| <!-- HTML header for doxygen 1.8.15--> |
| <!-- Remember to use version doxygen 1.8.15 +--> |
| <!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd"> |
| <html xmlns="http://www.w3.org/1999/xhtml"> |
| <head> |
| <meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/> |
| <meta http-equiv="X-UA-Compatible" content="IE=9"/> |
| <meta name="generator" content="Doxygen 1.8.15"/> |
| <meta name="robots" content="NOINDEX, NOFOLLOW" /> <!-- Prevent indexing by search engines --> |
| <title>Compute Library: arm_compute/core/NEON/NEAsymm.h Source File</title> |
| <link href="tabs.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="jquery.js"></script> |
| <script type="text/javascript" src="dynsections.js"></script> |
| <link href="navtree.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="resize.js"></script> |
| <script type="text/javascript" src="navtreedata.js"></script> |
| <script type="text/javascript" src="navtree.js"></script> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(document).ready(initResizable); |
| /* @license-end */</script> |
| <link href="search/search.css" rel="stylesheet" type="text/css"/> |
| <script type="text/javascript" src="search/searchdata.js"></script> |
| <script type="text/javascript" src="search/search.js"></script> |
| <script type="text/x-mathjax-config"> |
| MathJax.Hub.Config({ |
| extensions: ["tex2jax.js"], |
| jax: ["input/TeX","output/HTML-CSS"], |
| }); |
| </script><script type="text/javascript" async="async" src="http://cdn.mathjax.org/mathjax/latest/MathJax.js"></script> |
| <link href="doxygen.css" rel="stylesheet" type="text/css" /> |
| <link href="stylesheet.css" rel="stylesheet" type="text/css"/> |
| </head> |
| <body> |
| <div id="top"><!-- do not remove this div, it is closed by doxygen! --> |
| <div id="titlearea"> |
| <table cellspacing="0" cellpadding="0"> |
| <tbody> |
| <tr style="height: 56px;"> |
| <img alt="Compute Library" src="https://raw.githubusercontent.com/ARM-software/ComputeLibrary/gh-pages/ACL_logo.png" style="max-width: 100%;margin-top: 15px;margin-left: 10px"/> |
| <td style="padding-left: 0.5em;"> |
| <div id="projectname"> |
|  <span id="projectnumber">19.11.1</span> |
| </div> |
| </td> |
| </tr> |
| </tbody> |
| </table> |
| </div> |
| <!-- end header part --> |
| <!-- Generated by Doxygen 1.8.15 --> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| var searchBox = new SearchBox("searchBox", "search",false,'Search'); |
| /* @license-end */ |
| </script> |
| <script type="text/javascript" src="menudata.js"></script> |
| <script type="text/javascript" src="menu.js"></script> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(function() { |
| initMenu('',true,false,'search.php','Search'); |
| $(document).ready(function() { init_search(); }); |
| }); |
| /* @license-end */</script> |
| <div id="main-nav"></div> |
| </div><!-- top --> |
| <div id="side-nav" class="ui-resizable side-nav-resizable"> |
| <div id="nav-tree"> |
| <div id="nav-tree-contents"> |
| <div id="nav-sync" class="sync"></div> |
| </div> |
| </div> |
| <div id="splitbar" style="-moz-user-select:none;" |
| class="ui-resizable-handle"> |
| </div> |
| </div> |
| <script type="text/javascript"> |
| /* @license magnet:?xt=urn:btih:cf05388f2679ee054f2beb29a391d25f4e673ac3&dn=gpl-2.0.txt GPL-v2 */ |
| $(document).ready(function(){initNavTree('_n_e_asymm_8h_source.xhtml','');}); |
| /* @license-end */ |
| </script> |
| <div id="doc-content"> |
| <!-- window showing the filter options --> |
| <div id="MSearchSelectWindow" |
| onmouseover="return searchBox.OnSearchSelectShow()" |
| onmouseout="return searchBox.OnSearchSelectHide()" |
| onkeydown="return searchBox.OnSearchSelectKey(event)"> |
| </div> |
| |
| <!-- iframe showing the search results (closed by default) --> |
| <div id="MSearchResultsWindow"> |
| <iframe src="javascript:void(0)" frameborder="0" |
| name="MSearchResults" id="MSearchResults"> |
| </iframe> |
| </div> |
| |
| <div class="header"> |
| <div class="headertitle"> |
| <div class="title">NEAsymm.h</div> </div> |
| </div><!--header--> |
| <div class="contents"> |
| <a href="_n_e_asymm_8h.xhtml">Go to the documentation of this file.</a><div class="fragment"><div class="line"><a name="l00001"></a><span class="lineno"> 1</span> <span class="comment">/*</span></div><div class="line"><a name="l00002"></a><span class="lineno"> 2</span> <span class="comment"> * Copyright (c) 2017-2019 ARM Limited.</span></div><div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> *</span></div><div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> * SPDX-License-Identifier: MIT</span></div><div class="line"><a name="l00005"></a><span class="lineno"> 5</span> <span class="comment"> *</span></div><div class="line"><a name="l00006"></a><span class="lineno"> 6</span> <span class="comment"> * Permission is hereby granted, free of charge, to any person obtaining a copy</span></div><div class="line"><a name="l00007"></a><span class="lineno"> 7</span> <span class="comment"> * of this software and associated documentation files (the "Software"), to</span></div><div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <span class="comment"> * deal in the Software without restriction, including without limitation the</span></div><div class="line"><a name="l00009"></a><span class="lineno"> 9</span> <span class="comment"> * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or</span></div><div class="line"><a name="l00010"></a><span class="lineno"> 10</span> <span class="comment"> * sell copies of the Software, and to permit persons to whom the Software is</span></div><div class="line"><a name="l00011"></a><span class="lineno"> 11</span> <span class="comment"> * furnished to do so, subject to the following conditions:</span></div><div class="line"><a name="l00012"></a><span class="lineno"> 12</span> <span class="comment"> *</span></div><div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="comment"> * The above copyright notice and this permission notice shall be included in all</span></div><div class="line"><a name="l00014"></a><span class="lineno"> 14</span> <span class="comment"> * copies or substantial portions of the Software.</span></div><div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="comment"> *</span></div><div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="comment"> * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR</span></div><div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="comment"> * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,</span></div><div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="comment"> * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE</span></div><div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="comment"> * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER</span></div><div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="comment"> * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,</span></div><div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="comment"> * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE</span></div><div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="comment"> * SOFTWARE.</span></div><div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="comment"> */</span></div><div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#ifndef __ARM_COMPUTE_NEASYMM_H__</span></div><div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="preprocessor">#define __ARM_COMPUTE_NEASYMM_H__</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">#include "<a class="code" href="_n_e_math_8h.xhtml">arm_compute/core/NEON/NEMath.h</a>"</span></div><div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#include <arm_neon.h></span></div><div class="line"><a name="l00029"></a><span class="lineno"> 29</span> </div><div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="keyword">namespace </span><a class="code" href="namespacearm__compute.xhtml">arm_compute</a></div><div class="line"><a name="l00031"></a><span class="lineno"> 31</span> {</div><div class="line"><a name="l00032"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a70ecc99f2f6b646579b58463dc026671"> 32</a></span> <span class="keyword">using</span> <a class="code" href="namespacearm__compute.xhtml#a70ecc99f2f6b646579b58463dc026671">qasymm8x8_t</a> = uint8x8_t; <span class="comment">/**< 8 bit quantized asymmetric vector with 8 elements */</span></div><div class="line"><a name="l00033"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a805d740633d70a5f0b019d72c517e67d"> 33</a></span> <span class="keyword">using</span> <a class="code" href="namespacearm__compute.xhtml#a805d740633d70a5f0b019d72c517e67d">qasymm8x8x2_t</a> = uint8x8x2_t; <span class="comment">/**< 8 bit quantized asymmetric vector with 