This commit is contained in:
CircleCI Docs
2025-01-09 21:56:20 +00:00
parent 04b749a588
commit d8d647015b
2642 changed files with 137687 additions and 70861 deletions

View File

@@ -3,7 +3,7 @@
<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="generator" content="Doxygen 1.13.1"/>
<meta name="viewport" content="width=device-width, initial-scale=1"/>
<title>MLX: mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_masked.h Source File</title>
<link href="tabs.css" rel="stylesheet" type="text/css"/>
@@ -11,11 +11,18 @@
<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>
@@ -28,12 +35,24 @@
<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 1.12.0 -->
<!-- Generated by Doxygen 1.13.1 -->
<script type="text/javascript">
/* @license magnet:?xt=urn:btih:d3d9a9a6595521f9666a5e94cc830dab83b65699&amp;dn=expat.txt MIT */
var searchBox = new SearchBox("searchBox", "search/",'.html');
@@ -44,22 +63,23 @@ var searchBox = new SearchBox("searchBox", "search/",'.html');
$(function() { codefold.init(0); });
/* @license-end */
</script>
<script type="text/javascript" src="menudata.js"></script>
<script type="text/javascript" src="menu.js"></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() {
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); });
$(function(){initNavTree('steel__gemm__masked_8h_source.html',''); initResizable(true); });
/* @license-end */
</script>
<div id="doc-content">
<!-- window showing the filter options -->
<div id="MSearchSelectWindow"
onmouseover="return searchBox.OnSearchSelectShow()"
@@ -81,12 +101,6 @@ $(function(){ initResizable(false); });
</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><li class="navelem"><a class="el" href="dir_76215a6c54e2b67053e723fc2395583c.html">steel</a></li><li class="navelem"><a class="el" href="dir_6768c99e6145fb9510ccdb40db8ede25.html">gemm</a></li><li class="navelem"><a class="el" href="dir_9c555e3d0f5b8c3fb3a7397c81fd5bf9.html">kernels</a></li> </ul>
</div>
</div><!-- top -->
<div id="doc-content">
<div class="header">
<div class="headertitle"><div class="title">steel_gemm_masked.h</div></div>
</div><!--header-->
@@ -96,7 +110,7 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00003" name="l00003"></a><span class="lineno"> 3</span><span class="preprocessor">#include &quot;<a class="code" href="steel_2defines_8h.html">mlx/backend/metal/kernels/steel/defines.h</a>&quot;</span></div>
<div class="line"><a id="l00004" name="l00004"></a><span class="lineno"> 4</span><span class="keyword">using namespace </span><a class="code hl_namespace" href="namespacemetal.html">metal</a>;</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="namespacemlx_1_1steel.html">mlx::steel</a>;</div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span> </div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span></div>
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span><span class="comment">// GEMM kernels</span></div>
<div class="line"><a id="l00010" name="l00010"></a><span class="lineno"> 10</span> </div>
<div class="line"><a id="l00011" name="l00011"></a><span class="lineno"> 11</span><span class="keyword">struct </span><a class="code hl_struct" href="struct___no_mask.html">_NoMask</a> {</div>
@@ -135,7 +149,7 @@ $(function(){ initResizable(false); });
</div>
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span>};</div>
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> </div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span><span class="keyword">typedef</span> <span class="keyword">struct </span><a class="code hl_struct" href="struct___no_mask.html">_NoMask</a> <a class="code hl_struct" href="struct___no_mask.html">nomask_t</a>;</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span><span class="keyword">typedef</span> <span class="keyword">struct </span><a class="code hl_struct" href="struct___no_mask.html">_NoMask</a> <a class="code hl_typedef" href="kernels_2gemv__masked_8h.html#a1480c8cdff1cae1462a5a71632969bca">nomask_t</a>;</div>
<div class="line"><a id="l00038" name="l00038"></a><span class="lineno"> 38</span> </div>
<div class="line"><a id="l00039" name="l00039"></a><span class="lineno"> 39</span><span class="keyword">template</span> &lt;</div>
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> <span class="keyword">typename</span> T,</div>
@@ -152,13 +166,13 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span> <span class="keywordtype">bool</span> K_aligned&gt;</div>
