zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-11-04 18:48:15 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity
Files
721548a7996c1f68541092ddf6593cd447ccf1d7
mlx/docs/build/html/quantized_8h_source.html
CircleCI Docs 721548a799 rebase
2025-10-17 19:14:58 +00:00

2118 lines
320 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 1.12.0"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>MLX: mlx/backend/metal/kernels/quantized.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="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>
<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>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.12.0 -->
<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>
<script type="text/javascript" src="menudata.js"></script>
<script type="text/javascript" src="menu.js"></script>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function() {
initMenu('',true,false,'search.php','Search',false);
$(function() { init_search(); });
});
/* @license-end */
</script>
<div id="main-nav"></div>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function(){ initResizable(false); });
/* @license-end */
</script>
<!-- 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 id="nav-path" class="navpath">
<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> </ul>
</div>
</div><!-- top -->
<div id="doc-content">
<div class="header">
<div class="headertitle"><div class="title">quantized.h</div></div>
</div><!--header-->
<div class="contents">
<a href="quantized_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="preprocessor">#include &lt;metal_simdgroup&gt;</span></div>
<div class="line"><a id="l00004" name="l00004"></a><span class="lineno"> 4</span><span class="preprocessor">#include &lt;metal_stdlib&gt;</span></div>
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span> </div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span><span class="keyword">using namespace </span><a class="code hl_namespace" href="namespacemetal.html">metal</a>;</div>
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span> </div>
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"><a class="line" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd"> 8</a></span><span class="preprocessor">#define MLX_MTL_CONST static constant constexpr const</span></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"><a class="line" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99"> 10</a></span><a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <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="l00011" name="l00011"></a><span class="lineno"><a class="line" href="quantized_8h.html#a803e4d5a1459844ba647aea5b004e133"> 11</a></span><a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <span class="keywordtype">int</span> <a class="code hl_variable" href="quantized_8h.html#a803e4d5a1459844ba647aea5b004e133">QUAD_SIZE</a> = 4;</div>
<div class="line"><a id="l00012" name="l00012"></a><span class="lineno"> 12</span> </div>
<div class="line"><a id="l00013" name="l00013"></a><span class="lineno"> 13</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> U, <span class="keywordtype">int</span> values_per_thread, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00014" data-start="{" data-end="}">
<div class="line"><a id="l00014" name="l00014"></a><span class="lineno"><a class="line" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9"> 14</a></span><span class="keyword">inline</span> U <a class="code hl_function" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector</a>(<span class="keyword">const</span> device T* x, thread U* x_thread) {</div>
<div class="line"><a id="l00015" name="l00015"></a><span class="lineno"> 15</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00016" name="l00016"></a><span class="lineno"> 16</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00017" name="l00017"></a><span class="lineno"> 17</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00018" name="l00018"></a><span class="lineno"> 18</span> </div>
<div class="line"><a id="l00019" name="l00019"></a><span class="lineno"> 19</span> U sum = 0;</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="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00022" name="l00022"></a><span class="lineno"> 22</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_thread; i += 4) {</div>
<div class="line"><a id="l00023" name="l00023"></a><span class="lineno"> 23</span> sum += x[i] + x[i + 1] + x[i + 2] + x[i + 3];</div>
<div class="line"><a id="l00024" name="l00024"></a><span class="lineno"> 24</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> x_thread[i + 1] = x[i + 1] / 4.0f;</div>
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> x_thread[i + 2] = x[i + 2] / 16.0f;</div>
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> x_thread[i + 3] = x[i + 3] / 64.0f;</div>
<div class="line"><a id="l00028" name="l00028"></a><span class="lineno"> 28</span> }</div>
<div class="line"><a id="l00029" name="l00029"></a><span class="lineno"> 29</span> }</div>
<div class="line"><a id="l00030" name="l00030"></a><span class="lineno"> 30</span> </div>
<div class="line"><a id="l00031" name="l00031"></a><span class="lineno"> 31</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00032" name="l00032"></a><span class="lineno"> 32</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_thread; i += 4) {</div>
<div class="line"><a id="l00033" name="l00033"></a><span class="lineno"> 33</span> sum += x[i] + x[i + 1] + x[i + 2] + x[i + 3];</div>
<div class="line"><a id="l00034" name="l00034"></a><span class="lineno"> 34</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span> x_thread[i + 1] = x[i + 1] / 16.0f;</div>
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> x_thread[i + 2] = x[i + 2] / 256.0f;</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> x_thread[i + 3] = x[i + 3] / 4096.0f;</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> }</div>
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> </div>
<div class="line"><a id="l00041" name="l00041"></a><span class="lineno"> 41</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00042" name="l00042"></a><span class="lineno"> 42</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00043" name="l00043"></a><span class="lineno"> 43</span> sum += x[i];</div>
<div class="line"><a id="l00044" name="l00044"></a><span class="lineno"> 44</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00045" name="l00045"></a><span class="lineno"> 45</span> }</div>
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> }</div>
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> </div>
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> <span class="keywordflow">return</span> sum;</div>
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span>}</div>
</div>
<div class="line"><a id="l00050" name="l00050"></a><span class="lineno"> 50</span> </div>
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> U, <span class="keywordtype">int</span> values_per_thread, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00052" data-start="{" data-end="}">
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"><a class="line" href="quantized_8h.html#aa69e143d646fad332c1a53e8c9b337b7"> 52</a></span><span class="keyword">inline</span> U <a class="code hl_function" href="quantized_8h.html#aa69e143d646fad332c1a53e8c9b337b7">load_vector_safe</a>(<span class="keyword">const</span> device T* x, thread U* x_thread, <span class="keywordtype">int</span> N) {</div>
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00056" name="l00056"></a><span class="lineno"> 56</span> </div>
<div class="line"><a id="l00057" name="l00057"></a><span class="lineno"> 57</span> U sum = 0;</div>
<div class="line"><a id="l00058" name="l00058"></a><span class="lineno"> 58</span> </div>
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span> <span class="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N; i += 4) {</div>
<div class="line"><a id="l00061" name="l00061"></a><span class="lineno"> 61</span> sum += x[i] + x[i + 1] + x[i + 2] + x[i + 3];</div>
<div class="line"><a id="l00062" name="l00062"></a><span class="lineno"> 62</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00063" name="l00063"></a><span class="lineno"> 63</span> x_thread[i + 1] = x[i + 1] / 4.0f;</div>
<div class="line"><a id="l00064" name="l00064"></a><span class="lineno"> 64</span> x_thread[i + 2] = x[i + 2] / 16.0f;</div>
<div class="line"><a id="l00065" name="l00065"></a><span class="lineno"> 65</span> x_thread[i + 3] = x[i + 3] / 64.0f;</div>
<div class="line"><a id="l00066" name="l00066"></a><span class="lineno"> 66</span> }</div>
<div class="line"><a id="l00067" name="l00067"></a><span class="lineno"> 67</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = N; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00068" name="l00068"></a><span class="lineno"> 68</span> x_thread[i] = 0;</div>
<div class="line"><a id="l00069" name="l00069"></a><span class="lineno"> 69</span> }</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> </div>
<div class="line"><a id="l00072" name="l00072"></a><span class="lineno"> 72</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00073" name="l00073"></a><span class="lineno"> 73</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N; i += 4) {</div>
<div class="line"><a id="l00074" name="l00074"></a><span class="lineno"> 74</span> sum += x[i] + x[i + 1] + x[i + 2] + x[i + 3];</div>
<div class="line"><a id="l00075" name="l00075"></a><span class="lineno"> 75</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00076" name="l00076"></a><span class="lineno"> 76</span> x_thread[i + 1] = x[i + 1] / 16.0f;</div>
<div class="line"><a id="l00077" name="l00077"></a><span class="lineno"> 77</span> x_thread[i + 2] = x[i + 2] / 256.0f;</div>
<div class="line"><a id="l00078" name="l00078"></a><span class="lineno"> 78</span> x_thread[i + 3] = x[i + 3] / 4096.0f;</div>
<div class="line"><a id="l00079" name="l00079"></a><span class="lineno"> 79</span> }</div>
<div class="line"><a id="l00080" name="l00080"></a><span class="lineno"> 80</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = N; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00081" name="l00081"></a><span class="lineno"> 81</span> x_thread[i] = 0;</div>
<div class="line"><a id="l00082" name="l00082"></a><span class="lineno"> 82</span> }</div>
<div class="line"><a id="l00083" name="l00083"></a><span class="lineno"> 83</span> }</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="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00086" name="l00086"></a><span class="lineno"> 86</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N; i++) {</div>
<div class="line"><a id="l00087" name="l00087"></a><span class="lineno"> 87</span> sum += x[i];</div>
<div class="line"><a id="l00088" name="l00088"></a><span class="lineno"> 88</span> x_thread[i] = x[i];</div>
<div class="line"><a id="l00089" name="l00089"></a><span class="lineno"> 89</span> }</div>
<div class="line"><a id="l00090" name="l00090"></a><span class="lineno"> 90</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = N; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00091" name="l00091"></a><span class="lineno"> 91</span> x_thread[i] = 0;</div>
<div class="line"><a id="l00092" name="l00092"></a><span class="lineno"> 92</span> }</div>
<div class="line"><a id="l00093" name="l00093"></a><span class="lineno"> 93</span> }</div>
<div class="line"><a id="l00094" name="l00094"></a><span class="lineno"> 94</span> </div>
<div class="line"><a id="l00095" name="l00095"></a><span class="lineno"> 95</span> <span class="keywordflow">return</span> sum;</div>
<div class="line"><a id="l00096" name="l00096"></a><span class="lineno"> 96</span>}</div>
</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><span class="keyword">template</span> &lt;<span class="keyword">typename</span> U, <span class="keywordtype">int</span> values_per_thread, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00099" data-start="{" data-end="}">
<div class="line"><a id="l00099" name="l00099"></a><span class="lineno"><a class="line" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071"> 99</a></span><span class="keyword">inline</span> U <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot</a>(</div>
<div class="line"><a id="l00100" name="l00100"></a><span class="lineno"> 100</span> <span class="keyword">const</span> device uint8_t* w,</div>
<div class="line"><a id="l00101" name="l00101"></a><span class="lineno"> 101</span> <span class="keyword">const</span> thread U* x_thread,</div>
<div class="line"><a id="l00102" name="l00102"></a><span class="lineno"> 102</span> U scale,</div>
<div class="line"><a id="l00103" name="l00103"></a><span class="lineno"> 103</span> U bias,</div>
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> U sum) {</div>
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> </div>
<div class="line"><a id="l00109" name="l00109"></a><span class="lineno"> 109</span> U accum = 0;</div>
<div class="line"><a id="l00110" name="l00110"></a><span class="lineno"> 110</span> </div>
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> <span class="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (values_per_thread / 4); i++) {</div>
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> accum +=</div>
<div class="line"><a id="l00114" name="l00114"></a><span class="lineno"> 114</span> (x_thread[4 * i] * (w[i] &amp; 0x03) +</div>
<div class="line"><a id="l00115" name="l00115"></a><span class="lineno"> 115</span> x_thread[4 * i + 1] * (w[i] &amp; 0x0c) +</div>
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> x_thread[4 * i + 2] * (w[i] &amp; 0x30) +</div>
<div class="line"><a id="l00117" name="l00117"></a><span class="lineno"> 117</span> x_thread[4 * i + 3] * (w[i] &amp; 0xc0));</div>
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span> }</div>
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> }</div>
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> </div>
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00122" name="l00122"></a><span class="lineno"> 122</span> <span class="keyword">const</span> device uint16_t* ws = (<span class="keyword">const</span> device uint16_t*)w;</div>
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (values_per_thread / 4); i++) {</div>
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> accum +=</div>
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> (x_thread[4 * i] * (ws[i] &amp; 0x000f) +</div>
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> x_thread[4 * i + 1] * (ws[i] &amp; 0x00f0) +</div>
<div class="line"><a id="l00127" name="l00127"></a><span class="lineno"> 127</span> x_thread[4 * i + 2] * (ws[i] &amp; 0x0f00) +</div>
<div class="line"><a id="l00128" name="l00128"></a><span class="lineno"> 128</span> x_thread[4 * i + 3] * (ws[i] &amp; 0xf000));</div>
<div class="line"><a id="l00129" name="l00129"></a><span class="lineno"> 129</span> }</div>
<div class="line"><a id="l00130" name="l00130"></a><span class="lineno"> 130</span> }</div>
<div class="line"><a id="l00131" name="l00131"></a><span class="lineno"> 131</span> </div>
<div class="line"><a id="l00132" name="l00132"></a><span class="lineno"> 132</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> accum += x_thread[i] * w[i];</div>
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> }</div>
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> }</div>
<div class="line"><a id="l00137" name="l00137"></a><span class="lineno"> 137</span> </div>
<div class="line"><a id="l00138" name="l00138"></a><span class="lineno"> 138</span> <span class="keywordflow">return</span> scale * accum + sum * bias;</div>
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span>}</div>
</div>
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> </div>
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> U, <span class="keywordtype">int</span> values_per_thread, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00142" data-start="{" data-end="}">
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"><a class="line" href="quantized_8h.html#a07b26d2d0b0d65dfe925c452c453fa42"> 142</a></span><span class="keyword">inline</span> U <a class="code hl_function" href="quantized_8h.html#a07b26d2d0b0d65dfe925c452c453fa42">qdot_safe</a>(</div>
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> <span class="keyword">const</span> device uint8_t* w,</div>
<div class="line"><a id="l00144" name="l00144"></a><span class="lineno"> 144</span> <span class="keyword">const</span> thread U* x_thread,</div>
<div class="line"><a id="l00145" name="l00145"></a><span class="lineno"> 145</span> U scale,</div>
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> U bias,</div>
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> U sum,</div>
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> <span class="keywordtype">int</span> N) {</div>
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> </div>
<div class="line"><a id="l00153" name="l00153"></a><span class="lineno"> 153</span> U accum = 0;</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="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (N / 4); i++) {</div>
<div class="line"><a id="l00157" name="l00157"></a><span class="lineno"> 157</span> accum +=</div>
<div class="line"><a id="l00158" name="l00158"></a><span class="lineno"> 158</span> (x_thread[4 * i] * (w[i] &amp; 0x03) +</div>
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> x_thread[4 * i + 1] * (w[i] &amp; 0x0c) +</div>
<div class="line"><a id="l00160" name="l00160"></a><span class="lineno"> 160</span> x_thread[4 * i + 2] * (w[i] &amp; 0x30) +</div>
<div class="line"><a id="l00161" name="l00161"></a><span class="lineno"> 161</span> x_thread[4 * i + 3] * (w[i] &amp; 0xc0));</div>
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span> }</div>
<div class="line"><a id="l00163" name="l00163"></a><span class="lineno"> 163</span> }</div>
<div class="line"><a id="l00164" name="l00164"></a><span class="lineno"> 164</span> </div>
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"> 166</span> <span class="keyword">const</span> device uint16_t* ws = (<span class="keyword">const</span> device uint16_t*)w;</div>
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (N / 4); i++) {</div>
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> accum +=</div>
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> (x_thread[4 * i] * (ws[i] &amp; 0x000f) +</div>
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> x_thread[4 * i + 1] * (ws[i] &amp; 0x00f0) +</div>
<div class="line"><a id="l00171" name="l00171"></a><span class="lineno"> 171</span> x_thread[4 * i + 2] * (ws[i] &amp; 0x0f00) +</div>
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> x_thread[4 * i + 3] * (ws[i] &amp; 0xf000));</div>
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> }</div>
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> }</div>
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> </div>
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N; i++) {</div>
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> accum += x_thread[i] * w[i];</div>
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> }</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> </div>
<div class="line"><a id="l00182" name="l00182"></a><span class="lineno"> 182</span> <span class="keywordflow">return</span> scale * accum + sum * bias;</div>
<div class="line"><a id="l00183" name="l00183"></a><span class="lineno"> 183</span>}</div>
</div>
<div class="line"><a id="l00184" name="l00184"></a><span class="lineno"> 184</span> </div>
<div class="line"><a id="l00185" name="l00185"></a><span class="lineno"> 185</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> U, <span class="keywordtype">int</span> values_per_thread, <span class="keywordtype">int</span> bits&gt;</div>
<div class="line"><a id="l00186" name="l00186"></a><span class="lineno"> 186</span><span class="keyword">inline</span> <span class="keywordtype">void</span></div>
<div class="foldopen" id="foldopen00187" data-start="{" data-end="}">
<div class="line"><a id="l00187" name="l00187"></a><span class="lineno"><a class="line" href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58"> 187</a></span><a class="code hl_function" href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58">qouter</a>(<span class="keyword">const</span> thread uint8_t* w, U x, U scale, U bias, thread U* result) {</div>
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00190" name="l00190"></a><span class="lineno"> 190</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00191" name="l00191"></a><span class="lineno"> 191</span> </div>
<div class="line"><a id="l00192" name="l00192"></a><span class="lineno"> 192</span> <span class="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00193" name="l00193"></a><span class="lineno"> 193</span> U s[4] = {scale, scale / 4.0f, scale / 16.0f, scale / 64.0f};</div>
<div class="line"><a id="l00194" name="l00194"></a><span class="lineno"> 194</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (values_per_thread / 4); i++) {</div>
<div class="line"><a id="l00195" name="l00195"></a><span class="lineno"> 195</span> result[4 * i] += x * (s[0] * (w[i] &amp; 0x03) + bias);</div>
<div class="line"><a id="l00196" name="l00196"></a><span class="lineno"> 196</span> result[4 * i + 1] += x * (s[1] * (w[i] &amp; 0x0c) + bias);</div>
<div class="line"><a id="l00197" name="l00197"></a><span class="lineno"> 197</span> result[4 * i + 2] += x * (s[2] * (w[i] &amp; 0x30) + bias);</div>
<div class="line"><a id="l00198" name="l00198"></a><span class="lineno"> 198</span> result[4 * i + 3] += x * (s[3] * (w[i] &amp; 0xc0) + bias);</div>
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"> 199</span> }</div>
<div class="line"><a id="l00200" name="l00200"></a><span class="lineno"> 200</span> }</div>
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span> </div>
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> U s[2] = {scale, scale / 16.0f};</div>
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (values_per_thread / 2); i++) {</div>
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> result[2 * i] += x * (s[0] * (w[i] &amp; 0x0f) + bias);</div>
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> result[2 * i + 1] += x * (s[1] * (w[i] &amp; 0xf0) + bias);</div>
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> }</div>
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> }</div>
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> </div>
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_thread; i++) {</div>
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> result[i] += x * (scale * w[i] + bias);</div>
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> }</div>
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> }</div>
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span>}</div>
</div>
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> </div>
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> U, <span class="keywordtype">int</span> N, <span class="keywordtype">int</span> bits&gt;</div>
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span><span class="keyword">inline</span> <span class="keywordtype">void</span></div>
<div class="foldopen" id="foldopen00219" data-start="{" data-end="}">
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"><a class="line" href="quantized_8h.html#aecff265b63566d0d5689cfc4e5b037d2"> 219</a></span><a class="code hl_function" href="quantized_8h.html#aecff265b63566d0d5689cfc4e5b037d2">dequantize</a>(<span class="keyword">const</span> device uint8_t* w, U scale, U bias, threadgroup U* w_local) {</div>
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00223" name="l00223"></a><span class="lineno"> 223</span> </div>
<div class="line"><a id="l00224" name="l00224"></a><span class="lineno"> 224</span> <span class="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> U s[4] = {</div>
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> scale,</div>
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> scale / <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(4.0f),</div>
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> scale / <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(16.0f),</div>
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> scale / <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(64.0f)};</div>
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (N / 4); i++) {</div>
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> w_local[4 * i] = s[0] * (w[i] &amp; 0x03) + bias;</div>
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> w_local[4 * i + 1] = s[1] * (w[i] &amp; 0x0c) + bias;</div>
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> w_local[4 * i + 2] = s[2] * (w[i] &amp; 0x30) + bias;</div>
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> w_local[4 * i + 3] = s[3] * (w[i] &amp; 0xc0) + bias;</div>
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> }</div>
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> }</div>
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> </div>
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> U s[2] = {scale, scale / <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(16.0f)};</div>
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; (N / 2); i++) {</div>
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> w_local[2 * i] = s[0] * (w[i] &amp; 0x0f) + bias;</div>
<div class="line"><a id="l00242" name="l00242"></a><span class="lineno"> 242</span> w_local[2 * i + 1] = s[1] * (w[i] &amp; 0xf0) + bias;</div>
<div class="line"><a id="l00243" name="l00243"></a><span class="lineno"> 243</span> }</div>
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> }</div>
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> </div>
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l00247" name="l00247"></a><span class="lineno"> 247</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; N; i++) {</div>
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> w_local[i] = scale * w[i] + bias;</div>
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> }</div>
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> }</div>
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span>}</div>
</div>
<div class="line"><a id="l00252" name="l00252"></a><span class="lineno"> 252</span> </div>
<div class="line"><a id="l00253" name="l00253"></a><span class="lineno"> 253</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> <span class="keywordtype">short</span> BROWS,</div>
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> <span class="keywordtype">short</span> BCOLS,</div>
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> <span class="keywordtype">short</span> dst_ld,</div>
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> <span class="keywordtype">short</span> reduction_dim,</div>
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> <span class="keywordtype">short</span> tgp_size,</div>
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> <span class="keywordtype">short</span> group_size,</div>
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> <span class="keywordtype">short</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>&gt;</div>
<div class="foldopen" id="foldopen00262" data-start="{" data-end="};">
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html"> 262</a></span><span class="keyword">struct </span><a class="code hl_struct" href="struct_quantized_block_loader.html">QuantizedBlockLoader</a> {</div>
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> BCOLS &lt;= group_size,</div>
<div class="line"><a id="l00265" name="l00265"></a><span class="lineno"> 265</span> <span class="stringliteral">&quot;The group size should be larger than the columns&quot;</span>);</div>
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"> 266</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> group_size % BCOLS == 0,</div>
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> <span class="stringliteral">&quot;The group size should be divisible by the columns&quot;</span>);</div>
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> bits == 2 || bits == 4 || bits == 8,</div>
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span> <span class="stringliteral">&quot;Template undefined for bits not in {2, 4, 8}&quot;</span>);</div>
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> </div>
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d"> 273</a></span> <a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a> = 32 / bits;</div>
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb"> 274</a></span> <a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> = BCOLS / <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>;</div>
<div class="line"><a id="l00275" name="l00275"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb"> 275</a></span> <a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a> =</div>
<div class="line"><a id="l00276" name="l00276"></a><span class="lineno"> 276</span> (<a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> * BROWS &lt; tgp_size) ? 1 : (<a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> * BROWS) / tgp_size;</div>
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a31e14175f3d4902d9fe5ab5a219f61ba"> 277</a></span> <a class="code hl_define" href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a31e14175f3d4902d9fe5ab5a219f61ba">group_steps</a> = group_size / BCOLS;</div>
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> </div>
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e"> 279</a></span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a>;</div>
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#ac3f651c1a645291d1037a2cc8ded2320"> 280</a></span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ac3f651c1a645291d1037a2cc8ded2320">tile_stride</a>;</div>
<div class="line"><a id="l00281" name="l00281"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6"> 281</a></span> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">group_step_cnt</a>;</div>
<div class="line"><a id="l00282" name="l00282"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab"> 282</a></span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab">group_stride</a>;</div>
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> </div>
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475"> 284</a></span> <span class="keyword">const</span> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475">thread_idx</a>;</div>
<div class="line"><a id="l00285" name="l00285"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906"> 285</a></span> <span class="keyword">const</span> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a>;</div>
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00"> 286</a></span> <span class="keyword">const</span> <span class="keywordtype">short</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00">bj</a>;</div>
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> </div>
<div class="line"><a id="l00288" name="l00288"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a9857214690fe6abad0e19d1045152f83"> 288</a></span> threadgroup T* <a class="code hl_variable" href="struct_quantized_block_loader.html#a9857214690fe6abad0e19d1045152f83">dst</a>;</div>
<div class="line"><a id="l00289" name="l00289"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b"> 289</a></span> <span class="keyword">const</span> device uint32_t* <a class="code hl_variable" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">src</a>;</div>
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf"> 290</a></span> <span class="keyword">const</span> device T* <a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>;</div>
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd"> 291</a></span> <span class="keyword">const</span> device T* <a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>;</div>
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> </div>
<div class="foldopen" id="foldopen00293" data-start="{" data-end="}">
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#af59b054750a65e7e79c1cd05c4acac93"> 293</a></span> <a class="code hl_function" href="struct_quantized_block_loader.html#af59b054750a65e7e79c1cd05c4acac93">QuantizedBlockLoader</a>(</div>
