843 lines
127 KiB
HTML
843 lines
127 KiB
HTML
<!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 1.8.5"/>
|
|
<title>Faiss: IVFPQ.cu 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="search/search.css" rel="stylesheet" type="text/css"/>
|
|
<script type="text/javascript" src="search/search.js"></script>
|
|
<script type="text/javascript">
|
|
$(document).ready(function() { searchBox.OnSelectItem(0); });
|
|
</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 style="padding-left: 0.5em;">
|
|
<div id="projectname">Faiss
|
|
</div>
|
|
</td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
</div>
|
|
<!-- end header part -->
|
|
<!-- Generated by Doxygen 1.8.5 -->
|
|
<script type="text/javascript">
|
|
var searchBox = new SearchBox("searchBox", "search",false,'Search');
|
|
</script>
|
|
<div id="navrow1" class="tabs">
|
|
<ul class="tablist">
|
|
<li><a href="index.html"><span>Main Page</span></a></li>
|
|
<li><a href="namespaces.html"><span>Namespaces</span></a></li>
|
|
<li><a href="annotated.html"><span>Classes</span></a></li>
|
|
<li class="current"><a href="files.html"><span>Files</span></a></li>
|
|
<li>
|
|
<div id="MSearchBox" class="MSearchBoxInactive">
|
|
<span class="left">
|
|
<img id="MSearchSelect" src="search/mag_sel.png"
|
|
onmouseover="return searchBox.OnSearchSelectShow()"
|
|
onmouseout="return searchBox.OnSearchSelectHide()"
|
|
alt=""/>
|
|
<input type="text" id="MSearchField" value="Search" accesskey="S"
|
|
onfocus="searchBox.OnSearchFieldFocus(true)"
|
|
onblur="searchBox.OnSearchFieldFocus(false)"
|
|
onkeyup="searchBox.OnSearchFieldChange(event)"/>
|
|
</span><span class="right">
|
|
<a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.png" alt=""/></a>
|
|
</span>
|
|
</div>
|
|
</li>
|
|
</ul>
|
|
</div>
|
|
<div id="navrow2" class="tabs2">
|
|
<ul class="tablist">
|
|
<li><a href="files.html"><span>File List</span></a></li>
|
|
</ul>
|
|
</div>
|
|
<!-- window showing the filter options -->
|
|
<div id="MSearchSelectWindow"
|
|
onmouseover="return searchBox.OnSearchSelectShow()"
|
|
onmouseout="return searchBox.OnSearchSelectHide()"
|
|
onkeydown="return searchBox.OnSearchSelectKey(event)">
|
|
<a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(0)"><span class="SelectionMark"> </span>All</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(1)"><span class="SelectionMark"> </span>Classes</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(2)"><span class="SelectionMark"> </span>Namespaces</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(3)"><span class="SelectionMark"> </span>Functions</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(4)"><span class="SelectionMark"> </span>Variables</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(5)"><span class="SelectionMark"> </span>Typedefs</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(6)"><span class="SelectionMark"> </span>Enumerations</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(7)"><span class="SelectionMark"> </span>Enumerator</a><a class="SelectItem" href="javascript:void(0)" onclick="searchBox.OnSelectItem(8)"><span class="SelectionMark"> </span>Friends</a></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 id="nav-path" class="navpath">
|
|
<ul>
|
|
<li class="navelem"><a class="el" href="dir_6b3ae6988449b0834e9596fad5d75199.html">gpu</a></li><li class="navelem"><a class="el" href="dir_49d1182a3b8dfb62757c53ae905481ad.html">impl</a></li> </ul>
|
|
</div>
|
|
</div><!-- top -->
|
|
<div class="header">
|
|
<div class="headertitle">
|
|
<div class="title">IVFPQ.cu</div> </div>
|
|
</div><!--header-->
|
|
<div class="contents">
|
|
<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">/**</span></div>
|
|
<div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> * Copyright (c) 2015-present, Facebook, Inc.</span></div>
|
|
<div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> * All rights reserved.</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"> * This source code is licensed under the CC-by-NC license found in the</span></div>
|
|
<div class="line"><a name="l00007"></a><span class="lineno"> 7</span> <span class="comment"> * LICENSE file in the root directory of this source tree.</span></div>
|
|
<div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <span class="comment"> */</span></div>
|
|
<div class="line"><a name="l00009"></a><span class="lineno"> 9</span> </div>
|
|
<div class="line"><a name="l00010"></a><span class="lineno"> 10</span> <span class="comment">// Copyright 2004-present Facebook. All Rights Reserved.</span></div>
|
|
<div class="line"><a name="l00011"></a><span class="lineno"> 11</span> </div>
|
|
<div class="line"><a name="l00012"></a><span class="lineno"> 12</span> <span class="preprocessor">#include "IVFPQ.cuh"</span></div>
|
|
<div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="preprocessor">#include "../GpuResources.h"</span></div>
|
|
<div class="line"><a name="l00014"></a><span class="lineno"> 14</span> <span class="preprocessor">#include "BroadcastSum.cuh"</span></div>
|
|
<div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="preprocessor">#include "Distance.cuh"</span></div>
|
|
<div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="preprocessor">#include "FlatIndex.cuh"</span></div>
|
|
<div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="preprocessor">#include "InvertedListAppend.cuh"</span></div>
|
|
<div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="preprocessor">#include "L2Norm.cuh"</span></div>
|
|
<div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="preprocessor">#include "PQCodeDistances.cuh"</span></div>
|
|
<div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="preprocessor">#include "PQScanMultiPassNoPrecomputed.cuh"</span></div>
|
|
<div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="preprocessor">#include "PQScanMultiPassPrecomputed.cuh"</span></div>
|
|
<div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="preprocessor">#include "RemapIndices.h"</span></div>
|
|
<div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="preprocessor">#include "VectorResidual.cuh"</span></div>
|
|
<div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#include "../utils/DeviceDefs.cuh"</span></div>
|
|
<div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="preprocessor">#include "../utils/DeviceUtils.h"</span></div>
|
|
<div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="preprocessor">#include "../utils/HostTensor.cuh"</span></div>
|
|
<div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="preprocessor">#include "../utils/MatrixMult.cuh"</span></div>
|
|
<div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="preprocessor">#include "../utils/NoTypeTensor.cuh"</span></div>
|
|
<div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="preprocessor">#include "../utils/Transpose.cuh"</span></div>
|
|
<div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="preprocessor">#include <limits></span></div>
|
|
<div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="preprocessor">#include <thrust/host_vector.h></span></div>
|
|
<div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="preprocessor">#include <unordered_map></span></div>
|
|
<div class="line"><a name="l00033"></a><span class="lineno"> 33</span> </div>
|
|
<div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="keyword">namespace </span>faiss { <span class="keyword">namespace </span>gpu {</div>
|
|
<div class="line"><a name="l00035"></a><span class="lineno"> 35</span> </div>
|
|
<div class="line"><a name="l00036"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#acdcc64f9b72e661d39528d5d67e54536"> 36</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#acdcc64f9b72e661d39528d5d67e54536">IVFPQ::IVFPQ</a>(<a class="code" href="classfaiss_1_1gpu_1_1GpuResources.html">GpuResources</a>* resources,</div>
|
|
<div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  <a class="code" href="classfaiss_1_1gpu_1_1FlatIndex.html">FlatIndex</a>* quantizer,</div>
|
|
<div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  <span class="keywordtype">int</span> numSubQuantizers,</div>
|
|
<div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  <span class="keywordtype">int</span> bitsPerSubQuantizer,</div>
|
|
<div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  <span class="keywordtype">float</span>* pqCentroidData,</div>
|
|
<div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  IndicesOptions indicesOptions,</div>
|
|
<div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  <span class="keywordtype">bool</span> useFloat16LookupTables) :</div>
|
|
<div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html">IVFBase</a>(resources,</div>
|
|
<div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  quantizer,</div>
|
|
<div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  numSubQuantizers,</div>
|
|
<div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  indicesOptions),</div>
|
|
<div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  numSubQuantizers_(numSubQuantizers),</div>
|
|
<div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  bitsPerSubQuantizer_(bitsPerSubQuantizer),</div>
|
|
<div class="line"><a name="l00049"></a><span class="lineno"> 49</span>  numSubQuantizerCodes_(utils::pow2(bitsPerSubQuantizer_)),</div>
|
|
<div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  dimPerSubQuantizer_(dim_ / numSubQuantizers),</div>
|
|
<div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  precomputedCodes_(false),</div>
|
|
<div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  useFloat16LookupTables_(useFloat16LookupTables) {</div>
|
|
<div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  FAISS_ASSERT(pqCentroidData);</div>
|
|
<div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div>
|
|
<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  FAISS_ASSERT(bitsPerSubQuantizer_ <= 8);</div>
|
|
<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  FAISS_ASSERT(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#aba3e3cfa469e5187f2d553fff10e0250">dim_</a> % numSubQuantizers_ == 0);</div>
|
|
<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  FAISS_ASSERT(<a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#adb58eeacdceb0e0fde1820ca7f116e05">isSupportedPQCodeLength</a>(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>));</div>
|
|
<div class="line"><a name="l00058"></a><span class="lineno"> 58</span> </div>
|
|
<div class="line"><a name="l00059"></a><span class="lineno"> 59</span> <span class="preprocessor">#ifndef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00060"></a><span class="lineno"> 60</span> <span class="preprocessor"></span> FAISS_ASSERT(!useFloat16LookupTables_);</div>
|
|
<div class="line"><a name="l00061"></a><span class="lineno"> 61</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00062"></a><span class="lineno"> 62</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  setPQCentroids_(pqCentroidData);</div>
|
|
<div class="line"><a name="l00064"></a><span class="lineno"> 64</span> }</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> IVFPQ::~IVFPQ() {</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> </div>
|
|
<div class="line"><a name="l00070"></a><span class="lineno"> 70</span> <span class="keywordtype">bool</span></div>
|
|
<div class="line"><a name="l00071"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#adb58eeacdceb0e0fde1820ca7f116e05"> 71</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#adb58eeacdceb0e0fde1820ca7f116e05">IVFPQ::isSupportedPQCodeLength</a>(<span class="keywordtype">int</span> size) {</div>
|
|
<div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  <span class="keywordflow">switch</span> (size) {</div>
|
|
<div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  <span class="keywordflow">case</span> 1:</div>
|
|
<div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="keywordflow">case</span> 2:</div>
|
|
<div class="line"><a name="l00075"></a><span class="lineno"> 75</span>  <span class="keywordflow">case</span> 3:</div>
|
|
<div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  <span class="keywordflow">case</span> 4:</div>
|
|
<div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  <span class="keywordflow">case</span> 8:</div>
|
|
<div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="keywordflow">case</span> 12:</div>
|
|
<div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="keywordflow">case</span> 16:</div>
|
|
<div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  <span class="keywordflow">case</span> 20:</div>
|
|
<div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  <span class="keywordflow">case</span> 24:</div>
|
|
<div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="keywordflow">case</span> 28:</div>
|
|
<div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  <span class="keywordflow">case</span> 32:</div>
