFFSET)</span></div><div class="line"><a name="l02202"></a><span class="lineno"> 2202</span>  <span class="comment">// Compute the offset contribution due to A_OFFSET</span></div><div class="line"><a name="l02203"></a><span class="lineno"> 2203</span>  __global uchar *sum_row_addr = sum_row_ptr + sum_row_offset_first_element_in_bytes + y * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02204"></a><span class="lineno"> 2204</span> </div><div class="line"><a name="l02205"></a><span class="lineno"> 2205</span>  <span class="comment">// Compute the offset contribution due to B_OFFSET</span></div><div class="line"><a name="l02206"></a><span class="lineno"> 2206</span> <span class="preprocessor">#if defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)</span></div><div class="line"><a name="l02207"></a><span class="lineno"> 2207</span>  b_offset_s32 = (int4) * (((__global <span class="keywordtype">int</span> *)(sum_row_addr + batch_id * sum_row_stride_y)) + (z % (int)DEPTH_INPUT3D) * (int)HEIGHT_INPUT3D);</div><div class="line"><a name="l02208"></a><span class="lineno"> 2208</span> <span class="preprocessor">#else // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)</span></div><div class="line"><a name="l02209"></a><span class="lineno"> 2209</span>  b_offset_s32 = (int4) * (((__global <span class="keywordtype">int</span> *)(sum_row_addr + batch_id * sum_row_stride_y)));</div><div class="line"><a name="l02210"></a><span class="lineno"> 2210</span> <span class="preprocessor">#endif // defined(HEIGHT_INPUT3D) && defined(DEPTH_INPUT3D)</span></div><div class="line"><a name="l02211"></a><span class="lineno"> 2211</span>  b_offset_s32 *= (int4)B_OFFSET;</div><div class="line"><a name="l02212"></a><span class="lineno"> 2212</span> <span class="preprocessor">#endif // defined(B_OFFSET)</span></div><div class="line"><a name="l02213"></a><span class="lineno"> 2213</span> </div><div class="line"><a name="l02214"></a><span class="lineno"> 2214</span> <span class="preprocessor">#if defined(ADD_BIAS)</span></div><div class="line"><a name="l02215"></a><span class="lineno"> 2215</span>  <span class="comment">// Add bias</span></div><div class="line"><a name="l02216"></a><span class="lineno"> 2216</span>  __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02217"></a><span class="lineno"> 2217</span> </div><div class="line"><a name="l02218"></a><span class="lineno"> 2218</span>  int4 biases_values = vload4(0, (__global <span class="keywordtype">int</span> *)bias_addr);</div><div class="line"><a name="l02219"></a><span class="lineno"> 2219</span>  b_offset_s32 += (int4)biases_values;</div><div class="line"><a name="l02220"></a><span class="lineno"> 2220</span> <span class="preprocessor">#endif // defined(ADD_BIAS)</span></div><div class="line"><a name="l02221"></a><span class="lineno"> 2221</span> </div><div class="line"><a name="l02222"></a><span class="lineno"> 2222</span>  <span class="keywordflow">return</span> (int4)K_OFFSET + a_offset_s32 + b_offset_s32;</div><div class="line"><a name="l02223"></a><span class="lineno"> 2223</span> }</div><div class="line"><a name="l02224"></a><span class="lineno"> 2224</span> </div><div class="line"><a name="l02225"></a><span class="lineno"> 2225</span> <span class="comment">/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel. The computation is performed in-place</span></div><div class="line"><a name="l02226"></a><span class="lineno"> 2226</span> <span class="comment"> *</span></div><div class="line"><a name="l02227"></a><span class="lineno"> 2227</span> <span class="comment"> * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel),</span></div><div class="line"><a name="l02228"></a><span class="lineno"> 2228</span> <span class="comment"> * and adds to it the offset contribution of matrix A and matrix B in-place.</span></div><div class="line"><a name="l02229"></a><span class="lineno"> 2229</span> <span class="comment"> *</span></div><div class="line"><a name="l02230"></a><span class="lineno"> 2230</span> <span class="comment"> * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)</span></div><div class="line"><a name="l02231"></a><span class="lineno"> 2231</span> <span class="comment"> * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)</span></div><div class="line"><a name="l02232"></a><span class="lineno"> 2232</span> <span class="comment"> * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)</span></div><div class="line"><a name="l02233"></a><span class="lineno"> 2233</span> <span class="comment"> * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches</span></div><div class="line"><a name="l02234"></a><span class="lineno"> 2234</span> <span class="comment"> *</span></div><div class="line"><a name="l02235"></a><span class="lineno"> 2235</span> <span class="comment"> * The final result is:</span></div><div class="line"><a name="l02236"></a><span class="lineno"> 2236</span> <span class="comment"> *</span></div><div class="line"><a name="l02237"></a><span class="lineno"> 2237</span> <span class="comment"> * mm_result[i][k] = mm_result[i][k] +</span></div><div class="line"><a name="l02238"></a><span class="lineno"> 2238</span> <span class="comment"> * (sum_col[k] * A_OFFSET) +</span></div><div class="line"><a name="l02239"></a><span class="lineno"> 2239</span> <span class="comment"> * (sum_row[i] * B_OFFSET) +</span></div><div class="line"><a name="l02240"></a><span class="lineno"> 2240</span> <span class="comment"> * (K_OFFSET)</span></div><div class="line"><a name="l02241"></a><span class="lineno"> 2241</span> <span class="comment"> *</span></div><div class="line"><a name="l02242"></a><span class="lineno"> 2242</span> <span class="comment"> * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32</span></div><div class="line"><a name="l02243"></a><span class="lineno"> 2243</span> <span class="comment"> * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02244"></a><span class="lineno"> 2244</span> <span class="comment"> * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02245"></a><span class="lineno"> 2245</span> <span class="comment"> * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02246"></a><span class="lineno"> 2246</span> <span class="comment"> * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02247"></a><span class="lineno"> 2247</span> <span class="comment"> * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)</span></div><div class="line"><a name="l02248"></a><span class="lineno"> 2248</span> <span class="comment"> * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)</span></div><div class="line"><a name="l02249"></a><span class="lineno"> 2249</span> <span class="comment"> * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor</span></div><div class="line"><a name="l02250"></a><span class="lineno"> 2250</span> <span class="comment"> * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02251"></a><span class="lineno"> 2251</span> <span class="comment"> * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02252"></a><span class="lineno"> 2252</span> <span class="comment"> * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02253"></a><span class="lineno"> 2253</span> <span class="comment"> * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02254"></a><span class="lineno"> 2254</span> <span class="comment"> * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02255"></a><span class="lineno"> 2255</span> <span class="comment"> * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02256"></a><span class="lineno"> 2256</span> <span class="comment"> * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02257"></a><span class="lineno"> 2257</span> <span class="comment"> * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02258"></a><span class="lineno"> 2258</span> <span class="comment"> * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02259"></a><span class="lineno"> 2259</span> <span class="comment"> * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02260"></a><span class="lineno"> 2260</span> <span class="comment"> * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02261"></a><span class="lineno"> 2261</span> <span class="comment"> * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02262"></a><span class="lineno"> 2262</span> <span class="comment"> * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr</span></div><div class="line"><a name="l02263"></a><span class="lineno"> 2263</span> <span class="comment"> * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02264"></a><span class="lineno"> 2264</span> <span class="comment"> * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02265"></a><span class="lineno"> 2265</span> <span class="comment"> * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor</span></div><div