-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy patha00104_source.html
210 lines (208 loc) · 19.4 KB
/
a00104_source.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "https://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 1.8.16"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>tesseract: /tesseract/src/arch/intsimdmatrixsse.cpp 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>
<link href="doxygen.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;">
<td id="projectalign" style="padding-left: 0.5em;">
<div id="projectname">tesseract
 <span id="projectnumber">5.0.0-alpha-619-ge9db</span>
</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.8.16 -->
<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('a00104_source.html','');});
/* @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">intsimdmatrixsse.cpp</div> </div>
</div><!--header-->
<div class="contents">
<a href="a00104.html">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">// File: intsindmatrixsse.cpp</span></div>
<div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment">// Description: SSE implementation of 8-bit int SIMD matrix multiply.</span></div>
<div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment">// Author: Ray Smith</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">// (C) Copyright 2017, Google Inc.</span></div>
<div class="line"><a name="l00007"></a><span class="lineno"> 7</span> <span class="comment">// Licensed under the Apache License, Version 2.0 (the "License");</span></div>
<div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <span class="comment">// you may not use this file except in compliance with the License.</span></div>
<div class="line"><a name="l00009"></a><span class="lineno"> 9</span> <span class="comment">// You may obtain a copy of the License at</span></div>
<div class="line"><a name="l00010"></a><span class="lineno"> 10</span> <span class="comment">// http://www.apache.org/licenses/LICENSE-2.0</span></div>
<div class="line"><a name="l00011"></a><span class="lineno"> 11</span> <span class="comment">// Unless required by applicable law or agreed to in writing, software</span></div>
<div class="line"><a name="l00012"></a><span class="lineno"> 12</span> <span class="comment">// distributed under the License is distributed on an "AS IS" BASIS,</span></div>
<div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="comment">// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.</span></div>
<div class="line"><a name="l00014"></a><span class="lineno"> 14</span> <span class="comment">// See the License for the specific language governing permissions and</span></div>
<div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="comment">// limitations under the License.</span></div>
<div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="comment"></span> </div>
<div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="preprocessor">#if !defined(__SSE4_1__)</span></div>
<div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="preprocessor">#error Implementation only for SSE 4.1 capable architectures</span></div>
<div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="preprocessor">#endif</span></div>
<div class="line"><a name="l00021"></a><span class="lineno"> 21</span>  </div>
<div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="preprocessor">#include "<a class="code" href="a00098.html">intsimdmatrix.h</a>"</span></div>
<div class="line"><a name="l00023"></a><span class="lineno"> 23</span>  </div>
<div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#include <cstdint></span></div>
<div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="preprocessor">#include <emmintrin.h></span></div>
<div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#include <smmintrin.h></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="keyword">namespace </span><a class="code" href="a01818.html">tesseract</a> {</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="comment">// Computes and returns the dot product of the n-vectors u and v.</span></div>
<div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="comment">// Uses Intel SSE intrinsics to access the SIMD instruction set.</span></div>
<div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="keyword">static</span> int32_t IntDotProductSSE(<span class="keyword">const</span> int8_t* u, <span class="keyword">const</span> int8_t* v, <span class="keywordtype">int</span> n) {</div>
<div class="line"><a name="l00033"></a><span class="lineno"> 33</span>  <span class="keywordtype">int</span> max_offset = n - 8;</div>
<div class="line"><a name="l00034"></a><span class="lineno"> 34</span>  <span class="keywordtype">int</span> offset = 0;</div>
<div class="line"><a name="l00035"></a><span class="lineno"> 35</span>  <span class="comment">// Accumulate a set of 4 32-bit sums in sum, by loading 8 pairs of 8-bit</span></div>
<div class="line"><a name="l00036"></a><span class="lineno"> 36</span>  <span class="comment">// values, extending to 16 bit, multiplying to make 32 bit results.</span></div>
<div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  int32_t result = 0;</div>
<div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  <span class="keywordflow">if</span> (offset <= max_offset) {</div>
<div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  offset = 8;</div>
<div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  __m128i packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(u));</div>
<div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  __m128i packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(v));</div>
<div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  __m128i sum = _mm_cvtepi8_epi16(packed1);</div>
<div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  packed2 = _mm_cvtepi8_epi16(packed2);</div>
<div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  <span class="comment">// The magic _mm_add_epi16 is perfect here. It multiplies 8 pairs of 16 bit</span></div>
<div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  <span class="comment">// ints to make 32 bit results, which are then horizontally added in pairs</span></div>
<div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  <span class="comment">// to make 4 32 bit results that still fit in a 128 bit register.</span></div>
<div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  sum = _mm_madd_epi16(sum, packed2);</div>
<div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  <span class="keywordflow">while</span> (offset <= max_offset) {</div>
<div class="line"><a name="l00049"></a><span class="lineno"> 49</span>  packed1 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(u + offset));</div>
<div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  packed2 = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(v + offset));</div>
<div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  offset += 8;</div>
<div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  packed1 = _mm_cvtepi8_epi16(packed1);</div>
<div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  packed2 = _mm_cvtepi8_epi16(packed2);</div>
<div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  packed1 = _mm_madd_epi16(packed1, packed2);</div>
<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  sum = _mm_add_epi32(sum, packed1);</div>
<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  }</div>
<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  <span class="comment">// Sum the 4 packed 32 bit sums and extract the low result.</span></div>
<div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  sum = _mm_hadd_epi32(sum, sum);</div>
<div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  sum = _mm_hadd_epi32(sum, sum);</div>
<div class="line"><a name="l00060"></a><span class="lineno"> 60</span>  result = _mm_cvtsi128_si32(sum);</div>
<div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  }</div>
<div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  <span class="keywordflow">while</span> (offset < n) {</div>
<div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  result += u[offset] * v[offset];</div>
<div class="line"><a name="l00064"></a><span class="lineno"> 64</span>  ++offset;</div>
<div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  }</div>
<div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keywordflow">return</span> result;</div>
<div class="line"><a name="l00067"></a><span class="lineno"> 67</span> }</div>
<div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  </div>
<div class="line"><a name="l00069"></a><span class="lineno"> 69</span> <span class="comment">// Computes part of matrix.vector v = Wu. Computes 1 result.</span></div>
<div class="line"><a name="l00070"></a><span class="lineno"> 70</span> <span class="keyword">static</span> <span class="keywordtype">void</span> PartialMatrixDotVector1(<span class="keyword">const</span> int8_t* wi, <span class="keyword">const</span> <span class="keywordtype">double</span>* scales,</div>
<div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  <span class="keyword">const</span> int8_t* u, <span class="keywordtype">int</span> num_in,</div>
<div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  <span class="keywordtype">double</span>* v) {</div>
<div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  <span class="keywordtype">double</span> total = IntDotProductSSE(u, wi, num_in);</div>
<div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="comment">// Add in the bias and correct for integer values.</span></div>
<div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  *v = (total / INT8_MAX + wi[num_in]) * *scales;</div>
<div class="line"><a name="l00076"></a><span class="lineno"> 76</span> }</div>
<div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  </div>
<div class="line"><a name="l00078"></a><span class="lineno"> 78</span> <span class="keyword">static</span> <span class="keywordtype">void</span> matrixDotVector(<span class="keywordtype">int</span> dim1, <span class="keywordtype">int</span> dim2, <span class="keyword">const</span> int8_t* wi,</div>
<div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="keyword">const</span> <span class="keywordtype">double</span>* scales, <span class="keyword">const</span> int8_t* u, <span class="keywordtype">double</span>* v) {</div>
<div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> num_out = dim1;</div>
<div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  <span class="keyword">const</span> <span class="keywordtype">int</span> num_in = dim2 - 1;</div>
<div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="keywordtype">int</span> output = 0;</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="keywordflow">for</span> (; output < num_out; output++) {</div>
<div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  PartialMatrixDotVector1(wi, scales, u, num_in, v);</div>
<div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  wi += dim2;</div>
<div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  scales++;</div>
<div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  v++;</div>
<div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  }</div>
<div class="line"><a name="l00090"></a><span class="lineno"> 90</span> }</div>
<div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  </div>
<div class="line"><a name="l00092"></a><span class="lineno"> 92</span> <span class="keyword">const</span> IntSimdMatrix <a class="code" href="a02570.html#ad7aed52d7915512a59818a5efda68c1e">IntSimdMatrix::intSimdMatrixSSE</a> = {</div>
<div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  matrixDotVector,</div>
<div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  <span class="comment">// Number of 32 bit outputs held in each register.</span></div>
<div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  1,</div>
<div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  <span class="comment">// Maximum number of registers that we will use to hold outputs.</span></div>
<div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  1,</div>
<div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  <span class="comment">// Number of 8 bit inputs in the inputs register.</span></div>
<div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  1,</div>
<div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  <span class="comment">// Number of inputs in each weight group.</span></div>
<div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  1</div>
<div class="line"><a name="l00102"></a><span class="lineno"> 102</span> };</div>
<div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  </div>
<div class="line"><a name="l00104"></a><span class="lineno"> 104</span> } <span class="comment">// namespace tesseract.</span></div>
</div><!-- fragment --></div><!-- contents -->
</div><!-- doc-content -->
<div class="ttc" id="aa01818_html"><div class="ttname"><a href="a01818.html">tesseract</a></div><div class="ttdef"><b>Definition:</b> <a href="a00005_source.html#l00065">baseapi.h:65</a></div></div>
<div class="ttc" id="aa02570_html_ad7aed52d7915512a59818a5efda68c1e"><div class="ttname"><a href="a02570.html#ad7aed52d7915512a59818a5efda68c1e">tesseract::IntSimdMatrix::intSimdMatrixSSE</a></div><div class="ttdeci">static const IntSimdMatrix intSimdMatrixSSE</div><div class="ttdef"><b>Definition:</b> <a href="a00098_source.html#l00118">intsimdmatrix.h:118</a></div></div>
<div class="ttc" id="aa00098_html"><div class="ttname"><a href="a00098.html">intsimdmatrix.h</a></div></div>
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<li class="navelem"><a class="el" href="dir_68267d1309a1af8e8297ef4c3efbcdba.html">src</a></li><li class="navelem"><a class="el" href="dir_70d5e250c66d001b18da37689179cade.html">arch</a></li><li class="navelem"><a class="el" href="a00104.html">intsimdmatrixsse.cpp</a></li>
<li class="footer">Generated on Thu Jan 30 2020 14:22:19 for tesseract by
<a href="http://www.doxygen.org/index.html">
<img class="footer" src="doxygen.png" alt="doxygen"/></a> 1.8.16 </li>
</ul>
</div>
</body>
</html>