|
|
<div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  <span class="keywordflow">case</span> 40:</div>
|
|
<div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="keywordflow">case</span> 48:</div>
|
|
<div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="keywordflow">case</span> 56: <span class="comment">// only supported with float16</span></div>
|
|
<div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  <span class="keywordflow">case</span> 64: <span class="comment">// only supported with float16</span></div>
|
|
<div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  <span class="keywordflow">case</span> 96: <span class="comment">// only supported with float16</span></div>
|
|
<div class="line"><a name="l00089"></a><span class="lineno"> 89</span>  <span class="keywordflow">return</span> <span class="keyword">true</span>;</div>
|
|
<div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  <span class="keywordflow">default</span>:</div>
|
|
<div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  <span class="keywordflow">return</span> <span class="keyword">false</span>;</div>
|
|
<div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  }</div>
|
|
<div class="line"><a name="l00093"></a><span class="lineno"> 93</span> }</div>
|
|
<div class="line"><a name="l00094"></a><span class="lineno"> 94</span> </div>
|
|
<div class="line"><a name="l00095"></a><span class="lineno"> 95</span> <span class="keywordtype">bool</span></div>
|
|
<div class="line"><a name="l00096"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#a0eedf0295ad73125ee1254173a176674"> 96</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#a0eedf0295ad73125ee1254173a176674">IVFPQ::isSupportedNoPrecomputedSubDimSize</a>(<span class="keywordtype">int</span> dims) {</div>
|
|
<div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  <span class="keywordflow">return</span> faiss::gpu::isSupportedNoPrecomputedSubDimSize(dims);</div>
|
|
<div class="line"><a name="l00098"></a><span class="lineno"> 98</span> }</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> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00101"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#adcee5dbf48c3cb6b8a67f5f392e155fd"> 101</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#adcee5dbf48c3cb6b8a67f5f392e155fd">IVFPQ::setPrecomputedCodes</a>(<span class="keywordtype">bool</span> enable) {</div>
|
|
<div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  <span class="keywordflow">if</span> (precomputedCodes_ != enable) {</div>
|
|
<div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  precomputedCodes_ = enable;</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>  <span class="keywordflow">if</span> (precomputedCodes_) {</div>
|
|
<div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  precomputeCodes_();</div>
|
|
<div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00108"></a><span class="lineno"> 108</span>  <span class="comment">// Clear out old precomputed code data</span></div>
|
|
<div class="line"><a name="l00109"></a><span class="lineno"> 109</span>  precomputedCode_ = std::move(<a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 3, true></a>());</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> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00112"></a><span class="lineno"> 112</span> <span class="preprocessor"></span> precomputedCodeHalf_ = std::move(<a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<half, 3, true></a>());</div>
|
|
<div class="line"><a name="l00113"></a><span class="lineno"> 113</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00114"></a><span class="lineno"> 114</span> <span class="preprocessor"></span> }</div>
|
|
<div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  }</div>
|
|
<div class="line"><a name="l00116"></a><span class="lineno"> 116</span> }</div>
|
|
<div class="line"><a name="l00117"></a><span class="lineno"> 117</span> </div>
|
|
<div class="line"><a name="l00118"></a><span class="lineno"> 118</span> <span class="keywordtype">int</span></div>
|
|
<div class="line"><a name="l00119"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#ab1e07b04b25569cc58c5f3f033f4dab3"> 119</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#ab1e07b04b25569cc58c5f3f033f4dab3">IVFPQ::classifyAndAddVectors</a>(<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 2, true></a>& vecs,</div>
|
|
<div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<long, 1, true></a>& indices) {</div>
|
|
<div class="line"><a name="l00121"></a><span class="lineno"> 121</span>  FAISS_ASSERT(vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0) == indices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0));</div>
|
|
<div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  FAISS_ASSERT(vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(1) == <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#aba3e3cfa469e5187f2d553fff10e0250">dim_</a>);</div>
|
|
<div class="line"><a name="l00123"></a><span class="lineno"> 123</span> </div>
|
|
<div class="line"><a name="l00124"></a><span class="lineno"> 124</span>  FAISS_ASSERT(!<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>->getUseFloat16());</div>
|
|
<div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="keyword">auto</span>& coarseCentroids = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>-><a class="code" href="classfaiss_1_1gpu_1_1FlatIndex.html#a12058744ffb3fbcbb047872449269c06">getVectorsFloat32Ref</a>();</div>
|
|
<div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  <span class="keyword">auto</span>& mem = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getMemoryManagerCurrentDevice();</div>
|
|
<div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  <span class="keyword">auto</span> stream = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice();</div>
|
|
<div class="line"><a name="l00128"></a><span class="lineno"> 128</span> </div>
|
|
<div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  <span class="comment">// Number of valid vectors that we actually add; we return this</span></div>
|
|
<div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordtype">int</span> numAdded = 0;</div>
|
|
<div class="line"><a name="l00131"></a><span class="lineno"> 131</span> </div>
|
|
<div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  <span class="comment">// We don't actually need this</span></div>
|
|
<div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 2, true></a> listDistance(mem, {vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), 1}, stream);</div>
|
|
<div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  <span class="comment">// We use this</span></div>
|
|
<div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 2, true></a> listIds2d(mem, {vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), 1}, stream);</div>
|
|
<div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="keyword">auto</span> listIds = listIds2d.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">view</a><1>({vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0)});</div>
|
|
<div class="line"><a name="l00137"></a><span class="lineno"> 137</span> </div>
|
|
<div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>->query(vecs, 1, listDistance, listIds2d, <span class="keyword">false</span>);</div>
|
|
<div class="line"><a name="l00139"></a><span class="lineno"> 139</span> </div>
|
|
<div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="comment">// Copy the lists that we wish to append to back to the CPU</span></div>
|
|
<div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <span class="comment">// FIXME: really this can be into pinned memory and a true async</span></div>
|
|
<div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  <span class="comment">// copy on a different stream; we can start the copy early, but it's</span></div>
|
|
<div class="line"><a name="l00143"></a><span class="lineno"> 143</span>  <span class="comment">// tiny</span></div>
|
|
<div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <a class="code" href="classfaiss_1_1gpu_1_1HostTensor.html">HostTensor<int, 1, true></a> listIdsHost(listIds, stream);</div>
|
|
<div class="line"><a name="l00145"></a><span class="lineno"> 145</span> </div>
|
|
<div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="comment">// Calculate the residual for each closest centroid</span></div>
|
|
<div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 2, true></a> residuals(</div>
|
|
<div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  mem, {vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), vecs.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(1)}, stream);</div>
|
|
<div class="line"><a name="l00149"></a><span class="lineno"> 149</span> </div>
|
|
<div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  runCalcResidual(vecs, coarseCentroids, listIds, residuals, stream);</div>
|
|
<div class="line"><a name="l00151"></a><span class="lineno"> 151</span> </div>
|
|
<div class="line"><a name="l00152"></a><span class="lineno"> 152</span>  <span class="comment">// Residuals are in the form</span></div>
|
|
<div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  <span class="comment">// (vec x numSubQuantizer x dimPerSubQuantizer)</span></div>
|
|
<div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="comment">// transpose to</span></div>
|
|
<div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="comment">// (numSubQuantizer x vec x dimPerSubQuantizer)</span></div>
|
|
<div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  <span class="keyword">auto</span> residualsView = residuals.view<3>(</div>
|
|
<div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  {residuals.getSize(0), numSubQuantizers_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00158"></a><span class="lineno"> 158</span> </div>
|
|
<div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 3, true></a> residualsTranspose(</div>
|
|
<div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  mem,</div>
|
|
<div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  {numSubQuantizers_, residuals.getSize(0), dimPerSubQuantizer_},</div>
|
|
<div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  stream);</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>  runTransposeAny(residualsView, 0, 1, residualsTranspose, stream);</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">// Get the product quantizer centroids in the form</span></div>
|
|
<div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="comment">// (numSubQuantizer x numSubQuantizerCodes x dimPerSubQuantizer)</span></div>
|
|
<div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  <span class="comment">// which is pqCentroidsMiddleCode_</span></div>
|
|
<div class="line"><a name="l00169"></a><span class="lineno"> 169</span> </div>
|
|
<div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  <span class="comment">// We now have a batch operation to find the top-1 distances:</span></div>
|
|
<div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  <span class="comment">// batch size: numSubQuantizer</span></div>
|
|
<div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  <span class="comment">// centroids: (numSubQuantizerCodes x dimPerSubQuantizer)</span></div>
|
|
<div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <span class="comment">// residuals: (vec x dimPerSubQuantizer)</span></div>
|
|
<div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  <span class="comment">// => (numSubQuantizer x vec x 1)</span></div>
|
|
<div class="line"><a name="l00175"></a><span class="lineno"> 175</span> </div>
|
|
<div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 3, true></a> closestSubQDistance(</div>
|
|
<div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  mem, {numSubQuantizers_, residuals.getSize(0), 1}, stream);</div>
|
|
<div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 3, true></a> closestSubQIndex(</div>
|
|
<div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  mem, {numSubQuantizers_, residuals.getSize(0), 1}, stream);</div>
|
|
<div class="line"><a name="l00180"></a><span class="lineno"> 180</span> </div>
|
|
<div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> subQ = 0; subQ < numSubQuantizers_; ++subQ) {</div>
|
|
<div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  <span class="keyword">auto</span> closestSubQDistanceView = closestSubQDistance[subQ].<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">view</a>();</div>
|
|
<div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <span class="keyword">auto</span> closestSubQIndexView = closestSubQIndex[subQ].view();</div>
|
|
<div class="line"><a name="l00184"></a><span class="lineno"> 184</span> </div>
|
|