class="line"><a name="l02266"></a><span class="lineno"> 2266</span> <span class="comment"> */</span></div><div class="line"><a name="l02267"></a><span class="lineno"> 2267</span> __kernel <span class="keywordtype">void</span> gemmlowp_offset_contribution(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(mm_result)</div><div class="line"><a name="l02268"></a><span class="lineno"> 2268</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02269"></a><span class="lineno"> 2269</span>  ,</div><div class="line"><a name="l02270"></a><span class="lineno"> 2270</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_col)</div><div class="line"><a name="l02271"></a><span class="lineno"> 2271</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02272"></a><span class="lineno"> 2272</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02273"></a><span class="lineno"> 2273</span>  ,</div><div class="line"><a name="l02274"></a><span class="lineno"> 2274</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_row)</div><div class="line"><a name="l02275"></a><span class="lineno"> 2275</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02276"></a><span class="lineno"> 2276</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02277"></a><span class="lineno"> 2277</span>  ,</div><div class="line"><a name="l02278"></a><span class="lineno"> 2278</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases)</div><div class="line"><a name="l02279"></a><span class="lineno"> 2279</span> #endif <span class="comment">// defined(ADD_BIAS))</span></div><div class="line"><a name="l02280"></a><span class="lineno"> 2280</span>  )</div><div class="line"><a name="l02281"></a><span class="lineno"> 2281</span> {</div><div class="line"><a name="l02282"></a><span class="lineno"> 2282</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02283"></a><span class="lineno"> 2283</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02284"></a><span class="lineno"> 2284</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02285"></a><span class="lineno"> 2285</span> </div><div class="line"><a name="l02286"></a><span class="lineno"> 2286</span>  <span class="comment">// Compute offset contribution</span></div><div class="line"><a name="l02287"></a><span class="lineno"> 2287</span>  int4 offset_term_s32 = offset_contribution(</div><div class="line"><a name="l02288"></a><span class="lineno"> 2288</span>  x, y, z</div><div class="line"><a name="l02289"></a><span class="lineno"> 2289</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02290"></a><span class="lineno"> 2290</span>  ,</div><div class="line"><a name="l02291"></a><span class="lineno"> 2291</span>  sum_col_ptr,</div><div class="line"><a name="l02292"></a><span class="lineno"> 2292</span>  sum_col_stride_x,</div><div class="line"><a name="l02293"></a><span class="lineno"> 2293</span>  sum_col_step_x,</div><div class="line"><a name="l02294"></a><span class="lineno"> 2294</span>  sum_col_stride_y,</div><div class="line"><a name="l02295"></a><span class="lineno"> 2295</span>  sum_col_step_y,</div><div class="line"><a name="l02296"></a><span class="lineno"> 2296</span>  sum_col_offset_first_element_in_bytes</div><div class="line"><a name="l02297"></a><span class="lineno"> 2297</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02298"></a><span class="lineno"> 2298</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02299"></a><span class="lineno"> 2299</span>  ,</div><div class="line"><a name="l02300"></a><span class="lineno"> 2300</span>  sum_row_ptr,</div><div class="line"><a name="l02301"></a><span class="lineno"> 2301</span>  sum_row_stride_x,</div><div class="line"><a name="l02302"></a><span class="lineno"> 2302</span>  sum_row_step_x,</div><div class="line"><a name="l02303"></a><span class="lineno"> 2303</span>  sum_row_stride_y,</div><div class="line"><a name="l02304"></a><span class="lineno"> 2304</span>  sum_row_step_y,</div><div class="line"><a name="l02305"></a><span class="lineno"> 2305</span>  sum_row_offset_first_element_in_bytes</div><div class="line"><a name="l02306"></a><span class="lineno"> 2306</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02307"></a><span class="lineno"> 2307</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02308"></a><span class="lineno"> 2308</span>  ,</div><div class="line"><a name="l02309"></a><span class="lineno"> 2309</span>  biases_ptr,</div><div class="line"><a name="l02310"></a><span class="lineno"> 2310</span>  biases_stride_x,</div><div class="line"><a name="l02311"></a><span class="lineno"> 2311</span>  biases_step_x,</div><div class="line"><a name="l02312"></a><span class="lineno"> 2312</span>  biases_offset_first_element_in_bytes</div><div class="line"><a name="l02313"></a><span class="lineno"> 2313</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02314"></a><span class="lineno"> 2314</span>  );</div><div class="line"><a name="l02315"></a><span class="lineno"> 2315</span> </div><div class="line"><a name="l02316"></a><span class="lineno"> 2316</span>  __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * mm_result_stride_y + z * mm_result_stride_z;</div><div class="line"><a name="l02317"></a><span class="lineno"> 2317</span> </div><div class="line"><a name="l02318"></a><span class="lineno"> 2318</span>  int4 in_s32 = vload4(0, (__global <span class="keywordtype">int</span> *)mm_result_addr);</div><div class="line"><a name="l02319"></a><span class="lineno"> 2319</span> </div><div class="line"><a name="l02320"></a><span class="lineno"> 2320</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02321"></a><span class="lineno"> 2321</span>  in_s32 += offset_term_s32;</div><div class="line"><a name="l02322"></a><span class="lineno"> 2322</span> </div><div class="line"><a name="l02323"></a><span class="lineno"> 2323</span>  <span class="comment">// Store the result with the offset contribution</span></div><div class="line"><a name="l02324"></a><span class="lineno"> 2324</span>  vstore4(in_s32, 0, (__global <span class="keywordtype">int</span> *)mm_result_addr);</div><div class="line"><a name="l02325"></a><span class="lineno"> 2325</span> }</div><div class="line"><a name="l02326"></a><span class="lineno"> 2326</span> </div><div class="line"><a name="l02327"></a><span class="lineno"> 2327</span> <span class="preprocessor">#if defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02328"></a><span class="lineno"> 2328</span> <span class="comment">/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.</span></div><div class="line"><a name="l02329"></a><span class="lineno"> 2329</span> <span class="comment"> *</span></div><div class="line"><a name="l02330"></a><span class="lineno"> 2330</span> <span class="comment"> * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.</span></div><div class="line"><a name="l02331"></a><span class="lineno"> 2331</span> <span class="comment"> *</span></div><div class="line"><a name="l02332"></a><span class="lineno"> 2332</span> <span class="comment"> *</span></div><div class="line"><a name="l02333"></a><span class="lineno"> 2333</span> <span class="comment"> * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)</span></div><div class="line"><a name="l02334"></a><span class="lineno"> 2334</span> <span class="comment"> * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)</span></div><div class="line"><a name="l02335"></a><span class="lineno"> 2335</span> <span class="comment"> * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)</span></div><div class="line"><a name="l02336"></a><span class="lineno"> 2336</span> <span class="comment"> * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches</span></div><div class="line"><a name="l02337"></a><span class="lineno"> 2337</span> <span class="comment"> *</span></div><div class="line"><a name="l02338"></a><span class="lineno"> 2338</span> <span class="comment"> * The result before the output stage is:</span></div><div class="line"><a name="l02339"></a><span class="lineno"> 2339</span> <span class="comment"> *</span></div><div class="line"><a name="l02340"></a><span class="lineno"> 2340</span> <span class="comment"> * mm_result[i][k] = mm_result[i][k] +</span></div><div class="line"><a name="l02341"></a><span class="lineno"> 2341</span> <span class="comment"> * (sum_col[k] * A_OFFSET) +</span></div><div class="line"><a name="l02342"></a><span class="lineno"> 2342</span> <span class="comment"> * (sum_row[i] * B_OFFSET) +</span></div><div class="line"><a name="l02343"></a><span class="lineno"> 2343</span> <span class="comment"> * (K_OFFSET)</span></div><div class="line"><a name="l02344"></a><span class="lineno"> 2344</span> <span class="comment"> *</span></div><div class="line"><a