16 elements */</span></div><div class="line"><a name="l00034"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a0c6e16826765130b480af53eda3635be"> 34</a></span> <span class="keyword">using</span> <a class="code" href="namespacearm__compute.xhtml#a0c6e16826765130b480af53eda3635be">qasymm8x8x3_t</a> = uint8x8x3_t; <span class="comment">/**< 8 bit quantized asymmetric vector with 24 elements */</span></div><div class="line"><a name="l00035"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a41d3460ff424e2e21c629b3f27386a8b"> 35</a></span> <span class="keyword">using</span> <a class="code" href="namespacearm__compute.xhtml#a41d3460ff424e2e21c629b3f27386a8b">qasymm8x8x4_t</a> = uint8x8x4_t; <span class="comment">/**< 8 bit quantized asymmetric vector with 32 elements */</span></div><div class="line"><a name="l00036"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#ac7e1832b8c58f07a98fc8d390b16ac27"> 36</a></span> <span class="keyword">using</span> <a class="code" href="namespacearm__compute.xhtml#ac7e1832b8c58f07a98fc8d390b16ac27">qasymm8x16_t</a> = uint8x16_t; <span class="comment">/**< 8 bit quantized asymmetric vector with 16 elements */</span></div><div class="line"><a name="l00037"></a><span class="lineno"> 37</span> <span class="comment"></span></div><div class="line"><a name="l00038"></a><span class="lineno"> 38</span> <span class="comment">/** Perform a multiply-accumulate on all 16 components of a QASYMM8 vector</span></div><div class="line"><a name="l00039"></a><span class="lineno"> 39</span> <span class="comment"> *</span></div><div class="line"><a name="l00040"></a><span class="lineno"> 40</span> <span class="comment"> * vd*vs + vo</span></div><div class="line"><a name="l00041"></a><span class="lineno"> 41</span> <span class="comment"> *</span></div><div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="comment"> * @param[in] vd Input vector value in QASYMM8 format</span></div><div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="comment"> * @param[in] vs Vector multiplier in F32 format. The multiplier value must be duplicated across all four lanes.</span></div><div class="line"><a name="l00044"></a><span class="lineno"> 44</span> <span class="comment"> * @param[in] vo Vector addend in F32 format. The addend value must be duplicated across all four lanes.</span></div><div class="line"><a name="l00045"></a><span class="lineno"> 45</span> <span class="comment"> *</span></div><div class="line"><a name="l00046"></a><span class="lineno"> 46</span> <span class="comment"> * @return A 16-component vector in QASYMM8 format, saturated to fit</span></div><div class="line"><a name="l00047"></a><span class="lineno"> 47</span> <span class="comment"> */</span></div><div class="line"><a name="l00048"></a><span class="lineno"> 48</span> uint8x16_t <a class="code" href="namespacearm__compute.xhtml#aba8de9eaa10a7cb45c5aa603585d0e4c">vmlaq_qasymm8</a>(<a class="code" href="namespacearm__compute.xhtml#ac7e1832b8c58f07a98fc8d390b16ac27">qasymm8x16_t</a> vd, float32x4_t vs, float32x4_t vo);</div><div class="line"><a name="l00049"></a><span class="lineno"> 49</span> <span class="comment"></span></div><div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="comment">/** Performs final quantization step on 16 elements</span></div><div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="comment"> *</span></div><div class="line"><a name="l00052"></a><span class="lineno"> 52</span> <span class="comment"> * @tparam is_bounded_relu Specified if a fused bounded relu should be applied</span></div><div class="line"><a name="l00053"></a><span class="lineno"> 53</span> <span class="comment"> *</span></div><div class="line"><a name="l00054"></a><span class="lineno"> 54</span> <span class="comment"> * @param in_s32 Input to be quantized.</span></div><div class="line"><a name="l00055"></a><span class="lineno"> 55</span> <span class="comment"> * @param result_fixedpoint_multiplier Result multiplier parameter</span></div><div class="line"><a name="l00056"></a><span class="lineno"> 56</span> <span class="comment"> * @param result_shift Result shift parameter</span></div><div class="line"><a name="l00057"></a><span class="lineno"> 57</span> <span class="comment"> * @param result_offset_after_shift_s32 Result offset parameter</span></div><div class="line"><a name="l00058"></a><span class="lineno"> 58</span> <span class="comment"> * @param min_u8 Relu lower bound</span></div><div class="line"><a name="l00059"></a><span class="lineno"> 59</span> <span class="comment"> * @param max_u8 Relu upper bound</span></div><div class="line"><a name="l00060"></a><span class="lineno"> 60</span> <span class="comment"> *</span></div><div class="line"><a name="l00061"></a><span class="lineno"> 61</span> <span class="comment"> * @return Quantized values</span></div><div class="line"><a name="l00062"></a><span class="lineno"> 62</span> <span class="comment"> */</span></div><div class="line"><a name="l00063"></a><span class="lineno"> 63</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> is_bounded_relu></div><div class="line"><a name="l00064"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a695a8105065167f7e4596d31f23a3573"> 64</a></span> uint8x16_t <a class="code" href="namespacearm__compute.xhtml#a695a8105065167f7e4596d31f23a3573">finalize_quantization</a>(int32x4x4_t &in_s32,</div><div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keywordtype">int</span> result_fixedpoint_multiplier,</div><div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  int32_t result_shift,</div><div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  int32x4_t result_offset_after_shift_s32,</div><div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  uint8x16_t min_u8,</div><div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  uint8x16_t max_u8)</div><div class="line"><a name="l00070"></a><span class="lineno"> 70</span> {</div><div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> <span class="keyword">static</span> int32x4_t zero_s32 = vdupq_n_s32(0);</div><div class="line"><a name="l00072"></a><span class="lineno"> 72</span> </div><div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  <span class="comment">// Fixed point multiplication with vector saturating rounding doubling multiply high with scalar</span></div><div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);</div><div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);</div><div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  in_s32.val[2] = vqrdmulhq_n_s32(in_s32.val[2], result_fixedpoint_multiplier);</div><div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  in_s32.val[3] = vqrdmulhq_n_s32(in_s32.val[3], result_fixedpoint_multiplier);</div><div class="line"><a name="l00078"></a><span class="lineno"> 78</span> </div><div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="comment">// Round to the nearest division by a power-of-two using result_shift_s32</span></div><div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  in_s32.val[0] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[0], result_shift);</div><div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  in_s32.val[1] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[1], result_shift);</div><div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  in_s32.val[2] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[2], result_shift);</div><div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  in_s32.val[3] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[3], result_shift);</div><div class="line"><a name="l00084"></a><span class="lineno"> 84</span> </div><div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="comment">// Add the offset terms</span></div><div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_after_shift_s32);</div><div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_after_shift_s32);</div><div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_after_shift_s32);</div><div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_after_shift_s32);</div><div class="line"><a name="l00090"></a><span class="lineno"> 90</span> </div><div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="comment">// Saturate negative values</span></div><div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);</div><div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);</div><div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);</div><div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32);</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="comment">// Convert S32 to S16</span></div><div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  <span class="keyword">const</span> int16x8x2_t in_s16 =</div><div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  {</div><div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  {</div><div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),</div><div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))</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>  };</div><div class="line"><a name="l00105"></a><span class="lineno"> 105</span> </div><div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="comment">// Convert S16 to U8</span></div><div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));</div><div class="line"><a name="l00108"></a><span class="lineno"> 108</span> </div><div class="line"><a name="l00109"></a><span class="lineno"> 109</span>  <span class="keywordflow">if</span>(is_bounded_relu)</div><div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  {</div><div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  out_u8 = vmaxq_u8(out_u8, min_u8);</div><div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  out_u8 = vminq_u8(out_u8, max_u8);</div><div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  }</div><div class="line"><a name="l00114"></a><span class="lineno"> 114</span> </div><div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  <span class="keywordflow">return</span> out_u8;</div><div class="line"><a name="l00116"></a><span class="lineno"> 116</span> }</div><div class="line"><a name="l00117"></a><span class="lineno"> 117</span> <span class="comment"></span></div><div class="line"><a name="l00118"></a><span class="lineno"> 118</span> <span class="comment">/** Performs final quantization step on 16 elements for symmetric quantization</span></div><div class="line"><a name="l00119"></a><span class="lineno"> 119</span> <span class="comment"> *</span></div><div class="line"><a name="l00120"></a><span class="lineno"> 120</span> <span class="comment"> * @tparam is_bounded_relu Specified if a fused bounded relu should be applied</span></div><div class="line"><a name="l00121"></a><span class="lineno"> 121</span> <span class="comment"> *</span></div><div class="line"><a name="l00122"></a><span class="lineno"> 122</span> <span class="comment"> * @param in_s32 Input to be quantized.</span></div><div class="line"><a name="l00123"></a><span class="lineno"> 123</span> <span class="comment"> * @param result_fixedpoint_multiplier Result multiplier parameter</span></div><div class="line"><a name="l00124"></a><span class="lineno"> 124</span> <span class="comment"> * @param result_shift Result shift parameter</span></div><div class="line"><a name="l00125"></a><span class="lineno"> 125</span> <span class="comment"> * @param result_offset_after_shift_s32 Result offset parameter</span></div><div class="line"><a name="l00126"></a><span class="lineno"> 126</span> <span class="comment"> * @param min_s8 Relu lower bound</span></div><div class="line"><a name="l00127"></a><span class="lineno"> 127</span> <span class="comment"> * @param max_s8 Relu upper bound</span></div><div class="line"><a name="l00128"></a><span class="lineno"> 128</span> <span class="comment"> *</span></div><div class="line"><a name="l00129"></a><span class="lineno"> 129</span> <span class="comment"> * @return Quantized values</span></div><div class="line"><a name="l00130"></a><span class="lineno"> 130</span> <span class="comment"> */</span></div><div class="line"><a name="l00131"></a><span class="lineno"> 131</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> is_bounded_relu></div><div class="line"><a name="l00132"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a189eb85a92663706760a7afcc09f77a7"> 132</a></span> <span class="keyword">inline</span> int8x16_t <a class="code" href="namespacearm__compute.xhtml#a189eb85a92663706760a7afcc09f77a7">finalize_quantization_symm</a>(int32x4x4_t &in_s32,</div><div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  <span class="keyword">const</span> int32x4x4_t &result_fixedpoint_multiplier,</div><div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  <span class="keyword">const</span> int32x4x4_t &result_shift,</div><div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  <span class="keyword">const</span> int32x4_t &result_offset_after_shift_s32,</div><div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="keyword">const</span> int8x16_t &min_s8,</div><div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  <span class="keyword">const</span> int8x16_t &max_s8)</div><div class="line"><a name="l00138"></a><span class="lineno"> 138</span> {</div><div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  <span class="comment">// Fixed point multiplication with vector saturating rounding doubling multiply high with scalar</span></div><div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  in_s32.val[0] = vqrdmulhq_s32(in_s32.val[0], result_fixedpoint_multiplier.val[0]);</div><div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  in_s32.val[1] = vqrdmulhq_s32(in_s32.val[1], result_fixedpoint_multiplier.val[1]);</div><div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  in_s32.val[2] = vqrdmulhq_s32(in_s32.val[2], result_fixedpoint_multiplier.val[2]);</div><div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  in_s32.val[3] = vqrdmulhq_s32(in_s32.val[3], result_fixedpoint_multiplier.val[3]);</div><div class="line"><a name="l00144"></a><span class="lineno"> 144</span> </div><div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="comment">// Round to the nearest division by a power-of-two using result_shift_s32</span></div><div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  in_s32.val[0] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[0], result_shift.val[0]);</div><div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  in_s32.val[1] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[1], result_shift.val[1]);</div><div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  in_s32.val[2] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[2], result_shift.val[2]);</div><div class="line"><a name="l00149"></a><span class="lineno"> 149</span>  in_s32.val[3] = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_s32.val[3], result_shift.val[3]);</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="comment">// Add the offset terms</span></div><div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_after_shift_s32);</div><div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_after_shift_s32);</div><div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_after_shift_s32);</div><div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_after_shift_s32);</div><div class="line"><a name="l00156"></a><span class="lineno"> 156</span> </div><div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  <span class="comment">// Convert S32 to S16</span></div><div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  <span class="keyword">const</span> int16x8x2_t in_s16 =</div><div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  {</div><div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  {</div><div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),</div><div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))</div><div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  }</div><div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  };</div><div class="line"><a name="l00165"></a><span class="lineno"> 165</span> </div><div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="comment">// Convert S16 to S8</span></div><div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1]));</div><div class="line"><a name="l00168"></a><span class="lineno"> 168</span> </div><div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  <span class="keywordflow">if</span>(is_bounded_relu)</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>  out_s8 = vmaxq_s8(out_s8, min_s8);</div><div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  out_s8 = vminq_s8(out_s8, max_s8);</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> </div><div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <span class="keywordflow">return</span> out_s8;</div><div class="line"><a name="l00176"></a><span class="lineno"> 176</span> }</div><div class="line"><a name="l00177"></a><span class="lineno"> 177</span> <span class="comment"></span></div><div class="line"><a name="l00178"></a><span class="lineno"> 178</span> <span class="comment">/** Performs final quantization step on single element</span></div><div class="line"><a name="l00179"></a><span class="lineno"> 179</span> <span class="comment"> *</span></div><div class="line"><a name="l00180"></a><span class="lineno"> 180</span> <span class="comment"> * @tparam is_bounded_relu Specified if a fused bounded relu should be applied</span></div><div class="line"><a name="l00181"></a><span class="lineno"> 181</span> <span class="comment"> *</span></div><div class="line"><a name="l00182"></a><span class="lineno"> 182</span> <span class="comment"> * @param[in] in_value Input to be quantized.