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> <span class="keyword">const</span> device uint32_t* src_,</div>
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> <span class="keyword">const</span> device T* scales_,</div>
<div class="line"><a id="l00296" name="l00296"></a><span class="lineno"> 296</span> <span class="keyword">const</span> device T* biases_,</div>
<div class="line"><a id="l00297" name="l00297"></a><span class="lineno"> 297</span> <span class="keyword">const</span> <span class="keywordtype">int</span> src_ld_,</div>
<div class="line"><a id="l00298" name="l00298"></a><span class="lineno"> 298</span> threadgroup T* dst_,</div>
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> ushort simd_group_id [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> ushort simd_lane_id [[thread_index_in_simdgroup]])</div>
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> : <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a>(src_ld_),</div>
<div class="line"><a id="l00302" name="l00302"></a><span class="lineno"> 302</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ac3f651c1a645291d1037a2cc8ded2320">tile_stride</a>(</div>
<div class="line"><a id="l00303" name="l00303"></a><span class="lineno"> 303</span> reduction_dim ? <a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> : BROWS * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a> / <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>),</div>
<div class="line"><a id="l00304" name="l00304"></a><span class="lineno"> 304</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">group_step_cnt</a>(0),</div>
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab">group_stride</a>(BROWS * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a> / group_size),</div>
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475">thread_idx</a>(simd_group_id * 32 + simd_lane_id),</div>
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a>(<a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475">thread_idx</a> / <a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a>),</div>
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00">bj</a>((<a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475">thread_idx</a>) % <a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a>),</div>
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> dst(dst_ + <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> * dst_ld + <a class="code hl_variable" href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00">bj</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>),</div>
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">src</a>(src_ + <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a> / <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a> + <a class="code hl_variable" href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00">bj</a>),</div>
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>(scales_ + <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a> / group_size),</div>
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>(biases_ + <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">src_ld</a> / group_size) {}</div>
</div>
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> </div>
<div class="foldopen" id="foldopen00314" data-start="{" data-end="}">
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a86009527cb4b53e4c21fd6b1f78cfefc"> 314</a></span> <span class="keywordtype">void</span> <a class="code hl_function" href="struct_quantized_block_loader.html#a86009527cb4b53e4c21fd6b1f78cfefc">load_unsafe</a>()<span class="keyword"> const </span>{</div>
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> * BROWS &lt; tgp_size &amp;&amp; bi &gt;= BROWS) {</div>
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> }</div>
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> </div>
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> T scale = *<a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>;</div>
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> T bias = *<a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>;</div>
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; <a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a>; i++) {</div>
<div class="line"><a id="l00322" name="l00322"></a><span class="lineno"> 322</span> <a class="code hl_function" href="quantized_8h.html#aecff265b63566d0d5689cfc4e5b037d2">dequantize&lt;T, pack_factor, bits&gt;</a>(</div>
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> (device uint8_t*)(<a class="code hl_variable" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">src</a> + i), scale, bias, dst + i * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>);</div>
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> }</div>
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> }</div>
</div>
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> </div>
<div class="foldopen" id="foldopen00327" data-start="{" data-end="}">
<div class="line"><a id="l00327" name="l00327"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a699dc9aa284b8fbf870310bbb224465b"> 327</a></span> <span class="keywordtype">void</span> <a class="code hl_function" href="struct_quantized_block_loader.html#a699dc9aa284b8fbf870310bbb224465b">load_safe</a>(short2 src_tile_dim)<span class="keyword"> const </span>{</div>
<div class="line"><a id="l00328" name="l00328"></a><span class="lineno"> 328</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">BCOLS_PACKED</a> * BROWS &lt; tgp_size &amp;&amp; bi &gt;= BROWS) {</div>
<div class="line"><a id="l00329" name="l00329"></a><span class="lineno"> 329</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00330" name="l00330"></a><span class="lineno"> 330</span> }</div>
<div class="line"><a id="l00331" name="l00331"></a><span class="lineno"> 331</span> </div>
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> <span class="keywordflow">if</span> (reduction_dim == 1 &amp;&amp; <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> &gt;= src_tile_dim.y) {</div>
<div class="line"><a id="l00333" name="l00333"></a><span class="lineno"> 333</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; <a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>; i++) {</div>
<div class="line"><a id="l00334" name="l00334"></a><span class="lineno"> 334</span> dst[i] = T(0);</div>
<div class="line"><a id="l00335" name="l00335"></a><span class="lineno"> 335</span> }</div>
<div class="line"><a id="l00336" name="l00336"></a><span class="lineno"> 336</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00337" name="l00337"></a><span class="lineno"> 337</span> }</div>
<div class="line"><a id="l00338" name="l00338"></a><span class="lineno"> 338</span> </div>
<div class="line"><a id="l00339" name="l00339"></a><span class="lineno"> 339</span> <span class="keywordflow">if</span> (reduction_dim == 0 &amp;&amp; <a class="code hl_variable" href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">bi</a> &gt;= src_tile_dim.x) {</div>
<div class="line"><a id="l00340" name="l00340"></a><span class="lineno"> 340</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; <a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a> * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>; i++) {</div>
<div class="line"><a id="l00341" name="l00341"></a><span class="lineno"> 341</span> dst[i] = T(0);</div>
<div class="line"><a id="l00342" name="l00342"></a><span class="lineno"> 342</span> }</div>
<div class="line"><a id="l00343" name="l00343"></a><span class="lineno"> 343</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00344" name="l00344"></a><span class="lineno"> 344</span> }</div>
<div class="line"><a id="l00345" name="l00345"></a><span class="lineno"> 345</span> </div>
<div class="line"><a id="l00346" name="l00346"></a><span class="lineno"> 346</span> T scale = *<a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>;</div>
<div class="line"><a id="l00347" name="l00347"></a><span class="lineno"> 347</span> T bias = *<a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>;</div>
<div class="line"><a id="l00348" name="l00348"></a><span class="lineno"> 348</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; <a class="code hl_variable" href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">n_reads</a>; i++) {</div>
<div class="line"><a id="l00349" name="l00349"></a><span class="lineno"> 349</span> <a class="code hl_function" href="quantized_8h.html#aecff265b63566d0d5689cfc4e5b037d2">dequantize&lt;T, pack_factor, bits&gt;</a>(</div>
<div class="line"><a id="l00350" name="l00350"></a><span class="lineno"> 350</span> (device uint8_t*)(<a class="code hl_variable" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">src</a> + i), scale, bias, dst + i * <a class="code hl_variable" href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">pack_factor</a>);</div>
<div class="line"><a id="l00351" name="l00351"></a><span class="lineno"> 351</span> }</div>
<div class="line"><a id="l00352" name="l00352"></a><span class="lineno"> 352</span> }</div>
</div>
<div class="line"><a id="l00353" name="l00353"></a><span class="lineno"> 353</span> </div>
<div class="foldopen" id="foldopen00354" data-start="{" data-end="}">
<div class="line"><a id="l00354" name="l00354"></a><span class="lineno"><a class="line" href="struct_quantized_block_loader.html#a674138ef7c43cc45586ea9f8fd6f6bd9"> 354</a></span> <span class="keywordtype">void</span> <a class="code hl_function" href="struct_quantized_block_loader.html#a674138ef7c43cc45586ea9f8fd6f6bd9">next</a>() {</div>
<div class="line"><a id="l00355" name="l00355"></a><span class="lineno"> 355</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">src</a> += <a class="code hl_variable" href="struct_quantized_block_loader.html#ac3f651c1a645291d1037a2cc8ded2320">tile_stride</a>;</div>
<div class="line"><a id="l00356" name="l00356"></a><span class="lineno"> 356</span> <span class="keywordflow">if</span> (reduction_dim == 1) {</div>
<div class="line"><a id="l00357" name="l00357"></a><span class="lineno"> 357</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="struct_quantized_block_loader.html#a31e14175f3d4902d9fe5ab5a219f61ba">group_steps</a> &gt; 1) {</div>
<div class="line"><a id="l00358" name="l00358"></a><span class="lineno"> 358</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">group_step_cnt</a>++;</div>
<div class="line"><a id="l00359" name="l00359"></a><span class="lineno"> 359</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">group_step_cnt</a> == <a class="code hl_variable" href="struct_quantized_block_loader.html#a31e14175f3d4902d9fe5ab5a219f61ba">group_steps</a>) {</div>
<div class="line"><a id="l00360" name="l00360"></a><span class="lineno"> 360</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">group_step_cnt</a> = 0;</div>
<div class="line"><a id="l00361" name="l00361"></a><span class="lineno"> 361</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>++;</div>
<div class="line"><a id="l00362" name="l00362"></a><span class="lineno"> 362</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>++;</div>
<div class="line"><a id="l00363" name="l00363"></a><span class="lineno"> 363</span> }</div>
<div class="line"><a id="l00364" name="l00364"></a><span class="lineno"> 364</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00365" name="l00365"></a><span class="lineno"> 365</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a>++;</div>
<div class="line"><a id="l00366" name="l00366"></a><span class="lineno"> 366</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a>++;</div>
<div class="line"><a id="l00367" name="l00367"></a><span class="lineno"> 367</span> }</div>
<div class="line"><a id="l00368" name="l00368"></a><span class="lineno"> 368</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00369" name="l00369"></a><span class="lineno"> 369</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">scales</a> += <a class="code hl_variable" href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab">group_stride</a>;</div>
<div class="line"><a id="l00370" name="l00370"></a><span class="lineno"> 370</span> <a class="code hl_variable" href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">biases</a> += <a class="code hl_variable" href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab">group_stride</a>;</div>
<div class="line"><a id="l00371" name="l00371"></a><span class="lineno"> 371</span> }</div>
<div class="line"><a id="l00372" name="l00372"></a><span class="lineno"> 372</span> }</div>
</div>
<div class="line"><a id="l00373" name="l00373"></a><span class="lineno"> 373</span>};</div>
</div>
<div class="line"><a id="l00374" name="l00374"></a><span class="lineno"> 374</span> </div>
<div class="line"><a id="l00375" name="l00375"></a><span class="lineno"> 375</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits, <span class="keywordtype">int</span> D&gt;</div>
<div class="foldopen" id="foldopen00376" data-start="{" data-end="}">
<div class="line"><a id="l00376" name="l00376"></a><span class="lineno"><a class="line" href="quantized_8h.html#ad5cf1cf63656bc1780685d22169cd4ef"> 376</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#ad5cf1cf63656bc1780685d22169cd4ef">qmv_quad_impl</a>(</div>
<div class="line"><a id="l00377" name="l00377"></a><span class="lineno"> 377</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00378" name="l00378"></a><span class="lineno"> 378</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00379" name="l00379"></a><span class="lineno"> 379</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00380" name="l00380"></a><span class="lineno"> 380</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00381" name="l00381"></a><span class="lineno"> 381</span> device T* y,</div>
<div class="line"><a id="l00382" name="l00382"></a><span class="lineno"> 382</span> constant <span class="keywordtype">int</span>&amp; in_vec_size,</div>
<div class="line"><a id="l00383" name="l00383"></a><span class="lineno"> 383</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size,</div>
<div class="line"><a id="l00384" name="l00384"></a><span class="lineno"> 384</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00385" name="l00385"></a><span class="lineno"> 385</span> uint quad_gid [[quadgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00386" name="l00386"></a><span class="lineno"> 386</span> uint quad_lid [[thread_index_in_quadgroup]]) {</div>
<div class="line"><a id="l00387" name="l00387"></a><span class="lineno"> 387</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> quads_per_simd = <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a> / <a class="code hl_variable" href="quantized_8h.html#a803e4d5a1459844ba647aea5b004e133">QUAD_SIZE</a>;</div>
<div class="line"><a id="l00388" name="l00388"></a><span class="lineno"> 388</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00389" name="l00389"></a><span class="lineno"> 389</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> values_per_thread = D / <a class="code hl_variable" href="quantized_8h.html#a803e4d5a1459844ba647aea5b004e133">QUAD_SIZE</a>;</div>
<div class="line"><a id="l00390" name="l00390"></a><span class="lineno"> 390</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_thread = values_per_thread / pack_factor;</div>
<div class="line"><a id="l00391" name="l00391"></a><span class="lineno"> 391</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> scale_step_per_thread = group_size / values_per_thread;</div>
<div class="line"><a id="l00392" name="l00392"></a><span class="lineno"> 392</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> results_per_quadgroup = 8;</div>
<div class="line"><a id="l00393" name="l00393"></a><span class="lineno"> 393</span> </div>
<div class="line"><a id="l00394" name="l00394"></a><span class="lineno"> 394</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00395" name="l00395"></a><span class="lineno"> 395</span> </div>
<div class="line"><a id="l00396" name="l00396"></a><span class="lineno"> 396</span> thread U x_thread[values_per_thread];</div>
<div class="line"><a id="l00397" name="l00397"></a><span class="lineno"> 397</span> thread U result[results_per_quadgroup] = {0};</div>
<div class="line"><a id="l00398" name="l00398"></a><span class="lineno"> 398</span> </div>
<div class="line"><a id="l00399" name="l00399"></a><span class="lineno"> 399</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00400" name="l00400"></a><span class="lineno"> 400</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_w = in_vec_size / pack_factor;</div>
<div class="line"><a id="l00401" name="l00401"></a><span class="lineno"> 401</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_g = in_vec_size / group_size;</div>
<div class="line"><a id="l00402" name="l00402"></a><span class="lineno"> 402</span> <span class="keyword">const</span> <span class="keywordtype">int</span> out_row = tid.x * quads_per_simd * results_per_quadgroup + quad_gid;</div>
<div class="line"><a id="l00403" name="l00403"></a><span class="lineno"> 403</span> </div>
<div class="line"><a id="l00404" name="l00404"></a><span class="lineno"> 404</span> w += out_row * in_vec_size_w + quad_lid * packs_per_thread;</div>
<div class="line"><a id="l00405" name="l00405"></a><span class="lineno"> 405</span> scales += out_row * in_vec_size_g + quad_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00406" name="l00406"></a><span class="lineno"> 406</span> biases += out_row * in_vec_size_g + quad_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00407" name="l00407"></a><span class="lineno"> 407</span> x += tid.y * in_vec_size + quad_lid * values_per_thread;</div>
<div class="line"><a id="l00408" name="l00408"></a><span class="lineno"> 408</span> y += tid.y * out_vec_size + out_row;</div>
<div class="line"><a id="l00409" name="l00409"></a><span class="lineno"> 409</span> </div>
<div class="line"><a id="l00410" name="l00410"></a><span class="lineno"> 410</span> U sum = <a class="code hl_function" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread);</div>
<div class="line"><a id="l00411" name="l00411"></a><span class="lineno"> 411</span> </div>
<div class="line"><a id="l00412" name="l00412"></a><span class="lineno"> 412</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_quadgroup; row++) {</div>
<div class="line"><a id="l00413" name="l00413"></a><span class="lineno"> 413</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00414" name="l00414"></a><span class="lineno"> 414</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w * quads_per_simd);</div>
<div class="line"><a id="l00415" name="l00415"></a><span class="lineno"> 415</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g * quads_per_simd;</div>
<div class="line"><a id="l00416" name="l00416"></a><span class="lineno"> 416</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g * quads_per_simd;</div>
<div class="line"><a id="l00417" name="l00417"></a><span class="lineno"> 417</span> </div>
<div class="line"><a id="l00418" name="l00418"></a><span class="lineno"> 418</span> U s = sl[0];</div>
<div class="line"><a id="l00419" name="l00419"></a><span class="lineno"> 419</span> U b = bl[0];</div>
<div class="line"><a id="l00420" name="l00420"></a><span class="lineno"> 420</span> <span class="keywordflow">if</span> (row * quads_per_simd + out_row &lt; out_vec_size) {</div>
<div class="line"><a id="l00421" name="l00421"></a><span class="lineno"> 421</span> result[row] += <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot&lt;U, values_per_thread, bits&gt;</a>(wl, x_thread, s, b, sum);</div>
<div class="line"><a id="l00422" name="l00422"></a><span class="lineno"> 422</span> }</div>
<div class="line"><a id="l00423" name="l00423"></a><span class="lineno"> 423</span> }</div>
<div class="line"><a id="l00424" name="l00424"></a><span class="lineno"> 424</span> </div>
<div class="line"><a id="l00425" name="l00425"></a><span class="lineno"> 425</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_quadgroup; row++) {</div>
<div class="line"><a id="l00426" name="l00426"></a><span class="lineno"> 426</span> result[row] = quad_sum(result[row]);</div>
<div class="line"><a id="l00427" name="l00427"></a><span class="lineno"> 427</span> <span class="keywordflow">if</span> (quad_lid == 0 &amp;&amp; row * quads_per_simd + out_row &lt; out_vec_size) {</div>
<div class="line"><a id="l00428" name="l00428"></a><span class="lineno"> 428</span> y[row * quads_per_simd] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(result[row]);</div>
<div class="line"><a id="l00429" name="l00429"></a><span class="lineno"> 429</span> }</div>
<div class="line"><a id="l00430" name="l00430"></a><span class="lineno"> 430</span> }</div>
<div class="line"><a id="l00431" name="l00431"></a><span class="lineno"> 431</span>}</div>
</div>
<div class="line"><a id="l00432" name="l00432"></a><span class="lineno"> 432</span> </div>
<div class="line"><a id="l00433" name="l00433"></a><span class="lineno"> 433</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00434" data-start="{" data-end="}">
<div class="line"><a id="l00434" name="l00434"></a><span class="lineno"><a class="line" href="quantized_8h.html#aba7687e6f8f1d29c0a1b2a3db150bd81"> 434</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#aba7687e6f8f1d29c0a1b2a3db150bd81">qmv_fast_impl</a>(</div>
<div class="line"><a id="l00435" name="l00435"></a><span class="lineno"> 435</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00436" name="l00436"></a><span class="lineno"> 436</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00437" name="l00437"></a><span class="lineno"> 437</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00438" name="l00438"></a><span class="lineno"> 438</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00439" name="l00439"></a><span class="lineno"> 439</span> device T* y,</div>
<div class="line"><a id="l00440" name="l00440"></a><span class="lineno"> 440</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size,</div>
<div class="line"><a id="l00441" name="l00441"></a><span class="lineno"> 441</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size,</div>
<div class="line"><a id="l00442" name="l00442"></a><span class="lineno"> 442</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00443" name="l00443"></a><span class="lineno"> 443</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00444" name="l00444"></a><span class="lineno"> 444</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00445" name="l00445"></a><span class="lineno"> 445</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_thread = bits &gt; 2 ? 2 : 1;</div>
<div class="line"><a id="l00446" name="l00446"></a><span class="lineno"> 446</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> num_simdgroups = 2;</div>
<div class="line"><a id="l00447" name="l00447"></a><span class="lineno"> 447</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> results_per_simdgroup = 4;</div>
<div class="line"><a id="l00448" name="l00448"></a><span class="lineno"> 448</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00449" name="l00449"></a><span class="lineno"> 449</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> values_per_thread = pack_factor * packs_per_thread;</div>
<div class="line"><a id="l00450" name="l00450"></a><span class="lineno"> 450</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> block_size = values_per_thread * <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>;</div>
<div class="line"><a id="l00451" name="l00451"></a><span class="lineno"> 451</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> scale_step_per_thread = group_size / values_per_thread;</div>
<div class="line"><a id="l00452" name="l00452"></a><span class="lineno"> 452</span> </div>
<div class="line"><a id="l00453" name="l00453"></a><span class="lineno"> 453</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00454" name="l00454"></a><span class="lineno"> 454</span> </div>
<div class="line"><a id="l00455" name="l00455"></a><span class="lineno"> 455</span> thread U x_thread[values_per_thread];</div>
<div class="line"><a id="l00456" name="l00456"></a><span class="lineno"> 456</span> thread U result[results_per_simdgroup] = {0};</div>
<div class="line"><a id="l00457" name="l00457"></a><span class="lineno"> 457</span> </div>
<div class="line"><a id="l00458" name="l00458"></a><span class="lineno"> 458</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00459" name="l00459"></a><span class="lineno"> 459</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_w = in_vec_size / pack_factor;</div>
<div class="line"><a id="l00460" name="l00460"></a><span class="lineno"> 460</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_g = in_vec_size / group_size;</div>
<div class="line"><a id="l00461" name="l00461"></a><span class="lineno"> 461</span> <span class="keyword">const</span> <span class="keywordtype">int</span> out_row = tid.x * (num_simdgroups * results_per_simdgroup) +</div>
<div class="line"><a id="l00462" name="l00462"></a><span class="lineno"> 462</span> simd_gid * results_per_simdgroup;</div>
<div class="line"><a id="l00463" name="l00463"></a><span class="lineno"> 463</span> w += out_row * in_vec_size_w + simd_lid * packs_per_thread;</div>
<div class="line"><a id="l00464" name="l00464"></a><span class="lineno"> 464</span> scales += out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00465" name="l00465"></a><span class="lineno"> 465</span> biases += out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00466" name="l00466"></a><span class="lineno"> 466</span> x += tid.y * in_vec_size + simd_lid * values_per_thread;</div>
<div class="line"><a id="l00467" name="l00467"></a><span class="lineno"> 467</span> y += tid.y * out_vec_size + out_row;</div>
<div class="line"><a id="l00468" name="l00468"></a><span class="lineno"> 468</span> </div>
<div class="line"><a id="l00469" name="l00469"></a><span class="lineno"> 469</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; in_vec_size; k += block_size) {</div>
<div class="line"><a id="l00470" name="l00470"></a><span class="lineno"> 470</span> U sum = <a class="code hl_function" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread);</div>
<div class="line"><a id="l00471" name="l00471"></a><span class="lineno"> 471</span> </div>
<div class="line"><a id="l00472" name="l00472"></a><span class="lineno"> 472</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_simdgroup; row++) {</div>
<div class="line"><a id="l00473" name="l00473"></a><span class="lineno"> 473</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00474" name="l00474"></a><span class="lineno"> 474</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w);</div>
<div class="line"><a id="l00475" name="l00475"></a><span class="lineno"> 475</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g;</div>
<div class="line"><a id="l00476" name="l00476"></a><span class="lineno"> 476</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g;</div>
<div class="line"><a id="l00477" name="l00477"></a><span class="lineno"> 477</span> </div>
<div class="line"><a id="l00478" name="l00478"></a><span class="lineno"> 478</span> U s = sl[0];</div>
<div class="line"><a id="l00479" name="l00479"></a><span class="lineno"> 479</span> U b = bl[0];</div>
<div class="line"><a id="l00480" name="l00480"></a><span class="lineno"> 480</span> result[row] += <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot&lt;U, values_per_thread, bits&gt;</a>(wl, x_thread, s, b, sum);</div>
<div class="line"><a id="l00481" name="l00481"></a><span class="lineno"> 481</span> }</div>
<div class="line"><a id="l00482" name="l00482"></a><span class="lineno"> 482</span> </div>
<div class="line"><a id="l00483" name="l00483"></a><span class="lineno"> 483</span> w += block_size / pack_factor;</div>
<div class="line"><a id="l00484" name="l00484"></a><span class="lineno"> 484</span> scales += block_size / group_size;</div>
<div class="line"><a id="l00485" name="l00485"></a><span class="lineno"> 485</span> biases += block_size / group_size;</div>
<div class="line"><a id="l00486" name="l00486"></a><span class="lineno"> 486</span> x += block_size;</div>
<div class="line"><a id="l00487" name="l00487"></a><span class="lineno"> 487</span> }</div>
<div class="line"><a id="l00488" name="l00488"></a><span class="lineno"> 488</span> </div>
<div class="line"><a id="l00489" name="l00489"></a><span class="lineno"> 489</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_simdgroup; row++) {</div>
<div class="line"><a id="l00490" name="l00490"></a><span class="lineno"> 490</span> result[row] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(result[row]);</div>
<div class="line"><a id="l00491" name="l00491"></a><span class="lineno"> 491</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00492" name="l00492"></a><span class="lineno"> 492</span> y[row] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(result[row]);</div>
<div class="line"><a id="l00493" name="l00493"></a><span class="lineno"> 493</span> }</div>
<div class="line"><a id="l00494" name="l00494"></a><span class="lineno"> 494</span> }</div>
<div class="line"><a id="l00495" name="l00495"></a><span class="lineno"> 495</span>}</div>
</div>
<div class="line"><a id="l00496" name="l00496"></a><span class="lineno"> 496</span> </div>
<div class="line"><a id="l00497" name="l00497"></a><span class="lineno"> 497</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00498" data-start="{" data-end="}">
<div class="line"><a id="l00498" name="l00498"></a><span class="lineno"><a class="line" href="quantized_8h.html#a8e13c7d895624f738d2a6d9893b687fd"> 498</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a8e13c7d895624f738d2a6d9893b687fd">qmv_impl</a>(</div>
<div class="line"><a id="l00499" name="l00499"></a><span class="lineno"> 499</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00500" name="l00500"></a><span class="lineno"> 500</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00501" name="l00501"></a><span class="lineno"> 501</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00502" name="l00502"></a><span class="lineno"> 502</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00503" name="l00503"></a><span class="lineno"> 503</span> device T* y,</div>
<div class="line"><a id="l00504" name="l00504"></a><span class="lineno"> 504</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size,</div>
<div class="line"><a id="l00505" name="l00505"></a><span class="lineno"> 505</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size,</div>
<div class="line"><a id="l00506" name="l00506"></a><span class="lineno"> 506</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00507" name="l00507"></a><span class="lineno"> 507</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00508" name="l00508"></a><span class="lineno"> 508</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00509" name="l00509"></a><span class="lineno"> 509</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> num_simdgroups = 2;</div>
<div class="line"><a id="l00510" name="l00510"></a><span class="lineno"> 510</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> results_per_simdgroup = 4;</div>
<div class="line"><a id="l00511" name="l00511"></a><span class="lineno"> 511</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_thread = 1;</div>
<div class="line"><a id="l00512" name="l00512"></a><span class="lineno"> 512</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00513" name="l00513"></a><span class="lineno"> 513</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> values_per_thread = pack_factor * packs_per_thread;</div>
<div class="line"><a id="l00514" name="l00514"></a><span class="lineno"> 514</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> block_size = values_per_thread * <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>;</div>
<div class="line"><a id="l00515" name="l00515"></a><span class="lineno"> 515</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> scale_step_per_thread = group_size / values_per_thread;</div>
<div class="line"><a id="l00516" name="l00516"></a><span class="lineno"> 516</span> </div>
<div class="line"><a id="l00517" name="l00517"></a><span class="lineno"> 517</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00518" name="l00518"></a><span class="lineno"> 518</span> </div>
<div class="line"><a id="l00519" name="l00519"></a><span class="lineno"> 519</span> thread U x_thread[values_per_thread];</div>
<div class="line"><a id="l00520" name="l00520"></a><span class="lineno"> 520</span> thread U result[results_per_simdgroup] = {0};</div>
<div class="line"><a id="l00521" name="l00521"></a><span class="lineno"> 521</span> </div>
<div class="line"><a id="l00522" name="l00522"></a><span class="lineno"> 522</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00523" name="l00523"></a><span class="lineno"> 523</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_w = in_vec_size / pack_factor;</div>
<div class="line"><a id="l00524" name="l00524"></a><span class="lineno"> 524</span> <span class="keyword">const</span> <span class="keywordtype">int</span> in_vec_size_g = in_vec_size / group_size;</div>
<div class="line"><a id="l00525" name="l00525"></a><span class="lineno"> 525</span> <span class="keyword">const</span> <span class="keywordtype">int</span> out_row = tid.x * (num_simdgroups * results_per_simdgroup) +</div>
<div class="line"><a id="l00526" name="l00526"></a><span class="lineno"> 526</span> simd_gid * results_per_simdgroup;</div>
<div class="line"><a id="l00527" name="l00527"></a><span class="lineno"> 527</span> <span class="keyword">const</span> <span class="keywordtype">int</span> used_out_row = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(out_vec_size - results_per_simdgroup, out_row);</div>
<div class="line"><a id="l00528" name="l00528"></a><span class="lineno"> 528</span> </div>
<div class="line"><a id="l00529" name="l00529"></a><span class="lineno"> 529</span> <span class="keywordflow">if</span> (out_row &gt;= out_vec_size) {</div>
<div class="line"><a id="l00530" name="l00530"></a><span class="lineno"> 530</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00531" name="l00531"></a><span class="lineno"> 531</span> }</div>
<div class="line"><a id="l00532" name="l00532"></a><span class="lineno"> 532</span> </div>
<div class="line"><a id="l00533" name="l00533"></a><span class="lineno"> 533</span> <span class="comment">// In this case we need to properly guard all our reads because there isn&#39;t</span></div>
<div class="line"><a id="l00534" name="l00534"></a><span class="lineno"> 534</span> <span class="comment">// even 1 tile in the matrix</span></div>
<div class="line"><a id="l00535" name="l00535"></a><span class="lineno"> 535</span> <span class="keywordflow">if</span> (out_vec_size &lt; (num_simdgroups * results_per_simdgroup)) {</div>
<div class="line"><a id="l00536" name="l00536"></a><span class="lineno"> 536</span> w += out_row * in_vec_size_w + simd_lid * packs_per_thread;</div>
<div class="line"><a id="l00537" name="l00537"></a><span class="lineno"> 537</span> scales += out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00538" name="l00538"></a><span class="lineno"> 538</span> biases += out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00539" name="l00539"></a><span class="lineno"> 539</span> x += tid.y * in_vec_size + simd_lid * values_per_thread;</div>
<div class="line"><a id="l00540" name="l00540"></a><span class="lineno"> 540</span> y += tid.y * out_vec_size + out_row;</div>
<div class="line"><a id="l00541" name="l00541"></a><span class="lineno"> 541</span> </div>
<div class="line"><a id="l00542" name="l00542"></a><span class="lineno"> 542</span> <span class="keywordtype">int</span> k = 0;</div>
<div class="line"><a id="l00543" name="l00543"></a><span class="lineno"> 543</span> <span class="keywordflow">for</span> (; k &lt; in_vec_size - block_size; k += block_size) {</div>
<div class="line"><a id="l00544" name="l00544"></a><span class="lineno"> 544</span> U sum = <a class="code hl_function" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread);</div>
<div class="line"><a id="l00545" name="l00545"></a><span class="lineno"> 545</span> </div>
<div class="line"><a id="l00546" name="l00546"></a><span class="lineno"> 546</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; out_row + row &lt; out_vec_size; row++) {</div>
<div class="line"><a id="l00547" name="l00547"></a><span class="lineno"> 547</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00548" name="l00548"></a><span class="lineno"> 548</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w);</div>
<div class="line"><a id="l00549" name="l00549"></a><span class="lineno"> 549</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g;</div>
<div class="line"><a id="l00550" name="l00550"></a><span class="lineno"> 550</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g;</div>
<div class="line"><a id="l00551" name="l00551"></a><span class="lineno"> 551</span> </div>
<div class="line"><a id="l00552" name="l00552"></a><span class="lineno"> 552</span> U s = sl[0];</div>
<div class="line"><a id="l00553" name="l00553"></a><span class="lineno"> 553</span> U b = bl[0];</div>
<div class="line"><a id="l00554" name="l00554"></a><span class="lineno"> 554</span> result[row] +=</div>
<div class="line"><a id="l00555" name="l00555"></a><span class="lineno"> 555</span> <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot&lt;U, values_per_thread, bits&gt;</a>(wl, x_thread, s, b, sum);</div>
<div class="line"><a id="l00556" name="l00556"></a><span class="lineno"> 556</span> }</div>
<div class="line"><a id="l00557" name="l00557"></a><span class="lineno"> 557</span> </div>
<div class="line"><a id="l00558" name="l00558"></a><span class="lineno"> 558</span> w += block_size / pack_factor;</div>
<div class="line"><a id="l00559" name="l00559"></a><span class="lineno"> 559</span> scales += block_size / group_size;</div>
<div class="line"><a id="l00560" name="l00560"></a><span class="lineno"> 560</span> biases += block_size / group_size;</div>
<div class="line"><a id="l00561" name="l00561"></a><span class="lineno"> 561</span> x += block_size;</div>
<div class="line"><a id="l00562" name="l00562"></a><span class="lineno"> 562</span> }</div>
<div class="line"><a id="l00563" name="l00563"></a><span class="lineno"> 563</span> <span class="keyword">const</span> <span class="keywordtype">int</span> remaining = clamp(</div>
<div class="line"><a id="l00564" name="l00564"></a><span class="lineno"> 564</span> <span class="keyword">static_cast&lt;</span><span class="keywordtype">int</span><span class="keyword">&gt;</span>(in_vec_size - k - simd_lid * values_per_thread),</div>
<div class="line"><a id="l00565" name="l00565"></a><span class="lineno"> 565</span> 0,</div>
<div class="line"><a id="l00566" name="l00566"></a><span class="lineno"> 566</span> values_per_thread);</div>
<div class="line"><a id="l00567" name="l00567"></a><span class="lineno"> 567</span> U sum =</div>
<div class="line"><a id="l00568" name="l00568"></a><span class="lineno"> 568</span> <a class="code hl_function" href="quantized_8h.html#aa69e143d646fad332c1a53e8c9b337b7">load_vector_safe&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread, remaining);</div>
<div class="line"><a id="l00569" name="l00569"></a><span class="lineno"> 569</span> </div>
<div class="line"><a id="l00570" name="l00570"></a><span class="lineno"> 570</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; out_row + row &lt; out_vec_size; row++) {</div>
<div class="line"><a id="l00571" name="l00571"></a><span class="lineno"> 571</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00572" name="l00572"></a><span class="lineno"> 572</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w);</div>
<div class="line"><a id="l00573" name="l00573"></a><span class="lineno"> 573</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g;</div>
<div class="line"><a id="l00574" name="l00574"></a><span class="lineno"> 574</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g;</div>
<div class="line"><a id="l00575" name="l00575"></a><span class="lineno"> 575</span> </div>
<div class="line"><a id="l00576" name="l00576"></a><span class="lineno"> 576</span> U s = sl[0];</div>
<div class="line"><a id="l00577" name="l00577"></a><span class="lineno"> 577</span> U b = bl[0];</div>
<div class="line"><a id="l00578" name="l00578"></a><span class="lineno"> 578</span> result[row] += <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot&lt;U, values_per_thread, bits&gt;</a>(wl, x_thread, s, b, sum);</div>
<div class="line"><a id="l00579" name="l00579"></a><span class="lineno"> 579</span> }</div>
<div class="line"><a id="l00580" name="l00580"></a><span class="lineno"> 580</span> </div>
<div class="line"><a id="l00581" name="l00581"></a><span class="lineno"> 581</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; out_row + row &lt; out_vec_size; row++) {</div>
<div class="line"><a id="l00582" name="l00582"></a><span class="lineno"> 582</span> result[row] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(result[row]);</div>
<div class="line"><a id="l00583" name="l00583"></a><span class="lineno"> 583</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00584" name="l00584"></a><span class="lineno"> 584</span> y[row] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(result[row]);</div>
<div class="line"><a id="l00585" name="l00585"></a><span class="lineno"> 585</span> }</div>
<div class="line"><a id="l00586" name="l00586"></a><span class="lineno"> 586</span> }</div>
<div class="line"><a id="l00587" name="l00587"></a><span class="lineno"> 587</span> }</div>
<div class="line"><a id="l00588" name="l00588"></a><span class="lineno"> 588</span> </div>
<div class="line"><a id="l00589" name="l00589"></a><span class="lineno"> 589</span> <span class="comment">// In this case the last tile is moved back to redo some output values</span></div>
<div class="line"><a id="l00590" name="l00590"></a><span class="lineno"> 590</span> <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00591" name="l00591"></a><span class="lineno"> 591</span> w += used_out_row * in_vec_size_w + simd_lid * packs_per_thread;</div>
<div class="line"><a id="l00592" name="l00592"></a><span class="lineno"> 592</span> scales += used_out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00593" name="l00593"></a><span class="lineno"> 593</span> biases += used_out_row * in_vec_size_g + simd_lid / scale_step_per_thread;</div>
<div class="line"><a id="l00594" name="l00594"></a><span class="lineno"> 594</span> x += tid.y * in_vec_size + simd_lid * values_per_thread;</div>
<div class="line"><a id="l00595" name="l00595"></a><span class="lineno"> 595</span> y += tid.y * out_vec_size + used_out_row;</div>
<div class="line"><a id="l00596" name="l00596"></a><span class="lineno"> 596</span> </div>
<div class="line"><a id="l00597" name="l00597"></a><span class="lineno"> 597</span> <span class="keywordtype">int</span> k = 0;</div>
<div class="line"><a id="l00598" name="l00598"></a><span class="lineno"> 598</span> <span class="keywordflow">for</span> (; k &lt; in_vec_size - block_size; k += block_size) {</div>
<div class="line"><a id="l00599" name="l00599"></a><span class="lineno"> 599</span> U sum = <a class="code hl_function" href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread);</div>
<div class="line"><a id="l00600" name="l00600"></a><span class="lineno"> 600</span> </div>
<div class="line"><a id="l00601" name="l00601"></a><span class="lineno"> 601</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_simdgroup; row++) {</div>
<div class="line"><a id="l00602" name="l00602"></a><span class="lineno"> 602</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00603" name="l00603"></a><span class="lineno"> 603</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w);</div>
<div class="line"><a id="l00604" name="l00604"></a><span class="lineno"> 604</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g;</div>
<div class="line"><a id="l00605" name="l00605"></a><span class="lineno"> 605</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g;</div>
<div class="line"><a id="l00606" name="l00606"></a><span class="lineno"> 606</span> </div>
<div class="line"><a id="l00607" name="l00607"></a><span class="lineno"> 607</span> U s = sl[0];</div>
<div class="line"><a id="l00608" name="l00608"></a><span class="lineno"> 608</span> U b = bl[0];</div>
<div class="line"><a id="l00609" name="l00609"></a><span class="lineno"> 609</span> result[row] +=</div>
<div class="line"><a id="l00610" name="l00610"></a><span class="lineno"> 610</span> <a class="code hl_function" href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot&lt;U, values_per_thread, bits&gt;</a>(wl, x_thread, s, b, sum);</div>
<div class="line"><a id="l00611" name="l00611"></a><span class="lineno"> 611</span> }</div>
<div class="line"><a id="l00612" name="l00612"></a><span class="lineno"> 612</span> </div>
<div class="line"><a id="l00613" name="l00613"></a><span class="lineno"> 613</span> w += block_size / pack_factor;</div>
<div class="line"><a id="l00614" name="l00614"></a><span class="lineno"> 614</span> scales += block_size / group_size;</div>
<div class="line"><a id="l00615" name="l00615"></a><span class="lineno"> 615</span> biases += block_size / group_size;</div>
<div class="line"><a id="l00616" name="l00616"></a><span class="lineno"> 616</span> x += block_size;</div>
<div class="line"><a id="l00617" name="l00617"></a><span class="lineno"> 617</span> }</div>
<div class="line"><a id="l00618" name="l00618"></a><span class="lineno"> 618</span> <span class="keyword">const</span> <span class="keywordtype">int</span> remaining = clamp(</div>
<div class="line"><a id="l00619" name="l00619"></a><span class="lineno"> 619</span> <span class="keyword">static_cast&lt;</span><span class="keywordtype">int</span><span class="keyword">&gt;</span>(in_vec_size - k - simd_lid * values_per_thread),</div>
<div class="line"><a id="l00620" name="l00620"></a><span class="lineno"> 620</span> 0,</div>
<div class="line"><a id="l00621" name="l00621"></a><span class="lineno"> 621</span> values_per_thread);</div>
<div class="line"><a id="l00622" name="l00622"></a><span class="lineno"> 622</span> U sum =</div>
<div class="line"><a id="l00623" name="l00623"></a><span class="lineno"> 623</span> <a class="code hl_function" href="quantized_8h.html#aa69e143d646fad332c1a53e8c9b337b7">load_vector_safe&lt;T, U, values_per_thread, bits&gt;</a>(x, x_thread, remaining);</div>
<div class="line"><a id="l00624" name="l00624"></a><span class="lineno"> 624</span> </div>
<div class="line"><a id="l00625" name="l00625"></a><span class="lineno"> 625</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_simdgroup; row++) {</div>
<div class="line"><a id="l00626" name="l00626"></a><span class="lineno"> 626</span> <span class="keyword">const</span> device uint8_t* wl =</div>
<div class="line"><a id="l00627" name="l00627"></a><span class="lineno"> 627</span> (<span class="keyword">const</span> device uint8_t*)(w + row * in_vec_size_w);</div>
<div class="line"><a id="l00628" name="l00628"></a><span class="lineno"> 628</span> <span class="keyword">const</span> device T* sl = scales + row * in_vec_size_g;</div>
<div class="line"><a id="l00629" name="l00629"></a><span class="lineno"> 629</span> <span class="keyword">const</span> device T* bl = biases + row * in_vec_size_g;</div>
<div class="line"><a id="l00630" name="l00630"></a><span class="lineno"> 630</span> </div>
<div class="line"><a id="l00631" name="l00631"></a><span class="lineno"> 631</span> U s = sl[0];</div>
<div class="line"><a id="l00632" name="l00632"></a><span class="lineno"> 632</span> U b = bl[0];</div>
<div class="line"><a id="l00633" name="l00633"></a><span class="lineno"> 633</span> result[row] += <a class="code hl_function" href="quantized_8h.html#a07b26d2d0b0d65dfe925c452c453fa42">qdot_safe&lt;U, values_per_thread, bits&gt;</a>(</div>
<div class="line"><a id="l00634" name="l00634"></a><span class="lineno"> 634</span> wl, x_thread, s, b, sum, remaining);</div>
<div class="line"><a id="l00635" name="l00635"></a><span class="lineno"> 635</span> }</div>
<div class="line"><a id="l00636" name="l00636"></a><span class="lineno"> 636</span> </div>
<div class="line"><a id="l00637" name="l00637"></a><span class="lineno"> 637</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> row = 0; row &lt; results_per_simdgroup; row++) {</div>
<div class="line"><a id="l00638" name="l00638"></a><span class="lineno"> 638</span> result[row] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(result[row]);</div>
<div class="line"><a id="l00639" name="l00639"></a><span class="lineno"> 639</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00640" name="l00640"></a><span class="lineno"> 640</span> y[row] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(result[row]);</div>
<div class="line"><a id="l00641" name="l00641"></a><span class="lineno"> 641</span> }</div>
<div class="line"><a id="l00642" name="l00642"></a><span class="lineno"> 642</span> }</div>
<div class="line"><a id="l00643" name="l00643"></a><span class="lineno"> 643</span> }</div>
<div class="line"><a id="l00644" name="l00644"></a><span class="lineno"> 644</span>}</div>
</div>
<div class="line"><a id="l00645" name="l00645"></a><span class="lineno"> 645</span> </div>
<div class="line"><a id="l00646" name="l00646"></a><span class="lineno"> 646</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen00647" data-start="{" data-end="}">
<div class="line"><a id="l00647" name="l00647"></a><span class="lineno"><a class="line" href="quantized_8h.html#a4a8c8db7d5d480733726fd6d1a645e12"> 647</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a4a8c8db7d5d480733726fd6d1a645e12">qvm_impl</a>(</div>
<div class="line"><a id="l00648" name="l00648"></a><span class="lineno"> 648</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00649" name="l00649"></a><span class="lineno"> 649</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00650" name="l00650"></a><span class="lineno"> 650</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00651" name="l00651"></a><span class="lineno"> 651</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00652" name="l00652"></a><span class="lineno"> 652</span> device T* y,</div>
<div class="line"><a id="l00653" name="l00653"></a><span class="lineno"> 653</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size,</div>
<div class="line"><a id="l00654" name="l00654"></a><span class="lineno"> 654</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size,</div>
<div class="line"><a id="l00655" name="l00655"></a><span class="lineno"> 655</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00656" name="l00656"></a><span class="lineno"> 656</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00657" name="l00657"></a><span class="lineno"> 657</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00658" name="l00658"></a><span class="lineno"> 658</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> num_simdgroups = 2;</div>
<div class="line"><a id="l00659" name="l00659"></a><span class="lineno"> 659</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00660" name="l00660"></a><span class="lineno"> 660</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> tn = 32 / pack_factor;</div>
<div class="line"><a id="l00661" name="l00661"></a><span class="lineno"> 661</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> blocksize = <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>;</div>
<div class="line"><a id="l00662" name="l00662"></a><span class="lineno"> 662</span> </div>
<div class="line"><a id="l00663" name="l00663"></a><span class="lineno"> 663</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00664" name="l00664"></a><span class="lineno"> 664</span> <span class="keyword">typedef</span> <span class="keyword">struct </span>{</div>
<div class="line"><a id="l00665" name="l00665"></a><span class="lineno"> 665</span> uint32_t wi[tn];</div>
<div class="line"><a id="l00666" name="l00666"></a><span class="lineno"> 666</span> } vec_w;</div>
<div class="line"><a id="l00667" name="l00667"></a><span class="lineno"> 667</span> </div>
<div class="line"><a id="l00668" name="l00668"></a><span class="lineno"> 668</span> thread vec_w w_local;</div>
<div class="line"><a id="l00669" name="l00669"></a><span class="lineno"> 669</span> thread U result[tn * pack_factor] = {0};</div>
<div class="line"><a id="l00670" name="l00670"></a><span class="lineno"> 670</span> thread U scale = 1;</div>
<div class="line"><a id="l00671" name="l00671"></a><span class="lineno"> 671</span> thread U bias = 0;</div>
<div class="line"><a id="l00672" name="l00672"></a><span class="lineno"> 672</span> thread U x_local = 0;</div>
<div class="line"><a id="l00673" name="l00673"></a><span class="lineno"> 673</span> </div>
<div class="line"><a id="l00674" name="l00674"></a><span class="lineno"> 674</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00675" name="l00675"></a><span class="lineno"> 675</span> <span class="keyword">const</span> <span class="keywordtype">int</span> out_vec_size_w = out_vec_size / pack_factor;</div>
<div class="line"><a id="l00676" name="l00676"></a><span class="lineno"> 676</span> <span class="keyword">const</span> <span class="keywordtype">int</span> out_vec_size_g = out_vec_size / group_size;</div>
<div class="line"><a id="l00677" name="l00677"></a><span class="lineno"> 677</span> <span class="keywordtype">int</span> out_col =</div>
<div class="line"><a id="l00678" name="l00678"></a><span class="lineno"> 678</span> tid.x * (num_simdgroups * pack_factor * tn) + simd_gid * pack_factor * tn;</div>
<div class="line"><a id="l00679" name="l00679"></a><span class="lineno"> 679</span> w += out_col / pack_factor + simd_lid * out_vec_size_w;</div>
<div class="line"><a id="l00680" name="l00680"></a><span class="lineno"> 680</span> scales += out_col / group_size + simd_lid * out_vec_size_g;</div>
<div class="line"><a id="l00681" name="l00681"></a><span class="lineno"> 681</span> biases += out_col / group_size + simd_lid * out_vec_size_g;</div>
<div class="line"><a id="l00682" name="l00682"></a><span class="lineno"> 682</span> x += tid.y * in_vec_size + simd_lid;</div>
<div class="line"><a id="l00683" name="l00683"></a><span class="lineno"> 683</span> y += tid.y * out_vec_size + out_col;</div>
<div class="line"><a id="l00684" name="l00684"></a><span class="lineno"> 684</span> </div>
<div class="line"><a id="l00685" name="l00685"></a><span class="lineno"> 685</span> <span class="keywordflow">if</span> (out_col &gt;= out_vec_size) {</div>
<div class="line"><a id="l00686" name="l00686"></a><span class="lineno"> 686</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00687" name="l00687"></a><span class="lineno"> 687</span> }</div>
<div class="line"><a id="l00688" name="l00688"></a><span class="lineno"> 688</span> </div>
<div class="line"><a id="l00689" name="l00689"></a><span class="lineno"> 689</span> <span class="comment">// Loop over in_vec in blocks of blocksize</span></div>
<div class="line"><a id="l00690" name="l00690"></a><span class="lineno"> 690</span> <span class="keywordtype">int</span> remaining = in_vec_size % blocksize;</div>
<div class="line"><a id="l00691" name="l00691"></a><span class="lineno"> 691</span> <span class="keywordflow">if</span> (remaining == 0) {</div>
<div class="line"><a id="l00692" name="l00692"></a><span class="lineno"> 692</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; in_vec_size; i += blocksize) {</div>
<div class="line"><a id="l00693" name="l00693"></a><span class="lineno"> 693</span> x_local = *x;</div>
<div class="line"><a id="l00694" name="l00694"></a><span class="lineno"> 694</span> scale = *scales;</div>
<div class="line"><a id="l00695" name="l00695"></a><span class="lineno"> 695</span> bias = *biases;</div>
<div class="line"><a id="l00696" name="l00696"></a><span class="lineno"> 696</span> w_local = *((device vec_w*)w);</div>
<div class="line"><a id="l00697" name="l00697"></a><span class="lineno"> 697</span> </div>
<div class="line"><a id="l00698" name="l00698"></a><span class="lineno"> 698</span> <a class="code hl_function" href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58">qouter&lt;U, tn * pack_factor, bits&gt;</a>(</div>
<div class="line"><a id="l00699" name="l00699"></a><span class="lineno"> 699</span> (thread uint8_t*)&amp;w_local, x_local, scale, bias, result);</div>
<div class="line"><a id="l00700" name="l00700"></a><span class="lineno"> 700</span> </div>
<div class="line"><a id="l00701" name="l00701"></a><span class="lineno"> 701</span> x += blocksize;</div>
<div class="line"><a id="l00702" name="l00702"></a><span class="lineno"> 702</span> scales += blocksize * out_vec_size_g;</div>
<div class="line"><a id="l00703" name="l00703"></a><span class="lineno"> 703</span> biases += blocksize * out_vec_size_g;</div>
<div class="line"><a id="l00704" name="l00704"></a><span class="lineno"> 704</span> w += blocksize * out_vec_size_w;</div>
<div class="line"><a id="l00705" name="l00705"></a><span class="lineno"> 705</span> }</div>
<div class="line"><a id="l00706" name="l00706"></a><span class="lineno"> 706</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00707" name="l00707"></a><span class="lineno"> 707</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = blocksize; i &lt; in_vec_size; i += blocksize) {</div>
<div class="line"><a id="l00708" name="l00708"></a><span class="lineno"> 708</span> x_local = *x;</div>
<div class="line"><a id="l00709" name="l00709"></a><span class="lineno"> 709</span> scale = *scales;</div>
<div class="line"><a id="l00710" name="l00710"></a><span class="lineno"> 710</span> bias = *biases;</div>
<div class="line"><a id="l00711" name="l00711"></a><span class="lineno"> 711</span> w_local = *((device vec_w*)w);</div>
<div class="line"><a id="l00712" name="l00712"></a><span class="lineno"> 712</span> </div>
<div class="line"><a id="l00713" name="l00713"></a><span class="lineno"> 713</span> <a class="code hl_function" href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58">qouter&lt;U, tn * pack_factor, bits&gt;</a>(</div>
<div class="line"><a id="l00714" name="l00714"></a><span class="lineno"> 714</span> (thread uint8_t*)&amp;w_local, x_local, scale, bias, result);</div>
<div class="line"><a id="l00715" name="l00715"></a><span class="lineno"> 715</span> </div>
<div class="line"><a id="l00716" name="l00716"></a><span class="lineno"> 716</span> x += blocksize;</div>
<div class="line"><a id="l00717" name="l00717"></a><span class="lineno"> 717</span> scales += blocksize * out_vec_size_g;</div>
<div class="line"><a id="l00718" name="l00718"></a><span class="lineno"> 718</span> biases += blocksize * out_vec_size_g;</div>
<div class="line"><a id="l00719" name="l00719"></a><span class="lineno"> 719</span> w += blocksize * out_vec_size_w;</div>
<div class="line"><a id="l00720" name="l00720"></a><span class="lineno"> 720</span> }</div>
<div class="line"><a id="l00721" name="l00721"></a><span class="lineno"> 721</span> <span class="keywordflow">if</span> (<span class="keyword">static_cast&lt;</span><span class="keywordtype">int</span><span class="keyword">&gt;</span>(simd_lid) &lt; remaining) {</div>
<div class="line"><a id="l00722" name="l00722"></a><span class="lineno"> 722</span> x_local = *x;</div>
<div class="line"><a id="l00723" name="l00723"></a><span class="lineno"> 723</span> scale = *scales;</div>
<div class="line"><a id="l00724" name="l00724"></a><span class="lineno"> 724</span> bias = *biases;</div>
<div class="line"><a id="l00725" name="l00725"></a><span class="lineno"> 725</span> w_local = *((device vec_w*)w);</div>
<div class="line"><a id="l00726" name="l00726"></a><span class="lineno"> 726</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00727" name="l00727"></a><span class="lineno"> 727</span> x_local = 0;</div>
<div class="line"><a id="l00728" name="l00728"></a><span class="lineno"> 728</span> scale = 0;</div>
<div class="line"><a id="l00729" name="l00729"></a><span class="lineno"> 729</span> bias = 0;</div>
<div class="line"><a id="l00730" name="l00730"></a><span class="lineno"> 730</span> }</div>
<div class="line"><a id="l00731" name="l00731"></a><span class="lineno"> 731</span> <a class="code hl_function" href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58">qouter&lt;U, tn * pack_factor, bits&gt;</a>(</div>
<div class="line"><a id="l00732" name="l00732"></a><span class="lineno"> 732</span> (thread uint8_t*)&amp;w_local, x_local, scale, bias, result);</div>
<div class="line"><a id="l00733" name="l00733"></a><span class="lineno"> 733</span> }</div>
<div class="line"><a id="l00734" name="l00734"></a><span class="lineno"> 734</span> </div>
<div class="line"><a id="l00735" name="l00735"></a><span class="lineno"> 735</span><span class="comment">// Accumulate in the simdgroup</span></div>
<div class="line"><a id="l00736" name="l00736"></a><span class="lineno"> 736</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l00737" name="l00737"></a><span class="lineno"> 737</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; tn * pack_factor; k++) {</div>
<div class="line"><a id="l00738" name="l00738"></a><span class="lineno"> 738</span> result[k] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(result[k]);</div>
<div class="line"><a id="l00739" name="l00739"></a><span class="lineno"> 739</span> }</div>
<div class="line"><a id="l00740" name="l00740"></a><span class="lineno"> 740</span> </div>
<div class="line"><a id="l00741" name="l00741"></a><span class="lineno"> 741</span> <span class="comment">// Store the result</span></div>
<div class="line"><a id="l00742" name="l00742"></a><span class="lineno"> 742</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00743" name="l00743"></a><span class="lineno"> 743</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l00744" name="l00744"></a><span class="lineno"> 744</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; tn * pack_factor; k++) {</div>
<div class="line"><a id="l00745" name="l00745"></a><span class="lineno"> 745</span> y[k] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(result[k]);</div>
<div class="line"><a id="l00746" name="l00746"></a><span class="lineno"> 746</span> }</div>
<div class="line"><a id="l00747" name="l00747"></a><span class="lineno"> 747</span> }</div>
<div class="line"><a id="l00748" name="l00748"></a><span class="lineno"> 748</span>}</div>
</div>
<div class="line"><a id="l00749" name="l00749"></a><span class="lineno"> 749</span> </div>
<div class="line"><a id="l00750" name="l00750"></a><span class="lineno"> 750</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l00751" name="l00751"></a><span class="lineno"> 751</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l00752" name="l00752"></a><span class="lineno"> 752</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l00753" name="l00753"></a><span class="lineno"> 753</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l00754" name="l00754"></a><span class="lineno"> 754</span> <span class="keyword">const</span> <span class="keywordtype">bool</span> aligned_N,</div>
<div class="line"><a id="l00755" name="l00755"></a><span class="lineno"> 755</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l00756" name="l00756"></a><span class="lineno"> 756</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l00757" name="l00757"></a><span class="lineno"> 757</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen00758" data-start="{" data-end="}">
<div class="line"><a id="l00758" name="l00758"></a><span class="lineno"><a class="line" href="quantized_8h.html#af5750a35e8f5462218effba719f7f5b8"> 758</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#af5750a35e8f5462218effba719f7f5b8">qmm_t_impl</a>(</div>
<div class="line"><a id="l00759" name="l00759"></a><span class="lineno"> 759</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00760" name="l00760"></a><span class="lineno"> 760</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00761" name="l00761"></a><span class="lineno"> 761</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00762" name="l00762"></a><span class="lineno"> 762</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00763" name="l00763"></a><span class="lineno"> 763</span> device T* y,</div>
<div class="line"><a id="l00764" name="l00764"></a><span class="lineno"> 764</span> threadgroup T* Xs,</div>
<div class="line"><a id="l00765" name="l00765"></a><span class="lineno"> 765</span> threadgroup T* Ws,</div>
<div class="line"><a id="l00766" name="l00766"></a><span class="lineno"> 766</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K,</div>
<div class="line"><a id="l00767" name="l00767"></a><span class="lineno"> 767</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N,</div>
