mirror of
https://github.com/facebookresearch/faiss.git
synced 2025-06-03 21:54:02 +08:00
0df3f646fc
various bugfixes from github issues kmean with some frozen centroids GPU better tiling for large flat datasets default AVX for vector ops
370 lines
43 KiB
HTML
370 lines
43 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: /data/users/matthijs/github_faiss/faiss/gpu/impl/L2Norm.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">L2Norm.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) 2015-present, Facebook, Inc.</span></div>
|
|
<div class="line"><a name="l00003"></a><span class="lineno"> 3</span> <span class="comment"> * All rights reserved.</span></div>
|
|
<div class="line"><a name="l00004"></a><span class="lineno"> 4</span> <span class="comment"> *</span></div>
|
|
<div class="line"><a name="l00005"></a><span class="lineno"> 5</span> <span class="comment"> * This source code is licensed under the BSD+Patents license found in the</span></div>
|
|
<div class="line"><a name="l00006"></a><span class="lineno"> 6</span> <span class="comment"> * LICENSE file in the root directory of this source tree.</span></div>
|
|
<div class="line"><a name="l00007"></a><span class="lineno"> 7</span> <span class="comment"> */</span></div>
|
|
<div class="line"><a name="l00008"></a><span class="lineno"> 8</span> </div>
|
|
<div class="line"><a name="l00009"></a><span class="lineno"> 9</span> <span class="comment">// Copyright 2004-present Facebook. All Rights Reserved.</span></div>
|
|
<div class="line"><a name="l00010"></a><span class="lineno"> 10</span> </div>
|
|
<div class="line"><a name="l00011"></a><span class="lineno"> 11</span> <span class="preprocessor">#include "L2Norm.cuh"</span></div>
|
|
<div class="line"><a name="l00012"></a><span class="lineno"> 12</span> <span class="preprocessor">#include "../../FaissAssert.h"</span></div>
|
|
<div class="line"><a name="l00013"></a><span class="lineno"> 13</span> <span class="preprocessor">#include "../utils/ConversionOperators.cuh"</span></div>
|
|
<div class="line"><a name="l00014"></a><span class="lineno"> 14</span> <span class="preprocessor">#include "../utils/DeviceDefs.cuh"</span></div>
|
|
<div class="line"><a name="l00015"></a><span class="lineno"> 15</span> <span class="preprocessor">#include "../utils/DeviceUtils.h"</span></div>
|
|
<div class="line"><a name="l00016"></a><span class="lineno"> 16</span> <span class="preprocessor">#include "../utils/Float16.cuh"</span></div>
|
|
<div class="line"><a name="l00017"></a><span class="lineno"> 17</span> <span class="preprocessor">#include "../utils/MathOperators.cuh"</span></div>
|
|
<div class="line"><a name="l00018"></a><span class="lineno"> 18</span> <span class="preprocessor">#include "../utils/PtxUtils.cuh"</span></div>
|
|
<div class="line"><a name="l00019"></a><span class="lineno"> 19</span> <span class="preprocessor">#include "../utils/StaticUtils.h"</span></div>
|
|
<div class="line"><a name="l00020"></a><span class="lineno"> 20</span> <span class="preprocessor">#include "../utils/Reductions.cuh"</span></div>
|
|
<div class="line"><a name="l00021"></a><span class="lineno"> 21</span> </div>
|
|
<div class="line"><a name="l00022"></a><span class="lineno"> 22</span> <span class="keyword">namespace </span>faiss { <span class="keyword">namespace </span>gpu {</div>
|
|
<div class="line"><a name="l00023"></a><span class="lineno"> 23</span> </div>
|
|
<div class="line"><a name="l00024"></a><span class="lineno"> 24</span> <span class="comment">// Input: (batch x dim), # repeats</span></div>
|
|
<div class="line"><a name="l00025"></a><span class="lineno"> 25</span> <span class="comment">// Output: (# repeats, norm of batch vector)</span></div>
|
|
<div class="line"><a name="l00026"></a><span class="lineno"> 26</span> <span class="comment">// Done under the presumption that the dimension size is not too large</span></div>
|
|
<div class="line"><a name="l00027"></a><span class="lineno"> 27</span> <span class="comment">// (<10k or so), since there wouldn't be enough parallelism applying a</span></div>
|
|
<div class="line"><a name="l00028"></a><span class="lineno"> 28</span> <span class="comment">// single block to the problem. Also that each vector is large enough</span></div>
|
|
<div class="line"><a name="l00029"></a><span class="lineno"> 29</span> <span class="comment">// (>64), since a single block works on multiple rows' norms at the</span></div>
|
|
<div class="line"><a name="l00030"></a><span class="lineno"> 30</span> <span class="comment">// same time.</span></div>
|
|
<div class="line"><a name="l00031"></a><span class="lineno"> 31</span> <span class="comment">// T: the type we are doing the math in (e.g., float, half)</span></div>
|
|
<div class="line"><a name="l00032"></a><span class="lineno"> 32</span> <span class="comment">// TVec: the potentially vectorized type we are loading in (e.g.,</span></div>
|
|
<div class="line"><a name="l00033"></a><span class="lineno"> 33</span> <span class="comment">// float4, half2)</span></div>
|
|
<div class="line"><a name="l00034"></a><span class="lineno"> 34</span> <span class="keyword">template</span> <<span class="keyword">typename</span> T, <span class="keyword">typename</span> TVec, <span class="keyword">typename</span> TIndex,</div>
|
|
<div class="line"><a name="l00035"></a><span class="lineno"> 35</span>  <span class="keywordtype">int</span> RowTileSize, <span class="keywordtype">bool</span> NormLoop, <span class="keywordtype">bool</span> NormSquared></div>
|
|
<div class="line"><a name="l00036"></a><span class="lineno"> 36</span> __global__ <span class="keywordtype">void</span> l2Norm(Tensor<TVec, 2, true, TIndex> input,</div>
|
|
<div class="line"><a name="l00037"></a><span class="lineno"> 37</span>  Tensor<T, 1, true, TIndex> output) {</div>
|
|
<div class="line"><a name="l00038"></a><span class="lineno"> 38</span>  <span class="keyword">extern</span> __shared__ <span class="keywordtype">char</span> smemByte[]; <span class="comment">// #warps * RowTileSize elements</span></div>
|
|
<div class="line"><a name="l00039"></a><span class="lineno"> 39</span>  T* smem = (T*) smemByte;</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>  TIndex numWarps = utils::divUp(blockDim.x, kWarpSize);</div>
|
|
<div class="line"><a name="l00042"></a><span class="lineno"> 42</span>  TIndex laneId = getLaneId();</div>
|
|
<div class="line"><a name="l00043"></a><span class="lineno"> 43</span>  TIndex warpId = threadIdx.x / kWarpSize;</div>
|
|
<div class="line"><a name="l00044"></a><span class="lineno"> 44</span> </div>
|
|
<div class="line"><a name="l00045"></a><span class="lineno"> 45</span>  <span class="keywordtype">bool</span> lastRowTile = (blockIdx.x == (gridDim.x - 1));</div>
|
|
<div class="line"><a name="l00046"></a><span class="lineno"> 46</span>  TIndex rowStart = RowTileSize * blockIdx.x;</div>
|
|
<div class="line"><a name="l00047"></a><span class="lineno"> 47</span>  T rowNorm[RowTileSize];</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>  <span class="keywordflow">if</span> (lastRowTile) {</div>
|
|
<div class="line"><a name="l00050"></a><span class="lineno"> 50</span>  <span class="comment">// We are handling the very end of the input matrix rows</span></div>
|
|
<div class="line"><a name="l00051"></a><span class="lineno"> 51</span>  <span class="keywordflow">for</span> (TIndex row = 0; row < input.getSize(0) - rowStart; ++row) {</div>
|
|
<div class="line"><a name="l00052"></a><span class="lineno"> 52</span>  <span class="keywordflow">if</span> (NormLoop) {</div>
|
|
<div class="line"><a name="l00053"></a><span class="lineno"> 53</span>  rowNorm[0] = Math<T>::zero();</div>
|
|
<div class="line"><a name="l00054"></a><span class="lineno"> 54</span> </div>
|
|
<div class="line"><a name="l00055"></a><span class="lineno"> 55</span>  <span class="keywordflow">for</span> (TIndex col = threadIdx.x;</div>
|
|
<div class="line"><a name="l00056"></a><span class="lineno"> 56</span>  col < input.getSize(1); col += blockDim.x) {</div>
|
|
<div class="line"><a name="l00057"></a><span class="lineno"> 57</span>  TVec val = input[rowStart + row][col];</div>
|
|
<div class="line"><a name="l00058"></a><span class="lineno"> 58</span>  val = Math<TVec>::mul(val, val);</div>
|
|
<div class="line"><a name="l00059"></a><span class="lineno"> 59</span>  rowNorm[0] = Math<T>::add(rowNorm[0], Math<TVec>::reduceAdd(val));</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">else</span> {</div>
|
|
<div class="line"><a name="l00062"></a><span class="lineno"> 62</span>  TVec val = input[rowStart + row][threadIdx.x];</div>
|
|
<div class="line"><a name="l00063"></a><span class="lineno"> 63</span>  val = Math<TVec>::mul(val, val);</div>
|
|
<div class="line"><a name="l00064"></a><span class="lineno"> 64</span>  rowNorm[0] = <a class="code" href="structfaiss_1_1gpu_1_1Math.html#a4b17f0b5d014f300e76dde5b24af8014">Math<TVec>::reduceAdd</a>(val);</div>
|
|
<div class="line"><a name="l00065"></a><span class="lineno"> 65</span>  }</div>
|
|
<div class="line"><a name="l00066"></a><span class="lineno"> 66</span> </div>
|
|
<div class="line"><a name="l00067"></a><span class="lineno"> 67</span>  rowNorm[0] = warpReduceAllSum(rowNorm[0]);</div>
|
|
<div class="line"><a name="l00068"></a><span class="lineno"> 68</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00069"></a><span class="lineno"> 69</span>  smem[row * numWarps + warpId] = rowNorm[0];</div>
|
|
<div class="line"><a name="l00070"></a><span class="lineno"> 70</span>  }</div>
|
|
<div class="line"><a name="l00071"></a><span class="lineno"> 71</span>  }</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>  <span class="comment">// We are guaranteed that all RowTileSize rows are available in</span></div>
|
|
<div class="line"><a name="l00074"></a><span class="lineno"> 74</span>  <span class="comment">// [rowStart, rowStart + RowTileSize)</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>  <span class="keywordflow">if</span> (NormLoop) {</div>
|
|
<div class="line"><a name="l00077"></a><span class="lineno"> 77</span>  <span class="comment">// A single block of threads is not big enough to span each</span></div>
|
|
<div class="line"><a name="l00078"></a><span class="lineno"> 78</span>  <span class="comment">// vector</span></div>
|
|
<div class="line"><a name="l00079"></a><span class="lineno"> 79</span>  TVec tmp[RowTileSize];</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> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00082"></a><span class="lineno"> 82</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00083"></a><span class="lineno"> 83</span>  rowNorm[row] = Math<T>::zero();</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>  <span class="keywordflow">for</span> (TIndex col = threadIdx.x;</div>
|
|
<div class="line"><a name="l00087"></a><span class="lineno"> 87</span>  col < input.getSize(1); col += blockDim.x) {</div>
|
|
<div class="line"><a name="l00088"></a><span class="lineno"> 88</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00089"></a><span class="lineno"> 89</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00090"></a><span class="lineno"> 90</span>  tmp[row] = input[rowStart + row][col];</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> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00094"></a><span class="lineno"> 94</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00095"></a><span class="lineno"> 95</span>  tmp[row] = Math<TVec>::mul(tmp[row], tmp[row]);</div>
|
|
<div class="line"><a name="l00096"></a><span class="lineno"> 96</span>  }</div>
|
|
<div class="line"><a name="l00097"></a><span class="lineno"> 97</span> </div>
|
|
<div class="line"><a name="l00098"></a><span class="lineno"> 98</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00099"></a><span class="lineno"> 99</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00100"></a><span class="lineno"> 100</span>  rowNorm[row] = Math<T>::add(rowNorm[row],</div>
|
|
<div class="line"><a name="l00101"></a><span class="lineno"> 101</span>  Math<TVec>::reduceAdd(tmp[row]));</div>
|
|
<div class="line"><a name="l00102"></a><span class="lineno"> 102</span>  }</div>
|
|
<div class="line"><a name="l00103"></a><span class="lineno"> 103</span>  }</div>
|
|
<div class="line"><a name="l00104"></a><span class="lineno"> 104</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00105"></a><span class="lineno"> 105</span>  TVec tmp[RowTileSize];</div>
|
|
<div class="line"><a name="l00106"></a><span class="lineno"> 106</span> </div>
|
|
<div class="line"><a name="l00107"></a><span class="lineno"> 107</span>  <span class="comment">// A block of threads is the exact size of the vector</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> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00110"></a><span class="lineno"> 110</span>  tmp[row] = input[rowStart + row][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> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00115"></a><span class="lineno"> 115</span>  tmp[row] = Math<TVec>::mul(tmp[row], tmp[row]);</div>
|
|
<div class="line"><a name="l00116"></a><span class="lineno"> 116</span>  }</div>
|
|
<div class="line"><a name="l00117"></a><span class="lineno"> 117</span> </div>
|
|
<div class="line"><a name="l00118"></a><span class="lineno"> 118</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00119"></a><span class="lineno"> 119</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00120"></a><span class="lineno"> 120</span>  rowNorm[row] = <a class="code" href="structfaiss_1_1gpu_1_1Math.html#a4b17f0b5d014f300e76dde5b24af8014">Math<TVec>::reduceAdd</a>(tmp[row]);</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>  }</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="comment">// Sum up all parts in each warp</span></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> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00127"></a><span class="lineno"> 127</span>  rowNorm[row] = warpReduceAllSum(rowNorm[row]);</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>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00131"></a><span class="lineno"> 131</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00132"></a><span class="lineno"> 132</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00133"></a><span class="lineno"> 133</span>  smem[row * numWarps + warpId] = rowNorm[row];</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>  }</div>
|
|
<div class="line"><a name="l00136"></a><span class="lineno"> 136</span>  }</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>  __syncthreads();</div>
|
|
<div class="line"><a name="l00139"></a><span class="lineno"> 139</span> </div>
|
|
<div class="line"><a name="l00140"></a><span class="lineno"> 140</span>  <span class="comment">// Sum across warps</span></div>
|
|
<div class="line"><a name="l00141"></a><span class="lineno"> 141</span>  <span class="keywordflow">if</span> (warpId == 0) {</div>
|
|
<div class="line"><a name="l00142"></a><span class="lineno"> 142</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00143"></a><span class="lineno"> 143</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00144"></a><span class="lineno"> 144</span>  rowNorm[row] = laneId < numWarps ?</div>
|
|
<div class="line"><a name="l00145"></a><span class="lineno"> 145</span>  smem[row * numWarps + laneId] : Math<T>::zero();</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> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00149"></a><span class="lineno"> 149</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00150"></a><span class="lineno"> 150</span>  rowNorm[row] = warpReduceAllSum(rowNorm[row]);</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>  <span class="comment">// Write out answer</span></div>
|
|
<div class="line"><a name="l00154"></a><span class="lineno"> 154</span>  <span class="keywordflow">if</span> (laneId == 0) {</div>
|
|
<div class="line"><a name="l00155"></a><span class="lineno"> 155</span> <span class="preprocessor">#pragma unroll</span></div>
|
|
<div class="line"><a name="l00156"></a><span class="lineno"> 156</span> <span class="preprocessor"></span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row < RowTileSize; ++row) {</div>
|
|
<div class="line"><a name="l00157"></a><span class="lineno"> 157</span>  <span class="keywordtype">int</span> outCol = rowStart + row;</div>
|
|
<div class="line"><a name="l00158"></a><span class="lineno"> 158</span> </div>
|
|
<div class="line"><a name="l00159"></a><span class="lineno"> 159</span>  <span class="keywordflow">if</span> (lastRowTile) {</div>
|
|
<div class="line"><a name="l00160"></a><span class="lineno"> 160</span>  <span class="keywordflow">if</span> (outCol < output.getSize(0)) {</div>
|
|
<div class="line"><a name="l00161"></a><span class="lineno"> 161</span>  output[outCol] =</div>
|
|
<div class="line"><a name="l00162"></a><span class="lineno"> 162</span>  NormSquared ? rowNorm[row] :</div>
|
|
<div class="line"><a name="l00163"></a><span class="lineno"> 163</span>  ConvertTo<T>::to(</div>
|
|
<div class="line"><a name="l00164"></a><span class="lineno"> 164</span>  sqrtf(ConvertTo<float>::to(rowNorm[row])));</div>
|
|
<div class="line"><a name="l00165"></a><span class="lineno"> 165</span>  }</div>
|
|
<div class="line"><a name="l00166"></a><span class="lineno"> 166</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00167"></a><span class="lineno"> 167</span>  output[outCol] =</div>
|
|
<div class="line"><a name="l00168"></a><span class="lineno"> 168</span>  NormSquared ? rowNorm[row] :</div>
|
|
<div class="line"><a name="l00169"></a><span class="lineno"> 169</span>  ConvertTo<T>::to(</div>
|
|
<div class="line"><a name="l00170"></a><span class="lineno"> 170</span>  sqrtf(ConvertTo<float>::to(rowNorm[row])));</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>  }</div>
|
|
<div class="line"><a name="l00173"></a><span class="lineno"> 173</span>  }</div>
|
|
<div class="line"><a name="l00174"></a><span class="lineno"> 174</span>  }</div>
|
|
<div class="line"><a name="l00175"></a><span class="lineno"> 175</span> }</div>
|
|
<div class="line"><a name="l00176"></a><span class="lineno"> 176</span> </div>
|
|
<div class="line"><a name="l00177"></a><span class="lineno"> 177</span> <span class="keyword">template</span> <<span class="keyword">typename</span> T, <span class="keyword">typename</span> TVec, <span class="keyword">typename</span> TIndex></div>
|
|
<div class="line"><a name="l00178"></a><span class="lineno"> 178</span> <span class="keywordtype">void</span> runL2Norm(Tensor<T, 2, true, TIndex>& input,</div>
|
|
<div class="line"><a name="l00179"></a><span class="lineno"> 179</span>  Tensor<T, 1, true, TIndex>& output,</div>
|
|
<div class="line"><a name="l00180"></a><span class="lineno"> 180</span>  <span class="keywordtype">bool</span> normSquared,</div>
|
|
