| Jenkins | b9abeae | 2018-11-22 11:58:08 +0000 | [diff] [blame^] | 94 | <a href="direct__convolution5x5_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) 2016-2018 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> </div><div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#undef CONVERT_SAT</span></div><div class="line"><a name="l00027"></a><span class="lineno"> 27</span> </div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#if defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)</span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> </div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="preprocessor">#if STRIDE_X == 1</span></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr)</span></div><div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#elif STRIDE_X == 2 </span><span class="comment">/* STRIDE_X == 1 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00033"></a><span class="lineno"> 33</span> <span class="preprocessor">#define CONVOLUTION1x5(acc, src_row_ptr, weights_row_ptr) CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr)</span></div><div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="preprocessor">#else </span><span class="comment">/* STRIDE_X not equals 1 or 2 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00035"></a><span class="lineno"> 35</span> <span class="preprocessor">#error "STRIDE_X larger than 2 is not supported"</span></div><div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="preprocessor">#endif </span><span class="comment">/* STRIDE_X == 2 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> </div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="preprocessor">#define CONVOLUTION1x5_STRIDE1(acc, src_row_ptr, weights_row_ptr) \</span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> <span class="preprocessor"> weights_values0 = vload4(0, weights_row_ptr); \</span></div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="preprocessor"> DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \</span></div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 8) \</span></div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span> <span class="preprocessor"> src0 = vload8(0, src_row_ptr); \</span></div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> <span class="preprocessor"> src1 = vload4(0, src_row_ptr + 8); \</span></div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span> <span class="preprocessor"> \</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> <span class="preprocessor"> acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \</span></div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \</span></div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \</span></div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s45, src0.s67, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \</span></div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span> <span class="preprocessor"> })</span></div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span> <span class="preprocessor">#define CONVOLUTION1x5_STRIDE2(acc, src_row_ptr, weights_row_ptr) \</span></div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> <span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span> <span class="preprocessor"> weights_values0 = vload4(0, weights_row_ptr); \</span></div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span> <span class="preprocessor"> DATA_TYPE weights_value1 = *(weights_row_ptr + 4); \</span></div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 16) \</span></div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span> <span class="preprocessor"> src0 = vload16(0, src_row_ptr); \</span></div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span> <span class="preprocessor"> src1 = vload4(0, src_row_ptr + 16); \</span></div><div class="line"><a name="l00064"></a><span class="lineno"> 64</span> <span class="preprocessor"> acc += src0.even * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \</span></div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \</span></div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \</span></div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span> <span class="preprocessor"> \</span></div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \</span></div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \</span></div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> <span class="preprocessor"> })</span></div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span> </div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span> <span class="preprocessor">#if defined(DATA_LAYOUT_NHWC)</span></div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span> </div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span> <span class="preprocessor">#define PTR_TO_VALUE(PTR, DATA_TYPE) *((__global DATA_TYPE *)(PTR))</span></div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span> </div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span> <span class="preprocessor">#if STRIDE_X == 1</span></div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span> <span class="preprocessor">#define CONVOLUTION1x5_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x5_STRIDE1_NHWC(acc, row_ptr, weights_ptr)</span></div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> <span class="preprocessor">#elif STRIDE_X == 2 </span><span class="comment">/* STRIDE_X == 1 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span> <span class="preprocessor">#define CONVOLUTION1x5_NHWC(acc, row_ptr, weights_ptr) CONVOLUTION1x5_STRIDE2_NHWC(acc, row_ptr, weights_ptr)</span></div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span> <span class="preprocessor">#else </span><span class="comment">/* STRIDE_X not equals 1 or 2 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span> <span class="preprocessor">#error "STRIDE_X larger than 2 is not supported"</span></div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span> <span class="preprocessor">#endif </span><span