<!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: /data/users/hoss/faiss/gpu/impl/BinaryDistance.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_5956a3e80a20e8e03eb577bedb92689f.html">gpu</a></li><li class="navelem"><a class="el" href="dir_2be73404b46ec2282840cd36fdb9a907.html">impl</a></li> </ul>
</div>
</div><!-- top -->
<div class="header">
<div class="headertitle">
<div class="title">BinaryDistance.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"> * Copyright (c) Facebook, Inc. and its affiliates.</span></div>
<div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> *</span></div>
<div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> * This source code is licensed under the MIT license found in the</span></div>
<div class="line"><a name="l00005"></a><span class="lineno"> 5</span> <span class="comment"> * LICENSE file in the root directory of this source tree.</span></div>
<div class="line"><a name="l00006"></a><span class="lineno"> 6</span> <span class="comment"> */</span></div>
<div class="line"><a name="l00007"></a><span class="lineno"> 7</span> </div>
<div class="line"><a name="l00008"></a><span class="lineno"> 8</span> <span class="preprocessor">#include "../utils/DeviceTensor.cuh"</span></div>
<div class="line"><a name="l00009"></a><span class="lineno"> 9</span> <span class="preprocessor">#include "../utils/DeviceDefs.cuh"</span></div>
<div class="line"><a name="l00010"></a><span class="lineno"> 10</span> <span class="preprocessor">#include "../utils/DeviceUtils.h"</span></div>
<div class="line"><a name="l00011"></a><span class="lineno"> 11</span> <span class="preprocessor">#include "../utils/Select.cuh"</span></div>
<div class="line"><a name="l00012"></a><span class="lineno"> 12</span> </div>
<div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="keyword">namespace </span>faiss { <span class="keyword">namespace </span>gpu {</div>
<div class="line"><a name="l00014"></a><span class="lineno"> 14</span> </div>
<div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="comment">// Number of warps that the kernel is instantiated with</span></div>
<div class="line"><a name="l00016"></a><span class="lineno"> 16</span> constexpr <span class="keywordtype">int</span> kWarps = 8;</div>
<div class="line"><a name="l00017"></a><span class="lineno"> 17</span> constexpr <span class="keywordtype">int</span> kLanes = kWarpSize;</div>
<div class="line"><a name="l00018"></a><span class="lineno"> 18</span> </div>
<div class="line"><a name="l00019"></a><span class="lineno"> 19</span> constexpr <span class="keywordtype">int</span> kMaxDistance = std::numeric_limits<int>::max();</div>
<div class="line"><a name="l00020"></a><span class="lineno"> 20</span> </div>
<div class="line"><a name="l00021"></a><span class="lineno"> 21</span> <span class="comment">// Performs a binary matrix multiplication, returning the lowest k results in</span></div>
<div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="comment">// `vecs` for each `query` in terms of Hamming distance (a fused kernel)</span></div>
<div class="line"><a name="l00023"></a><span class="lineno"> 23</span> <span class="comment">// Each warp calculates distance for a single query</span></div>
<div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="keyword">template</span> <<span class="keywordtype">int</span> NumWarpQ,</div>
<div class="line"><a name="l00025"></a><span class="lineno"> 25</span>  <span class="keywordtype">int</span> NumThreadQ,</div>
<div class="line"><a name="l00026"></a><span class="lineno"> 26</span>  <span class="keyword">typename</span> BinaryType></div>
<div class="line"><a name="l00027"></a><span class="lineno"> 27</span> __launch_bounds__(kWarps * kLanes)</div>
<div class="line"><a name="l00028"></a><span class="lineno"> 28</span> __global__ <span class="keywordtype">void</span> binaryDistanceAnySize(<span class="keyword">const</span> Tensor<BinaryType, 2, true> vecs,</div>
<div class="line"><a name="l00029"></a><span class="lineno"> 29</span>  <span class="keyword">const</span> Tensor<BinaryType, 2, true> query,</div>
<div class="line"><a name="l00030"></a><span class="lineno"> 30</span>  Tensor<int, 2, true> outK,</div>
<div class="line"><a name="l00031"></a><span class="lineno"> 31</span>  Tensor<int, 2, true> outV,</div>
<div class="line"><a name="l00032"></a><span class="lineno"> 32</span>  <span class="keywordtype">int</span> k) {</div>
<div class="line"><a name="l00033"></a><span class="lineno"> 33</span>  <span class="comment">// A matrix tile (query, k)</span></div>
<div class="line"><a name="l00034"></a><span class="lineno"> 34</span>  __shared__ BinaryType queryTile[kWarps][kLanes + 1]; <span class="comment">// avoid bank conflict</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="comment">// B matrix tile (vec, k)</span></div>
<div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  __shared__ BinaryType vecTile[kLanes][kLanes + 1]; <span class="comment">// avoid bank conflict</span></div>
<div class="line"><a name="l00038"></a><span class="lineno"> 38</span> </div>
<div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  WarpSelect<int, int, false, Comparator<int>,</div>
<div class="line"><a name="l00040"></a><span class="lineno"> 40</span>  NumWarpQ, NumThreadQ, kWarps * kLanes></div>
<div class="line"><a name="l00041"></a><span class="lineno"> 41</span>  heap(kMaxDistance, -1, k);</div>
<div class="line"><a name="l00042"></a><span class="lineno"> 42</span> </div>
<div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  <span class="keywordtype">int</span> warpId = threadIdx.y;</div>
<div class="line"><a name="l00044"></a><span class="lineno"> 44</span>  <span class="keywordtype">int</span> laneId = threadIdx.x;</div>
<div class="line"><a name="l00045"></a><span class="lineno"> 45</span> </div>
<div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  <span class="comment">// Each warp handles a single query</span></div>
<div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  <span class="keywordtype">int</span> warpQuery = blockIdx.x * kWarps + warpId;</div>
<div class="line"><a name="l00048"></a><span class="lineno"> 48</span>  <span class="keywordtype">bool</span> queryInBounds = warpQuery < query.getSize(0);</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">// Each warp loops through the entire chunk of vectors</span></div>
<div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> blockVec = 0; blockVec < vecs.getSize(0); blockVec += kLanes) {</div>
<div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  <span class="keywordtype">int</span> threadDistance = 0;</div>
<div class="line"><a name="l00053"></a><span class="lineno"> 53</span> </div>
<div class="line"><a name="l00054"></a><span class="lineno"> 54</span>  <span class="comment">// Reduction dimension</span></div>
<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> blockK = 0; blockK < vecs.getSize(1); blockK += kLanes) {</div>
<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  <span class="keywordtype">int</span> laneK = blockK + laneId;</div>
<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  <span class="keywordtype">bool</span> kInBounds = laneK < vecs.getSize(1);</div>
<div class="line"><a name="l00058"></a><span class="lineno"> 58</span> </div>
<div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  queryTile[warpId][laneId] = queryInBounds && kInBounds ?</div>
<div class="line"><a name="l00060"></a><span class="lineno"> 60</span>  query[warpQuery][laneK] : 0;</div>
<div class="line"><a name="l00061"></a><span class="lineno"> 61</span> </div>
<div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  <span class="comment">// kWarps warps are responsible for loading 32 vecs</span></div>
<div class="line"><a name="l00063"></a><span class="lineno"> 63</span> <span class="preprocessor">#pragma unroll</span></div>
<div class="line"><a name="l00064"></a><span class="lineno"> 64</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < kLanes / kWarps; ++i) {</div>
<div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  <span class="keywordtype">int</span> warpVec = i * kWarps + warpId;</div>
<div class="line"><a name="l00066"></a><span class="lineno"> 66</span>  <span class="keywordtype">int</span> vec = blockVec + warpVec;</div>
<div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  <span class="keywordtype">bool</span> vecInBounds = vec < vecs.getSize(0);</div>
<div class="line"><a name="l00068"></a><span class="lineno"> 68</span> </div>
<div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  vecTile[warpVec][laneId] = vecInBounds && kInBounds ?</div>
<div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  vecs[vec][laneK] : 0;</div>
<div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  }</div>
<div class="line"><a name="l00072"></a><span class="lineno"> 72</span> </div>
<div class="line"><a name="l00073"></a><span class="lineno"> 73</span>  __syncthreads();</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>  <span class="comment">// Compare distances</span></div>
<div class="line"><a name="l00076"></a><span class="lineno"> 76</span> <span class="preprocessor">#pragma unroll</span></div>
<div class="line"><a name="l00077"></a><span class="lineno"> 77</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < kLanes; ++i) {</div>
<div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  threadDistance += __popc(queryTile[warpId][i] ^ vecTile[laneId][i]);</div>
<div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  }</div>
<div class="line"><a name="l00080"></a><span class="lineno"> 80</span> </div>
<div class="line"><a name="l00081"></a><span class="lineno"> 81</span>  __syncthreads();</div>
<div class="line"><a name="l00082"></a><span class="lineno"> 82</span>  }</div>
<div class="line"><a name="l00083"></a><span class="lineno"> 83</span> </div>
<div class="line"><a name="l00084"></a><span class="lineno"> 84</span>  <span class="comment">// Lanes within a warp are different vec results against the same query</span></div>
<div class="line"><a name="l00085"></a><span class="lineno"> 85</span>  <span class="comment">// Only submit distances which represent real (query, vec) pairs</span></div>
<div class="line"><a name="l00086"></a><span class="lineno"> 86</span>  <span class="keywordtype">bool</span> valInBounds = queryInBounds && (blockVec + laneId < vecs.getSize(0));</div>
<div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  threadDistance = valInBounds ? threadDistance : kMaxDistance;</div>
<div class="line"><a name="l00088"></a><span class="lineno"> 88</span>  <span class="keywordtype">int</span> <span class="keywordtype">id</span> = valInBounds ? blockVec + laneId : -1;</div>
<div class="line"><a name="l00089"></a><span class="lineno"> 89</span> </div>
<div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  heap.add(threadDistance, <span class="keywordtype">id</span>);</div>
<div class="line"><a name="l00091"></a><span class="lineno"> 91</span>  }</div>
<div class="line"><a name="l00092"></a><span class="lineno"> 92</span> </div>
<div class="line"><a name="l00093"></a><span class="lineno"> 93</span>  heap.reduce();</div>
<div class="line"><a name="l00094"></a><span class="lineno"> 94</span> </div>
<div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  <span class="keywordflow">if</span> (warpQuery < query.getSize(0)) {</div>
<div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  heap.writeOut(outK[warpQuery].data(),</div>
<div class="line"><a name="l00097"></a><span class="lineno"> 97</span>  outV[warpQuery].data(),</div>
<div class="line"><a name="l00098"></a><span class="lineno"> 98</span>  k);</div>
<div class="line"><a name="l00099"></a><span class="lineno"> 99</span>  }</div>
<div class="line"><a name="l00100"></a><span class="lineno"> 100</span> }</div>
<div class="line"><a name="l00101"></a><span class="lineno"> 101</span> </div>
<div class="line"><a name="l00102"></a><span class="lineno"> 102</span> <span class="comment">// Version of the kernel that avoids a loop over the reduction dimension, and</span></div>
<div class="line"><a name="l00103"></a><span class="lineno"> 103</span> <span class="comment">// thus avoids reloading the query vectors</span></div>
<div class="line"><a name="l00104"></a><span class="lineno"> 104</span> <span class="keyword">template</span> <<span class="keywordtype">int</span> NumWarpQ,</div>
<div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  <span class="keywordtype">int</span> NumThreadQ,</div>
<div class="line"><a name="l00106"></a><span class="lineno"> 106</span>  <span class="keyword">typename</span> BinaryType,</div>
<div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="keywordtype">int</span> ReductionLimit = kLanes></div>
<div class="line"><a name="l00108"></a><span class="lineno"> 108</span> __global__ <span class="keywordtype">void</span></div>
<div class="line"><a name="l00109"></a><span class="lineno"> 109</span> __launch_bounds__(kWarps * kLanes)</div>
<div class="line"><a name="l00110"></a><span class="lineno"> 110</span> binaryDistanceLimitSize(<span class="keyword">const</span> Tensor<BinaryType, 2, true> vecs,</div>
<div class="line"><a name="l00111"></a><span class="lineno"> 111</span>  <span class="keyword">const</span> Tensor<BinaryType, 2, true> query,</div>
<div class="line"><a name="l00112"></a><span class="lineno"> 112</span>  Tensor<int, 2, true> outK,</div>
<div class="line"><a name="l00113"></a><span class="lineno"> 113</span>  Tensor<int, 2, true> outV,</div>
<div class="line"><a name="l00114"></a><span class="lineno"> 114</span>  <span class="keywordtype">int</span> k) {</div>
<div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  <span class="comment">// A matrix tile (query, k)</span></div>
<div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  __shared__ BinaryType queryTile[kWarps][kLanes + 1]; <span class="comment">// avoid bank conflict</span></div>
<div class="line"><a name="l00117"></a><span class="lineno"> 117</span> </div>
<div class="line"><a name="l00118"></a><span class="lineno"> 118</span>  <span class="comment">// B matrix tile (vec, k)</span></div>
<div class="line"><a name="l00119"></a><span class="lineno"> 119</span>  __shared__ BinaryType vecTile[kLanes][kLanes + 1]; <span class="comment">// avoid bank conflict</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>  WarpSelect<int, int, false, Comparator<int>,</div>
<div class="line"><a name="l00122"></a><span class="lineno"> 122</span>  NumWarpQ, NumThreadQ, kWarps * kLanes></div>
<div class="line"><a name="l00123"></a><span class="lineno"> 123</span>  heap(kMaxDistance, -1, k);</div>
<div class="line"><a name="l00124"></a><span class="lineno"> 124</span> </div>
<div class="line"><a name="l00125"></a><span class="lineno"> 125</span>  <span class="keywordtype">int</span> warpId = threadIdx.y;</div>
<div class="line"><a name="l00126"></a><span class="lineno"> 126</span>  <span class="keywordtype">int</span> laneId = threadIdx.x;</div>
<div class="line"><a name="l00127"></a><span class="lineno"> 127</span> </div>
<div class="line"><a name="l00128"></a><span class="lineno"> 128</span>  <span class="comment">// Each warp handles a single query</span></div>
<div class="line"><a name="l00129"></a><span class="lineno"> 129</span>  <span class="keywordtype">int</span> laneK = laneId;</div>
<div class="line"><a name="l00130"></a><span class="lineno"> 130</span>  <span class="keywordtype">int</span> warpQuery = blockIdx.x * kWarps + warpId;</div>
<div class="line"><a name="l00131"></a><span class="lineno"> 131</span>  <span class="keywordtype">bool</span> kInBounds = laneK < vecs.getSize(1);</div>
<div class="line"><a name="l00132"></a><span class="lineno"> 132</span>  <span class="keywordtype">bool</span> queryInBounds = warpQuery < query.getSize(0);</div>
<div class="line"><a name="l00133"></a><span class="lineno"> 133</span> </div>
<div class="line"><a name="l00134"></a><span class="lineno"> 134</span> </div>
<div class="line"><a name="l00135"></a><span class="lineno"> 135</span>  queryTile[warpId][laneId] = queryInBounds && kInBounds ?</div>
<div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  query[warpQuery][laneK] : 0;</div>
<div class="line"><a name="l00137"></a><span class="lineno"> 137</span> </div>
<div class="line"><a name="l00138"></a><span class="lineno"> 138</span>  <span class="comment">// Each warp loops through the entire chunk of vectors</span></div>
<div class="line"><a name="l00139"></a><span class="lineno"> 139</span>  <span class="keywordflow">for</span> (<span class="keywordtype">int</span> blockVec = 0; blockVec < vecs.getSize(0); blockVec += kLanes) {</div>
<div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="keywordtype">int</span> threadDistance = 0;</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>  <span class="comment">// kWarps warps are responsible for loading 32 vecs</span></div>
<div class="line"><a name="l00143"></a><span class="lineno"> 143</span> <span class="preprocessor">#pragma unroll</span></div>
<div class="line"><a name="l00144"></a><span class="lineno"> 144</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < kLanes / kWarps; ++i) {</div>
<div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  <span class="keywordtype">int</span> warpVec = i * kWarps + warpId;</div>
<div class="line"><a name="l00146"></a><span class="lineno"> 146</span>  <span class="keywordtype">int</span> vec = blockVec + warpVec;</div>
<div class="line"><a name="l00147"></a><span class="lineno"> 147</span>  <span class="keywordtype">bool</span> vecInBounds = vec < vecs.getSize(0);</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>  vecTile[warpVec][laneId] = vecInBounds && kInBounds ?</div>
<div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  vecs[vec][laneK] : 0;</div>
<div class="line"><a name="l00151"></a><span class="lineno"> 151</span>  }</div>
<div class="line"><a name="l00152"></a><span class="lineno"> 152</span> </div>
<div class="line"><a name="l00153"></a><span class="lineno"> 153</span>  __syncthreads();</div>
<div class="line"><a name="l00154"></a><span class="lineno"> 154</span> </div>
<div class="line"><a name="l00155"></a><span class="lineno"> 155</span>  <span class="comment">// Compare distances</span></div>
<div class="line"><a name="l00156"></a><span class="lineno"> 156</span> <span class="preprocessor">#pragma unroll</span></div>
<div class="line"><a name="l00157"></a><span class="lineno"> 157</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < ReductionLimit; ++i) {</div>
<div class="line"><a name="l00158"></a><span class="lineno"> 158</span>  threadDistance += __popc(queryTile[warpId][i] ^ vecTile[laneId][i]);</div>
<div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  }</div>
<div class="line"><a name="l00160"></a><span class="lineno"> 160</span> </div>
<div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  __syncthreads();</div>
<div class="line"><a name="l00162"></a><span class="lineno"> 162</span> </div>
<div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  <span class="comment">// Lanes within a warp are different vec results against the same query</span></div>
<div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  <span class="comment">// Only submit distances which represent real (query, vec) pairs</span></div>
<div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  <span class="keywordtype">bool</span> valInBounds = queryInBounds && (blockVec + laneId < vecs.getSize(0));</div>
<div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  threadDistance = valInBounds ? threadDistance : kMaxDistance;</div>
<div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  <span class="keywordtype">int</span> <span class="keywordtype">id</span> = valInBounds ? blockVec + laneId : -1;</div>
<div class="line"><a name="l00168"></a><span class="lineno"> 168</span> </div>
<div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  heap.add(threadDistance, <span class="keywordtype">id</span>);</div>
<div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  }</div>
<div class="line"><a name="l00171"></a><span class="lineno"> 171</span> </div>
<div class="line"><a name="l00172"></a><span class="lineno"> 172</span>  heap.reduce();</div>
<div class="line"><a name="l00173"></a><span class="lineno"> 173</span> </div>
<div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  <span class="keywordflow">if</span> (warpQuery < query.getSize(0)) {</div>
<div class="line"><a name="l00175"></a><span class="lineno"> 175</span>  heap.writeOut(outK[warpQuery].data(),</div>
<div class="line"><a name="l00176"></a><span class="lineno"> 176</span>  outV[warpQuery].data(),</div>
<div class="line"><a name="l00177"></a><span class="lineno"> 177</span>  k);</div>
<div class="line"><a name="l00178"></a><span class="lineno"> 178</span>  }</div>
<div class="line"><a name="l00179"></a><span class="lineno"> 179</span> }</div>
<div class="line"><a name="l00180"></a><span class="lineno"> 180</span> </div>
<div class="line"><a name="l00181"></a><span class="lineno"> 181</span> <span class="keyword">template</span> <<span class="keyword">typename</span> BinaryType></div>
<div class="line"><a name="l00182"></a><span class="lineno"> 182</span> <span class="keywordtype">void</span> runBinaryDistanceAnySize(Tensor<BinaryType, 2, true>& vecs,</div>
<div class="line"><a name="l00183"></a><span class="lineno"> 183</span>  Tensor<BinaryType, 2, true>& query,</div>
<div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  Tensor<int, 2, true>& outK,</div>
<div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  Tensor<int, 2, true>& outV,</div>
<div class="line"><a name="l00186"></a><span class="lineno"> 186</span>  <span class="keywordtype">int</span> k, cudaStream_t stream) {</div>
<div class="line"><a name="l00187"></a><span class="lineno"> 187</span>  dim3 grid(utils::divUp(query.getSize(0), kWarps));</div>
<div class="line"><a name="l00188"></a><span class="lineno"> 188</span>  dim3 block(kLanes, kWarps);</div>
<div class="line"><a name="l00189"></a><span class="lineno"> 189</span> </div>
<div class="line"><a name="l00190"></a><span class="lineno"> 190</span>  <span class="keywordflow">if</span> (k == 1) {</div>
<div class="line"><a name="l00191"></a><span class="lineno"> 191</span>  binaryDistanceAnySize<1, 1, BinaryType></div>
<div class="line"><a name="l00192"></a><span class="lineno"> 192</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00193"></a><span class="lineno"> 193</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00194"></a><span class="lineno"> 194</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 32) {</div>
<div class="line"><a name="l00195"></a><span class="lineno"> 195</span>  binaryDistanceAnySize<32, 2, BinaryType></div>
<div class="line"><a name="l00196"></a><span class="lineno"> 196</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00197"></a><span class="lineno"> 197</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00198"></a><span class="lineno"> 198</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 64) {</div>
<div class="line"><a name="l00199"></a><span class="lineno"> 199</span>  binaryDistanceAnySize<64, 3, BinaryType></div>
<div class="line"><a name="l00200"></a><span class="lineno"> 200</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00201"></a><span class="lineno"> 201</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00202"></a><span class="lineno"> 202</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 128) {</div>
<div class="line"><a name="l00203"></a><span class="lineno"> 203</span>  binaryDistanceAnySize<128, 3, BinaryType></div>
<div class="line"><a name="l00204"></a><span class="lineno"> 204</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00205"></a><span class="lineno"> 205</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00206"></a><span class="lineno"> 206</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 256) {</div>
<div class="line"><a name="l00207"></a><span class="lineno"> 207</span>  binaryDistanceAnySize<256, 4, BinaryType></div>
<div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 512) {</div>
<div class="line"><a name="l00211"></a><span class="lineno"> 211</span>  binaryDistanceAnySize<512, 8, BinaryType></div>
<div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 1024) {</div>
<div class="line"><a name="l00215"></a><span class="lineno"> 215</span>  binaryDistanceAnySize<1024, 8, BinaryType></div>
<div class="line"><a name="l00216"></a><span class="lineno"> 216</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00218"></a><span class="lineno"> 218</span>  }</div>
<div class="line"><a name="l00219"></a><span class="lineno"> 219</span> <span class="preprocessor">#if GPU_MAX_SELECTION_K >= 2048</span></div>
<div class="line"><a name="l00220"></a><span class="lineno"> 220</span> <span class="preprocessor"></span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 2048) {</div>
<div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  binaryDistanceAnySize<2048, 8, BinaryType></div>
<div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  vecs, query, outK, outV, k);</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> <span class="preprocessor">#endif</span></div>
<div class="line"><a name="l00226"></a><span class="lineno"> 226</span> <span class="preprocessor"></span>}</div>
<div class="line"><a name="l00227"></a><span class="lineno"> 227</span> </div>
<div class="line"><a name="l00228"></a><span class="lineno"> 228</span> <span class="keyword">template</span> <<span class="keyword">typename</span> BinaryType, <span class="keywordtype">int</span> ReductionLimit></div>
<div class="line"><a name="l00229"></a><span class="lineno"> 229</span> <span class="keywordtype">void</span> runBinaryDistanceLimitSize(Tensor<BinaryType, 2, true>& vecs,</div>
<div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  Tensor<BinaryType, 2, true>& query,</div>
<div class="line"><a name="l00231"></a><span class="lineno"> 231</span>  Tensor<int, 2, true>& outK,</div>
<div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  Tensor<int, 2, true>& outV,</div>
<div class="line"><a name="l00233"></a><span class="lineno"> 233</span>  <span class="keywordtype">int</span> k, cudaStream_t stream) {</div>
<div class="line"><a name="l00234"></a><span class="lineno"> 234</span>  dim3 grid(utils::divUp(query.getSize(0), kWarps));</div>
<div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  dim3 block(kLanes, kWarps);</div>
<div class="line"><a name="l00236"></a><span class="lineno"> 236</span> </div>
<div class="line"><a name="l00237"></a><span class="lineno"> 237</span>  <span class="keywordflow">if</span> (k == 1) {</div>
<div class="line"><a name="l00238"></a><span class="lineno"> 238</span>  binaryDistanceLimitSize<1, 1, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00240"></a><span class="lineno"> 240</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00241"></a><span class="lineno"> 241</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 32) {</div>
<div class="line"><a name="l00242"></a><span class="lineno"> 242</span>  binaryDistanceLimitSize<32, 2, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 64) {</div>
<div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  binaryDistanceLimitSize<64, 3, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 128) {</div>
<div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  binaryDistanceLimitSize<128, 3, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00253"></a><span class="lineno"> 253</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 256) {</div>
<div class="line"><a name="l00254"></a><span class="lineno"> 254</span>  binaryDistanceLimitSize<256, 4, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00255"></a><span class="lineno"> 255</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00256"></a><span class="lineno"> 256</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 512) {</div>
<div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  binaryDistanceLimitSize<512, 8, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 1024) {</div>
<div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  binaryDistanceLimitSize<1024, 8, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  vecs, query, outK, outV, k);</div>
<div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  }</div>
<div class="line"><a name="l00266"></a><span class="lineno"> 266</span> <span class="preprocessor">#if GPU_MAX_SELECTION_K >= 2048</span></div>
<div class="line"><a name="l00267"></a><span class="lineno"> 267</span> <span class="preprocessor"></span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (k <= 2048) {</div>
<div class="line"><a name="l00268"></a><span class="lineno"> 268</span>  binaryDistanceLimitSize<2048, 8, BinaryType, ReductionLimit></div>
<div class="line"><a name="l00269"></a><span class="lineno"> 269</span>  <<<grid, block, 0, stream>>>(</div>
<div class="line"><a name="l00270"></a><span class="lineno"> 270</span>  vecs, query, outK, outV, k);</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> <span class="preprocessor">#endif</span></div>
<div class="line"><a name="l00273"></a><span class="lineno"> 273</span> <span class="preprocessor"></span>}</div>
<div class="line"><a name="l00274"></a><span class="lineno"> 274</span> </div>
<div class="line"><a name="l00275"></a><span class="lineno"> 275</span> <span class="keywordtype">void</span> runBinaryDistance(Tensor<unsigned char, 2, true>& vecs,</div>
<div class="line"><a name="l00276"></a><span class="lineno"> 276</span>  Tensor<unsigned char, 2, true>& query,</div>
<div class="line"><a name="l00277"></a><span class="lineno"> 277</span>  Tensor<int, 2, true>& outK,</div>
<div class="line"><a name="l00278"></a><span class="lineno"> 278</span>  Tensor<int, 2, true>& outV,</div>
<div class="line"><a name="l00279"></a><span class="lineno"> 279</span>  <span class="keywordtype">int</span> k, cudaStream_t stream) {</div>
<div class="line"><a name="l00280"></a><span class="lineno"> 280</span>  FAISS_ASSERT(k <= GPU_MAX_SELECTION_K);</div>
<div class="line"><a name="l00281"></a><span class="lineno"> 281</span>  FAISS_ASSERT(vecs.getSize(1) == query.getSize(1));</div>
<div class="line"><a name="l00282"></a><span class="lineno"> 282</span> </div>
<div class="line"><a name="l00283"></a><span class="lineno"> 283</span>  FAISS_ASSERT(outK.getSize(1) == k);</div>
<div class="line"><a name="l00284"></a><span class="lineno"> 284</span>  FAISS_ASSERT(outV.getSize(1) == k);</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>  <span class="comment">// For the optimized uint32 kernel, we handle 32 * 8 = 256 max dims</span></div>
<div class="line"><a name="l00287"></a><span class="lineno"> 287</span>  constexpr <span class="keywordtype">int</span> kReductionLimit32 = 8;</div>
<div class="line"><a name="l00288"></a><span class="lineno"> 288</span> </div>
<div class="line"><a name="l00289"></a><span class="lineno"> 289</span>  <span class="comment">// For the optimized uint8 kernel, we handle 8 * 16 = 128 max dims</span></div>
<div class="line"><a name="l00290"></a><span class="lineno"> 290</span>  constexpr <span class="keywordtype">int</span> kReductionLimit8 = 16;</div>
<div class="line"><a name="l00291"></a><span class="lineno"> 291</span> </div>
<div class="line"><a name="l00292"></a><span class="lineno"> 292</span>  <span class="comment">// All other cases (large or small) go through the general kernel</span></div>
<div class="line"><a name="l00293"></a><span class="lineno"> 293</span> </div>
<div class="line"><a name="l00294"></a><span class="lineno"> 294</span>  <span class="keywordflow">if</span> (vecs.getSize(1) % <span class="keyword">sizeof</span>(<span class="keywordtype">unsigned</span> int) == 0 &&</div>
<div class="line"><a name="l00295"></a><span class="lineno"> 295</span>  (vecs.getSize(1) / <span class="keyword">sizeof</span>(<span class="keywordtype">unsigned</span> int)) <= kReductionLimit32) {</div>
<div class="line"><a name="l00296"></a><span class="lineno"> 296</span>  <span class="keyword">auto</span> vecs32 = vecs.castResize<<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>>();</div>
<div class="line"><a name="l00297"></a><span class="lineno"> 297</span>  <span class="keyword">auto</span> query32 = query.castResize<<span class="keywordtype">unsigned</span> <span class="keywordtype">int</span>>();</div>
<div class="line"><a name="l00298"></a><span class="lineno"> 298</span> </div>
<div class="line"><a name="l00299"></a><span class="lineno"> 299</span>  <span class="comment">// Optimize for vectors with dimensions a multiple of 32 that are less than</span></div>
<div class="line"><a name="l00300"></a><span class="lineno"> 300</span>  <span class="comment">// 32 * kReductionLimit (256) dimensions in size</span></div>
<div class="line"><a name="l00301"></a><span class="lineno"> 301</span>  runBinaryDistanceLimitSize<unsigned int, kReductionLimit32>(</div>
<div class="line"><a name="l00302"></a><span class="lineno"> 302</span>  vecs32, query32, outK, outV, k, stream);</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>  } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (vecs.getSize(1) <= kReductionLimit8) {</div>
<div class="line"><a name="l00305"></a><span class="lineno"> 305</span>  <span class="comment">// Optimize for vectors with dimensions a multiple of 32 that are less than</span></div>
<div class="line"><a name="l00306"></a><span class="lineno"> 306</span>  <span class="comment">// 32 * kReductionLimit (256) dimensions in size</span></div>
<div class="line"><a name="l00307"></a><span class="lineno"> 307</span>  runBinaryDistanceLimitSize<unsigned char, kReductionLimit8>(</div>
<div class="line"><a name="l00308"></a><span class="lineno"> 308</span>  vecs, query, outK, outV, k, stream);</div>
<div class="line"><a name="l00309"></a><span class="lineno"> 309</span>  } <span class="keywordflow">else</span> {</div>
<div class="line"><a name="l00310"></a><span class="lineno"> 310</span>  <span class="comment">// Arbitrary size kernel</span></div>
<div class="line"><a name="l00311"></a><span class="lineno"> 311</span>  runBinaryDistanceAnySize<unsigned char>(</div>
<div class="line"><a name="l00312"></a><span class="lineno"> 312</span>  vecs, query, outK, outV, k, stream);</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> }</div>
<div class="line"><a name="l00315"></a><span class="lineno"> 315</span> </div>
<div class="line"><a name="l00316"></a><span class="lineno"> 316</span> } } <span class="comment">// namespace</span></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>