<div class="line"><a name="l00181"></a><span class="lineno"> 181</span>  cudaStream_t stream) {</div>
|
|
<div class="line"><a name="l00182"></a><span class="lineno"> 182</span>  FAISS_ASSERT(input.getSize(0) == output.getSize(0));</div>
|
|
<div class="line"><a name="l00183"></a><span class="lineno"> 183</span> </div>
|
|
<div class="line"><a name="l00184"></a><span class="lineno"> 184</span>  TIndex maxThreads = (TIndex) getMaxThreadsCurrentDevice();</div>
|
|
<div class="line"><a name="l00185"></a><span class="lineno"> 185</span>  constexpr <span class="keywordtype">int</span> rowTileSize = 8;</div>
|
|
<div class="line"><a name="l00186"></a><span class="lineno"> 186</span> </div>
|
|
<div class="line"><a name="l00187"></a><span class="lineno"> 187</span> <span class="preprocessor">#define RUN_L2(TYPE_T, TYPE_TVEC, INPUT) \</span></div>
|
|
<div class="line"><a name="l00188"></a><span class="lineno"> 188</span> <span class="preprocessor"> do { \</span></div>
|
|
<div class="line"><a name="l00189"></a><span class="lineno"> 189</span> <span class="preprocessor"> if (normLoop) { \</span></div>
|
|
<div class="line"><a name="l00190"></a><span class="lineno"> 190</span> <span class="preprocessor"> if (normSquared) { \</span></div>
|
|
<div class="line"><a name="l00191"></a><span class="lineno"> 191</span> <span class="preprocessor"> l2Norm<TYPE_T, TYPE_TVEC, TIndex, rowTileSize, true, true> \</span></div>
|
|
<div class="line"><a name="l00192"></a><span class="lineno"> 192</span> <span class="preprocessor"> <<<grid, block, smem, stream>>>(INPUT, output); \</span></div>
|
|
<div class="line"><a name="l00193"></a><span class="lineno"> 193</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00194"></a><span class="lineno"> 194</span> <span class="preprocessor"> l2Norm<TYPE_T, TYPE_TVEC, TIndex, rowTileSize, true, false> \</span></div>
|
|
<div class="line"><a name="l00195"></a><span class="lineno"> 195</span> <span class="preprocessor"> <<<grid, block, smem, stream>>>(INPUT, output); \</span></div>
|
|
<div class="line"><a name="l00196"></a><span class="lineno"> 196</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00197"></a><span class="lineno"> 197</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00198"></a><span class="lineno"> 198</span> <span class="preprocessor"> if (normSquared) { \</span></div>
|
|
<div class="line"><a name="l00199"></a><span class="lineno"> 199</span> <span class="preprocessor"> l2Norm<TYPE_T, TYPE_TVEC, TIndex, rowTileSize, false, true> \</span></div>
|
|
<div class="line"><a name="l00200"></a><span class="lineno"> 200</span> <span class="preprocessor"> <<<grid, block, smem, stream>>>(INPUT, output); \</span></div>
|
|
<div class="line"><a name="l00201"></a><span class="lineno"> 201</span> <span class="preprocessor"> } else { \</span></div>
|
|
<div class="line"><a name="l00202"></a><span class="lineno"> 202</span> <span class="preprocessor"> l2Norm<TYPE_T, TYPE_TVEC, TIndex, rowTileSize, false, false> \</span></div>
|
|
<div class="line"><a name="l00203"></a><span class="lineno"> 203</span> <span class="preprocessor"> <<<grid, block, smem, stream>>>(INPUT, output); \</span></div>
|
|
<div class="line"><a name="l00204"></a><span class="lineno"> 204</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00205"></a><span class="lineno"> 205</span> <span class="preprocessor"> } \</span></div>
|
|
<div class="line"><a name="l00206"></a><span class="lineno"> 206</span> <span class="preprocessor"> } while (0)</span></div>
|
|
<div class="line"><a name="l00207"></a><span class="lineno"> 207</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00208"></a><span class="lineno"> 208</span>  <span class="keywordflow">if</span> (input.template canCastResize<TVec>()) {</div>
|
|