name="l02345"></a><span class="lineno"> 2345</span> <span class="comment"> * This result is quantized down to uint8 using the output stage. The output stage computes the following operations:</span></div><div class="line"><a name="l02346"></a><span class="lineno"> 2346</span> <span class="comment"> *</span></div><div class="line"><a name="l02347"></a><span class="lineno"> 2347</span> <span class="comment"> * -# Add offset terms to final result</span></div><div class="line"><a name="l02348"></a><span class="lineno"> 2348</span> <span class="comment"> * -# Multiply each entry of result by result_mult_int</span></div><div class="line"><a name="l02349"></a><span class="lineno"> 2349</span> <span class="comment"> * -# Add bias to final result (if -DADD_BIAS is passed at compile time)</span></div><div class="line"><a name="l02350"></a><span class="lineno"> 2350</span> <span class="comment"> * -# Shift the int32 accumulator by result_shift</span></div><div class="line"><a name="l02351"></a><span class="lineno"> 2351</span> <span class="comment"> * -# Clamp the value between the specified min and max bounds (if -DMIN_BOUND and/or -DMAX_BOUND are passed at compile time)</span></div><div class="line"><a name="l02352"></a><span class="lineno"> 2352</span> <span class="comment"> * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.</span></div><div class="line"><a name="l02353"></a><span class="lineno"> 2353</span> <span class="comment"> *</span></div><div class="line"><a name="l02354"></a><span class="lineno"> 2354</span> <span class="comment"> * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT</span></div><div class="line"><a name="l02355"></a><span class="lineno"> 2355</span> <span class="comment"> *</span></div><div class="line"><a name="l02356"></a><span class="lineno"> 2356</span> <span class="comment"> * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time</span></div><div class="line"><a name="l02357"></a><span class="lineno"> 2357</span> <span class="comment"> * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.</span></div><div class="line"><a name="l02358"></a><span class="lineno"> 2358</span> <span class="comment"> * These values can be used to implement "rectified linear unit" activation functions</span></div><div class="line"><a name="l02359"></a><span class="lineno"> 2359</span> <span class="comment"> *</span></div><div class="line"><a name="l02360"></a><span class="lineno"> 2360</span> <span class="comment"> * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32</span></div><div class="line"><a name="l02361"></a><span class="lineno"> 2361</span> <span class="comment"> * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02362"></a><span class="lineno"> 2362</span> <span class="comment"> * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02363"></a><span class="lineno"> 2363</span> <span class="comment"> * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02364"></a><span class="lineno"> 2364</span> <span class="comment"> * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02365"></a><span class="lineno"> 2365</span> <span class="comment"> * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)</span></div><div class="line"><a name="l02366"></a><span class="lineno"> 2366</span> <span class="comment"> * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)</span></div><div class="line"><a name="l02367"></a><span class="lineno"> 2367</span> <span class="comment"> * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor</span></div><div class="line"><a name="l02368"></a><span class="lineno"> 2368</span> <span class="comment"> * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02369"></a><span class="lineno"> 2369</span> <span class="comment"> * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02370"></a><span class="lineno"> 2370</span> <span class="comment"> * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02371"></a><span class="lineno"> 2371</span> <span class="comment"> * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02372"></a><span class="lineno"> 2372</span> <span class="comment"> * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02373"></a><span class="lineno"> 2373</span> <span class="comment"> * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02374"></a><span class="lineno"> 2374</span> <span class="comment"> * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02375"></a><span class="lineno"> 2375</span> <span class="comment"> * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02376"></a><span class="lineno"> 2376</span> <span class="comment"> * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02377"></a><span class="lineno"> 2377</span> <span class="comment"> * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02378"></a><span class="lineno"> 2378</span> <span class="comment"> * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02379"></a><span class="lineno"> 2379</span> <span class="comment"> * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02380"></a><span class="lineno"> 2380</span> <span class="comment"> * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr</span></div><div class="line"><a name="l02381"></a><span class="lineno"> 2381</span> <span class="comment"> * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02382"></a><span class="lineno"> 2382</span> <span class="comment"> * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02383"></a><span class="lineno"> 2383</span> <span class="comment"> * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor</span></div><div class="line"><a name="l02384"></a><span class="lineno"> 2384</span> <span class="comment"> * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8</span></div><div class="line"><a name="l02385"></a><span class="lineno"> 2385</span> <span class="comment"> * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02386"></a><span class="lineno"> 2386</span> <span class="comment"> * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02387"></a><span class="lineno"> 2387</span> <span class="comment"> * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02388"></a><span class="lineno"> 2388</span> <span class="comment"> * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02389"></a><span class="lineno"> 2389</span> <span class="comment"> * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)</span></div><div class="line"><a name="l02390"></a><span class="lineno"> 2390</span> <span class="comment"> * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)</span></div><div class="line"><a name="l02391"></a><span class="lineno"> 2391</span> <span class="comment"> * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor</span></div><div class="line"><a name="l02392"></a><span class="lineno"> 2392</span> <span class="comment"> */</span></div><div class="line"><a name="l02393"></a><span class="lineno"> 2393</span> __kernel <span class="keywordtype">void</span> gemmlowp_offset_contribution_quantize_down(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(mm_result)</div><div class="line"><a name="l02394"></a><span class="lineno"> 2394</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02395"></a><span class="lineno"> 2395</span>  ,</div><div class="line"><a name="l02396"></a><span class="lineno"> 2396</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_col)</div><div class="line"><a name="l02397"></a><span class="lineno"> 2397</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02398"></a><span class="lineno"> 2398</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02399"></a><span class="lineno"> 2399</span>  ,</div><div class="line"><a name="l02400"></a><span class="lineno"> 2400</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_row)</div><div class="line"><a name="l02401"></a><span class="lineno"> 2401</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02402"></a><span class="lineno"> 2402</span>  ,</div><div class="line"><a name="l02403"></a><span class="lineno"> 2403</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02404"></a><span class="lineno"> 2404</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l02405"></a><span class="lineno"> 2405</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02406"></a><span class="lineno"> 2406</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l02407"></a><span class="lineno"> 2407</span> {</div><div class="line"><a name="l02408"></a><span class="lineno"> 2408</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02409"></a><span class="lineno"> 2409</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02410"></a><span class="lineno"> 2410</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02411"></a><span class="lineno"> 2411</span> </div><div class="line"><a