</span></div><div class="line"><a name="l00183"></a><span class="lineno"> 183</span> <span class="comment"> * @param[in] result_fixedpoint_multiplier Result multiplier parameter</span></div><div class="line"><a name="l00184"></a><span class="lineno"> 184</span> <span class="comment"> * @param[in] result_shift Result shift parameter</span></div><div class="line"><a name="l00185"></a><span class="lineno"> 185</span> <span class="comment"> * @param[in] result_offset_after_shift_s32 Result offset parameter</span></div><div class="line"><a name="l00186"></a><span class="lineno"> 186</span> <span class="comment"> * @param[in] min_u8 Relu lower bound</span></div><div class="line"><a name="l00187"></a><span class="lineno"> 187</span> <span class="comment"> * @param[in] max_u8 Relu upper bound</span></div><div class="line"><a name="l00188"></a><span class="lineno"> 188</span> <span class="comment"> *</span></div><div class="line"><a name="l00189"></a><span class="lineno"> 189</span> <span class="comment"> * @return Quantized value</span></div><div class="line"><a name="l00190"></a><span class="lineno"> 190</span> <span class="comment"> */</span></div><div class="line"><a name="l00191"></a><span class="lineno"> 191</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> is_bounded_relu></div><div class="line"><a name="l00192"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a86f8bf876974759725e39496be2cc7f9"> 192</a></span> <span class="keyword">inline</span> uint8_t <a class="code" href="namespacearm__compute.xhtml#a695a8105065167f7e4596d31f23a3573">finalize_quantization</a>(int32_t in_value, <span class="keywordtype">int</span> result_fixedpoint_multiplier,</div><div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  int32_t result_shift, int32_t result_offset_after_shift_s32,</div><div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  uint8_t min_u8, uint8_t max_u8)</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>  int32x4_t in_s32 = vdupq_n_s32(in_value);</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="comment">// Fixed point multiplication with vector saturating rounding doubling multiply high with scalar</span></div><div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  in_value = vgetq_lane_s32(vqrdmulhq_n_s32(in_s32, result_fixedpoint_multiplier), 0);</div><div class="line"><a name="l00200"></a><span class="lineno"> 200</span> </div><div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  <span class="comment">// Shift value by result_shift_s32</span></div><div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  in_value = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_value, result_shift);</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">// Add the offset term</span></div><div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  in_value += result_offset_after_shift_s32;</div><div class="line"><a name="l00206"></a><span class="lineno"> 206</span> </div><div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  <span class="comment">// Bound the result</span></div><div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  uint8_t out_u8 = static_cast<uint8_t>(std::max<int32_t>(0, std::min<int32_t>(255, in_value)));</div><div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="keywordflow">if</span>(is_bounded_relu)</div><div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  {</div><div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  out_u8 = static_cast<uint8_t>(std::max(min_u8, std::min(max_u8, out_u8)));</div><div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  }</div><div class="line"><a name="l00213"></a><span class="lineno"> 213</span> </div><div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="keywordflow">return</span> out_u8;</div><div class="line"><a name="l00215"></a><span class="lineno"> 215</span> }</div><div class="line"><a name="l00216"></a><span class="lineno"> 216</span> <span class="comment"></span></div><div class="line"><a name="l00217"></a><span class="lineno"> 217</span> <span class="comment">/** Performs final quantization step on single element</span></div><div class="line"><a name="l00218"></a><span class="lineno"> 218</span> <span class="comment"> *</span></div><div class="line"><a name="l00219"></a><span class="lineno"> 219</span> <span class="comment"> * @tparam is_bounded_relu Specified if a fused bounded relu should be applied</span></div><div class="line"><a name="l00220"></a><span class="lineno"> 220</span> <span class="comment"> *</span></div><div class="line"><a name="l00221"></a><span class="lineno"> 221</span> <span class="comment"> * @param[in] in_value Input to be quantized.</span></div><div class="line"><a name="l00222"></a><span class="lineno"> 222</span> <span class="comment"> * @param[in] result_fixedpoint_multiplier Result multiplier parameter</span></div><div class="line"><a name="l00223"></a><span class="lineno"> 223</span> <span class="comment"> * @param[in] result_shift Result shift parameter</span></div><div class="line"><a name="l00224"></a><span class="lineno"> 224</span> <span class="comment"> * @param[in] result_offset_after_shift_s32 Result offset parameter</span></div><div class="line"><a name="l00225"></a><span class="lineno"> 225</span> <span class="comment"> * @param[in] min_s8 Relu lower bound</span></div><div class="line"><a name="l00226"></a><span class="lineno"> 226</span> <span class="comment"> * @param[in] max_s8 Relu upper bound</span></div><div class="line"><a name="l00227"></a><span class="lineno"> 227</span> <span class="comment"> *</span></div><div class="line"><a name="l00228"></a><span class="lineno"> 228</span> <span class="comment"> * @return Quantized value</span></div><div class="line"><a name="l00229"></a><span class="lineno"> 229</span> <span class="comment"> */</span></div><div class="line"><a name="l00230"></a><span class="lineno"> 230</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> is_bounded_relu></div><div class="line"><a name="l00231"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#ae83b33be1e3199b67bf22a771eb298a2"> 231</a></span> <span class="keyword">inline</span> int8_t <a class="code" href="namespacearm__compute.xhtml#a695a8105065167f7e4596d31f23a3573">finalize_quantization</a>(int32_t in_value, <span class="keywordtype">int</span> result_fixedpoint_multiplier,</div><div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  int32_t result_shift, int32_t result_offset_after_shift_s32,</div><div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  int8_t min_s8, int8_t max_s8)</div><div class="line"><a name="l00234"></a><span class="lineno"> 234</span> {</div><div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  int32x4_t in_s32 = vdupq_n_s32(in_value);</div><div class="line"><a name="l00236"></a><span class="lineno"> 236</span> </div><div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="comment">// Fixed point multiplication with vector saturating rounding doubling multiply high with scalar</span></div><div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  in_value = vgetq_lane_s32(vqrdmulhq_n_s32(in_s32, result_fixedpoint_multiplier), 0);</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>  <span class="comment">// Shift value by result_shift_s32</span></div><div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  in_value = <a class="code" href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">rounding_divide_by_pow2</a>(in_value, result_shift);</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">// Add the offset term</span></div><div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  in_value += result_offset_after_shift_s32;</div><div class="line"><a name="l00245"></a><span class="lineno"> 245</span> </div><div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="comment">// Bound the result</span></div><div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  int8_t out_s8 = static_cast<int8_t>(std::max<int32_t>(-128, std::min<int32_t>(127, in_value)));</div><div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  <span class="keywordflow">if</span>(is_bounded_relu)</div><div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  {</div><div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  out_s8 = static_cast<int8_t>(std::max(min_s8, std::min(max_s8, out_s8)));</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> </div><div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  <span class="keywordflow">return</span> out_s8;</div><div class="line"><a name="l00254"></a><span class="lineno"> 254</span> }</div><div class="line"><a name="l00255"></a><span class="lineno"> 255</span> <span class="comment"></span></div><div class="line"><a name="l00256"></a><span class="lineno"> 256</span> <span class="comment">/** Dequantize a neon vector holding 8 quantized values.