<div class="line"><a name="l00209"></a><span class="lineno"> 209</span>  <span class="comment">// Can load using the vectorized type</span></div>
|
|
<div class="line"><a name="l00210"></a><span class="lineno"> 210</span>  <span class="keyword">auto</span> inputV = input.template castResize<TVec>();</div>
|
|
<div class="line"><a name="l00211"></a><span class="lineno"> 211</span> </div>
|
|
<div class="line"><a name="l00212"></a><span class="lineno"> 212</span>  <span class="keyword">auto</span> dim = inputV.getSize(1);</div>
|
|
<div class="line"><a name="l00213"></a><span class="lineno"> 213</span>  <span class="keywordtype">bool</span> normLoop = dim > maxThreads;</div>
|
|
<div class="line"><a name="l00214"></a><span class="lineno"> 214</span>  <span class="keyword">auto</span> numThreads = min(dim, maxThreads);</div>
|
|
<div class="line"><a name="l00215"></a><span class="lineno"> 215</span> </div>
|
|
<div class="line"><a name="l00216"></a><span class="lineno"> 216</span>  <span class="keyword">auto</span> grid = dim3(utils::divUp(inputV.getSize(0), rowTileSize));</div>
|
|
<div class="line"><a name="l00217"></a><span class="lineno"> 217</span>  <span class="keyword">auto</span> block = dim3(numThreads);</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="keyword">auto</span> smem = <span class="keyword">sizeof</span>(T) * rowTileSize * utils::divUp(numThreads, kWarpSize);</div>
|
|
<div class="line"><a name="l00220"></a><span class="lineno"> 220</span> </div>
|
|
<div class="line"><a name="l00221"></a><span class="lineno"> 221</span>  RUN_L2(T, TVec, inputV);</div>
|
|
<div class="line"><a name="l00222"></a><span class="lineno"> 222</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00223"></a><span class="lineno"> 223</span>  <span class="comment">// Can't load using the vectorized type</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>  <span class="keyword">auto</span> dim = input.getSize(1);</div>
|
|
<div class="line"><a name="l00226"></a><span class="lineno"> 226</span>  <span class="keywordtype">bool</span> normLoop = dim > maxThreads;</div>
|
|
<div class="line"><a name="l00227"></a><span class="lineno"> 227</span>  <span class="keyword">auto</span> numThreads = min(dim, maxThreads);</div>
|
|
<div class="line"><a name="l00228"></a><span class="lineno"> 228</span> </div>
|
|
<div class="line"><a name="l00229"></a><span class="lineno"> 229</span>  <span class="keyword">auto</span> grid = dim3(utils::divUp(input.getSize(0), rowTileSize));</div>
|
|
<div class="line"><a name="l00230"></a><span class="lineno"> 230</span>  <span class="keyword">auto</span> block = dim3(numThreads);</div>
|
|
<div class="line"><a name="l00231"></a><span class="lineno"> 231</span> </div>
|
|
<div class="line"><a name="l00232"></a><span class="lineno"> 232</span>  <span class="keyword">auto</span> smem = <span class="keyword">sizeof</span>(T) * rowTileSize * utils::divUp(numThreads, kWarpSize);</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>  RUN_L2(T, T, input);</div>
|
|
<div class="line"><a name="l00235"></a><span class="lineno"> 235</span>  }</div>
|
|
<div class="line"><a name="l00236"></a><span class="lineno"> 236</span> </div>
|
|
<div class="line"><a name="l00237"></a><span class="lineno"> 237</span> <span class="preprocessor">#undef RUN_L2</span></div>
|
|
<div class="line"><a name="l00238"></a><span class="lineno"> 238</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00239"></a><span class="lineno"> 239</span>  CUDA_TEST_ERROR();</div>
|
|
<div class="line"><a name="l00240"></a><span class="lineno"> 240</span> }</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">void</span> runL2Norm(Tensor<float, 2, true>& input,</div>
|
|
<div class="line"><a name="l00243"></a><span class="lineno"> 243</span>  Tensor<float, 1, true>& output,</div>
|
|
<div class="line"><a name="l00244"></a><span class="lineno"> 244</span>  <span class="keywordtype">bool</span> normSquared,</div>
|
|
<div class="line"><a name="l00245"></a><span class="lineno"> 245</span>  cudaStream_t stream) {</div>
|
|
<div class="line"><a name="l00246"></a><span class="lineno"> 246</span>  <span class="keywordflow">if</span> (input.canUseIndexType<<span class="keywordtype">int</span>>()) {</div>
|
|
<div class="line"><a name="l00247"></a><span class="lineno"> 247</span>  runL2Norm<float, float4, int>(input, output, normSquared, stream);</div>
|
|
<div class="line"><a name="l00248"></a><span class="lineno"> 248</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00249"></a><span class="lineno"> 249</span>  <span class="keyword">auto</span> inputCast = input.castIndexType<<span class="keywordtype">long</span>>();</div>
|
|
<div class="line"><a name="l00250"></a><span class="lineno"> 250</span>  <span class="keyword">auto</span> outputCast = output.castIndexType<<span class="keywordtype">long</span>>();</div>
|
|
<div class="line"><a name="l00251"></a><span class="lineno"> 251</span>  runL2Norm<float, float4, long>(inputCast, outputCast, normSquared, stream);</div>
|
|
<div class="line"><a name="l00252"></a><span class="lineno"> 252</span>  }</div>
|
|
<div class="line"><a name="l00253"></a><span class="lineno"> 253</span> }</div>
|
|
<div class="line"><a name="l00254"></a><span class="lineno"> 254</span> </div>
|
|
<div class="line"><a name="l00255"></a><span class="lineno"> 255</span> <span class="preprocessor">#ifdef FAISS_USE_FLOAT16</span></div>
|
|
<div class="line"><a name="l00256"></a><span class="lineno"> 256</span> <span class="preprocessor"></span><span class="keywordtype">void</span> runL2Norm(Tensor<half, 2, true>& input,</div>
|
|
<div class="line"><a name="l00257"></a><span class="lineno"> 257</span>  Tensor<half, 1, true>& output,</div>
|
|
<div class="line"><a name="l00258"></a><span class="lineno"> 258</span>  <span class="keywordtype">bool</span> normSquared,</div>
|
|
<div class="line"><a name="l00259"></a><span class="lineno"> 259</span>  cudaStream_t stream) {</div>
|
|
<div class="line"><a name="l00260"></a><span class="lineno"> 260</span>  <span class="keywordflow">if</span> (input.canUseIndexType<<span class="keywordtype">int</span>>()) {</div>
|
|
<div class="line"><a name="l00261"></a><span class="lineno"> 261</span>  runL2Norm<half, half2, int>(input, output, normSquared, stream);</div>
|
|
<div class="line"><a name="l00262"></a><span class="lineno"> 262</span>  } <span class="keywordflow">else</span> {</div>
|
|
<div class="line"><a name="l00263"></a><span class="lineno"> 263</span>  <span class="keyword">auto</span> inputCast = input.castIndexType<<span class="keywordtype">long</span>>();</div>
|
|
<div class="line"><a name="l00264"></a><span class="lineno"> 264</span>  <span class="keyword">auto</span> outputCast = output.castIndexType<<span class="keywordtype">long</span>>();</div>
|
|
<div class="line"><a name="l00265"></a><span class="lineno"> 265</span>  runL2Norm<half, half2, long>(inputCast, outputCast, normSquared, stream);</div>
|
|
<div class="line"><a name="l00266"></a><span class="lineno"> 266</span>  }</div>
|
|
<div class="line"><a name="l00267"></a><span class="lineno"> 267</span> }</div>
|
|
<div class="line"><a name="l00268"></a><span class="lineno"> 268</span> <span class="preprocessor">#endif</span></div>
|
|
<div class="line"><a name="l00269"></a><span class="lineno"> 269</span> <span class="preprocessor"></span></div>
|
|
<div class="line"><a name="l00270"></a><span class="lineno"> 270</span> } } <span class="comment">// namespace</span></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#l00044">MathOperators.cuh:44</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>
|