name="l02412"></a><span class="lineno"> 2412</span>  __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;</div><div class="line"><a name="l02413"></a><span class="lineno"> 2413</span> </div><div class="line"><a name="l02414"></a><span class="lineno"> 2414</span>  <span class="comment">// Compute offset contribution</span></div><div class="line"><a name="l02415"></a><span class="lineno"> 2415</span>  int4 offset_term_s32 = offset_contribution(</div><div class="line"><a name="l02416"></a><span class="lineno"> 2416</span>  x, y, z</div><div class="line"><a name="l02417"></a><span class="lineno"> 2417</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02418"></a><span class="lineno"> 2418</span>  ,</div><div class="line"><a name="l02419"></a><span class="lineno"> 2419</span>  sum_col_ptr,</div><div class="line"><a name="l02420"></a><span class="lineno"> 2420</span>  sum_col_stride_x,</div><div class="line"><a name="l02421"></a><span class="lineno"> 2421</span>  sum_col_step_x,</div><div class="line"><a name="l02422"></a><span class="lineno"> 2422</span>  sum_col_stride_y,</div><div class="line"><a name="l02423"></a><span class="lineno"> 2423</span>  sum_col_step_y,</div><div class="line"><a name="l02424"></a><span class="lineno"> 2424</span>  sum_col_offset_first_element_in_bytes</div><div class="line"><a name="l02425"></a><span class="lineno"> 2425</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02426"></a><span class="lineno"> 2426</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02427"></a><span class="lineno"> 2427</span>  ,</div><div class="line"><a name="l02428"></a><span class="lineno"> 2428</span>  sum_row_ptr,</div><div class="line"><a name="l02429"></a><span class="lineno"> 2429</span>  sum_row_stride_x,</div><div class="line"><a name="l02430"></a><span class="lineno"> 2430</span>  sum_row_step_x,</div><div class="line"><a name="l02431"></a><span class="lineno"> 2431</span>  sum_row_stride_y,</div><div class="line"><a name="l02432"></a><span class="lineno"> 2432</span>  sum_row_step_y,</div><div class="line"><a name="l02433"></a><span class="lineno"> 2433</span>  sum_row_offset_first_element_in_bytes</div><div class="line"><a name="l02434"></a><span class="lineno"> 2434</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02435"></a><span class="lineno"> 2435</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02436"></a><span class="lineno"> 2436</span>  ,</div><div class="line"><a name="l02437"></a><span class="lineno"> 2437</span>  biases_ptr,</div><div class="line"><a name="l02438"></a><span class="lineno"> 2438</span>  biases_stride_x,</div><div class="line"><a name="l02439"></a><span class="lineno"> 2439</span>  biases_step_x,</div><div class="line"><a name="l02440"></a><span class="lineno"> 2440</span>  biases_offset_first_element_in_bytes</div><div class="line"><a name="l02441"></a><span class="lineno"> 2441</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02442"></a><span class="lineno"> 2442</span>  );</div><div class="line"><a name="l02443"></a><span class="lineno"> 2443</span> </div><div class="line"><a name="l02444"></a><span class="lineno"> 2444</span>  __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * mm_result_stride_y + z * mm_result_stride_z;</div><div class="line"><a name="l02445"></a><span class="lineno"> 2445</span> </div><div class="line"><a name="l02446"></a><span class="lineno"> 2446</span>  int4 in_s32 = vload4(0, (__global <span class="keywordtype">int</span> *)mm_result_addr);</div><div class="line"><a name="l02447"></a><span class="lineno"> 2447</span> </div><div class="line"><a name="l02448"></a><span class="lineno"> 2448</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02449"></a><span class="lineno"> 2449</span>  in_s32 += offset_term_s32;</div><div class="line"><a name="l02450"></a><span class="lineno"> 2450</span> </div><div class="line"><a name="l02451"></a><span class="lineno"> 2451</span>  <span class="comment">// -------------- OUTPUT STAGE</span></div><div class="line"><a name="l02452"></a><span class="lineno"> 2452</span> </div><div class="line"><a name="l02453"></a><span class="lineno"> 2453</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02454"></a><span class="lineno"> 2454</span>  in_s32 += (int4)RESULT_OFFSET;</div><div class="line"><a name="l02455"></a><span class="lineno"> 2455</span> </div><div class="line"><a name="l02456"></a><span class="lineno"> 2456</span>  <span class="comment">// Multiply by result_mult_int and shift</span></div><div class="line"><a name="l02457"></a><span class="lineno"> 2457</span>  in_s32 *= RESULT_MULTIPLIER;</div><div class="line"><a name="l02458"></a><span class="lineno"> 2458</span> </div><div class="line"><a name="l02459"></a><span class="lineno"> 2459</span>  in_s32 >>= RESULT_SHIFT;</div><div class="line"><a name="l02460"></a><span class="lineno"> 2460</span> </div><div class="line"><a name="l02461"></a><span class="lineno"> 2461</span>  uchar4 res = convert_uchar4_sat(in_s32);</div><div class="line"><a name="l02462"></a><span class="lineno"> 2462</span> </div><div class="line"><a name="l02463"></a><span class="lineno"> 2463</span> <span class="preprocessor">#if defined(MIN_BOUND)</span></div><div class="line"><a name="l02464"></a><span class="lineno"> 2464</span>  res = max(res, (uchar4)MIN_BOUND);</div><div class="line"><a name="l02465"></a><span class="lineno"> 2465</span> <span class="preprocessor">#endif // defined(MIN_BOUND)</span></div><div class="line"><a name="l02466"></a><span class="lineno"> 2466</span> <span class="preprocessor">#if defined(MAX_BOUND)</span></div><div class="line"><a name="l02467"></a><span class="lineno"> 2467</span>  res = min(res, (uchar4)MAX_BOUND);</div><div class="line"><a name="l02468"></a><span class="lineno"> 2468</span> <span class="preprocessor">#endif // defined(MAX_BOUND)</span></div><div class="line"><a name="l02469"></a><span class="lineno"> 2469</span> </div><div class="line"><a name="l02470"></a><span class="lineno"> 2470</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l02471"></a><span class="lineno"> 2471</span>  vstore4(res, 0, dst_addr);</div><div class="line"><a name="l02472"></a><span class="lineno"> 2472</span> }</div><div class="line"><a name="l02473"></a><span class="lineno"> 2473</span> </div><div class="line"><a name="l02474"></a><span class="lineno"> 2474</span> <span class="comment">/* OpenCL kernel used to add the offset contribution after @ref CLGEMMLowpMatrixMultiplyKernel and it quantizes down to uint8.</span></div><div class="line"><a name="l02475"></a><span class="lineno"> 2475</span> <span class="comment"> *</span></div><div class="line"><a name="l02476"></a><span class="lineno"> 2476</span> <span class="comment"> * This kernel takes a final int32 accumulator value (the output of @CLGEMMLowpMatrixMultiplyKernel), adds to it the offset contribution of matrix A and matrix B and quantizes to uint8 through the output stage.</span></div><div class="line"><a name="l02477"></a><span class="lineno"> 2477</span> <span class="comment"> *</span></div><div class="line"><a name="l02478"></a><span class="lineno"> 2478</span> <span class="comment"> *</span></div><div class="line"><a name="l02479"></a><span class="lineno"> 2479</span> <span class="comment"> * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)</span></div><div class="line"><a name="l02480"></a><span class="lineno"> 2480</span> <span class="comment"> * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)</span></div><div class="line"><a name="l02481"></a><span class="lineno"> 2481</span> <span class="comment"> * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)</span></div><div class="line"><a name="l02482"></a><span class="lineno"> 2482</span> <span class="comment"> * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches</span></div><div class="line"><a name="l02483"></a><span class="lineno"> 2483</span> <span class="comment"> *</span></div><div class="line"><a name="l02484"></a><span class="lineno"> 2484</span> <span class="comment"> * The result before the output stage is:</span></div><div class="line"><a name="l02485"></a><span class="lineno"> 2485</span> <span class="comment"> *</span></div><div class="line"><a name="l02486"></a><span class="lineno"> 2486</span> <span class="comment"> * mm_result[i][k] = mm_result[i][k] +</span></div><div class="line"><a name="l02487"></a><span class="lineno"> 2487</span> <span class="comment"> * (sum_col[k] * A_OFFSET) +</span></div><div class="line"><a name="l02488"></a><span class="lineno"> 2488</span> <span class="comment"> * (sum_row[i] * B_OFFSET) +</span></div><div class="line"><a name="l02489"></a><span class="lineno"> 2489</span> <span class="comment"> * (K_OFFSET)</span></div><div class="line"><a name="l02490"></a><span class="lineno"> 2490</span> <span class="comment"> *</span></div><div class="line"><a name="l02491"></a><span class="lineno"> 2491</span> <span class="comment"> * This result is quantized down to uint8 using the output stage. The output stage computes the following operations:</span></div><div class="line"><a name="l02492"></a><span class="lineno"> 2492</span> <span class="comment"> *</span></div><div class="line"><a name="l02493"></a><span class="lineno"> 2493</span> <span class="comment"> * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier</span></div><div class="line"><a name="l02494"></a><span class="lineno"> 2494</span> <span class="comment"> * -# Add bias to final result if bias tensor is not a nullptr</span></div><div class="line"><a name="l02495"></a><span class="lineno"> 2495</span> <span class="comment"> * -# Round to nearest division by a power-of-two using result_shift</span></div><div class="line"><a name="l02496"></a><span class="lineno"> 2496</span> <span class="comment"> * -# Add offset to each result</span></div><div class="line"><a name="l02497"></a><span class="lineno"> 2497</span> <span class="comment"> * -# Clamp the value between the specified min and max bounds</span></div><div class="line"><a name="l02498"></a><span class="lineno"> 2498</span> <span class="comment"> * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8.</span></div><div class="line"><a name="l02499"></a><span class="lineno"> 2499</span> <span class="comment"> *</span></div><div class="line"><a name="l02500"></a><span class="lineno"> 2500</span> <span class="comment"> * @attention The offset, scalar scale factor and number of bits to shift right of output tensor must be passed at compile time using -DRESULT_OFFSET, -RESULT_MULT_INT and -DRESULT_SHIFT</span></div><div class="line"><a name="l02501"></a><span class="lineno"> 2501</span> <span class="comment"> *</span></div><div class="line"><a name="l02502"></a><span class="lineno"> 2502</span> <span class="comment"> * @note In case the addition of int32 biases is required, -DADD_BIAS should be passed at compile time</span></div><div class="line"><a name="l02503"></a><span class="lineno"> 2503</span> <span class="comment"> * @note In case the clamping of the result is required, the min and max bounds can be passed at compile time using -DMIN_BOUND and -DMAX_BOUND.</span></div><div class="line"><a name="l02504"></a><span class="lineno"> 2504</span> <span class="comment"> * These values can be used to implement "rectified linear unit" activation functions</span></div><div class="line"><a name="l02505"></a><span class="lineno"> 2505</span> <span class="comment"> *</span></div><div class="line"><a name="l02506"></a><span class="lineno"> 2506</span> <span class="comment"> * @param[in] mm_result_ptr Pointer to the source tensor. Supported data type: S32</span></div><div class="line"><a name="l02507"></a><span class="lineno"> 2507</span> <span class="comment"> * @param[in] mm_result_stride_x Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02508"></a><span class="lineno"> 2508</span> <span class="comment"> * @param[in] mm_result_step_x mm_result_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02509"></a><span class="lineno"> 2509</span> <span class="comment"> * @param[in] mm_result_stride_y Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02510"></a><span class="lineno"> 2510</span> <span class="comment"> * @param[in] mm_result_step_y mm_result_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02511"></a><span class="lineno"> 2511</span> <span class="comment"> * @param[in] mm_result_stride_z Stride of the source tensor in Z dimension (in bytes)</span></div><div class="line"><a name="l02512"></a><span class="lineno"> 2512</span> <span class="comment"> * @param[in] mm_result_step_z mm_result_stride_z * number of elements along Z processed per workitem(in bytes)</span></div><div class="line"><a name="l02513"></a><span class="lineno"> 2513</span> <span class="comment"> * @param[in] mm_result_offset_first_element_in_bytes The offset of the first element in the source tensor</span></div><div class="line"><a name="l02514"></a><span class="lineno"> 2514</span> <span class="comment"> * @param[in] sum_col_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02515"></a><span class="lineno"> 2515</span> <span class="comment"> * @param[in] sum_col_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02516"></a><span class="lineno"> 2516</span> <span class="comment"> * @param[in] sum_col_step_x (Optional) sum_col_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02517"></a><span class="lineno"> 2517</span> <span class="comment"> * @param[in] sum_col_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02518"></a><span class="lineno"> 2518</span> <span class="comment"> * @param[in] sum_col_step_y (Optional) sum_col_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02519"></a><span class="lineno"> 2519</span> <span class="comment"> * @param[in] sum_col_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02520"></a><span class="lineno"> 2520</span> <span class="comment"> * @param[in] sum_row_ptr (Optional) Pointer to the source tensor. Supported data type: same as @p mm_result_ptr</span></div><div class="line"><a name="l02521"></a><span class="lineno"> 2521</span> <span class="comment"> * @param[in] sum_row_stride_x (Optional) Stride of the source tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02522"></a><span class="lineno"> 2522</span> <span class="comment"> * @param[in] sum_row_step_x (Optional) sum_row_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02523"></a><span class="lineno"> 2523</span> <span class="comment"> * @param[in] sum_row_stride_y (Optional) Stride of the source tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02524"></a><span class="lineno"> 2524</span> <span class="comment"> * @param[in] sum_row_step_y (Optional) sum_row_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02525"></a><span class="lineno"> 2525</span> <span class="comment"> * @param[in] sum_row_offset_first_element_in_bytes (Optional) The offset of the first element in the source tensor</span></div><div class="line"><a name="l02526"></a><span class="lineno"> 2526</span> <span class="comment"> * @param[in] biases_ptr (Optional) Pointer to the biases tensor. Supported data type: same as @p src_ptr</span></div><div class="line"><a name="l02527"></a><span class="lineno"> 2527</span> <span class="comment"> * @param[in] biases_stride_x (Optional) Stride of the biases tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02528"></a><span class="lineno"> 2528</span> <span class="comment"> * @param[in] biases_step_x (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02529"></a><span class="lineno"> 2529</span> <span class="comment"> * @param[in] biases_offset_first_element_in_bytes (Optional) The offset of the first element in the biases tensor</span></div><div class="line"><a name="l02530"></a><span class="lineno"> 2530</span> <span class="comment"> * @param[out] dst_ptr Pointer to the destination tensor Supported data type: QASYMM8</span></div><div class="line"><a name="l02531"></a><span class="lineno"> 2531</span> <span class="comment"> * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)</span></div><div class="line"><a name="l02532"></a><span class="lineno"> 2532</span> <span class="comment"> * @param[in] dst_step_x dst_gx_stride_x * number of elements along X processed per workitem(in bytes)</span></div><div class="line"><a name="l02533"></a><span class="lineno"> 2533</span> <span class="comment"> * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)</span></div><div class="line"><a name="l02534"></a><span class="lineno"> 2534</span> <span class="comment"> * @param[in] dst_step_y dst_gx_stride_y * number of elements along Y processed per workitem(in bytes)</span></div><div class="line"><a name="l02535"></a><span class="lineno"> 2535</span> <span class="comment"> * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes)</span></div><div class="line"><a name="l02536"></a><span class="lineno"> 2536</span> <span class="comment"> * @param[in] dst_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)</span></div><div class="line"><a name="l02537"></a><span class="lineno"> 2537</span> <span class="comment"> * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor</span></div><div class="line"><a name="l02538"></a><span class="lineno"> 2538</span> <span class="comment"> */</span></div><div class="line"><a name="l02539"></a><span class="lineno"> 2539</span> __kernel <span class="keywordtype">void</span> gemmlowp_offset_contribution_quantize_down_fixedpoint(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(mm_result)</div><div