zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-06-26 02:33:21 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity
c7080f89ca
mlx/docs/build/html/kernels_2softmax_8h_source.html
CircleCI Docs c7080f89ca rebase
2025-06-04 01:01:55 +00:00

322 lines
36 KiB
HTML
Raw Blame History

<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "https://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml" lang="en-US">
<head>
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=11"/>
<meta name="generator" content="Doxygen 1.13.2"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>MLX: mlx/backend/metal/kernels/softmax.h 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>
<script type="text/javascript" src="clipboard.js"></script>
<link href="navtree.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="navtreedata.js"></script>
<script type="text/javascript" src="navtree.js"></script>
<script type="text/javascript" src="resize.js"></script>
<script type="text/javascript" src="cookie.js"></script>
<link href="search/search.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="search/searchdata.js"></script>
<script type="text/javascript" src="search/search.js"></script>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function() { init_search(); });
/* @license-end */
</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 id="projectrow">
<td id="projectalign">
<div id="projectname">MLX
</div>
</td>
<td> <div id="MSearchBox" class="MSearchBoxInactive">
<span class="left">
<span id="MSearchSelect" onmouseover="return searchBox.OnSearchSelectShow()" onmouseout="return searchBox.OnSearchSelectHide()">&#160;</span>
<input type="text" id="MSearchField" value="" placeholder="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.svg" alt=""/></a>
</span>
</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.13.2 -->
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
var searchBox = new SearchBox("searchBox", "search/",'.html');
/* @license-end */
</script>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function() { codefold.init(0); });
/* @license-end */
</script>
</div><!-- top -->
<div id="side-nav" class="ui-resizable side-nav-resizable">
<div id="nav-tree">
<div id="nav-tree-contents">
<div id="nav-sync" class="sync"></div>
</div>
</div>
<div id="splitbar" style="-moz-user-select:none;"
class="ui-resizable-handle">
</div>
</div>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function(){initNavTree('kernels_2softmax_8h_source.html',''); initResizable(true); });
/* @license-end */
</script>
<div id="doc-content">
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
onkeydown="return searchBox.OnSearchSelectKey(event)">
</div>
<!-- iframe showing the search results (closed by default) -->
<div id="MSearchResultsWindow">
<div id="MSearchResults">
<div class="SRPage">
<div id="SRIndex">
<div id="SRResults"></div>
<div class="SRStatus" id="Loading">Loading...</div>
<div class="SRStatus" id="Searching">Searching...</div>
<div class="SRStatus" id="NoMatches">No Matches</div>
</div>
</div>
</div>
</div>
<div class="header">
<div class="headertitle"><div class="title">softmax.h</div></div>
</div><!--header-->
<div class="contents">
<a href="kernels_2softmax_8h.html">Go to the documentation of this file.</a><div class="fragment"><div class="line"><a id="l00001" name="l00001"></a><span class="lineno"> 1</span><span class="comment">// Copyright © 2023-2024 Apple Inc.</span></div>
<div class="line"><a id="l00002" name="l00002"></a><span class="lineno"> 2</span> </div>
<div class="line"><a id="l00003" name="l00003"></a><span class="lineno"> 3</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T&gt;</div>
<div class="foldopen" id="foldopen00004" data-start="{" data-end="}">
<div class="line"><a id="l00004" name="l00004"></a><span class="lineno"><a class="line" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b"> 4</a></span><span class="keyword">inline</span> T <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(T x) {</div>
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span> <span class="comment">// Softmax doesn&#39;t need high precision exponential cause x is gonna be in</span></div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span> <span class="comment">// (-oo, 0] anyway and subsequently it will be divided by sum(exp(x_i)).</span></div>
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span> <span class="keywordflow">return</span> fast::exp(x);</div>
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span>}</div>
</div>
<div class="line"><a id="l00009" name="l00009"></a><span class="lineno"> 9</span> </div>
<div class="line"><a id="l00010" name="l00010"></a><span class="lineno"> 10</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> AccT = T, <span class="keywordtype">int</span> N_READS = SOFTMAX_N_READS&gt;</div>
<div class="foldopen" id="foldopen00011" data-start="{" data-end="}">
<div class="line"><a id="l00011" name="l00011"></a><span class="lineno"><a class="line" href="kernels_2softmax_8h.html#a815fe70f879f318e5d6e99acf043f52b"> 11</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="kernels_2softmax_8h.html#a815fe70f879f318e5d6e99acf043f52b">softmax_single_row</a>(</div>
<div class="line"><a id="l00012" name="l00012"></a><span class="lineno"> 12</span> <span class="keyword">const</span> device T* in,</div>
<div class="line"><a id="l00013" name="l00013"></a><span class="lineno"> 13</span> device T* out,</div>
<div class="line"><a id="l00014" name="l00014"></a><span class="lineno"> 14</span> constant <span class="keywordtype">int</span>&amp; axis_size,</div>
<div class="line"><a id="l00015" name="l00015"></a><span class="lineno"> 15</span> uint gid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00016" name="l00016"></a><span class="lineno"> 16</span> uint _lid [[thread_position_in_threadgroup]],</div>
<div class="line"><a id="l00017" name="l00017"></a><span class="lineno"> 17</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
<div class="line"><a id="l00018" name="l00018"></a><span class="lineno"> 18</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
<div class="line"><a id="l00019" name="l00019"></a><span class="lineno"> 19</span> <span class="keywordtype">int</span> lid = _lid;</div>
<div class="line"><a id="l00020" name="l00020"></a><span class="lineno"> 20</span> </div>
<div class="line"><a id="l00021" name="l00021"></a><span class="lineno"> 21</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a> = 32;</div>
<div class="line"><a id="l00022" name="l00022"></a><span class="lineno"> 22</span> </div>
<div class="line"><a id="l00023" name="l00023"></a><span class="lineno"> 23</span> threadgroup AccT local_max[<a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>];</div>
<div class="line"><a id="l00024" name="l00024"></a><span class="lineno"> 24</span> threadgroup AccT local_normalizer[<a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>];</div>
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> </div>
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> AccT ld[N_READS];</div>
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> </div>
<div class="line"><a id="l00028" name="l00028"></a><span class="lineno"> 28</span> in += gid * size_t(axis_size) + lid * N_READS;</div>
<div class="line"><a id="l00029" name="l00029"></a><span class="lineno"> 29</span> <span class="keywordflow">if</span> (lid * N_READS + N_READS &lt;= axis_size) {</div>
<div class="line"><a id="l00030" name="l00030"></a><span class="lineno"> 30</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00031" name="l00031"></a><span class="lineno"> 31</span> ld[i] = AccT(in[i]);</div>
<div class="line"><a id="l00032" name="l00032"></a><span class="lineno"> 32</span> }</div>
<div class="line"><a id="l00033" name="l00033"></a><span class="lineno"> 33</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00034" name="l00034"></a><span class="lineno"> 34</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span> ld[i] =</div>
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> ((lid * N_READS + i) &lt; axis_size) ? AccT(in[i]) : <a class="code hl_variable" href="struct_limits.html#a6e81584ba65a4dc6ff9366b458e3a20e">Limits&lt;AccT&gt;::min</a>;</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> }</div>
<div class="line"><a id="l00038" name="l00038"></a><span class="lineno"> 38</span> }</div>
<div class="line"><a id="l00039" name="l00039"></a><span class="lineno"> 39</span> <span class="keywordflow">if</span> (simd_group_id == 0) {</div>
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> local_max[simd_lane_id] = <a class="code hl_variable" href="struct_limits.html#a6e81584ba65a4dc6ff9366b458e3a20e">Limits&lt;AccT&gt;::min</a>;</div>
<div class="line"><a id="l00041" name="l00041"></a><span class="lineno"> 41</span> local_normalizer[simd_lane_id] = 0;</div>
<div class="line"><a id="l00042" name="l00042"></a><span class="lineno"> 42</span> }</div>
<div class="line"><a id="l00043" name="l00043"></a><span class="lineno"> 43</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00044" name="l00044"></a><span class="lineno"> 44</span> </div>
<div class="line"><a id="l00045" name="l00045"></a><span class="lineno"> 45</span> <span class="comment">// Get the max</span></div>
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> AccT maxval = <a class="code hl_variable" href="struct_limits.html#ae7469d21f2688797ca3e388d919ef05e">Limits&lt;AccT&gt;::finite_min</a>;</div>
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> maxval = (maxval &lt; ld[i]) ? ld[i] : maxval;</div>
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span> }</div>
<div class="line"><a id="l00050" name="l00050"></a><span class="lineno"> 50</span> maxval = simd_max(maxval);</div>
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"> 52</span> local_max[simd_group_id] = maxval;</div>
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> }</div>
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> <span class="keywordflow">if</span> (simd_group_id == 0) {</div>
<div class="line"><a id="l00056" name="l00056"></a><span class="lineno"> 56</span> maxval = simd_max(local_max[simd_lane_id]);</div>
<div class="line"><a id="l00057" name="l00057"></a><span class="lineno"> 57</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00058" name="l00058"></a><span class="lineno"> 58</span> local_max[0] = maxval;</div>
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span> }</div>
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span> }</div>
<div class="line"><a id="l00061" name="l00061"></a><span class="lineno"> 61</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00062" name="l00062"></a><span class="lineno"> 62</span> maxval = local_max[0];</div>
<div class="line"><a id="l00063" name="l00063"></a><span class="lineno"> 63</span> </div>
<div class="line"><a id="l00064" name="l00064"></a><span class="lineno"> 64</span> <span class="comment">// Compute exp(x_i - maxval) and store the partial sums in local_normalizer</span></div>
<div class="line"><a id="l00065" name="l00065"></a><span class="lineno"> 65</span> AccT normalizer = 0;</div>
<div class="line"><a id="l00066" name="l00066"></a><span class="lineno"> 66</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00067" name="l00067"></a><span class="lineno"> 67</span> AccT exp_x = <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(ld[i] - maxval);</div>
<div class="line"><a id="l00068" name="l00068"></a><span class="lineno"> 68</span> ld[i] = exp_x;</div>
<div class="line"><a id="l00069" name="l00069"></a><span class="lineno"> 69</span> normalizer += exp_x;</div>
<div class="line"><a id="l00070" name="l00070"></a><span class="lineno"> 70</span> }</div>
<div class="line"><a id="l00071" name="l00071"></a><span class="lineno"> 71</span> normalizer = simd_sum(normalizer);</div>
<div class="line"><a id="l00072" name="l00072"></a><span class="lineno"> 72</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00073" name="l00073"></a><span class="lineno"> 73</span> local_normalizer[simd_group_id] = normalizer;</div>
<div class="line"><a id="l00074" name="l00074"></a><span class="lineno"> 74</span> }</div>
<div class="line"><a id="l00075" name="l00075"></a><span class="lineno"> 75</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00076" name="l00076"></a><span class="lineno"> 76</span> <span class="keywordflow">if</span> (simd_group_id == 0) {</div>
<div class="line"><a id="l00077" name="l00077"></a><span class="lineno"> 77</span> normalizer = simd_sum(local_normalizer[simd_lane_id]);</div>
<div class="line"><a id="l00078" name="l00078"></a><span class="lineno"> 78</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00079" name="l00079"></a><span class="lineno"> 79</span> local_normalizer[0] = normalizer;</div>
<div class="line"><a id="l00080" name="l00080"></a><span class="lineno"> 80</span> }</div>
<div class="line"><a id="l00081" name="l00081"></a><span class="lineno"> 81</span> }</div>
<div class="line"><a id="l00082" name="l00082"></a><span class="lineno"> 82</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00083" name="l00083"></a><span class="lineno"> 83</span> normalizer = 1 / local_normalizer[0];</div>
<div class="line"><a id="l00084" name="l00084"></a><span class="lineno"> 84</span> </div>
<div class="line"><a id="l00085" name="l00085"></a><span class="lineno"> 85</span> <span class="comment">// Normalize and write to the output</span></div>
<div class="line"><a id="l00086" name="l00086"></a><span class="lineno"> 86</span> out += gid * size_t(axis_size) + lid * N_READS;</div>
<div class="line"><a id="l00087" name="l00087"></a><span class="lineno"> 87</span> <span class="keywordflow">if</span> (lid * N_READS + N_READS &lt;= axis_size) {</div>
<div class="line"><a id="l00088" name="l00088"></a><span class="lineno"> 88</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00089" name="l00089"></a><span class="lineno"> 89</span> out[i] = T(ld[i] * normalizer);</div>
<div class="line"><a id="l00090" name="l00090"></a><span class="lineno"> 90</span> }</div>
<div class="line"><a id="l00091" name="l00091"></a><span class="lineno"> 91</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00092" name="l00092"></a><span class="lineno"> 92</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00093" name="l00093"></a><span class="lineno"> 93</span> <span class="keywordflow">if</span> ((lid * N_READS + i) &lt; axis_size) {</div>
<div class="line"><a id="l00094" name="l00094"></a><span class="lineno"> 94</span> out[i] = T(ld[i] * normalizer);</div>
<div class="line"><a id="l00095" name="l00095"></a><span class="lineno"> 95</span> }</div>
<div class="line"><a id="l00096" name="l00096"></a><span class="lineno"> 96</span> }</div>
<div class="line"><a id="l00097" name="l00097"></a><span class="lineno"> 97</span> }</div>
<div class="line"><a id="l00098" name="l00098"></a><span class="lineno"> 98</span>}</div>
</div>
<div class="line"><a id="l00099" name="l00099"></a><span class="lineno"> 99</span> </div>
<div class="line"><a id="l00100" name="l00100"></a><span class="lineno"> 100</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> AccT = T, <span class="keywordtype">int</span> N_READS = SOFTMAX_N_READS&gt;</div>
<div class="foldopen" id="foldopen00101" data-start="{" data-end="}">
<div class="line"><a id="l00101" name="l00101"></a><span class="lineno"><a class="line" href="kernels_2softmax_8h.html#a8c47b0924ebfeebcca25f3dd17373276"> 101</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="kernels_2softmax_8h.html#a8c47b0924ebfeebcca25f3dd17373276">softmax_looped</a>(</div>
<div class="line"><a id="l00102" name="l00102"></a><span class="lineno"> 102</span> <span class="keyword">const</span> device T* in,</div>
<div class="line"><a id="l00103" name="l00103"></a><span class="lineno"> 103</span> device T* out,</div>
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> constant <span class="keywordtype">int</span>&amp; axis_size,</div>
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> uint gid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> uint lid [[thread_position_in_threadgroup]],</div>
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> uint lsize [[threads_per_threadgroup]],</div>
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
<div class="line"><a id="l00109" name="l00109"></a><span class="lineno"> 109</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
<div class="line"><a id="l00110" name="l00110"></a><span class="lineno"> 110</span> in += gid * size_t(axis_size);</div>
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> </div>
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a> = 32;</div>
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> </div>
<div class="line"><a id="l00114" name="l00114"></a><span class="lineno"> 114</span> threadgroup AccT local_max[<a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>];</div>
<div class="line"><a id="l00115" name="l00115"></a><span class="lineno"> 115</span> threadgroup AccT local_normalizer[<a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>];</div>
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> </div>
<div class="line"><a id="l00117" name="l00117"></a><span class="lineno"> 117</span> <span class="comment">// Get the max and the normalizer in one go</span></div>
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span> AccT prevmax;</div>
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> AccT maxval = <a class="code hl_variable" href="struct_limits.html#ae7469d21f2688797ca3e388d919ef05e">Limits&lt;AccT&gt;::finite_min</a>;</div>
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> AccT normalizer = 0;</div>
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> r = 0; r &lt; static_cast&lt;int&gt;(<a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8e5a4b0fb5d018d7b078d147efe4f1e3">ceildiv</a>(axis_size, N_READS * lsize));</div>
<div class="line"><a id="l00122" name="l00122"></a><span class="lineno"> 122</span> r++) {</div>
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span> <span class="keywordtype">int</span> offset = r * lsize * N_READS + lid * N_READS;</div>
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> AccT vals[N_READS];</div>
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> <span class="keywordflow">if</span> (offset + N_READS &lt;= axis_size) {</div>
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00127" name="l00127"></a><span class="lineno"> 127</span> vals[i] = AccT(in[offset + i]);</div>
<div class="line"><a id="l00128" name="l00128"></a><span class="lineno"> 128</span> }</div>
<div class="line"><a id="l00129" name="l00129"></a><span class="lineno"> 129</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00130" name="l00130"></a><span class="lineno"> 130</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00131" name="l00131"></a><span class="lineno"> 131</span> vals[i] = (offset + i &lt; axis_size) ? AccT(in[offset + i])</div>
<div class="line"><a id="l00132" name="l00132"></a><span class="lineno"> 132</span> : <a class="code hl_variable" href="struct_limits.html#ae7469d21f2688797ca3e388d919ef05e">Limits&lt;AccT&gt;::finite_min</a>;</div>