</span></div><div class="line"><a name="l00257"></a><span class="lineno"> 257</span> <span class="comment"> *</span></div><div class="line"><a name="l00258"></a><span class="lineno"> 258</span> <span class="comment"> * @param[in] qv Input values to be dequantized.</span></div><div class="line"><a name="l00259"></a><span class="lineno"> 259</span> <span class="comment"> * @param[in] qi Quantization information to be used in the computation.</span></div><div class="line"><a name="l00260"></a><span class="lineno"> 260</span> <span class="comment"> *</span></div><div class="line"><a name="l00261"></a><span class="lineno"> 261</span> <span class="comment"> * @return Dequantized values in a neon vector</span></div><div class="line"><a name="l00262"></a><span class="lineno"> 262</span> <span class="comment"> */</span></div><div class="line"><a name="l00263"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7"> 263</a></span> <span class="keyword">inline</span> float32x4x2_t <a class="code" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">vdequantize</a>(<span class="keyword">const</span> uint8x8_t &qv, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml">UniformQuantizationInfo</a> &qi)</div><div class="line"><a name="l00264"></a><span class="lineno"> 264</span> {</div><div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  <span class="keyword">const</span> <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">scale</a>;</div><div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">offset</a>;</div><div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  <span class="keyword">const</span> int32x4_t voffset = vdupq_n_s32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  <span class="keyword">const</span> float32x4_t vscale = vdupq_n_f32(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  <span class="keyword">const</span> float32x4x2_t vdequantized_input =</div><div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  {</div><div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  {</div><div class="line"><a name="l00272"></a><span class="lineno"> 272</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(qv)))), voffset)), vscale),</div><div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(qv)))), voffset)), vscale),</div><div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  }</div><div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  };</div><div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  <span class="keywordflow">return</span> vdequantized_input;</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="comment"></span></div><div class="line"><a name="l00279"></a><span class="lineno"> 279</span> <span class="comment">/** Dequantize a neon vector holding 16 quantized values.</span></div><div class="line"><a name="l00280"></a><span class="lineno"> 280</span> <span class="comment"> *</span></div><div class="line"><a name="l00281"></a><span class="lineno"> 281</span> <span class="comment"> * @param[in] qv Input values to be dequantized.</span></div><div class="line"><a name="l00282"></a><span class="lineno"> 282</span> <span class="comment"> * @param[in] qi Quantization information to be used in the computation.</span></div><div class="line"><a name="l00283"></a><span class="lineno"> 283</span> <span class="comment"> *</span></div><div class="line"><a name="l00284"></a><span class="lineno"> 284</span> <span class="comment"> * @return Dequantized values in a neon vector</span></div><div class="line"><a name="l00285"></a><span class="lineno"> 285</span> <span class="comment"> */</span></div><div class="line"><a name="l00286"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a7048ece2af58ce75587256d7980db171"> 286</a></span> <span class="keyword">inline</span> float32x4x4_t <a class="code" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">vdequantize</a>(<span class="keyword">const</span> uint8x16_t &qv, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml">UniformQuantizationInfo</a> &qi)</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="keyword">const</span> <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">scale</a>;</div><div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">offset</a>;</div><div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  <span class="keyword">const</span> int32x4_t voffset = vdupq_n_s32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <span class="keyword">const</span> float32x4_t vscale = vdupq_n_f32(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <span class="keyword">const</span> float32x4x4_t vdequantized_input =</div><div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  {</div><div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  {</div><div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00298"></a><span class="lineno"> 298</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),</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>  };</div><div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  <span class="keywordflow">return</span> vdequantized_input;</div><div class="line"><a name="l00302"></a><span class="lineno"> 302</span> }</div><div class="line"><a name="l00303"></a><span class="lineno"> 303</span> <span class="comment"></span></div><div class="line"><a name="l00304"></a><span class="lineno"> 304</span> <span class="comment">/** Dequantize following an asymmetric quantization scheme a neon vector holding 16 quantized values.</span></div><div class="line"><a name="l00305"></a><span class="lineno"> 305</span> <span class="comment"> *</span></div><div class="line"><a name="l00306"></a><span class="lineno"> 306</span> <span class="comment"> * @param[in] qv Input values to be dequantized.</span></div><div class="line"><a name="l00307"></a><span class="lineno"> 307</span> <span class="comment"> * @param[in] scale Quantization scaling factor.</span></div><div class="line"><a name="l00308"></a><span class="lineno"> 308</span> <span class="comment"> * @param[in] offset Zero quantization offset.</span></div><div class="line"><a name="l00309"></a><span class="lineno"> 309</span> <span class="comment"> *</span></div><div class="line"><a name="l00310"></a><span class="lineno"> 310</span> <span class="comment"> * @return Dequantized values in a neon vector</span></div><div class="line"><a name="l00311"></a><span class="lineno"> 311</span> <span class="comment"> */</span></div><div class="line"><a name="l00312"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#afefd32b0cf61750c523e6d15790a561e"> 312</a></span> <span class="keyword">inline</span> float32x4x4_t <a class="code" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">vdequantize</a>(<span class="keyword">const</span> uint8x16_t &qv, <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>, int32_t <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>)</div><div class="line"><a name="l00313"></a><span class="lineno"> 313</span> {</div><div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  <span class="keyword">const</span> int32x4_t voffset = vdupq_n_s32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  <span class="keyword">const</span> float32x4_t vscale = vdupq_n_f32(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  <span class="keyword">const</span> float32x4x4_t vdequantized_input =</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>  {</div><div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_low_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00321"></a><span class="lineno"> 321</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  vmulq_f32(vcvtq_f32_s32(vsubq_s32(vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(vmovl_u8(vget_high_u8(qv))))), voffset)), vscale),</div><div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  }</div><div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  };</div><div class="line"><a name="l00325"></a><span class="lineno"> 325</span>  <span class="keywordflow">return</span> vdequantized_input;</div><div class="line"><a name="l00326"></a><span class="lineno"> 326</span> }</div><div class="line"><a name="l00327"></a><span class="lineno"> 327</span> <span class="comment"></span></div><div class="line"><a name="l00328"></a><span class="lineno"> 328</span> <span class="comment">/** Dequantize following symmetric quantization scheme a neon vector holding 16 quantized values.</span></div><div class="line"><a name="l00329"></a><span class="lineno"> 329</span> <span class="comment"> *</span></div><div class="line"><a name="l00330"></a><span class="lineno"> 330</span> <span class="comment"> * @param[in] qv Input values to be dequantized.</span></div><div class="line"><a name="l00331"></a><span class="lineno"> 331</span> <span class="comment"> * @param[in] vscale Vector containing quantization scaling factors.</span></div><div class="line"><a name="l00332"></a><span class="lineno"> 332</span> <span class="comment"> *</span></div><div class="line"><a name="l00333"></a><span class="lineno"> 333</span> <span class="comment"> * @return Dequantized values in a neon vector</span></div><div class="line"><a name="l00334"></a><span class="lineno"> 334</span> <span class="comment"> */</span></div><div class="line"><a name="l00335"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a3838d1ee6a765a3b7ff52dc040edf5c0"> 335</a></span> <span class="keyword">inline</span> float32x4x4_t <a class="code" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">vdequantize</a>(<span class="keyword">const</span> int8x16_t &qv, <span class="keyword">const</span> float32x4x4_t vscale)</div><div class="line"><a name="l00336"></a><span class="lineno"> 336</span> {</div><div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  <span class="keyword">const</span> float32x4x4_t vdequantized_input =</div><div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  {</div><div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  {</div><div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(qv))))), vscale.val[0]),</div><div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(qv))))), vscale.val[1]),</div><div class="line"><a name="l00342"></a><span class="lineno"> 342</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(qv))))), vscale.val[2]),</div><div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(qv))))), vscale.val[3]),</div><div class="line"><a name="l00344"></a><span class="lineno"> 344</span>  }</div><div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  };</div><div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  <span class="keywordflow">return</span> vdequantized_input;</div><div class="line"><a name="l00347"></a><span class="lineno"> 347</span> }</div><div class="line"><a name="l00348"></a><span class="lineno"> 348</span> <span class="comment"></span></div><div class="line"><a name="l00349"></a><span class="lineno"> 349</span> <span class="comment">/** Dequantize following a symmetric quantization scheme a neon vector holding 16 quantized values.</span></div><div class="line"><a name="l00350"></a><span class="lineno"> 350</span> <span class="comment"> *</span></div><div class="line"><a name="l00351"></a><span class="lineno"> 351</span> <span class="comment"> * @param[in] qv Input values to be dequantized.</span></div><div class="line"><a name="l00352"></a><span class="lineno"> 352</span> <span class="comment"> * @param[in] scale Quantization scaling factor.</span></div><div class="line"><a name="l00353"></a><span class="lineno"> 353</span> <span class="comment"> *</span></div><div class="line"><a name="l00354"></a><span class="lineno"> 354</span> <span class="comment"> * @return Dequantized values in a neon vector</span></div><div class="line"><a name="l00355"></a><span class="lineno"> 355</span> <span class="comment"> */</span></div><div class="line"><a name="l00356"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a496aa10035fdc8a0e86dde37b64f8228"> 356</a></span> <span class="keyword">inline</span> float32x4x4_t <a class="code" href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">vdequantize</a>(<span class="keyword">const</span> int8x16_t &qv, <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>)</div><div class="line"><a name="l00357"></a><span class="lineno"> 357</span> {</div><div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  <span class="keyword">const</span> float32x4_t vscale = vdupq_n_f32(<a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00359"></a><span class="lineno"> 359</span>  <span class="keyword">const</span> float32x4x4_t vdequantized_input =</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>  {</div><div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_low_s8(qv))))), vscale),</div><div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_low_s8(qv))))), vscale),</div><div class="line"><a name="l00364"></a><span class="lineno"> 364</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_low_s16(vmovl_s8(vget_high_s8(qv))))), vscale),</div><div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  vmulq_f32(vcvtq_f32_s32(vmovl_s16(vget_high_s16(vmovl_s8(vget_high_s8(qv))))), vscale),</div><div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  }</div><div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  };</div><div class="line"><a name="l00368"></a><span class="lineno"> 368</span>  <span class="keywordflow">return</span> vdequantized_input;</div><div class="line"><a name="l00369"></a><span class="lineno"> 369</span> }</div><div class="line"><a name="l00370"></a><span class="lineno"> 370</span> <span class="comment"></span></div><div class="line"><a name="l00371"></a><span class="lineno"> 371</span> <span class="comment">/** Quantize a neon vector holding 8 floating point values.</span></div><div class="line"><a name="l00372"></a><span class="lineno"> 372</span> <span class="comment"> *</span></div><div class="line"><a name="l00373"></a><span class="lineno"> 373</span> <span class="comment"> * @param[in] qv Input values to be quantized.</span></div><div class="line"><a name="l00374"></a><span class="lineno"> 374</span> <span class="comment"> * @param[in] qi Quantization information to be used in the computation.</span></div><div class="line"><a name="l00375"></a><span class="lineno"> 375</span> <span class="comment"> *</span></div><div class="line"><a name="l00376"></a><span class="lineno"> 376</span> <span class="comment"> * @return A neon vector holding the quantized values</span></div><div class="line"><a name="l00377"></a><span class="lineno"> 377</span> <span class="comment"> */</span></div><div class="line"><a name="l00378"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a4ff1249e423ae9bfab075f5317c4d004"> 378</a></span> <span class="keyword">inline</span> uint8x8_t <a class="code" href="namespacearm__compute.xhtml#a4ff1249e423ae9bfab075f5317c4d004">vquantize</a>(<span class="keyword">const</span> float32x4x2_t &qv, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml">UniformQuantizationInfo</a> &qi)</div><div class="line"><a name="l00379"></a><span class="lineno"> 379</span> {</div><div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="keyword">const</span> <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">scale</a>;</div><div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">offset</a>;</div><div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <span class="keyword">const</span> float32x4_t voffset = vdupq_n_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  <span class="keyword">const</span> float32x4_t vinvscale = vdupq_n_f32(1.f / <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00384"></a><span class="lineno"> 384</span>  <span class="keyword">const</span> int32x4x4_t rf =</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>  {</div><div class="line"><a name="l00387"></a><span class="lineno"> 387</span> <span class="preprocessor">#ifdef __aarch64__</span></div><div class="line"><a name="l00388"></a><span class="lineno"> 388</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00390"></a><span class="lineno"> 390</span> <span class="preprocessor">#else //__aarch64__</span></div><div class="line"><a name="l00391"></a><span class="lineno"> 391</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00393"></a><span class="lineno"> 393</span> <span class="preprocessor">#endif //__aarch64__</span></div><div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  }</div><div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  };</div><div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  <span class="keywordflow">return</span> vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));</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> <span class="comment"></span></div><div class="line"><a name="l00399"></a><span class="lineno"> 399</span> <span class="comment">/** Quantize a neon vector holding 16 floating point values.</span></div><div class="line"><a name="l00400"></a><span class="lineno"> 400</span> <span class="comment"> *</span></div><div class="line"><a name="l00401"></a><span class="lineno"> 401</span> <span class="comment"> * @param[in] qv Input values to be quantized.</span></div><div class="line"><a name="l00402"></a><span class="lineno"> 402</span> <span class="comment"> * @param[in] qi Quantization information to be used in the computation.</span></div><div class="line"><a name="l00403"></a><span class="lineno"> 403</span> <span class="comment"> *</span></div><div class="line"><a name="l00404"></a><span class="lineno"> 404</span> <span class="comment"> * @return A neon vector holding the quantized values</span></div><div class="line"><a name="l00405"></a><span class="lineno"> 405</span> <span class="comment"> */</span></div><div class="line"><a name="l00406"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a42c040989e7f451dd9f2efa7384b89f6"> 406</a></span> <span class="keyword">inline</span> uint8x16_t <a class="code" href="namespacearm__compute.xhtml#a4ff1249e423ae9bfab075f5317c4d004">vquantize</a>(<span class="keyword">const</span> float32x4x4_t &qv, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml">UniformQuantizationInfo</a> &qi)</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="keyword">const</span> <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">scale</a>;</div><div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">offset</a>;</div><div class="line"><a name="l00410"></a><span class="lineno"> 410</span>  <span class="keyword">const</span> float32x4_t voffset = vdupq_n_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00411"></a><span class="lineno"> 411</span>  <span class="keyword">const</span> float32x4_t vinvscale = vdupq_n_f32(1.f / <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="keyword">const</span> int32x4x4_t rf =</div><div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  {</div><div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  {</div><div class="line"><a name="l00415"></a><span class="lineno"> 415</span> <span class="preprocessor">#ifdef __aarch64__</span></div><div class="line"><a name="l00416"></a><span class="lineno"> 416</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00418"></a><span class="lineno"> 418</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),</div><div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),</div><div class="line"><a name="l00420"></a><span class="lineno"> 420</span> <span class="preprocessor">#else //__aarch64__</span></div><div class="line"><a name="l00421"></a><span class="lineno"> 421</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00422"></a><span class="lineno"> 422</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),</div><div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),</div><div class="line"><a name="l00425"></a><span class="lineno"> 425</span> <span class="preprocessor">#endif //__aarch64__</span></div><div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  }</div><div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  };</div><div class="line"><a name="l00428"></a><span class="lineno"> 428</span>  <span class="keyword">const</span> uint8x8_t pa = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));</div><div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  <span class="keyword">const</span> uint8x8_t pb = vqmovun_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));</div><div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  <span class="keywordflow">return</span> vcombine_u8(pa, pb);</div><div class="line"><a name="l00431"></a><span class="lineno"> 431</span> }</div><div class="line"><a name="l00432"></a><span class="lineno"> 432</span> <span class="comment"></span></div><div class="line"><a name="l00433"></a><span class="lineno"> 433</span> <span class="comment">/** Quantize to QASYMM16 a neon vector holding 16 floating point values.</span></div><div class="line"><a name="l00434"></a><span class="lineno"> 434</span> <span class="comment"> *</span></div><div class="line"><a name="l00435"></a><span class="lineno"> 435</span> <span class="comment"> * @param[in] qv Input values to be quantized.</span></div><div class="line"><a name="l00436"></a><span class="lineno"> 436</span> <span class="comment"> * @param[in] qi Quantization information to be used in the computation.</span></div><div class="line"><a name="l00437"></a><span class="lineno"> 437</span> <span class="comment"> *</span></div><div class="line"><a name="l00438"></a><span class="lineno"> 438</span> <span class="comment"> * @return A neon vector holding the quantized values</span></div><div class="line"><a name="l00439"></a><span class="lineno"> 439</span> <span class="comment"> */</span></div><div class="line"><a name="l00440"></a><span class="lineno"><a class="line" href="namespacearm__compute.xhtml#a687dee3fd66555374211e3ff676ce5f1"> 440</a></span> <span class="keyword">inline</span> uint16x8x2_t <a class="code" href="namespacearm__compute.xhtml#a687dee3fd66555374211e3ff676ce5f1">vquantize_qasymm16</a>(<span class="keyword">const</span> float32x4x4_t &qv, <span class="keyword">const</span> <a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml">UniformQuantizationInfo</a> &qi)</div><div class="line"><a name="l00441"></a><span class="lineno"> 441</span> {</div><div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  <span class="keyword">const</span> <span class="keywordtype">float</span> <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">scale</a>;</div><div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a> = qi.<a class="code" href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">offset</a>;</div><div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  <span class="keyword">const</span> float32x4_t voffset = vdupq_n_f32(<a class="code" href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a>);</div><div class="line"><a name="l00445"></a><span class="lineno"> 445</span>  <span class="keyword">const</span> float32x4_t vinvscale = vdupq_n_f32(1.f / <a class="code" href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">scale</a>);</div><div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  <span class="keyword">const</span> int32x4x4_t rf =</div><div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  {</div><div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  {</div><div class="line"><a name="l00449"></a><span class="lineno"> 449</span> <span class="preprocessor">#ifdef __aarch64__</span></div><div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),</div><div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  vcvtnq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),</div><div class="line"><a name="l00454"></a><span class="lineno"> 454</span> <span class="preprocessor">#else //__aarch64__</span></div><div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[0], vinvscale)),</div><div class="line"><a name="l00456"></a><span class="lineno"> 456</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[1], vinvscale)),</div><div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[2], vinvscale)),</div><div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  vcvtq_s32_f32(vmlaq_f32(voffset, qv.val[3], vinvscale)),</div><div class="line"><a name="l00459"></a><span class="lineno"> 459</span> <span class="preprocessor">#endif //__aarch64__</span></div><div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  }</div><div class="line"><a name="l00461"></a><span class="lineno"> 461</span>  };</div><div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  <span class="keyword">const</span> uint16x8_t pa = vcombine_u16(vqmovun_s32(rf.val[0]), vqmovun_s32(rf.val[1]));</div><div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  <span class="keyword">const</span> uint16x8_t pb = vcombine_u16(vqmovun_s32(rf.val[2]), vqmovun_s32(rf.val[3]));</div><div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  <span class="keywordflow">return</span> { pa, pb };</div><div class="line"><a name="l00465"></a><span class="lineno"> 465</span> }</div><div class="line"><a name="l00466"></a><span class="lineno"> 466</span> } <span class="comment">// namespace arm_compute</span></div><div class="line"><a name="l00467"></a><span class="lineno"> 467</span> <span class="preprocessor">#include "<a class="code" href="_n_e_asymm_8inl.xhtml">arm_compute/core/NEON/NEAsymm.inl</a>"</span></div><div class="line"><a name="l00468"></a><span class="lineno"> 468</span> <span class="preprocessor">#endif // __ARM_COMPUTE_NEASYMM_H__</span></div><div class="ttc" id="src_2core_2_c_l_2cl__kernels_2_helpers_8h_xhtml_a009469e4d9b8fce3b6d5e97d2077827d"><div class="ttname"><a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h.xhtml#a009469e4d9b8fce3b6d5e97d2077827d">offset</a></div><div class="ttdeci">__global uchar * offset(const Image *img, int x, int y)</div><div class="ttdoc">Get the pointer position of a Image.</div><div class="ttdef"><b>Definition:</b> <a href="src_2core_2_c_l_2cl__kernels_2_helpers_8h_source.xhtml#l00510">helpers.h:510</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a805d740633d70a5f0b019d72c517e67d"><div class="ttname"><a href="namespacearm__compute.xhtml#a805d740633d70a5f0b019d72c517e67d">arm_compute::qasymm8x8x2_t</a></div><div class="ttdeci">uint8x8x2_t qasymm8x8x2_t</div><div class="ttdoc">8 bit quantized asymmetric vector with 16 elements</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00033">NEAsymm.h:33</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a687dee3fd66555374211e3ff676ce5f1"><div class="ttname"><a href="namespacearm__compute.xhtml#a687dee3fd66555374211e3ff676ce5f1">arm_compute::vquantize_qasymm16</a></div><div class="ttdeci">uint16x8x2_t vquantize_qasymm16(const float32x4x4_t &qv, const UniformQuantizationInfo &qi)</div><div class="ttdoc">Quantize to QASYMM16 a neon vector holding 16 floating point values.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00440">NEAsymm.h:440</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a485ae62f3ade6c8d33c5c2e85b29c3a7"><div class="ttname"><a href="namespacearm__compute.xhtml#a485ae62f3ade6c8d33c5c2e85b29c3a7">arm_compute::vdequantize</a></div><div class="ttdeci">float32x4x2_t vdequantize(const uint8x8_t &qv, const UniformQuantizationInfo &qi)</div><div class="ttdoc">Dequantize a neon vector holding 8 quantized values.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00263">NEAsymm.h:263</a></div></div> |
| <div class="ttc" id="structarm__compute_1_1_uniform_quantization_info_xhtml"><div class="ttname"><a href="structarm__compute_1_1_uniform_quantization_info.xhtml">arm_compute::UniformQuantizationInfo</a></div><div class="ttdoc">Quantization info when assuming per layer quantization.</div><div class="ttdef"><b>Definition:</b> <a href="_quantization_info_8h_source.xhtml#l00040">QuantizationInfo.h:40</a></div></div> |
| <div class="ttc" id="structarm__compute_1_1_uniform_quantization_info_xhtml_a1d28dec57cce925ad92342891bd71e7c"><div class="ttname"><a href="structarm__compute_1_1_uniform_quantization_info.xhtml#a1d28dec57cce925ad92342891bd71e7c">arm_compute::UniformQuantizationInfo::scale</a></div><div class="ttdeci">float scale</div><div class="ttdef"><b>Definition:</b> <a href="_quantization_info_8h_source.xhtml#l00062">QuantizationInfo.h:62</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml"><div class="ttname"><a href="namespacearm__compute.xhtml">arm_compute</a></div><div class="ttdoc">Copyright (c) 2017-2019 ARM Limited.</div><div class="ttdef"><b>Definition:</b> <a href="00__introduction_8dox_source.xhtml#l00024">00_introduction.dox:24</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_aba8de9eaa10a7cb45c5aa603585d0e4c"><div class="ttname"><a href="namespacearm__compute.xhtml#aba8de9eaa10a7cb45c5aa603585d0e4c">arm_compute::vmlaq_qasymm8</a></div><div class="ttdeci">uint8x16_t vmlaq_qasymm8(qasymm8x16_t vd, float32x4_t vs, float32x4_t vo)</div><div class="ttdoc">Perform a multiply-accumulate on all 16 components of a QASYMM8 vector.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8inl_source.xhtml#l00026">NEAsymm.inl:26</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a0c6e16826765130b480af53eda3635be"><div class="ttname"><a href="namespacearm__compute.xhtml#a0c6e16826765130b480af53eda3635be">arm_compute::qasymm8x8x3_t</a></div><div class="ttdeci">uint8x8x3_t qasymm8x8x3_t</div><div class="ttdoc">8 bit quantized asymmetric vector with 24 elements</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00034">NEAsymm.h:34</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a41d3460ff424e2e21c629b3f27386a8b"><div class="ttname"><a href="namespacearm__compute.xhtml#a41d3460ff424e2e21c629b3f27386a8b">arm_compute::qasymm8x8x4_t</a></div><div class="ttdeci">uint8x8x4_t qasymm8x8x4_t</div><div class="ttdoc">8 bit quantized asymmetric vector with 32 elements</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00035">NEAsymm.h:35</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a70ecc99f2f6b646579b58463dc026671"><div class="ttname"><a href="namespacearm__compute.xhtml#a70ecc99f2f6b646579b58463dc026671">arm_compute::qasymm8x8_t</a></div><div class="ttdeci">uint8x8_t qasymm8x8_t</div><div class="ttdoc">8 bit quantized asymmetric vector with 8 elements</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00032">NEAsymm.h:32</a></div></div> |
| <div class="ttc" id="namespacearm__compute_1_1test_1_1validation_xhtml_acec6d8ad52a28972fa74e071c1a63b6a"><div class="ttname"><a href="namespacearm__compute_1_1test_1_1validation.xhtml#acec6d8ad52a28972fa74e071c1a63b6a">arm_compute::test::validation::scale</a></div><div class="ttdeci">scale</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_o_n_2_pixel_wise_multiplication_8cpp_source.xhtml#l00317">PixelWiseMultiplication.cpp:317</a></div></div> |
| <div class="ttc" id="_n_e_asymm_8inl_xhtml"><div class="ttname"><a href="_n_e_asymm_8inl.xhtml">NEAsymm.inl</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a77e02c19b498c3ff4ee373b0825cfd19"><div class="ttname"><a href="namespacearm__compute.xhtml#a77e02c19b498c3ff4ee373b0825cfd19">arm_compute::rounding_divide_by_pow2</a></div><div class="ttdeci">int32x4_t rounding_divide_by_pow2(int32x4_t x, int32x4_t exponent)</div><div class="ttdoc">Round to the nearest division by a power-of-two using exponent.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_math_8inl_source.xhtml#l00297">NEMath.inl:297</a></div></div> |
| <div class="ttc" id="structarm__compute_1_1_uniform_quantization_info_xhtml_a97bd6c077f3c7769f575b82988b9b668"><div class="ttname"><a href="structarm__compute_1_1_uniform_quantization_info.xhtml#a97bd6c077f3c7769f575b82988b9b668">arm_compute::UniformQuantizationInfo::offset</a></div><div class="ttdeci">int32_t offset</div><div class="ttdef"><b>Definition:</b> <a href="_quantization_info_8h_source.xhtml#l00063">QuantizationInfo.h:63</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a695a8105065167f7e4596d31f23a3573"><div class="ttname"><a href="namespacearm__compute.xhtml#a695a8105065167f7e4596d31f23a3573">arm_compute::finalize_quantization</a></div><div class="ttdeci">uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int32x4_t result_offset_after_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8)</div><div class="ttdoc">Performs final quantization step on 16 elements.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00064">NEAsymm.h:64</a></div></div> |
| <div class="ttc" id="_n_e_math_8h_xhtml"><div class="ttname"><a href="_n_e_math_8h.xhtml">NEMath.h</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a4ff1249e423ae9bfab075f5317c4d004"><div class="ttname"><a href="namespacearm__compute.xhtml#a4ff1249e423ae9bfab075f5317c4d004">arm_compute::vquantize</a></div><div class="ttdeci">uint8x8_t vquantize(const float32x4x2_t &qv, const UniformQuantizationInfo &qi)</div><div class="ttdoc">Quantize a neon vector holding 8 floating point values.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00378">NEAsymm.h:378</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_ac7e1832b8c58f07a98fc8d390b16ac27"><div class="ttname"><a href="namespacearm__compute.xhtml#ac7e1832b8c58f07a98fc8d390b16ac27">arm_compute::qasymm8x16_t</a></div><div class="ttdeci">uint8x16_t qasymm8x16_t</div><div class="ttdoc">8 bit quantized asymmetric vector with 16 elements</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00036">NEAsymm.h:36</a></div></div> |
| <div class="ttc" id="namespacearm__compute_xhtml_a189eb85a92663706760a7afcc09f77a7"><div class="ttname"><a href="namespacearm__compute.xhtml#a189eb85a92663706760a7afcc09f77a7">arm_compute::finalize_quantization_symm</a></div><div class="ttdeci">int8x16_t finalize_quantization_symm(int32x4x4_t &in_s32, const int32x4x4_t &result_fixedpoint_multiplier, const int32x4x4_t &result_shift, const int32x4_t &result_offset_after_shift_s32, const int8x16_t &min_s8, const int8x16_t &max_s8)</div><div class="ttdoc">Performs final quantization step on 16 elements for symmetric quantization.</div><div class="ttdef"><b>Definition:</b> <a href="_n_e_asymm_8h_source.xhtml#l00132">NEAsymm.h:132</a></div></div> |
| </div><!-- fragment --></div><!-- contents --> |
| </div><!-- doc-content --> |
| <!-- start footer part --> |
| <div id="nav-path" class="navpath"><!-- id is needed for treeview function! --> |
| <ul> |
| <li class="navelem"><a class="el" href="dir_214608ef36d61442cb2b0c1c4e9a7def.xhtml">arm_compute</a></li><li class="navelem"><a class="el" href="dir_1fb090f0c6070330bfaccc4236d3ca0d.xhtml">core</a></li><li class="navelem"><a class="el" href="dir_315f6cf1ec0b2df3ae747ff8286a19f5.xhtml">NEON</a></li><li class="navelem"><a class="el" href="_n_e_asymm_8h.xhtml">NEAsymm.h</a></li> |
| <li class="footer">Generated on Wed Jan 22 2020 18:07:36 for Compute Library by |
| <a href="http://www.doxygen.org/index.html"> |
| <img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.15 </li> |
| </ul> |
| </div> |
| </body> |
| </html> |