class="line"><a name="l02540"></a><span class="lineno"> 2540</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02541"></a><span class="lineno"> 2541</span>  ,</div><div class="line"><a name="l02542"></a><span class="lineno"> 2542</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_col)</div><div class="line"><a name="l02543"></a><span class="lineno"> 2543</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02544"></a><span class="lineno"> 2544</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02545"></a><span class="lineno"> 2545</span>  ,</div><div class="line"><a name="l02546"></a><span class="lineno"> 2546</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a22f42fcf2077d951271df83b55c1a71a">IMAGE_DECLARATION</a>(sum_row)</div><div class="line"><a name="l02547"></a><span class="lineno"> 2547</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02548"></a><span class="lineno"> 2548</span>  ,</div><div class="line"><a name="l02549"></a><span class="lineno"> 2549</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02550"></a><span class="lineno"> 2550</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l02551"></a><span class="lineno"> 2551</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02552"></a><span class="lineno"> 2552</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l02553"></a><span class="lineno"> 2553</span> {</div><div class="line"><a name="l02554"></a><span class="lineno"> 2554</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02555"></a><span class="lineno"> 2555</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02556"></a><span class="lineno"> 2556</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02557"></a><span class="lineno"> 2557</span> </div><div class="line"><a name="l02558"></a><span class="lineno"> 2558</span>  <span class="comment">// Compute offset contribution</span></div><div class="line"><a name="l02559"></a><span class="lineno"> 2559</span>  int4 offset_term_s32 = offset_contribution(</div><div class="line"><a name="l02560"></a><span class="lineno"> 2560</span>  x, y, z</div><div class="line"><a name="l02561"></a><span class="lineno"> 2561</span> #<span class="keywordflow">if</span> defined(A_OFFSET)</div><div class="line"><a name="l02562"></a><span class="lineno"> 2562</span>  ,</div><div class="line"><a name="l02563"></a><span class="lineno"> 2563</span>  sum_col_ptr,</div><div class="line"><a name="l02564"></a><span class="lineno"> 2564</span>  sum_col_stride_x,</div><div class="line"><a name="l02565"></a><span class="lineno"> 2565</span>  sum_col_step_x,</div><div class="line"><a name="l02566"></a><span class="lineno"> 2566</span>  sum_col_stride_y,</div><div class="line"><a name="l02567"></a><span class="lineno"> 2567</span>  sum_col_step_y,</div><div class="line"><a name="l02568"></a><span class="lineno"> 2568</span>  sum_col_offset_first_element_in_bytes</div><div class="line"><a name="l02569"></a><span class="lineno"> 2569</span> #endif <span class="comment">// defined(A_OFFSET)</span></div><div class="line"><a name="l02570"></a><span class="lineno"> 2570</span> #<span class="keywordflow">if</span> defined(B_OFFSET)</div><div class="line"><a name="l02571"></a><span class="lineno"> 2571</span>  ,</div><div class="line"><a name="l02572"></a><span class="lineno"> 2572</span>  sum_row_ptr,</div><div class="line"><a name="l02573"></a><span class="lineno"> 2573</span>  sum_row_stride_x,</div><div class="line"><a name="l02574"></a><span class="lineno"> 2574</span>  sum_row_step_x,</div><div class="line"><a name="l02575"></a><span class="lineno"> 2575</span>  sum_row_stride_y,</div><div class="line"><a name="l02576"></a><span class="lineno"> 2576</span>  sum_row_step_y,</div><div class="line"><a name="l02577"></a><span class="lineno"> 2577</span>  sum_row_offset_first_element_in_bytes</div><div class="line"><a name="l02578"></a><span class="lineno"> 2578</span> #endif <span class="comment">// defined(B_OFFSET)</span></div><div class="line"><a name="l02579"></a><span class="lineno"> 2579</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02580"></a><span class="lineno"> 2580</span>  ,</div><div class="line"><a name="l02581"></a><span class="lineno"> 2581</span>  biases_ptr,</div><div class="line"><a name="l02582"></a><span class="lineno"> 2582</span>  biases_stride_x,</div><div class="line"><a name="l02583"></a><span class="lineno"> 2583</span>  biases_step_x,</div><div class="line"><a name="l02584"></a><span class="lineno"> 2584</span>  biases_offset_first_element_in_bytes</div><div class="line"><a name="l02585"></a><span class="lineno"> 2585</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02586"></a><span class="lineno"> 2586</span>  );</div><div class="line"><a name="l02587"></a><span class="lineno"> 2587</span> </div><div class="line"><a name="l02588"></a><span class="lineno"> 2588</span>  __global uchar *mm_result_addr = mm_result_ptr + mm_result_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * mm_result_stride_y + z * mm_result_stride_z;</div><div class="line"><a name="l02589"></a><span class="lineno"> 2589</span> </div><div class="line"><a name="l02590"></a><span class="lineno"> 2590</span>  __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;</div><div class="line"><a name="l02591"></a><span class="lineno"> 2591</span> </div><div class="line"><a name="l02592"></a><span class="lineno"> 2592</span>  int4 in_s32 = vload4(0, (__global <span class="keywordtype">int</span> *)mm_result_addr);</div><div class="line"><a name="l02593"></a><span class="lineno"> 2593</span> </div><div class="line"><a name="l02594"></a><span class="lineno"> 2594</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02595"></a><span class="lineno"> 2595</span>  in_s32 += offset_term_s32;</div><div class="line"><a name="l02596"></a><span class="lineno"> 2596</span> </div><div class="line"><a name="l02597"></a><span class="lineno"> 2597</span>  <span class="comment">// -------------- OUTPUT STAGE</span></div><div class="line"><a name="l02598"></a><span class="lineno"> 2598</span> </div><div class="line"><a name="l02599"></a><span class="lineno"> 2599</span>  <span class="comment">// Multiply by result_mult_int and shift</span></div><div class="line"><a name="l02600"></a><span class="lineno"> 2600</span>  in_s32 = <a class="code" href="helpers__asymm_8h.xhtml#a86de0ffca367bfcf27a8ae5dd1fdef2d">ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE</a>(in_s32, RESULT_MULTIPLIER, RESULT_SHIFT, 4);</div><div class="line"><a name="l02601"></a><span class="lineno"> 2601</span> </div><div class="line"><a name="l02602"></a><span class="lineno"> 2602</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02603"></a><span class="lineno"> 2603</span>  in_s32 += (int4)RESULT_OFFSET;</div><div class="line"><a name="l02604"></a><span class="lineno"> 2604</span> </div><div class="line"><a name="l02605"></a><span class="lineno"> 2605</span>  uchar4 res = convert_uchar4_sat(in_s32);</div><div class="line"><a name="l02606"></a><span class="lineno"> 2606</span> </div><div class="line"><a name="l02607"></a><span class="lineno"> 2607</span> <span class="preprocessor">#if defined(MIN_BOUND)</span></div><div class="line"><a name="l02608"></a><span class="lineno"> 2608</span>  res = max(res, (uchar4)MIN_BOUND);</div><div class="line"><a name="l02609"></a><span class="lineno"> 2609</span> <span class="preprocessor">#endif // defined(MIN_BOUND)</span></div><div class="line"><a name="l02610"></a><span class="lineno"> 2610</span> <span class="preprocessor">#if defined(MAX_BOUND)</span></div><div class="line"><a name="l02611"></a><span class="lineno"> 2611</span>  res = min(res, (uchar4)MAX_BOUND);</div><div class="line"><a name="l02612"></a><span class="lineno"> 2612</span> <span class="preprocessor">#endif // defined(MAX_BOUND)</span></div><div class="line"><a name="l02613"></a><span class="lineno"> 2613</span> </div><div class="line"><a name="l02614"></a><span class="lineno"> 2614</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l02615"></a><span class="lineno"> 2615</span>  vstore4(res, 0, dst_addr);</div><div class="line"><a name="l02616"></a><span class="lineno"> 2616</span> }</div><div class="line"><a name="l02617"></a><span class="lineno"> 2617</span> <span class="preprocessor">#endif // defined(K_OFFSET) && defined(RESULT_OFFSET) && defined(RESULT_MULTIPLIER) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02618"></a><span class="lineno"> 2618</span> <span class="preprocessor">#endif // defined(K_OFFSET)</span></div><div class="line"><a name="l02619"></a><span class="lineno"> 2619</span> </div><div class="line"><a name="l02620"></a><span class="lineno"> 2620</span> <span class="preprocessor">#if defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02621"></a><span class="lineno"> 2621</span> </div><div class="line"><a name="l02660"></a><span class="lineno"> 2660</span> __kernel <span class="keywordtype">void</span> gemmlowp_output_stage_quantize_down(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l02661"></a><span class="lineno"> 2661</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02662"></a><span class="lineno"> 2662</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l02663"></a><span class="lineno"> 2663</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02664"></a><span class="lineno"> 2664</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l02665"></a><span class="lineno"> 2665</span> {</div><div class="line"><a name="l02666"></a><span class="lineno"> 2666</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02667"></a><span class="lineno"> 2667</span>  <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02668"></a><span class="lineno"> 2668</span>  <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02669"></a><span class="lineno"> 2669</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02670"></a><span class="lineno"> 2670</span> </div><div class="line"><a name="l02671"></a><span class="lineno"> 2671</span>  __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * src_stride_y + z * src_stride_z;</div><div class="line"><a name="l02672"></a><span class="lineno"> 2672</span> </div><div class="line"><a name="l02673"></a><span class="lineno"> 2673</span>  __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;</div><div class="line"><a name="l02674"></a><span class="lineno"> 2674</span> </div><div class="line"><a name="l02675"></a><span class="lineno"> 2675</span>  int4 input_values = vload4(0, (__global <span class="keywordtype">int</span> *)src_addr);</div><div class="line"><a name="l02676"></a><span class="lineno"> 2676</span> </div><div class="line"><a name="l02677"></a><span class="lineno"> 2677</span> <span class="preprocessor">#if defined(ADD_BIAS)</span></div><div class="line"><a name="l02678"></a><span class="lineno"> 2678</span>  <span class="comment">// Add bias</span></div><div class="line"><a name="l02679"></a><span class="lineno"> 2679</span>  __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02680"></a><span class="lineno"> 2680</span> </div><div class="line"><a name="l02681"></a><span class="lineno"> 2681</span>  int4 biases_values = vload4(0, (__global <span class="keywordtype">int</span> *)bias_addr);</div><div class="line"><a name="l02682"></a><span class="lineno"> 2682</span>  input_values += (int4)biases_values;</div><div class="line"><a name="l02683"></a><span class="lineno"> 2683</span> <span class="preprocessor">#endif // defined(ADD_BIAS)</span></div><div class="line"><a name="l02684"></a><span class="lineno"> 2684</span> </div><div class="line"><a name="l02685"></a><span class="lineno"> 2685</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02686"></a><span class="lineno"> 2686</span>  input_values += (int4)RESULT_OFFSET;</div><div class="line"><a name="l02687"></a><span class="lineno"> 2687</span> </div><div class="line"><a name="l02688"></a><span class="lineno"> 2688</span>  <span class="comment">// Multiply by result_mult_int and shift</span></div><div class="line"><a name="l02689"></a><span class="lineno"> 2689</span>  input_values *= RESULT_MULT_INT;</div><div class="line"><a name="l02690"></a><span class="lineno"> 2690</span> </div><div class="line"><a name="l02691"></a><span class="lineno"> 2691</span>  input_values >>= RESULT_SHIFT;</div><div class="line"><a name="l02692"></a><span class="lineno"> 2692</span> </div><div class="line"><a name="l02693"></a><span class="lineno"> 2693</span>  uchar4 res = convert_uchar4_sat(input_values);</div><div class="line"><a name="l02694"></a><span class="lineno"> 2694</span> </div><div class="line"><a name="l02695"></a><span class="lineno"> 2695</span> <span class="preprocessor">#if defined(MIN_BOUND)</span></div><div class="line"><a name="l02696"></a><span class="lineno"> 2696</span>  res = max(res, (uchar4)MIN_BOUND);</div><div class="line"><a name="l02697"></a><span class="lineno"> 2697</span> <span class="preprocessor">#endif // defined(MIN_BOUND)</span></div><div class="line"><a name="l02698"></a><span class="lineno"> 2698</span> <span class="preprocessor">#if defined(MAX_BOUND)</span></div><div class="line"><a name="l02699"></a><span class="lineno"> 2699</span>  res = min(res, (uchar4)MAX_BOUND);</div><div class="line"><a name="l02700"></a><span class="lineno"> 2700</span> <span class="preprocessor">#endif // defined(MAX_BOUND)</span></div><div class="line"><a name="l02701"></a><span class="lineno"> 2701</span> </div><div class="line"><a name="l02702"></a><span class="lineno"> 2702</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l02703"></a><span class="lineno"> 2703</span>  vstore4(res, 0, dst_addr);</div><div class="line"><a name="l02704"></a><span class="lineno"> 2704</span> }</div><div class="line"><a name="l02705"></a><span class="lineno"> 2705</span> <span class="preprocessor">#endif // defined(RESULT_OFFSET) && defined(RESULT_MULT_INT) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02706"></a><span class="lineno"> 2706</span> </div><div class="line"><a name="l02707"></a><span class="lineno"> 2707</span> <span class="preprocessor">#if defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02708"></a><span class="lineno"> 2708</span> </div><div class="line"><a name="l02747"></a><span class="lineno"> 2747</span> __kernel <span class="keywordtype">void</span> gemmlowp_output_stage_quantize_down_fixedpoint(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l02748"></a><span class="lineno"> 2748</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02749"></a><span class="lineno"> 2749</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l02750"></a><span class="lineno"> 2750</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02751"></a><span class="lineno"> 2751</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l02752"></a><span class="lineno"> 2752</span> {</div><div class="line"><a name="l02753"></a><span class="lineno"> 2753</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02754"></a><span class="lineno"> 2754</span>  <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02755"></a><span class="lineno"> 2755</span>  <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02756"></a><span class="lineno"> 2756</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02757"></a><span class="lineno"> 2757</span> </div><div class="line"><a name="l02758"></a><span class="lineno"> 2758</span>  __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * src_stride_y + z * src_stride_z;</div><div class="line"><a name="l02759"></a><span class="lineno"> 2759</span> </div><div class="line"><a name="l02760"></a><span class="lineno"> 2760</span>  __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;</div><div class="line"><a name="l02761"></a><span class="lineno"> 2761</span> </div><div class="line"><a name="l02762"></a><span class="lineno"> 2762</span>  int4 input_values = vload4(0, (__global <span class="keywordtype">int</span> *)src_addr);</div><div class="line"><a name="l02763"></a><span class="lineno"> 2763</span> </div><div class="line"><a name="l02764"></a><span class="lineno"> 2764</span> <span class="preprocessor">#if defined(ADD_BIAS)</span></div><div class="line"><a name="l02765"></a><span class="lineno"> 2765</span>  <span class="comment">// Add bias</span></div><div class="line"><a name="l02766"></a><span class="lineno"> 2766</span>  __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02767"></a><span class="lineno"> 2767</span> </div><div class="line"><a name="l02768"></a><span class="lineno"> 2768</span>  int4 biases_values = vload4(0, (__global <span class="keywordtype">int</span> *)bias_addr);</div><div class="line"><a name="l02769"></a><span class="lineno"> 2769</span>  input_values += (int4)biases_values;</div><div class="line"><a name="l02770"></a><span class="lineno"> 2770</span> <span class="preprocessor">#endif // defined(ADD_BIAS)</span></div><div class="line"><a name="l02771"></a><span class="lineno"> 2771</span> </div><div class="line"><a name="l02772"></a><span class="lineno"> 2772</span>  <span class="comment">// Multiply by result_mult_int and shift</span></div><div class="line"><a name="l02773"></a><span class="lineno"> 2773</span>  input_values = <a class="code" href="helpers__asymm_8h.xhtml#a86de0ffca367bfcf27a8ae5dd1fdef2d">ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE</a>(input_values, RESULT_FIXEDPOINT_MULTIPLIER, RESULT_SHIFT, 4);</div><div class="line"><a name="l02774"></a><span class="lineno"> 2774</span> </div><div class="line"><a name="l02775"></a><span class="lineno"> 2775</span>  <span class="comment">// Add the offset terms to GEMM's result</span></div><div class="line"><a name="l02776"></a><span class="lineno"> 2776</span>  input_values += (int4)RESULT_OFFSET_AFTER_SHIFT;</div><div class="line"><a name="l02777"></a><span class="lineno"> 2777</span> </div><div class="line"><a name="l02778"></a><span class="lineno"> 2778</span>  uchar4 res = convert_uchar4_sat(input_values);</div><div class="line"><a name="l02779"></a><span class="lineno"> 2779</span> </div><div class="line"><a name="l02780"></a><span class="lineno"> 2780</span> <span class="preprocessor">#if defined(MIN_BOUND)</span></div><div class="line"><a name="l02781"></a><span class="lineno"> 2781</span>  res = max(res, (uchar4)MIN_BOUND);</div><div class="line"><a name="l02782"></a><span class="lineno"> 2782</span> <span class="preprocessor">#endif // defined(MIN_BOUND)</span></div><div class="line"><a name="l02783"></a><span class="lineno"> 2783</span> <span class="preprocessor">#if defined(MAX_BOUND)</span></div><div class="line"><a name="l02784"></a><span class="lineno"> 2784</span>  res = min(res, (uchar4)MAX_BOUND);</div><div class="line"><a name="l02785"></a><span class="lineno"> 2785</span> <span class="preprocessor">#endif // defined(MAX_BOUND)</span></div><div class="line"><a name="l02786"></a><span class="lineno"> 2786</span> </div><div class="line"><a name="l02787"></a><span class="lineno"> 2787</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l02788"></a><span class="lineno"> 2788</span>  vstore4(res, 0, dst_addr);</div><div class="line"><a name="l02789"></a><span class="lineno"> 2789</span> }</div><div class="line"><a name="l02790"></a><span class="lineno"> 2790</span> <span class="preprocessor">#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)</span></div><div class="line"><a name="l02791"></a><span class="lineno"> 2791</span> </div><div class="line"><a name="l02792"></a><span class="lineno"> 2792</span> <span class="preprocessor">#if defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)</span></div><div class="line"><a name="l02793"></a><span class="lineno"> 2793</span> </div><div class="line"><a name="l02834"></a><span class="lineno"> 2834</span> __kernel <span class="keywordtype">void</span> gemmlowp_output_stage_quantize_down_float(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(src),</div><div class="line"><a name="l02835"></a><span class="lineno"> 2835</span> #<span class="keywordflow">if</span> defined(ADD_BIAS)</div><div class="line"><a name="l02836"></a><span class="lineno"> 2836</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a40a6eb9f2a7712f08d6bb8ff6c9e6ca7">VECTOR_DECLARATION</a>(biases),</div><div class="line"><a name="l02837"></a><span class="lineno"> 2837</span> #endif <span class="comment">// defined(ADD_BIAS)</span></div><div class="line"><a name="l02838"></a><span class="lineno"> 2838</span> #<span class="keywordflow">if</span> defined(DST_HEIGHT)</div><div class="line"><a name="l02839"></a><span class="lineno"> 2839</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a481bdc6d61b3df9dcdbdb244f0f97790">TENSOR4D_DECLARATION</a>(dst))</div><div class="line"><a name="l02840"></a><span class="lineno"> 2840</span> <span class="preprocessor">#else // defined(DST_HEIGHT)</span></div><div class="line"><a name="l02841"></a><span class="lineno"> 2841</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(dst))</div><div class="line"><a name="l02842"></a><span class="lineno"> 2842</span> #endif <span class="comment">// defined(DST_HEIGHT)</span></div><div class="line"><a name="l02843"></a><span class="lineno"> 2843</span> {</div><div class="line"><a name="l02844"></a><span class="lineno"> 2844</span>  <span class="comment">// Compute source and destination addresses</span></div><div class="line"><a name="l02845"></a><span class="lineno"> 2845</span>  <span class="keywordtype">int</span> x = get_global_id(0) * 4;</div><div class="line"><a name="l02846"></a><span class="lineno"> 2846</span>  <span class="keywordtype">int</span> y = get_global_id(1);</div><div class="line"><a name="l02847"></a><span class="lineno"> 2847</span>  <span class="keywordtype">int</span> z = get_global_id(2);</div><div class="line"><a name="l02848"></a><span class="lineno"> 2848</span> </div><div class="line"><a name="l02849"></a><span class="lineno"> 2849</span>  __global uchar *src_addr = src_ptr + src_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int) + y * src_stride_y + z * src_stride_z;</div><div class="line"><a name="l02850"></a><span class="lineno"> 2850</span> </div><div class="line"><a name="l02851"></a><span class="lineno"> 2851</span>  __global uchar *dst_addr = dst_ptr + dst_offset_first_element_in_bytes + x + y * dst_stride_y + z * dst_stride_z;</div><div class="line"><a name="l02852"></a><span class="lineno"> 2852</span> </div><div class="line"><a name="l02853"></a><span class="lineno"> 2853</span>  int4 input_values = vload4(0, (__global <span class="keywordtype">int</span> *)src_addr);</div><div class="line"><a name="l02854"></a><span class="lineno"> 2854</span> </div><div class="line"><a name="l02855"></a><span class="lineno"> 2855</span> <span class="preprocessor">#if defined(ADD_BIAS)</span></div><div class="line"><a name="l02856"></a><span class="lineno"> 2856</span>  <span class="comment">// Add bias</span></div><div class="line"><a name="l02857"></a><span class="lineno"> 2857</span>  __global uchar *bias_addr = biases_ptr + biases_offset_first_element_in_bytes + x * <span class="keyword">sizeof</span>(int);</div><div class="line"><a name="l02858"></a><span class="lineno"> 2858</span> </div><div class="line"><a name="l02859"></a><span class="lineno"> 2859</span>  int4 biases_values = vload4(0, (__global <span class="keywordtype">int</span> *)bias_addr);</div><div class="line"><a name="l02860"></a><span class="lineno"> 2860</span>  input_values += (int4)biases_values;</div><div class="line"><a name="l02861"></a><span class="lineno"> 2861</span> <span class="preprocessor">#endif // defined(ADD_BIAS)</span></div><div class="line"><a name="l02862"></a><span class="lineno"> 2862</span> </div><div class="line"><a name="l02863"></a><span class="lineno"> 2863</span>  <span class="comment">// Convert to float</span></div><div class="line"><a name="l02864"></a><span class="lineno"> 2864</span>  float16 input_values_f = convert_float4(input_values);</div><div class="line"><a name="l02865"></a><span class="lineno"> 2865</span>  input_values_f = <a class="code" href="namespacearm__compute.xhtml#aaae2b6b1c3f4404121346a4c27b22647">round</a>(input_values_f * (<span class="keywordtype">float</span>)REAL_MULTIPLIER + (<span class="keywordtype">float</span>)OUTPUT_OFFSET);</div><div class="line"><a name="l02866"></a><span class="lineno"> 2866</span> </div><div class="line"><a name="l02867"></a><span class="lineno"> 2867</span>  uchar4 res = convert_uchar4_sat(input_values_f);</div><div class="line"><a name="l02868"></a><span class="lineno"> 2868</span> </div><div class="line"><a name="l02869"></a><span class="lineno"> 2869</span> <span class="preprocessor">#if defined(MIN_BOUND)</span></div><div class="line"><a name="l02870"></a><span class="lineno"> 2870</span>  res = max(res, (uchar4)MIN_BOUND);</div><div class="line"><a name="l02871"></a><span class="lineno"> 2871</span> <span class="preprocessor">#endif // defined(MIN_BOUND)</span></div><div class="line"><a name="l02872"></a><span class="lineno"> 2872</span> <span class="preprocessor">#if defined(MAX_BOUND)</span></div><div class="line"><a name="l02873"></a><span class="lineno"> 2873</span>  res = min(res, (uchar4)MAX_BOUND);</div><div class="line"><a name="l02874"></a><span class="lineno"> 2874</span> <span class="preprocessor">#endif // defined(MAX_BOUND)</span></div><div class="line"><a name="l02875"></a><span class="lineno"> 2875</span> </div><div class="line"><a name="l02876"></a><span class="lineno"> 2876</span>  <span class="comment">// Store the result</span></div><div class="line"><a name="l02877"></a><span class="lineno"> 2877</span>  vstore4(res, 0, dst_addr);</div><div class="line"><a name="l02878"></a><span class="lineno"> 2878</span> }</div><div class="line"><a name="l02879"></a><span class="lineno"> 2879</span> <span class="preprocessor">#endif // defined(REAL_MULTIPLIER) && defined(OUTPUT_OFFSET)</span></div><div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_aa8d95ba04fc73845abc6045952cae5be"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aa8d95ba04fc73845abc6045952cae5be">CONVERT</a></div><div class="ttdeci">#define CONVERT(x, type)</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00060">helpers.h:60</a></div></div> |