<div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  <span class="keyword">auto</span> pqCentroidsMiddleCodeView = pqCentroidsMiddleCode_[subQ].<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">view</a>();</div>
|
|
<div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="keyword">auto</span> residualsTransposeView = residualsTranspose[subQ].view();</div>
|
|
<div class="line"><a name="l00187"></a><span class="lineno"> 187</span> </div>
|
|
<div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  runL2Distance(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>,</div>
|
|
<div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  pqCentroidsMiddleCodeView,</div>
|
|
<div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  <span class="keyword">nullptr</span>, <span class="comment">// no precomputed norms</span></div>
|
|
<div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  residualsTransposeView,</div>
|
|
<div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  1,</div>
|
|
<div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  closestSubQDistanceView,</div>
|
|
<div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  closestSubQIndexView,</div>
|
|
<div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  <span class="comment">// We don't care about distances</span></div>
|
|
<div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <span class="keyword">true</span>,</div>
|
|
<div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  <span class="comment">// Much larger tile size, since these vectors are a</span></div>
|
|
<div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  <span class="comment">// lot smaller than query vectors</span></div>
|
|
<div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  1024);</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> </div>
|
|
<div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  <span class="comment">// Now, we have the nearest sub-q centroid for each slice of the</span></div>
|
|
<div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  <span class="comment">// residual vector.</span></div>
|
|
<div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  <span class="keyword">auto</span> closestSubQIndexView = closestSubQIndex.view<2>(</div>
|
|
<div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  {numSubQuantizers_, residuals.getSize(0)});</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">// Transpose this for easy use</span></div>
|
|
<div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 2, true></a> encodings(</div>
|
|
<div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  mem, {residuals.getSize(0), numSubQuantizers_}, stream);</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>  runTransposeAny(closestSubQIndexView, 0, 1, encodings, stream);</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>  <span class="comment">// Now we add the encoded vectors to the individual lists</span></div>
|
|
<div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="comment">// First, make sure that there is space available for adding the new</span></div>
|
|
<div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  <span class="comment">// encoded vectors and indices</span></div>
|
|
<div class="line"><a name="l00216"></a><span class="lineno"> 216</span> </div>
|
|
<div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  <span class="comment">// list id -> # being added</span></div>
|
|
<div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  std::unordered_map<int, int> assignCounts;</div>
|
|
<div class="line"><a name="l00219"></a><span class="lineno"> 219</span> </div>
|
|
<div class="line"><a name="l00220"></a><span class="lineno"> 220</span>  <span class="comment">// vector id -> offset in list</span></div>
|
|
<div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  <span class="comment">// (we already have vector id -> list id in listIds)</span></div>
|
|
<div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  <a class="code" href="classfaiss_1_1gpu_1_1HostTensor.html">HostTensor<int, 1, true></a> listOffsetHost({listIdsHost.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0)});</div>
|
|
<div class="line"><a name="l00223"></a><span class="lineno"> 223</span> </div>
|
|
<div class="line"><a name="l00224"></a><span class="lineno"> 224</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < listIdsHost.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0); ++i) {</div>
|
|
<div class="line"><a name="l00225"></a><span class="lineno"> 225</span>  <span class="keywordtype">int</span> listId = listIdsHost[i];</div>
|
|
<div class="line"><a name="l00226"></a><span class="lineno"> 226</span> </div>
|
|
<div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  <span class="comment">// Add vector could be invalid (contains NaNs etc)</span></div>
|
|
<div class="line"><a name="l00228"></a><span class="lineno"> 228</span>  <span class="keywordflow">if</span> (listId < 0) {</div>
|
|
<div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  listOffsetHost[i] = -1;</div>
|
|
<div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  <span class="keywordflow">continue</span>;</div>
|
|
<div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  }</div>
|
|
<div class="line"><a name="l00232"></a><span class="lineno"> 232</span> </div>
|
|
<div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  FAISS_ASSERT(listId < <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#accc4d96c14643e5f471220cb1e92ac70">numLists_</a>);</div>
|
|
<div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  ++numAdded;</div>
|
|
<div class="line"><a name="l00235"></a><span class="lineno"> 235</span> </div>
|
|
<div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  <span class="keywordtype">int</span> offset = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>[listId]->size() / <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>;</div>
|
|
<div class="line"><a name="l00237"></a><span class="lineno"> 237</span> </div>
|
|
<div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  <span class="keyword">auto</span> it = assignCounts.find(listId);</div>
|
|
<div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  <span class="keywordflow">if</span> (it != assignCounts.end()) {</div>
|
|
<div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  offset += it->second;</div>
|
|
<div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  it->second++;</div>
|
|
<div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  assignCounts[listId] = 1;</div>
|
|
<div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  }</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>  listOffsetHost[i] = offset;</div>
|
|
<div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  }</div>
|
|
<div class="line"><a name="l00248"></a><span class="lineno"> 248</span> </div>
|
|
<div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  <span class="comment">// If we didn't add anything (all invalid vectors), no need to</span></div>
|
|
<div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="comment">// continue</span></div>
|
|
<div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  <span class="keywordflow">if</span> (numAdded == 0) {</div>
|
|
<div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  <span class="keywordflow">return</span> 0;</div>
|
|
<div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  }</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">// We need to resize the data structures for the inverted lists on</span></div>
|
|
<div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  <span class="comment">// the GPUs, which means that they might need reallocation, which</span></div>
|
|
<div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  <span class="comment">// means that their base address may change. Figure out the new base</span></div>
|
|
<div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  <span class="comment">// addresses, and update those in a batch on the device</span></div>
|
|
<div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  {</div>
|
|
<div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="comment">// Resize all of the lists that we are appending to</span></div>
|
|
<div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  <span class="keywordflow">for</span> (<span class="keyword">auto</span>& counts : assignCounts) {</div>
|
|
<div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  <span class="keyword">auto</span>& codes = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>[counts.first];</div>
|
|
<div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  codes->resize(codes->size() + counts.second * <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>,</div>
|
|
<div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  stream);</div>
|
|
<div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  <span class="keywordtype">int</span> newNumVecs = (int) (codes->size() / <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>);</div>
|
|
<div class="line"><a name="l00266"></a><span class="lineno"> 266</span> </div>
|
|
<div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  <span class="keyword">auto</span>& indices = deviceListIndices_[counts.first];</div>
|
|
<div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  <span class="keywordflow">if</span> ((<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_32_BIT) ||</div>
|
|
<div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_64_BIT)) {</div>
|
|
<div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  <span class="keywordtype">size_t</span> indexSize =</div>
|
|
<div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_32_BIT) ? <span class="keyword">sizeof</span>(<span class="keywordtype">int</span>) : <span class="keyword">sizeof</span>(long);</div>
|
|
<div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div>
|
|
<div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  indices->resize(indices->size() + counts.second * indexSize, stream);</div>
|
|
<div class="line"><a name="l00274"></a><span class="lineno"> 274</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_CPU) {</div>
|
|
<div class="line"><a name="l00275"></a><span class="lineno"> 275</span>  <span class="comment">// indices are stored on the CPU side</span></div>
|
|
<div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  FAISS_ASSERT(counts.first < <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">listOffsetToUserIndex_</a>.size());</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="keyword">auto</span>& userIndices = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">listOffsetToUserIndex_</a>[counts.first];</div>
|
|
<div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  userIndices.resize(newNumVecs);</div>
|
|
<div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  <span class="comment">// indices are not stored on the GPU or CPU side</span></div>
|
|
<div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  FAISS_ASSERT(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_IVF);</div>
|
|
<div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  }</div>
|
|
<div class="line"><a name="l00284"></a><span class="lineno"> 284</span> </div>
|
|
<div class="line"><a name="l00285"></a><span class="lineno"> 285</span>  <span class="comment">// This is used by the multi-pass query to decide how much scratch</span></div>
|
|
<div class="line"><a name="l00286"></a><span class="lineno"> 286</span>  <span class="comment">// space to allocate for intermediate results</span></div>
|
|
<div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a> = std::max(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a>, newNumVecs);</div>
|
|
<div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  }</div>
|
|
<div class="line"><a name="l00289"></a><span class="lineno"> 289</span> </div>
|
|
<div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  <span class="comment">// Update all pointers and sizes on the device for lists that we</span></div>
|
|
<div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <span class="comment">// appended to</span></div>
|
|
<div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  {</div>
|
|
<div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  std::vector<int> listIds(assignCounts.size());</div>
|
|
<div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <span class="keywordtype">int</span> i = 0;</div>
|
|
<div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  <span class="keywordflow">for</span> (<span class="keyword">auto</span>& counts : assignCounts) {</div>
|
|
<div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  listIds[i++] = counts.first;</div>
|
|
<div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  }</div>
|
|
<div class="line"><a name="l00298"></a><span class="lineno"> 298</span> </div>
|
|