class="comment">/* STRIDE_X == 2 */</span><span class="preprocessor"></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="preprocessor">#define CONVOLUTION1x5_STRIDE1_NHWC(acc, row_ptr, weights_ptr) \</span></div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span> <span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 8) \</span></div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span> <span class="preprocessor"> src0 = (VEC_DATA_TYPE(DATA_TYPE, 8))( \</span></div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span> <span class="preprocessor"> src1 = (VEC_DATA_TYPE(DATA_TYPE, 4))( \</span></div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00096"></a><span class="lineno"> 96</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00097"></a><span class="lineno"> 97</span> <span class="preprocessor"> weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 4))( \</span></div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span> <span class="preprocessor"> PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span> <span class="preprocessor"> PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span> <span class="preprocessor"> DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE); \</span></div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span> <span class="preprocessor"> acc += src0 * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \</span></div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1234, src0.s567, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \</span></div><div class="line"><a name="l00103"></a><span class="lineno"> 103</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s234, src0.s567, src1.s01) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \</span></div><div class="line"><a name="l00104"></a><span class="lineno"> 104</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s345, src0.s67, src1.s012) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \</span></div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s45, src0.s67, src1.s0123) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \</span></div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span> <span class="preprocessor"> })</span></div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span> </div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span> <span class="preprocessor">#define CONVOLUTION1x5_STRIDE2_NHWC(acc, row_ptr, weights_ptr) \</span></div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span> <span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 16) \</span></div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span> <span class="preprocessor"> src0 = (VEC_DATA_TYPE(DATA_TYPE, 16))( \</span></div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 0 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 1 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 2 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 3 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 4 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 5 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 6 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 7 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 8 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 9 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 10 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 11 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 12 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 13 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 14 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 15 * src_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> <span class="preprocessor"> src1 = (VEC_DATA_TYPE(DATA_TYPE, 4))( \</span></div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 16 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 17 * src_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span> <span class="preprocessor"> PTR_TO_VALUE(row_ptr + 18 * src_stride_y, DATA_TYPE), PTR_TO_VALUE(row_ptr + 19 * src_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span> <span class="preprocessor"> VEC_DATA_TYPE(DATA_TYPE, 4) \</span></div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span> <span class="preprocessor"> weights_values0 = (VEC_DATA_TYPE(DATA_TYPE, 4))( \</span></div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span> <span class="preprocessor"> PTR_TO_VALUE(weights_ptr + 0 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 1 * weights_stride_y, DATA_TYPE), \</span></div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span> <span class="preprocessor"> PTR_TO_VALUE(weights_ptr + 2 * weights_stride_y, DATA_TYPE), PTR_TO_VALUE(weights_ptr + 3 * weights_stride_y, DATA_TYPE)); \</span></div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span> <span class="preprocessor"> DATA_TYPE weights_value1 = PTR_TO_VALUE(weights_ptr + 4 * weights_stride_y, DATA_TYPE); \</span></div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span> <span class="preprocessor"> acc += src0.s02468ACE * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s0; \</span></div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s1357, src0.s9BDF) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s1; \</span></div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s2468, src0.sACE, src1.s0) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s2; \</span></div><div class="line"><a name="l00132"></a><span class="lineno"> 132</span> <span class="preprocessor"> \</span></div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s3579, src0.sBDF, src1.s1) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_values0.s3; \</span></div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span> <span class="preprocessor"> acc += (VEC_DATA_TYPE(DATA_TYPE, 8))(src0.s468a, src0.sCE, src1.s02) * (VEC_DATA_TYPE(DATA_TYPE, 8))weights_value1; \</span></div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span> <span class="preprocessor"> })</span></div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span> </div><div