1011 lines
129 KiB
HTML
1011 lines
129 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: IVFFlatScan.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">IVFFlatScan.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 "IVFFlatScan.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 "IVFUtils.cuh"</span></div>
|
|
<div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="preprocessor">#include "../utils/ConversionOperators.cuh"</span></div>
|
|
<div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="preprocessor">#include "../utils/DeviceDefs.cuh"</span></div>
|
|
<div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="preprocessor">#include "../utils/DeviceUtils.h"</span></div>
|
|
<div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="preprocessor">#include "../utils/DeviceTensor.cuh"</span></div>
|
|
<div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="preprocessor">#include "../utils/Float16.cuh"</span></div>
|
|
<div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="preprocessor">#include "../utils/MathOperators.cuh"</span></div>
|
|
<div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="preprocessor">#include "../utils/LoadStoreOperators.cuh"</span></div>
|
|
<div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="preprocessor">#include "../utils/PtxUtils.cuh"</span></div>
|
|
<div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="preprocessor">#include "../utils/Reductions.cuh"</span></div>
|
|
<div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="preprocessor">#include "../utils/StaticUtils.h"</span></div>
|
|
<div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="preprocessor">#include <thrust/host_vector.h></span></div>
|
|
<div class="line"><a name="l00026"></a><span class="lineno"> 26</span> </div>
|
|
<div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="keyword">namespace </span>faiss { <span class="keyword">namespace </span>gpu {</div>
|
|
<div class="line"><a name="l00028"></a><span class="lineno"> 28</span> </div>
|
|
<div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="keyword">template</span> <<span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="keyword">inline</span> __device__ <span class="keyword">typename</span> Math<T>::ScalarType l2Distance(T a, T b) {</div>
|
|
<div class="line"><a name="l00031"></a><span class="lineno"> 31</span>  a = Math<T>::sub(a, b);</div>
|
|
<div class="line"><a name="l00032"></a><span class="lineno"> 32</span>  a = Math<T>::mul(a, a);</div>
|
|
<div class="line"><a name="l00033"></a><span class="lineno"> 33</span>  <span class="keywordflow">return</span> <a class="code" href="structfaiss_1_1gpu_1_1Math.html#a4b17f0b5d014f300e76dde5b24af8014">Math<T>::reduceAdd</a>(a);</div>
|
|
<div class="line"><a name="l00034"></a><span class="lineno"> 34</span> }</div>
|
|
<div class="line"><a name="l00035"></a><span class="lineno"> 35</span> </div>
|
|
<div class="line"><a name="l00036"></a><span class="lineno"> 36</span> <span class="keyword">template</span> <<span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00037"></a><span class="lineno"> 37</span> <span class="keyword">inline</span> __device__ <span class="keyword">typename</span> Math<T>::ScalarType ipDistance(T a, T b) {</div>
|
|
<div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  <span class="keywordflow">return</span> <a class="code" href="structfaiss_1_1gpu_1_1Math.html#a4b17f0b5d014f300e76dde5b24af8014">Math<T>::reduceAdd</a>(Math<T>::mul(a, b));</div>
|
|
<div class="line"><a name="l00039"></a><span class="lineno"> 39</span> }</div>
|
|
<div class="line"><a name="l00040"></a><span class="lineno"> 40</span> </div>
|
|
<div class="line"><a name="l00041"></a><span class="lineno"> 41</span> <span class="comment">// For list scanning, even if the input data is `half`, we perform all</span></div>
|
|
<div class="line"><a name="l00042"></a><span class="lineno"> 42</span> <span class="comment">// math in float32, because the code is memory b/w bound, and the</span></div>
|
|
<div class="line"><a name="l00043"></a><span class="lineno"> 43</span> <span class="comment">// added precision for accumulation is useful</span></div>
|
|
<div class="line"><a name="l00044"></a><span class="lineno"> 44</span> <span class="comment"></span></div>
|
|
<div class="line"><a name="l00045"></a><span class="lineno"> 45</span> <span class="comment">/// The class that we use to provide scan specializations</span></div>
|
|
<div class="line"><a name="l00046"></a><span class="lineno"> 46</span> <span class="comment"></span><span class="keyword">template</span> <<span class="keywordtype">int</span> Dims, <span class="keywordtype">bool</span> L2, <span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00047"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan.html"> 47</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a> {</div>
|
|
<div class="line"><a name="l00048"></a><span class="lineno"> 48</span> };</div>
|
|
<div class="line"><a name="l00049"></a><span class="lineno"> 49</span> </div>
|
|
<div class="line"><a name="l00050"></a><span class="lineno"> 50</span> <span class="comment">// Fallback implementation: works for any dimension size</span></div>
|
|
<div class="line"><a name="l00051"></a><span class="lineno"> 51</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2, <span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00052"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan_3-1_00_01L2_00_01T_01_4.html"> 52</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><-1, L2, T> {</div>
|
|
<div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  <span class="keyword">extern</span> __shared__ <span class="keywordtype">float</span> smem[];</div>
|
|
<div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  T* vecs = (T*) vecData;</div>
|
|
<div class="line"><a name="l00060"></a><span class="lineno"> 60</span> </div>
|
|
<div class="line"><a name="l00061"></a><span class="lineno"> 61</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> vec = 0; vec < numVecs; ++vec) {</div>
|
|
<div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  <span class="comment">// Reduce in dist</span></div>
|
|
<div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  <span class="keywordtype">float</span> dist = 0.0f;</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>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> d = threadIdx.x; d < dim; d += blockDim.x) {</div>
|
|
<div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keywordtype">float</span> vecVal = <a class="code" href="structfaiss_1_1gpu_1_1ConvertTo.html">ConvertTo<float>::to</a>(vecs[vec * dim + d]);</div>
|
|
<div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  <span class="keywordtype">float</span> queryVal = query[d];</div>
|
|
<div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keywordtype">float</span> curDist;</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="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  curDist = l2Distance(queryVal, vecVal);</div>
|
|
<div class="line"><a name="l00072"></a><span class="lineno"> 72</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  curDist = ipDistance(queryVal, vecVal);</div>
|
|
<div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  }</div>
|
|
<div class="line"><a name="l00075"></a><span class="lineno"> 75</span> </div>
|
|
<div class="line"><a name="l00076"></a><span class="lineno"> 76</span>  dist += curDist;</div>
|
|
<div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  }</div>
|
|
<div class="line"><a name="l00078"></a><span class="lineno"> 78</span> </div>
|
|
<div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  <span class="comment">// Reduce distance within block</span></div>
|
|
<div class="line"><a name="l00080"></a><span class="lineno"> 80</span>  dist = blockReduceAllSum<float, false, true>(dist, smem);</div>
|
|
<div class="line"><a name="l00081"></a><span class="lineno"> 81</span> </div>
|
|
<div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  <span class="keywordflow">if</span> (threadIdx.x == 0) {</div>
|
|
<div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  distanceOut[vec] = dist;</div>
|
|
<div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  }</div>
|
|
<div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  }</div>
|
|
<div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  }</div>
|
|
<div class="line"><a name="l00087"></a><span class="lineno"> 87</span> };</div>
|
|
<div class="line"><a name="l00088"></a><span class="lineno"> 88</span> </div>
|
|
<div class="line"><a name="l00089"></a><span class="lineno"> 89</span> <span class="comment">// implementation: works for # dims == blockDim.x</span></div>
|
|
<div class="line"><a name="l00090"></a><span class="lineno"> 90</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2, <span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00091"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan_3_010_00_01L2_00_01T_01_4.html"> 91</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><0, L2, T> {</div>
|
|
<div class="line"><a name="l00092"></a><span class="lineno"> 92</span>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00094"></a><span class="lineno"> 94</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  <span class="keyword">extern</span> __shared__ <span class="keywordtype">float</span> smem[];</div>
|
|
<div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  T* vecs = (T*) vecData;</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">float</span> queryVal = query[threadIdx.x];</div>
|
|
<div class="line"><a name="l00101"></a><span class="lineno"> 101</span> </div>
|
|
<div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</div>
|
|
<div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  <span class="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll);</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">for</span> (<span class="keywordtype">int</span> i = 0; i < limit; i += kUnroll) {</div>
|
|
<div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keywordtype">float</span> vecVal[kUnroll];</div>
|
|
<div class="line"><a name="l00107"></a><span class="lineno"> 107</span> </div>
|
|
<div class="line"><a name="l00108"></a><span class="lineno"> 108</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00109"></a><span class="lineno"> 109</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  vecVal[j] = <a class="code" href="structfaiss_1_1gpu_1_1ConvertTo.html">ConvertTo<float>::to</a>(vecs[(i + j) * dim + threadIdx.x]);</div>
|
|
<div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  }</div>
|
|
<div class="line"><a name="l00112"></a><span class="lineno"> 112</span> </div>
|
|
<div class="line"><a name="l00113"></a><span class="lineno"> 113</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00114"></a><span class="lineno"> 114</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  vecVal[j] = l2Distance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00117"></a><span class="lineno"> 117</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  vecVal[j] = ipDistance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  }</div>
|
|
<div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  }</div>
|
|
<div class="line"><a name="l00121"></a><span class="lineno"> 121</span> </div>
|
|
<div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  blockReduceAllSum<kUnroll, float, false, true>(vecVal, smem);</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>  <span class="keywordflow">if</span> (threadIdx.x == 0) {</div>
|
|
<div class="line"><a name="l00125"></a><span class="lineno"> 125</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00126"></a><span class="lineno"> 126</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  distanceOut[i + j] = vecVal[j];</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>  }</div>
|
|
<div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  }</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">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit; i < numVecs; ++i) {</div>
|
|
<div class="line"><a name="l00134"></a><span class="lineno"> 134</span>  <span class="keywordtype">float</span> vecVal = <a class="code" href="structfaiss_1_1gpu_1_1ConvertTo.html">ConvertTo<float>::to</a>(vecs[i * dim + threadIdx.x]);</div>
|
|
<div class="line"><a name="l00135"></a><span class="lineno"> 135</span> </div>
|
|
<div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00137"></a><span class="lineno"> 137</span>  vecVal = l2Distance(queryVal, vecVal);</div>
|
|
<div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  vecVal = ipDistance(queryVal, vecVal);</div>
|
|
<div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  }</div>
|
|
<div class="line"><a name="l00141"></a><span class="lineno"> 141</span> </div>
|
|
<div class="line"><a name="l00142"></a><span class="lineno"> 142</span>  vecVal = blockReduceAllSum<float, false, true>(vecVal, smem);</div>
|
|
<div class="line"><a name="l00143"></a><span class="lineno"> 143</span> </div>
|
|
<div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  <span class="keywordflow">if</span> (threadIdx.x == 0) {</div>
|
|
<div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  distanceOut[i] = vecVal;</div>
|
|
<div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  }</div>
|
|
<div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  }</div>
|
|
<div class="line"><a name="l00148"></a><span class="lineno"> 148</span>  }</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> </div>
|
|
<div class="line"><a name="l00151"></a><span class="lineno"> 151</span> <span class="comment">// 64-d float32 implementation</span></div>
|
|
<div class="line"><a name="l00152"></a><span class="lineno"> 152</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00153"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan_3_0164_00_01L2_00_01float_01_4.html"> 153</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><64, L2, float> {</div>
|
|
<div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 64;</div>
|
|
<div class="line"><a name="l00155"></a><span class="lineno"> 155</span> </div>
|
|
<div class="line"><a name="l00156"></a><span class="lineno"> 156</span>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  <span class="comment">// Each warp reduces a single 64-d vector; each lane loads a float2</span></div>
|
|
<div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  <span class="keywordtype">float</span>* vecs = (<span class="keywordtype">float</span>*) vecData;</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>  <span class="keywordtype">int</span> laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="keywordtype">int</span> warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  <span class="keywordtype">int</span> numWarps = blockDim.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00167"></a><span class="lineno"> 167</span> </div>
|
|
<div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  float2 queryVal = *(float2*) &query[laneId * 2];</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>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</div>
|
|
<div class="line"><a name="l00171"></a><span class="lineno"> 171</span>  float2 vecVal[kUnroll];</div>
|
|
<div class="line"><a name="l00172"></a><span class="lineno"> 172</span> </div>
|
|
<div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  <span class="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll * numWarps);</div>
|
|
<div class="line"><a name="l00174"></a><span class="lineno"> 174</span> </div>
|
|
<div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = warpId; i < limit; i += kUnroll * numWarps) {</div>
|
|
<div class="line"><a name="l00176"></a><span class="lineno"> 176</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00177"></a><span class="lineno"> 177</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  <span class="comment">// Vector we are loading from is i</span></div>
|
|
<div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  <span class="comment">// Dim we are loading from is laneId * 2</span></div>
|
|
<div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  vecVal[j] = *(float2*) &vecs[(i + j * numWarps) * kDims + laneId * 2];</div>
|
|
<div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  }</div>
|
|
<div class="line"><a name="l00182"></a><span class="lineno"> 182</span> </div>
|
|
<div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  <span class="keywordtype">float</span> dist[kUnroll];</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="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00186"></a><span class="lineno"> 186</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  dist[j] = l2Distance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00189"></a><span class="lineno"> 189</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  dist[j] = ipDistance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  }</div>
|
|
<div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  }</div>
|
|
<div class="line"><a name="l00193"></a><span class="lineno"> 193</span> </div>
|
|
<div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  <span class="comment">// Reduce within the warp</span></div>
|
|
<div class="line"><a name="l00195"></a><span class="lineno"> 195</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00196"></a><span class="lineno"> 196</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  dist[j] = warpReduceAllSum(dist[j]);</div>
|
|
<div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  }</div>
|
|
<div class="line"><a name="l00199"></a><span class="lineno"> 199</span> </div>
|
|
<div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00201"></a><span class="lineno"> 201</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00202"></a><span class="lineno"> 202</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  distanceOut[i + j * numWarps] = dist[j];</div>
|
|
<div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  }</div>
|
|
<div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  }</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> </div>
|
|
<div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <span class="comment">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit + warpId; i < numVecs; i += numWarps) {</div>
|
|
<div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  vecVal[0] = *(float2*) &vecs[i * kDims + laneId * 2];</div>
|
|
<div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  <span class="keywordtype">float</span> dist;</div>
|
|
<div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  dist = l2Distance(queryVal, vecVal[0]);</div>
|
|
<div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  dist = ipDistance(queryVal, vecVal[0]);</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> </div>
|
|
<div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  dist = warpReduceAllSum(dist);</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="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  distanceOut[i] = dist;</div>
|
|
<div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  }</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>  }</div>
|
|
<div class="line"><a name="l00225"></a><span class="lineno"> 225</span> };</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="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00228"></a><span class="lineno"> 228</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00229"></a><span class="lineno"> 229</span> <span class="comment">// float16 implementation</span></div>
|
|
<div class="line"><a name="l00230"></a><span class="lineno"> 230</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00231"></a><span class="lineno"> 231</span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><64, L2, half> {</div>
|
|
<div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 64;</div>
|
|
<div class="line"><a name="l00233"></a><span class="lineno"> 233</span> </div>
|
|
<div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00236"></a><span class="lineno"> 236</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  <span class="comment">// Each warp reduces a single 64-d vector; each lane loads a half2</span></div>
|
|
<div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  half* vecs = (half*) vecData;</div>
|
|
<div class="line"><a name="l00241"></a><span class="lineno"> 241</span> </div>
|
|
<div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  <span class="keywordtype">int</span> laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  <span class="keywordtype">int</span> warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  <span class="keywordtype">int</span> numWarps = blockDim.x / kWarpSize;</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>  float2 queryVal = *(float2*) &query[laneId * 2];</div>
|
|
<div class="line"><a name="l00247"></a><span class="lineno"> 247</span> </div>
|
|
<div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</div>
|
|
<div class="line"><a name="l00249"></a><span class="lineno"> 249</span> </div>
|
|
<div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  half2 vecVal[kUnroll];</div>
|
|
<div class="line"><a name="l00251"></a><span class="lineno"> 251</span> </div>
|
|
<div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  <span class="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll * numWarps);</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>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = warpId; i < limit; i += kUnroll * numWarps) {</div>
|
|
<div class="line"><a name="l00255"></a><span class="lineno"> 255</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00256"></a><span class="lineno"> 256</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  <span class="comment">// Vector we are loading from is i</span></div>
|
|
<div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  <span class="comment">// Dim we are loading from is laneId * 2</span></div>
|
|
<div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  vecVal[j] = *(half2*) &vecs[(i + j * numWarps) * kDims + laneId * 2];</div>
|
|
<div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  }</div>
|
|
<div class="line"><a name="l00261"></a><span class="lineno"> 261</span> </div>
|
|
<div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  <span class="keywordtype">float</span> dist[kUnroll];</div>
|
|
<div class="line"><a name="l00263"></a><span class="lineno"> 263</span> </div>
|
|
<div class="line"><a name="l00264"></a><span class="lineno"> 264</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00265"></a><span class="lineno"> 265</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00267"></a><span class="lineno"> 267</span>  dist[j] = l2Distance(queryVal, __half22float2(vecVal[j]));</div>
|
|
<div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  dist[j] = ipDistance(queryVal, __half22float2(vecVal[j]));</div>
|
|
<div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  }</div>
|
|
<div class="line"><a name="l00271"></a><span class="lineno"> 271</span>  }</div>
|
|
<div class="line"><a name="l00272"></a><span class="lineno"> 272</span> </div>
|
|
<div class="line"><a name="l00273"></a><span class="lineno"> 273</span>  <span class="comment">// Reduce within the warp</span></div>
|
|
<div class="line"><a name="l00274"></a><span class="lineno"> 274</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00275"></a><span class="lineno"> 275</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  dist[j] = warpReduceAllSum(dist[j]);</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> </div>
|
|
<div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00280"></a><span class="lineno"> 280</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00281"></a><span class="lineno"> 281</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00282"></a><span class="lineno"> 282</span>  distanceOut[i + j * numWarps] = dist[j];</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>  }</div>
|
|
<div class="line"><a name="l00286"></a><span class="lineno"> 286</span> </div>
|
|
<div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  <span class="comment">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00288"></a><span class="lineno"> 288</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit + warpId; i < numVecs; i += numWarps) {</div>
|
|
<div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  vecVal[0] = *(half2*) &vecs[i * kDims + laneId * 2];</div>
|
|
<div class="line"><a name="l00290"></a><span class="lineno"> 290</span> </div>
|
|
<div class="line"><a name="l00291"></a><span class="lineno"> 291</span>  <span class="keywordtype">float</span> dist;</div>
|
|
<div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00293"></a><span class="lineno"> 293</span>  dist = l2Distance(queryVal, __half22float2(vecVal[0]));</div>
|
|
<div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  dist = ipDistance(queryVal, __half22float2(vecVal[0]));</div>
|
|
<div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  }</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>  dist = warpReduceAllSum(dist);</div>
|
|
<div class="line"><a name="l00299"></a><span class="lineno"> 299</span> </div>
|
|
<div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  distanceOut[i] = dist;</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>  }</div>
|
|
<div class="line"><a name="l00304"></a><span class="lineno"> 304</span>  }</div>
|
|
<div class="line"><a name="l00305"></a><span class="lineno"> 305</span> };</div>
|
|
<div class="line"><a name="l00306"></a><span class="lineno"> 306</span> </div>
|
|
<div class="line"><a name="l00307"></a><span class="lineno"> 307</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00308"></a><span class="lineno"> 308</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00309"></a><span class="lineno"> 309</span> <span class="comment">// 128-d float32 implementation</span></div>
|
|
<div class="line"><a name="l00310"></a><span class="lineno"> 310</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00311"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan_3_01128_00_01L2_00_01float_01_4.html"> 311</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><128, L2, float> {</div>
|
|
<div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 128;</div>
|
|
<div class="line"><a name="l00313"></a><span class="lineno"> 313</span> </div>
|
|
<div class="line"><a name="l00314"></a><span class="lineno"> 314</span>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00315"></a><span class="lineno"> 315</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00316"></a><span class="lineno"> 316</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00317"></a><span class="lineno"> 317</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00318"></a><span class="lineno"> 318</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00319"></a><span class="lineno"> 319</span>  <span class="comment">// Each warp reduces a single 128-d vector; each lane loads a float4</span></div>
|
|
<div class="line"><a name="l00320"></a><span class="lineno"> 320</span>  <span class="keywordtype">float</span>* vecs = (<span class="keywordtype">float</span>*) vecData;</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>  <span class="keywordtype">int</span> laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00323"></a><span class="lineno"> 323</span>  <span class="keywordtype">int</span> warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00324"></a><span class="lineno"> 324</span>  <span class="keywordtype">int</span> numWarps = blockDim.x / kWarpSize;</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>  float4 queryVal = *(float4*) &query[laneId * 4];</div>
|
|
<div class="line"><a name="l00327"></a><span class="lineno"> 327</span> </div>
|
|
<div class="line"><a name="l00328"></a><span class="lineno"> 328</span>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</div>
|
|
<div class="line"><a name="l00329"></a><span class="lineno"> 329</span>  float4 vecVal[kUnroll];</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="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll * numWarps);</div>
|
|
<div class="line"><a name="l00332"></a><span class="lineno"> 332</span> </div>
|
|
<div class="line"><a name="l00333"></a><span class="lineno"> 333</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = warpId; i < limit; i += kUnroll * numWarps) {</div>
|
|
<div class="line"><a name="l00334"></a><span class="lineno"> 334</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00335"></a><span class="lineno"> 335</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00336"></a><span class="lineno"> 336</span>  <span class="comment">// Vector we are loading from is i</span></div>
|
|
<div class="line"><a name="l00337"></a><span class="lineno"> 337</span>  <span class="comment">// Dim we are loading from is laneId * 4</span></div>
|
|
<div class="line"><a name="l00338"></a><span class="lineno"> 338</span>  vecVal[j] = *(float4*) &vecs[(i + j * numWarps) * kDims + laneId * 4];</div>
|
|
<div class="line"><a name="l00339"></a><span class="lineno"> 339</span>  }</div>
|
|
<div class="line"><a name="l00340"></a><span class="lineno"> 340</span> </div>
|
|
<div class="line"><a name="l00341"></a><span class="lineno"> 341</span>  <span class="keywordtype">float</span> dist[kUnroll];</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="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00344"></a><span class="lineno"> 344</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00345"></a><span class="lineno"> 345</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00346"></a><span class="lineno"> 346</span>  dist[j] = l2Distance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00347"></a><span class="lineno"> 347</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00348"></a><span class="lineno"> 348</span>  dist[j] = ipDistance(queryVal, vecVal[j]);</div>
|
|
<div class="line"><a name="l00349"></a><span class="lineno"> 349</span>  }</div>
|
|
<div class="line"><a name="l00350"></a><span class="lineno"> 350</span>  }</div>
|
|
<div class="line"><a name="l00351"></a><span class="lineno"> 351</span> </div>
|
|
<div class="line"><a name="l00352"></a><span class="lineno"> 352</span>  <span class="comment">// Reduce within the warp</span></div>
|
|
<div class="line"><a name="l00353"></a><span class="lineno"> 353</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00354"></a><span class="lineno"> 354</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00355"></a><span class="lineno"> 355</span>  dist[j] = warpReduceAllSum(dist[j]);</div>
|
|
<div class="line"><a name="l00356"></a><span class="lineno"> 356</span>  }</div>
|
|
<div class="line"><a name="l00357"></a><span class="lineno"> 357</span> </div>
|
|
<div class="line"><a name="l00358"></a><span class="lineno"> 358</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00359"></a><span class="lineno"> 359</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00360"></a><span class="lineno"> 360</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00361"></a><span class="lineno"> 361</span>  distanceOut[i + j * numWarps] = dist[j];</div>
|
|
<div class="line"><a name="l00362"></a><span class="lineno"> 362</span>  }</div>
|
|
<div class="line"><a name="l00363"></a><span class="lineno"> 363</span>  }</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> </div>
|
|
<div class="line"><a name="l00366"></a><span class="lineno"> 366</span>  <span class="comment">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00367"></a><span class="lineno"> 367</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit + warpId; i < numVecs; i += numWarps) {</div>
|
|
<div class="line"><a name="l00368"></a><span class="lineno"> 368</span>  vecVal[0] = *(float4*) &vecs[i * kDims + laneId * 4];</div>
|
|
<div class="line"><a name="l00369"></a><span class="lineno"> 369</span>  <span class="keywordtype">float</span> dist;</div>
|
|
<div class="line"><a name="l00370"></a><span class="lineno"> 370</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00371"></a><span class="lineno"> 371</span>  dist = l2Distance(queryVal, vecVal[0]);</div>
|
|
<div class="line"><a name="l00372"></a><span class="lineno"> 372</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00373"></a><span class="lineno"> 373</span>  dist = ipDistance(queryVal, vecVal[0]);</div>
|
|
<div class="line"><a name="l00374"></a><span class="lineno"> 374</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>  dist = warpReduceAllSum(dist);</div>
|
|
<div class="line"><a name="l00377"></a><span class="lineno"> 377</span> </div>
|
|
<div class="line"><a name="l00378"></a><span class="lineno"> 378</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00379"></a><span class="lineno"> 379</span>  distanceOut[i] = dist;</div>
|
|
<div class="line"><a name="l00380"></a><span class="lineno"> 380</span>  }</div>
|
|
<div class="line"><a name="l00381"></a><span class="lineno"> 381</span>  }</div>
|
|
<div class="line"><a name="l00382"></a><span class="lineno"> 382</span>  }</div>
|
|
<div class="line"><a name="l00383"></a><span class="lineno"> 383</span> };</div>
|
|
<div class="line"><a name="l00384"></a><span class="lineno"> 384</span> </div>
|
|
<div class="line"><a name="l00385"></a><span class="lineno"> 385</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00386"></a><span class="lineno"> 386</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00387"></a><span class="lineno"> 387</span> <span class="comment">// float16 implementation</span></div>
|
|
<div class="line"><a name="l00388"></a><span class="lineno"> 388</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00389"></a><span class="lineno"> 389</span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><128, L2, half> {</div>
|
|
<div class="line"><a name="l00390"></a><span class="lineno"> 390</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 128;</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="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00393"></a><span class="lineno"> 393</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00394"></a><span class="lineno"> 394</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00395"></a><span class="lineno"> 395</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00396"></a><span class="lineno"> 396</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00397"></a><span class="lineno"> 397</span>  <span class="comment">// Each warp reduces a single 128-d vector; each lane loads a Half4</span></div>
|
|
<div class="line"><a name="l00398"></a><span class="lineno"> 398</span>  half* vecs = (half*) vecData;</div>
|
|
<div class="line"><a name="l00399"></a><span class="lineno"> 399</span> </div>
|
|
<div class="line"><a name="l00400"></a><span class="lineno"> 400</span>  <span class="keywordtype">int</span> laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00401"></a><span class="lineno"> 401</span>  <span class="keywordtype">int</span> warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00402"></a><span class="lineno"> 402</span>  <span class="keywordtype">int</span> numWarps = blockDim.x / kWarpSize;</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>  float4 queryVal = *(float4*) &query[laneId * 4];</div>
|
|
<div class="line"><a name="l00405"></a><span class="lineno"> 405</span> </div>
|
|
<div class="line"><a name="l00406"></a><span class="lineno"> 406</span>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</div>
|
|
<div class="line"><a name="l00407"></a><span class="lineno"> 407</span> </div>
|
|
<div class="line"><a name="l00408"></a><span class="lineno"> 408</span>  Half4 vecVal[kUnroll];</div>
|
|
<div class="line"><a name="l00409"></a><span class="lineno"> 409</span> </div>
|
|
<div class="line"><a name="l00410"></a><span class="lineno"> 410</span>  <span class="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll * numWarps);</div>
|
|
<div class="line"><a name="l00411"></a><span class="lineno"> 411</span> </div>
|
|
<div class="line"><a name="l00412"></a><span class="lineno"> 412</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = warpId; i < limit; i += kUnroll * numWarps) {</div>
|
|
<div class="line"><a name="l00413"></a><span class="lineno"> 413</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00414"></a><span class="lineno"> 414</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00415"></a><span class="lineno"> 415</span>  <span class="comment">// Vector we are loading from is i</span></div>
|
|
<div class="line"><a name="l00416"></a><span class="lineno"> 416</span>  <span class="comment">// Dim we are loading from is laneId * 4</span></div>
|
|
<div class="line"><a name="l00417"></a><span class="lineno"> 417</span>  vecVal[j] =</div>
|
|
<div class="line"><a name="l00418"></a><span class="lineno"> 418</span>  <a class="code" href="structfaiss_1_1gpu_1_1LoadStore.html">LoadStore<Half4>::load</a>(</div>
|
|
<div class="line"><a name="l00419"></a><span class="lineno"> 419</span>  &vecs[(i + j * numWarps) * kDims + laneId * 4]);</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> </div>
|
|
<div class="line"><a name="l00422"></a><span class="lineno"> 422</span>  <span class="keywordtype">float</span> dist[kUnroll];</div>
|
|
<div class="line"><a name="l00423"></a><span class="lineno"> 423</span> </div>
|
|
<div class="line"><a name="l00424"></a><span class="lineno"> 424</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00425"></a><span class="lineno"> 425</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00426"></a><span class="lineno"> 426</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00427"></a><span class="lineno"> 427</span>  dist[j] = l2Distance(queryVal, half4ToFloat4(vecVal[j]));</div>
|
|
<div class="line"><a name="l00428"></a><span class="lineno"> 428</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00429"></a><span class="lineno"> 429</span>  dist[j] = ipDistance(queryVal, half4ToFloat4(vecVal[j]));</div>
|
|
<div class="line"><a name="l00430"></a><span class="lineno"> 430</span>  }</div>
|
|
<div class="line"><a name="l00431"></a><span class="lineno"> 431</span>  }</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="comment">// Reduce within the warp</span></div>
|
|
<div class="line"><a name="l00434"></a><span class="lineno"> 434</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00435"></a><span class="lineno"> 435</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00436"></a><span class="lineno"> 436</span>  dist[j] = warpReduceAllSum(dist[j]);</div>
|
|
<div class="line"><a name="l00437"></a><span class="lineno"> 437</span>  }</div>
|
|
<div class="line"><a name="l00438"></a><span class="lineno"> 438</span> </div>
|
|
<div class="line"><a name="l00439"></a><span class="lineno"> 439</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00440"></a><span class="lineno"> 440</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00441"></a><span class="lineno"> 441</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00442"></a><span class="lineno"> 442</span>  distanceOut[i + j * numWarps] = dist[j];</div>
|
|
<div class="line"><a name="l00443"></a><span class="lineno"> 443</span>  }</div>
|
|
<div class="line"><a name="l00444"></a><span class="lineno"> 444</span>  }</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> </div>
|
|
<div class="line"><a name="l00447"></a><span class="lineno"> 447</span>  <span class="comment">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00448"></a><span class="lineno"> 448</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit + warpId; i < numVecs; i += numWarps) {</div>
|
|
<div class="line"><a name="l00449"></a><span class="lineno"> 449</span>  vecVal[0] = LoadStore<Half4>::load(&vecs[i * kDims + laneId * 4]);</div>
|
|
<div class="line"><a name="l00450"></a><span class="lineno"> 450</span> </div>
|
|
<div class="line"><a name="l00451"></a><span class="lineno"> 451</span>  <span class="keywordtype">float</span> dist;</div>
|
|
<div class="line"><a name="l00452"></a><span class="lineno"> 452</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00453"></a><span class="lineno"> 453</span>  dist = l2Distance(queryVal, half4ToFloat4(vecVal[0]));</div>
|
|
<div class="line"><a name="l00454"></a><span class="lineno"> 454</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00455"></a><span class="lineno"> 455</span>  dist = ipDistance(queryVal, half4ToFloat4(vecVal[0]));</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> </div>
|
|
<div class="line"><a name="l00458"></a><span class="lineno"> 458</span>  dist = warpReduceAllSum(dist);</div>
|
|
<div class="line"><a name="l00459"></a><span class="lineno"> 459</span> </div>
|
|
<div class="line"><a name="l00460"></a><span class="lineno"> 460</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00461"></a><span class="lineno"> 461</span>  distanceOut[i] = dist;</div>
|
|
<div class="line"><a name="l00462"></a><span class="lineno"> 462</span>  }</div>
|
|
<div class="line"><a name="l00463"></a><span class="lineno"> 463</span>  }</div>
|
|
<div class="line"><a name="l00464"></a><span class="lineno"> 464</span>  }</div>
|
|
<div class="line"><a name="l00465"></a><span class="lineno"> 465</span> };</div>
|
|
<div class="line"><a name="l00466"></a><span class="lineno"> 466</span> </div>
|
|
<div class="line"><a name="l00467"></a><span class="lineno"> 467</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00468"></a><span class="lineno"> 468</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00469"></a><span class="lineno"> 469</span> <span class="comment">// 256-d float32 implementation</span></div>
|
|
<div class="line"><a name="l00470"></a><span class="lineno"> 470</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00471"></a><span class="lineno"><a class="line" href="structfaiss_1_1gpu_1_1IVFFlatScan_3_01256_00_01L2_00_01float_01_4.html"> 471</a></span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><256, L2, float> {</div>
|
|
<div class="line"><a name="l00472"></a><span class="lineno"> 472</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 256;</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>  <span class="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00475"></a><span class="lineno"> 475</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00476"></a><span class="lineno"> 476</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00477"></a><span class="lineno"> 477</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00478"></a><span class="lineno"> 478</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00479"></a><span class="lineno"> 479</span>  <span class="comment">// A specialization here to load per-warp seems to be worse, since</span></div>
|
|
<div class="line"><a name="l00480"></a><span class="lineno"> 480</span>  <span class="comment">// we're already running at near memory b/w peak</span></div>
|
|
<div class="line"><a name="l00481"></a><span class="lineno"> 481</span>  <a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan<0, L2, float>::scan</a>(query,</div>
|
|
<div class="line"><a name="l00482"></a><span class="lineno"> 482</span>  vecData,</div>
|
|
<div class="line"><a name="l00483"></a><span class="lineno"> 483</span>  numVecs,</div>
|
|
<div class="line"><a name="l00484"></a><span class="lineno"> 484</span>  dim,</div>
|
|
<div class="line"><a name="l00485"></a><span class="lineno"> 485</span>  distanceOut);</div>
|
|
<div class="line"><a name="l00486"></a><span class="lineno"> 486</span>  }</div>
|
|
<div class="line"><a name="l00487"></a><span class="lineno"> 487</span> };</div>
|
|
<div class="line"><a name="l00488"></a><span class="lineno"> 488</span> </div>
|
|
<div class="line"><a name="l00489"></a><span class="lineno"> 489</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00490"></a><span class="lineno"> 490</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00491"></a><span class="lineno"> 491</span> <span class="comment">// float16 implementation</span></div>
|
|
<div class="line"><a name="l00492"></a><span class="lineno"> 492</span> <span class="keyword">template</span> <<span class="keywordtype">bool</span> L2></div>
|
|
<div class="line"><a name="l00493"></a><span class="lineno"> 493</span> <span class="keyword">struct </span><a class="code" href="structfaiss_1_1gpu_1_1IVFFlatScan.html">IVFFlatScan</a><256, L2, half> {</div>
|
|
<div class="line"><a name="l00494"></a><span class="lineno"> 494</span>  <span class="keyword">static</span> constexpr <span class="keywordtype">int</span> kDims = 256;</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="keyword">static</span> __device__ <span class="keywordtype">void</span> scan(<span class="keywordtype">float</span>* query,</div>
|
|
<div class="line"><a name="l00497"></a><span class="lineno"> 497</span>  <span class="keywordtype">void</span>* vecData,</div>
|
|
<div class="line"><a name="l00498"></a><span class="lineno"> 498</span>  <span class="keywordtype">int</span> numVecs,</div>
|
|
<div class="line"><a name="l00499"></a><span class="lineno"> 499</span>  <span class="keywordtype">int</span> dim,</div>
|
|
<div class="line"><a name="l00500"></a><span class="lineno"> 500</span>  <span class="keywordtype">float</span>* distanceOut) {</div>
|
|
<div class="line"><a name="l00501"></a><span class="lineno"> 501</span>  <span class="comment">// Each warp reduces a single 256-d vector; each lane loads a Half8</span></div>
|
|
<div class="line"><a name="l00502"></a><span class="lineno"> 502</span>  half* vecs = (half*) vecData;</div>
|
|
<div class="line"><a name="l00503"></a><span class="lineno"> 503</span> </div>
|
|
<div class="line"><a name="l00504"></a><span class="lineno"> 504</span>  <span class="keywordtype">int</span> laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00505"></a><span class="lineno"> 505</span>  <span class="keywordtype">int</span> warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00506"></a><span class="lineno"> 506</span>  <span class="keywordtype">int</span> numWarps = blockDim.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00507"></a><span class="lineno"> 507</span> </div>
|
|
<div class="line"><a name="l00508"></a><span class="lineno"> 508</span>  <span class="comment">// This is not a contiguous load, but we only have to load these two</span></div>
|
|
<div class="line"><a name="l00509"></a><span class="lineno"> 509</span>  <span class="comment">// values, so that we can load by Half8 below</span></div>
|
|
<div class="line"><a name="l00510"></a><span class="lineno"> 510</span>  float4 queryValA = *(float4*) &query[laneId * 8];</div>
|
|
<div class="line"><a name="l00511"></a><span class="lineno"> 511</span>  float4 queryValB = *(float4*) &query[laneId * 8 + 4];</div>
|
|
<div class="line"><a name="l00512"></a><span class="lineno"> 512</span> </div>
|
|
<div class="line"><a name="l00513"></a><span class="lineno"> 513</span>  constexpr <span class="keywordtype">int</span> kUnroll = 4;</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>  Half8 vecVal[kUnroll];</div>
|
|
<div class="line"><a name="l00516"></a><span class="lineno"> 516</span> </div>
|
|
<div class="line"><a name="l00517"></a><span class="lineno"> 517</span>  <span class="keywordtype">int</span> limit = utils::roundDown(numVecs, kUnroll * numWarps);</div>
|
|
<div class="line"><a name="l00518"></a><span class="lineno"> 518</span> </div>
|
|
<div class="line"><a name="l00519"></a><span class="lineno"> 519</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = warpId; i < limit; i += kUnroll * numWarps) {</div>
|
|
<div class="line"><a name="l00520"></a><span class="lineno"> 520</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00521"></a><span class="lineno"> 521</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00522"></a><span class="lineno"> 522</span>  <span class="comment">// Vector we are loading from is i</span></div>
|
|
<div class="line"><a name="l00523"></a><span class="lineno"> 523</span>  <span class="comment">// Dim we are loading from is laneId * 8</span></div>
|
|
<div class="line"><a name="l00524"></a><span class="lineno"> 524</span>  vecVal[j] =</div>
|
|
<div class="line"><a name="l00525"></a><span class="lineno"> 525</span>  <a class="code" href="structfaiss_1_1gpu_1_1LoadStore.html">LoadStore<Half8>::load</a>(</div>
|
|
<div class="line"><a name="l00526"></a><span class="lineno"> 526</span>  &vecs[(i + j * numWarps) * kDims + laneId * 8]);</div>
|
|
<div class="line"><a name="l00527"></a><span class="lineno"> 527</span>  }</div>
|
|
<div class="line"><a name="l00528"></a><span class="lineno"> 528</span> </div>
|
|
<div class="line"><a name="l00529"></a><span class="lineno"> 529</span>  <span class="keywordtype">float</span> dist[kUnroll];</div>
|
|
<div class="line"><a name="l00530"></a><span class="lineno"> 530</span> </div>
|
|
<div class="line"><a name="l00531"></a><span class="lineno"> 531</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00532"></a><span class="lineno"> 532</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00533"></a><span class="lineno"> 533</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00534"></a><span class="lineno"> 534</span>  dist[j] = l2Distance(queryValA, half4ToFloat4(vecVal[j].a));</div>
|
|
<div class="line"><a name="l00535"></a><span class="lineno"> 535</span>  dist[j] += l2Distance(queryValB, half4ToFloat4(vecVal[j].b));</div>
|
|
<div class="line"><a name="l00536"></a><span class="lineno"> 536</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00537"></a><span class="lineno"> 537</span>  dist[j] = ipDistance(queryValA, half4ToFloat4(vecVal[j].a));</div>
|
|
<div class="line"><a name="l00538"></a><span class="lineno"> 538</span>  dist[j] += ipDistance(queryValB, half4ToFloat4(vecVal[j].b));</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>  }</div>
|
|
<div class="line"><a name="l00541"></a><span class="lineno"> 541</span> </div>
|
|
<div class="line"><a name="l00542"></a><span class="lineno"> 542</span>  <span class="comment">// Reduce within the warp</span></div>
|
|
<div class="line"><a name="l00543"></a><span class="lineno"> 543</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00544"></a><span class="lineno"> 544</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00545"></a><span class="lineno"> 545</span>  dist[j] = warpReduceAllSum(dist[j]);</div>
|
|
<div class="line"><a name="l00546"></a><span class="lineno"> 546</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> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00549"></a><span class="lineno"> 549</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00550"></a><span class="lineno"> 550</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < kUnroll; ++j) {</div>
|
|
<div class="line"><a name="l00551"></a><span class="lineno"> 551</span>  distanceOut[i + j * numWarps] = dist[j];</div>
|
|
<div class="line"><a name="l00552"></a><span class="lineno"> 552</span>  }</div>
|
|
<div class="line"><a name="l00553"></a><span class="lineno"> 553</span>  }</div>
|
|
<div class="line"><a name="l00554"></a><span class="lineno"> 554</span>  }</div>
|
|
<div class="line"><a name="l00555"></a><span class="lineno"> 555</span> </div>
|
|
<div class="line"><a name="l00556"></a><span class="lineno"> 556</span>  <span class="comment">// Handle remainder</span></div>
|
|
<div class="line"><a name="l00557"></a><span class="lineno"> 557</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = limit + warpId; i < numVecs; i += numWarps) {</div>
|
|
<div class="line"><a name="l00558"></a><span class="lineno"> 558</span>  vecVal[0] = LoadStore<Half8>::load(&vecs[i * kDims + laneId * 8]);</div>
|
|
<div class="line"><a name="l00559"></a><span class="lineno"> 559</span> </div>
|
|
<div class="line"><a name="l00560"></a><span class="lineno"> 560</span>  <span class="keywordtype">float</span> dist;</div>
|
|
<div class="line"><a name="l00561"></a><span class="lineno"> 561</span>  <span class="keywordflow">if</span> (L2) {</div>
|
|
<div class="line"><a name="l00562"></a><span class="lineno"> 562</span>  dist = l2Distance(queryValA, half4ToFloat4(vecVal[0].a));</div>
|
|
<div class="line"><a name="l00563"></a><span class="lineno"> 563</span>  dist += l2Distance(queryValB, half4ToFloat4(vecVal[0].b));</div>
|
|
<div class="line"><a name="l00564"></a><span class="lineno"> 564</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00565"></a><span class="lineno"> 565</span>  dist = ipDistance(queryValA, half4ToFloat4(vecVal[0].a));</div>
|
|
<div class="line"><a name="l00566"></a><span class="lineno"> 566</span>  dist += ipDistance(queryValB, half4ToFloat4(vecVal[0].b));</div>
|
|
<div class="line"><a name="l00567"></a><span class="lineno"> 567</span>  }</div>
|
|
<div class="line"><a name="l00568"></a><span class="lineno"> 568</span> </div>
|
|
<div class="line"><a name="l00569"></a><span class="lineno"> 569</span>  dist = warpReduceAllSum(dist);</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>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00572"></a><span class="lineno"> 572</span>  distanceOut[i] = dist;</div>
|
|
<div class="line"><a name="l00573"></a><span class="lineno"> 573</span>  }</div>
|
|
<div class="line"><a name="l00574"></a><span class="lineno"> 574</span>  }</div>
|
|
<div class="line"><a name="l00575"></a><span class="lineno"> 575</span>  }</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> </div>
|
|
<div class="line"><a name="l00578"></a><span class="lineno"> 578</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00579"></a><span class="lineno"> 579</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00580"></a><span class="lineno"> 580</span> <span class="keyword">template</span> <<span class="keywordtype">int</span> Dims, <span class="keywordtype">bool</span> L2, <span class="keyword">typename</span> T></div>
|
|
<div class="line"><a name="l00581"></a><span class="lineno"> 581</span> __global__ <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00582"></a><span class="lineno"> 582</span> ivfFlatScan(Tensor<float, 2, true> queries,</div>
|
|
<div class="line"><a name="l00583"></a><span class="lineno"> 583</span>  Tensor<int, 2, true> listIds,</div>
|
|
<div class="line"><a name="l00584"></a><span class="lineno"> 584</span>  <span class="keywordtype">void</span>** allListData,</div>
|
|
<div class="line"><a name="l00585"></a><span class="lineno"> 585</span>  <span class="keywordtype">int</span>* listLengths,</div>
|
|
<div class="line"><a name="l00586"></a><span class="lineno"> 586</span>  Tensor<int, 2, true> prefixSumOffsets,</div>
|
|
<div class="line"><a name="l00587"></a><span class="lineno"> 587</span>  Tensor<float, 1, true> distance) {</div>
|
|
<div class="line"><a name="l00588"></a><span class="lineno"> 588</span>  <span class="keyword">auto</span> queryId = blockIdx.y;</div>
|
|
<div class="line"><a name="l00589"></a><span class="lineno"> 589</span>  <span class="keyword">auto</span> probeId = blockIdx.x;</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>  <span class="comment">// This is where we start writing out data</span></div>
|
|
<div class="line"><a name="l00592"></a><span class="lineno"> 592</span>  <span class="comment">// We ensure that before the array (at offset -1), there is a 0 value</span></div>
|
|
<div class="line"><a name="l00593"></a><span class="lineno"> 593</span>  <span class="keywordtype">int</span> outBase = *(prefixSumOffsets[queryId][probeId].data() - 1);</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>  <span class="keyword">auto</span> listId = listIds[queryId][probeId];</div>
|
|
<div class="line"><a name="l00596"></a><span class="lineno"> 596</span>  <span class="comment">// Safety guard in case NaNs in input cause no list ID to be generated</span></div>
|
|
<div class="line"><a name="l00597"></a><span class="lineno"> 597</span>  <span class="keywordflow">if</span> (listId == -1) {</div>
|
|
<div class="line"><a name="l00598"></a><span class="lineno"> 598</span>  <span class="keywordflow">return</span>;</div>
|
|
<div class="line"><a name="l00599"></a><span class="lineno"> 599</span>  }</div>
|
|
<div class="line"><a name="l00600"></a><span class="lineno"> 600</span> </div>
|
|
<div class="line"><a name="l00601"></a><span class="lineno"> 601</span>  <span class="keyword">auto</span> query = queries[queryId].data();</div>
|
|
<div class="line"><a name="l00602"></a><span class="lineno"> 602</span>  <span class="keyword">auto</span> vecs = allListData[listId];</div>
|
|
<div class="line"><a name="l00603"></a><span class="lineno"> 603</span>  <span class="keyword">auto</span> numVecs = listLengths[listId];</div>
|
|
<div class="line"><a name="l00604"></a><span class="lineno"> 604</span>  <span class="keyword">auto</span> dim = queries.getSize(1);</div>
|
|
<div class="line"><a name="l00605"></a><span class="lineno"> 605</span>  <span class="keyword">auto</span> distanceOut = distance[outBase].data();</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>  IVFFlatScan<Dims, L2, T>::scan(query, vecs, numVecs, dim, distanceOut);</div>
|
|
<div class="line"><a name="l00608"></a><span class="lineno"> 608</span> }</div>
|
|
<div class="line"><a name="l00609"></a><span class="lineno"> 609</span> </div>
|
|
<div class="line"><a name="l00610"></a><span class="lineno"> 610</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00611"></a><span class="lineno"> 611</span> runIVFFlatScanTile(Tensor<float, 2, true>& queries,</div>
|
|
<div class="line"><a name="l00612"></a><span class="lineno"> 612</span>  Tensor<int, 2, true>& listIds,</div>
|
|
<div class="line"><a name="l00613"></a><span class="lineno"> 613</span>  thrust::device_vector<void*>& listData,</div>
|
|
<div class="line"><a name="l00614"></a><span class="lineno"> 614</span>  thrust::device_vector<void*>& listIndices,</div>
|
|
<div class="line"><a name="l00615"></a><span class="lineno"> 615</span>  IndicesOptions indicesOptions,</div>
|
|
<div class="line"><a name="l00616"></a><span class="lineno"> 616</span>  thrust::device_vector<int>& listLengths,</div>
|
|
<div class="line"><a name="l00617"></a><span class="lineno"> 617</span>  Tensor<char, 1, true>& thrustMem,</div>
|
|
<div class="line"><a name="l00618"></a><span class="lineno"> 618</span>  Tensor<int, 2, true>& prefixSumOffsets,</div>
|
|
<div class="line"><a name="l00619"></a><span class="lineno"> 619</span>  Tensor<float, 1, true>& allDistances,</div>
|
|
<div class="line"><a name="l00620"></a><span class="lineno"> 620</span>  Tensor<float, 3, true>& heapDistances,</div>
|
|
<div class="line"><a name="l00621"></a><span class="lineno"> 621</span>  Tensor<int, 3, true>& heapIndices,</div>
|
|
<div class="line"><a name="l00622"></a><span class="lineno"> 622</span>  <span class="keywordtype">int</span> k,</div>
|
|
<div class="line"><a name="l00623"></a><span class="lineno"> 623</span>  <span class="keywordtype">bool</span> l2Distance,</div>
|
|
<div class="line"><a name="l00624"></a><span class="lineno"> 624</span>  <span class="keywordtype">bool</span> useFloat16,</div>
|
|
<div class="line"><a name="l00625"></a><span class="lineno"> 625</span>  Tensor<float, 2, true>& outDistances,</div>
|
|
<div class="line"><a name="l00626"></a><span class="lineno"> 626</span>  Tensor<long, 2, true>& outIndices,</div>
|
|
<div class="line"><a name="l00627"></a><span class="lineno"> 627</span>  cudaStream_t stream) {</div>
|
|
<div class="line"><a name="l00628"></a><span class="lineno"> 628</span>  <span class="comment">// Calculate offset lengths, so we know where to write out</span></div>
|
|
<div class="line"><a name="l00629"></a><span class="lineno"> 629</span>  <span class="comment">// intermediate results</span></div>
|
|
<div class="line"><a name="l00630"></a><span class="lineno"> 630</span>  runCalcListOffsets(listIds, listLengths, prefixSumOffsets, thrustMem, 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>  <span class="comment">// Calculate distances for vectors within our chunk of lists</span></div>
|
|
<div class="line"><a name="l00633"></a><span class="lineno"> 633</span>  constexpr <span class="keywordtype">int</span> kMaxThreadsIVF = 512;</div>
|
|
<div class="line"><a name="l00634"></a><span class="lineno"> 634</span> </div>
|
|
<div class="line"><a name="l00635"></a><span class="lineno"> 635</span>  <span class="comment">// FIXME: if `half` and # dims is multiple of 2, halve the</span></div>
|
|
<div class="line"><a name="l00636"></a><span class="lineno"> 636</span>  <span class="comment">// threadblock size</span></div>
|
|
<div class="line"><a name="l00637"></a><span class="lineno"> 637</span> </div>
|
|
<div class="line"><a name="l00638"></a><span class="lineno"> 638</span>  <span class="keywordtype">int</span> dim = queries.getSize(1);</div>
|
|
<div class="line"><a name="l00639"></a><span class="lineno"> 639</span>  <span class="keywordtype">int</span> numThreads = std::min(dim, kMaxThreadsIVF);</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>  <span class="keyword">auto</span> grid = dim3(listIds.getSize(1),</div>
|
|
<div class="line"><a name="l00642"></a><span class="lineno"> 642</span>  listIds.getSize(0));</div>
|
|
<div class="line"><a name="l00643"></a><span class="lineno"> 643</span>  <span class="keyword">auto</span> block = dim3(numThreads);</div>
|
|
<div class="line"><a name="l00644"></a><span class="lineno"> 644</span>  <span class="comment">// All exact dim kernels are unrolled by 4, hence the `4`</span></div>
|
|
<div class="line"><a name="l00645"></a><span class="lineno"> 645</span>  <span class="keyword">auto</span> smem = <span class="keyword">sizeof</span>(float) * utils::divUp(numThreads, kWarpSize) * 4;</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="preprocessor">#define RUN_IVF_FLAT(DIMS, L2, T) \</span></div>
|
|
<div class="line"><a name="l00648"></a><span class="lineno"> 648</span> <span class="preprocessor"> do { \</span></div>
|
|
<div class="line"><a name="l00649"></a><span class="lineno"> 649</span> <span class="preprocessor"> ivfFlatScan<DIMS, L2, T> \</span></div>
|
|
<div class="line"><a name="l00650"></a><span class="lineno"> 650</span> <span class="preprocessor"> <<<grid, block, smem, stream>>>( \</span></div>
|
|
<div class="line"><a name="l00651"></a><span class="lineno"> 651</span> <span class="preprocessor"> queries, \</span></div>
|
|
<div class="line"><a name="l00652"></a><span class="lineno"> 652</span> <span class="preprocessor"> listIds, \</span></div>
|
|
<div class="line"><a name="l00653"></a><span class="lineno"> 653</span> <span class="preprocessor"> listData.data().get(), \</span></div>
|
|
<div class="line"><a name="l00654"></a><span class="lineno"> 654</span> <span class="preprocessor"> listLengths.data().get(), \</span></div>
|
|
<div class="line"><a name="l00655"></a><span class="lineno"> 655</span> <span class="preprocessor"> prefixSumOffsets, \</span></div>
|
|
<div class="line"><a name="l00656"></a><span class="lineno"> 656</span> <span class="preprocessor"> allDistances); \</span></div>
|
|
<div class="line"><a name="l00657"></a><span class="lineno"> 657</span> <span class="preprocessor"> } while (0)</span></div>
|
|
<div class="line"><a name="l00658"></a><span class="lineno"> 658</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00659"></a><span class="lineno"> 659</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00660"></a><span class="lineno"> 660</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00661"></a><span class="lineno"> 661</span> <span class="preprocessor">#define HANDLE_DIM_CASE(DIMS) \</span></div>
|
|
<div class="line"><a name="l00662"></a><span class="lineno"> 662</span> <span class="preprocessor"> do { \</span></div>
|
|
<div class="line"><a name="l00663"></a><span class="lineno"> 663</span> <span class="preprocessor"> if (l2Distance) { \</span></div>
|
|
<div class="line"><a name="l00664"></a><span class="lineno"> 664</span> <span class="preprocessor"> if (useFloat16) { \</span></div>
|
|
<div class="line"><a name="l00665"></a><span class="lineno"> 665</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, true, half); \</span></div>
|
|
<div class="line"><a name="l00666"></a><span class="lineno"> 666</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00667"></a><span class="lineno"> 667</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, true, float); \</span></div>
|
|
<div class="line"><a name="l00668"></a><span class="lineno"> 668</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00669"></a><span class="lineno"> 669</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00670"></a><span class="lineno"> 670</span> <span class="preprocessor"> if (useFloat16) { \</span></div>
|
|
<div class="line"><a name="l00671"></a><span class="lineno"> 671</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, false, half); \</span></div>
|
|
<div class="line"><a name="l00672"></a><span class="lineno"> 672</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00673"></a><span class="lineno"> 673</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, false, float); \</span></div>
|
|
<div class="line"><a name="l00674"></a><span class="lineno"> 674</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00675"></a><span class="lineno"> 675</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00676"></a><span class="lineno"> 676</span> <span class="preprocessor"> } while (0)</span></div>
|
|
<div class="line"><a name="l00677"></a><span class="lineno"> 677</span> <span class="preprocessor"></span><span class="preprocessor">#else</span></div>
|
|
<div class="line"><a name="l00678"></a><span class="lineno"> 678</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00679"></a><span class="lineno"> 679</span> <span class="preprocessor">#define HANDLE_DIM_CASE(DIMS) \</span></div>
|
|
<div class="line"><a name="l00680"></a><span class="lineno"> 680</span> <span class="preprocessor"> do { \</span></div>
|
|
<div class="line"><a name="l00681"></a><span class="lineno"> 681</span> <span class="preprocessor"> if (l2Distance) { \</span></div>
|
|
<div class="line"><a name="l00682"></a><span class="lineno"> 682</span> <span class="preprocessor"> if (useFloat16) { \</span></div>
|
|
<div class="line"><a name="l00683"></a><span class="lineno"> 683</span> <span class="preprocessor"> FAISS_ASSERT(false); \</span></div>
|
|
<div class="line"><a name="l00684"></a><span class="lineno"> 684</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00685"></a><span class="lineno"> 685</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, true, float); \</span></div>
|
|
<div class="line"><a name="l00686"></a><span class="lineno"> 686</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00687"></a><span class="lineno"> 687</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00688"></a><span class="lineno"> 688</span> <span class="preprocessor"> if (useFloat16) { \</span></div>
|
|
<div class="line"><a name="l00689"></a><span class="lineno"> 689</span> <span class="preprocessor"> FAISS_ASSERT(false); \</span></div>
|
|
<div class="line"><a name="l00690"></a><span class="lineno"> 690</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00691"></a><span class="lineno"> 691</span> <span class="preprocessor"> RUN_IVF_FLAT(DIMS, false, float); \</span></div>
|
|
<div class="line"><a name="l00692"></a><span class="lineno"> 692</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00693"></a><span class="lineno"> 693</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00694"></a><span class="lineno"> 694</span> <span class="preprocessor"> } while (0)</span></div>
|
|
<div class="line"><a name="l00695"></a><span class="lineno"> 695</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00696"></a><span class="lineno"> 696</span> <span class="preprocessor">#endif // FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00697"></a><span class="lineno"> 697</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00698"></a><span class="lineno"> 698</span>  <span class="keywordflow">if</span> (dim == 64) {</div>
|
|
<div class="line"><a name="l00699"></a><span class="lineno"> 699</span>  HANDLE_DIM_CASE(64);</div>
|
|
<div class="line"><a name="l00700"></a><span class="lineno"> 700</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (dim == 128) {</div>
|
|
<div class="line"><a name="l00701"></a><span class="lineno"> 701</span>  HANDLE_DIM_CASE(128);</div>
|
|
<div class="line"><a name="l00702"></a><span class="lineno"> 702</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (dim == 256) {</div>
|
|
<div class="line"><a name="l00703"></a><span class="lineno"> 703</span>  HANDLE_DIM_CASE(256);</div>
|
|
<div class="line"><a name="l00704"></a><span class="lineno"> 704</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (dim <= kMaxThreadsIVF) {</div>
|
|
<div class="line"><a name="l00705"></a><span class="lineno"> 705</span>  HANDLE_DIM_CASE(0);</div>
|
|
<div class="line"><a name="l00706"></a><span class="lineno"> 706</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00707"></a><span class="lineno"> 707</span>  HANDLE_DIM_CASE(-1);</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> </div>
|
|
<div class="line"><a name="l00710"></a><span class="lineno"> 710</span> <span class="preprocessor">#undef HANDLE_DIM_CASE</span></div>
|
|
<div class="line"><a name="l00711"></a><span class="lineno"> 711</span> <span class="preprocessor"></span><span class="preprocessor">#undef RUN_IVF_FLAT</span></div>
|
|
<div class="line"><a name="l00712"></a><span class="lineno"> 712</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00713"></a><span class="lineno"> 713</span>  <span class="comment">// k-select the output in chunks, to increase parallelism</span></div>
|
|
<div class="line"><a name="l00714"></a><span class="lineno"> 714</span>  runPass1SelectLists(prefixSumOffsets,</div>
|
|
<div class="line"><a name="l00715"></a><span class="lineno"> 715</span>  allDistances,</div>
|
|
<div class="line"><a name="l00716"></a><span class="lineno"> 716</span>  listIds.getSize(1),</div>
|
|
<div class="line"><a name="l00717"></a><span class="lineno"> 717</span>  k,</div>
|
|
<div class="line"><a name="l00718"></a><span class="lineno"> 718</span>  !l2Distance, <span class="comment">// L2 distance chooses smallest</span></div>
|
|
<div class="line"><a name="l00719"></a><span class="lineno"> 719</span>  heapDistances,</div>
|
|
<div class="line"><a name="l00720"></a><span class="lineno"> 720</span>  heapIndices,</div>
|
|
<div class="line"><a name="l00721"></a><span class="lineno"> 721</span>  stream);</div>
|
|
<div class="line"><a name="l00722"></a><span class="lineno"> 722</span> </div>
|
|
<div class="line"><a name="l00723"></a><span class="lineno"> 723</span>  <span class="comment">// k-select final output</span></div>
|
|
<div class="line"><a name="l00724"></a><span class="lineno"> 724</span>  <span class="keyword">auto</span> flatHeapDistances = heapDistances.downcastInner<2>();</div>
|
|
<div class="line"><a name="l00725"></a><span class="lineno"> 725</span>  <span class="keyword">auto</span> flatHeapIndices = heapIndices.downcastInner<2>();</div>
|
|
<div class="line"><a name="l00726"></a><span class="lineno"> 726</span> </div>
|
|
<div class="line"><a name="l00727"></a><span class="lineno"> 727</span>  runPass2SelectLists(flatHeapDistances,</div>
|
|
<div class="line"><a name="l00728"></a><span class="lineno"> 728</span>  flatHeapIndices,</div>
|
|
<div class="line"><a name="l00729"></a><span class="lineno"> 729</span>  listIndices,</div>
|
|
<div class="line"><a name="l00730"></a><span class="lineno"> 730</span>  indicesOptions,</div>
|
|
<div class="line"><a name="l00731"></a><span class="lineno"> 731</span>  prefixSumOffsets,</div>
|
|
<div class="line"><a name="l00732"></a><span class="lineno"> 732</span>  listIds,</div>
|
|
<div class="line"><a name="l00733"></a><span class="lineno"> 733</span>  k,</div>
|
|
<div class="line"><a name="l00734"></a><span class="lineno"> 734</span>  !l2Distance, <span class="comment">// L2 distance chooses smallest</span></div>
|
|
<div class="line"><a name="l00735"></a><span class="lineno"> 735</span>  outDistances,</div>
|
|
<div class="line"><a name="l00736"></a><span class="lineno"> 736</span>  outIndices,</div>
|
|
<div class="line"><a name="l00737"></a><span class="lineno"> 737</span>  stream);</div>
|
|
<div class="line"><a name="l00738"></a><span class="lineno"> 738</span> </div>
|
|
<div class="line"><a name="l00739"></a><span class="lineno"> 739</span>  CUDA_VERIFY(cudaGetLastError());</div>
|
|
<div class="line"><a name="l00740"></a><span class="lineno"> 740</span> }</div>
|
|
<div class="line"><a name="l00741"></a><span class="lineno"> 741</span> </div>
|
|
<div class="line"><a name="l00742"></a><span class="lineno"> 742</span> <span class="keywordtype">void</span></div>
|
|
<div class="line"><a name="l00743"></a><span class="lineno"> 743</span> runIVFFlatScan(Tensor<float, 2, true>& queries,</div>
|
|
<div class="line"><a name="l00744"></a><span class="lineno"> 744</span>  Tensor<int, 2, true>& listIds,</div>
|
|
<div class="line"><a name="l00745"></a><span class="lineno"> 745</span>  thrust::device_vector<void*>& listData,</div>
|
|
<div class="line"><a name="l00746"></a><span class="lineno"> 746</span>  thrust::device_vector<void*>& listIndices,</div>
|
|
<div class="line"><a name="l00747"></a><span class="lineno"> 747</span>  IndicesOptions indicesOptions,</div>
|
|
<div class="line"><a name="l00748"></a><span class="lineno"> 748</span>  thrust::device_vector<int>& listLengths,</div>
|
|
<div class="line"><a name="l00749"></a><span class="lineno"> 749</span>  <span class="keywordtype">int</span> maxListLength,</div>
|
|
<div class="line"><a name="l00750"></a><span class="lineno"> 750</span>  <span class="keywordtype">int</span> k,</div>
|
|
<div class="line"><a name="l00751"></a><span class="lineno"> 751</span>  <span class="keywordtype">bool</span> l2Distance,</div>
|
|
<div class="line"><a name="l00752"></a><span class="lineno"> 752</span>  <span class="keywordtype">bool</span> useFloat16,</div>
|
|
<div class="line"><a name="l00753"></a><span class="lineno"> 753</span>  <span class="comment">// output</span></div>
|
|
<div class="line"><a name="l00754"></a><span class="lineno"> 754</span>  Tensor<float, 2, true>& outDistances,</div>
|
|
<div class="line"><a name="l00755"></a><span class="lineno"> 755</span>  <span class="comment">// output</span></div>
|
|
<div class="line"><a name="l00756"></a><span class="lineno"> 756</span>  Tensor<long, 2, true>& outIndices,</div>
|
|
<div class="line"><a name="l00757"></a><span class="lineno"> 757</span>  GpuResources* res) {</div>
|
|
<div class="line"><a name="l00758"></a><span class="lineno"> 758</span>  constexpr <span class="keywordtype">int</span> kMinQueryTileSize = 8;</div>
|
|
<div class="line"><a name="l00759"></a><span class="lineno"> 759</span>  constexpr <span class="keywordtype">int</span> kMaxQueryTileSize = 128;</div>
|
|
<div class="line"><a name="l00760"></a><span class="lineno"> 760</span>  constexpr <span class="keywordtype">int</span> kThrustMemSize = 16384;</div>
|
|
<div class="line"><a name="l00761"></a><span class="lineno"> 761</span> </div>
|
|
<div class="line"><a name="l00762"></a><span class="lineno"> 762</span>  <span class="keywordtype">int</span> nprobe = listIds.getSize(1);</div>
|
|
<div class="line"><a name="l00763"></a><span class="lineno"> 763</span> </div>
|
|
<div class="line"><a name="l00764"></a><span class="lineno"> 764</span>  <span class="keyword">auto</span>& mem = res->getMemoryManagerCurrentDevice();</div>
|
|
<div class="line"><a name="l00765"></a><span class="lineno"> 765</span>  <span class="keyword">auto</span> stream = res->getDefaultStreamCurrentDevice();</div>
|
|
<div class="line"><a name="l00766"></a><span class="lineno"> 766</span> </div>
|
|
<div class="line"><a name="l00767"></a><span class="lineno"> 767</span>  <span class="comment">// Make a reservation for Thrust to do its dirty work (global memory</span></div>
|
|
<div class="line"><a name="l00768"></a><span class="lineno"> 768</span>  <span class="comment">// cross-block reduction space); hopefully this is large enough.</span></div>
|
|
<div class="line"><a name="l00769"></a><span class="lineno"> 769</span>  DeviceTensor<char, 1, true> thrustMem1(</div>
|
|
<div class="line"><a name="l00770"></a><span class="lineno"> 770</span>  mem, {kThrustMemSize}, stream);</div>
|
|
<div class="line"><a name="l00771"></a><span class="lineno"> 771</span>  DeviceTensor<char, 1, true> thrustMem2(</div>
|
|
<div class="line"><a name="l00772"></a><span class="lineno"> 772</span>  mem, {kThrustMemSize}, stream);</div>
|
|
<div class="line"><a name="l00773"></a><span class="lineno"> 773</span>  DeviceTensor<char, 1, true>* thrustMem[2] =</div>
|
|
<div class="line"><a name="l00774"></a><span class="lineno"> 774</span>  {&thrustMem1, &thrustMem2};</div>
|
|
<div class="line"><a name="l00775"></a><span class="lineno"> 775</span> </div>
|
|
<div class="line"><a name="l00776"></a><span class="lineno"> 776</span>  <span class="comment">// How much temporary storage is available?</span></div>
|
|
<div class="line"><a name="l00777"></a><span class="lineno"> 777</span>  <span class="comment">// If possible, we'd like to fit within the space available.</span></div>
|
|
<div class="line"><a name="l00778"></a><span class="lineno"> 778</span>  <span class="keywordtype">size_t</span> sizeAvailable = mem.getSizeAvailable();</div>
|
|
<div class="line"><a name="l00779"></a><span class="lineno"> 779</span> </div>
|
|
<div class="line"><a name="l00780"></a><span class="lineno"> 780</span>  <span class="comment">// We run two passes of heap selection</span></div>
|
|
<div class="line"><a name="l00781"></a><span class="lineno"> 781</span>  <span class="comment">// This is the size of the first-level heap passes</span></div>
|
|
<div class="line"><a name="l00782"></a><span class="lineno"> 782</span>  constexpr <span class="keywordtype">int</span> kNProbeSplit = 8;</div>
|
|
<div class="line"><a name="l00783"></a><span class="lineno"> 783</span>  <span class="keywordtype">int</span> pass2Chunks = std::min(nprobe, kNProbeSplit);</div>
|
|
<div class="line"><a name="l00784"></a><span class="lineno"> 784</span> </div>
|
|
<div class="line"><a name="l00785"></a><span class="lineno"> 785</span>  <span class="keywordtype">size_t</span> sizeForFirstSelectPass =</div>
|
|
<div class="line"><a name="l00786"></a><span class="lineno"> 786</span>  pass2Chunks * k * (<span class="keyword">sizeof</span>(float) + <span class="keyword">sizeof</span>(<span class="keywordtype">int</span>));</div>
|
|
<div class="line"><a name="l00787"></a><span class="lineno"> 787</span> </div>
|
|
<div class="line"><a name="l00788"></a><span class="lineno"> 788</span>  <span class="comment">// How much temporary storage we need per each query</span></div>
|
|
<div class="line"><a name="l00789"></a><span class="lineno"> 789</span>  <span class="keywordtype">size_t</span> sizePerQuery =</div>
|
|
<div class="line"><a name="l00790"></a><span class="lineno"> 790</span>  2 * <span class="comment">// # streams</span></div>
|
|
<div class="line"><a name="l00791"></a><span class="lineno"> 791</span>  ((nprobe * <span class="keyword">sizeof</span>(int) + <span class="keyword">sizeof</span>(<span class="keywordtype">int</span>)) + <span class="comment">// prefixSumOffsets</span></div>
|
|
<div class="line"><a name="l00792"></a><span class="lineno"> 792</span>  nprobe * maxListLength * <span class="keyword">sizeof</span>(<span class="keywordtype">float</span>) + <span class="comment">// allDistances</span></div>
|
|
<div class="line"><a name="l00793"></a><span class="lineno"> 793</span>  sizeForFirstSelectPass);</div>
|
|
<div class="line"><a name="l00794"></a><span class="lineno"> 794</span> </div>
|
|
<div class="line"><a name="l00795"></a><span class="lineno"> 795</span>  <span class="keywordtype">int</span> queryTileSize = (int) (sizeAvailable / sizePerQuery);</div>
|
|
<div class="line"><a name="l00796"></a><span class="lineno"> 796</span> </div>
|
|
<div class="line"><a name="l00797"></a><span class="lineno"> 797</span>  <span class="keywordflow">if</span> (queryTileSize < kMinQueryTileSize) {</div>
|
|
<div class="line"><a name="l00798"></a><span class="lineno"> 798</span>  queryTileSize = kMinQueryTileSize;</div>
|
|
<div class="line"><a name="l00799"></a><span class="lineno"> 799</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (queryTileSize > kMaxQueryTileSize) {</div>
|
|
<div class="line"><a name="l00800"></a><span class="lineno"> 800</span>  queryTileSize = kMaxQueryTileSize;</div>
|
|
<div class="line"><a name="l00801"></a><span class="lineno"> 801</span>  }</div>
|
|
<div class="line"><a name="l00802"></a><span class="lineno"> 802</span> </div>
|
|
<div class="line"><a name="l00803"></a><span class="lineno"> 803</span>  <span class="comment">// FIXME: we should adjust queryTileSize to deal with this, since</span></div>
|
|
<div class="line"><a name="l00804"></a><span class="lineno"> 804</span>  <span class="comment">// indexing is in int32</span></div>
|
|
<div class="line"><a name="l00805"></a><span class="lineno"> 805</span>  FAISS_ASSERT(queryTileSize * nprobe * maxListLength <</div>
|
|
<div class="line"><a name="l00806"></a><span class="lineno"> 806</span>  std::numeric_limits<int>::max());</div>
|
|
<div class="line"><a name="l00807"></a><span class="lineno"> 807</span> </div>
|
|
<div class="line"><a name="l00808"></a><span class="lineno"> 808</span>  <span class="comment">// Temporary memory buffers</span></div>
|
|
<div class="line"><a name="l00809"></a><span class="lineno"> 809</span>  <span class="comment">// Make sure there is space prior to the start which will be 0, and</span></div>
|
|
<div class="line"><a name="l00810"></a><span class="lineno"> 810</span>  <span class="comment">// will handle the boundary condition without branches</span></div>
|
|
<div class="line"><a name="l00811"></a><span class="lineno"> 811</span>  DeviceTensor<int, 1, true> prefixSumOffsetSpace1(</div>
|
|
<div class="line"><a name="l00812"></a><span class="lineno"> 812</span>  mem, {queryTileSize * nprobe + 1}, stream);</div>
|
|
<div class="line"><a name="l00813"></a><span class="lineno"> 813</span>  DeviceTensor<int, 1, true> prefixSumOffsetSpace2(</div>
|
|
<div class="line"><a name="l00814"></a><span class="lineno"> 814</span>  mem, {queryTileSize * nprobe + 1}, stream);</div>
|
|
<div class="line"><a name="l00815"></a><span class="lineno"> 815</span> </div>
|
|
<div class="line"><a name="l00816"></a><span class="lineno"> 816</span>  DeviceTensor<int, 2, true> prefixSumOffsets1(</div>
|
|
<div class="line"><a name="l00817"></a><span class="lineno"> 817</span>  prefixSumOffsetSpace1[1].data(),</div>
|
|
<div class="line"><a name="l00818"></a><span class="lineno"> 818</span>  {queryTileSize, nprobe});</div>
|
|
<div class="line"><a name="l00819"></a><span class="lineno"> 819</span>  DeviceTensor<int, 2, true> prefixSumOffsets2(</div>
|
|
<div class="line"><a name="l00820"></a><span class="lineno"> 820</span>  prefixSumOffsetSpace2[1].data(),</div>
|
|
<div class="line"><a name="l00821"></a><span class="lineno"> 821</span>  {queryTileSize, nprobe});</div>
|
|
<div class="line"><a name="l00822"></a><span class="lineno"> 822</span>  DeviceTensor<int, 2, true>* prefixSumOffsets[2] =</div>
|
|
<div class="line"><a name="l00823"></a><span class="lineno"> 823</span>  {&prefixSumOffsets1, &prefixSumOffsets2};</div>
|
|
<div class="line"><a name="l00824"></a><span class="lineno"> 824</span> </div>
|
|
<div class="line"><a name="l00825"></a><span class="lineno"> 825</span>  <span class="comment">// Make sure the element before prefixSumOffsets is 0, since we</span></div>
|
|
<div class="line"><a name="l00826"></a><span class="lineno"> 826</span>  <span class="comment">// depend upon simple, boundary-less indexing to get proper results</span></div>
|
|
<div class="line"><a name="l00827"></a><span class="lineno"> 827</span>  CUDA_VERIFY(cudaMemsetAsync(prefixSumOffsetSpace1.data(),</div>
|
|
<div class="line"><a name="l00828"></a><span class="lineno"> 828</span>  0,</div>
|
|
<div class="line"><a name="l00829"></a><span class="lineno"> 829</span>  <span class="keyword">sizeof</span>(int),</div>
|
|
<div class="line"><a name="l00830"></a><span class="lineno"> 830</span>  stream));</div>
|
|
<div class="line"><a name="l00831"></a><span class="lineno"> 831</span>  CUDA_VERIFY(cudaMemsetAsync(prefixSumOffsetSpace2.data(),</div>
|
|
<div class="line"><a name="l00832"></a><span class="lineno"> 832</span>  0,</div>
|
|
<div class="line"><a name="l00833"></a><span class="lineno"> 833</span>  <span class="keyword">sizeof</span>(int),</div>
|
|
<div class="line"><a name="l00834"></a><span class="lineno"> 834</span>  stream));</div>
|
|
<div class="line"><a name="l00835"></a><span class="lineno"> 835</span> </div>
|
|
<div class="line"><a name="l00836"></a><span class="lineno"> 836</span>  DeviceTensor<float, 1, true> allDistances1(</div>
|
|
<div class="line"><a name="l00837"></a><span class="lineno"> 837</span>  mem, {queryTileSize * nprobe * maxListLength}, stream);</div>
|
|
<div class="line"><a name="l00838"></a><span class="lineno"> 838</span>  DeviceTensor<float, 1, true> allDistances2(</div>
|
|
<div class="line"><a name="l00839"></a><span class="lineno"> 839</span>  mem, {queryTileSize * nprobe * maxListLength}, stream);</div>
|
|
<div class="line"><a name="l00840"></a><span class="lineno"> 840</span>  DeviceTensor<float, 1, true>* allDistances[2] =</div>
|
|
<div class="line"><a name="l00841"></a><span class="lineno"> 841</span>  {&allDistances1, &allDistances2};</div>
|
|
<div class="line"><a name="l00842"></a><span class="lineno"> 842</span> </div>
|
|
<div class="line"><a name="l00843"></a><span class="lineno"> 843</span>  DeviceTensor<float, 3, true> heapDistances1(</div>
|
|
<div class="line"><a name="l00844"></a><span class="lineno"> 844</span>  mem, {queryTileSize, pass2Chunks, k}, stream);</div>
|
|
<div class="line"><a name="l00845"></a><span class="lineno"> 845</span>  DeviceTensor<float, 3, true> heapDistances2(</div>
|
|
<div class="line"><a name="l00846"></a><span class="lineno"> 846</span>  mem, {queryTileSize, pass2Chunks, k}, stream);</div>
|
|
<div class="line"><a name="l00847"></a><span class="lineno"> 847</span>  DeviceTensor<float, 3, true>* heapDistances[2] =</div>
|
|
<div class="line"><a name="l00848"></a><span class="lineno"> 848</span>  {&heapDistances1, &heapDistances2};</div>
|
|
<div class="line"><a name="l00849"></a><span class="lineno"> 849</span> </div>
|
|
<div class="line"><a name="l00850"></a><span class="lineno"> 850</span>  DeviceTensor<int, 3, true> heapIndices1(</div>
|
|
<div class="line"><a name="l00851"></a><span class="lineno"> 851</span>  mem, {queryTileSize, pass2Chunks, k}, stream);</div>
|
|
<div class="line"><a name="l00852"></a><span class="lineno"> 852</span>  DeviceTensor<int, 3, true> heapIndices2(</div>
|
|
<div class="line"><a name="l00853"></a><span class="lineno"> 853</span>  mem, {queryTileSize, pass2Chunks, k}, stream);</div>
|
|
<div class="line"><a name="l00854"></a><span class="lineno"> 854</span>  DeviceTensor<int, 3, true>* heapIndices[2] =</div>
|
|
<div class="line"><a name="l00855"></a><span class="lineno"> 855</span>  {&heapIndices1, &heapIndices2};</div>
|
|
<div class="line"><a name="l00856"></a><span class="lineno"> 856</span> </div>
|
|
<div class="line"><a name="l00857"></a><span class="lineno"> 857</span>  <span class="keyword">auto</span> streams = res->getAlternateStreamsCurrentDevice();</div>
|
|
<div class="line"><a name="l00858"></a><span class="lineno"> 858</span>  streamWait(streams, {stream});</div>
|
|
<div class="line"><a name="l00859"></a><span class="lineno"> 859</span> </div>
|
|
<div class="line"><a name="l00860"></a><span class="lineno"> 860</span>  <span class="keywordtype">int</span> curStream = 0;</div>
|
|
<div class="line"><a name="l00861"></a><span class="lineno"> 861</span> </div>
|
|
<div class="line"><a name="l00862"></a><span class="lineno"> 862</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> query = 0; query < queries.getSize(0); query += queryTileSize) {</div>
|
|
<div class="line"><a name="l00863"></a><span class="lineno"> 863</span>  <span class="keywordtype">int</span> numQueriesInTile =</div>
|
|
<div class="line"><a name="l00864"></a><span class="lineno"> 864</span>  std::min(queryTileSize, queries.getSize(0) - query);</div>
|
|
<div class="line"><a name="l00865"></a><span class="lineno"> 865</span> </div>
|
|
<div class="line"><a name="l00866"></a><span class="lineno"> 866</span>  <span class="keyword">auto</span> prefixSumOffsetsView =</div>
|
|
<div class="line"><a name="l00867"></a><span class="lineno"> 867</span>  prefixSumOffsets[curStream]->narrowOutermost(0, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00868"></a><span class="lineno"> 868</span> </div>
|
|
<div class="line"><a name="l00869"></a><span class="lineno"> 869</span>  <span class="keyword">auto</span> listIdsView =</div>
|
|
<div class="line"><a name="l00870"></a><span class="lineno"> 870</span>  listIds.narrowOutermost(query, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00871"></a><span class="lineno"> 871</span>  <span class="keyword">auto</span> queryView =</div>
|
|
<div class="line"><a name="l00872"></a><span class="lineno"> 872</span>  queries.narrowOutermost(query, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00873"></a><span class="lineno"> 873</span> </div>
|
|
<div class="line"><a name="l00874"></a><span class="lineno"> 874</span>  <span class="keyword">auto</span> heapDistancesView =</div>
|
|
<div class="line"><a name="l00875"></a><span class="lineno"> 875</span>  heapDistances[curStream]->narrowOutermost(0, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00876"></a><span class="lineno"> 876</span>  <span class="keyword">auto</span> heapIndicesView =</div>
|
|
<div class="line"><a name="l00877"></a><span class="lineno"> 877</span>  heapIndices[curStream]->narrowOutermost(0, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00878"></a><span class="lineno"> 878</span> </div>
|
|
<div class="line"><a name="l00879"></a><span class="lineno"> 879</span>  <span class="keyword">auto</span> outDistanceView =</div>
|
|
<div class="line"><a name="l00880"></a><span class="lineno"> 880</span>  outDistances.narrowOutermost(query, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00881"></a><span class="lineno"> 881</span>  <span class="keyword">auto</span> outIndicesView =</div>
|
|
<div class="line"><a name="l00882"></a><span class="lineno"> 882</span>  outIndices.narrowOutermost(query, numQueriesInTile);</div>
|
|
<div class="line"><a name="l00883"></a><span class="lineno"> 883</span> </div>
|
|
<div class="line"><a name="l00884"></a><span class="lineno"> 884</span>  runIVFFlatScanTile(queryView,</div>
|
|
<div class="line"><a name="l00885"></a><span class="lineno"> 885</span>  listIdsView,</div>
|
|
<div class="line"><a name="l00886"></a><span class="lineno"> 886</span>  listData,</div>
|
|
<div class="line"><a name="l00887"></a><span class="lineno"> 887</span>  listIndices,</div>
|
|
<div class="line"><a name="l00888"></a><span class="lineno"> 888</span>  indicesOptions,</div>
|
|
<div class="line"><a name="l00889"></a><span class="lineno"> 889</span>  listLengths,</div>
|
|
<div class="line"><a name="l00890"></a><span class="lineno"> 890</span>  *thrustMem[curStream],</div>
|
|
<div class="line"><a name="l00891"></a><span class="lineno"> 891</span>  prefixSumOffsetsView,</div>
|
|
<div class="line"><a name="l00892"></a><span class="lineno"> 892</span>  *allDistances[curStream],</div>
|
|
<div class="line"><a name="l00893"></a><span class="lineno"> 893</span>  heapDistancesView,</div>
|
|
<div class="line"><a name="l00894"></a><span class="lineno"> 894</span>  heapIndicesView,</div>
|
|
<div class="line"><a name="l00895"></a><span class="lineno"> 895</span>  k,</div>
|
|
<div class="line"><a name="l00896"></a><span class="lineno"> 896</span>  l2Distance,</div>
|
|
<div class="line"><a name="l00897"></a><span class="lineno"> 897</span>  useFloat16,</div>
|
|
<div class="line"><a name="l00898"></a><span class="lineno"> 898</span>  outDistanceView,</div>
|
|
<div class="line"><a name="l00899"></a><span class="lineno"> 899</span>  outIndicesView,</div>
|
|
<div class="line"><a name="l00900"></a><span class="lineno"> 900</span>  streams[curStream]);</div>
|
|
<div class="line"><a name="l00901"></a><span class="lineno"> 901</span> </div>
|
|
<div class="line"><a name="l00902"></a><span class="lineno"> 902</span>  curStream = (curStream + 1) % 2;</div>
|
|
<div class="line"><a name="l00903"></a><span class="lineno"> 903</span>  }</div>
|
|
<div class="line"><a name="l00904"></a><span class="lineno"> 904</span> </div>
|
|
<div class="line"><a name="l00905"></a><span class="lineno"> 905</span>  streamWait({stream}, streams);</div>
|
|
<div class="line"><a name="l00906"></a><span class="lineno"> 906</span> }</div>
|
|
<div class="line"><a name="l00907"></a><span class="lineno"> 907</span> </div>
|
|
<div class="line"><a name="l00908"></a><span class="lineno"> 908</span> } } <span class="comment">// namespace</span></div>
|
|
<div class="ttc" id="structfaiss_1_1gpu_1_1LoadStore_html"><div class="ttname"><a href="structfaiss_1_1gpu_1_1LoadStore.html">faiss::gpu::LoadStore</a></div><div class="ttdef"><b>Definition:</b> <a href="LoadStoreOperators_8cuh_source.html#l00025">LoadStoreOperators.cuh:25</a></div></div>
|
|
<div class="ttc" id="structfaiss_1_1gpu_1_1Math_html_a4b17f0b5d014f300e76dde5b24af8014"><div class="ttname"><a href="structfaiss_1_1gpu_1_1Math.html#a4b17f0b5d014f300e76dde5b24af8014">faiss::gpu::Math::reduceAdd</a></div><div class="ttdeci">static __device__ T reduceAdd(T v)</div><div class="ttdoc">For a vector type, this is a horizontal add, returning sum(v_i) </div><div class="ttdef"><b>Definition:</b> <a href="MathOperators_8cuh_source.html#l00045">MathOperators.cuh:45</a></div></div>
|
|
<div class="ttc" id="structfaiss_1_1gpu_1_1ConvertTo_html"><div class="ttname"><a href="structfaiss_1_1gpu_1_1ConvertTo.html">faiss::gpu::ConvertTo</a></div><div class="ttdef"><b>Definition:</b> <a href="ConversionOperators_8cuh_source.html#l00024">ConversionOperators.cuh:24</a></div></div>
|
|
<div class="ttc" id="structfaiss_1_1gpu_1_1IVFFlatScan_html"><div class="ttname"><a href="structfaiss_1_1gpu_1_1IVFFlatScan.html">faiss::gpu::IVFFlatScan</a></div><div class="ttdoc">The class that we use to provide scan specializations. </div><div class="ttdef"><b>Definition:</b> <a href="IVFFlatScan_8cu_source.html#l00047">IVFFlatScan.cu:47</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>
|