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"> 52</span>[[kernel, max_total_threads_per_threadgroup(WM* WN * 32)]] <span class="keywordtype">void</span></div>
<div class="foldopen" id="foldopen00053" data-start="{" data-end="}">
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"><a class="line" href="steel__gemm__masked_8h.html#af805e998b2046ee30c2b4be813e3af97"> 53</a></span><a class="code hl_function" href="steel__gemm__masked_8h.html#af805e998b2046ee30c2b4be813e3af97">block_masked_gemm</a>(</div>
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"><a class="line" href="steel__gemm__masked_8h.html#a3642df61cba8eff47a08a20834137eac"> 53</a></span><a class="code hl_function" href="steel__gemm__masked_8h.html#a3642df61cba8eff47a08a20834137eac">block_masked_gemm</a>(</div>
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> <span class="keyword">const</span> device T* A [[buffer(0)]],</div>
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> <span class="keyword">const</span> device T* B [[buffer(1)]],</div>
<div class="line"><a id="l00056" name="l00056"></a><span class="lineno"> 56</span> device T* D [[buffer(3)]],</div>
<div class="line"><a id="l00057" name="l00057"></a><span class="lineno"> 57</span> <span class="keyword">const</span> constant <a class="code hl_struct" href="structmlx_1_1steel_1_1_g_e_m_m_params.html">GEMMParams</a>* params [[buffer(4)]],</div>
<div class="line"><a id="l00058" name="l00058"></a><span class="lineno"> 58</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(6)]],</div>
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* batch_strides [[buffer(7)]],</div>
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span> <span class="keyword">const</span> constant int64_t* batch_strides [[buffer(7)]],</div>
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span> <span class="keyword">const</span> device out_mask_t* out_mask [[buffer(10)]],</div>
<div class="line"><a id="l00061" name="l00061"></a><span class="lineno"> 61</span> <span class="keyword">const</span> device op_mask_t* lhs_mask [[buffer(11)]],</div>
<div class="line"><a id="l00062" name="l00062"></a><span class="lineno"> 62</span> <span class="keyword">const</span> device op_mask_t* rhs_mask [[buffer(12)]],</div>
@@ -206,23 +220,23 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> <span class="keywordflow">return</span>;</div>
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> }</div>
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> </div>
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_batch_strides =</div>
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_batch_strides =</div>
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> batch_strides + 2 * params-&gt;batch_ndim;</div>
<div class="line"><a id="l00109" name="l00109"></a><span class="lineno"> 109</span> </div>
<div class="line"><a id="l00110" name="l00110"></a><span class="lineno"> 110</span> <span class="keywordflow">if</span> (params-&gt;batch_ndim &gt; 1) {</div>
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> <span class="keywordflow">if</span> (has_output_mask) {</div>
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> out_mask += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc</a>(</div>
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> out_mask += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc</a>(</div>
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> tid.z, batch_shape, mask_batch_strides, params-&gt;batch_ndim);</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> mask_batch_strides += params-&gt;batch_ndim;</div>
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> }</div>
<div class="line"><a id="l00117" name="l00117"></a><span class="lineno"> 117</span> </div>
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span> <span class="keywordflow">if</span> (has_operand_mask) {</div>
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_strides_lhs = mask_batch_strides;</div>
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_strides_rhs =</div>
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_strides_lhs = mask_batch_strides;</div>
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_strides_rhs =</div>
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> mask_strides_lhs + params-&gt;batch_ndim;</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> ulong2 batch_offsets = <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="l00123" name="l00123"></a><span class="lineno"> 123</span> ulong2 batch_offsets = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#af62bacceef7d93f8c1ba4fcf5b1adfe6">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> tid.z,</div>
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> batch_shape,</div>
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> mask_strides_lhs,</div>
@@ -246,10 +260,10 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00144" name="l00144"></a><span class="lineno"> 144</span> </div>
<div class="line"><a id="l00145" name="l00145"></a><span class="lineno"> 145</span> <span class="comment">// Adjust for batch</span></div>
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> <span class="keywordflow">if</span> (params-&gt;batch_ndim &gt; 1) {</div>
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* A_bstrides = batch_strides;</div>
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* B_bstrides = batch_strides + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* A_bstrides = batch_strides;</div>
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* B_bstrides = batch_strides + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> </div>
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> ulong2 batch_offsets = <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="l00150" name="l00150"></a><span class="lineno"> 150</span> ulong2 batch_offsets = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#af62bacceef7d93f8c1ba4fcf5b1adfe6">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> tid.z, batch_shape, A_bstrides, B_bstrides, params-&gt;batch_ndim);</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> A += batch_offsets.x;</div>
@@ -358,7 +372,7 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> MN_aligned ? short(BN) : short(<a class="code hl_function" href="namespacemetal.html#a6653b28c9473087141eddce39878d4d3">min</a>(BN, params-&gt;N - c_col));</div>
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> </div>
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> <span class="keywordtype">int</span> gemm_k_iterations = params-&gt;gemm_k_iterations_aligned;</div>
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> </div>
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span></div>
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> <span class="comment">// Do unaligned K iterations first</span></div>
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"> 262</span> <span class="keywordflow">if</span> (!K_aligned) {</div>
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> <span class="keyword">const</span> <span class="keywordtype">int</span> k_last = params-&gt;gemm_k_iterations_aligned * BK;</div>
@@ -408,7 +422,7 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> loader_b.src -= k_jump_b;</div>
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> }</div>
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> }</div>
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> </div>
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span></div>
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> <span class="comment">// MNK aligned loop</span></div>
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> <span class="keywordflow">if</span> (MN_aligned) {</div>
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> <span class="keywordflow">for</span> (; gemm_k_iterations &gt; 0; gemm_k_iterations--) {</div>
@@ -537,13 +551,13 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00437" name="l00437"></a><span class="lineno"> 437</span> <span class="keywordtype">bool</span> has_operand_mask = <span class="keyword">false</span>&gt;</div>
<div class="line"><a id="l00438" name="l00438"></a><span class="lineno"> 438</span>[[kernel, max_total_threads_per_threadgroup(WM* WN * 32)]] <span class="keywordtype">void</span></div>
<div class="foldopen" id="foldopen00439" data-start="{" data-end="}">
<div class="line"><a id="l00439" name="l00439"></a><span class="lineno"><a class="line" href="steel__gemm__masked_8h.html#a477932e2ae9d49366f7ede6db63f9cac"> 439</a></span><a class="code hl_function" href="steel__gemm__masked_8h.html#af805e998b2046ee30c2b4be813e3af97">block_masked_gemm</a>(</div>
<div class="line"><a id="l00439" name="l00439"></a><span class="lineno"><a class="line" href="steel__gemm__masked_8h.html#ac59086cee8cf396751da691fe80cda7b"> 439</a></span><a class="code hl_function" href="steel__gemm__masked_8h.html#a3642df61cba8eff47a08a20834137eac">block_masked_gemm</a>(</div>
<div class="line"><a id="l00440" name="l00440"></a><span class="lineno"> 440</span> <span class="keyword">const</span> device T* A [[buffer(0)]],</div>
<div class="line"><a id="l00441" name="l00441"></a><span class="lineno"> 441</span> <span class="keyword">const</span> device T* B [[buffer(1)]],</div>
<div class="line"><a id="l00442" name="l00442"></a><span class="lineno"> 442</span> device T* D [[buffer(3)]],</div>
<div class="line"><a id="l00443" name="l00443"></a><span class="lineno"> 443</span> <span class="keyword">const</span> constant <a class="code hl_struct" href="structmlx_1_1steel_1_1_g_e_m_m_params.html">GEMMParams</a>* params [[buffer(4)]],</div>
<div class="line"><a id="l00444" name="l00444"></a><span class="lineno"> 444</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* batch_shape [[buffer(6)]],</div>
<div class="line"><a id="l00445" name="l00445"></a><span class="lineno"> 445</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* batch_strides [[buffer(7)]],</div>
<div class="line"><a id="l00445" name="l00445"></a><span class="lineno"> 445</span> <span class="keyword">const</span> constant int64_t* batch_strides [[buffer(7)]],</div>
<div class="line"><a id="l00446" name="l00446"></a><span class="lineno"> 446</span> <span class="keyword">const</span> device <span class="keywordtype">bool</span>* out_mask [[buffer(10)]],</div>
<div class="line"><a id="l00447" name="l00447"></a><span class="lineno"> 447</span> <span class="keyword">const</span> device <span class="keywordtype">bool</span>* lhs_mask [[buffer(11)]],</div>
<div class="line"><a id="l00448" name="l00448"></a><span class="lineno"> 448</span> <span class="keyword">const</span> device <span class="keywordtype">bool</span>* rhs_mask [[buffer(12)]],</div>
@@ -577,18 +591,18 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00476" name="l00476"></a><span class="lineno"> 476</span> }</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> <span class="keywordflow">if</span> (params-&gt;batch_ndim &gt; 1) {</div>
<div class="line"><a id="l00479" name="l00479"></a><span class="lineno"> 479</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_batch_strides =</div>
<div class="line"><a id="l00479" name="l00479"></a><span class="lineno"> 479</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_batch_strides =</div>
<div class="line"><a id="l00480" name="l00480"></a><span class="lineno"> 480</span> batch_strides + 2 * params-&gt;batch_ndim;</div>
<div class="line"><a id="l00481" name="l00481"></a><span class="lineno"> 481</span> out_mask +=</div>
<div class="line"><a id="l00482" name="l00482"></a><span class="lineno"> 482</span> <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc</a>(tid.z, batch_shape, mask_batch_strides, params-&gt;batch_ndim);</div>
<div class="line"><a id="l00482" name="l00482"></a><span class="lineno"> 482</span> <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc</a>(tid.z, batch_shape, mask_batch_strides, params-&gt;batch_ndim);</div>
<div class="line"><a id="l00483" name="l00483"></a><span class="lineno"> 483</span> </div>
<div class="line"><a id="l00484" name="l00484"></a><span class="lineno"> 484</span> <span class="keywordflow">if</span> (has_operand_mask) {</div>
<div class="line"><a id="l00485" name="l00485"></a><span class="lineno"> 485</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_strides_lhs =</div>
<div class="line"><a id="l00485" name="l00485"></a><span class="lineno"> 485</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_strides_lhs =</div>
<div class="line"><a id="l00486" name="l00486"></a><span class="lineno"> 486</span> mask_batch_strides + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00487" name="l00487"></a><span class="lineno"> 487</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* mask_strides_rhs =</div>
<div class="line"><a id="l00487" name="l00487"></a><span class="lineno"> 487</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* mask_strides_rhs =</div>
<div class="line"><a id="l00488" name="l00488"></a><span class="lineno"> 488</span> mask_strides_lhs + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00489" name="l00489"></a><span class="lineno"> 489</span> </div>
<div class="line"><a id="l00490" name="l00490"></a><span class="lineno"> 490</span> ulong2 batch_offsets = <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="l00490" name="l00490"></a><span class="lineno"> 490</span> ulong2 batch_offsets = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#af62bacceef7d93f8c1ba4fcf5b1adfe6">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l00491" name="l00491"></a><span class="lineno"> 491</span> tid.z,</div>
<div class="line"><a id="l00492" name="l00492"></a><span class="lineno"> 492</span> batch_shape,</div>
<div class="line"><a id="l00493" name="l00493"></a><span class="lineno"> 493</span> mask_strides_lhs,</div>
@@ -608,10 +622,10 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00507" name="l00507"></a><span class="lineno"> 507</span> </div>
<div class="line"><a id="l00508" name="l00508"></a><span class="lineno"> 508</span> <span class="comment">// Adjust for batch</span></div>
<div class="line"><a id="l00509" name="l00509"></a><span class="lineno"> 509</span> <span class="keywordflow">if</span> (params-&gt;batch_ndim &gt; 1) {</div>
<div class="line"><a id="l00510" name="l00510"></a><span class="lineno"> 510</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* A_bstrides = batch_strides;</div>
<div class="line"><a id="l00511" name="l00511"></a><span class="lineno"> 511</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* B_bstrides = batch_strides + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00510" name="l00510"></a><span class="lineno"> 510</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* A_bstrides = batch_strides;</div>
<div class="line"><a id="l00511" name="l00511"></a><span class="lineno"> 511</span> <span class="keyword">const</span> constant <span class="keyword">auto</span>* B_bstrides = batch_strides + params-&gt;batch_ndim;</div>
<div class="line"><a id="l00512" name="l00512"></a><span class="lineno"> 512</span> </div>
<div class="line"><a id="l00513" name="l00513"></a><span class="lineno"> 513</span> ulong2 batch_offsets = <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="l00513" name="l00513"></a><span class="lineno"> 513</span> ulong2 batch_offsets = <a class="code hl_function" href="backend_2metal_2kernels_2steel_2utils_8h.html#af62bacceef7d93f8c1ba4fcf5b1adfe6">elem_to_loc_broadcast</a>(</div>
<div class="line"><a id="l00514" name="l00514"></a><span class="lineno"> 514</span> tid.z, batch_shape, A_bstrides, B_bstrides, params-&gt;batch_ndim);</div>
<div class="line"><a id="l00515" name="l00515"></a><span class="lineno"> 515</span> </div>
<div class="line"><a id="l00516" name="l00516"></a><span class="lineno"> 516</span> A += batch_offsets.x;</div>
@@ -689,7 +703,7 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00588" name="l00588"></a><span class="lineno"> 588</span> A, params-&gt;lda, As, simd_group_id, simd_lane_id);</div>
<div class="line"><a id="l00589" name="l00589"></a><span class="lineno"> 589</span> thread <span class="keyword">typename</span> gemm_kernel::loader_b_t loader_b(</div>
<div class="line"><a id="l00590" name="l00590"></a><span class="lineno"> 590</span> B, params-&gt;ldb, Bs, simd_group_id, simd_lane_id);</div>
<div class="line"><a id="l00591" name="l00591"></a><span class="lineno"> 591</span> </div>
<div class="line"><a id="l00591" name="l00591"></a><span class="lineno"> 591</span></div>
<div class="line"><a id="l00593" name="l00593"></a><span class="lineno"> 593</span> <span class="comment">// MNK aligned loop</span></div>
<div class="line"><a id="l00594" name="l00594"></a><span class="lineno"> 594</span> <span class="keywordflow">if</span> (MN_aligned) {</div>
<div class="line"><a id="l00595" name="l00595"></a><span class="lineno"> 595</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> k = 0; k &lt; gemm_k_iterations; k++) {</div>
@@ -817,14 +831,15 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00718" name="l00718"></a><span class="lineno"> 718</span> }</div>
<div class="line"><a id="l00719" name="l00719"></a><span class="lineno"> 719</span>}</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_a22eaa505dbc7dd2a63a895f2e16712f5"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc</a></div><div class="ttdeci">METAL_FUNC IdxT elem_to_loc(uint elem, constant const int *shape, constant const StrideT *strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:93</div></div>
<div class="ttc" id="abackend_2metal_2kernels_2steel_2utils_8h_html_af62bacceef7d93f8c1ba4fcf5b1adfe6"><div class="ttname"><a href="backend_2metal_2kernels_2steel_2utils_8h.html#af62bacceef7d93f8c1ba4fcf5b1adfe6">elem_to_loc_broadcast</a></div><div class="ttdeci">METAL_FUNC ulong2 elem_to_loc_broadcast(uint elem, constant const int *shape, constant const int64_t *a_strides, constant const int64_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_a497dd9f1a00c8a4303d8782158a0812a"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc</a></div><div class="ttdeci">METAL_FUNC IdxT elem_to_loc(IdxT elem, constant const int *shape, constant const int64_t *strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:93</div></div>
<div class="ttc" id="akernels_2gemv__masked_8h_html_a1480c8cdff1cae1462a5a71632969bca"><div class="ttname"><a href="kernels_2gemv__masked_8h.html#a1480c8cdff1cae1462a5a71632969bca">nomask_t</a></div><div class="ttdeci">struct _NoMask nomask_t</div><div class="ttdef"><b>Definition</b> gemv_masked.h:27</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_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:232</div></div>
<div class="ttc" id="anamespacemlx_1_1steel_html"><div class="ttname"><a href="namespacemlx_1_1steel.html">mlx::steel</a></div><div class="ttdef"><b>Definition</b> attn.h:19</div></div>
<div class="ttc" id="asteel_2defines_8h_html"><div class="ttname"><a href="steel_2defines_8h.html">defines.h</a></div></div>
<div class="ttc" id="asteel_2defines_8h_html_a5a5c3095b132a7589bc19cd5cb80e2c6"><div class="ttname"><a href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div><div class="ttdeci">#define STEEL_PRAGMA_UNROLL</div><div class="ttdef"><b>Definition</b> defines.h:4</div></div>
<div class="ttc" id="asteel__gemm__masked_8h_html_af805e998b2046ee30c2b4be813e3af97"><div class="ttname"><a href="steel__gemm__masked_8h.html#af805e998b2046ee30c2b4be813e3af97">block_masked_gemm</a></div><div class="ttdeci">void block_masked_gemm(const device T *A, const device T *B, device T *D, const constant GEMMParams *params, const constant int *batch_shape, const constant size_t *batch_strides, const device out_mask_t *out_mask, const device op_mask_t *lhs_mask, const device op_mask_t *rhs_mask, const constant int *mask_strides, uint simd_lane_id, uint simd_group_id, uint3 tid, uint3 lid)</div><div class="ttdef"><b>Definition</b> steel_gemm_masked.h:53</div></div>
<div class="ttc" id="asteel__gemm__masked_8h_html_a3642df61cba8eff47a08a20834137eac"><div class="ttname"><a href="steel__gemm__masked_8h.html#a3642df61cba8eff47a08a20834137eac">block_masked_gemm</a></div><div class="ttdeci">void block_masked_gemm(const device T *A, const device T *B, device T *D, const constant GEMMParams *params, const constant int *batch_shape, const constant int64_t *batch_strides, const device out_mask_t *out_mask, const device op_mask_t *lhs_mask, const device op_mask_t *rhs_mask, const constant int *mask_strides, uint simd_lane_id, uint simd_group_id, uint3 tid, uint3 lid)</div><div class="ttdef"><b>Definition</b> steel_gemm_masked.h:53</div></div>
<div class="ttc" id="astruct___no_mask_html"><div class="ttname"><a href="struct___no_mask.html">_NoMask</a></div><div class="ttdef"><b>Definition</b> gemv_masked.h:10</div></div>
<div class="ttc" id="astruct___no_mask_html_a0c4a4557d5c97ceafe3a2c4e521cdf7e"><div class="ttname"><a href="struct___no_mask.html#a0c4a4557d5c97ceafe3a2c4e521cdf7e">_NoMask::x</a></div><div class="ttdeci">char x</div><div class="ttdef"><b>Definition</b> gemv_masked.h:11</div></div>
<div class="ttc" id="astruct_scale_op_html"><div class="ttname"><a href="struct_scale_op.html">ScaleOp</a></div><div class="ttdef"><b>Definition</b> gemv_masked.h:30</div></div>
@@ -833,10 +848,13 @@ $(function(){ initResizable(false); });
<div class="ttc" id="astructmlx_1_1steel_1_1_g_e_m_m_kernel_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_g_e_m_m_kernel.html">mlx::steel::GEMMKernel</a></div><div class="ttdef"><b>Definition</b> gemm.h:37</div></div>
<div class="ttc" id="astructmlx_1_1steel_1_1_g_e_m_m_params_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_g_e_m_m_params.html">mlx::steel::GEMMParams</a></div><div class="ttdef"><b>Definition</b> params.h:12</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 -->
<!-- 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="dir_76215a6c54e2b67053e723fc2395583c.html">steel</a></li><li class="navelem"><a class="el" href="dir_6768c99e6145fb9510ccdb40db8ede25.html">gemm</a></li><li class="navelem"><a class="el" href="dir_9c555e3d0f5b8c3fb3a7397c81fd5bf9.html">kernels</a></li><li class="navelem"><a class="el" href="steel__gemm__masked_8h.html">steel_gemm_masked.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> 1.13.1 </li>
</ul>
</div>
</body>
</html>