<div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#acc695610c9513952b8d234dc0db78e5c">updateDeviceListInfo_</a>(listIds, stream);</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>  }</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">// If we're maintaining the indices on the CPU side, update our</span></div>
|
|
<div class="line"><a name="l00304"></a><span class="lineno"> 304</span>  <span class="comment">// map. We already resized our map above.</span></div>
|
|
<div class="line"><a name="l00305"></a><span class="lineno"> 305</span>  <span class="keywordflow">if</span> (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_CPU) {</div>
|
|
<div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  <span class="comment">// We need to maintain the indices on the CPU side</span></div>
|
|
<div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  <a class="code" href="classfaiss_1_1gpu_1_1HostTensor.html">HostTensor<long, 1, true></a> hostIndices(indices, stream);</div>
|
|
<div class="line"><a name="l00308"></a><span class="lineno"> 308</span> </div>
|
|
<div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < hostIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0); ++i) {</div>
|
|
<div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  <span class="keywordtype">int</span> listId = listIdsHost[i];</div>
|
|
<div class="line"><a name="l00311"></a><span class="lineno"> 311</span> </div>
|
|
<div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  <span class="comment">// Add vector could be invalid (contains NaNs etc)</span></div>
|
|
<div class="line"><a name="l00313"></a><span class="lineno"> 313</span>  <span class="keywordflow">if</span> (listId < 0) {</div>
|
|
<div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  <span class="keywordflow">continue</span>;</div>
|
|
<div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  }</div>
|
|
<div class="line"><a name="l00316"></a><span class="lineno"> 316</span> </div>
|
|
<div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  <span class="keywordtype">int</span> offset = listOffsetHost[i];</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>  FAISS_ASSERT(listId < <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">listOffsetToUserIndex_</a>.size());</div>
|
|
<div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  <span class="keyword">auto</span>& userIndices = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">listOffsetToUserIndex_</a>[listId];</div>
|
|
<div class="line"><a name="l00321"></a><span class="lineno"> 321</span> </div>
|
|
<div class="line"><a name="l00322"></a><span class="lineno"> 322</span>  FAISS_ASSERT(offset < userIndices.size());</div>
|
|
<div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  userIndices[offset] = hostIndices[i];</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>  }</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">// We similarly need to actually append the new encoded vectors</span></div>
|
|
<div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  {</div>
|
|
<div class="line"><a name="l00329"></a><span class="lineno"> 329</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 1, true></a> listOffset(mem, listOffsetHost, stream);</div>
|
|
<div class="line"><a name="l00330"></a><span class="lineno"> 330</span> </div>
|
|
<div class="line"><a name="l00331"></a><span class="lineno"> 331</span>  <span class="comment">// This kernel will handle appending each encoded vector + index to</span></div>
|
|
<div class="line"><a name="l00332"></a><span class="lineno"> 332</span>  <span class="comment">// the appropriate list</span></div>
|
|
<div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  runIVFPQInvertedListAppend(listIds,</div>
|
|
<div class="line"><a name="l00334"></a><span class="lineno"> 334</span>  listOffset,</div>
|
|
<div class="line"><a name="l00335"></a><span class="lineno"> 335</span>  encodings,</div>
|
|
<div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  indices,</div>
|
|
<div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a96240a08b42bd1913e2286d7d514fc56">deviceListDataPointers_</a>,</div>
|
|
<div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a9aedcf0e6a20b908980ae96d73461f4c">deviceListIndexPointers_</a>,</div>
|
|
<div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a>,</div>
|
|
<div class="line"><a name="l00340"></a><span class="lineno"> 340</span>  stream);</div>
|
|
<div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  }</div>
|
|
<div class="line"><a name="l00342"></a><span class="lineno"> 342</span> </div>
|
|
<div class="line"><a name="l00343"></a><span class="lineno"> 343</span>  <span class="keywordflow">return</span> numAdded;</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="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00347"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#a9992b38226dc8f92ca2691582fabb675"> 347</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#a9992b38226dc8f92ca2691582fabb675">IVFPQ::addCodeVectorsFromCpu</a>(<span class="keywordtype">int</span> listId,</div>
|
|
<div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  <span class="keyword">const</span> <span class="keywordtype">void</span>* codes,</div>
|
|
<div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  <span class="keyword">const</span> <span class="keywordtype">long</span>* indices,</div>
|
|
<div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  <span class="keywordtype">size_t</span> numVecs) {</div>
|
|
<div class="line"><a name="l00351"></a><span class="lineno"> 351</span>  <span class="comment">// This list must already exist</span></div>
|
|
<div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  FAISS_ASSERT(listId < <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>.size());</div>
|
|
<div class="line"><a name="l00353"></a><span class="lineno"> 353</span>  <span class="keyword">auto</span> stream = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice();</div>
|
|
<div class="line"><a name="l00354"></a><span class="lineno"> 354</span> </div>
|
|
<div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  <span class="comment">// If there's nothing to add, then there's nothing we have to do</span></div>
|
|
<div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  <span class="keywordflow">if</span> (numVecs == 0) {</div>
|
|
<div class="line"><a name="l00357"></a><span class="lineno"> 357</span>  <span class="keywordflow">return</span>;</div>
|
|
<div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  }</div>
|
|
<div class="line"><a name="l00359"></a><span class="lineno"> 359</span> </div>
|
|
<div class="line"><a name="l00360"></a><span class="lineno"> 360</span>  <span class="keywordtype">size_t</span> lengthInBytes = numVecs * <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>;</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>  <span class="keyword">auto</span>& listCodes = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>[listId];</div>
|
|
<div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  <span class="keyword">auto</span> prevCodeData = listCodes->data();</div>
|
|
<div class="line"><a name="l00364"></a><span class="lineno"> 364</span> </div>
|
|
<div class="line"><a name="l00365"></a><span class="lineno"> 365</span>  <span class="comment">// We only have int32 length representations on the GPU per each</span></div>
|
|
<div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  <span class="comment">// list; the length is in sizeof(char)</span></div>
|
|
<div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  FAISS_ASSERT(listCodes->size() % bytesPerVector_ == 0);</div>
|
|
<div class="line"><a name="l00368"></a><span class="lineno"> 368</span>  FAISS_ASSERT(listCodes->size() + lengthInBytes <=</div>
|
|
<div class="line"><a name="l00369"></a><span class="lineno"> 369</span>  (size_t) std::numeric_limits<int>::max());</div>
|
|
<div class="line"><a name="l00370"></a><span class="lineno"> 370</span> </div>
|
|
<div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  listCodes->append((<span class="keywordtype">unsigned</span> <span class="keywordtype">char</span>*) codes,</div>
|
|
<div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  lengthInBytes,</div>
|
|
<div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  stream,</div>
|
|
<div class="line"><a name="l00374"></a><span class="lineno"> 374</span>  <span class="keyword">true</span> <span class="comment">/* exact reserved size */</span>);</div>
|
|
<div class="line"><a name="l00375"></a><span class="lineno"> 375</span> </div>
|
|
<div class="line"><a name="l00376"></a><span class="lineno"> 376</span>  <span class="comment">// Handle the indices as well</span></div>
|
|
<div class="line"><a name="l00377"></a><span class="lineno"> 377</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a5027720549de98f4e609d6339099df35">addIndicesFromCpu_</a>(listId, indices, numVecs);</div>
|
|
<div class="line"><a name="l00378"></a><span class="lineno"> 378</span> </div>
|
|
<div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  <span class="comment">// This list address may have changed due to vector resizing, but</span></div>
|
|
<div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  <span class="comment">// only bother updating it on the device if it has changed</span></div>
|
|
<div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  <span class="keywordflow">if</span> (prevCodeData != listCodes->data()) {</div>
|
|
<div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a96240a08b42bd1913e2286d7d514fc56">deviceListDataPointers_</a>[listId] = listCodes->data();</div>
|
|
<div class="line"><a name="l00383"></a><span class="lineno"> 383</span>  }</div>
|
|
<div class="line"><a name="l00384"></a><span class="lineno"> 384</span> </div>
|
|
<div class="line"><a name="l00385"></a><span class="lineno"> 385</span>  <span class="comment">// And our size has changed too</span></div>
|
|
<div class="line"><a name="l00386"></a><span class="lineno"> 386</span>  <span class="keywordtype">int</span> listLength = listCodes->size() / <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>;</div>
|
|
<div class="line"><a name="l00387"></a><span class="lineno"> 387</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a3a1c2031a4763f7d55bc8a400c63af66">deviceListLengths_</a>[listId] = listLength;</div>
|
|
<div class="line"><a name="l00388"></a><span class="lineno"> 388</span> </div>
|
|
<div class="line"><a name="l00389"></a><span class="lineno"> 389</span>  <span class="comment">// We update this as well, since the multi-pass algorithm uses it</span></div>
|
|
<div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a> = std::max(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a>, listLength);</div>
|
|
<div class="line"><a name="l00391"></a><span class="lineno"> 391</span> </div>
|
|
<div class="line"><a name="l00392"></a><span class="lineno"> 392</span>  <span class="comment">// device_vector add is potentially happening on a different stream</span></div>
|
|
<div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  <span class="comment">// than our default stream</span></div>
|
|
<div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  <span class="keywordflow">if</span> (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice() != 0) {</div>
|
|
<div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  streamWait({stream}, {0});</div>
|
|
<div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  }</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> </div>
|
|
<div class="line"><a name="l00399"></a><span class="lineno"> 399</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00400"></a><span class="lineno"> 400</span> IVFPQ::setPQCentroids_(<span class="keywordtype">float</span>* data) {</div>
|
|
<div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  <span class="keywordtype">size_t</span> pqSize =</div>
|
|
<div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  numSubQuantizers_ * numSubQuantizerCodes_ * dimPerSubQuantizer_;</div>
|
|
<div class="line"><a name="l00403"></a><span class="lineno"> 403</span> </div>
|
|
<div class="line"><a name="l00404"></a><span class="lineno"> 404</span>  <span class="comment">// Make sure the data is on the host</span></div>
|
|
<div class="line"><a name="l00405"></a><span class="lineno"> 405</span>  <span class="comment">// FIXME: why are we doing this?</span></div>
|
|
<div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  thrust::host_vector<float> hostMemory;</div>
|
|
<div class="line"><a name="l00407"></a><span class="lineno"> 407</span>  hostMemory.insert(hostMemory.end(), data, data + pqSize);</div>
|
|
<div class="line"><a name="l00408"></a><span class="lineno"> 408</span> </div>
|
|
<div class="line"><a name="l00409"></a><span class="lineno"> 409</span>  <a class="code" href="classfaiss_1_1gpu_1_1HostTensor.html">HostTensor<float, 3, true></a> pqHost(</div>