<div class="line"><a id="l00768" name="l00768"></a><span class="lineno"> 768</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M,</div>
<div class="line"><a id="l00769" name="l00769"></a><span class="lineno"> 769</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00770" name="l00770"></a><span class="lineno"> 770</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l00771" name="l00771"></a><span class="lineno"> 771</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00772" name="l00772"></a><span class="lineno"> 772</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00773" name="l00773"></a><span class="lineno"> 773</span> <span class="keyword">static_assert</span>(BK &gt;= <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>, <span class="stringliteral">&quot;BK should be larger than SIMD_SIZE&quot;</span>);</div>
<div class="line"><a id="l00774" name="l00774"></a><span class="lineno"> 774</span> <span class="keyword">static_assert</span>(BK % <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a> == 0, <span class="stringliteral">&quot;BK should be divisible by SIMD_SIZE&quot;</span>);</div>
<div class="line"><a id="l00775" name="l00775"></a><span class="lineno"> 775</span> </div>
<div class="line"><a id="l00776" name="l00776"></a><span class="lineno"> 776</span> (void)lid;</div>
<div class="line"><a id="l00777" name="l00777"></a><span class="lineno"> 777</span> </div>
<div class="line"><a id="l00778" name="l00778"></a><span class="lineno"> 778</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> WM = 2;</div>
<div class="line"><a id="l00779" name="l00779"></a><span class="lineno"> 779</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> WN = 2;</div>
<div class="line"><a id="l00780" name="l00780"></a><span class="lineno"> 780</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00781" name="l00781"></a><span class="lineno"> 781</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l00782" name="l00782"></a><span class="lineno"> 782</span> </div>
<div class="line"><a id="l00783" name="l00783"></a><span class="lineno"> 783</span> <span class="comment">// Instantiate the appropriate BlockMMA and Loader</span></div>
<div class="line"><a id="l00784" name="l00784"></a><span class="lineno"> 784</span> <span class="keyword">using </span>mma_t = mlx::steel::</div>
<div class="line"><a id="l00785" name="l00785"></a><span class="lineno"> 785</span> BlockMMA&lt;T, T, BM, BN, BK, WM, WN, false, true, BK_padded, BK_padded&gt;;</div>
<div class="line"><a id="l00786" name="l00786"></a><span class="lineno"> 786</span> <span class="keyword">using </span>loader_x_t =</div>
<div class="line"><a id="l00787" name="l00787"></a><span class="lineno"> 787</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader.html">mlx::steel::BlockLoader&lt;T, BM, BK, BK_padded, 1, WM * WN * SIMD_SIZE&gt;</a>;</div>
<div class="line"><a id="l00788" name="l00788"></a><span class="lineno"> 788</span> <span class="keyword">using </span>loader_w_t = <a class="code hl_struct" href="struct_quantized_block_loader.html">QuantizedBlockLoader</a>&lt;</div>
<div class="line"><a id="l00789" name="l00789"></a><span class="lineno"> 789</span> T,</div>
<div class="line"><a id="l00790" name="l00790"></a><span class="lineno"> 790</span> BN,</div>
<div class="line"><a id="l00791" name="l00791"></a><span class="lineno"> 791</span> BK,</div>
<div class="line"><a id="l00792" name="l00792"></a><span class="lineno"> 792</span> BK_padded,</div>
<div class="line"><a id="l00793" name="l00793"></a><span class="lineno"> 793</span> 1,</div>
<div class="line"><a id="l00794" name="l00794"></a><span class="lineno"> 794</span> WM * WN * <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>,</div>
<div class="line"><a id="l00795" name="l00795"></a><span class="lineno"> 795</span> group_size,</div>
<div class="line"><a id="l00796" name="l00796"></a><span class="lineno"> 796</span> bits&gt;;</div>
<div class="line"><a id="l00797" name="l00797"></a><span class="lineno"> 797</span> </div>
<div class="line"><a id="l00798" name="l00798"></a><span class="lineno"> 798</span> <span class="comment">// Set the block</span></div>
<div class="line"><a id="l00799" name="l00799"></a><span class="lineno"> 799</span> <span class="keyword">const</span> <span class="keywordtype">int</span> K_w = K / pack_factor;</div>
<div class="line"><a id="l00800" name="l00800"></a><span class="lineno"> 800</span> <span class="keyword">const</span> <span class="keywordtype">int</span> K_g = K / group_size;</div>
<div class="line"><a id="l00801" name="l00801"></a><span class="lineno"> 801</span> <span class="keyword">const</span> <span class="keywordtype">int</span> y_row = tid.y * BM;</div>
<div class="line"><a id="l00802" name="l00802"></a><span class="lineno"> 802</span> <span class="keyword">const</span> <span class="keywordtype">int</span> y_col = tid.x * BN;</div>
<div class="line"><a id="l00803" name="l00803"></a><span class="lineno"> 803</span> </div>
<div class="line"><a id="l00804" name="l00804"></a><span class="lineno"> 804</span> x += y_row * K;</div>
<div class="line"><a id="l00805" name="l00805"></a><span class="lineno"> 805</span> w += y_col * K_w;</div>
<div class="line"><a id="l00806" name="l00806"></a><span class="lineno"> 806</span> scales += y_col * K_g;</div>
<div class="line"><a id="l00807" name="l00807"></a><span class="lineno"> 807</span> biases += y_col * K_g;</div>
<div class="line"><a id="l00808" name="l00808"></a><span class="lineno"> 808</span> y += y_row * N + y_col;</div>
<div class="line"><a id="l00809" name="l00809"></a><span class="lineno"> 809</span> </div>
<div class="line"><a id="l00810" name="l00810"></a><span class="lineno"> 810</span> <span class="comment">// Make the x loader and mma operation</span></div>
<div class="line"><a id="l00811" name="l00811"></a><span class="lineno"> 811</span> <span class="keyword">const</span> <span class="keywordtype">short</span> num_els = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(BM, M - y_row);</div>
<div class="line"><a id="l00812" name="l00812"></a><span class="lineno"> 812</span> <span class="keyword">const</span> <span class="keywordtype">short</span> num_outs = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(BN, N - y_col);</div>
<div class="line"><a id="l00813" name="l00813"></a><span class="lineno"> 813</span> loader_x_t loader_x(x, K, Xs, simd_gid, simd_lid);</div>
<div class="line"><a id="l00814" name="l00814"></a><span class="lineno"> 814</span> loader_w_t loader_w(w, scales, biases, K, Ws, simd_gid, simd_lid);</div>
<div class="line"><a id="l00815" name="l00815"></a><span class="lineno"> 815</span> mma_t mma_op(simd_gid, simd_lid);</div>
<div class="line"><a id="l00816" name="l00816"></a><span class="lineno"> 816</span> </div>
<div class="line"><a id="l00817" name="l00817"></a><span class="lineno"> 817</span> <span class="keywordflow">if</span> (num_els &lt; BM) {</div>
<div class="line"><a id="l00818" name="l00818"></a><span class="lineno"> 818</span> <span class="keywordflow">if</span> (!aligned_N &amp;&amp; num_outs &lt; BN) {</div>
<div class="line"><a id="l00819" name="l00819"></a><span class="lineno"> 819</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00820" name="l00820"></a><span class="lineno"> 820</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00821" name="l00821"></a><span class="lineno"> 821</span> loader_x.load_safe(short2(BK, num_els));</div>
<div class="line"><a id="l00822" name="l00822"></a><span class="lineno"> 822</span> loader_w.load_safe(short2(BK, num_outs));</div>
<div class="line"><a id="l00823" name="l00823"></a><span class="lineno"> 823</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00824" name="l00824"></a><span class="lineno"> 824</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00825" name="l00825"></a><span class="lineno"> 825</span> loader_x.next();</div>
<div class="line"><a id="l00826" name="l00826"></a><span class="lineno"> 826</span> loader_w.next();</div>
<div class="line"><a id="l00827" name="l00827"></a><span class="lineno"> 827</span> }</div>
<div class="line"><a id="l00828" name="l00828"></a><span class="lineno"> 828</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00829" name="l00829"></a><span class="lineno"> 829</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00830" name="l00830"></a><span class="lineno"> 830</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00831" name="l00831"></a><span class="lineno"> 831</span> loader_x.load_safe(short2(BK, num_els));</div>
<div class="line"><a id="l00832" name="l00832"></a><span class="lineno"> 832</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00833" name="l00833"></a><span class="lineno"> 833</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00834" name="l00834"></a><span class="lineno"> 834</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00835" name="l00835"></a><span class="lineno"> 835</span> loader_x.next();</div>
<div class="line"><a id="l00836" name="l00836"></a><span class="lineno"> 836</span> loader_w.next();</div>
<div class="line"><a id="l00837" name="l00837"></a><span class="lineno"> 837</span> }</div>
<div class="line"><a id="l00838" name="l00838"></a><span class="lineno"> 838</span> }</div>
<div class="line"><a id="l00839" name="l00839"></a><span class="lineno"> 839</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00840" name="l00840"></a><span class="lineno"> 840</span> <span class="keywordflow">if</span> (!aligned_N &amp;&amp; num_outs &lt; BN) {</div>
<div class="line"><a id="l00841" name="l00841"></a><span class="lineno"> 841</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00842" name="l00842"></a><span class="lineno"> 842</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00843" name="l00843"></a><span class="lineno"> 843</span> loader_x.load_unsafe();</div>
<div class="line"><a id="l00844" name="l00844"></a><span class="lineno"> 844</span> loader_w.load_safe(short2(BK, num_outs));</div>
<div class="line"><a id="l00845" name="l00845"></a><span class="lineno"> 845</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00846" name="l00846"></a><span class="lineno"> 846</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00847" name="l00847"></a><span class="lineno"> 847</span> loader_x.next();</div>
<div class="line"><a id="l00848" name="l00848"></a><span class="lineno"> 848</span> loader_w.next();</div>
<div class="line"><a id="l00849" name="l00849"></a><span class="lineno"> 849</span> }</div>
<div class="line"><a id="l00850" name="l00850"></a><span class="lineno"> 850</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00851" name="l00851"></a><span class="lineno"> 851</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00852" name="l00852"></a><span class="lineno"> 852</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00853" name="l00853"></a><span class="lineno"> 853</span> loader_x.load_unsafe();</div>
<div class="line"><a id="l00854" name="l00854"></a><span class="lineno"> 854</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00855" name="l00855"></a><span class="lineno"> 855</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00856" name="l00856"></a><span class="lineno"> 856</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00857" name="l00857"></a><span class="lineno"> 857</span> loader_x.next();</div>
<div class="line"><a id="l00858" name="l00858"></a><span class="lineno"> 858</span> loader_w.next();</div>
<div class="line"><a id="l00859" name="l00859"></a><span class="lineno"> 859</span> }</div>
<div class="line"><a id="l00860" name="l00860"></a><span class="lineno"> 860</span> }</div>
<div class="line"><a id="l00861" name="l00861"></a><span class="lineno"> 861</span> }</div>
<div class="line"><a id="l00862" name="l00862"></a><span class="lineno"> 862</span> </div>
<div class="line"><a id="l00863" name="l00863"></a><span class="lineno"> 863</span> <span class="comment">// Store results to device memory</span></div>
<div class="line"><a id="l00864" name="l00864"></a><span class="lineno"> 864</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00865" name="l00865"></a><span class="lineno"> 865</span> <span class="keywordflow">if</span> (num_els &lt; BM || num_outs &lt; BN) {</div>
<div class="line"><a id="l00866" name="l00866"></a><span class="lineno"> 866</span> mma_op.store_result_safe(y, N, short2(num_outs, num_els));</div>
<div class="line"><a id="l00867" name="l00867"></a><span class="lineno"> 867</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00868" name="l00868"></a><span class="lineno"> 868</span> mma_op.store_result(y, N);</div>
<div class="line"><a id="l00869" name="l00869"></a><span class="lineno"> 869</span> }</div>
<div class="line"><a id="l00870" name="l00870"></a><span class="lineno"> 870</span>}</div>
</div>
<div class="line"><a id="l00871" name="l00871"></a><span class="lineno"> 871</span> </div>
<div class="line"><a id="l00872" name="l00872"></a><span class="lineno"> 872</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l00873" name="l00873"></a><span class="lineno"> 873</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l00874" name="l00874"></a><span class="lineno"> 874</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l00875" name="l00875"></a><span class="lineno"> 875</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l00876" name="l00876"></a><span class="lineno"> 876</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l00877" name="l00877"></a><span class="lineno"> 877</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l00878" name="l00878"></a><span class="lineno"> 878</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen00879" data-start="{" data-end="}">
<div class="line"><a id="l00879" name="l00879"></a><span class="lineno"><a class="line" href="quantized_8h.html#a0ba59096494f1001c195312571523ae9"> 879</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a0ba59096494f1001c195312571523ae9">qmm_n_impl</a>(</div>
<div class="line"><a id="l00880" name="l00880"></a><span class="lineno"> 880</span> <span class="keyword">const</span> device uint32_t* w,</div>
<div class="line"><a id="l00881" name="l00881"></a><span class="lineno"> 881</span> <span class="keyword">const</span> device T* scales,</div>
<div class="line"><a id="l00882" name="l00882"></a><span class="lineno"> 882</span> <span class="keyword">const</span> device T* biases,</div>
<div class="line"><a id="l00883" name="l00883"></a><span class="lineno"> 883</span> <span class="keyword">const</span> device T* x,</div>
<div class="line"><a id="l00884" name="l00884"></a><span class="lineno"> 884</span> device T* y,</div>
<div class="line"><a id="l00885" name="l00885"></a><span class="lineno"> 885</span> threadgroup T* Xs,</div>
<div class="line"><a id="l00886" name="l00886"></a><span class="lineno"> 886</span> threadgroup T* Ws,</div>
<div class="line"><a id="l00887" name="l00887"></a><span class="lineno"> 887</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K,</div>
<div class="line"><a id="l00888" name="l00888"></a><span class="lineno"> 888</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N,</div>
<div class="line"><a id="l00889" name="l00889"></a><span class="lineno"> 889</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M,</div>
<div class="line"><a id="l00890" name="l00890"></a><span class="lineno"> 890</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00891" name="l00891"></a><span class="lineno"> 891</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l00892" name="l00892"></a><span class="lineno"> 892</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00893" name="l00893"></a><span class="lineno"> 893</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00894" name="l00894"></a><span class="lineno"> 894</span> <span class="keyword">static_assert</span>(BK &gt;= <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>, <span class="stringliteral">&quot;BK should be larger than SIMD_SIZE&quot;</span>);</div>
<div class="line"><a id="l00895" name="l00895"></a><span class="lineno"> 895</span> <span class="keyword">static_assert</span>(BK % <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a> == 0, <span class="stringliteral">&quot;BK should be divisible by SIMD_SIZE&quot;</span>);</div>
<div class="line"><a id="l00896" name="l00896"></a><span class="lineno"> 896</span> </div>
<div class="line"><a id="l00897" name="l00897"></a><span class="lineno"> 897</span> (void)lid;</div>
<div class="line"><a id="l00898" name="l00898"></a><span class="lineno"> 898</span> </div>
<div class="line"><a id="l00899" name="l00899"></a><span class="lineno"> 899</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> WM = 2;</div>
<div class="line"><a id="l00900" name="l00900"></a><span class="lineno"> 900</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> WN = 2;</div>
<div class="line"><a id="l00901" name="l00901"></a><span class="lineno"> 901</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> pack_factor = 32 / bits;</div>
<div class="line"><a id="l00902" name="l00902"></a><span class="lineno"> 902</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l00903" name="l00903"></a><span class="lineno"> 903</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN_padded = (BN + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l00904" name="l00904"></a><span class="lineno"> 904</span> </div>
<div class="line"><a id="l00905" name="l00905"></a><span class="lineno"> 905</span> <span class="comment">// Instantiate the appropriate BlockMMA and Loader</span></div>
<div class="line"><a id="l00906" name="l00906"></a><span class="lineno"> 906</span> <span class="keyword">using </span>mma_t = mlx::steel::</div>
<div class="line"><a id="l00907" name="l00907"></a><span class="lineno"> 907</span> BlockMMA&lt;T, T, BM, BN, BK, WM, WN, false, false, BK_padded, BN_padded&gt;;</div>
<div class="line"><a id="l00908" name="l00908"></a><span class="lineno"> 908</span> <span class="keyword">using </span>loader_x_t = mlx::steel::</div>
<div class="line"><a id="l00909" name="l00909"></a><span class="lineno"> 909</span> BlockLoader&lt;T, BM, BK, BK_padded, 1, WM * WN * SIMD_SIZE, 1, 4&gt;;</div>
<div class="line"><a id="l00910" name="l00910"></a><span class="lineno"> 910</span> <span class="keyword">using </span>loader_w_t = <a class="code hl_struct" href="struct_quantized_block_loader.html">QuantizedBlockLoader</a>&lt;</div>
<div class="line"><a id="l00911" name="l00911"></a><span class="lineno"> 911</span> T,</div>
<div class="line"><a id="l00912" name="l00912"></a><span class="lineno"> 912</span> BK,</div>
<div class="line"><a id="l00913" name="l00913"></a><span class="lineno"> 913</span> BN,</div>
<div class="line"><a id="l00914" name="l00914"></a><span class="lineno"> 914</span> BN_padded,</div>
<div class="line"><a id="l00915" name="l00915"></a><span class="lineno"> 915</span> 0,</div>
<div class="line"><a id="l00916" name="l00916"></a><span class="lineno"> 916</span> WM * WN * <a class="code hl_variable" href="quantized_8h.html#a62969a218d93680f5e35d0c61b160b99">SIMD_SIZE</a>,</div>
<div class="line"><a id="l00917" name="l00917"></a><span class="lineno"> 917</span> group_size,</div>
<div class="line"><a id="l00918" name="l00918"></a><span class="lineno"> 918</span> bits&gt;;</div>
<div class="line"><a id="l00919" name="l00919"></a><span class="lineno"> 919</span> </div>
<div class="line"><a id="l00920" name="l00920"></a><span class="lineno"> 920</span> <span class="comment">// Set the block</span></div>
<div class="line"><a id="l00921" name="l00921"></a><span class="lineno"> 921</span> <span class="keyword">const</span> <span class="keywordtype">int</span> y_row = tid.y * BM;</div>
<div class="line"><a id="l00922" name="l00922"></a><span class="lineno"> 922</span> <span class="keyword">const</span> <span class="keywordtype">int</span> y_col = tid.x * BN;</div>
<div class="line"><a id="l00923" name="l00923"></a><span class="lineno"> 923</span> x += y_row * K;</div>
<div class="line"><a id="l00924" name="l00924"></a><span class="lineno"> 924</span> w += y_col / pack_factor;</div>
<div class="line"><a id="l00925" name="l00925"></a><span class="lineno"> 925</span> scales += y_col / group_size;</div>
<div class="line"><a id="l00926" name="l00926"></a><span class="lineno"> 926</span> biases += y_col / group_size;</div>
<div class="line"><a id="l00927" name="l00927"></a><span class="lineno"> 927</span> y += y_row * N + y_col;</div>
<div class="line"><a id="l00928" name="l00928"></a><span class="lineno"> 928</span> </div>
<div class="line"><a id="l00929" name="l00929"></a><span class="lineno"> 929</span> <span class="comment">// Make the x loader and mma operation</span></div>
<div class="line"><a id="l00930" name="l00930"></a><span class="lineno"> 930</span> <span class="keyword">const</span> <span class="keywordtype">short</span> num_els = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(BM, M - y_row);</div>
<div class="line"><a id="l00931" name="l00931"></a><span class="lineno"> 931</span> loader_x_t loader_x(x, K, Xs, simd_gid, simd_lid);</div>
<div class="line"><a id="l00932" name="l00932"></a><span class="lineno"> 932</span> loader_w_t loader_w(w, scales, biases, N, Ws, simd_gid, simd_lid);</div>
<div class="line"><a id="l00933" name="l00933"></a><span class="lineno"> 933</span> mma_t mma_op(simd_gid, simd_lid);</div>
<div class="line"><a id="l00934" name="l00934"></a><span class="lineno"> 934</span> </div>
<div class="line"><a id="l00935" name="l00935"></a><span class="lineno"> 935</span> <span class="keywordflow">if</span> (num_els &lt; BM) {</div>
<div class="line"><a id="l00936" name="l00936"></a><span class="lineno"> 936</span> <span class="keywordflow">if</span> ((K % BK) != 0) {</div>
<div class="line"><a id="l00937" name="l00937"></a><span class="lineno"> 937</span> <span class="keyword">const</span> <span class="keywordtype">int</span> k_blocks = K / BK;</div>
<div class="line"><a id="l00938" name="l00938"></a><span class="lineno"> 938</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; k_blocks; k++) {</div>
<div class="line"><a id="l00939" name="l00939"></a><span class="lineno"> 939</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00940" name="l00940"></a><span class="lineno"> 940</span> loader_x.load_safe(short2(BK, num_els));</div>
<div class="line"><a id="l00941" name="l00941"></a><span class="lineno"> 941</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00942" name="l00942"></a><span class="lineno"> 942</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00943" name="l00943"></a><span class="lineno"> 943</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00944" name="l00944"></a><span class="lineno"> 944</span> loader_x.next();</div>
<div class="line"><a id="l00945" name="l00945"></a><span class="lineno"> 945</span> loader_w.next();</div>
<div class="line"><a id="l00946" name="l00946"></a><span class="lineno"> 946</span> }</div>
<div class="line"><a id="l00947" name="l00947"></a><span class="lineno"> 947</span> <span class="keyword">const</span> <span class="keywordtype">short</span> num_k = K - k_blocks * BK;</div>
<div class="line"><a id="l00948" name="l00948"></a><span class="lineno"> 948</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00949" name="l00949"></a><span class="lineno"> 949</span> loader_x.load_safe(short2(num_k, num_els));</div>
<div class="line"><a id="l00950" name="l00950"></a><span class="lineno"> 950</span> loader_w.load_safe(short2(BN, num_k));</div>
<div class="line"><a id="l00951" name="l00951"></a><span class="lineno"> 951</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00952" name="l00952"></a><span class="lineno"> 952</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00953" name="l00953"></a><span class="lineno"> 953</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00954" name="l00954"></a><span class="lineno"> 954</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00955" name="l00955"></a><span class="lineno"> 955</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00956" name="l00956"></a><span class="lineno"> 956</span> loader_x.load_safe(short2(BK, num_els));</div>
<div class="line"><a id="l00957" name="l00957"></a><span class="lineno"> 957</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00958" name="l00958"></a><span class="lineno"> 958</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00959" name="l00959"></a><span class="lineno"> 959</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00960" name="l00960"></a><span class="lineno"> 960</span> loader_x.next();</div>
<div class="line"><a id="l00961" name="l00961"></a><span class="lineno"> 961</span> loader_w.next();</div>
<div class="line"><a id="l00962" name="l00962"></a><span class="lineno"> 962</span> }</div>
<div class="line"><a id="l00963" name="l00963"></a><span class="lineno"> 963</span> }</div>
<div class="line"><a id="l00964" name="l00964"></a><span class="lineno"> 964</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00965" name="l00965"></a><span class="lineno"> 965</span> <span class="keywordflow">if</span> ((K % BK) != 0) {</div>
<div class="line"><a id="l00966" name="l00966"></a><span class="lineno"> 966</span> <span class="keyword">const</span> <span class="keywordtype">int</span> k_blocks = K / BK;</div>
<div class="line"><a id="l00967" name="l00967"></a><span class="lineno"> 967</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; k_blocks; k++) {</div>
<div class="line"><a id="l00968" name="l00968"></a><span class="lineno"> 968</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00969" name="l00969"></a><span class="lineno"> 969</span> loader_x.load_unsafe();</div>
<div class="line"><a id="l00970" name="l00970"></a><span class="lineno"> 970</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00971" name="l00971"></a><span class="lineno"> 971</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00972" name="l00972"></a><span class="lineno"> 972</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00973" name="l00973"></a><span class="lineno"> 973</span> loader_x.next();</div>
<div class="line"><a id="l00974" name="l00974"></a><span class="lineno"> 974</span> loader_w.next();</div>
<div class="line"><a id="l00975" name="l00975"></a><span class="lineno"> 975</span> }</div>
<div class="line"><a id="l00976" name="l00976"></a><span class="lineno"> 976</span> <span class="keyword">const</span> <span class="keywordtype">short</span> num_k = K - k_blocks * BK;</div>
<div class="line"><a id="l00977" name="l00977"></a><span class="lineno"> 977</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00978" name="l00978"></a><span class="lineno"> 978</span> loader_x.load_safe(short2(num_k, BM));</div>
<div class="line"><a id="l00979" name="l00979"></a><span class="lineno"> 979</span> loader_w.load_safe(short2(BN, num_k));</div>
<div class="line"><a id="l00980" name="l00980"></a><span class="lineno"> 980</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00981" name="l00981"></a><span class="lineno"> 981</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00982" name="l00982"></a><span class="lineno"> 982</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00983" name="l00983"></a><span class="lineno"> 983</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; K; k += BK) {</div>
<div class="line"><a id="l00984" name="l00984"></a><span class="lineno"> 984</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00985" name="l00985"></a><span class="lineno"> 985</span> loader_x.load_unsafe();</div>
<div class="line"><a id="l00986" name="l00986"></a><span class="lineno"> 986</span> loader_w.load_unsafe();</div>
<div class="line"><a id="l00987" name="l00987"></a><span class="lineno"> 987</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00988" name="l00988"></a><span class="lineno"> 988</span> mma_op.mma(Xs, Ws);</div>
<div class="line"><a id="l00989" name="l00989"></a><span class="lineno"> 989</span> loader_x.next();</div>
<div class="line"><a id="l00990" name="l00990"></a><span class="lineno"> 990</span> loader_w.next();</div>
<div class="line"><a id="l00991" name="l00991"></a><span class="lineno"> 991</span> }</div>
<div class="line"><a id="l00992" name="l00992"></a><span class="lineno"> 992</span> }</div>
<div class="line"><a id="l00993" name="l00993"></a><span class="lineno"> 993</span> }</div>
<div class="line"><a id="l00994" name="l00994"></a><span class="lineno"> 994</span> </div>
<div class="line"><a id="l00995" name="l00995"></a><span class="lineno"> 995</span> <span class="comment">// Store results to device memory</span></div>
<div class="line"><a id="l00996" name="l00996"></a><span class="lineno"> 996</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00997" name="l00997"></a><span class="lineno"> 997</span> <span class="keywordflow">if</span> (num_els &lt; BM) {</div>
<div class="line"><a id="l00998" name="l00998"></a><span class="lineno"> 998</span> mma_op.store_result_safe(y, N, short2(BN, num_els));</div>
<div class="line"><a id="l00999" name="l00999"></a><span class="lineno"> 999</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01000" name="l01000"></a><span class="lineno"> 1000</span> mma_op.store_result(y, N);</div>
<div class="line"><a id="l01001" name="l01001"></a><span class="lineno"> 1001</span> }</div>
<div class="line"><a id="l01002" name="l01002"></a><span class="lineno"> 1002</span>}</div>
</div>
<div class="line"><a id="l01003" name="l01003"></a><span class="lineno"> 1003</span> </div>
<div class="line"><a id="l01004" name="l01004"></a><span class="lineno"> 1004</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T&gt;</div>
<div class="foldopen" id="foldopen01005" data-start="{" data-end="}">
<div class="line"><a id="l01005" name="l01005"></a><span class="lineno"><a class="line" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be"> 1005</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets</a>(</div>
<div class="line"><a id="l01006" name="l01006"></a><span class="lineno"> 1006</span> <span class="keyword">const</span> device T*&amp; x,</div>
<div class="line"><a id="l01007" name="l01007"></a><span class="lineno"> 1007</span> <span class="keyword">const</span> device uint32_t*&amp; w,</div>
<div class="line"><a id="l01008" name="l01008"></a><span class="lineno"> 1008</span> <span class="keyword">const</span> device T*&amp; scales,</div>
<div class="line"><a id="l01009" name="l01009"></a><span class="lineno"> 1009</span> <span class="keyword">const</span> device T*&amp; biases,</div>
<div class="line"><a id="l01010" name="l01010"></a><span class="lineno"> 1010</span> device T*&amp; y,</div>
<div class="line"><a id="l01011" name="l01011"></a><span class="lineno"> 1011</span> <span class="keywordtype">int</span> output_stride,</div>
<div class="line"><a id="l01012" name="l01012"></a><span class="lineno"> 1012</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims,</div>
<div class="line"><a id="l01013" name="l01013"></a><span class="lineno"> 1013</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape,</div>
<div class="line"><a id="l01014" name="l01014"></a><span class="lineno"> 1014</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides,</div>
<div class="line"><a id="l01015" name="l01015"></a><span class="lineno"> 1015</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims,</div>
<div class="line"><a id="l01016" name="l01016"></a><span class="lineno"> 1016</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape,</div>
<div class="line"><a id="l01017" name="l01017"></a><span class="lineno"> 1017</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides,</div>
<div class="line"><a id="l01018" name="l01018"></a><span class="lineno"> 1018</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides,</div>
<div class="line"><a id="l01019" name="l01019"></a><span class="lineno"> 1019</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides,</div>
<div class="line"><a id="l01020" name="l01020"></a><span class="lineno"> 1020</span> uint3 tid [[threadgroup_position_in_grid]]) {</div>
<div class="line"><a id="l01021" name="l01021"></a><span class="lineno"> 1021</span> <span class="comment">// Set the input/output matrices</span></div>
<div class="line"><a id="l01022" name="l01022"></a><span class="lineno"> 1022</span> uint32_t x_idx = tid.z;</div>
<div class="line"><a id="l01023" name="l01023"></a><span class="lineno"> 1023</span> uint32_t w_idx = tid.z;</div>
<div class="line"><a id="l01024" name="l01024"></a><span class="lineno"> 1024</span> <span class="keywordflow">if</span> (x_batch_ndims == 1) {</div>
<div class="line"><a id="l01025" name="l01025"></a><span class="lineno"> 1025</span> x += x_idx * x_strides[0];</div>
<div class="line"><a id="l01026" name="l01026"></a><span class="lineno"> 1026</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01027" name="l01027"></a><span class="lineno"> 1027</span> x += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a>(x_idx, x_shape, x_strides, x_batch_ndims);</div>
<div class="line"><a id="l01028" name="l01028"></a><span class="lineno"> 1028</span> }</div>
<div class="line"><a id="l01029" name="l01029"></a><span class="lineno"> 1029</span> <span class="keywordflow">if</span> (w_batch_ndims == 1) {</div>
<div class="line"><a id="l01030" name="l01030"></a><span class="lineno"> 1030</span> w += w_idx * w_strides[0];</div>
<div class="line"><a id="l01031" name="l01031"></a><span class="lineno"> 1031</span> scales += w_idx * s_strides[0];</div>
<div class="line"><a id="l01032" name="l01032"></a><span class="lineno"> 1032</span> biases += w_idx * b_strides[0];</div>
<div class="line"><a id="l01033" name="l01033"></a><span class="lineno"> 1033</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01034" name="l01034"></a><span class="lineno"> 1034</span> ulong3 idx = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#aaf4974425147d6f26d031691e321637f">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l01035" name="l01035"></a><span class="lineno"> 1035</span> w_idx, w_shape, w_strides, s_strides, b_strides, w_batch_ndims);</div>
<div class="line"><a id="l01036" name="l01036"></a><span class="lineno"> 1036</span> w += idx.x;</div>
<div class="line"><a id="l01037" name="l01037"></a><span class="lineno"> 1037</span> scales += idx.y;</div>
<div class="line"><a id="l01038" name="l01038"></a><span class="lineno"> 1038</span> biases += idx.z;</div>
<div class="line"><a id="l01039" name="l01039"></a><span class="lineno"> 1039</span> }</div>
<div class="line"><a id="l01040" name="l01040"></a><span class="lineno"> 1040</span> y += tid.z * output_stride;</div>
<div class="line"><a id="l01041" name="l01041"></a><span class="lineno"> 1041</span>}</div>
</div>
<div class="line"><a id="l01042" name="l01042"></a><span class="lineno"> 1042</span> </div>
<div class="line"><a id="l01043" name="l01043"></a><span class="lineno"> 1043</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T&gt;</div>
<div class="foldopen" id="foldopen01044" data-start="{" data-end="}">
<div class="line"><a id="l01044" name="l01044"></a><span class="lineno"><a class="line" href="quantized_8h.html#a3ab400746ad77be89c30d25638e01698"> 1044</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets</a>(</div>
<div class="line"><a id="l01045" name="l01045"></a><span class="lineno"> 1045</span> <span class="keyword">const</span> device T*&amp; x,</div>
<div class="line"><a id="l01046" name="l01046"></a><span class="lineno"> 1046</span> <span class="keyword">const</span> device uint32_t*&amp; w,</div>
<div class="line"><a id="l01047" name="l01047"></a><span class="lineno"> 1047</span> <span class="keyword">const</span> device T*&amp; scales,</div>
<div class="line"><a id="l01048" name="l01048"></a><span class="lineno"> 1048</span> <span class="keyword">const</span> device T*&amp; biases,</div>
<div class="line"><a id="l01049" name="l01049"></a><span class="lineno"> 1049</span> <span class="keyword">const</span> device uint32_t* lhs_indices,</div>
<div class="line"><a id="l01050" name="l01050"></a><span class="lineno"> 1050</span> <span class="keyword">const</span> device uint32_t* rhs_indices,</div>
<div class="line"><a id="l01051" name="l01051"></a><span class="lineno"> 1051</span> device T*&amp; y,</div>
<div class="line"><a id="l01052" name="l01052"></a><span class="lineno"> 1052</span> <span class="keywordtype">int</span> output_stride,</div>
<div class="line"><a id="l01053" name="l01053"></a><span class="lineno"> 1053</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims,</div>
<div class="line"><a id="l01054" name="l01054"></a><span class="lineno"> 1054</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape,</div>
<div class="line"><a id="l01055" name="l01055"></a><span class="lineno"> 1055</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides,</div>
<div class="line"><a id="l01056" name="l01056"></a><span class="lineno"> 1056</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides,</div>
<div class="line"><a id="l01057" name="l01057"></a><span class="lineno"> 1057</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims,</div>
<div class="line"><a id="l01058" name="l01058"></a><span class="lineno"> 1058</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape,</div>
<div class="line"><a id="l01059" name="l01059"></a><span class="lineno"> 1059</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides,</div>
<div class="line"><a id="l01060" name="l01060"></a><span class="lineno"> 1060</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims,</div>
<div class="line"><a id="l01061" name="l01061"></a><span class="lineno"> 1061</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape,</div>
<div class="line"><a id="l01062" name="l01062"></a><span class="lineno"> 1062</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides,</div>
<div class="line"><a id="l01063" name="l01063"></a><span class="lineno"> 1063</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides,</div>
<div class="line"><a id="l01064" name="l01064"></a><span class="lineno"> 1064</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides,</div>
<div class="line"><a id="l01065" name="l01065"></a><span class="lineno"> 1065</span> uint3 tid [[threadgroup_position_in_grid]]) {</div>
<div class="line"><a id="l01066" name="l01066"></a><span class="lineno"> 1066</span> <span class="comment">// Set the input/output matrices</span></div>
<div class="line"><a id="l01067" name="l01067"></a><span class="lineno"> 1067</span> uint32_t x_idx;</div>
<div class="line"><a id="l01068" name="l01068"></a><span class="lineno"> 1068</span> uint32_t w_idx;</div>
<div class="line"><a id="l01069" name="l01069"></a><span class="lineno"> 1069</span> <span class="keywordflow">if</span> (batch_ndims == 1) {</div>
<div class="line"><a id="l01070" name="l01070"></a><span class="lineno"> 1070</span> x_idx = lhs_indices[tid.z * lhs_strides[0]];</div>
<div class="line"><a id="l01071" name="l01071"></a><span class="lineno"> 1071</span> w_idx = rhs_indices[tid.z * rhs_strides[0]];</div>
<div class="line"><a id="l01072" name="l01072"></a><span class="lineno"> 1072</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01073" name="l01073"></a><span class="lineno"> 1073</span> ulong2 idx = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#aaf4974425147d6f26d031691e321637f">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l01074" name="l01074"></a><span class="lineno"> 1074</span> tid.z, batch_shape, lhs_strides, rhs_strides, batch_ndims);</div>
<div class="line"><a id="l01075" name="l01075"></a><span class="lineno"> 1075</span> x_idx = lhs_indices[idx.x];</div>
<div class="line"><a id="l01076" name="l01076"></a><span class="lineno"> 1076</span> w_idx = rhs_indices[idx.y];</div>
<div class="line"><a id="l01077" name="l01077"></a><span class="lineno"> 1077</span> }</div>
<div class="line"><a id="l01078" name="l01078"></a><span class="lineno"> 1078</span> <span class="keywordflow">if</span> (x_batch_ndims == 1) {</div>
<div class="line"><a id="l01079" name="l01079"></a><span class="lineno"> 1079</span> x += x_idx * x_strides[0];</div>
<div class="line"><a id="l01080" name="l01080"></a><span class="lineno"> 1080</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01081" name="l01081"></a><span class="lineno"> 1081</span> x += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a>(x_idx, x_shape, x_strides, x_batch_ndims);</div>
<div class="line"><a id="l01082" name="l01082"></a><span class="lineno"> 1082</span> }</div>
<div class="line"><a id="l01083" name="l01083"></a><span class="lineno"> 1083</span> <span class="keywordflow">if</span> (w_batch_ndims == 1) {</div>
<div class="line"><a id="l01084" name="l01084"></a><span class="lineno"> 1084</span> w += w_idx * w_strides[0];</div>
<div class="line"><a id="l01085" name="l01085"></a><span class="lineno"> 1085</span> scales += w_idx * s_strides[0];</div>
<div class="line"><a id="l01086" name="l01086"></a><span class="lineno"> 1086</span> biases += w_idx * b_strides[0];</div>
<div class="line"><a id="l01087" name="l01087"></a><span class="lineno"> 1087</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01088" name="l01088"></a><span class="lineno"> 1088</span> ulong3 idx = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#aaf4974425147d6f26d031691e321637f">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l01089" name="l01089"></a><span class="lineno"> 1089</span> w_idx, w_shape, w_strides, s_strides, b_strides, w_batch_ndims);</div>
<div class="line"><a id="l01090" name="l01090"></a><span class="lineno"> 1090</span> w += idx.x;</div>
<div class="line"><a id="l01091" name="l01091"></a><span class="lineno"> 1091</span> scales += idx.y;</div>
<div class="line"><a id="l01092" name="l01092"></a><span class="lineno"> 1092</span> biases += idx.z;</div>
<div class="line"><a id="l01093" name="l01093"></a><span class="lineno"> 1093</span> }</div>
<div class="line"><a id="l01094" name="l01094"></a><span class="lineno"> 1094</span> y += tid.z * output_stride;</div>
<div class="line"><a id="l01095" name="l01095"></a><span class="lineno"> 1095</span>}</div>
</div>
<div class="line"><a id="l01096" name="l01096"></a><span class="lineno"> 1096</span> </div>
<div class="line"><a id="l01097" name="l01097"></a><span class="lineno"> 1097</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits, <span class="keywordtype">int</span> D, <span class="keywordtype">bool</span> batched&gt;</div>
<div class="foldopen" id="foldopen01098" data-start="{" data-end="}">
<div class="line"><a id="l01098" name="l01098"></a><span class="lineno"><a class="line" href="quantized_8h.html#a7ce5f53a4d6d1555e9402d545408d0ad"> 1098</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a7ce5f53a4d6d1555e9402d545408d0ad">qmv_quad</a>(</div>
<div class="line"><a id="l01099" name="l01099"></a><span class="lineno"> 1099</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01100" name="l01100"></a><span class="lineno"> 1100</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01101" name="l01101"></a><span class="lineno"> 1101</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01102" name="l01102"></a><span class="lineno"> 1102</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01103" name="l01103"></a><span class="lineno"> 1103</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01104" name="l01104"></a><span class="lineno"> 1104</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01105" name="l01105"></a><span class="lineno"> 1105</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01106" name="l01106"></a><span class="lineno"> 1106</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01107" name="l01107"></a><span class="lineno"> 1107</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01108" name="l01108"></a><span class="lineno"> 1108</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01109" name="l01109"></a><span class="lineno"> 1109</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01110" name="l01110"></a><span class="lineno"> 1110</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01111" name="l01111"></a><span class="lineno"> 1111</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01112" name="l01112"></a><span class="lineno"> 1112</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01113" name="l01113"></a><span class="lineno"> 1113</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01114" name="l01114"></a><span class="lineno"> 1114</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01115" name="l01115"></a><span class="lineno"> 1115</span> uint quad_gid [[quadgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01116" name="l01116"></a><span class="lineno"> 1116</span> uint quad_lid [[thread_index_in_quadgroup]]) {</div>
<div class="line"><a id="l01117" name="l01117"></a><span class="lineno"> 1117</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01118" name="l01118"></a><span class="lineno"> 1118</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01119" name="l01119"></a><span class="lineno"> 1119</span> x,</div>
<div class="line"><a id="l01120" name="l01120"></a><span class="lineno"> 1120</span> w,</div>
<div class="line"><a id="l01121" name="l01121"></a><span class="lineno"> 1121</span> scales,</div>
<div class="line"><a id="l01122" name="l01122"></a><span class="lineno"> 1122</span> biases,</div>
<div class="line"><a id="l01123" name="l01123"></a><span class="lineno"> 1123</span> y,</div>
<div class="line"><a id="l01124" name="l01124"></a><span class="lineno"> 1124</span> out_vec_size,</div>
<div class="line"><a id="l01125" name="l01125"></a><span class="lineno"> 1125</span> x_batch_ndims,</div>
<div class="line"><a id="l01126" name="l01126"></a><span class="lineno"> 1126</span> x_shape,</div>
<div class="line"><a id="l01127" name="l01127"></a><span class="lineno"> 1127</span> x_strides,</div>
<div class="line"><a id="l01128" name="l01128"></a><span class="lineno"> 1128</span> w_batch_ndims,</div>
<div class="line"><a id="l01129" name="l01129"></a><span class="lineno"> 1129</span> w_shape,</div>
<div class="line"><a id="l01130" name="l01130"></a><span class="lineno"> 1130</span> w_strides,</div>
<div class="line"><a id="l01131" name="l01131"></a><span class="lineno"> 1131</span> s_strides,</div>
<div class="line"><a id="l01132" name="l01132"></a><span class="lineno"> 1132</span> b_strides,</div>
<div class="line"><a id="l01133" name="l01133"></a><span class="lineno"> 1133</span> tid);</div>
<div class="line"><a id="l01134" name="l01134"></a><span class="lineno"> 1134</span> }</div>
<div class="line"><a id="l01135" name="l01135"></a><span class="lineno"> 1135</span> <a class="code hl_function" href="quantized_8h.html#ad5cf1cf63656bc1780685d22169cd4ef">qmv_quad_impl&lt;T, group_size, bits, D&gt;</a>(</div>
<div class="line"><a id="l01136" name="l01136"></a><span class="lineno"> 1136</span> w,</div>
<div class="line"><a id="l01137" name="l01137"></a><span class="lineno"> 1137</span> scales,</div>
<div class="line"><a id="l01138" name="l01138"></a><span class="lineno"> 1138</span> biases,</div>
<div class="line"><a id="l01139" name="l01139"></a><span class="lineno"> 1139</span> x,</div>
<div class="line"><a id="l01140" name="l01140"></a><span class="lineno"> 1140</span> y,</div>
<div class="line"><a id="l01141" name="l01141"></a><span class="lineno"> 1141</span> in_vec_size,</div>
<div class="line"><a id="l01142" name="l01142"></a><span class="lineno"> 1142</span> out_vec_size,</div>
<div class="line"><a id="l01143" name="l01143"></a><span class="lineno"> 1143</span> tid,</div>
<div class="line"><a id="l01144" name="l01144"></a><span class="lineno"> 1144</span> quad_gid,</div>
<div class="line"><a id="l01145" name="l01145"></a><span class="lineno"> 1145</span> quad_lid);</div>
<div class="line"><a id="l01146" name="l01146"></a><span class="lineno"> 1146</span>}</div>
</div>
<div class="line"><a id="l01147" name="l01147"></a><span class="lineno"> 1147</span> </div>
<div class="line"><a id="l01148" name="l01148"></a><span class="lineno"> 1148</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits, <span class="keywordtype">bool</span> batched&gt;</div>
<div class="foldopen" id="foldopen01149" data-start="{" data-end="}">
<div class="line"><a id="l01149" name="l01149"></a><span class="lineno"><a class="line" href="quantized_8h.html#a7bd1d9f17c86c8fd34ec13678cff755f"> 1149</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a7bd1d9f17c86c8fd34ec13678cff755f">qmv_fast</a>(</div>
<div class="line"><a id="l01150" name="l01150"></a><span class="lineno"> 1150</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01151" name="l01151"></a><span class="lineno"> 1151</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01152" name="l01152"></a><span class="lineno"> 1152</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01153" name="l01153"></a><span class="lineno"> 1153</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01154" name="l01154"></a><span class="lineno"> 1154</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01155" name="l01155"></a><span class="lineno"> 1155</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01156" name="l01156"></a><span class="lineno"> 1156</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01157" name="l01157"></a><span class="lineno"> 1157</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01158" name="l01158"></a><span class="lineno"> 1158</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01159" name="l01159"></a><span class="lineno"> 1159</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01160" name="l01160"></a><span class="lineno"> 1160</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01161" name="l01161"></a><span class="lineno"> 1161</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01162" name="l01162"></a><span class="lineno"> 1162</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01163" name="l01163"></a><span class="lineno"> 1163</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01164" name="l01164"></a><span class="lineno"> 1164</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01165" name="l01165"></a><span class="lineno"> 1165</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01166" name="l01166"></a><span class="lineno"> 1166</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01167" name="l01167"></a><span class="lineno"> 1167</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01168" name="l01168"></a><span class="lineno"> 1168</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01169" name="l01169"></a><span class="lineno"> 1169</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01170" name="l01170"></a><span class="lineno"> 1170</span> x,</div>
<div class="line"><a id="l01171" name="l01171"></a><span class="lineno"> 1171</span> w,</div>
<div class="line"><a id="l01172" name="l01172"></a><span class="lineno"> 1172</span> scales,</div>
<div class="line"><a id="l01173" name="l01173"></a><span class="lineno"> 1173</span> biases,</div>
<div class="line"><a id="l01174" name="l01174"></a><span class="lineno"> 1174</span> y,</div>
<div class="line"><a id="l01175" name="l01175"></a><span class="lineno"> 1175</span> out_vec_size,</div>
<div class="line"><a id="l01176" name="l01176"></a><span class="lineno"> 1176</span> x_batch_ndims,</div>
<div class="line"><a id="l01177" name="l01177"></a><span class="lineno"> 1177</span> x_shape,</div>
<div class="line"><a id="l01178" name="l01178"></a><span class="lineno"> 1178</span> x_strides,</div>
<div class="line"><a id="l01179" name="l01179"></a><span class="lineno"> 1179</span> w_batch_ndims,</div>
<div class="line"><a id="l01180" name="l01180"></a><span class="lineno"> 1180</span> w_shape,</div>
<div class="line"><a id="l01181" name="l01181"></a><span class="lineno"> 1181</span> w_strides,</div>
<div class="line"><a id="l01182" name="l01182"></a><span class="lineno"> 1182</span> s_strides,</div>
<div class="line"><a id="l01183" name="l01183"></a><span class="lineno"> 1183</span> b_strides,</div>
<div class="line"><a id="l01184" name="l01184"></a><span class="lineno"> 1184</span> tid);</div>
<div class="line"><a id="l01185" name="l01185"></a><span class="lineno"> 1185</span> }</div>
<div class="line"><a id="l01186" name="l01186"></a><span class="lineno"> 1186</span> <a class="code hl_function" href="quantized_8h.html#aba7687e6f8f1d29c0a1b2a3db150bd81">qmv_fast_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01187" name="l01187"></a><span class="lineno"> 1187</span> w,</div>
<div class="line"><a id="l01188" name="l01188"></a><span class="lineno"> 1188</span> scales,</div>
<div class="line"><a id="l01189" name="l01189"></a><span class="lineno"> 1189</span> biases,</div>
<div class="line"><a id="l01190" name="l01190"></a><span class="lineno"> 1190</span> x,</div>
<div class="line"><a id="l01191" name="l01191"></a><span class="lineno"> 1191</span> y,</div>
<div class="line"><a id="l01192" name="l01192"></a><span class="lineno"> 1192</span> in_vec_size,</div>
<div class="line"><a id="l01193" name="l01193"></a><span class="lineno"> 1193</span> out_vec_size,</div>
<div class="line"><a id="l01194" name="l01194"></a><span class="lineno"> 1194</span> tid,</div>
<div class="line"><a id="l01195" name="l01195"></a><span class="lineno"> 1195</span> simd_gid,</div>
<div class="line"><a id="l01196" name="l01196"></a><span class="lineno"> 1196</span> simd_lid);</div>
<div class="line"><a id="l01197" name="l01197"></a><span class="lineno"> 1197</span>}</div>
</div>
<div class="line"><a id="l01198" name="l01198"></a><span class="lineno"> 1198</span> </div>
<div class="line"><a id="l01199" name="l01199"></a><span class="lineno"> 1199</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits, <span class="keywordtype">bool</span> batched&gt;</div>
<div class="foldopen" id="foldopen01200" data-start="{" data-end="}">
<div class="line"><a id="l01200" name="l01200"></a><span class="lineno"><a class="line" href="quantized_8h.html#a639c50a08b5cf57e8be5279a116274bd"> 1200</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a639c50a08b5cf57e8be5279a116274bd">qmv</a>(</div>
<div class="line"><a id="l01201" name="l01201"></a><span class="lineno"> 1201</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01202" name="l01202"></a><span class="lineno"> 1202</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01203" name="l01203"></a><span class="lineno"> 1203</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01204" name="l01204"></a><span class="lineno"> 1204</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01205" name="l01205"></a><span class="lineno"> 1205</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01206" name="l01206"></a><span class="lineno"> 1206</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01207" name="l01207"></a><span class="lineno"> 1207</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01208" name="l01208"></a><span class="lineno"> 1208</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01209" name="l01209"></a><span class="lineno"> 1209</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01210" name="l01210"></a><span class="lineno"> 1210</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01211" name="l01211"></a><span class="lineno"> 1211</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01212" name="l01212"></a><span class="lineno"> 1212</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01213" name="l01213"></a><span class="lineno"> 1213</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01214" name="l01214"></a><span class="lineno"> 1214</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01215" name="l01215"></a><span class="lineno"> 1215</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01216" name="l01216"></a><span class="lineno"> 1216</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01217" name="l01217"></a><span class="lineno"> 1217</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01218" name="l01218"></a><span class="lineno"> 1218</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01219" name="l01219"></a><span class="lineno"> 1219</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01220" name="l01220"></a><span class="lineno"> 1220</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01221" name="l01221"></a><span class="lineno"> 1221</span> x,</div>
<div class="line"><a id="l01222" name="l01222"></a><span class="lineno"> 1222</span> w,</div>
<div class="line"><a id="l01223" name="l01223"></a><span class="lineno"> 1223</span> scales,</div>
<div class="line"><a id="l01224" name="l01224"></a><span class="lineno"> 1224</span> biases,</div>
<div class="line"><a id="l01225" name="l01225"></a><span class="lineno"> 1225</span> y,</div>
<div class="line"><a id="l01226" name="l01226"></a><span class="lineno"> 1226</span> out_vec_size,</div>
<div class="line"><a id="l01227" name="l01227"></a><span class="lineno"> 1227</span> x_batch_ndims,</div>
<div class="line"><a id="l01228" name="l01228"></a><span class="lineno"> 1228</span> x_shape,</div>
<div class="line"><a id="l01229" name="l01229"></a><span class="lineno"> 1229</span> x_strides,</div>
<div class="line"><a id="l01230" name="l01230"></a><span class="lineno"> 1230</span> w_batch_ndims,</div>
<div class="line"><a id="l01231" name="l01231"></a><span class="lineno"> 1231</span> w_shape,</div>
<div class="line"><a id="l01232" name="l01232"></a><span class="lineno"> 1232</span> w_strides,</div>
<div class="line"><a id="l01233" name="l01233"></a><span class="lineno"> 1233</span> s_strides,</div>
<div class="line"><a id="l01234" name="l01234"></a><span class="lineno"> 1234</span> b_strides,</div>
<div class="line"><a id="l01235" name="l01235"></a><span class="lineno"> 1235</span> tid);</div>
<div class="line"><a id="l01236" name="l01236"></a><span class="lineno"> 1236</span> }</div>
<div class="line"><a id="l01237" name="l01237"></a><span class="lineno"> 1237</span> <a class="code hl_function" href="quantized_8h.html#a8e13c7d895624f738d2a6d9893b687fd">qmv_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01238" name="l01238"></a><span class="lineno"> 1238</span> w,</div>
<div class="line"><a id="l01239" name="l01239"></a><span class="lineno"> 1239</span> scales,</div>
<div class="line"><a id="l01240" name="l01240"></a><span class="lineno"> 1240</span> biases,</div>
<div class="line"><a id="l01241" name="l01241"></a><span class="lineno"> 1241</span> x,</div>
<div class="line"><a id="l01242" name="l01242"></a><span class="lineno"> 1242</span> y,</div>
<div class="line"><a id="l01243" name="l01243"></a><span class="lineno"> 1243</span> in_vec_size,</div>
<div class="line"><a id="l01244" name="l01244"></a><span class="lineno"> 1244</span> out_vec_size,</div>
<div class="line"><a id="l01245" name="l01245"></a><span class="lineno"> 1245</span> tid,</div>
<div class="line"><a id="l01246" name="l01246"></a><span class="lineno"> 1246</span> simd_gid,</div>
<div class="line"><a id="l01247" name="l01247"></a><span class="lineno"> 1247</span> simd_lid);</div>
<div class="line"><a id="l01248" name="l01248"></a><span class="lineno"> 1248</span>}</div>
</div>
<div class="line"><a id="l01249" name="l01249"></a><span class="lineno"> 1249</span> </div>
<div class="line"><a id="l01250" name="l01250"></a><span class="lineno"> 1250</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits, <span class="keywordtype">bool</span> batched&gt;</div>
<div class="foldopen" id="foldopen01251" data-start="{" data-end="}">
<div class="line"><a id="l01251" name="l01251"></a><span class="lineno"><a class="line" href="quantized_8h.html#ad84f7d5ab9e32dbbe3ca759ae5d5d5c5"> 1251</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#ad84f7d5ab9e32dbbe3ca759ae5d5d5c5">qvm</a>(</div>
<div class="line"><a id="l01252" name="l01252"></a><span class="lineno"> 1252</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01253" name="l01253"></a><span class="lineno"> 1253</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01254" name="l01254"></a><span class="lineno"> 1254</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01255" name="l01255"></a><span class="lineno"> 1255</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01256" name="l01256"></a><span class="lineno"> 1256</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01257" name="l01257"></a><span class="lineno"> 1257</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01258" name="l01258"></a><span class="lineno"> 1258</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01259" name="l01259"></a><span class="lineno"> 1259</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01260" name="l01260"></a><span class="lineno"> 1260</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01261" name="l01261"></a><span class="lineno"> 1261</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01262" name="l01262"></a><span class="lineno"> 1262</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01263" name="l01263"></a><span class="lineno"> 1263</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01264" name="l01264"></a><span class="lineno"> 1264</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01265" name="l01265"></a><span class="lineno"> 1265</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01266" name="l01266"></a><span class="lineno"> 1266</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01267" name="l01267"></a><span class="lineno"> 1267</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01268" name="l01268"></a><span class="lineno"> 1268</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01269" name="l01269"></a><span class="lineno"> 1269</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01270" name="l01270"></a><span class="lineno"> 1270</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01271" name="l01271"></a><span class="lineno"> 1271</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01272" name="l01272"></a><span class="lineno"> 1272</span> x,</div>
<div class="line"><a id="l01273" name="l01273"></a><span class="lineno"> 1273</span> w,</div>
<div class="line"><a id="l01274" name="l01274"></a><span class="lineno"> 1274</span> scales,</div>
<div class="line"><a id="l01275" name="l01275"></a><span class="lineno"> 1275</span> biases,</div>
<div class="line"><a id="l01276" name="l01276"></a><span class="lineno"> 1276</span> y,</div>
<div class="line"><a id="l01277" name="l01277"></a><span class="lineno"> 1277</span> out_vec_size,</div>
<div class="line"><a id="l01278" name="l01278"></a><span class="lineno"> 1278</span> x_batch_ndims,</div>
<div class="line"><a id="l01279" name="l01279"></a><span class="lineno"> 1279</span> x_shape,</div>
<div class="line"><a id="l01280" name="l01280"></a><span class="lineno"> 1280</span> x_strides,</div>
<div class="line"><a id="l01281" name="l01281"></a><span class="lineno"> 1281</span> w_batch_ndims,</div>
<div class="line"><a id="l01282" name="l01282"></a><span class="lineno"> 1282</span> w_shape,</div>
<div class="line"><a id="l01283" name="l01283"></a><span class="lineno"> 1283</span> w_strides,</div>
<div class="line"><a id="l01284" name="l01284"></a><span class="lineno"> 1284</span> s_strides,</div>
<div class="line"><a id="l01285" name="l01285"></a><span class="lineno"> 1285</span> b_strides,</div>
<div class="line"><a id="l01286" name="l01286"></a><span class="lineno"> 1286</span> tid);</div>
<div class="line"><a id="l01287" name="l01287"></a><span class="lineno"> 1287</span> }</div>
<div class="line"><a id="l01288" name="l01288"></a><span class="lineno"> 1288</span> <a class="code hl_function" href="quantized_8h.html#a4a8c8db7d5d480733726fd6d1a645e12">qvm_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01289" name="l01289"></a><span class="lineno"> 1289</span> w,</div>
<div class="line"><a id="l01290" name="l01290"></a><span class="lineno"> 1290</span> scales,</div>
<div class="line"><a id="l01291" name="l01291"></a><span class="lineno"> 1291</span> biases,</div>
<div class="line"><a id="l01292" name="l01292"></a><span class="lineno"> 1292</span> x,</div>
<div class="line"><a id="l01293" name="l01293"></a><span class="lineno"> 1293</span> y,</div>
<div class="line"><a id="l01294" name="l01294"></a><span class="lineno"> 1294</span> in_vec_size,</div>
<div class="line"><a id="l01295" name="l01295"></a><span class="lineno"> 1295</span> out_vec_size,</div>
<div class="line"><a id="l01296" name="l01296"></a><span class="lineno"> 1296</span> tid,</div>
<div class="line"><a id="l01297" name="l01297"></a><span class="lineno"> 1297</span> simd_gid,</div>
<div class="line"><a id="l01298" name="l01298"></a><span class="lineno"> 1298</span> simd_lid);</div>
<div class="line"><a id="l01299" name="l01299"></a><span class="lineno"> 1299</span>}</div>
</div>
<div class="line"><a id="l01300" name="l01300"></a><span class="lineno"> 1300</span> </div>
<div class="line"><a id="l01301" name="l01301"></a><span class="lineno"> 1301</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l01302" name="l01302"></a><span class="lineno"> 1302</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l01303" name="l01303"></a><span class="lineno"> 1303</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l01304" name="l01304"></a><span class="lineno"> 1304</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l01305" name="l01305"></a><span class="lineno"> 1305</span> <span class="keyword">const</span> <span class="keywordtype">bool</span> aligned_N,</div>
<div class="line"><a id="l01306" name="l01306"></a><span class="lineno"> 1306</span> <span class="keyword">const</span> <span class="keywordtype">bool</span> batched,</div>
<div class="line"><a id="l01307" name="l01307"></a><span class="lineno"> 1307</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l01308" name="l01308"></a><span class="lineno"> 1308</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l01309" name="l01309"></a><span class="lineno"> 1309</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen01310" data-start="{" data-end="}">
<div class="line"><a id="l01310" name="l01310"></a><span class="lineno"><a class="line" href="quantized_8h.html#abe2e3ef0ee4ec2cb61dc5330ad463d10"> 1310</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#abe2e3ef0ee4ec2cb61dc5330ad463d10">qmm_t</a>(</div>
<div class="line"><a id="l01311" name="l01311"></a><span class="lineno"> 1311</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01312" name="l01312"></a><span class="lineno"> 1312</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01313" name="l01313"></a><span class="lineno"> 1313</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01314" name="l01314"></a><span class="lineno"> 1314</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01315" name="l01315"></a><span class="lineno"> 1315</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01316" name="l01316"></a><span class="lineno"> 1316</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K [[buffer(5)]],</div>
<div class="line"><a id="l01317" name="l01317"></a><span class="lineno"> 1317</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N [[buffer(6)]],</div>
<div class="line"><a id="l01318" name="l01318"></a><span class="lineno"> 1318</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M [[buffer(7)]],</div>
<div class="line"><a id="l01319" name="l01319"></a><span class="lineno"> 1319</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(8)]],</div>
<div class="line"><a id="l01320" name="l01320"></a><span class="lineno"> 1320</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(9)]],</div>
<div class="line"><a id="l01321" name="l01321"></a><span class="lineno"> 1321</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(10)]],</div>
<div class="line"><a id="l01322" name="l01322"></a><span class="lineno"> 1322</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(11)]],</div>
<div class="line"><a id="l01323" name="l01323"></a><span class="lineno"> 1323</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(12)]],</div>
<div class="line"><a id="l01324" name="l01324"></a><span class="lineno"> 1324</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(13)]],</div>
<div class="line"><a id="l01325" name="l01325"></a><span class="lineno"> 1325</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(14)]],</div>
<div class="line"><a id="l01326" name="l01326"></a><span class="lineno"> 1326</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(15)]],</div>
<div class="line"><a id="l01327" name="l01327"></a><span class="lineno"> 1327</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01328" name="l01328"></a><span class="lineno"> 1328</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l01329" name="l01329"></a><span class="lineno"> 1329</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01330" name="l01330"></a><span class="lineno"> 1330</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01331" name="l01331"></a><span class="lineno"> 1331</span> (void)lid;</div>
<div class="line"><a id="l01332" name="l01332"></a><span class="lineno"> 1332</span> </div>
<div class="line"><a id="l01333" name="l01333"></a><span class="lineno"> 1333</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01334" name="l01334"></a><span class="lineno"> 1334</span> </div>
<div class="line"><a id="l01335" name="l01335"></a><span class="lineno"> 1335</span> threadgroup T Xs[BM * BK_padded];</div>
<div class="line"><a id="l01336" name="l01336"></a><span class="lineno"> 1336</span> threadgroup T Ws[BN * BK_padded];</div>
<div class="line"><a id="l01337" name="l01337"></a><span class="lineno"> 1337</span> </div>
<div class="line"><a id="l01338" name="l01338"></a><span class="lineno"> 1338</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01339" name="l01339"></a><span class="lineno"> 1339</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01340" name="l01340"></a><span class="lineno"> 1340</span> x,</div>
<div class="line"><a id="l01341" name="l01341"></a><span class="lineno"> 1341</span> w,</div>
<div class="line"><a id="l01342" name="l01342"></a><span class="lineno"> 1342</span> scales,</div>
<div class="line"><a id="l01343" name="l01343"></a><span class="lineno"> 1343</span> biases,</div>
<div class="line"><a id="l01344" name="l01344"></a><span class="lineno"> 1344</span> y,</div>
<div class="line"><a id="l01345" name="l01345"></a><span class="lineno"> 1345</span> M * N,</div>
<div class="line"><a id="l01346" name="l01346"></a><span class="lineno"> 1346</span> x_batch_ndims,</div>
<div class="line"><a id="l01347" name="l01347"></a><span class="lineno"> 1347</span> x_shape,</div>
<div class="line"><a id="l01348" name="l01348"></a><span class="lineno"> 1348</span> x_strides,</div>
<div class="line"><a id="l01349" name="l01349"></a><span class="lineno"> 1349</span> w_batch_ndims,</div>
<div class="line"><a id="l01350" name="l01350"></a><span class="lineno"> 1350</span> w_shape,</div>
<div class="line"><a id="l01351" name="l01351"></a><span class="lineno"> 1351</span> w_strides,</div>
<div class="line"><a id="l01352" name="l01352"></a><span class="lineno"> 1352</span> s_strides,</div>
<div class="line"><a id="l01353" name="l01353"></a><span class="lineno"> 1353</span> b_strides,</div>
<div class="line"><a id="l01354" name="l01354"></a><span class="lineno"> 1354</span> tid);</div>
<div class="line"><a id="l01355" name="l01355"></a><span class="lineno"> 1355</span> }</div>
<div class="line"><a id="l01356" name="l01356"></a><span class="lineno"> 1356</span> <a class="code hl_function" href="quantized_8h.html#af5750a35e8f5462218effba719f7f5b8">qmm_t_impl&lt;T, group_size, bits, aligned_N, BM, BK, BN&gt;</a>(</div>
<div class="line"><a id="l01357" name="l01357"></a><span class="lineno"> 1357</span> w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);</div>
<div class="line"><a id="l01358" name="l01358"></a><span class="lineno"> 1358</span>}</div>
</div>
<div class="line"><a id="l01359" name="l01359"></a><span class="lineno"> 1359</span> </div>
<div class="line"><a id="l01360" name="l01360"></a><span class="lineno"> 1360</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l01361" name="l01361"></a><span class="lineno"> 1361</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l01362" name="l01362"></a><span class="lineno"> 1362</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l01363" name="l01363"></a><span class="lineno"> 1363</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l01364" name="l01364"></a><span class="lineno"> 1364</span> <span class="keyword">const</span> <span class="keywordtype">bool</span> batched,</div>
<div class="line"><a id="l01365" name="l01365"></a><span class="lineno"> 1365</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l01366" name="l01366"></a><span class="lineno"> 1366</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l01367" name="l01367"></a><span class="lineno"> 1367</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen01368" data-start="{" data-end="}">
<div class="line"><a id="l01368" name="l01368"></a><span class="lineno"><a class="line" href="quantized_8h.html#a2ce135e392dbf9a3e5180fb083792ed7"> 1368</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a2ce135e392dbf9a3e5180fb083792ed7">qmm_n</a>(</div>
<div class="line"><a id="l01369" name="l01369"></a><span class="lineno"> 1369</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01370" name="l01370"></a><span class="lineno"> 1370</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01371" name="l01371"></a><span class="lineno"> 1371</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01372" name="l01372"></a><span class="lineno"> 1372</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01373" name="l01373"></a><span class="lineno"> 1373</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01374" name="l01374"></a><span class="lineno"> 1374</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K [[buffer(5)]],</div>
<div class="line"><a id="l01375" name="l01375"></a><span class="lineno"> 1375</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N [[buffer(6)]],</div>
<div class="line"><a id="l01376" name="l01376"></a><span class="lineno"> 1376</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M [[buffer(7)]],</div>
<div class="line"><a id="l01377" name="l01377"></a><span class="lineno"> 1377</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(8)]],</div>
<div class="line"><a id="l01378" name="l01378"></a><span class="lineno"> 1378</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(9)]],</div>
<div class="line"><a id="l01379" name="l01379"></a><span class="lineno"> 1379</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(10)]],</div>
<div class="line"><a id="l01380" name="l01380"></a><span class="lineno"> 1380</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(11)]],</div>
<div class="line"><a id="l01381" name="l01381"></a><span class="lineno"> 1381</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(12)]],</div>
<div class="line"><a id="l01382" name="l01382"></a><span class="lineno"> 1382</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(13)]],</div>
<div class="line"><a id="l01383" name="l01383"></a><span class="lineno"> 1383</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(14)]],</div>
<div class="line"><a id="l01384" name="l01384"></a><span class="lineno"> 1384</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(15)]],</div>
<div class="line"><a id="l01385" name="l01385"></a><span class="lineno"> 1385</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01386" name="l01386"></a><span class="lineno"> 1386</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l01387" name="l01387"></a><span class="lineno"> 1387</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01388" name="l01388"></a><span class="lineno"> 1388</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01389" name="l01389"></a><span class="lineno"> 1389</span> (void)lid;</div>
<div class="line"><a id="l01390" name="l01390"></a><span class="lineno"> 1390</span> </div>
<div class="line"><a id="l01391" name="l01391"></a><span class="lineno"> 1391</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01392" name="l01392"></a><span class="lineno"> 1392</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN_padded = (BN + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01393" name="l01393"></a><span class="lineno"> 1393</span> </div>
<div class="line"><a id="l01394" name="l01394"></a><span class="lineno"> 1394</span> threadgroup T Xs[BM * BK_padded];</div>
<div class="line"><a id="l01395" name="l01395"></a><span class="lineno"> 1395</span> threadgroup T Ws[BK * BN_padded];</div>
<div class="line"><a id="l01396" name="l01396"></a><span class="lineno"> 1396</span> </div>
<div class="line"><a id="l01397" name="l01397"></a><span class="lineno"> 1397</span> <span class="keywordflow">if</span> (batched) {</div>
<div class="line"><a id="l01398" name="l01398"></a><span class="lineno"> 1398</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01399" name="l01399"></a><span class="lineno"> 1399</span> x,</div>
<div class="line"><a id="l01400" name="l01400"></a><span class="lineno"> 1400</span> w,</div>
<div class="line"><a id="l01401" name="l01401"></a><span class="lineno"> 1401</span> scales,</div>
<div class="line"><a id="l01402" name="l01402"></a><span class="lineno"> 1402</span> biases,</div>
<div class="line"><a id="l01403" name="l01403"></a><span class="lineno"> 1403</span> y,</div>
<div class="line"><a id="l01404" name="l01404"></a><span class="lineno"> 1404</span> M * N,</div>
<div class="line"><a id="l01405" name="l01405"></a><span class="lineno"> 1405</span> x_batch_ndims,</div>
<div class="line"><a id="l01406" name="l01406"></a><span class="lineno"> 1406</span> x_shape,</div>
<div class="line"><a id="l01407" name="l01407"></a><span class="lineno"> 1407</span> x_strides,</div>
<div class="line"><a id="l01408" name="l01408"></a><span class="lineno"> 1408</span> w_batch_ndims,</div>
<div class="line"><a id="l01409" name="l01409"></a><span class="lineno"> 1409</span> w_shape,</div>
<div class="line"><a id="l01410" name="l01410"></a><span class="lineno"> 1410</span> w_strides,</div>
<div class="line"><a id="l01411" name="l01411"></a><span class="lineno"> 1411</span> s_strides,</div>
<div class="line"><a id="l01412" name="l01412"></a><span class="lineno"> 1412</span> b_strides,</div>
<div class="line"><a id="l01413" name="l01413"></a><span class="lineno"> 1413</span> tid);</div>
<div class="line"><a id="l01414" name="l01414"></a><span class="lineno"> 1414</span> }</div>
<div class="line"><a id="l01415" name="l01415"></a><span class="lineno"> 1415</span> </div>
<div class="line"><a id="l01416" name="l01416"></a><span class="lineno"> 1416</span> <a class="code hl_function" href="quantized_8h.html#a0ba59096494f1001c195312571523ae9">qmm_n_impl&lt;T, group_size, bits, BM, BK, BN&gt;</a>(</div>
<div class="line"><a id="l01417" name="l01417"></a><span class="lineno"> 1417</span> w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);</div>
<div class="line"><a id="l01418" name="l01418"></a><span class="lineno"> 1418</span>}</div>
</div>
<div class="line"><a id="l01419" name="l01419"></a><span class="lineno"> 1419</span> </div>
<div class="line"><a id="l01420" name="l01420"></a><span class="lineno"> 1420</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01421" data-start="{" data-end="}">
<div class="line"><a id="l01421" name="l01421"></a><span class="lineno"><a class="line" href="quantized_8h.html#a530b720e123e59d73ea89a0a2d0946b7"> 1421</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a530b720e123e59d73ea89a0a2d0946b7">bs_qmv_fast</a>(</div>
<div class="line"><a id="l01422" name="l01422"></a><span class="lineno"> 1422</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01423" name="l01423"></a><span class="lineno"> 1423</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01424" name="l01424"></a><span class="lineno"> 1424</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01425" name="l01425"></a><span class="lineno"> 1425</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01426" name="l01426"></a><span class="lineno"> 1426</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01427" name="l01427"></a><span class="lineno"> 1427</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01428" name="l01428"></a><span class="lineno"> 1428</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01429" name="l01429"></a><span class="lineno"> 1429</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01430" name="l01430"></a><span class="lineno"> 1430</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01431" name="l01431"></a><span class="lineno"> 1431</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01432" name="l01432"></a><span class="lineno"> 1432</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01433" name="l01433"></a><span class="lineno"> 1433</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01434" name="l01434"></a><span class="lineno"> 1434</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01435" name="l01435"></a><span class="lineno"> 1435</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01436" name="l01436"></a><span class="lineno"> 1436</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01437" name="l01437"></a><span class="lineno"> 1437</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims [[buffer(15)]],</div>
<div class="line"><a id="l01438" name="l01438"></a><span class="lineno"> 1438</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(16)]],</div>
<div class="line"><a id="l01439" name="l01439"></a><span class="lineno"> 1439</span> <span class="keyword">const</span> device uint32_t* lhs_indices [[buffer(17)]],</div>
<div class="line"><a id="l01440" name="l01440"></a><span class="lineno"> 1440</span> <span class="keyword">const</span> device uint32_t* rhs_indices [[buffer(18)]],</div>
<div class="line"><a id="l01441" name="l01441"></a><span class="lineno"> 1441</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides [[buffer(19)]],</div>
<div class="line"><a id="l01442" name="l01442"></a><span class="lineno"> 1442</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides [[buffer(20)]],</div>
<div class="line"><a id="l01443" name="l01443"></a><span class="lineno"> 1443</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01444" name="l01444"></a><span class="lineno"> 1444</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01445" name="l01445"></a><span class="lineno"> 1445</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01446" name="l01446"></a><span class="lineno"> 1446</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01447" name="l01447"></a><span class="lineno"> 1447</span> x,</div>
<div class="line"><a id="l01448" name="l01448"></a><span class="lineno"> 1448</span> w,</div>
<div class="line"><a id="l01449" name="l01449"></a><span class="lineno"> 1449</span> scales,</div>
<div class="line"><a id="l01450" name="l01450"></a><span class="lineno"> 1450</span> biases,</div>
<div class="line"><a id="l01451" name="l01451"></a><span class="lineno"> 1451</span> lhs_indices,</div>
<div class="line"><a id="l01452" name="l01452"></a><span class="lineno"> 1452</span> rhs_indices,</div>
<div class="line"><a id="l01453" name="l01453"></a><span class="lineno"> 1453</span> y,</div>
<div class="line"><a id="l01454" name="l01454"></a><span class="lineno"> 1454</span> out_vec_size,</div>
<div class="line"><a id="l01455" name="l01455"></a><span class="lineno"> 1455</span> batch_ndims,</div>
<div class="line"><a id="l01456" name="l01456"></a><span class="lineno"> 1456</span> batch_shape,</div>
<div class="line"><a id="l01457" name="l01457"></a><span class="lineno"> 1457</span> lhs_strides,</div>
<div class="line"><a id="l01458" name="l01458"></a><span class="lineno"> 1458</span> rhs_strides,</div>
<div class="line"><a id="l01459" name="l01459"></a><span class="lineno"> 1459</span> x_batch_ndims,</div>
<div class="line"><a id="l01460" name="l01460"></a><span class="lineno"> 1460</span> x_shape,</div>
<div class="line"><a id="l01461" name="l01461"></a><span class="lineno"> 1461</span> x_strides,</div>
<div class="line"><a id="l01462" name="l01462"></a><span class="lineno"> 1462</span> w_batch_ndims,</div>
<div class="line"><a id="l01463" name="l01463"></a><span class="lineno"> 1463</span> w_shape,</div>
<div class="line"><a id="l01464" name="l01464"></a><span class="lineno"> 1464</span> w_strides,</div>
<div class="line"><a id="l01465" name="l01465"></a><span class="lineno"> 1465</span> s_strides,</div>
<div class="line"><a id="l01466" name="l01466"></a><span class="lineno"> 1466</span> b_strides,</div>
<div class="line"><a id="l01467" name="l01467"></a><span class="lineno"> 1467</span> tid);</div>
<div class="line"><a id="l01468" name="l01468"></a><span class="lineno"> 1468</span> <a class="code hl_function" href="quantized_8h.html#aba7687e6f8f1d29c0a1b2a3db150bd81">qmv_fast_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01469" name="l01469"></a><span class="lineno"> 1469</span> w,</div>
<div class="line"><a id="l01470" name="l01470"></a><span class="lineno"> 1470</span> scales,</div>
<div class="line"><a id="l01471" name="l01471"></a><span class="lineno"> 1471</span> biases,</div>
<div class="line"><a id="l01472" name="l01472"></a><span class="lineno"> 1472</span> x,</div>
<div class="line"><a id="l01473" name="l01473"></a><span class="lineno"> 1473</span> y,</div>
<div class="line"><a id="l01474" name="l01474"></a><span class="lineno"> 1474</span> in_vec_size,</div>
<div class="line"><a id="l01475" name="l01475"></a><span class="lineno"> 1475</span> out_vec_size,</div>
<div class="line"><a id="l01476" name="l01476"></a><span class="lineno"> 1476</span> tid,</div>
<div class="line"><a id="l01477" name="l01477"></a><span class="lineno"> 1477</span> simd_gid,</div>
<div class="line"><a id="l01478" name="l01478"></a><span class="lineno"> 1478</span> simd_lid);</div>
<div class="line"><a id="l01479" name="l01479"></a><span class="lineno"> 1479</span>}</div>
</div>
<div class="line"><a id="l01480" name="l01480"></a><span class="lineno"> 1480</span> </div>
<div class="line"><a id="l01481" name="l01481"></a><span class="lineno"> 1481</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01482" data-start="{" data-end="}">
<div class="line"><a id="l01482" name="l01482"></a><span class="lineno"><a class="line" href="quantized_8h.html#acf4c7fc77821a83b31aedfb48443d3ed"> 1482</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#acf4c7fc77821a83b31aedfb48443d3ed">bs_qmv</a>(</div>
<div class="line"><a id="l01483" name="l01483"></a><span class="lineno"> 1483</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01484" name="l01484"></a><span class="lineno"> 1484</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01485" name="l01485"></a><span class="lineno"> 1485</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01486" name="l01486"></a><span class="lineno"> 1486</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01487" name="l01487"></a><span class="lineno"> 1487</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01488" name="l01488"></a><span class="lineno"> 1488</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01489" name="l01489"></a><span class="lineno"> 1489</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01490" name="l01490"></a><span class="lineno"> 1490</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01491" name="l01491"></a><span class="lineno"> 1491</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01492" name="l01492"></a><span class="lineno"> 1492</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01493" name="l01493"></a><span class="lineno"> 1493</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01494" name="l01494"></a><span class="lineno"> 1494</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01495" name="l01495"></a><span class="lineno"> 1495</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01496" name="l01496"></a><span class="lineno"> 1496</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01497" name="l01497"></a><span class="lineno"> 1497</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01498" name="l01498"></a><span class="lineno"> 1498</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims [[buffer(15)]],</div>
<div class="line"><a id="l01499" name="l01499"></a><span class="lineno"> 1499</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(16)]],</div>
<div class="line"><a id="l01500" name="l01500"></a><span class="lineno"> 1500</span> <span class="keyword">const</span> device uint32_t* lhs_indices [[buffer(17)]],</div>
<div class="line"><a id="l01501" name="l01501"></a><span class="lineno"> 1501</span> <span class="keyword">const</span> device uint32_t* rhs_indices [[buffer(18)]],</div>
<div class="line"><a id="l01502" name="l01502"></a><span class="lineno"> 1502</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides [[buffer(19)]],</div>
<div class="line"><a id="l01503" name="l01503"></a><span class="lineno"> 1503</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides [[buffer(20)]],</div>
<div class="line"><a id="l01504" name="l01504"></a><span class="lineno"> 1504</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01505" name="l01505"></a><span class="lineno"> 1505</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01506" name="l01506"></a><span class="lineno"> 1506</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01507" name="l01507"></a><span class="lineno"> 1507</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01508" name="l01508"></a><span class="lineno"> 1508</span> x,</div>
<div class="line"><a id="l01509" name="l01509"></a><span class="lineno"> 1509</span> w,</div>
<div class="line"><a id="l01510" name="l01510"></a><span class="lineno"> 1510</span> scales,</div>
<div class="line"><a id="l01511" name="l01511"></a><span class="lineno"> 1511</span> biases,</div>
<div class="line"><a id="l01512" name="l01512"></a><span class="lineno"> 1512</span> lhs_indices,</div>
<div class="line"><a id="l01513" name="l01513"></a><span class="lineno"> 1513</span> rhs_indices,</div>
<div class="line"><a id="l01514" name="l01514"></a><span class="lineno"> 1514</span> y,</div>
<div class="line"><a id="l01515" name="l01515"></a><span class="lineno"> 1515</span> out_vec_size,</div>
<div class="line"><a id="l01516" name="l01516"></a><span class="lineno"> 1516</span> batch_ndims,</div>
<div class="line"><a id="l01517" name="l01517"></a><span class="lineno"> 1517</span> batch_shape,</div>
<div class="line"><a id="l01518" name="l01518"></a><span class="lineno"> 1518</span> lhs_strides,</div>
<div class="line"><a id="l01519" name="l01519"></a><span class="lineno"> 1519</span> rhs_strides,</div>
<div class="line"><a id="l01520" name="l01520"></a><span class="lineno"> 1520</span> x_batch_ndims,</div>
<div class="line"><a id="l01521" name="l01521"></a><span class="lineno"> 1521</span> x_shape,</div>
<div class="line"><a id="l01522" name="l01522"></a><span class="lineno"> 1522</span> x_strides,</div>
<div class="line"><a id="l01523" name="l01523"></a><span class="lineno"> 1523</span> w_batch_ndims,</div>
<div class="line"><a id="l01524" name="l01524"></a><span class="lineno"> 1524</span> w_shape,</div>
<div class="line"><a id="l01525" name="l01525"></a><span class="lineno"> 1525</span> w_strides,</div>
<div class="line"><a id="l01526" name="l01526"></a><span class="lineno"> 1526</span> s_strides,</div>
<div class="line"><a id="l01527" name="l01527"></a><span class="lineno"> 1527</span> b_strides,</div>
<div class="line"><a id="l01528" name="l01528"></a><span class="lineno"> 1528</span> tid);</div>
<div class="line"><a id="l01529" name="l01529"></a><span class="lineno"> 1529</span> <a class="code hl_function" href="quantized_8h.html#a8e13c7d895624f738d2a6d9893b687fd">qmv_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01530" name="l01530"></a><span class="lineno"> 1530</span> w,</div>
<div class="line"><a id="l01531" name="l01531"></a><span class="lineno"> 1531</span> scales,</div>
<div class="line"><a id="l01532" name="l01532"></a><span class="lineno"> 1532</span> biases,</div>
<div class="line"><a id="l01533" name="l01533"></a><span class="lineno"> 1533</span> x,</div>
<div class="line"><a id="l01534" name="l01534"></a><span class="lineno"> 1534</span> y,</div>
<div class="line"><a id="l01535" name="l01535"></a><span class="lineno"> 1535</span> in_vec_size,</div>
<div class="line"><a id="l01536" name="l01536"></a><span class="lineno"> 1536</span> out_vec_size,</div>
<div class="line"><a id="l01537" name="l01537"></a><span class="lineno"> 1537</span> tid,</div>
<div class="line"><a id="l01538" name="l01538"></a><span class="lineno"> 1538</span> simd_gid,</div>
<div class="line"><a id="l01539" name="l01539"></a><span class="lineno"> 1539</span> simd_lid);</div>
<div class="line"><a id="l01540" name="l01540"></a><span class="lineno"> 1540</span>}</div>
</div>
<div class="line"><a id="l01541" name="l01541"></a><span class="lineno"> 1541</span> </div>
<div class="line"><a id="l01542" name="l01542"></a><span class="lineno"> 1542</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> group_size, <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01543" data-start="{" data-end="}">
<div class="line"><a id="l01543" name="l01543"></a><span class="lineno"><a class="line" href="quantized_8h.html#a6d6e3c31e44f232e58ae9d605e1f4494"> 1543</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a6d6e3c31e44f232e58ae9d605e1f4494">bs_qvm</a>(</div>
<div class="line"><a id="l01544" name="l01544"></a><span class="lineno"> 1544</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01545" name="l01545"></a><span class="lineno"> 1545</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01546" name="l01546"></a><span class="lineno"> 1546</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01547" name="l01547"></a><span class="lineno"> 1547</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01548" name="l01548"></a><span class="lineno"> 1548</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01549" name="l01549"></a><span class="lineno"> 1549</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; in_vec_size [[buffer(5)]],</div>
<div class="line"><a id="l01550" name="l01550"></a><span class="lineno"> 1550</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; out_vec_size [[buffer(6)]],</div>
<div class="line"><a id="l01551" name="l01551"></a><span class="lineno"> 1551</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(7)]],</div>
<div class="line"><a id="l01552" name="l01552"></a><span class="lineno"> 1552</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(8)]],</div>
<div class="line"><a id="l01553" name="l01553"></a><span class="lineno"> 1553</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(9)]],</div>
<div class="line"><a id="l01554" name="l01554"></a><span class="lineno"> 1554</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(10)]],</div>
<div class="line"><a id="l01555" name="l01555"></a><span class="lineno"> 1555</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(11)]],</div>
<div class="line"><a id="l01556" name="l01556"></a><span class="lineno"> 1556</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(12)]],</div>
<div class="line"><a id="l01557" name="l01557"></a><span class="lineno"> 1557</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(13)]],</div>
<div class="line"><a id="l01558" name="l01558"></a><span class="lineno"> 1558</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(14)]],</div>
<div class="line"><a id="l01559" name="l01559"></a><span class="lineno"> 1559</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims [[buffer(15)]],</div>
<div class="line"><a id="l01560" name="l01560"></a><span class="lineno"> 1560</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(16)]],</div>
<div class="line"><a id="l01561" name="l01561"></a><span class="lineno"> 1561</span> <span class="keyword">const</span> device uint32_t* lhs_indices [[buffer(17)]],</div>
<div class="line"><a id="l01562" name="l01562"></a><span class="lineno"> 1562</span> <span class="keyword">const</span> device uint32_t* rhs_indices [[buffer(18)]],</div>
<div class="line"><a id="l01563" name="l01563"></a><span class="lineno"> 1563</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides [[buffer(19)]],</div>
<div class="line"><a id="l01564" name="l01564"></a><span class="lineno"> 1564</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides [[buffer(20)]],</div>
<div class="line"><a id="l01565" name="l01565"></a><span class="lineno"> 1565</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01566" name="l01566"></a><span class="lineno"> 1566</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01567" name="l01567"></a><span class="lineno"> 1567</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01568" name="l01568"></a><span class="lineno"> 1568</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01569" name="l01569"></a><span class="lineno"> 1569</span> x,</div>
<div class="line"><a id="l01570" name="l01570"></a><span class="lineno"> 1570</span> w,</div>
<div class="line"><a id="l01571" name="l01571"></a><span class="lineno"> 1571</span> scales,</div>
<div class="line"><a id="l01572" name="l01572"></a><span class="lineno"> 1572</span> biases,</div>
<div class="line"><a id="l01573" name="l01573"></a><span class="lineno"> 1573</span> lhs_indices,</div>
<div class="line"><a id="l01574" name="l01574"></a><span class="lineno"> 1574</span> rhs_indices,</div>
<div class="line"><a id="l01575" name="l01575"></a><span class="lineno"> 1575</span> y,</div>
<div class="line"><a id="l01576" name="l01576"></a><span class="lineno"> 1576</span> out_vec_size,</div>
<div class="line"><a id="l01577" name="l01577"></a><span class="lineno"> 1577</span> batch_ndims,</div>
<div class="line"><a id="l01578" name="l01578"></a><span class="lineno"> 1578</span> batch_shape,</div>
<div class="line"><a id="l01579" name="l01579"></a><span class="lineno"> 1579</span> lhs_strides,</div>
<div class="line"><a id="l01580" name="l01580"></a><span class="lineno"> 1580</span> rhs_strides,</div>
<div class="line"><a id="l01581" name="l01581"></a><span class="lineno"> 1581</span> x_batch_ndims,</div>
<div class="line"><a id="l01582" name="l01582"></a><span class="lineno"> 1582</span> x_shape,</div>
<div class="line"><a id="l01583" name="l01583"></a><span class="lineno"> 1583</span> x_strides,</div>
<div class="line"><a id="l01584" name="l01584"></a><span class="lineno"> 1584</span> w_batch_ndims,</div>
<div class="line"><a id="l01585" name="l01585"></a><span class="lineno"> 1585</span> w_shape,</div>
<div class="line"><a id="l01586" name="l01586"></a><span class="lineno"> 1586</span> w_strides,</div>
<div class="line"><a id="l01587" name="l01587"></a><span class="lineno"> 1587</span> s_strides,</div>
<div class="line"><a id="l01588" name="l01588"></a><span class="lineno"> 1588</span> b_strides,</div>
<div class="line"><a id="l01589" name="l01589"></a><span class="lineno"> 1589</span> tid);</div>
<div class="line"><a id="l01590" name="l01590"></a><span class="lineno"> 1590</span> <a class="code hl_function" href="quantized_8h.html#a4a8c8db7d5d480733726fd6d1a645e12">qvm_impl&lt;T, group_size, bits&gt;</a>(</div>
<div class="line"><a id="l01591" name="l01591"></a><span class="lineno"> 1591</span> w,</div>
<div class="line"><a id="l01592" name="l01592"></a><span class="lineno"> 1592</span> scales,</div>
<div class="line"><a id="l01593" name="l01593"></a><span class="lineno"> 1593</span> biases,</div>
<div class="line"><a id="l01594" name="l01594"></a><span class="lineno"> 1594</span> x,</div>
<div class="line"><a id="l01595" name="l01595"></a><span class="lineno"> 1595</span> y,</div>
<div class="line"><a id="l01596" name="l01596"></a><span class="lineno"> 1596</span> in_vec_size,</div>
<div class="line"><a id="l01597" name="l01597"></a><span class="lineno"> 1597</span> out_vec_size,</div>
<div class="line"><a id="l01598" name="l01598"></a><span class="lineno"> 1598</span> tid,</div>
<div class="line"><a id="l01599" name="l01599"></a><span class="lineno"> 1599</span> simd_gid,</div>
<div class="line"><a id="l01600" name="l01600"></a><span class="lineno"> 1600</span> simd_lid);</div>
<div class="line"><a id="l01601" name="l01601"></a><span class="lineno"> 1601</span>}</div>
</div>
<div class="line"><a id="l01602" name="l01602"></a><span class="lineno"> 1602</span> </div>
<div class="line"><a id="l01603" name="l01603"></a><span class="lineno"> 1603</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l01604" name="l01604"></a><span class="lineno"> 1604</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l01605" name="l01605"></a><span class="lineno"> 1605</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l01606" name="l01606"></a><span class="lineno"> 1606</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l01607" name="l01607"></a><span class="lineno"> 1607</span> <span class="keyword">const</span> <span class="keywordtype">bool</span> aligned_N,</div>
<div class="line"><a id="l01608" name="l01608"></a><span class="lineno"> 1608</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l01609" name="l01609"></a><span class="lineno"> 1609</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l01610" name="l01610"></a><span class="lineno"> 1610</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen01611" data-start="{" data-end="}">
<div class="line"><a id="l01611" name="l01611"></a><span class="lineno"><a class="line" href="quantized_8h.html#ab1ae143eba2afceb8df63f38b26f9a84"> 1611</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#ab1ae143eba2afceb8df63f38b26f9a84">bs_qmm_t</a>(</div>
<div class="line"><a id="l01612" name="l01612"></a><span class="lineno"> 1612</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01613" name="l01613"></a><span class="lineno"> 1613</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01614" name="l01614"></a><span class="lineno"> 1614</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01615" name="l01615"></a><span class="lineno"> 1615</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01616" name="l01616"></a><span class="lineno"> 1616</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01617" name="l01617"></a><span class="lineno"> 1617</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K [[buffer(5)]],</div>
<div class="line"><a id="l01618" name="l01618"></a><span class="lineno"> 1618</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N [[buffer(6)]],</div>
<div class="line"><a id="l01619" name="l01619"></a><span class="lineno"> 1619</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M [[buffer(7)]],</div>
<div class="line"><a id="l01620" name="l01620"></a><span class="lineno"> 1620</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(8)]],</div>
<div class="line"><a id="l01621" name="l01621"></a><span class="lineno"> 1621</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(9)]],</div>
<div class="line"><a id="l01622" name="l01622"></a><span class="lineno"> 1622</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(10)]],</div>
<div class="line"><a id="l01623" name="l01623"></a><span class="lineno"> 1623</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(11)]],</div>
<div class="line"><a id="l01624" name="l01624"></a><span class="lineno"> 1624</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(12)]],</div>
<div class="line"><a id="l01625" name="l01625"></a><span class="lineno"> 1625</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(13)]],</div>
<div class="line"><a id="l01626" name="l01626"></a><span class="lineno"> 1626</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(14)]],</div>
<div class="line"><a id="l01627" name="l01627"></a><span class="lineno"> 1627</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(15)]],</div>
<div class="line"><a id="l01628" name="l01628"></a><span class="lineno"> 1628</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims [[buffer(16)]],</div>
<div class="line"><a id="l01629" name="l01629"></a><span class="lineno"> 1629</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(17)]],</div>
<div class="line"><a id="l01630" name="l01630"></a><span class="lineno"> 1630</span> <span class="keyword">const</span> device uint32_t* lhs_indices [[buffer(18)]],</div>
<div class="line"><a id="l01631" name="l01631"></a><span class="lineno"> 1631</span> <span class="keyword">const</span> device uint32_t* rhs_indices [[buffer(19)]],</div>
<div class="line"><a id="l01632" name="l01632"></a><span class="lineno"> 1632</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides [[buffer(20)]],</div>
<div class="line"><a id="l01633" name="l01633"></a><span class="lineno"> 1633</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides [[buffer(21)]],</div>
<div class="line"><a id="l01634" name="l01634"></a><span class="lineno"> 1634</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01635" name="l01635"></a><span class="lineno"> 1635</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l01636" name="l01636"></a><span class="lineno"> 1636</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01637" name="l01637"></a><span class="lineno"> 1637</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01638" name="l01638"></a><span class="lineno"> 1638</span> (void)lid;</div>
<div class="line"><a id="l01639" name="l01639"></a><span class="lineno"> 1639</span> </div>
<div class="line"><a id="l01640" name="l01640"></a><span class="lineno"> 1640</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01641" name="l01641"></a><span class="lineno"> 1641</span> </div>
<div class="line"><a id="l01642" name="l01642"></a><span class="lineno"> 1642</span> threadgroup T Xs[BM * BK_padded];</div>
<div class="line"><a id="l01643" name="l01643"></a><span class="lineno"> 1643</span> threadgroup T Ws[BN * BK_padded];</div>
<div class="line"><a id="l01644" name="l01644"></a><span class="lineno"> 1644</span> </div>
<div class="line"><a id="l01645" name="l01645"></a><span class="lineno"> 1645</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01646" name="l01646"></a><span class="lineno"> 1646</span> x,</div>
<div class="line"><a id="l01647" name="l01647"></a><span class="lineno"> 1647</span> w,</div>
<div class="line"><a id="l01648" name="l01648"></a><span class="lineno"> 1648</span> scales,</div>
<div class="line"><a id="l01649" name="l01649"></a><span class="lineno"> 1649</span> biases,</div>
<div class="line"><a id="l01650" name="l01650"></a><span class="lineno"> 1650</span> lhs_indices,</div>
<div class="line"><a id="l01651" name="l01651"></a><span class="lineno"> 1651</span> rhs_indices,</div>
<div class="line"><a id="l01652" name="l01652"></a><span class="lineno"> 1652</span> y,</div>
<div class="line"><a id="l01653" name="l01653"></a><span class="lineno"> 1653</span> M * N,</div>
<div class="line"><a id="l01654" name="l01654"></a><span class="lineno"> 1654</span> batch_ndims,</div>
<div class="line"><a id="l01655" name="l01655"></a><span class="lineno"> 1655</span> batch_shape,</div>
<div class="line"><a id="l01656" name="l01656"></a><span class="lineno"> 1656</span> lhs_strides,</div>
<div class="line"><a id="l01657" name="l01657"></a><span class="lineno"> 1657</span> rhs_strides,</div>
<div class="line"><a id="l01658" name="l01658"></a><span class="lineno"> 1658</span> x_batch_ndims,</div>
<div class="line"><a id="l01659" name="l01659"></a><span class="lineno"> 1659</span> x_shape,</div>
<div class="line"><a id="l01660" name="l01660"></a><span class="lineno"> 1660</span> x_strides,</div>
<div class="line"><a id="l01661" name="l01661"></a><span class="lineno"> 1661</span> w_batch_ndims,</div>
<div class="line"><a id="l01662" name="l01662"></a><span class="lineno"> 1662</span> w_shape,</div>
<div class="line"><a id="l01663" name="l01663"></a><span class="lineno"> 1663</span> w_strides,</div>
<div class="line"><a id="l01664" name="l01664"></a><span class="lineno"> 1664</span> s_strides,</div>
<div class="line"><a id="l01665" name="l01665"></a><span class="lineno"> 1665</span> b_strides,</div>
<div class="line"><a id="l01666" name="l01666"></a><span class="lineno"> 1666</span> tid);</div>
<div class="line"><a id="l01667" name="l01667"></a><span class="lineno"> 1667</span> <a class="code hl_function" href="quantized_8h.html#af5750a35e8f5462218effba719f7f5b8">qmm_t_impl&lt;T, group_size, bits, aligned_N, BM, BK, BN&gt;</a>(</div>
<div class="line"><a id="l01668" name="l01668"></a><span class="lineno"> 1668</span> w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);</div>
<div class="line"><a id="l01669" name="l01669"></a><span class="lineno"> 1669</span>}</div>
</div>
<div class="line"><a id="l01670" name="l01670"></a><span class="lineno"> 1670</span> </div>
<div class="line"><a id="l01671" name="l01671"></a><span class="lineno"> 1671</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l01672" name="l01672"></a><span class="lineno"> 1672</span> <span class="keyword">typename</span> T,</div>
<div class="line"><a id="l01673" name="l01673"></a><span class="lineno"> 1673</span> <span class="keyword">const</span> <span class="keywordtype">int</span> group_size,</div>
<div class="line"><a id="l01674" name="l01674"></a><span class="lineno"> 1674</span> <span class="keyword">const</span> <span class="keywordtype">int</span> <a class="code hl_function" href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">bits</a>,</div>
<div class="line"><a id="l01675" name="l01675"></a><span class="lineno"> 1675</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BM = 32,</div>
<div class="line"><a id="l01676" name="l01676"></a><span class="lineno"> 1676</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BK = 32,</div>
<div class="line"><a id="l01677" name="l01677"></a><span class="lineno"> 1677</span> <span class="keyword">const</span> <span class="keywordtype">int</span> BN = 32&gt;</div>
<div class="foldopen" id="foldopen01678" data-start="{" data-end="}">
<div class="line"><a id="l01678" name="l01678"></a><span class="lineno"><a class="line" href="quantized_8h.html#a1a66b061c46383952a0f067c3848971f"> 1678</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a1a66b061c46383952a0f067c3848971f">bs_qmm_n</a>(</div>
<div class="line"><a id="l01679" name="l01679"></a><span class="lineno"> 1679</span> <span class="keyword">const</span> device uint32_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01680" name="l01680"></a><span class="lineno"> 1680</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01681" name="l01681"></a><span class="lineno"> 1681</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01682" name="l01682"></a><span class="lineno"> 1682</span> <span class="keyword">const</span> device T* x [[buffer(3)]],</div>
<div class="line"><a id="l01683" name="l01683"></a><span class="lineno"> 1683</span> device T* y [[buffer(4)]],</div>
<div class="line"><a id="l01684" name="l01684"></a><span class="lineno"> 1684</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; K [[buffer(5)]],</div>
<div class="line"><a id="l01685" name="l01685"></a><span class="lineno"> 1685</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N [[buffer(6)]],</div>
<div class="line"><a id="l01686" name="l01686"></a><span class="lineno"> 1686</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; M [[buffer(7)]],</div>
<div class="line"><a id="l01687" name="l01687"></a><span class="lineno"> 1687</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; x_batch_ndims [[buffer(8)]],</div>
<div class="line"><a id="l01688" name="l01688"></a><span class="lineno"> 1688</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* x_shape [[buffer(9)]],</div>
<div class="line"><a id="l01689" name="l01689"></a><span class="lineno"> 1689</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* x_strides [[buffer(10)]],</div>
<div class="line"><a id="l01690" name="l01690"></a><span class="lineno"> 1690</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; w_batch_ndims [[buffer(11)]],</div>
<div class="line"><a id="l01691" name="l01691"></a><span class="lineno"> 1691</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* w_shape [[buffer(12)]],</div>
<div class="line"><a id="l01692" name="l01692"></a><span class="lineno"> 1692</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* w_strides [[buffer(13)]],</div>
<div class="line"><a id="l01693" name="l01693"></a><span class="lineno"> 1693</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* s_strides [[buffer(14)]],</div>
<div class="line"><a id="l01694" name="l01694"></a><span class="lineno"> 1694</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* b_strides [[buffer(15)]],</div>
<div class="line"><a id="l01695" name="l01695"></a><span class="lineno"> 1695</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; batch_ndims [[buffer(16)]],</div>
<div class="line"><a id="l01696" name="l01696"></a><span class="lineno"> 1696</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(17)]],</div>
<div class="line"><a id="l01697" name="l01697"></a><span class="lineno"> 1697</span> <span class="keyword">const</span> device uint32_t* lhs_indices [[buffer(18)]],</div>
<div class="line"><a id="l01698" name="l01698"></a><span class="lineno"> 1698</span> <span class="keyword">const</span> device uint32_t* rhs_indices [[buffer(19)]],</div>
<div class="line"><a id="l01699" name="l01699"></a><span class="lineno"> 1699</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* lhs_strides [[buffer(20)]],</div>
<div class="line"><a id="l01700" name="l01700"></a><span class="lineno"> 1700</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* rhs_strides [[buffer(21)]],</div>
<div class="line"><a id="l01701" name="l01701"></a><span class="lineno"> 1701</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l01702" name="l01702"></a><span class="lineno"> 1702</span> uint lid [[thread_index_in_threadgroup]],</div>
<div class="line"><a id="l01703" name="l01703"></a><span class="lineno"> 1703</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l01704" name="l01704"></a><span class="lineno"> 1704</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l01705" name="l01705"></a><span class="lineno"> 1705</span> (void)lid;</div>
<div class="line"><a id="l01706" name="l01706"></a><span class="lineno"> 1706</span> </div>
<div class="line"><a id="l01707" name="l01707"></a><span class="lineno"> 1707</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BK_padded = (BK + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01708" name="l01708"></a><span class="lineno"> 1708</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN_padded = (BN + 16 / <span class="keyword">sizeof</span>(T));</div>
<div class="line"><a id="l01709" name="l01709"></a><span class="lineno"> 1709</span> </div>
<div class="line"><a id="l01710" name="l01710"></a><span class="lineno"> 1710</span> threadgroup T Xs[BM * BK_padded];</div>
<div class="line"><a id="l01711" name="l01711"></a><span class="lineno"> 1711</span> threadgroup T Ws[BK * BN_padded];</div>
<div class="line"><a id="l01712" name="l01712"></a><span class="lineno"> 1712</span> </div>
<div class="line"><a id="l01713" name="l01713"></a><span class="lineno"> 1713</span> <a class="code hl_function" href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets&lt;T&gt;</a>(</div>
<div class="line"><a id="l01714" name="l01714"></a><span class="lineno"> 1714</span> x,</div>
<div class="line"><a id="l01715" name="l01715"></a><span class="lineno"> 1715</span> w,</div>
<div class="line"><a id="l01716" name="l01716"></a><span class="lineno"> 1716</span> scales,</div>
<div class="line"><a id="l01717" name="l01717"></a><span class="lineno"> 1717</span> biases,</div>
<div class="line"><a id="l01718" name="l01718"></a><span class="lineno"> 1718</span> lhs_indices,</div>
<div class="line"><a id="l01719" name="l01719"></a><span class="lineno"> 1719</span> rhs_indices,</div>
<div class="line"><a id="l01720" name="l01720"></a><span class="lineno"> 1720</span> y,</div>
<div class="line"><a id="l01721" name="l01721"></a><span class="lineno"> 1721</span> M * N,</div>
<div class="line"><a id="l01722" name="l01722"></a><span class="lineno"> 1722</span> batch_ndims,</div>
<div class="line"><a id="l01723" name="l01723"></a><span class="lineno"> 1723</span> batch_shape,</div>
<div class="line"><a id="l01724" name="l01724"></a><span class="lineno"> 1724</span> lhs_strides,</div>
<div class="line"><a id="l01725" name="l01725"></a><span class="lineno"> 1725</span> rhs_strides,</div>
<div class="line"><a id="l01726" name="l01726"></a><span class="lineno"> 1726</span> x_batch_ndims,</div>
<div class="line"><a id="l01727" name="l01727"></a><span class="lineno"> 1727</span> x_shape,</div>
<div class="line"><a id="l01728" name="l01728"></a><span class="lineno"> 1728</span> x_strides,</div>
<div class="line"><a id="l01729" name="l01729"></a><span class="lineno"> 1729</span> w_batch_ndims,</div>
<div class="line"><a id="l01730" name="l01730"></a><span class="lineno"> 1730</span> w_shape,</div>
<div class="line"><a id="l01731" name="l01731"></a><span class="lineno"> 1731</span> w_strides,</div>
<div class="line"><a id="l01732" name="l01732"></a><span class="lineno"> 1732</span> s_strides,</div>
<div class="line"><a id="l01733" name="l01733"></a><span class="lineno"> 1733</span> b_strides,</div>
<div class="line"><a id="l01734" name="l01734"></a><span class="lineno"> 1734</span> tid);</div>
<div class="line"><a id="l01735" name="l01735"></a><span class="lineno"> 1735</span> <a class="code hl_function" href="quantized_8h.html#a0ba59096494f1001c195312571523ae9">qmm_n_impl&lt;T, group_size, bits, BM, BK, BN&gt;</a>(</div>
<div class="line"><a id="l01736" name="l01736"></a><span class="lineno"> 1736</span> w, scales, biases, x, y, Xs, Ws, K, N, M, tid, lid, simd_gid, simd_lid);</div>
<div class="line"><a id="l01737" name="l01737"></a><span class="lineno"> 1737</span>}</div>
</div>
<div class="line"><a id="l01738" name="l01738"></a><span class="lineno"> 1738</span> </div>
<div class="line"><a id="l01739" name="l01739"></a><span class="lineno"> 1739</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01740" data-start="{" data-end="}">
<div class="line"><a id="l01740" name="l01740"></a><span class="lineno"><a class="line" href="quantized_8h.html#a47bcf4a14566e01e14bd3c155811db59"> 1740</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a47bcf4a14566e01e14bd3c155811db59">affine_quantize</a>(</div>
<div class="line"><a id="l01741" name="l01741"></a><span class="lineno"> 1741</span> <span class="keyword">const</span> device T* w [[buffer(0)]],</div>
<div class="line"><a id="l01742" name="l01742"></a><span class="lineno"> 1742</span> device uint8_t* out [[buffer(1)]],</div>
<div class="line"><a id="l01743" name="l01743"></a><span class="lineno"> 1743</span> device T* scales [[buffer(2)]],</div>
<div class="line"><a id="l01744" name="l01744"></a><span class="lineno"> 1744</span> device T* biases [[buffer(3)]],</div>
<div class="line"><a id="l01745" name="l01745"></a><span class="lineno"> 1745</span> uint2 index [[thread_position_in_grid]],</div>
<div class="line"><a id="l01746" name="l01746"></a><span class="lineno"> 1746</span> uint2 grid_dim [[threads_per_grid]]) {</div>
<div class="line"><a id="l01747" name="l01747"></a><span class="lineno"> 1747</span> <span class="keyword">constexpr</span> T eps = T(1e-7);</div>
<div class="line"><a id="l01748" name="l01748"></a><span class="lineno"> 1748</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a> = 32;</div>
<div class="line"><a id="l01749" name="l01749"></a><span class="lineno"> 1749</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> uint8_bits = 8;</div>
<div class="line"><a id="l01750" name="l01750"></a><span class="lineno"> 1750</span> <span class="keyword">constexpr</span> T n_bins = (1 &lt;&lt; bits) - 1;</div>
<div class="line"><a id="l01751" name="l01751"></a><span class="lineno"> 1751</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_int = uint8_bits / bits;</div>
<div class="line"><a id="l01752" name="l01752"></a><span class="lineno"> 1752</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> values_per_reduce = group_size / <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a>;</div>
<div class="line"><a id="l01753" name="l01753"></a><span class="lineno"> 1753</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> writes_per_reduce = packs_per_int / values_per_reduce;</div>
<div class="line"><a id="l01754" name="l01754"></a><span class="lineno"> 1754</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> writes_per_pack =</div>
<div class="line"><a id="l01755" name="l01755"></a><span class="lineno"> 1755</span> writes_per_reduce &gt; 1 ? 1 : values_per_reduce / packs_per_int;</div>
<div class="line"><a id="l01756" name="l01756"></a><span class="lineno"> 1756</span> </div>
<div class="line"><a id="l01757" name="l01757"></a><span class="lineno"> 1757</span> <span class="keyword">static_assert</span>(</div>
<div class="line"><a id="l01758" name="l01758"></a><span class="lineno"> 1758</span> group_size % <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a> == 0,</div>
<div class="line"><a id="l01759" name="l01759"></a><span class="lineno"> 1759</span> <span class="stringliteral">&quot;Group size must be divisible by simd size.&quot;</span>);</div>
<div class="line"><a id="l01760" name="l01760"></a><span class="lineno"> 1760</span> </div>
<div class="line"><a id="l01761" name="l01761"></a><span class="lineno"> 1761</span> <span class="keywordtype">size_t</span> offset = index.x + grid_dim.x * size_t(index.y);</div>
<div class="line"><a id="l01762" name="l01762"></a><span class="lineno"> 1762</span> <span class="keywordtype">size_t</span> in_index = offset * values_per_reduce;</div>
<div class="line"><a id="l01763" name="l01763"></a><span class="lineno"> 1763</span> <span class="keywordtype">size_t</span> out_index = offset * writes_per_pack;</div>
<div class="line"><a id="l01764" name="l01764"></a><span class="lineno"> 1764</span> </div>
<div class="line"><a id="l01765" name="l01765"></a><span class="lineno"> 1765</span> T w_thread[values_per_reduce];</div>
<div class="line"><a id="l01766" name="l01766"></a><span class="lineno"> 1766</span> T w_min = <a class="code hl_struct" href="struct_limits.html">Limits&lt;T&gt;::max</a>;</div>
<div class="line"><a id="l01767" name="l01767"></a><span class="lineno"> 1767</span> T w_max = 0;</div>
<div class="line"><a id="l01768" name="l01768"></a><span class="lineno"> 1768</span> </div>
<div class="line"><a id="l01769" name="l01769"></a><span class="lineno"> 1769</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l01770" name="l01770"></a><span class="lineno"> 1770</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_reduce; i++) {</div>
<div class="line"><a id="l01771" name="l01771"></a><span class="lineno"> 1771</span> T val = w[in_index + i];</div>
<div class="line"><a id="l01772" name="l01772"></a><span class="lineno"> 1772</span> w_thread[i] = val;</div>
<div class="line"><a id="l01773" name="l01773"></a><span class="lineno"> 1773</span> w_min = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(w_min, val);</div>
<div class="line"><a id="l01774" name="l01774"></a><span class="lineno"> 1774</span> w_max = <a class="code hl_function" href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">max</a>(w_max, val);</div>
<div class="line"><a id="l01775" name="l01775"></a><span class="lineno"> 1775</span> }</div>
<div class="line"><a id="l01776" name="l01776"></a><span class="lineno"> 1776</span> </div>
<div class="line"><a id="l01777" name="l01777"></a><span class="lineno"> 1777</span> w_min = <a class="code hl_function" href="namespacemetal.html#ae9e2a23e00724ba2d7868bc4112b386b">simd_min</a>(w_min);</div>
<div class="line"><a id="l01778" name="l01778"></a><span class="lineno"> 1778</span> w_max = <a class="code hl_function" href="namespacemetal.html#a048cad0aca52cb737ebf103e76bd1c49">simd_max</a>(w_max);</div>
<div class="line"><a id="l01779" name="l01779"></a><span class="lineno"> 1779</span> </div>
<div class="line"><a id="l01780" name="l01780"></a><span class="lineno"> 1780</span> T scale = <a class="code hl_function" href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">max</a>((w_max - w_min) / n_bins, eps);</div>
<div class="line"><a id="l01781" name="l01781"></a><span class="lineno"> 1781</span> <span class="keywordtype">bool</span> side = <a class="code hl_function" href="namespacemetal.html#a87c5122c60f9a12afceb9925a5b78ffb">abs</a>(w_min) &gt; <a class="code hl_function" href="namespacemetal.html#a87c5122c60f9a12afceb9925a5b78ffb">abs</a>(w_max);</div>
<div class="line"><a id="l01782" name="l01782"></a><span class="lineno"> 1782</span> scale = side ? scale : -scale;</div>
<div class="line"><a id="l01783" name="l01783"></a><span class="lineno"> 1783</span> T edge = side ? w_min : w_max;</div>
<div class="line"><a id="l01784" name="l01784"></a><span class="lineno"> 1784</span> T q0 = <a class="code hl_function" href="namespacemetal.html#a46c667e169ff9d51a9204a045305442f">round</a>(edge / scale);</div>
<div class="line"><a id="l01785" name="l01785"></a><span class="lineno"> 1785</span> <span class="keywordtype">bool</span> at_zero = q0 == 0.0f;</div>
<div class="line"><a id="l01786" name="l01786"></a><span class="lineno"> 1786</span> scale = at_zero ? scale : edge / q0;</div>
<div class="line"><a id="l01787" name="l01787"></a><span class="lineno"> 1787</span> T bias = at_zero ? T(0) : edge;</div>
<div class="line"><a id="l01788" name="l01788"></a><span class="lineno"> 1788</span> </div>
<div class="line"><a id="l01789" name="l01789"></a><span class="lineno"> 1789</span> <span class="comment">// Write out the scales and biases</span></div>
<div class="line"><a id="l01790" name="l01790"></a><span class="lineno"> 1790</span> <span class="keywordtype">size_t</span> gindex = in_index / group_size;</div>
<div class="line"><a id="l01791" name="l01791"></a><span class="lineno"> 1791</span> <span class="keywordflow">if</span> (in_index % group_size == 0) {</div>
<div class="line"><a id="l01792" name="l01792"></a><span class="lineno"> 1792</span> scales[gindex] = scale;</div>
<div class="line"><a id="l01793" name="l01793"></a><span class="lineno"> 1793</span> biases[gindex] = bias;</div>
<div class="line"><a id="l01794" name="l01794"></a><span class="lineno"> 1794</span> }</div>
<div class="line"><a id="l01795" name="l01795"></a><span class="lineno"> 1795</span> </div>
<div class="line"><a id="l01796" name="l01796"></a><span class="lineno"> 1796</span> uint8_t output = 0;</div>
<div class="line"><a id="l01797" name="l01797"></a><span class="lineno"> 1797</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l01798" name="l01798"></a><span class="lineno"> 1798</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; values_per_reduce; i++) {</div>
<div class="line"><a id="l01799" name="l01799"></a><span class="lineno"> 1799</span> uint8_t val = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(<a class="code hl_function" href="namespacemetal.html#a46c667e169ff9d51a9204a045305442f">round</a>((w_thread[i] - bias) / scale), n_bins);</div>
<div class="line"><a id="l01800" name="l01800"></a><span class="lineno"> 1800</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l01801" name="l01801"></a><span class="lineno"> 1801</span> output = val;</div>
<div class="line"><a id="l01802" name="l01802"></a><span class="lineno"> 1802</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01803" name="l01803"></a><span class="lineno"> 1803</span> output += val &lt;&lt; (bits * (i % packs_per_int));</div>
<div class="line"><a id="l01804" name="l01804"></a><span class="lineno"> 1804</span> }</div>
<div class="line"><a id="l01805" name="l01805"></a><span class="lineno"> 1805</span> </div>
<div class="line"><a id="l01806" name="l01806"></a><span class="lineno"> 1806</span> <span class="keywordflow">if</span> (packs_per_int &lt; values_per_reduce &amp;&amp;</div>
<div class="line"><a id="l01807" name="l01807"></a><span class="lineno"> 1807</span> i % packs_per_int == packs_per_int - 1) {</div>
<div class="line"><a id="l01808" name="l01808"></a><span class="lineno"> 1808</span> out[out_index + i / packs_per_int] = output;</div>
<div class="line"><a id="l01809" name="l01809"></a><span class="lineno"> 1809</span> output = 0;</div>
<div class="line"><a id="l01810" name="l01810"></a><span class="lineno"> 1810</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01811" name="l01811"></a><span class="lineno"> 1811</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l01812" name="l01812"></a><span class="lineno"> 1812</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j &lt; writes_per_reduce - 1; j++) {</div>
<div class="line"><a id="l01813" name="l01813"></a><span class="lineno"> 1813</span> uint8_t sval = <a class="code hl_function" href="namespacemetal.html#af6e2dd7ae087aba6abac4f0350b7611c">simd_shuffle_down</a>(val, j + 1);</div>
<div class="line"><a id="l01814" name="l01814"></a><span class="lineno"> 1814</span> output += sval &lt;&lt; (bits * (values_per_reduce + j + i));</div>
<div class="line"><a id="l01815" name="l01815"></a><span class="lineno"> 1815</span> }</div>
<div class="line"><a id="l01816" name="l01816"></a><span class="lineno"> 1816</span> }</div>
<div class="line"><a id="l01817" name="l01817"></a><span class="lineno"> 1817</span> }</div>
<div class="line"><a id="l01818" name="l01818"></a><span class="lineno"> 1818</span> <span class="keywordflow">if</span> (writes_per_reduce &gt; 0 &amp;&amp; out_index % writes_per_reduce == 0) {</div>
<div class="line"><a id="l01819" name="l01819"></a><span class="lineno"> 1819</span> out[out_index / writes_per_reduce] = output;</div>
<div class="line"><a id="l01820" name="l01820"></a><span class="lineno"> 1820</span> }</div>
<div class="line"><a id="l01821" name="l01821"></a><span class="lineno"> 1821</span>}</div>
</div>
<div class="line"><a id="l01822" name="l01822"></a><span class="lineno"> 1822</span> </div>
<div class="line"><a id="l01823" name="l01823"></a><span class="lineno"> 1823</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01824" data-start="{" data-end="}">
<div class="line"><a id="l01824" name="l01824"></a><span class="lineno"><a class="line" href="quantized_8h.html#a7561acefd7b55e7e2b25393be08bb99c"> 1824</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a7561acefd7b55e7e2b25393be08bb99c">affine_quantize_scales_biases</a>(</div>
<div class="line"><a id="l01825" name="l01825"></a><span class="lineno"> 1825</span> <span class="keyword">const</span> device T* w [[buffer(0)]],</div>
<div class="line"><a id="l01826" name="l01826"></a><span class="lineno"> 1826</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01827" name="l01827"></a><span class="lineno"> 1827</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01828" name="l01828"></a><span class="lineno"> 1828</span> device uint8_t* out [[buffer(3)]],</div>
<div class="line"><a id="l01829" name="l01829"></a><span class="lineno"> 1829</span> uint2 index [[thread_position_in_grid]],</div>
<div class="line"><a id="l01830" name="l01830"></a><span class="lineno"> 1830</span> uint2 grid_dim [[threads_per_grid]]) {</div>
<div class="line"><a id="l01831" name="l01831"></a><span class="lineno"> 1831</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> uint8_bits = 8;</div>
<div class="line"><a id="l01832" name="l01832"></a><span class="lineno"> 1832</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_int = uint8_bits / bits;</div>
<div class="line"><a id="l01833" name="l01833"></a><span class="lineno"> 1833</span> <span class="keyword">constexpr</span> T n_bins = (1 &lt;&lt; bits) - 1;</div>
<div class="line"><a id="l01834" name="l01834"></a><span class="lineno"> 1834</span> </div>
<div class="line"><a id="l01835" name="l01835"></a><span class="lineno"> 1835</span> <span class="keywordtype">size_t</span> offset = index.x + grid_dim.x * size_t(index.y);</div>
<div class="line"><a id="l01836" name="l01836"></a><span class="lineno"> 1836</span> <span class="keywordtype">size_t</span> in_index = offset * packs_per_int;</div>
<div class="line"><a id="l01837" name="l01837"></a><span class="lineno"> 1837</span> <span class="keywordtype">size_t</span> gindex = in_index / group_size;</div>
<div class="line"><a id="l01838" name="l01838"></a><span class="lineno"> 1838</span> </div>
<div class="line"><a id="l01839" name="l01839"></a><span class="lineno"> 1839</span> T scale = scales[gindex];</div>
<div class="line"><a id="l01840" name="l01840"></a><span class="lineno"> 1840</span> T bias = biases[gindex];</div>
<div class="line"><a id="l01841" name="l01841"></a><span class="lineno"> 1841</span> </div>
<div class="line"><a id="l01842" name="l01842"></a><span class="lineno"> 1842</span> uint8_t output = 0;</div>
<div class="line"><a id="l01843" name="l01843"></a><span class="lineno"> 1843</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l01844" name="l01844"></a><span class="lineno"> 1844</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; packs_per_int; i++) {</div>
<div class="line"><a id="l01845" name="l01845"></a><span class="lineno"> 1845</span> uint8_t val = <a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(<a class="code hl_function" href="namespacemetal.html#a46c667e169ff9d51a9204a045305442f">round</a>((w[in_index + i] - bias) / scale), n_bins);</div>
<div class="line"><a id="l01846" name="l01846"></a><span class="lineno"> 1846</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l01847" name="l01847"></a><span class="lineno"> 1847</span> output = val;</div>
<div class="line"><a id="l01848" name="l01848"></a><span class="lineno"> 1848</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l01849" name="l01849"></a><span class="lineno"> 1849</span> output += val &lt;&lt; (bits * i);</div>
<div class="line"><a id="l01850" name="l01850"></a><span class="lineno"> 1850</span> }</div>
<div class="line"><a id="l01851" name="l01851"></a><span class="lineno"> 1851</span> }</div>
<div class="line"><a id="l01852" name="l01852"></a><span class="lineno"> 1852</span> out[offset] = output;</div>
<div class="line"><a id="l01853" name="l01853"></a><span class="lineno"> 1853</span>}</div>
</div>
<div class="line"><a id="l01854" name="l01854"></a><span class="lineno"> 1854</span> </div>
<div class="line"><a id="l01855" name="l01855"></a><span class="lineno"> 1855</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, const <span class="keywordtype">int</span> group_size, const <span class="keywordtype">int</span> bits&gt;</div>
<div class="foldopen" id="foldopen01856" data-start="{" data-end="}">
<div class="line"><a id="l01856" name="l01856"></a><span class="lineno"><a class="line" href="quantized_8h.html#a6076203615038eb06816158f7b3869c6"> 1856</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="quantized_8h.html#a6076203615038eb06816158f7b3869c6">affine_dequantize</a>(</div>
<div class="line"><a id="l01857" name="l01857"></a><span class="lineno"> 1857</span> <span class="keyword">const</span> device uint8_t* w [[buffer(0)]],</div>
<div class="line"><a id="l01858" name="l01858"></a><span class="lineno"> 1858</span> <span class="keyword">const</span> device T* scales [[buffer(1)]],</div>
<div class="line"><a id="l01859" name="l01859"></a><span class="lineno"> 1859</span> <span class="keyword">const</span> device T* biases [[buffer(2)]],</div>
<div class="line"><a id="l01860" name="l01860"></a><span class="lineno"> 1860</span> device T* out [[buffer(3)]],</div>
<div class="line"><a id="l01861" name="l01861"></a><span class="lineno"> 1861</span> uint2 index [[thread_position_in_grid]],</div>
<div class="line"><a id="l01862" name="l01862"></a><span class="lineno"> 1862</span> uint2 grid_dim [[threads_per_grid]]) {</div>
<div class="line"><a id="l01863" name="l01863"></a><span class="lineno"> 1863</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> uint8_bits = 8;</div>
<div class="line"><a id="l01864" name="l01864"></a><span class="lineno"> 1864</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> packs_per_int = uint8_bits / bits;</div>
<div class="line"><a id="l01865" name="l01865"></a><span class="lineno"> 1865</span> </div>
<div class="line"><a id="l01866" name="l01866"></a><span class="lineno"> 1866</span> <span class="keywordtype">size_t</span> offset = index.x + grid_dim.x * size_t(index.y);</div>
<div class="line"><a id="l01867" name="l01867"></a><span class="lineno"> 1867</span> <span class="keywordtype">size_t</span> oindex = offset * packs_per_int;</div>
<div class="line"><a id="l01868" name="l01868"></a><span class="lineno"> 1868</span> <span class="keywordtype">size_t</span> gindex = oindex / group_size;</div>
<div class="line"><a id="l01869" name="l01869"></a><span class="lineno"> 1869</span> T scale = scales[gindex];</div>
<div class="line"><a id="l01870" name="l01870"></a><span class="lineno"> 1870</span> T bias = biases[gindex];</div>
<div class="line"><a id="l01871" name="l01871"></a><span class="lineno"> 1871</span> uint val = w[offset];</div>
<div class="line"><a id="l01872" name="l01872"></a><span class="lineno"> 1872</span> </div>
<div class="line"><a id="l01873" name="l01873"></a><span class="lineno"> 1873</span><span class="preprocessor">#pragma clang loop unroll(full)</span></div>
<div class="line"><a id="l01874" name="l01874"></a><span class="lineno"> 1874</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; packs_per_int; i++) {</div>
<div class="line"><a id="l01875" name="l01875"></a><span class="lineno"> 1875</span> uint8_t d;</div>
<div class="line"><a id="l01876" name="l01876"></a><span class="lineno"> 1876</span> <span class="keywordflow">if</span> (bits == 2) {</div>
<div class="line"><a id="l01877" name="l01877"></a><span class="lineno"> 1877</span> d = (val &gt;&gt; (bits * i)) &amp; 0x03;</div>
<div class="line"><a id="l01878" name="l01878"></a><span class="lineno"> 1878</span> } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 4) {</div>
<div class="line"><a id="l01879" name="l01879"></a><span class="lineno"> 1879</span> d = (val &gt;&gt; (bits * i)) &amp; 0x0f;</div>
<div class="line"><a id="l01880" name="l01880"></a><span class="lineno"> 1880</span> } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (bits == 8) {</div>
<div class="line"><a id="l01881" name="l01881"></a><span class="lineno"> 1881</span> d = val;</div>
<div class="line"><a id="l01882" name="l01882"></a><span class="lineno"> 1882</span> }</div>
<div class="line"><a id="l01883" name="l01883"></a><span class="lineno"> 1883</span> out[oindex + i] = scale * d + bias;</div>
<div class="line"><a id="l01884" name="l01884"></a><span class="lineno"> 1884</span> }</div>
<div class="line"><a id="l01885" name="l01885"></a><span class="lineno"> 1885</span>}</div>
</div>
<div class="ttc" id="abackend_2metal_2kernels_2reduction_2ops_8h_html_a515b75d563a93d3c09ee677948dc83e3"><div class="ttname"><a href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a></div><div class="ttdeci">static constant constexpr const uint8_t simd_size</div><div class="ttdef"><b>Definition</b> ops.h:22</div></div>
<div class="ttc" id="abackend_2metal_2kernels_2steel_2utils_8h_html_aaf4974425147d6f26d031691e321637f"><div class="ttname"><a href="backend_2metal_2kernels_2steel_2utils_8h.html#aaf4974425147d6f26d031691e321637f">elem_to_loc_broadcast</a></div><div class="ttdeci">METAL_FUNC ulong2 elem_to_loc_broadcast(uint elem, constant const int *shape, constant const size_t *a_strides, constant const size_t *b_strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:7</div></div>
<div class="ttc" id="abackend_2metal_2kernels_2utils_8h_html_a8fd0c8fc6058e650fc99bca8b6acd7d1"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a></div><div class="ttdeci">METAL_FUNC stride_t elem_to_loc(uint elem, constant const int *shape, constant const stride_t *strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:87</div></div>
<div class="ttc" id="anamespacemetal_html"><div class="ttname"><a href="namespacemetal.html">metal</a></div><div class="ttdef"><b>Definition</b> bf16.h:265</div></div>
<div class="ttc" id="anamespacemetal_html_a048cad0aca52cb737ebf103e76bd1c49"><div class="ttname"><a href="namespacemetal.html#a048cad0aca52cb737ebf103e76bd1c49">metal::simd_max</a></div><div class="ttdeci">METAL_FUNC bfloat16_t simd_max(bfloat16_t data)</div><div class="ttdef"><b>Definition</b> bf16_math.h:392</div></div>
<div class="ttc" id="anamespacemetal_html_a46c667e169ff9d51a9204a045305442f"><div class="ttname"><a href="namespacemetal.html#a46c667e169ff9d51a9204a045305442f">metal::round</a></div><div class="ttdeci">METAL_FUNC bfloat16_t round(bfloat16_t x)</div><div class="ttdef"><b>Definition</b> bf16_math.h:234</div></div>
<div class="ttc" id="anamespacemetal_html_a6653b28c9473087141eddce39878d4d3"><div class="ttname"><a href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">metal::min</a></div><div class="ttdeci">METAL_FUNC bfloat16_t min(bfloat16_t x, bfloat16_t y)</div><div class="ttdef"><b>Definition</b> bf16_math.h:234</div></div>
<div class="ttc" id="anamespacemetal_html_a85181e37a00cb4a4217f1bb25389bce5"><div class="ttname"><a href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">metal::simd_sum</a></div><div class="ttdeci">METAL_FUNC bfloat16_t simd_sum(bfloat16_t data)</div><div class="ttdef"><b>Definition</b> bf16_math.h:392</div></div>
<div class="ttc" id="anamespacemetal_html_a853c80479ab2264d9c4587c7bcac767b"><div class="ttname"><a href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">metal::max</a></div><div class="ttdeci">METAL_FUNC bfloat16_t max(bfloat16_t x, bfloat16_t y)</div><div class="ttdef"><b>Definition</b> bf16_math.h:234</div></div>
<div class="ttc" id="anamespacemetal_html_a87c5122c60f9a12afceb9925a5b78ffb"><div class="ttname"><a href="namespacemetal.html#a87c5122c60f9a12afceb9925a5b78ffb">metal::abs</a></div><div class="ttdeci">METAL_FUNC bfloat16_t abs(bfloat16_t x)</div><div class="ttdef"><b>Definition</b> bf16_math.h:234</div></div>
<div class="ttc" id="anamespacemetal_html_ae9e2a23e00724ba2d7868bc4112b386b"><div class="ttname"><a href="namespacemetal.html#ae9e2a23e00724ba2d7868bc4112b386b">metal::simd_min</a></div><div class="ttdeci">METAL_FUNC bfloat16_t simd_min(bfloat16_t data)</div><div class="ttdef"><b>Definition</b> bf16_math.h:392</div></div>
<div class="ttc" id="anamespacemetal_html_af6e2dd7ae087aba6abac4f0350b7611c"><div class="ttname"><a href="namespacemetal.html#af6e2dd7ae087aba6abac4f0350b7611c">metal::simd_shuffle_down</a></div><div class="ttdeci">METAL_FUNC bfloat16_t simd_shuffle_down(bfloat16_t data, ushort delta)</div><div class="ttdef"><b>Definition</b> bf16_math.h:391</div></div>
<div class="ttc" id="anamespacemlx_1_1core_1_1random_html_abb895baa477f5a06b5f88e69245f1825"><div class="ttname"><a href="namespacemlx_1_1core_1_1random.html#abb895baa477f5a06b5f88e69245f1825">mlx::core::random::bits</a></div><div class="ttdeci">array bits(const std::vector&lt; int &gt; &amp;shape, int width, const std::optional&lt; array &gt; &amp;key=std::nullopt, StreamOrDevice s={})</div><div class="ttdoc">Generate an array with type uint32 filled with random bits.</div></div>
<div class="ttc" id="aquantized_8h_html_a0386011c52d03e60885a31e6fbd903dd"><div class="ttname"><a href="quantized_8h.html#a0386011c52d03e60885a31e6fbd903dd">MLX_MTL_CONST</a></div><div class="ttdeci">#define MLX_MTL_CONST</div><div class="ttdef"><b>Definition</b> quantized.h:8</div></div>
<div class="ttc" id="aquantized_8h_html_a07b26d2d0b0d65dfe925c452c453fa42"><div class="ttname"><a href="quantized_8h.html#a07b26d2d0b0d65dfe925c452c453fa42">qdot_safe</a></div><div class="ttdeci">U qdot_safe(const device uint8_t *w, const thread U *x_thread, U scale, U bias, U sum, int N)</div><div class="ttdef"><b>Definition</b> quantized.h:142</div></div>
<div class="ttc" id="aquantized_8h_html_a0ba59096494f1001c195312571523ae9"><div class="ttname"><a href="quantized_8h.html#a0ba59096494f1001c195312571523ae9">qmm_n_impl</a></div><div class="ttdeci">METAL_FUNC void qmm_n_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, threadgroup T *Xs, threadgroup T *Ws, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:879</div></div>
<div class="ttc" id="aquantized_8h_html_a1a66b061c46383952a0f067c3848971f"><div class="ttname"><a href="quantized_8h.html#a1a66b061c46383952a0f067c3848971f">bs_qmm_n</a></div><div class="ttdeci">void bs_qmm_n(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, const constant int &amp;batch_ndims, const constant int *batch_shape, const device uint32_t *lhs_indices, const device uint32_t *rhs_indices, const constant size_t *lhs_strides, const constant size_t *rhs_strides, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1678</div></div>
<div class="ttc" id="aquantized_8h_html_a2ce135e392dbf9a3e5180fb083792ed7"><div class="ttname"><a href="quantized_8h.html#a2ce135e392dbf9a3e5180fb083792ed7">qmm_n</a></div><div class="ttdeci">void qmm_n(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1368</div></div>
<div class="ttc" id="aquantized_8h_html_a47bcf4a14566e01e14bd3c155811db59"><div class="ttname"><a href="quantized_8h.html#a47bcf4a14566e01e14bd3c155811db59">affine_quantize</a></div><div class="ttdeci">void affine_quantize(const device T *w, device uint8_t *out, device T *scales, device T *biases, uint2 index, uint2 grid_dim)</div><div class="ttdef"><b>Definition</b> quantized.h:1740</div></div>
<div class="ttc" id="aquantized_8h_html_a4a8c8db7d5d480733726fd6d1a645e12"><div class="ttname"><a href="quantized_8h.html#a4a8c8db7d5d480733726fd6d1a645e12">qvm_impl</a></div><div class="ttdeci">METAL_FUNC void qvm_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:647</div></div>
<div class="ttc" id="aquantized_8h_html_a530b720e123e59d73ea89a0a2d0946b7"><div class="ttname"><a href="quantized_8h.html#a530b720e123e59d73ea89a0a2d0946b7">bs_qmv_fast</a></div><div class="ttdeci">void bs_qmv_fast(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, const constant int &amp;batch_ndims, const constant int *batch_shape, const device uint32_t *lhs_indices, const device uint32_t *rhs_indices, const constant size_t *lhs_strides, const constant size_t *rhs_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1421</div></div>
<div class="ttc" id="aquantized_8h_html_a6076203615038eb06816158f7b3869c6"><div class="ttname"><a href="quantized_8h.html#a6076203615038eb06816158f7b3869c6">affine_dequantize</a></div><div class="ttdeci">void affine_dequantize(const device uint8_t *w, const device T *scales, const device T *biases, device T *out, uint2 index, uint2 grid_dim)</div><div class="ttdef"><b>Definition</b> quantized.h:1856</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="aquantized_8h_html_a639c50a08b5cf57e8be5279a116274bd"><div class="ttname"><a href="quantized_8h.html#a639c50a08b5cf57e8be5279a116274bd">qmv</a></div><div class="ttdeci">void qmv(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1200</div></div>
<div class="ttc" id="aquantized_8h_html_a6d6e3c31e44f232e58ae9d605e1f4494"><div class="ttname"><a href="quantized_8h.html#a6d6e3c31e44f232e58ae9d605e1f4494">bs_qvm</a></div><div class="ttdeci">void bs_qvm(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, const constant int &amp;batch_ndims, const constant int *batch_shape, const device uint32_t *lhs_indices, const device uint32_t *rhs_indices, const constant size_t *lhs_strides, const constant size_t *rhs_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1543</div></div>
<div class="ttc" id="aquantized_8h_html_a7561acefd7b55e7e2b25393be08bb99c"><div class="ttname"><a href="quantized_8h.html#a7561acefd7b55e7e2b25393be08bb99c">affine_quantize_scales_biases</a></div><div class="ttdeci">void affine_quantize_scales_biases(const device T *w, const device T *scales, const device T *biases, device uint8_t *out, uint2 index, uint2 grid_dim)</div><div class="ttdef"><b>Definition</b> quantized.h:1824</div></div>
<div class="ttc" id="aquantized_8h_html_a7bd1d9f17c86c8fd34ec13678cff755f"><div class="ttname"><a href="quantized_8h.html#a7bd1d9f17c86c8fd34ec13678cff755f">qmv_fast</a></div><div class="ttdeci">void qmv_fast(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1149</div></div>
<div class="ttc" id="aquantized_8h_html_a7ce5f53a4d6d1555e9402d545408d0ad"><div class="ttname"><a href="quantized_8h.html#a7ce5f53a4d6d1555e9402d545408d0ad">qmv_quad</a></div><div class="ttdeci">void qmv_quad(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint quad_gid, uint quad_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1098</div></div>
<div class="ttc" id="aquantized_8h_html_a803e4d5a1459844ba647aea5b004e133"><div class="ttname"><a href="quantized_8h.html#a803e4d5a1459844ba647aea5b004e133">QUAD_SIZE</a></div><div class="ttdeci">static constant constexpr const int QUAD_SIZE</div><div class="ttdef"><b>Definition</b> quantized.h:11</div></div>
<div class="ttc" id="aquantized_8h_html_a8dbace41de9e1e21dd59d016db11b3e9"><div class="ttname"><a href="quantized_8h.html#a8dbace41de9e1e21dd59d016db11b3e9">load_vector</a></div><div class="ttdeci">U load_vector(const device T *x, thread U *x_thread)</div><div class="ttdef"><b>Definition</b> quantized.h:14</div></div>
<div class="ttc" id="aquantized_8h_html_a8e13c7d895624f738d2a6d9893b687fd"><div class="ttname"><a href="quantized_8h.html#a8e13c7d895624f738d2a6d9893b687fd">qmv_impl</a></div><div class="ttdeci">METAL_FUNC void qmv_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:498</div></div>
<div class="ttc" id="aquantized_8h_html_aa69e143d646fad332c1a53e8c9b337b7"><div class="ttname"><a href="quantized_8h.html#aa69e143d646fad332c1a53e8c9b337b7">load_vector_safe</a></div><div class="ttdeci">U load_vector_safe(const device T *x, thread U *x_thread, int N)</div><div class="ttdef"><b>Definition</b> quantized.h:52</div></div>
<div class="ttc" id="aquantized_8h_html_ab1ae143eba2afceb8df63f38b26f9a84"><div class="ttname"><a href="quantized_8h.html#ab1ae143eba2afceb8df63f38b26f9a84">bs_qmm_t</a></div><div class="ttdeci">void bs_qmm_t(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, const constant int &amp;batch_ndims, const constant int *batch_shape, const device uint32_t *lhs_indices, const device uint32_t *rhs_indices, const constant size_t *lhs_strides, const constant size_t *rhs_strides, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1611</div></div>
<div class="ttc" id="aquantized_8h_html_ab364d58ab652e3ad87a8f80910556071"><div class="ttname"><a href="quantized_8h.html#ab364d58ab652e3ad87a8f80910556071">qdot</a></div><div class="ttdeci">U qdot(const device uint8_t *w, const thread U *x_thread, U scale, U bias, U sum)</div><div class="ttdef"><b>Definition</b> quantized.h:99</div></div>
<div class="ttc" id="aquantized_8h_html_aba7687e6f8f1d29c0a1b2a3db150bd81"><div class="ttname"><a href="quantized_8h.html#aba7687e6f8f1d29c0a1b2a3db150bd81">qmv_fast_impl</a></div><div class="ttdeci">METAL_FUNC void qmv_fast_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:434</div></div>
<div class="ttc" id="aquantized_8h_html_abe2e3ef0ee4ec2cb61dc5330ad463d10"><div class="ttname"><a href="quantized_8h.html#abe2e3ef0ee4ec2cb61dc5330ad463d10">qmm_t</a></div><div class="ttdeci">void qmm_t(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1310</div></div>
<div class="ttc" id="aquantized_8h_html_accab1f9e17a65242347c051f98e4c0be"><div class="ttname"><a href="quantized_8h.html#accab1f9e17a65242347c051f98e4c0be">adjust_matrix_offsets</a></div><div class="ttdeci">METAL_FUNC void adjust_matrix_offsets(const device T *&amp;x, const device uint32_t *&amp;w, const device T *&amp;scales, const device T *&amp;biases, device T *&amp;y, int output_stride, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid)</div><div class="ttdef"><b>Definition</b> quantized.h:1005</div></div>
<div class="ttc" id="aquantized_8h_html_acf4c7fc77821a83b31aedfb48443d3ed"><div class="ttname"><a href="quantized_8h.html#acf4c7fc77821a83b31aedfb48443d3ed">bs_qmv</a></div><div class="ttdeci">void bs_qmv(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, const constant int &amp;batch_ndims, const constant int *batch_shape, const device uint32_t *lhs_indices, const device uint32_t *rhs_indices, const constant size_t *lhs_strides, const constant size_t *rhs_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1482</div></div>
<div class="ttc" id="aquantized_8h_html_ad5cf1cf63656bc1780685d22169cd4ef"><div class="ttname"><a href="quantized_8h.html#ad5cf1cf63656bc1780685d22169cd4ef">qmv_quad_impl</a></div><div class="ttdeci">METAL_FUNC void qmv_quad_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, constant int &amp;in_vec_size, const constant int &amp;out_vec_size, uint3 tid, uint quad_gid, uint quad_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:376</div></div>
<div class="ttc" id="aquantized_8h_html_ad84f7d5ab9e32dbbe3ca759ae5d5d5c5"><div class="ttname"><a href="quantized_8h.html#ad84f7d5ab9e32dbbe3ca759ae5d5d5c5">qvm</a></div><div class="ttdeci">void qvm(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, const constant int &amp;in_vec_size, const constant int &amp;out_vec_size, const constant int &amp;x_batch_ndims, const constant int *x_shape, const constant size_t *x_strides, const constant int &amp;w_batch_ndims, const constant int *w_shape, const constant size_t *w_strides, const constant size_t *s_strides, const constant size_t *b_strides, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:1251</div></div>
<div class="ttc" id="aquantized_8h_html_ae756f6817b584c60f5dcdd1d9c6b4f58"><div class="ttname"><a href="quantized_8h.html#ae756f6817b584c60f5dcdd1d9c6b4f58">qouter</a></div><div class="ttdeci">void qouter(const thread uint8_t *w, U x, U scale, U bias, thread U *result)</div><div class="ttdef"><b>Definition</b> quantized.h:187</div></div>
<div class="ttc" id="aquantized_8h_html_aecff265b63566d0d5689cfc4e5b037d2"><div class="ttname"><a href="quantized_8h.html#aecff265b63566d0d5689cfc4e5b037d2">dequantize</a></div><div class="ttdeci">void dequantize(const device uint8_t *w, U scale, U bias, threadgroup U *w_local)</div><div class="ttdef"><b>Definition</b> quantized.h:219</div></div>
<div class="ttc" id="aquantized_8h_html_af5750a35e8f5462218effba719f7f5b8"><div class="ttname"><a href="quantized_8h.html#af5750a35e8f5462218effba719f7f5b8">qmm_t_impl</a></div><div class="ttdeci">METAL_FUNC void qmm_t_impl(const device uint32_t *w, const device T *scales, const device T *biases, const device T *x, device T *y, threadgroup T *Xs, threadgroup T *Ws, const constant int &amp;K, const constant int &amp;N, const constant int &amp;M, uint3 tid, uint lid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> quantized.h:758</div></div>
<div class="ttc" id="astruct_limits_html"><div class="ttname"><a href="struct_limits.html">Limits</a></div><div class="ttdef"><b>Definition</b> utils.h:17</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html"><div class="ttname"><a href="struct_quantized_block_loader.html">QuantizedBlockLoader</a></div><div class="ttdef"><b>Definition</b> quantized.h:262</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a0ace7e3762ecfa5a4106e7dee7e1b6ab"><div class="ttname"><a href="struct_quantized_block_loader.html#a0ace7e3762ecfa5a4106e7dee7e1b6ab">QuantizedBlockLoader::group_stride</a></div><div class="ttdeci">const int group_stride</div><div class="ttdef"><b>Definition</b> quantized.h:282</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a1392a5278cf6e090ea80ebe7c4ac5fbb"><div class="ttname"><a href="struct_quantized_block_loader.html#a1392a5278cf6e090ea80ebe7c4ac5fbb">QuantizedBlockLoader::BCOLS_PACKED</a></div><div class="ttdeci">static constant constexpr const short BCOLS_PACKED</div><div class="ttdef"><b>Definition</b> quantized.h:274</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a17d01a6aba0833b073586ef2c09d0fbd"><div class="ttname"><a href="struct_quantized_block_loader.html#a17d01a6aba0833b073586ef2c09d0fbd">QuantizedBlockLoader::biases</a></div><div class="ttdeci">const device T * biases</div><div class="ttdef"><b>Definition</b> quantized.h:291</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a234feacde36a4afc0d740332a3769fb6"><div class="ttname"><a href="struct_quantized_block_loader.html#a234feacde36a4afc0d740332a3769fb6">QuantizedBlockLoader::group_step_cnt</a></div><div class="ttdeci">short group_step_cnt</div><div class="ttdef"><b>Definition</b> quantized.h:281</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a31e14175f3d4902d9fe5ab5a219f61ba"><div class="ttname"><a href="struct_quantized_block_loader.html#a31e14175f3d4902d9fe5ab5a219f61ba">QuantizedBlockLoader::group_steps</a></div><div class="ttdeci">static constant constexpr const short group_steps</div><div class="ttdef"><b>Definition</b> quantized.h:277</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a50821537ea747bc03295a09bb0eef475"><div class="ttname"><a href="struct_quantized_block_loader.html#a50821537ea747bc03295a09bb0eef475">QuantizedBlockLoader::thread_idx</a></div><div class="ttdeci">const short thread_idx</div><div class="ttdef"><b>Definition</b> quantized.h:284</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a6123e4a9209d6eacb58b2c2344ed1ecf"><div class="ttname"><a href="struct_quantized_block_loader.html#a6123e4a9209d6eacb58b2c2344ed1ecf">QuantizedBlockLoader::scales</a></div><div class="ttdeci">const device T * scales</div><div class="ttdef"><b>Definition</b> quantized.h:290</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a6213479f7a6d9314d8879f8856b0b6fb"><div class="ttname"><a href="struct_quantized_block_loader.html#a6213479f7a6d9314d8879f8856b0b6fb">QuantizedBlockLoader::n_reads</a></div><div class="ttdeci">static constant constexpr const short n_reads</div><div class="ttdef"><b>Definition</b> quantized.h:275</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a674138ef7c43cc45586ea9f8fd6f6bd9"><div class="ttname"><a href="struct_quantized_block_loader.html#a674138ef7c43cc45586ea9f8fd6f6bd9">QuantizedBlockLoader::next</a></div><div class="ttdeci">void next()</div><div class="ttdef"><b>Definition</b> quantized.h:354</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a699dc9aa284b8fbf870310bbb224465b"><div class="ttname"><a href="struct_quantized_block_loader.html#a699dc9aa284b8fbf870310bbb224465b">QuantizedBlockLoader::load_safe</a></div><div class="ttdeci">void load_safe(short2 src_tile_dim) const</div><div class="ttdef"><b>Definition</b> quantized.h:327</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a8050977d473d1a24fae5c833e609839e"><div class="ttname"><a href="struct_quantized_block_loader.html#a8050977d473d1a24fae5c833e609839e">QuantizedBlockLoader::src_ld</a></div><div class="ttdeci">const int src_ld</div><div class="ttdef"><b>Definition</b> quantized.h:279</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a85041d72225a2095659c70509291a906"><div class="ttname"><a href="struct_quantized_block_loader.html#a85041d72225a2095659c70509291a906">QuantizedBlockLoader::bi</a></div><div class="ttdeci">const short bi</div><div class="ttdef"><b>Definition</b> quantized.h:285</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a86009527cb4b53e4c21fd6b1f78cfefc"><div class="ttname"><a href="struct_quantized_block_loader.html#a86009527cb4b53e4c21fd6b1f78cfefc">QuantizedBlockLoader::load_unsafe</a></div><div class="ttdeci">void load_unsafe() const</div><div class="ttdef"><b>Definition</b> quantized.h:314</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a8eae73a0c04bf1e41fb96131f6aa500d"><div class="ttname"><a href="struct_quantized_block_loader.html#a8eae73a0c04bf1e41fb96131f6aa500d">QuantizedBlockLoader::pack_factor</a></div><div class="ttdeci">static constant constexpr const short pack_factor</div><div class="ttdef"><b>Definition</b> quantized.h:273</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_a9857214690fe6abad0e19d1045152f83"><div class="ttname"><a href="struct_quantized_block_loader.html#a9857214690fe6abad0e19d1045152f83">QuantizedBlockLoader::dst</a></div><div class="ttdeci">threadgroup T * dst</div><div class="ttdef"><b>Definition</b> quantized.h:288</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_ac3f651c1a645291d1037a2cc8ded2320"><div class="ttname"><a href="struct_quantized_block_loader.html#ac3f651c1a645291d1037a2cc8ded2320">QuantizedBlockLoader::tile_stride</a></div><div class="ttdeci">const int tile_stride</div><div class="ttdef"><b>Definition</b> quantized.h:280</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_ad85c6b7e07c81307b3b91eb4dd7be30b"><div class="ttname"><a href="struct_quantized_block_loader.html#ad85c6b7e07c81307b3b91eb4dd7be30b">QuantizedBlockLoader::src</a></div><div class="ttdeci">const device uint32_t * src</div><div class="ttdef"><b>Definition</b> quantized.h:289</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_ae2add92b2aaf3414e91f0470b9b0cc00"><div class="ttname"><a href="struct_quantized_block_loader.html#ae2add92b2aaf3414e91f0470b9b0cc00">QuantizedBlockLoader::bj</a></div><div class="ttdeci">const short bj</div><div class="ttdef"><b>Definition</b> quantized.h:286</div></div>
<div class="ttc" id="astruct_quantized_block_loader_html_af59b054750a65e7e79c1cd05c4acac93"><div class="ttname"><a href="struct_quantized_block_loader.html#af59b054750a65e7e79c1cd05c4acac93">QuantizedBlockLoader::QuantizedBlockLoader</a></div><div class="ttdeci">QuantizedBlockLoader(const device uint32_t *src_, const device T *scales_, const device T *biases_, const int src_ld_, threadgroup T *dst_, ushort simd_group_id, ushort simd_lane_id)</div><div class="ttdef"><b>Definition</b> quantized.h:293</div></div>
<div class="ttc" id="astructmlx_1_1steel_1_1_block_loader_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_block_loader.html">mlx::steel::BlockLoader</a></div><div class="ttdef"><b>Definition</b> loader.h:25</div></div>
</div><!-- fragment --></div><!-- contents -->
<!-- start footer part -->
<hr class="footer"/><address class="footer"><small>
Generated by&#160;<a href="https://www.doxygen.org/index.html"><img class="footer" src="doxygen.svg" width="104" height="31" alt="doxygen"/></a> 1.12.0
</small></address>
</div><!-- doc-content -->
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink