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
5a14f96fb92403f986ae661514421299f09fef06
mlx/docs/build/html/sdpa__vector_8h_source.html
CircleCI Docs 3bac79c1d6 rebase
2025-10-17 19:15:00 +00:00

455 lines
57 KiB
HTML
Raw Blame History

<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "https://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">
<html xmlns="http://www.w3.org/1999/xhtml" lang="en-US">
<head>
<meta http-equiv="Content-Type" content="text/xhtml;charset=UTF-8"/>
<meta http-equiv="X-UA-Compatible" content="IE=11"/>
<meta name="generator" content="Doxygen 1.13.2"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>MLX: mlx/backend/metal/kernels/sdpa_vector.h Source File</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="jquery.js"></script>
<script type="text/javascript" src="dynsections.js"></script>
<script type="text/javascript" src="clipboard.js"></script>
<link href="navtree.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="navtreedata.js"></script>
<script type="text/javascript" src="navtree.js"></script>
<script type="text/javascript" src="resize.js"></script>
<script type="text/javascript" src="cookie.js"></script>
<link href="search/search.css" rel="stylesheet" type="text/css"/>
<script type="text/javascript" src="search/searchdata.js"></script>
<script type="text/javascript" src="search/search.js"></script>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function() { init_search(); });
/* @license-end */
</script>
<link href="doxygen.css" rel="stylesheet" type="text/css" />
</head>
<body>
<div id="top"><!-- do not remove this div, it is closed by doxygen! -->
<div id="titlearea">
<table cellspacing="0" cellpadding="0">
<tbody>
<tr id="projectrow">
<td id="projectalign">
<div id="projectname">MLX
</div>
</td>
<td> <div id="MSearchBox" class="MSearchBoxInactive">
<span class="left">
<span id="MSearchSelect" onmouseover="return searchBox.OnSearchSelectShow()" onmouseout="return searchBox.OnSearchSelectHide()">&#160;</span>
<input type="text" id="MSearchField" value="" placeholder="Search" accesskey="S"
onfocus="searchBox.OnSearchFieldFocus(true)"
onblur="searchBox.OnSearchFieldFocus(false)"
onkeyup="searchBox.OnSearchFieldChange(event)"/>
</span><span class="right">
<a id="MSearchClose" href="javascript:searchBox.CloseResultsWindow()"><img id="MSearchCloseImg" border="0" src="search/close.svg" alt=""/></a>
</span>
</div>
</td>
</tr>
</tbody>
</table>
</div>
<!-- end header part -->
<!-- Generated by Doxygen 1.13.2 -->
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
var searchBox = new SearchBox("searchBox", "search/",'.html');
/* @license-end */
</script>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function() { codefold.init(0); });
/* @license-end */
</script>
</div><!-- top -->
<div id="side-nav" class="ui-resizable side-nav-resizable">
<div id="nav-tree">
<div id="nav-tree-contents">
<div id="nav-sync" class="sync"></div>
</div>
</div>
<div id="splitbar" style="-moz-user-select:none;"
class="ui-resizable-handle">
</div>
</div>
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
$(function(){initNavTree('sdpa__vector_8h_source.html',''); initResizable(true); });
/* @license-end */
</script>
<div id="doc-content">
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
onmouseout="return searchBox.OnSearchSelectHide()"
onkeydown="return searchBox.OnSearchSelectKey(event)">
</div>
<!-- iframe showing the search results (closed by default) -->
<div id="MSearchResultsWindow">
<div id="MSearchResults">
<div class="SRPage">
<div id="SRIndex">
<div id="SRResults"></div>
<div class="SRStatus" id="Loading">Loading...</div>
<div class="SRStatus" id="Searching">Searching...</div>
<div class="SRStatus" id="NoMatches">No Matches</div>
</div>
</div>
</div>
</div>
<div class="header">
<div class="headertitle"><div class="title">sdpa_vector.h</div></div>
</div><!--header-->
<div class="contents">
<a href="sdpa__vector_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 © 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> </div>
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span><span class="keyword">using namespace </span><a class="code hl_namespace" href="namespacemetal.html">metal</a>;</div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span> </div>
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"><a class="line" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81"> 7</a></span>constant <span class="keywordtype">bool</span> <a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a> [[function_constant(20)]];</div>
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span> </div>
<div class="line"><a id="l00009" name="l00009"></a><span class="lineno"> 9</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> D, <span class="keywordtype">int</span> V = D&gt;</div>
<div class="foldopen" id="foldopen00010" data-start="{" data-end="}">
<div class="line"><a id="l00010" name="l00010"></a><span class="lineno"><a class="line" href="sdpa__vector_8h.html#a826f7a3c7ab843abc0842241db3e57b3"> 10</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="sdpa__vector_8h.html#a826f7a3c7ab843abc0842241db3e57b3">sdpa_vector</a>(</div>
<div class="line"><a id="l00011" name="l00011"></a><span class="lineno"> 11</span> <span class="keyword">const</span> device T* queries [[buffer(0)]],</div>
<div class="line"><a id="l00012" name="l00012"></a><span class="lineno"> 12</span> <span class="keyword">const</span> device T* keys [[buffer(1)]],</div>
<div class="line"><a id="l00013" name="l00013"></a><span class="lineno"> 13</span> <span class="keyword">const</span> device T* values [[buffer(2)]],</div>
<div class="line"><a id="l00014" name="l00014"></a><span class="lineno"> 14</span> device T* out [[buffer(3)]],</div>
<div class="line"><a id="l00015" name="l00015"></a><span class="lineno"> 15</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; gqa_factor,</div>
<div class="line"><a id="l00016" name="l00016"></a><span class="lineno"> 16</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N,</div>
<div class="line"><a id="l00017" name="l00017"></a><span class="lineno"> 17</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>&amp; k_stride,</div>
<div class="line"><a id="l00018" name="l00018"></a><span class="lineno"> 18</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>&amp; v_stride,</div>
<div class="line"><a id="l00019" name="l00019"></a><span class="lineno"> 19</span> <span class="keyword">const</span> constant <span class="keywordtype">float</span>&amp; scale,</div>
<div class="line"><a id="l00020" name="l00020"></a><span class="lineno"> 20</span> <span class="keyword">const</span> device <span class="keywordtype">bool</span>* mask [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00021" name="l00021"></a><span class="lineno"> 21</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; mask_seq_stride [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00022" name="l00022"></a><span class="lineno"> 22</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; mask_head_stride [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00023" name="l00023"></a><span class="lineno"> 23</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00024" name="l00024"></a><span class="lineno"> 24</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN = 32;</div>
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BD = 32;</div>
<div class="line"><a id="l00028" name="l00028"></a><span class="lineno"> 28</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> qk_per_thread = D / BD;</div>
<div class="line"><a id="l00029" name="l00029"></a><span class="lineno"> 29</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> v_per_thread = V / BD;</div>
<div class="line"><a id="l00030" name="l00030"></a><span class="lineno"> 30</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> inner_k_stride = BN * D;</div>
<div class="line"><a id="l00031" name="l00031"></a><span class="lineno"> 31</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> inner_v_stride = BN * V;</div>
<div class="line"><a id="l00032" name="l00032"></a><span class="lineno"> 32</span> </div>
<div class="line"><a id="l00033" name="l00033"></a><span class="lineno"> 33</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00034" name="l00034"></a><span class="lineno"> 34</span> </div>
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span> thread U q[qk_per_thread];</div>
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> thread U k[qk_per_thread];</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> thread U o[v_per_thread];</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> threadgroup U outputs[BN * BD];</div>
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> threadgroup U max_scores[BN];</div>
<div class="line"><a id="l00041" name="l00041"></a><span class="lineno"> 41</span> threadgroup U sum_exp_scores[BN];</div>
<div class="line"><a id="l00042" name="l00042"></a><span class="lineno"> 42</span> </div>
<div class="line"><a id="l00043" name="l00043"></a><span class="lineno"> 43</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00044" name="l00044"></a><span class="lineno"> 44</span> <span class="keyword">const</span> <span class="keywordtype">int</span> head_idx = tid.y;</div>
<div class="line"><a id="l00045" name="l00045"></a><span class="lineno"> 45</span> <span class="keyword">const</span> <span class="keywordtype">int</span> kv_head_idx = head_idx / gqa_factor;</div>
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> queries += head_idx * D + simd_lid * qk_per_thread;</div>
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> keys += kv_head_idx * k_stride + simd_gid * D + simd_lid * qk_per_thread;</div>
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> values += kv_head_idx * v_stride + simd_gid * V + simd_lid * v_per_thread;</div>
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>) {</div>
<div class="line"><a id="l00050" name="l00050"></a><span class="lineno"> 50</span> mask += head_idx * mask_head_stride + simd_gid * mask_seq_stride;</div>
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span> }</div>
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"> 52</span> out += head_idx * V + simd_gid * v_per_thread;</div>
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> </div>
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> <span class="comment">// Read the query and 0 the output accumulator</span></div>
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; qk_per_thread; i++) {</div>
<div class="line"><a id="l00056" name="l00056"></a><span class="lineno"> 56</span> q[i] = <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(scale) * queries[i];</div>
<div class="line"><a id="l00057" name="l00057"></a><span class="lineno"> 57</span> }</div>
<div class="line"><a id="l00058" name="l00058"></a><span class="lineno"> 58</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span> o[i] = 0;</div>
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span> }</div>
<div class="line"><a id="l00061" name="l00061"></a><span class="lineno"> 61</span> </div>
<div class="line"><a id="l00062" name="l00062"></a><span class="lineno"> 62</span> U max_score = -INFINITY;</div>
<div class="line"><a id="l00063" name="l00063"></a><span class="lineno"> 63</span> U sum_exp_score = 0;</div>
<div class="line"><a id="l00064" name="l00064"></a><span class="lineno"> 64</span> </div>
<div class="line"><a id="l00065" name="l00065"></a><span class="lineno"> 65</span> <span class="comment">// For each key</span></div>
<div class="line"><a id="l00066" name="l00066"></a><span class="lineno"> 66</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = simd_gid; i &lt; N; i += BN) {</div>
<div class="line"><a id="l00067" name="l00067"></a><span class="lineno"> 67</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a> || mask[0]) {</div>
<div class="line"><a id="l00068" name="l00068"></a><span class="lineno"> 68</span> <span class="comment">// Read the key</span></div>
<div class="line"><a id="l00069" name="l00069"></a><span class="lineno"> 69</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j &lt; qk_per_thread; j++) {</div>
<div class="line"><a id="l00070" name="l00070"></a><span class="lineno"> 70</span> k[j] = keys[j];</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> </div>
<div class="line"><a id="l00073" name="l00073"></a><span class="lineno"> 73</span> <span class="comment">// Compute the i-th score</span></div>
<div class="line"><a id="l00074" name="l00074"></a><span class="lineno"> 74</span> U score = 0;</div>
<div class="line"><a id="l00075" name="l00075"></a><span class="lineno"> 75</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j &lt; qk_per_thread; j++) {</div>
<div class="line"><a id="l00076" name="l00076"></a><span class="lineno"> 76</span> score += q[j] * k[j];</div>
<div class="line"><a id="l00077" name="l00077"></a><span class="lineno"> 77</span> }</div>
<div class="line"><a id="l00078" name="l00078"></a><span class="lineno"> 78</span> score = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(score);</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="comment">// Update the accumulators</span></div>
<div class="line"><a id="l00081" name="l00081"></a><span class="lineno"> 81</span> U new_max = <a class="code hl_function" href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">max</a>(max_score, score);</div>
<div class="line"><a id="l00082" name="l00082"></a><span class="lineno"> 82</span> U factor = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_score - new_max);</div>
<div class="line"><a id="l00083" name="l00083"></a><span class="lineno"> 83</span> U exp_score = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(score - new_max);</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> max_score = new_max;</div>
<div class="line"><a id="l00086" name="l00086"></a><span class="lineno"> 86</span> sum_exp_score = sum_exp_score * factor + exp_score;</div>
<div class="line"><a id="l00087" name="l00087"></a><span class="lineno"> 87</span> </div>
<div class="line"><a id="l00088" name="l00088"></a><span class="lineno"> 88</span> <span class="comment">// Update the output accumulator</span></div>
<div class="line"><a id="l00089" name="l00089"></a><span class="lineno"> 89</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j &lt; v_per_thread; j++) {</div>
<div class="line"><a id="l00090" name="l00090"></a><span class="lineno"> 90</span> o[j] = o[j] * factor + exp_score * values[j];</div>
<div class="line"><a id="l00091" name="l00091"></a><span class="lineno"> 91</span> }</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> <span class="comment">// Move the pointers to the next kv</span></div>
<div class="line"><a id="l00095" name="l00095"></a><span class="lineno"> 95</span> keys += inner_k_stride;</div>
<div class="line"><a id="l00096" name="l00096"></a><span class="lineno"> 96</span> values += inner_v_stride;</div>
<div class="line"><a id="l00097" name="l00097"></a><span class="lineno"> 97</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>) {</div>
<div class="line"><a id="l00098" name="l00098"></a><span class="lineno"> 98</span> mask += BN * mask_seq_stride;</div>
<div class="line"><a id="l00099" name="l00099"></a><span class="lineno"> 99</span> }</div>
<div class="line"><a id="l00100" name="l00100"></a><span class="lineno"> 100</span> }</div>
<div class="line"><a id="l00101" name="l00101"></a><span class="lineno"> 101</span> </div>
<div class="line"><a id="l00102" name="l00102"></a><span class="lineno"> 102</span> <span class="comment">// Each thread has a partial part of the output so we need to combine them.</span></div>
<div class="line"><a id="l00103" name="l00103"></a><span class="lineno"> 103</span> </div>
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> <span class="comment">// First let&#39;s communicate the max and sum_exp</span></div>
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> max_scores[simd_gid] = max_score;</div>
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> sum_exp_scores[simd_gid] = sum_exp_score;</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> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00110" name="l00110"></a><span class="lineno"> 110</span> max_score = max_scores[simd_lid];</div>
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> U new_max = <a class="code hl_function" href="namespacemetal.html#a048cad0aca52cb737ebf103e76bd1c49">simd_max</a>(max_score);</div>
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> U factor = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_score - new_max);</div>
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> sum_exp_score = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(sum_exp_scores[simd_lid] * factor);</div>
<div class="line"><a id="l00114" name="l00114"></a><span class="lineno"> 114</span> </div>
<div class="line"><a id="l00115" name="l00115"></a><span class="lineno"> 115</span> <span class="comment">// Now we need to aggregate all the outputs</span></div>
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00117" name="l00117"></a><span class="lineno"> 117</span> outputs[simd_lid * BD + simd_gid] = o[i];</div>
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> o[i] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(outputs[simd_gid * BD + simd_lid] * factor) / sum_exp_score;</div>
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> }</div>
<div class="line"><a id="l00122" name="l00122"></a><span class="lineno"> 122</span> </div>
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span> <span class="comment">// And write the output</span></div>
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> out[i] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(o[i]);</div>
<div class="line"><a id="l00127" name="l00127"></a><span class="lineno"> 127</span> }</div>
<div class="line"><a id="l00128" name="l00128"></a><span class="lineno"> 128</span> }</div>
<div class="line"><a id="l00129" name="l00129"></a><span class="lineno"> 129</span>}</div>
</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><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> D, <span class="keywordtype">int</span> V = D&gt;</div>
<div class="foldopen" id="foldopen00132" data-start="{" data-end="}">
<div class="line"><a id="l00132" name="l00132"></a><span class="lineno"><a class="line" href="sdpa__vector_8h.html#aae1a2f23b03e24734805b08ebc5c1a59"> 132</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="sdpa__vector_8h.html#aae1a2f23b03e24734805b08ebc5c1a59">sdpa_vector_2pass_1</a>(</div>
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> <span class="keyword">const</span> device T* queries [[buffer(0)]],</div>
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> <span class="keyword">const</span> device T* keys [[buffer(1)]],</div>
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> <span class="keyword">const</span> device T* values [[buffer(2)]],</div>
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> device <span class="keywordtype">float</span>* out [[buffer(3)]],</div>
<div class="line"><a id="l00137" name="l00137"></a><span class="lineno"> 137</span> device <span class="keywordtype">float</span>* sums [[buffer(4)]],</div>
<div class="line"><a id="l00138" name="l00138"></a><span class="lineno"> 138</span> device <span class="keywordtype">float</span>* maxs [[buffer(5)]],</div>
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; gqa_factor,</div>
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; N,</div>
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>&amp; k_stride,</div>
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"> 142</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>&amp; v_stride,</div>
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> <span class="keyword">const</span> constant <span class="keywordtype">float</span>&amp; scale,</div>
<div class="line"><a id="l00144" name="l00144"></a><span class="lineno"> 144</span> <span class="keyword">const</span> device <span class="keywordtype">bool</span>* mask [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00145" name="l00145"></a><span class="lineno"> 145</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; mask_seq_stride [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>&amp; mask_head_stride [[function_constant(<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>)]],</div>
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN = 8;</div>
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BD = 32;</div>
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> qk_per_thread = D / BD;</div>
<div class="line"><a id="l00153" name="l00153"></a><span class="lineno"> 153</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> v_per_thread = V / BD;</div>
<div class="line"><a id="l00154" name="l00154"></a><span class="lineno"> 154</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> inner_k_stride = BN * D;</div>
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> inner_v_stride = BN * V;</div>
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> blocks = 32;</div>
<div class="line"><a id="l00157" name="l00157"></a><span class="lineno"> 157</span> </div>
<div class="line"><a id="l00158" name="l00158"></a><span class="lineno"> 158</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> </div>
<div class="line"><a id="l00160" name="l00160"></a><span class="lineno"> 160</span> thread U q[qk_per_thread];</div>
<div class="line"><a id="l00161" name="l00161"></a><span class="lineno"> 161</span> thread U k[qk_per_thread];</div>
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span> thread U o[v_per_thread];</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> threadgroup U outputs[BN * BD];</div>
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span> threadgroup U max_scores[BN];</div>
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"> 166</span> threadgroup U sum_exp_scores[BN];</div>
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> </div>
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> <span class="keyword">const</span> <span class="keywordtype">int</span> block_idx = tid.z;</div>
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> <span class="keyword">const</span> <span class="keywordtype">int</span> head_idx = tid.y;</div>
<div class="line"><a id="l00171" name="l00171"></a><span class="lineno"> 171</span> <span class="keyword">const</span> <span class="keywordtype">int</span> kv_head_idx = head_idx / gqa_factor;</div>
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> queries += head_idx * D + simd_lid * qk_per_thread;</div>
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> keys += kv_head_idx * k_stride + (block_idx * BN + simd_gid) * D +</div>
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> simd_lid * qk_per_thread;</div>
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> values += kv_head_idx * v_stride + (block_idx * BN + simd_gid) * V +</div>
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> simd_lid * v_per_thread;</div>
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> out += head_idx * blocks * V + block_idx * V + simd_lid * v_per_thread;</div>
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>) {</div>
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> mask += head_idx * mask_head_stride +</div>
<div class="line"><a id="l00180" name="l00180"></a><span class="lineno"> 180</span> (block_idx * BN + simd_gid) * mask_seq_stride;</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> sums += head_idx * blocks + block_idx;</div>
<div class="line"><a id="l00183" name="l00183"></a><span class="lineno"> 183</span> maxs += head_idx * blocks + block_idx;</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="comment">// Read the query and 0 the output accumulator</span></div>
<div class="line"><a id="l00186" name="l00186"></a><span class="lineno"> 186</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; qk_per_thread; i++) {</div>
<div class="line"><a id="l00187" name="l00187"></a><span class="lineno"> 187</span> q[i] = <span class="keyword">static_cast&lt;</span>U<span class="keyword">&gt;</span>(scale) * queries[i];</div>
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span> }</div>
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00190" name="l00190"></a><span class="lineno"> 190</span> o[i] = 0;</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> </div>
<div class="line"><a id="l00193" name="l00193"></a><span class="lineno"> 193</span> U max_score = -1e9;</div>
<div class="line"><a id="l00194" name="l00194"></a><span class="lineno"> 194</span> U sum_exp_score = 0;</div>
<div class="line"><a id="l00195" name="l00195"></a><span class="lineno"> 195</span> </div>
<div class="line"><a id="l00196" name="l00196"></a><span class="lineno"> 196</span> <span class="comment">// For each key</span></div>
<div class="line"><a id="l00197" name="l00197"></a><span class="lineno"> 197</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = block_idx * BN + simd_gid; i &lt; N; i += blocks * BN) {</div>
<div class="line"><a id="l00198" name="l00198"></a><span class="lineno"> 198</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a> || mask[0]) {</div>
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"> 199</span> <span class="comment">// Read the key</span></div>
<div class="line"><a id="l00200" name="l00200"></a><span class="lineno"> 200</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; qk_per_thread; i++) {</div>
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span> k[i] = keys[i];</div>
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> }</div>
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> </div>
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> <span class="comment">// Compute the i-th score</span></div>
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> U score = 0;</div>
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; qk_per_thread; i++) {</div>
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> score += q[i] * k[i];</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> score = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(score);</div>
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> </div>
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> <span class="comment">// Update the accumulators</span></div>
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> U new_max = <a class="code hl_function" href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">max</a>(max_score, score);</div>
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> U factor = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_score - new_max);</div>
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> U exp_score = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(score - new_max);</div>
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> </div>
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> max_score = new_max;</div>
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> sum_exp_score = sum_exp_score * factor + exp_score;</div>
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> </div>
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"> 219</span> <span class="comment">// Update the output accumulator</span></div>
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> o[i] = o[i] * factor + exp_score * values[i];</div>
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</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> </div>
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> <span class="comment">// Move the pointers to the next kv</span></div>
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> keys += blocks * inner_k_stride;</div>
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> values += blocks * inner_v_stride;</div>
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> <span class="keywordflow">if</span> (<a class="code hl_variable" href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a>) {</div>
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> mask += BN * blocks * mask_seq_stride;</div>
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> }</div>
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> }</div>
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> </div>
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> <span class="comment">// Each thread has a partial part of the output so we need to combine them.</span></div>
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> </div>
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> <span class="comment">// First let&#39;s communicate the max and sum_exp</span></div>
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> max_scores[simd_gid] = max_score;</div>
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> sum_exp_scores[simd_gid] = sum_exp_score;</div>
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> }</div>
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> max_score = (simd_lid &lt; BN) ? max_scores[simd_lid] : -1e9;</div>
<div class="line"><a id="l00242" name="l00242"></a><span class="lineno"> 242</span> U new_max = <a class="code hl_function" href="namespacemetal.html#a048cad0aca52cb737ebf103e76bd1c49">simd_max</a>(max_score);</div>
<div class="line"><a id="l00243" name="l00243"></a><span class="lineno"> 243</span> U factor = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_score - new_max);</div>
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> sum_exp_score = (simd_lid &lt; BN) ? sum_exp_scores[simd_lid] : 0;</div>
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> sum_exp_score = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(sum_exp_score * factor);</div>
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> </div>
<div class="line"><a id="l00247" name="l00247"></a><span class="lineno"> 247</span> <span class="comment">// Write the sum and new max</span></div>
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> <span class="keywordflow">if</span> (simd_gid == 0) {</div>
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> sums[0] = sum_exp_score;</div>
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> maxs[0] = new_max;</div>
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> }</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="comment">// Now we need to aggregate all the outputs</span></div>
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; v_per_thread; i++) {</div>
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> outputs[simd_lid * BN + simd_gid] =</div>
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> o[i] * <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_scores[simd_gid] - new_max);</div>
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> </div>
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> <span class="comment">// And write the output</span></div>
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> <span class="keywordflow">if</span> (simd_gid == 0) {</div>
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> U output = outputs[simd_lid * BN];</div>
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"> 262</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 1; j &lt; BN; j++) {</div>
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> output += outputs[simd_lid * BN + j];</div>
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> }</div>
<div class="line"><a id="l00265" name="l00265"></a><span class="lineno"> 265</span> out[i] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(output);</div>
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"> 266</span> }</div>
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> }</div>
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span>}</div>
</div>
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> </div>
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keywordtype">int</span> D&gt;</div>
<div class="foldopen" id="foldopen00272" data-start="{" data-end="}">
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"><a class="line" href="sdpa__vector_8h.html#a1368cf3618a4e03dbf743b3463205efe"> 272</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="sdpa__vector_8h.html#a1368cf3618a4e03dbf743b3463205efe">sdpa_vector_2pass_2</a>(</div>
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> <span class="keyword">const</span> device <span class="keywordtype">float</span>* partials [[buffer(0)]],</div>
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"> 274</span> <span class="keyword">const</span> device <span class="keywordtype">float</span>* sums [[buffer(1)]],</div>
<div class="line"><a id="l00275" name="l00275"></a><span class="lineno"> 275</span> <span class="keyword">const</span> device <span class="keywordtype">float</span>* maxs [[buffer(2)]],</div>
<div class="line"><a id="l00276" name="l00276"></a><span class="lineno"> 276</span> device T* out [[buffer(3)]],</div>
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"> 277</span> uint3 tid [[threadgroup_position_in_grid]],</div>
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> uint simd_gid [[simdgroup_index_in_threadgroup]],</div>
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"> 279</span> uint simd_lid [[thread_index_in_simdgroup]]) {</div>
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"> 280</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BN = 32;</div>
<div class="line"><a id="l00281" name="l00281"></a><span class="lineno"> 281</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> BD = 32;</div>
<div class="line"><a id="l00282" name="l00282"></a><span class="lineno"> 282</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> elem_per_thread = D / BD;</div>
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> blocks = 32;</div>
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"> 284</span> </div>
<div class="line"><a id="l00285" name="l00285"></a><span class="lineno"> 285</span> <span class="keyword">typedef</span> <span class="keywordtype">float</span> U;</div>
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"> 286</span> </div>
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> thread U o[elem_per_thread];</div>
<div class="line"><a id="l00288" name="l00288"></a><span class="lineno"> 288</span> threadgroup U outputs[BN * BD];</div>
<div class="line"><a id="l00289" name="l00289"></a><span class="lineno"> 289</span> </div>
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"> 290</span> <span class="comment">// Adjust positions</span></div>
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"> 291</span> <span class="keyword">const</span> <span class="keywordtype">int</span> head_idx = tid.y;</div>
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> partials += head_idx * blocks * D + simd_gid * D + simd_lid * elem_per_thread;</div>
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"> 293</span> sums += head_idx * blocks;</div>
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> maxs += head_idx * blocks;</div>
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> out += head_idx * D + simd_gid * elem_per_thread;</div>
<div class="line"><a id="l00296" name="l00296"></a><span class="lineno"> 296</span> </div>
<div class="line"><a id="l00297" name="l00297"></a><span class="lineno"> 297</span> <span class="comment">// First everybody reads the max and sum_exp</span></div>
<div class="line"><a id="l00298" name="l00298"></a><span class="lineno"> 298</span> U max_score = maxs[simd_lid];</div>
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> U new_max = <a class="code hl_function" href="namespacemetal.html#a048cad0aca52cb737ebf103e76bd1c49">simd_max</a>(max_score);</div>
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> U factor = <a class="code hl_function" href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">fast::exp</a>(max_score - new_max);</div>
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> U sum_exp_score = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(sums[simd_lid] * factor);</div>
<div class="line"><a id="l00302" name="l00302"></a><span class="lineno"> 302</span> </div>
<div class="line"><a id="l00303" name="l00303"></a><span class="lineno"> 303</span> <span class="comment">// Now read the block into registers and then use shared memory to transpose</span></div>
<div class="line"><a id="l00304" name="l00304"></a><span class="lineno"> 304</span> <span class="comment">// it</span></div>
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; elem_per_thread; i++) {</div>
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> o[i] = partials[i];</div>
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> }</div>
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; elem_per_thread; i++) {</div>
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> outputs[simd_lid * BD + simd_gid] = o[i];</div>
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> o[i] = <a class="code hl_function" href="namespacemetal.html#a85181e37a00cb4a4217f1bb25389bce5">simd_sum</a>(outputs[simd_gid * BD + simd_lid] * factor) / sum_exp_score;</div>
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> }</div>
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> </div>
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> <span class="comment">// And write the output</span></div>
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> <span class="keywordflow">if</span> (simd_lid == 0) {</div>
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; elem_per_thread; i++) {</div>
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> out[i] = <span class="keyword">static_cast&lt;</span>T<span class="keyword">&gt;</span>(o[i]);</div>
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> }</div>
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> }</div>
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span>}</div>
</div>
<div class="ttc" id="anamespacemetal_1_1fast_html_ad3dbd387b63373c29e3449609f763ede"><div class="ttname"><a href="namespacemetal_1_1fast.html#ad3dbd387b63373c29e3449609f763ede">metal::fast::exp</a></div><div class="ttdeci">METAL_FUNC bfloat16_t exp(bfloat16_t x)</div><div class="ttdef"><b>Definition</b> bf16_math.h:240</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_math.h:226</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:378</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:378</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:232</div></div>
<div class="ttc" id="asdpa__vector_8h_html_a1368cf3618a4e03dbf743b3463205efe"><div class="ttname"><a href="sdpa__vector_8h.html#a1368cf3618a4e03dbf743b3463205efe">sdpa_vector_2pass_2</a></div><div class="ttdeci">void sdpa_vector_2pass_2(const device float *partials, const device float *sums, const device float *maxs, device T *out, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> sdpa_vector.h:272</div></div>
<div class="ttc" id="asdpa__vector_8h_html_a6ed0dd113fe7d471fc0b869b8c028c81"><div class="ttname"><a href="sdpa__vector_8h.html#a6ed0dd113fe7d471fc0b869b8c028c81">has_mask</a></div><div class="ttdeci">constant bool has_mask</div><div class="ttdef"><b>Definition</b> sdpa_vector.h:7</div></div>
<div class="ttc" id="asdpa__vector_8h_html_a826f7a3c7ab843abc0842241db3e57b3"><div class="ttname"><a href="sdpa__vector_8h.html#a826f7a3c7ab843abc0842241db3e57b3">sdpa_vector</a></div><div class="ttdeci">void sdpa_vector(const device T *queries, const device T *keys, const device T *values, device T *out, const constant int &amp;gqa_factor, const constant int &amp;N, const constant size_t &amp;k_stride, const constant size_t &amp;v_stride, const constant float &amp;scale, const device bool *mask, const constant int &amp;mask_seq_stride, const constant int &amp;mask_head_stride, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> sdpa_vector.h:10</div></div>
<div class="ttc" id="asdpa__vector_8h_html_aae1a2f23b03e24734805b08ebc5c1a59"><div class="ttname"><a href="sdpa__vector_8h.html#aae1a2f23b03e24734805b08ebc5c1a59">sdpa_vector_2pass_1</a></div><div class="ttdeci">void sdpa_vector_2pass_1(const device T *queries, const device T *keys, const device T *values, device float *out, device float *sums, device float *maxs, const constant int &amp;gqa_factor, const constant int &amp;N, const constant size_t &amp;k_stride, const constant size_t &amp;v_stride, const constant float &amp;scale, const device bool *mask, const constant int &amp;mask_seq_stride, const constant int &amp;mask_head_stride, uint3 tid, uint simd_gid, uint simd_lid)</div><div class="ttdef"><b>Definition</b> sdpa_vector.h:132</div></div>
</div><!-- fragment --></div><!-- contents -->
</div><!-- doc-content -->
<!-- start footer part -->
<div id="nav-path" class="navpath"><!-- id is needed for treeview function! -->
<ul>
<li class="navelem"><a class="el" href="dir_938ab0ecf10b8b860ff766c820f665fd.html">mlx</a></li><li class="navelem"><a class="el" href="dir_1d446c9bd3c99228254c9484e0bc5c06.html">backend</a></li><li class="navelem"><a class="el" href="dir_d0c977ea65824390717cdb7efc36c157.html">metal</a></li><li class="navelem"><a class="el" href="dir_70a37effa88bcbd6b791977fa1e64356.html">kernels</a></li><li class="navelem"><a class="el" href="sdpa__vector_8h.html">sdpa_vector.h</a></li>
<li class="footer">Generated by <a href="https://www.doxygen.org/index.html"><img class="footer" src="doxygen.svg" width="104" height="31" alt="doxygen"/></a> 1.13.2 </li>
</ul>
</div>
</body>
</html>
Reference in New Issue View Git Blame Copy Permalink