|
|
<div class="line"><a name="l00410"></a><span class="lineno"> 410</span>  hostMemory.data(),</div>
|
|
<div class="line"><a name="l00411"></a><span class="lineno"> 411</span>  {numSubQuantizers_, numSubQuantizerCodes_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  DeviceTensor<float, 3, true> pqDevice(</div>
|
|
<div class="line"><a name="l00413"></a><span class="lineno"> 413</span>  pqHost,</div>
|
|
<div class="line"><a name="l00414"></a><span class="lineno"> 414</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00415"></a><span class="lineno"> 415</span> </div>
|
|
<div class="line"><a name="l00416"></a><span class="lineno"> 416</span>  DeviceTensor<float, 3, true> pqDeviceTranspose(</div>
|
|
<div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  {numSubQuantizers_, dimPerSubQuantizer_, numSubQuantizerCodes_});</div>
|
|
<div class="line"><a name="l00418"></a><span class="lineno"> 418</span>  runTransposeAny(pqDevice, 1, 2, pqDeviceTranspose,</div>
|
|
<div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00420"></a><span class="lineno"> 420</span> </div>
|
|
<div class="line"><a name="l00421"></a><span class="lineno"> 421</span>  pqCentroidsInnermostCode_ = std::move(pqDeviceTranspose);</div>
|
|
<div class="line"><a name="l00422"></a><span class="lineno"> 422</span> </div>
|
|
<div class="line"><a name="l00423"></a><span class="lineno"> 423</span>  <span class="comment">// Also maintain the PQ centroids in the form</span></div>
|
|
<div class="line"><a name="l00424"></a><span class="lineno"> 424</span>  <span class="comment">// (sub q)(code id)(sub dim)</span></div>
|
|
<div class="line"><a name="l00425"></a><span class="lineno"> 425</span>  DeviceTensor<float, 3, true> pqCentroidsMiddleCode(</div>
|
|
<div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  {numSubQuantizers_, numSubQuantizerCodes_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  runTransposeAny(pqCentroidsInnermostCode_, 1, 2, pqCentroidsMiddleCode,</div>
|
|
<div class="line"><a name="l00428"></a><span class="lineno"> 428</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00429"></a><span class="lineno"> 429</span> </div>
|
|
<div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  pqCentroidsMiddleCode_ = std::move(pqCentroidsMiddleCode);</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> </div>
|
|
<div class="line"><a name="l00433"></a><span class="lineno"> 433</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00434"></a><span class="lineno"> 434</span> IVFPQ::precomputeCodes_() {</div>
|
|
<div class="line"><a name="l00435"></a><span class="lineno"> 435</span>  <span class="comment">//</span></div>
|
|
<div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  <span class="comment">// d = || x - y_C ||^2 + || y_R ||^2 + 2 * (y_C|y_R) - 2 * (x|y_R)</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">// term 1 term 2 term 3</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"> 440</span> </div>
|
|
<div class="line"><a name="l00441"></a><span class="lineno"> 441</span>  <span class="comment">// Terms 1 and 3 are available only at query time. We compute term 2</span></div>
|
|
<div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  <span class="comment">// here.</span></div>
|
|
<div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  FAISS_ASSERT(!<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>->getUseFloat16());</div>
|
|
<div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  <span class="keyword">auto</span>& coarseCentroids = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>-><a class="code" href="classfaiss_1_1gpu_1_1FlatIndex.html#a12058744ffb3fbcbb047872449269c06">getVectorsFloat32Ref</a>();</div>
|
|
<div class="line"><a name="l00445"></a><span class="lineno"> 445</span> </div>
|
|
<div class="line"><a name="l00446"></a><span class="lineno"> 446</span>  <span class="comment">// Compute ||y_R||^2 by treating</span></div>
|
|
<div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <span class="comment">// (sub q)(code id)(sub dim) as (sub q * code id)(sub dim)</span></div>
|
|
<div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  <span class="keyword">auto</span> pqCentroidsMiddleCodeView =</div>
|
|
<div class="line"><a name="l00449"></a><span class="lineno"> 449</span>  pqCentroidsMiddleCode_.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">view</a><2>(</div>
|
|
<div class="line"><a name="l00450"></a><span class="lineno"> 450</span>  {numSubQuantizers_ * numSubQuantizerCodes_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  DeviceTensor<float, 1, true> subQuantizerNorms(</div>
|
|
<div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  {numSubQuantizers_ * numSubQuantizerCodes_});</div>
|
|
<div class="line"><a name="l00453"></a><span class="lineno"> 453</span> </div>
|
|
<div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  runL2Norm(pqCentroidsMiddleCodeView, subQuantizerNorms, <span class="keyword">true</span>,</div>
|
|
<div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00456"></a><span class="lineno"> 456</span> </div>
|
|
<div class="line"><a name="l00457"></a><span class="lineno"> 457</span>  <span class="comment">// Compute 2 * (y_C|y_R) via batch matrix multiplication</span></div>
|
|
<div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  <span class="comment">// batch size (sub q) x {(centroid id)(sub dim) x (code id)(sub dim)'}</span></div>
|
|
<div class="line"><a name="l00459"></a><span class="lineno"> 459</span>  <span class="comment">// => (sub q) x {(centroid id)(code id)}</span></div>
|
|
<div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  <span class="comment">// => (sub q)(centroid id)(code id)</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="comment">// View (centroid id)(dim) as</span></div>
|
|
<div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  <span class="comment">// (centroid id)(sub q)(dim)</span></div>
|
|
<div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  <span class="comment">// Transpose (centroid id)(sub q)(sub dim) to</span></div>
|
|
<div class="line"><a name="l00465"></a><span class="lineno"> 465</span>  <span class="comment">// (sub q)(centroid id)(sub dim)</span></div>
|
|
<div class="line"><a name="l00466"></a><span class="lineno"> 466</span>  <span class="keyword">auto</span> centroidView = coarseCentroids.view<3>(</div>
|
|
<div class="line"><a name="l00467"></a><span class="lineno"> 467</span>  {coarseCentroids.getSize(0), numSubQuantizers_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00468"></a><span class="lineno"> 468</span>  DeviceTensor<float, 3, true> centroidsTransposed(</div>
|
|
<div class="line"><a name="l00469"></a><span class="lineno"> 469</span>  {numSubQuantizers_, coarseCentroids.getSize(0), dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00470"></a><span class="lineno"> 470</span> </div>
|
|
<div class="line"><a name="l00471"></a><span class="lineno"> 471</span>  runTransposeAny(centroidView, 0, 1, centroidsTransposed,</div>
|
|
<div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00473"></a><span class="lineno"> 473</span> </div>
|
|
<div class="line"><a name="l00474"></a><span class="lineno"> 474</span>  DeviceTensor<float, 3, true> coarsePQProduct(</div>
|
|
<div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  {numSubQuantizers_, coarseCentroids.getSize(0), numSubQuantizerCodes_});</div>
|
|
<div class="line"><a name="l00476"></a><span class="lineno"> 476</span> </div>
|
|
<div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  runIteratedMatrixMult(coarsePQProduct, <span class="keyword">false</span>,</div>
|
|
<div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  centroidsTransposed, <span class="keyword">false</span>,</div>
|
|
<div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  pqCentroidsMiddleCode_, <span class="keyword">true</span>,</div>
|
|
<div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  2.0f, 0.0f,</div>
|
|
<div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getBlasHandleCurrentDevice(),</div>
|
|
<div class="line"><a name="l00482"></a><span class="lineno"> 482</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00483"></a><span class="lineno"> 483</span> </div>
|
|
<div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  <span class="comment">// Transpose (sub q)(centroid id)(code id) to</span></div>
|
|
<div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  <span class="comment">// (centroid id)(sub q)(code id)</span></div>
|
|
<div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  DeviceTensor<float, 3, true> coarsePQProductTransposed(</div>
|
|
<div class="line"><a name="l00487"></a><span class="lineno"> 487</span>  {coarseCentroids.getSize(0), numSubQuantizers_, numSubQuantizerCodes_});</div>
|
|
<div class="line"><a name="l00488"></a><span class="lineno"> 488</span>  runTransposeAny(coarsePQProduct, 0, 1, coarsePQProductTransposed,</div>
|
|
<div class="line"><a name="l00489"></a><span class="lineno"> 489</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00490"></a><span class="lineno"> 490</span> </div>
|
|
<div class="line"><a name="l00491"></a><span class="lineno"> 491</span>  <span class="comment">// View (centroid id)(sub q)(code id) as</span></div>
|
|
<div class="line"><a name="l00492"></a><span class="lineno"> 492</span>  <span class="comment">// (centroid id)(sub q * code id)</span></div>
|
|
<div class="line"><a name="l00493"></a><span class="lineno"> 493</span>  <span class="keyword">auto</span> coarsePQProductTransposedView = coarsePQProductTransposed.view<2>(</div>
|
|
<div class="line"><a name="l00494"></a><span class="lineno"> 494</span>  {coarseCentroids.getSize(0), numSubQuantizers_ * numSubQuantizerCodes_});</div>
|
|
<div class="line"><a name="l00495"></a><span class="lineno"> 495</span> </div>
|
|
<div class="line"><a name="l00496"></a><span class="lineno"> 496</span>  <span class="comment">// Sum || y_R ||^2 + 2 * (y_C|y_R)</span></div>
|
|
<div class="line"><a name="l00497"></a><span class="lineno"> 497</span>  <span class="comment">// i.e., add norms (sub q * code id)</span></div>
|
|
<div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="comment">// along columns of inner product (centroid id)(sub q * code id)</span></div>
|
|
<div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  runSumAlongColumns(subQuantizerNorms, coarsePQProductTransposedView,</div>
|
|
<div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00501"></a><span class="lineno"> 501</span> </div>
|
|
<div class="line"><a name="l00502"></a><span class="lineno"> 502</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00503"></a><span class="lineno"> 503</span> <span class="preprocessor"></span> <span class="keywordflow">if</span> (useFloat16LookupTables_) {</div>
|
|
<div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  precomputedCodeHalf_ = toHalf(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>,</div>
|
|
<div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice(),</div>
|
|
<div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  coarsePQProductTransposed);</div>
|
|
<div class="line"><a name="l00507"></a><span class="lineno"> 507</span>  <span class="keywordflow">return</span>;</div>
|
|
<div class="line"><a name="l00508"></a><span class="lineno"> 508</span>  }</div>
|
|
<div class="line"><a name="l00509"></a><span class="lineno"> 509</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00510"></a><span class="lineno"> 510</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  <span class="comment">// We added into the view, so `coarsePQProductTransposed` is now our</span></div>
|
|
<div class="line"><a name="l00512"></a><span class="lineno"> 512</span>  <span class="comment">// precomputed term 2.</span></div>
|
|
<div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  precomputedCode_ = std::move(coarsePQProductTransposed);</div>
|
|
<div class="line"><a name="l00514"></a><span class="lineno"> 514</span> }</div>
|
|
<div class="line"><a name="l00515"></a><span class="lineno"> 515</span> </div>
|
|
<div class="line"><a name="l00516"></a><span class="lineno"> 516</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00517"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#ab0c458aab9a3d903f31b0e63ce16e623"> 517</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#ab0c458aab9a3d903f31b0e63ce16e623">IVFPQ::query</a>(<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 2, true></a>& queries,</div>
|
|
<div class="line"><a name="l00518"></a><span class="lineno"> 518</span>  <span class="keywordtype">int</span> nprobe,</div>
|
|
<div class="line"><a name="l00519"></a><span class="lineno"> 519</span>  <span class="keywordtype">int</span> k,</div>
|
|
<div class="line"><a name="l00520"></a><span class="lineno"> 520</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 2, true></a>& outDistances,</div>