class="line"><a name="l00173"></a><span class="lineno"> 173</span> __kernel <span class="keywordtype">void</span> direct_convolution5x5_nhwc(</div><div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  <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="l00175"></a><span class="lineno"> 175</span>  <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#adbf67dcee294e673cf796f1ed8aeb6a4">dst</a>),</div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights),</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span> #ifdef HAS_BIAS</div><div class="line"><a name="l00178"></a><span class="lineno"> 178</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="l00179"></a><span class="lineno"> 179</span> #endif <span class="comment">/* defined(HAS_BIAS) */</span></div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weights_stride_w)</div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span> {</div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <a class="code" href="struct_image.xhtml">Image</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#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</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#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span> </div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 8)</div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  values0 = 0;</div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span> </div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  const <span class="keywordtype">int</span> id0 = get_global_id(0);</div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  const <span class="keywordtype">int</span> id1 = get_global_id(1);</div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  const <span class="keywordtype">int</span> id2 = get_global_id(2);</div><div class="line"><a name="l00192"></a><span class="lineno"> 192</span> </div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  __global uchar *weights_addr = (__global uchar *)<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a2101b2fe0193ce227ae4e0945e321d85">tensor3D_offset</a>(&weights, 0, 0, 0);</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  __global uchar *src_addr = (__global uchar *)<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0) - src_stride_x * id0 + ((id2 * STRIDE_Y) - PAD_TOP) * (<span class="keywordtype">int</span>)src_stride_z;</div><div class="line"><a name="l00195"></a><span class="lineno"> 195</span> </div><div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  weights_addr += id0 * weights_stride_w;</div><div class="line"><a name="l00197"></a><span class="lineno"> 197</span> </div><div class="line"><a name="l00198"></a><span class="lineno"> 198</span> <span class="preprocessor">#if(PAD_TOP == 1)</span></div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> coordy = id2 - PAD_TOP;</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  <span class="keywordflow">for</span>(<span class="keyword">volatile</span> <span class="keywordtype">int</span> d = 0; d < WEIGHTS_DEPTH; ++d)</div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  {</div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  <span class="keywordflow">if</span>(coordy < 0) <span class="comment">// special case Z = -1 doesn't exists</span></div><div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  {</div><div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  <span class="comment">//skip first row and load the two next ones</span></div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  }</div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(coordy == (DST_HEIGHT - PAD_TOP - 1))</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  {</div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="comment">// special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the</span></div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  <span class="comment">// Z axis has no padding at all.</span></div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  }</div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  {</div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  }</div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  src_addr += src_stride_x;</div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  weights_addr += weights_stride_x;</div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  }</div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span> <span class="preprocessor">#elif(PAD_TOP == 2)</span></div><div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> coordy = id2 * STRIDE_Y;</div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  <span class="keywordflow">for</span>(<span class="keyword">volatile</span> <span class="keywordtype">int</span> d = 0; d < WEIGHTS_DEPTH; ++d)</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  {</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  <span class="keywordflow">if</span>(coordy == 0) <span class="comment">// special case Z = -2 doesn't exists</span></div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  {</div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  <span class="comment">//skip first row and load the two next ones</span></div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  }</div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(coordy == 1) <span class="comment">// special case Z = -1 doesn't exists</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>  <span class="comment">//skip first row and load the two next ones</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  }</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  <span class="keywordflow">else</span> <span class="keywordflow">if</span>(coordy == (SRC_HEIGHT - 1))</div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  {</div><div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  <span class="comment">// special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the</span></div><div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  <span class="comment">// Z axis has no padding at all.</span></div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</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="keywordflow">else</span> <span class="keywordflow">if</span>(coordy == (SRC_HEIGHT - 2))</div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  {</div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  <span class="comment">// special case when computing the last row of the output we must read the last three rows from the input buffer (including padding) but the</span></div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="comment">// Z axis has no padding at all.</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  }</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <span class="keywordflow">else</span></div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  {</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  }</div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  src_addr += src_stride_x;</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  weights_addr += weights_stride_x;</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  }</div><div class="line"><a name="l00277"></a><span class="lineno"> 277</span> </div><div class="line"><a name="l00278"></a><span class="lineno"> 278</span> <span class="preprocessor">#else </span><span class="comment">/* PAD_TOP == 2 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  <span class="keywordflow">for</span>(<span class="keyword">volatile</span> <span class="keywordtype">int</span> d = 0; d < WEIGHTS_DEPTH; ++d)</div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  {</div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  CONVOLUTION1x5_NHWC(values0, src_addr, weights_addr);</div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 1 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 1 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 2 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 2 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 3 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 3 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span>  CONVOLUTION1x5_NHWC(values0, (src_addr + 4 * (<span class="keywordtype">int</span>)src_stride_z), (weights_addr + 4 * (<span class="keywordtype">int</span>)weights_stride_z));</div><div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  src_addr += src_stride_x;</div><div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  weights_addr += weights_stride_x;</div><div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  }</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span> <span class="preprocessor">#endif </span><span class="comment">/* PAD_TOP == 1 */</span><span class="preprocessor"></span></div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span> </div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span> <span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  values0 += (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 8)) * ((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&biases, id0)));</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span> <span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span> </div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y)) = values0.s0;</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y)) = values0.s1;</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 2 * dst_stride_y)) = values0.s2;</div><div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 3 * dst_stride_y)) = values0.s3;</div><div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 4 * dst_stride_y)) = values0.s4;</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 5 * dst_stride_y)) = values0.s5;</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 6 * dst_stride_y)) = values0.s6;</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span>  *((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 7 * dst_stride_y)) = values0.s7;</div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> }</div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> </div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> <span class="preprocessor">#endif // defined(DATA_LAYOUT_NHWC)</span></div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span> </div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span> __kernel <span class="keywordtype">void</span> direct_convolution5x5(</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <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="l00346"></a><span class="lineno"> 346</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="l00347"></a><span class="lineno"> 347</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights),</div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span> #ifdef HAS_BIAS</div><div class="line"><a name="l00349"></a><span class="lineno"> 349</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="l00350"></a><span class="lineno"> 350</span> #endif <span class="comment">/* defined(HAS_BIAS) */</span></div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weights_stride_w)</div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span> {</div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> weights = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a79e8e562daa6599317d2d1cd86ef1bf2">CONVERT_TO_TENSOR3D_STRUCT_NO_STEP</a>(weights);</div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l00356"></a><span class="lineno"> 356</span> </div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 8)</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  values0 = 0;</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span> </div><div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  __global uchar *weights_addr = (__global uchar *)<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a2101b2fe0193ce227ae4e0945e321d85">tensor3D_offset</a>(&weights, 0, 0, 0);</div><div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  __global uchar *src_addr = (__global