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> }</div>
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> }</div>
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> prevmax = maxval;</div>
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00137" name="l00137"></a><span class="lineno"> 137</span> maxval = (maxval &lt; vals[i]) ? vals[i] : maxval;</div>
<div class="line"><a id="l00138" name="l00138"></a><span class="lineno"> 138</span> }</div>
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span> normalizer *= <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(prevmax - maxval);</div>
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span> normalizer += <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(vals[i] - maxval);</div>
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"> 142</span> }</div>
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> }</div>
<div class="line"><a id="l00144" name="l00144"></a><span class="lineno"> 144</span> <span class="comment">// Now we got partial normalizer of N_READS * ceildiv(axis_size, N_READS *</span></div>
<div class="line"><a id="l00145" name="l00145"></a><span class="lineno"> 145</span> <span class="comment">// lsize) parts. We need to combine them.</span></div>
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> <span class="comment">// 1. We start by finding the max across simd groups</span></div>
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> <span class="comment">// 2. We then change the partial normalizers to account for a possible</span></div>
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> <span class="comment">// change in max</span></div>
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> <span class="comment">// 3. We sum all normalizers</span></div>
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> prevmax = maxval;</div>
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> maxval = simd_max(maxval);</div>
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> normalizer *= <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(prevmax - maxval);</div>
<div class="line"><a id="l00153" name="l00153"></a><span class="lineno"> 153</span> normalizer = simd_sum(normalizer);</div>
<div class="line"><a id="l00154" name="l00154"></a><span class="lineno"> 154</span> </div>
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> <span class="comment">// Now the normalizer and max value is correct for each simdgroup. We write</span></div>
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> <span class="comment">// them shared memory and combine them.</span></div>
<div class="line"><a id="l00157" name="l00157"></a><span class="lineno"> 157</span> prevmax = maxval;</div>
<div class="line"><a id="l00158" name="l00158"></a><span class="lineno"> 158</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> local_max[simd_group_id] = maxval;</div>
<div class="line"><a id="l00160" name="l00160"></a><span class="lineno"> 160</span> }</div>
<div class="line"><a id="l00161" name="l00161"></a><span class="lineno"> 161</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span> maxval = simd_max(local_max[simd_lane_id]);</div>
<div class="line"><a id="l00163" name="l00163"></a><span class="lineno"> 163</span> normalizer *= <a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(prevmax - maxval);</div>
<div class="line"><a id="l00164" name="l00164"></a><span class="lineno"> 164</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span> local_normalizer[simd_group_id] = normalizer;</div>
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"> 166</span> }</div>
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> normalizer = simd_sum(local_normalizer[simd_lane_id]);</div>
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> normalizer = 1 / normalizer;</div>
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> </div>
<div class="line"><a id="l00171" name="l00171"></a><span class="lineno"> 171</span> <span class="comment">// Finally given the normalizer and max value we can directly write the</span></div>
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> <span class="comment">// softmax output</span></div>
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> out += gid * size_t(axis_size);</div>
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> r = 0; r &lt; static_cast&lt;int&gt;(<a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8e5a4b0fb5d018d7b078d147efe4f1e3">ceildiv</a>(axis_size, N_READS * lsize));</div>
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> r++) {</div>
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> <span class="keywordtype">int</span> offset = r * lsize * N_READS + lid * N_READS;</div>