|
|
<div class="line"><a name="l00521"></a><span class="lineno"> 521</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<long, 2, true></a>& outIndices) {</div>
|
|
<div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  <span class="comment">// Validate these at a top level</span></div>
|
|
<div class="line"><a name="l00523"></a><span class="lineno"> 523</span>  FAISS_ASSERT(nprobe <= 1024);</div>
|
|
<div class="line"><a name="l00524"></a><span class="lineno"> 524</span>  FAISS_ASSERT(k <= 1024);</div>
|
|
<div class="line"><a name="l00525"></a><span class="lineno"> 525</span> </div>
|
|
<div class="line"><a name="l00526"></a><span class="lineno"> 526</span>  <span class="keyword">auto</span>& mem = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getMemoryManagerCurrentDevice();</div>
|
|
<div class="line"><a name="l00527"></a><span class="lineno"> 527</span>  <span class="keyword">auto</span> stream = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice();</div>
|
|
<div class="line"><a name="l00528"></a><span class="lineno"> 528</span>  nprobe = std::min(nprobe, <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>-><a class="code" href="classfaiss_1_1gpu_1_1FlatIndex.html#a6988df17792dae30f24cc859728777e6">getSize</a>());</div>
|
|
<div class="line"><a name="l00529"></a><span class="lineno"> 529</span> </div>
|
|
<div class="line"><a name="l00530"></a><span class="lineno"> 530</span>  FAISS_ASSERT(queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(1) == <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#aba3e3cfa469e5187f2d553fff10e0250">dim_</a>);</div>
|
|
<div class="line"><a name="l00531"></a><span class="lineno"> 531</span>  FAISS_ASSERT(outDistances.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0) == queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0));</div>
|
|
<div class="line"><a name="l00532"></a><span class="lineno"> 532</span>  FAISS_ASSERT(outIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0) == queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0));</div>
|
|
<div class="line"><a name="l00533"></a><span class="lineno"> 533</span> </div>
|
|
<div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  <span class="comment">// Reserve space for the closest coarse centroids</span></div>
|
|
<div class="line"><a name="l00535"></a><span class="lineno"> 535</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 2, true></a></div>
|
|
<div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  coarseDistances(mem, {queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), nprobe}, stream);</div>
|
|
<div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 2, true></a></div>
|
|
<div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  coarseIndices(mem, {queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), nprobe}, stream);</div>
|
|
<div class="line"><a name="l00539"></a><span class="lineno"> 539</span> </div>
|
|
<div class="line"><a name="l00540"></a><span class="lineno"> 540</span>  <span class="comment">// Find the `nprobe` closest coarse centroids; we can use int</span></div>
|
|
<div class="line"><a name="l00541"></a><span class="lineno"> 541</span>  <span class="comment">// indices both internally and externally</span></div>
|
|
<div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>->query(queries,</div>
|
|
<div class="line"><a name="l00543"></a><span class="lineno"> 543</span>  nprobe,</div>
|
|
<div class="line"><a name="l00544"></a><span class="lineno"> 544</span>  coarseDistances,</div>
|
|
<div class="line"><a name="l00545"></a><span class="lineno"> 545</span>  coarseIndices,</div>
|
|
<div class="line"><a name="l00546"></a><span class="lineno"> 546</span>  <span class="keyword">true</span>);</div>
|
|
<div class="line"><a name="l00547"></a><span class="lineno"> 547</span> </div>
|
|
<div class="line"><a name="l00548"></a><span class="lineno"> 548</span>  <span class="keywordflow">if</span> (precomputedCodes_) {</div>
|
|
<div class="line"><a name="l00549"></a><span class="lineno"> 549</span>  runPQPrecomputedCodes_(queries,</div>
|
|
<div class="line"><a name="l00550"></a><span class="lineno"> 550</span>  coarseDistances,</div>
|
|
<div class="line"><a name="l00551"></a><span class="lineno"> 551</span>  coarseIndices,</div>
|
|
<div class="line"><a name="l00552"></a><span class="lineno"> 552</span>  k,</div>
|
|
<div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  outDistances,</div>
|
|
<div class="line"><a name="l00554"></a><span class="lineno"> 554</span>  outIndices);</div>
|
|
<div class="line"><a name="l00555"></a><span class="lineno"> 555</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  runPQNoPrecomputedCodes_(queries,</div>
|
|
<div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  coarseDistances,</div>
|
|
<div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  coarseIndices,</div>
|
|
<div class="line"><a name="l00559"></a><span class="lineno"> 559</span>  k,</div>
|
|
<div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  outDistances,</div>
|
|
<div class="line"><a name="l00561"></a><span class="lineno"> 561</span>  outIndices);</div>
|
|
<div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  }</div>
|
|
<div class="line"><a name="l00563"></a><span class="lineno"> 563</span> </div>
|
|
<div class="line"><a name="l00564"></a><span class="lineno"> 564</span>  <span class="comment">// If the GPU isn't storing indices (they are on the CPU side), we</span></div>
|
|
<div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  <span class="comment">// need to perform the re-mapping here</span></div>
|
|
<div class="line"><a name="l00566"></a><span class="lineno"> 566</span>  <span class="comment">// FIXME: we might ultimately be calling this function with inputs</span></div>
|
|
<div class="line"><a name="l00567"></a><span class="lineno"> 567</span>  <span class="comment">// from the CPU, these are unnecessary copies</span></div>
|
|
<div class="line"><a name="l00568"></a><span class="lineno"> 568</span>  <span class="keywordflow">if</span> (<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a> == INDICES_CPU) {</div>
|
|
<div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  <a class="code" href="classfaiss_1_1gpu_1_1HostTensor.html">HostTensor<long, 2, true></a> hostOutIndices(outIndices, stream);</div>
|
|
<div class="line"><a name="l00570"></a><span class="lineno"> 570</span> </div>
|
|
<div class="line"><a name="l00571"></a><span class="lineno"> 571</span>  ivfOffsetToUserIndex(hostOutIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a37d3ac0dffcaad29f09d6e85fb07b335">data</a>(),</div>
|
|
<div class="line"><a name="l00572"></a><span class="lineno"> 572</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#accc4d96c14643e5f471220cb1e92ac70">numLists_</a>,</div>
|
|
<div class="line"><a name="l00573"></a><span class="lineno"> 573</span>  hostOutIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0),</div>
|
|
<div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  hostOutIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(1),</div>
|
|
<div class="line"><a name="l00575"></a><span class="lineno"> 575</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">listOffsetToUserIndex_</a>);</div>
|
|
<div class="line"><a name="l00576"></a><span class="lineno"> 576</span> </div>
|
|
<div class="line"><a name="l00577"></a><span class="lineno"> 577</span>  <span class="comment">// Copy back to GPU, since the input to this function is on the</span></div>
|
|
<div class="line"><a name="l00578"></a><span class="lineno"> 578</span>  <span class="comment">// GPU</span></div>
|
|
<div class="line"><a name="l00579"></a><span class="lineno"> 579</span>  outIndices.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#ae981a94263044f38be89d690dd958426">copyFrom</a>(hostOutIndices, stream);</div>
|
|
<div class="line"><a name="l00580"></a><span class="lineno"> 580</span>  }</div>
|
|
<div class="line"><a name="l00581"></a><span class="lineno"> 581</span> }</div>
|
|
<div class="line"><a name="l00582"></a><span class="lineno"> 582</span> </div>
|
|
<div class="line"><a name="l00583"></a><span class="lineno"> 583</span> std::vector<unsigned char></div>
|
|
<div class="line"><a name="l00584"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#a5b349dd021b11b5f48531825359b0657"> 584</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#a5b349dd021b11b5f48531825359b0657">IVFPQ::getListCodes</a>(<span class="keywordtype">int</span> listId)<span class="keyword"> const </span>{</div>
|
|
<div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  FAISS_ASSERT(listId < <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>.size());</div>
|
|
<div class="line"><a name="l00586"></a><span class="lineno"> 586</span> </div>
|
|
<div class="line"><a name="l00587"></a><span class="lineno"> 587</span>  <span class="keywordflow">return</span> <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">deviceListData_</a>[listId]->copyToHost<<span class="keywordtype">unsigned</span> <span class="keywordtype">char</span>>(</div>
|
|
<div class="line"><a name="l00588"></a><span class="lineno"> 588</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice());</div>
|
|
<div class="line"><a name="l00589"></a><span class="lineno"> 589</span> }</div>
|
|
<div class="line"><a name="l00590"></a><span class="lineno"> 590</span> </div>
|
|
<div class="line"><a name="l00591"></a><span class="lineno"> 591</span> <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 3, true></a></div>
|
|
<div class="line"><a name="l00592"></a><span class="lineno"><a class="line" href="classfaiss_1_1gpu_1_1IVFPQ.html#a3e8bff50f894c243c62e832f923e88e7"> 592</a></span> <a class="code" href="classfaiss_1_1gpu_1_1IVFPQ.html#a3e8bff50f894c243c62e832f923e88e7">IVFPQ::getPQCentroids</a>() {</div>
|
|
<div class="line"><a name="l00593"></a><span class="lineno"> 593</span>  <span class="keywordflow">return</span> pqCentroidsMiddleCode_;</div>
|
|
<div class="line"><a name="l00594"></a><span class="lineno"> 594</span> }</div>
|
|
<div class="line"><a name="l00595"></a><span class="lineno"> 595</span> </div>
|
|
<div class="line"><a name="l00596"></a><span class="lineno"> 596</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00597"></a><span class="lineno"> 597</span> IVFPQ::runPQPrecomputedCodes_(</div>
|
|
<div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 2, true></a>& queries,</div>
|
|
<div class="line"><a name="l00599"></a><span class="lineno"> 599</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 2, true></a>& coarseDistances,</div>
|
|
<div class="line"><a name="l00600"></a><span class="lineno"> 600</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<int, 2, true></a>& coarseIndices,</div>
|
|
<div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  <span class="keywordtype">int</span> k,</div>
|
|
<div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<float, 2, true></a>& outDistances,</div>
|
|
<div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  <a class="code" href="classfaiss_1_1gpu_1_1Tensor.html">Tensor<long, 2, true></a>& outIndices) {</div>
|
|
<div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  <span class="keyword">auto</span>& mem = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getMemoryManagerCurrentDevice();</div>
|
|
<div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  <span class="keyword">auto</span> stream = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getDefaultStreamCurrentDevice();</div>
|
|
<div class="line"><a name="l00606"></a><span class="lineno"> 606</span> </div>
|
|
<div class="line"><a name="l00607"></a><span class="lineno"> 607</span>  <span class="comment">// Compute precomputed code term 3, - 2 * (x|y_R)</span></div>
|
|
<div class="line"><a name="l00608"></a><span class="lineno"> 608</span>  <span class="comment">// This is done via batch MM</span></div>
|
|
<div class="line"><a name="l00609"></a><span class="lineno"> 609</span>  <span class="comment">// {sub q} x {(query id)(sub dim) * (code id)(sub dim)'} =></span></div>
|
|
<div class="line"><a name="l00610"></a><span class="lineno"> 610</span>  <span class="comment">// {sub q} x {(query id)(code id)}</span></div>
|
|
<div class="line"><a name="l00611"></a><span class="lineno"> 611</span>  <a class="code" href="classfaiss_1_1gpu_1_1DeviceTensor.html">DeviceTensor<float, 3, true></a> term3Transposed(</div>
|
|
<div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  mem,</div>
|
|
<div class="line"><a name="l00613"></a><span class="lineno"> 613</span>  {queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), numSubQuantizers_, numSubQuantizerCodes_},</div>
|
|
<div class="line"><a name="l00614"></a><span class="lineno"> 614</span>  stream);</div>
|
|
<div class="line"><a name="l00615"></a><span class="lineno"> 615</span> </div>
|
|
<div class="line"><a name="l00616"></a><span class="lineno"> 616</span>  <span class="comment">// These allocations within are only temporary, so release them when</span></div>
|
|
<div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  <span class="comment">// we're done to maximize free space</span></div>
|
|
<div class="line"><a name="l00618"></a><span class="lineno"> 618</span>  {</div>
|
|
<div class="line"><a name="l00619"></a><span class="lineno"> 619</span>  <span class="keyword">auto</span> querySubQuantizerView = queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">view</a><3>(</div>
|
|
<div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  {queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), numSubQuantizers_, dimPerSubQuantizer_});</div>
|
|
<div class="line"><a name="l00621"></a><span class="lineno"> 621</span>  DeviceTensor<float, 3, true> queriesTransposed(</div>
|
|
<div class="line"><a name="l00622"></a><span class="lineno"> 622</span>  mem,</div>
|
|
<div class="line"><a name="l00623"></a><span class="lineno"> 623</span>  {numSubQuantizers_, queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), dimPerSubQuantizer_},</div>
|
|
<div class="line"><a name="l00624"></a><span class="lineno"> 624</span>  stream);</div>
|
|
<div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  runTransposeAny(querySubQuantizerView, 0, 1, queriesTransposed, stream);</div>
|
|
<div class="line"><a name="l00626"></a><span class="lineno"> 626</span> </div>
|
|
<div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  DeviceTensor<float, 3, true> term3(</div>
|
|
<div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  mem,</div>
|
|
<div class="line"><a name="l00629"></a><span class="lineno"> 629</span>  {numSubQuantizers_, queries.<a class="code" href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">getSize</a>(0), numSubQuantizerCodes_},</div>
|
|
<div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  stream);</div>
|
|
<div class="line"><a name="l00631"></a><span class="lineno"> 631</span> </div>
|
|
<div class="line"><a name="l00632"></a><span class="lineno"> 632</span>  runIteratedMatrixMult(term3, <span class="keyword">false</span>,</div>
|
|
<div class="line"><a name="l00633"></a><span class="lineno"> 633</span>  queriesTransposed, <span class="keyword">false</span>,</div>
|
|
<div class="line"><a name="l00634"></a><span class="lineno"> 634</span>  pqCentroidsMiddleCode_, <span class="keyword">true</span>,</div>
|
|
<div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  -2.0f, 0.0f,</div>
|
|
<div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>->getBlasHandleCurrentDevice(),</div>
|
|
<div class="line"><a name="l00637"></a><span class="lineno"> 637</span>  stream);</div>
|
|
<div class="line"><a name="l00638"></a><span class="lineno"> 638</span> </div>
|
|
<div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  runTransposeAny(term3, 0, 1, term3Transposed, stream);</div>
|
|
<div class="line"><a name="l00640"></a><span class="lineno"> 640</span>  }</div>
|
|
<div class="line"><a name="l00641"></a><span class="lineno"> 641</span> </div>
|
|
<div class="line"><a name="l00642"></a><span class="lineno"> 642</span>  NoTypeTensor<3, true> term2;</div>
|
|
<div class="line"><a name="l00643"></a><span class="lineno"> 643</span>  NoTypeTensor<3, true> term3;</div>
|
|
<div class="line"><a name="l00644"></a><span class="lineno"> 644</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00645"></a><span class="lineno"> 645</span> <span class="preprocessor"></span> DeviceTensor<half, 3, true> term3Half;</div>
|
|
<div class="line"><a name="l00646"></a><span class="lineno"> 646</span> </div>
|
|
<div class="line"><a name="l00647"></a><span class="lineno"> 647</span>  <span class="keywordflow">if</span> (useFloat16LookupTables_) {</div>
|
|
<div class="line"><a name="l00648"></a><span class="lineno"> 648</span>  term3Half = toHalf(<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>, stream, term3Transposed);</div>
|
|
<div class="line"><a name="l00649"></a><span class="lineno"> 649</span>  term2 = NoTypeTensor<3, true>(precomputedCodeHalf_);</div>
|
|
<div class="line"><a name="l00650"></a><span class="lineno"> 650</span>  term3 = NoTypeTensor<3, true>(term3Half);</div>
|
|
<div class="line"><a name="l00651"></a><span class="lineno"> 651</span>  }</div>
|
|
<div class="line"><a name="l00652"></a><span class="lineno"> 652</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00653"></a><span class="lineno"> 653</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00654"></a><span class="lineno"> 654</span>  <span class="keywordflow">if</span> (!useFloat16LookupTables_) {</div>
|
|
<div class="line"><a name="l00655"></a><span class="lineno"> 655</span>  term2 = NoTypeTensor<3, true>(precomputedCode_);</div>
|
|
<div class="line"><a name="l00656"></a><span class="lineno"> 656</span>  term3 = NoTypeTensor<3, true>(term3Transposed);</div>
|
|
<div class="line"><a name="l00657"></a><span class="lineno"> 657</span>  }</div>
|
|
<div class="line"><a name="l00658"></a><span class="lineno"> 658</span> </div>
|
|
<div class="line"><a name="l00659"></a><span class="lineno"> 659</span>  runPQScanMultiPassPrecomputed(queries,</div>
|
|
<div class="line"><a name="l00660"></a><span class="lineno"> 660</span>  coarseDistances, <span class="comment">// term 1</span></div>
|
|
<div class="line"><a name="l00661"></a><span class="lineno"> 661</span>  term2, <span class="comment">// term 2</span></div>
|
|
<div class="line"><a name="l00662"></a><span class="lineno"> 662</span>  term3, <span class="comment">// term 3</span></div>
|
|
<div class="line"><a name="l00663"></a><span class="lineno"> 663</span>  coarseIndices,</div>
|
|
<div class="line"><a name="l00664"></a><span class="lineno"> 664</span>  useFloat16LookupTables_,</div>
|
|
<div class="line"><a name="l00665"></a><span class="lineno"> 665</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>,</div>
|
|
<div class="line"><a name="l00666"></a><span class="lineno"> 666</span>  numSubQuantizers_,</div>
|
|
<div class="line"><a name="l00667"></a><span class="lineno"> 667</span>  numSubQuantizerCodes_,</div>
|
|
<div class="line"><a name="l00668"></a><span class="lineno"> 668</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a96240a08b42bd1913e2286d7d514fc56">deviceListDataPointers_</a>,</div>
|
|
<div class="line"><a name="l00669"></a><span class="lineno"> 669</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a9aedcf0e6a20b908980ae96d73461f4c">deviceListIndexPointers_</a>,</div>
|
|
<div class="line"><a name="l00670"></a><span class="lineno"> 670</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a>,</div>
|
|
<div class="line"><a name="l00671"></a><span class="lineno"> 671</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a3a1c2031a4763f7d55bc8a400c63af66">deviceListLengths_</a>,</div>
|
|
<div class="line"><a name="l00672"></a><span class="lineno"> 672</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a>,</div>
|
|
<div class="line"><a name="l00673"></a><span class="lineno"> 673</span>  k,</div>
|
|
<div class="line"><a name="l00674"></a><span class="lineno"> 674</span>  outDistances,</div>
|
|
<div class="line"><a name="l00675"></a><span class="lineno"> 675</span>  outIndices,</div>
|
|
<div class="line"><a name="l00676"></a><span class="lineno"> 676</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>);</div>
|
|
<div class="line"><a name="l00677"></a><span class="lineno"> 677</span> }</div>
|
|
<div class="line"><a name="l00678"></a><span class="lineno"> 678</span> </div>
|
|
<div class="line"><a name="l00679"></a><span class="lineno"> 679</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00680"></a><span class="lineno"> 680</span> IVFPQ::runPQNoPrecomputedCodes_(</div>
|
|
<div class="line"><a name="l00681"></a><span class="lineno"> 681</span>  Tensor<float, 2, true>& queries,</div>
|
|
<div class="line"><a name="l00682"></a><span class="lineno"> 682</span>  DeviceTensor<float, 2, true>& coarseDistances,</div>
|
|
<div class="line"><a name="l00683"></a><span class="lineno"> 683</span>  DeviceTensor<int, 2, true>& coarseIndices,</div>
|
|
<div class="line"><a name="l00684"></a><span class="lineno"> 684</span>  <span class="keywordtype">int</span> k,</div>
|
|
<div class="line"><a name="l00685"></a><span class="lineno"> 685</span>  Tensor<float, 2, true>& outDistances,</div>
|
|
<div class="line"><a name="l00686"></a><span class="lineno"> 686</span>  Tensor<long, 2, true>& outIndices) {</div>
|
|
<div class="line"><a name="l00687"></a><span class="lineno"> 687</span>  FAISS_ASSERT(!<a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>->getUseFloat16());</div>
|
|
<div class="line"><a name="l00688"></a><span class="lineno"> 688</span>  <span class="keyword">auto</span>& coarseCentroids = <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">quantizer_</a>-><a class="code" href="classfaiss_1_1gpu_1_1FlatIndex.html#a12058744ffb3fbcbb047872449269c06">getVectorsFloat32Ref</a>();</div>
|
|
<div class="line"><a name="l00689"></a><span class="lineno"> 689</span> </div>
|
|
<div class="line"><a name="l00690"></a><span class="lineno"> 690</span>  runPQScanMultiPassNoPrecomputed(queries,</div>
|
|
<div class="line"><a name="l00691"></a><span class="lineno"> 691</span>  coarseCentroids,</div>
|
|
<div class="line"><a name="l00692"></a><span class="lineno"> 692</span>  pqCentroidsInnermostCode_,</div>
|
|
<div class="line"><a name="l00693"></a><span class="lineno"> 693</span>  coarseIndices,</div>
|
|
<div class="line"><a name="l00694"></a><span class="lineno"> 694</span>  useFloat16LookupTables_,</div>
|
|
<div class="line"><a name="l00695"></a><span class="lineno"> 695</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">bytesPerVector_</a>,</div>
|
|
<div class="line"><a name="l00696"></a><span class="lineno"> 696</span>  numSubQuantizers_,</div>
|
|
<div class="line"><a name="l00697"></a><span class="lineno"> 697</span>  numSubQuantizerCodes_,</div>
|
|
<div class="line"><a name="l00698"></a><span class="lineno"> 698</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a96240a08b42bd1913e2286d7d514fc56">deviceListDataPointers_</a>,</div>
|
|
<div class="line"><a name="l00699"></a><span class="lineno"> 699</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a9aedcf0e6a20b908980ae96d73461f4c">deviceListIndexPointers_</a>,</div>
|
|
<div class="line"><a name="l00700"></a><span class="lineno"> 700</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">indicesOptions_</a>,</div>
|
|
<div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a3a1c2031a4763f7d55bc8a400c63af66">deviceListLengths_</a>,</div>
|
|
<div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">maxListLength_</a>,</div>
|
|
<div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  k,</div>
|
|
<div class="line"><a name="l00704"></a><span class="lineno"> 704</span>  outDistances,</div>
|
|
<div class="line"><a name="l00705"></a><span class="lineno"> 705</span>  outIndices,</div>
|
|
<div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  <a class="code" href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">resources_</a>);</div>
|
|
<div class="line"><a name="l00707"></a><span class="lineno"> 707</span> }</div>
|
|
<div class="line"><a name="l00708"></a><span class="lineno"> 708</span> </div>
|
|
<div class="line"><a name="l00709"></a><span class="lineno"> 709</span> } } <span class="comment">// namespace</span></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_accc4d96c14643e5f471220cb1e92ac70"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#accc4d96c14643e5f471220cb1e92ac70">faiss::gpu::IVFBase::numLists_</a></div><div class="ttdeci">const int numLists_</div><div class="ttdoc">Number of inverted lists we maintain. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00090">IVFBase.cuh:90</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_ae25ea0901fb628844868413f51c85bda"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#ae25ea0901fb628844868413f51c85bda">faiss::gpu::IVFBase::maxListLength_</a></div><div class="ttdeci">int maxListLength_</div><div class="ttdoc">Maximum list length seen. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00111">IVFBase.cuh:111</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_a9992b38226dc8f92ca2691582fabb675"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#a9992b38226dc8f92ca2691582fabb675">faiss::gpu::IVFPQ::addCodeVectorsFromCpu</a></div><div class="ttdeci">void addCodeVectorsFromCpu(int listId, const void *codes, const long *indices, size_t numVecs)</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00347">IVFPQ.cu:347</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1FlatIndex_html_a6988df17792dae30f24cc859728777e6"><div class="ttname"><a href="classfaiss_1_1gpu_1_1FlatIndex.html#a6988df17792dae30f24cc859728777e6">faiss::gpu::FlatIndex::getSize</a></div><div class="ttdeci">int getSize() const </div><div class="ttdoc">Returns the number of vectors we contain. </div><div class="ttdef"><b>Definition:</b> <a href="FlatIndex_8cu_source.html#l00040">FlatIndex.cu:40</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a53f3c382a79b7f89630a85dfbc3a1fed"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a53f3c382a79b7f89630a85dfbc3a1fed">faiss::gpu::IVFBase::listOffsetToUserIndex_</a></div><div class="ttdeci">std::vector< std::vector< long > > listOffsetToUserIndex_</div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00123">IVFBase.cuh:123</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1FlatIndex_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1FlatIndex.html">faiss::gpu::FlatIndex</a></div><div class="ttdoc">Holder of GPU resources for a particular flat index. </div><div class="ttdef"><b>Definition:</b> <a href="FlatIndex_8cuh_source.html#l00023">FlatIndex.cuh:23</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1Tensor_html_a37d3ac0dffcaad29f09d6e85fb07b335"><div class="ttname"><a href="classfaiss_1_1gpu_1_1Tensor.html#a37d3ac0dffcaad29f09d6e85fb07b335">faiss::gpu::Tensor::data</a></div><div class="ttdeci">__host__ __device__ DataPtrType data()</div><div class="ttdoc">Returns a raw pointer to the start of our data. </div><div class="ttdef"><b>Definition:</b> <a href="Tensor_8cuh_source.html#l00162">Tensor.cuh:162</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html">faiss::gpu::IVFBase</a></div><div class="ttdoc">Base inverted list functionality for IVFFlat and IVFPQ. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00027">IVFBase.cuh:27</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1GpuResources_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1GpuResources.html">faiss::gpu::GpuResources</a></div><div class="ttdef"><b>Definition:</b> <a href="GpuResources_8h_source.html#l00024">GpuResources.h:24</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_acdcc64f9b72e661d39528d5d67e54536"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#acdcc64f9b72e661d39528d5d67e54536">faiss::gpu::IVFPQ::IVFPQ</a></div><div class="ttdeci">IVFPQ(GpuResources *resources, FlatIndex *quantizer, int numSubQuantizers, int bitsPerSubQuantizer, float *pqCentroidData, IndicesOptions indicesOptions, bool useFloat16LookupTables)</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00036">IVFPQ.cu:36</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a3a1c2031a4763f7d55bc8a400c63af66"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a3a1c2031a4763f7d55bc8a400c63af66">faiss::gpu::IVFBase::deviceListLengths_</a></div><div class="ttdeci">thrust::device_vector< int > deviceListLengths_</div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00108">IVFBase.cuh:108</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_adb58eeacdceb0e0fde1820ca7f116e05"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#adb58eeacdceb0e0fde1820ca7f116e05">faiss::gpu::IVFPQ::isSupportedPQCodeLength</a></div><div class="ttdeci">static bool isSupportedPQCodeLength(int size)</div><div class="ttdoc">Returns true if we support PQ in this size. </div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00071">IVFPQ.cu:71</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a9aedcf0e6a20b908980ae96d73461f4c"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a9aedcf0e6a20b908980ae96d73461f4c">faiss::gpu::IVFBase::deviceListIndexPointers_</a></div><div class="ttdeci">thrust::device_vector< void * > deviceListIndexPointers_</div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00104">IVFBase.cuh:104</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_ab1e07b04b25569cc58c5f3f033f4dab3"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#ab1e07b04b25569cc58c5f3f033f4dab3">faiss::gpu::IVFPQ::classifyAndAddVectors</a></div><div class="ttdeci">int classifyAndAddVectors(Tensor< float, 2, true > &vecs, Tensor< long, 1, true > &indices)</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00119">IVFPQ.cu:119</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_ab0c458aab9a3d903f31b0e63ce16e623"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#ab0c458aab9a3d903f31b0e63ce16e623">faiss::gpu::IVFPQ::query</a></div><div class="ttdeci">void query(Tensor< float, 2, true > &queries, int nprobe, int k, Tensor< float, 2, true > &outDistances, Tensor< long, 2, true > &outIndices)</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00517">IVFPQ.cu:517</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_a3e8bff50f894c243c62e832f923e88e7"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#a3e8bff50f894c243c62e832f923e88e7">faiss::gpu::IVFPQ::getPQCentroids</a></div><div class="ttdeci">Tensor< float, 3, true > getPQCentroids()</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00592">IVFPQ.cu:592</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a878114abdba07c9cf7735f9c0ed594c3"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a878114abdba07c9cf7735f9c0ed594c3">faiss::gpu::IVFBase::quantizer_</a></div><div class="ttdeci">FlatIndex * quantizer_</div><div class="ttdoc">Quantizer object. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00084">IVFBase.cuh:84</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_adcee5dbf48c3cb6b8a67f5f392e155fd"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#adcee5dbf48c3cb6b8a67f5f392e155fd">faiss::gpu::IVFPQ::setPrecomputedCodes</a></div><div class="ttdeci">void setPrecomputedCodes(bool enable)</div><div class="ttdoc">Enable or disable pre-computed codes. </div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00101">IVFPQ.cu:101</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_a5b349dd021b11b5f48531825359b0657"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#a5b349dd021b11b5f48531825359b0657">faiss::gpu::IVFPQ::getListCodes</a></div><div class="ttdeci">std::vector< unsigned char > getListCodes(int listId) const </div><div class="ttdoc">Return the list codes of a particular list back to the CPU. </div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00584">IVFPQ.cu:584</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a96240a08b42bd1913e2286d7d514fc56"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a96240a08b42bd1913e2286d7d514fc56">faiss::gpu::IVFBase::deviceListDataPointers_</a></div><div class="ttdeci">thrust::device_vector< void * > deviceListDataPointers_</div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00100">IVFBase.cuh:100</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a05e6400358ec1f529a67209d3f24cc63"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a05e6400358ec1f529a67209d3f24cc63">faiss::gpu::IVFBase::resources_</a></div><div class="ttdeci">GpuResources * resources_</div><div class="ttdoc">Collection of GPU resources that we use. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00081">IVFBase.cuh:81</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1Tensor_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1Tensor.html">faiss::gpu::Tensor</a></div><div class="ttdoc">Our tensor type. </div><div class="ttdef"><b>Definition:</b> <a href="Tensor_8cuh_source.html#l00031">Tensor.cuh:31</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1Tensor_html_a3095eaec5711fe697c16c21598a8ddc1"><div class="ttname"><a href="classfaiss_1_1gpu_1_1Tensor.html#a3095eaec5711fe697c16c21598a8ddc1">faiss::gpu::Tensor::view</a></div><div class="ttdeci">__host__ __device__ Tensor< T, SubDim, Contig, IndexT, PtrTraits > view(DataPtrType at)</div><div class="ttdef"><b>Definition:</b> <a href="Tensor-inl_8cuh_source.html#l00526">Tensor-inl.cuh:526</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a319568b832518392fed33ea4f8bfc613"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a319568b832518392fed33ea4f8bfc613">faiss::gpu::IVFBase::bytesPerVector_</a></div><div class="ttdeci">const int bytesPerVector_</div><div class="ttdoc">Number of bytes per vector in the list. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00093">IVFBase.cuh:93</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_acc695610c9513952b8d234dc0db78e5c"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#acc695610c9513952b8d234dc0db78e5c">faiss::gpu::IVFBase::updateDeviceListInfo_</a></div><div class="ttdeci">void updateDeviceListInfo_(cudaStream_t stream)</div><div class="ttdoc">Update all device-side list pointer and size information. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cu_source.html#l00137">IVFBase.cu:137</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1Tensor_html_a2d276c97faf432cdc9f3552da63c0d3c"><div class="ttname"><a href="classfaiss_1_1gpu_1_1Tensor.html#a2d276c97faf432cdc9f3552da63c0d3c">faiss::gpu::Tensor::getSize</a></div><div class="ttdeci">__host__ __device__ IndexT getSize(int i) const </div><div class="ttdef"><b>Definition:</b> <a href="Tensor_8cuh_source.html#l00210">Tensor.cuh:210</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1Tensor_html_ae981a94263044f38be89d690dd958426"><div class="ttname"><a href="classfaiss_1_1gpu_1_1Tensor.html#ae981a94263044f38be89d690dd958426">faiss::gpu::Tensor::copyFrom</a></div><div class="ttdeci">__host__ void copyFrom(Tensor< T, Dim, Contig, IndexT, PtrTraits > &t, cudaStream_t stream)</div><div class="ttdoc">Copies a tensor into ourselves; sizes must match. </div><div class="ttdef"><b>Definition:</b> <a href="Tensor-inl_8cuh_source.html#l00101">Tensor-inl.cuh:101</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1FlatIndex_html_a12058744ffb3fbcbb047872449269c06"><div class="ttname"><a href="classfaiss_1_1gpu_1_1FlatIndex.html#a12058744ffb3fbcbb047872449269c06">faiss::gpu::FlatIndex::getVectorsFloat32Ref</a></div><div class="ttdeci">Tensor< float, 2, true > & getVectorsFloat32Ref()</div><div class="ttdoc">Returns a reference to our vectors currently in use. </div><div class="ttdef"><b>Definition:</b> <a href="FlatIndex_8cu_source.html#l00061">FlatIndex.cu:61</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_afb6d10e23d6448c10f472b9234e0bcab"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#afb6d10e23d6448c10f472b9234e0bcab">faiss::gpu::IVFBase::indicesOptions_</a></div><div class="ttdeci">const IndicesOptions indicesOptions_</div><div class="ttdoc">How are user indices stored on the GPU? </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00096">IVFBase.cuh:96</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1HostTensor_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1HostTensor.html">faiss::gpu::HostTensor</a></div><div class="ttdef"><b>Definition:</b> <a href="HostTensor_8cuh_source.html#l00023">HostTensor.cuh:23</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a2facc7285107de1f24d3471cbcf15f26"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a2facc7285107de1f24d3471cbcf15f26">faiss::gpu::IVFBase::deviceListData_</a></div><div class="ttdeci">std::vector< std::unique_ptr< DeviceVector< unsigned char > > > deviceListData_</div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00117">IVFBase.cuh:117</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1DeviceTensor_html"><div class="ttname"><a href="classfaiss_1_1gpu_1_1DeviceTensor.html">faiss::gpu::DeviceTensor< float, 3, true ></a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_aba3e3cfa469e5187f2d553fff10e0250"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#aba3e3cfa469e5187f2d553fff10e0250">faiss::gpu::IVFBase::dim_</a></div><div class="ttdeci">const int dim_</div><div class="ttdoc">Expected dimensionality of the vectors. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cuh_source.html#l00087">IVFBase.cuh:87</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFBase_html_a5027720549de98f4e609d6339099df35"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFBase.html#a5027720549de98f4e609d6339099df35">faiss::gpu::IVFBase::addIndicesFromCpu_</a></div><div class="ttdeci">void addIndicesFromCpu_(int listId, const long *indices, size_t numVecs)</div><div class="ttdoc">Shared function to copy indices from CPU to GPU. </div><div class="ttdef"><b>Definition:</b> <a href="IVFBase_8cu_source.html#l00244">IVFBase.cu:244</a></div></div>
|
|
<div class="ttc" id="classfaiss_1_1gpu_1_1IVFPQ_html_a0eedf0295ad73125ee1254173a176674"><div class="ttname"><a href="classfaiss_1_1gpu_1_1IVFPQ.html#a0eedf0295ad73125ee1254173a176674">faiss::gpu::IVFPQ::isSupportedNoPrecomputedSubDimSize</a></div><div class="ttdeci">static bool isSupportedNoPrecomputedSubDimSize(int dims)</div><div class="ttdef"><b>Definition:</b> <a href="IVFPQ_8cu_source.html#l00096">IVFPQ.cu:96</a></div></div>
|
|
</div><!-- fragment --></div><!-- contents -->
|
|
<!-- start footer part -->
|
|
<hr class="footer"/><address class="footer"><small>
|
|
Generated by  <a href="http://www.doxygen.org/index.html">
|
|
<img class="footer" src="doxygen.png" alt="doxygen"/>
|
|
</a> 1.8.5
|
|
</small></address>
|
|
</body>
|
|
</html>
|