uchar *)<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0);</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span> </div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  const <span class="keywordtype">int</span> kernel_index = get_global_id(2);</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  weights_addr += kernel_index * weights_stride_w;</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span> </div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  for(volatile <span class="keywordtype">int</span> d = 0; d < WEIGHTS_DEPTH; ++d)</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  {</div><div class="line"><a name="l00368"></a><span class="lineno"> 368</span>  CONVOLUTION1x5(values0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)src_addr, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)weights_addr);</div><div class="line"><a name="l00369"></a><span class="lineno"> 369</span>  CONVOLUTION1x5(values0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + 1 * src_stride_y), (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  CONVOLUTION1x5(values0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + 2 * src_stride_y), (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  CONVOLUTION1x5(values0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + 3 * src_stride_y), (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights_addr + 3 * weights_stride_y));</div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  CONVOLUTION1x5(values0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(src_addr + 4 * src_stride_y), (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(weights_addr + 4 * weights_stride_y));</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>  src_addr += src_stride_z;</div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span>  weights_addr += weights_stride_z;</div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  }</div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span> </div><div class="line"><a name="l00378"></a><span class="lineno"> 378</span> <span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span> </div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  values0 += (<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a36f754c05b6fddf6df0d8d0a74f8159f">VEC_DATA_TYPE</a>(<a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a>, 8)) * ((__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&biases, kernel_index)));</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span> <span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span> </div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  vstore8(values0, 0, (__global <a class="code" href="convolution3x3_8cl.xhtml#afb8c72ce35c4a1f4a2588d6573e54aa1">DATA_TYPE</a> *)dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a>);</div><div class="line"><a name="l00385"></a><span class="lineno"> 385</span> }</div><div class="line"><a name="l00386"></a><span class="lineno"> 386</span> <span class="preprocessor">#endif // defined(DATA_TYPE) && defined(STRIDE_X) && defined(WEIGHTS_DEPTH)</span></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> <span class="preprocessor">#if defined(WEIGHTS_DEPTH)</span></div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span> </div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span> <span class="preprocessor">#define CONVOLUTION1x5_BIFROST(acc, src0, weights_row00, weights_row01) \</span></div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span> <span class="preprocessor"> ({ \</span></div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span> <span class="preprocessor"> acc.s0 = mad(src0.s0, weights_row00.s0, acc.s0); \</span></div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span> <span class="preprocessor"> acc.s1 = mad(src0.s1, weights_row00.s0, acc.s1); \</span></div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span> <span class="preprocessor"> acc.s2 = mad(src0.s2, weights_row00.s0, acc.s2); \</span></div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span> <span class="preprocessor"> acc.s3 = mad(src0.s3, weights_row00.s0, acc.s3); \</span></div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span> <span class="preprocessor"> acc.s0 = mad(src0.s1, weights_row00.s1, acc.s0); \</span></div><div class="line"><a name="l00397"></a><span class="lineno"> 397</span> <span class="preprocessor"> acc.s1 = mad(src0.s2, weights_row00.s1, acc.s1); \</span></div><div class="line"><a name="l00398"></a><span class="lineno"> 398</span> <span class="preprocessor"> acc.s2 = mad(src0.s3, weights_row00.s1, acc.s2); \</span></div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span> <span class="preprocessor"> acc.s3 = mad(src0.s4, weights_row00.s1, acc.s3); \</span></div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span> <span class="preprocessor"> acc.s0 = mad(src0.s2, weights_row00.s2, acc.s0); \</span></div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span> <span class="preprocessor"> acc.s1 = mad(src0.s3, weights_row00.s2, acc.s1); \</span></div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span> <span class="preprocessor"> acc.s2 = mad(src0.s4, weights_row00.s2, acc.s2); \</span></div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span> <span class="preprocessor"> acc.s3 = mad(src0.s5, weights_row00.s2, acc.s3); \</span></div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span> <span class="preprocessor"> acc.s0 = mad(src0.s3, weights_row00.s3, acc.s0); \</span></div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span> <span class="preprocessor"> acc.s1 = mad(src0.s4, weights_row00.s3, acc.s1); \</span></div><div class="line"><a name="l00406"></a><span class="lineno"> 406</span> <span class="preprocessor"> acc.s2 = mad(src0.s5, weights_row00.s3, acc.s2); \</span></div><div class="line"><a name="l00407"></a><span class="lineno"> 407</span> <span class="preprocessor"> acc.s3 = mad(src0.s6, weights_row00.s3, acc.s3); \</span></div><div class="line"><a name="l00408"></a><span class="lineno"> 408</span> <span class="preprocessor"> acc.s0 = mad(src0.s4, weights_row01, acc.s0); \</span></div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span> <span class="preprocessor"> acc.s1 = mad(src0.s5, weights_row01, acc.s1); \</span></div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span> <span class="preprocessor"> acc.s2 = mad(src0.s6, weights_row01, acc.s2); \</span></div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span> <span class="preprocessor"> acc.s3 = mad(src0.s7, weights_row01, acc.s3); \</span></div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span> <span class="preprocessor"> })</span></div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span> </div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span> __kernel <span class="keywordtype">void</span> direct_convolution5x5_f32_bifrost(</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <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="l00452"></a><span class="lineno"> 452</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="l00453"></a><span class="lineno"> 453</span>  <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a6b83038822d1ae7ab619b684ed3b7fc0">TENSOR3D_DECLARATION</a>(weights),</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span> #ifdef HAS_BIAS</div><div class="line"><a name="l00455"></a><span class="lineno"> 455</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="l00456"></a><span class="lineno"> 456</span> #endif <span class="comment">/* defined(HAS_BIAS) */</span></div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  <span class="keywordtype">unsigned</span> <span class="keywordtype">int</span> weights_stride_w)</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span> {</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  <span class="comment">// Get the kernel index</span></div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> kernel_index = get_global_id(2);</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>  <a class="code" href="struct_image.xhtml">Image</a> src = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#aebe814363556c244be043b13e7969197">CONVERT_TO_IMAGE_STRUCT</a>(src);</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  <a class="code" href="struct_tensor3_d.xhtml">Tensor3D</a> dst = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a31c8c760f08fb1a331b16b7c204321dc">CONVERT_TO_TENSOR3D_STRUCT</a>(dst);</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span> </div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  float4 values0 = 0.0f;</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  float4 values1 = 0.0f;</div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span> </div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  __global uchar *weights_addr = (__global uchar *)(weights_ptr + weights_offset_first_element_in_bytes + kernel_index * weights_stride_w);</div><div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  __global uchar *src_addr = (__global uchar *)<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>(&src, 0, 0);</div><div class="line"><a name="l00470"></a><span class="lineno"> 470</span> </div><div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  <span class="comment">// Note: Since each work-item computes 4x2 elements, we need to load 6 rows from the input tensor</span></div><div class="line"><a name="l00472"></a><span class="lineno"> 472</span> </div><div class="line"><a name="l00473"></a><span class="lineno"> 473</span>  <span class="keywordflow">for</span>(ushort d = 0; d < (ushort)WEIGHTS_DEPTH; ++d)</div><div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  {</div><div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  <span class="comment">// Load the weights from row0 and row1</span></div><div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  float4 weights_row00 = vload4(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 0 * weights_stride_y));</div><div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  <span class="keywordtype">float</span> weights_row01 = *((__global <span class="keywordtype">float</span> *)(weights_addr + 0 * weights_stride_y) + 4);</div><div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  float4 weights_row10 = vload4(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 1 * weights_stride_y));</div><div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  <span class="keywordtype">float</span> weights_row11 = *((__global <span class="keywordtype">float</span> *)(weights_addr + 1 * weights_stride_y) + 4);</div><div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  float8 src0;</div><div class="line"><a name="l00481"></a><span class="lineno"> 481</span> </div><div class="line"><a name="l00482"></a><span class="lineno"> 482</span>  <span class="comment">// Load values from row0 of input tensor</span></div><div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 0 * src_stride_y));</div><div class="line"><a name="l00484"></a><span class="lineno"> 484</span> </div><div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00487"></a><span class="lineno"> 487</span> </div><div class="line"><a name="l00488"></a><span class="lineno"> 488</span>  <span class="comment">// Load values from row1 of input tensor</span></div><div class="line"><a name="l00489"></a><span class="lineno"> 489</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 1 * src_stride_y));</div><div class="line"><a name="l00490"></a><span class="lineno"> 490</span> </div><div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11);</div><div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00494"></a><span class="lineno"> 494</span> </div><div class="line"><a name="l00495"></a><span class="lineno"> 495</span>  <span class="comment">// Load values from row2 of input tensor</span></div><div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 2 * src_stride_y));</div><div class="line"><a name="l00497"></a><span class="lineno"> 497</span> </div><div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="comment">// Load weights from row2</span></div><div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  weights_row00 = vload4(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 2 * weights_stride_y));</div><div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  weights_row01 = *((__global <span class="keywordtype">float</span> *)(weights_addr + 2 * weights_stride_y) + 4);</div><div class="line"><a name="l00501"></a><span class="lineno"> 501</span> </div><div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00503"></a><span class="lineno"> 503</span>  CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11);</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">// Load values from row3 of input tensor</span></div><div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 3 * src_stride_y));</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="comment">// Load weights from row3</span></div><div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  weights_row10 = vload4(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 3 * weights_stride_y));</div><div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  weights_row11 = *((__global <span class="keywordtype">float</span> *)(weights_addr + 3 * weights_stride_y) + 4);</div><div class="line"><a name="l00512"></a><span class="lineno"> 512</span> </div><div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00514"></a><span class="lineno"> 514</span>  CONVOLUTION1x5_BIFROST(values0, src0, weights_row10, weights_row11);</div><div class="line"><a name="l00515"></a><span class="lineno"> 515</span>  CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00516"></a><span class="lineno"> 516</span> </div><div class="line"><a name="l00517"></a><span class="lineno"> 517</span>  <span class="comment">// Load values from row4 of input tensor</span></div><div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 4 * src_stride_y));</div><div class="line"><a name="l00519"></a><span class="lineno"> 519</span> </div><div class="line"><a name="l00520"></a><span class="lineno"> 520</span>  <span class="comment">// Load weights from row4</span></div><div class="line"><a name="l00521"></a><span class="lineno"> 521</span>  weights_row00 = vload4(0, (__global <span class="keywordtype">float</span> *)(weights_addr + 4 * weights_stride_y));</div><div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  weights_row01 = *((__global <span class="keywordtype">float</span> *)(weights_addr + 4 * weights_stride_y) + 4);</div><div class="line"><a name="l00523"></a><span class="lineno"> 523</span> </div><div class="line"><a name="l00524"></a><span class="lineno"> 524</span>  CONVOLUTION1x5_BIFROST(values0, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00525"></a><span class="lineno"> 525</span>  CONVOLUTION1x5_BIFROST(values1, src0, weights_row10, weights_row11);</div><div class="line"><a name="l00526"></a><span class="lineno"> 526</span> </div><div class="line"><a name="l00527"></a><span class="lineno"> 527</span>  <span class="comment">// Load values from row5 of input tensor</span></div><div class="line"><a name="l00528"></a><span class="lineno"> 528</span>  src0 = vload8(0, (__global <span class="keywordtype">float</span> *)(src_addr + 5 * src_stride_y));</div><div class="line"><a name="l00529"></a><span class="lineno"> 529</span> </div><div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  <span class="comment">// Accumulate</span></div><div class="line"><a name="l00531"></a><span class="lineno"> 531</span>  CONVOLUTION1x5_BIFROST(values1, src0, weights_row00, weights_row01);</div><div class="line"><a name="l00532"></a><span class="lineno"> 532</span> </div><div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  src_addr += src_stride_z;</div><div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  weights_addr += weights_stride_z;</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> </div><div class="line"><a name="l00537"></a><span class="lineno"> 537</span> <span class="preprocessor">#ifdef HAS_BIAS</span></div><div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  <a class="code" href="struct_vector.xhtml">Vector</a> biases = <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a64d779f80eeb923e0ab2313433f7b40b">CONVERT_TO_VECTOR_STRUCT_NO_STEP</a>(biases);</div><div class="line"><a name="l00539"></a><span class="lineno"> 539</span> </div><div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  float4 bias = (float4) * ((__global <span class="keywordtype">float</span> *)(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a7e4940407322d6f0ccb8b6b86b856019">vector_offset</a>(&biases, kernel_index)));</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>  values0 += bias;</div><div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  values1 += bias;</div><div class="line"><a name="l00544"></a><span class="lineno"> 544</span> <span class="preprocessor">#endif </span><span class="comment">/* defined(HAS_BIAS) */</span><span class="preprocessor"></span></div><div class="line"><a name="l00545"></a><span class="lineno"> 545</span> </div><div class="line"><a name="l00546"></a><span class="lineno"> 546</span>  vstore4(values0, 0, (__global <span class="keywordtype">float</span> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 0 * dst_stride_y));</div><div class="line"><a name="l00547"></a><span class="lineno"> 547</span>  vstore4(values1, 0, (__global <span class="keywordtype">float</span> *)(dst.<a class="code" href="struct_tensor3_d.xhtml#acf52c23cbd7424606c10a606524e3e32">ptr</a> + 1 * dst_stride_y));</div><div class="line"><a name="l00548"></a><span class="lineno"> 548</span> }</div><div class="line"><a name="l00549"></a><span class="lineno"> 549</span> <span class="preprocessor">#endif // defined(WEIGHTS_DEPTH)</span></div><div class="ttc" id="struct_vector_xhtml"><div class="ttname"><a href="struct_vector.xhtml">Vector</a></div><div class="ttdoc">Structure to hold Vector information. </div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00140">helpers.h:140</a></div></div> |