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> <span class="keywordflow">if</span> (offset + N_READS &lt;= axis_size) {</div>
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> out[offset + i] = T(<a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(in[offset + i] - maxval) * normalizer);</div>
<div class="line"><a id="l00180" name="l00180"></a><span class="lineno"> 180</span> }</div>
<div class="line"><a id="l00181" name="l00181"></a><span class="lineno"> 181</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00182" name="l00182"></a><span class="lineno"> 182</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N_READS; i++) {</div>
<div class="line"><a id="l00183" name="l00183"></a><span class="lineno"> 183</span> <span class="keywordflow">if</span> (offset + i &lt; axis_size) {</div>
<div class="line"><a id="l00184" name="l00184"></a><span class="lineno"> 184</span> out[offset + i] =</div>
<div class="line"><a id="l00185" name="l00185"></a><span class="lineno"> 185</span> T(<a class="code hl_function" href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a>(in[offset + i] - maxval) * normalizer);</div>
<div class="line"><a id="l00186" name="l00186"></a><span class="lineno"> 186</span> }</div>
<div class="line"><a id="l00187" name="l00187"></a><span class="lineno"> 187</span> }</div>
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span> }</div>
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> }</div>
<div class="line"><a id="l00190" name="l00190"></a><span class="lineno"> 190</span>}</div>
</div>
<div class="ttc" id="abackend_2metal_2kernels_2utils_8h_html_a8e5a4b0fb5d018d7b078d147efe4f1e3"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a8e5a4b0fb5d018d7b078d147efe4f1e3">ceildiv</a></div><div class="ttdeci">T ceildiv(T N, U M)</div><div class="ttdoc">Compute ceil((float)N/(float)M)</div><div class="ttdef"><b>Definition</b> utils.h:302</div></div>
<div class="ttc" id="akernels_2softmax_8h_html_a440d4031ee5e86159a4dd715e44a438b"><div class="ttname"><a href="kernels_2softmax_8h.html#a440d4031ee5e86159a4dd715e44a438b">softmax_exp</a></div><div class="ttdeci">T softmax_exp(T x)</div><div class="ttdef"><b>Definition</b> softmax.h:4</div></div>
<div class="ttc" id="akernels_2softmax_8h_html_a815fe70f879f318e5d6e99acf043f52b"><div class="ttname"><a href="kernels_2softmax_8h.html#a815fe70f879f318e5d6e99acf043f52b">softmax_single_row</a></div><div class="ttdeci">void softmax_single_row(const device T *in, device T *out, constant int &amp;axis_size, uint gid, uint _lid, uint simd_lane_id, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> softmax.h:11</div></div>
<div class="ttc" id="akernels_2softmax_8h_html_a8c47b0924ebfeebcca25f3dd17373276"><div class="ttname"><a href="kernels_2softmax_8h.html#a8c47b0924ebfeebcca25f3dd17373276">softmax_looped</a></div><div class="ttdeci">void softmax_looped(const device T *in, device T *out, constant int &amp;axis_size, uint gid, uint lid, uint lsize, uint simd_lane_id, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> softmax.h:101</div></div>
<div class="ttc" id="aquantized_8h_html_a62969a218d93680f5e35d0c61b160b99"><div class="ttname"><a href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a></div><div class="ttdeci">static constant constexpr const int SIMD_SIZE</div><div class="ttdef"><b>Definition</b> quantized.h:10</div></div>
<div class="ttc" id="astruct_limits_html_a6e81584ba65a4dc6ff9366b458e3a20e"><div class="ttname"><a href="struct_limits.html#a6e81584ba65a4dc6ff9366b458e3a20e">Limits::min</a></div><div class="ttdeci">static const constant U min</div><div class="ttdef"><b>Definition</b> utils.h:25</div></div>
<div class="ttc" id="astruct_limits_html_ae7469d21f2688797ca3e388d919ef05e"><div class="ttname"><a href="struct_limits.html#ae7469d21f2688797ca3e388d919ef05e">Limits::finite_min</a></div><div class="ttdeci">static const constant U finite_min</div><div class="ttdef"><b>Definition</b> utils.h:27</div></div>
</div><!-- fragment --></div><!-- contents -->
</div><!-- doc-content -->
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<li class="navelem"><a class="el" href="dir_938ab0ecf10b8b860ff766c820f665fd.html">mlx</a></li><li class="navelem"><a class="el" href="dir_1d446c9bd3c99228254c9484e0bc5c06.html">backend</a></li><li class="navelem"><a class="el" href="dir_d0c977ea65824390717cdb7efc36c157.html">metal</a></li><li class="navelem"><a class="el" href="dir_70a37effa88bcbd6b791977fa1e64356.html">kernels</a></li><li class="navelem"><a class="el" href="kernels_2softmax_8h.html">softmax.h</a></li>
<li class="footer">Generated by <a href="https://www.doxygen.org/index.html"><img class="footer" src="doxygen.svg" width="104" height="31" alt="doxygen"/></a> 1.13.2 </li>
</ul>
</div>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink