mirror of
https://github.com/ml-explore/mlx.git
synced 2025-09-20 20:18:15 +08:00
rebase
This commit is contained in:
211
docs/build/html/reduce__row_8h_source.html
vendored
211
docs/build/html/reduce__row_8h_source.html
vendored
@@ -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/reduction/reduce_row.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&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()"> </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&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&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&dn=expat.txt MIT */
|
||||
$(function(){ initResizable(false); });
|
||||
$(function(){initNavTree('reduce__row_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_f60cd69d27fd3faa641c79056fff0e2d.html">reduction</a></li> </ul>
|
||||
</div>
|
||||
</div><!-- top -->
|
||||
<div id="doc-content">
|
||||
<div class="header">
|
||||
<div class="headertitle"><div class="title">reduce_row.h</div></div>
|
||||
</div><!--header-->
|
||||
@@ -98,7 +112,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span><span class="comment">// - `threadgroup_reduce` collaborative reduction in the threadgroup such that</span></div>
|
||||
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span><span class="comment">// lid.x == 0 holds the reduced value</span></div>
|
||||
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span><span class="comment">// - `thread_reduce` simple loop and reduce the row</span></div>
|
||||
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span> </div>
|
||||
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span></div>
|
||||
<div class="line"><a id="l00013" name="l00013"></a><span class="lineno"> 13</span><span class="keyword">template</span> <</div>
|
||||
<div class="line"><a id="l00014" name="l00014"></a><span class="lineno"> 14</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00015" name="l00015"></a><span class="lineno"> 15</span> <span class="keyword">typename</span> U,</div>
|
||||
@@ -113,7 +127,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00023" name="l00023"></a><span class="lineno"> 23</span> <span class="keywordtype">int</span> extra,</div>
|
||||
<div class="line"><a id="l00024" name="l00024"></a><span class="lineno"> 24</span> uint lsize_x,</div>
|
||||
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> uint lid_x) {</div>
|
||||
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> Op op;</div>
|
||||
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> </div>
|
||||
<div class="line"><a id="l00028" name="l00028"></a><span class="lineno"> 28</span> <span class="comment">// Set up the accumulator registers</span></div>
|
||||
<div class="line"><a id="l00029" name="l00029"></a><span class="lineno"> 29</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_WRITES; i++) {</div>
|
||||
@@ -124,7 +138,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00034" name="l00034"></a><span class="lineno"> 34</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < blocks; i++) {</div>
|
||||
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < N_WRITES; j++) {</div>
|
||||
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_READS; i++) {</div>
|
||||
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> totals[j] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> totals[j] = op(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00038" name="l00038"></a><span class="lineno"> 38</span> }</div>
|
||||
<div class="line"><a id="l00039" name="l00039"></a><span class="lineno"> 39</span> </div>
|
||||
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> inputs[j] += lsize_x * N_READS;</div>
|
||||
@@ -136,19 +150,19 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> <span class="keywordflow">if</span> (index + N_READS <= extra) {</div>
|
||||
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < N_WRITES; j++) {</div>
|
||||
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_READS; i++) {</div>
|
||||
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span> totals[j] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span> totals[j] = op(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00050" name="l00050"></a><span class="lineno"> 50</span> }</div>
|
||||
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span> }</div>
|
||||
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"> 52</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < N_WRITES; j++) {</div>
|
||||
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; index + i < extra; i++) {</div>
|
||||
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> totals[j] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span> totals[j] = op(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(inputs[j][i]), totals[j]);</div>
|
||||
<div class="line"><a id="l00056" name="l00056"></a><span class="lineno"> 56</span> }</div>
|
||||
<div class="line"><a id="l00057" name="l00057"></a><span class="lineno"> 57</span> }</div>
|
||||
<div class="line"><a id="l00058" name="l00058"></a><span class="lineno"> 58</span> }</div>
|
||||
<div class="line"><a id="l00059" name="l00059"></a><span class="lineno"> 59</span>}</div>
|
||||
</div>
|
||||
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span> </div>
|
||||
<div class="line"><a id="l00060" name="l00060"></a><span class="lineno"> 60</span></div>
|
||||
<div class="line"><a id="l00064" name="l00064"></a><span class="lineno"> 64</span><span class="keyword">template</span> <</div>
|
||||
<div class="line"><a id="l00065" name="l00065"></a><span class="lineno"> 65</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00066" name="l00066"></a><span class="lineno"> 66</span> <span class="keyword">typename</span> U,</div>
|
||||
@@ -175,7 +189,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00086" name="l00086"></a><span class="lineno"> 86</span> totals, inputs, blocks, extra, lsize_x, lid_x);</div>
|
||||
<div class="line"><a id="l00087" name="l00087"></a><span class="lineno"> 87</span>}</div>
|
||||
</div>
|
||||
<div class="line"><a id="l00088" name="l00088"></a><span class="lineno"> 88</span> </div>
|
||||
<div class="line"><a id="l00088" name="l00088"></a><span class="lineno"> 88</span></div>
|
||||
<div class="line"><a id="l00092" name="l00092"></a><span class="lineno"> 92</span><span class="keyword">template</span> <</div>
|
||||
<div class="line"><a id="l00093" name="l00093"></a><span class="lineno"> 93</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00094" name="l00094"></a><span class="lineno"> 94</span> <span class="keyword">typename</span> U,</div>
|
||||
@@ -183,14 +197,14 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00096" name="l00096"></a><span class="lineno"> 96</span> <span class="keywordtype">int</span> N_READS = <a class="code hl_variable" href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a>,</div>
|
||||
<div class="line"><a id="l00097" name="l00097"></a><span class="lineno"> 97</span> <span class="keywordtype">int</span> N_WRITES = <a class="code hl_variable" href="defines_8h.html#a68c33274e15a2f163f7631a36280d82f">REDUCE_N_WRITES</a>></div>
|
||||
<div class="foldopen" id="foldopen00098" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00098" name="l00098"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#a4d00c44e5f4a13be529ff8b664a0a342"> 98</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce</a>(</div>
|
||||
<div class="line"><a id="l00098" name="l00098"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#aee3cf6c55bc9a6d0046030f7dfd26e6b"> 98</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce</a>(</div>
|
||||
<div class="line"><a id="l00099" name="l00099"></a><span class="lineno"> 99</span> thread U totals[N_WRITES],</div>
|
||||
<div class="line"><a id="l00100" name="l00100"></a><span class="lineno"> 100</span> <span class="keyword">const</span> device T* in,</div>
|
||||
<div class="line"><a id="l00101" name="l00101"></a><span class="lineno"> 101</span> <span class="keyword">const</span> <span class="keywordtype">size_t</span> row_idx,</div>
|
||||
<div class="line"><a id="l00101" name="l00101"></a><span class="lineno"> 101</span> <span class="keyword">const</span> int64_t row_idx,</div>
|
||||
<div class="line"><a id="l00102" name="l00102"></a><span class="lineno"> 102</span> <span class="keywordtype">int</span> blocks,</div>
|
||||
<div class="line"><a id="l00103" name="l00103"></a><span class="lineno"> 103</span> <span class="keywordtype">int</span> extra,</div>
|
||||
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* shape,</div>
|
||||
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* strides,</div>
|
||||
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> <span class="keyword">const</span> constant int64_t* strides,</div>
|
||||
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& ndim,</div>
|
||||
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> uint lsize_x,</div>
|
||||
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> uint lid_x) {</div>
|
||||
@@ -198,14 +212,14 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00110" name="l00110"></a><span class="lineno"> 110</span> <span class="keyword">const</span> device T* inputs[N_WRITES];</div>
|
||||
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> in += lid_x * N_READS;</div>
|
||||
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_READS; i++) {</div>
|
||||
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> inputs[i] = in + <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc</a>(row_idx + i, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> inputs[i] = in + <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc</a>(row_idx + i, shape, strides, 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> </div>
|
||||
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> <a class="code hl_function" href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce<T, U, Op, N_READS, N_WRITES></a>(</div>
|
||||
<div class="line"><a id="l00117" name="l00117"></a><span class="lineno"> 117</span> totals, inputs, blocks, extra, lsize_x, lid_x);</div>
|
||||
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span>}</div>
|
||||
</div>
|
||||
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> </div>
|
||||
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span></div>
|
||||
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span><span class="keyword">template</span> <</div>
|
||||
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> <span class="keyword">typename</span> U,</div>
|
||||
@@ -220,11 +234,11 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
|
||||
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> uint simd_per_group [[simdgroups_per_threadgroup]],</div>
|
||||
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
|
||||
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> Op op;</div>
|
||||
<div class="line"><a id="l00137" name="l00137"></a><span class="lineno"> 137</span> </div>
|
||||
<div class="line"><a id="l00138" name="l00138"></a><span class="lineno"> 138</span> <span class="comment">// Simdgroup first</span></div>
|
||||
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_WRITES; i++) {</div>
|
||||
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.simd_reduce(totals[i]);</div>
|
||||
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> totals[i] = op.simd_reduce(totals[i]);</div>
|
||||
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span> }</div>
|
||||
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"> 142</span> </div>
|
||||
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> <span class="comment">// Across simdgroups</span></div>
|
||||
@@ -239,11 +253,11 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> U values[N_WRITES];</div>
|
||||
<div class="line"><a id="l00153" name="l00153"></a><span class="lineno"> 153</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_WRITES; i++) {</div>
|
||||
<div class="line"><a id="l00154" name="l00154"></a><span class="lineno"> 154</span> values[i] = (lid.x < simd_per_group) ? shared_vals[lid.x * N_WRITES + i]</div>
|
||||
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> : <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.init;</div>
|
||||
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> : op.init;</div>
|
||||
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> }</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="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < N_WRITES; i++) {</div>
|
||||
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.simd_reduce(values[i]);</div>
|
||||
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> totals[i] = op.simd_reduce(values[i]);</div>
|
||||
<div class="line"><a id="l00160" name="l00160"></a><span class="lineno"> 160</span> }</div>
|
||||
<div class="line"><a id="l00161" name="l00161"></a><span class="lineno"> 161</span> }</div>
|
||||
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span>}</div>
|
||||
@@ -253,19 +267,19 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span>METAL_FUNC <span class="keywordtype">void</span></div>
|
||||
<div class="foldopen" id="foldopen00166" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#afd80a25fa84e6cc884dcc8698859ade1"> 166</a></span><a class="code hl_function" href="reduce__row_8h.html#afd80a25fa84e6cc884dcc8698859ade1">thread_reduce</a>(thread U& total, <span class="keyword">const</span> device T* row, <span class="keywordtype">int</span> blocks, <span class="keywordtype">int</span> extra) {</div>
|
||||
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> Op op;</div>
|
||||
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < blocks; i++) {</div>
|
||||
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> U vals[N_READS];</div>
|
||||
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < N_READS; j++) {</div>
|
||||
<div class="line"><a id="l00171" name="l00171"></a><span class="lineno"> 171</span> vals[j] = row[j];</div>
|
||||
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> }</div>
|
||||
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 0; j < N_READS; j++) {</div>
|
||||
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> total = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(vals[j], total);</div>
|
||||
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> total = op(vals[j], total);</div>
|
||||
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> }</div>
|
||||
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> row += N_READS;</div>
|
||||
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> }</div>
|
||||
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < extra; i++) {</div>
|
||||
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> total = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(*row++, total);</div>
|
||||
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> total = op(*row++, total);</div>
|
||||
<div class="line"><a id="l00180" name="l00180"></a><span class="lineno"> 180</span> }</div>
|
||||
<div class="line"><a id="l00181" name="l00181"></a><span class="lineno"> 181</span>}</div>
|
||||
</div>
|
||||
@@ -287,23 +301,23 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00197" name="l00197"></a><span class="lineno"> 197</span> <span class="keywordtype">int</span> NDIMS,</div>
|
||||
<div class="line"><a id="l00198" name="l00198"></a><span class="lineno"> 198</span> <span class="keywordtype">int</span> N_READS = <a class="code hl_variable" href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a>></div>
|
||||
<div class="foldopen" id="foldopen00199" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#aeb49e89f1163cb3093770bb710df9f5e"> 199</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#aeb49e89f1163cb3093770bb710df9f5e">row_reduce_small</a>(</div>
|
||||
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#a1e1b59fa73d2b0be978494a759f2c6ee"> 199</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#a1e1b59fa73d2b0be978494a759f2c6ee">row_reduce_small</a>(</div>
|
||||
<div class="line"><a id="l00200" name="l00200"></a><span class="lineno"> 200</span> <span class="keyword">const</span> device T* in [[buffer(0)]],</div>
|
||||
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span> device U* out [[buffer(1)]],</div>
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& row_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& non_row_reductions [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> <span class="keyword">const</span> constant int64_t& row_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> <span class="keyword">const</span> constant int64_t& non_row_reductions [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* shape [[buffer(4)]],</div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="keyword">const</span> constant int64_t* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& ndim [[buffer(6)]],</div>
|
||||
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* reduce_shape [[buffer(7)]],</div>
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> <span class="keyword">const</span> constant int64_t* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& reduce_ndim [[buffer(9)]],</div>
|
||||
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
|
||||
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> uint3 gid [[threadgroup_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> uint3 gsize [[threadgroups_per_grid]],</div>
|
||||
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> uint3 tid [[thread_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> uint3 tsize [[threads_per_grid]]) {</div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> Op op;</div>
|
||||
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> </div>
|
||||
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> U total_val = Op::init;</div>
|
||||
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> <a class="code hl_struct" href="struct_looped_elem_to_loc.html">LoopedElemToLoc</a><NDIMS, IdxT, (NDIMS > 2)> loop(reduce_ndim);</div>
|
||||
@@ -316,7 +330,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> <span class="keywordflow">if</span> ((non_row_reductions < 32 && row_size <= 8) || non_row_reductions <= 8) {</div>
|
||||
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> <span class="comment">// Simple loop over non_row_reductions and reduce the row in the thread.</span></div>
|
||||
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> IdxT out_idx = tid.x + tsize.y * IdxT(tid.y);</div>
|
||||
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc<size_t, IdxT></a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc<IdxT></a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> </div>
|
||||
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> <span class="keywordflow">for</span> (uint r = 0; r < non_row_reductions; r++) {</div>
|
||||
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> row = in + loop.location();</div>
|
||||
@@ -329,7 +343,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> <span class="comment">// Collaboratively reduce over non_row_reductions in the simdgroup. Each</span></div>
|
||||
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> <span class="comment">// thread reduces every 32nd row and then a simple simd reduce.</span></div>
|
||||
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> IdxT out_idx = gid.y + gsize.y * IdxT(gid.z);</div>
|
||||
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc<size_t, IdxT></a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc<IdxT></a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00242" name="l00242"></a><span class="lineno"> 242</span> </div>
|
||||
<div class="line"><a id="l00243" name="l00243"></a><span class="lineno"> 243</span> loop.next(simd_lane_id, reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> </div>
|
||||
@@ -339,7 +353,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> loop.next(<a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a>, reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> }</div>
|
||||
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> </div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> total_val = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.simd_reduce(total_val);</div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> total_val = op.simd_reduce(total_val);</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="keywordflow">if</span> (simd_lane_id == 0) {</div>
|
||||
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> out[out_idx] = total_val;</div>
|
||||
@@ -352,15 +366,15 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> <span class="keyword">typename</span> U,</div>
|
||||
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"> 262</span> <span class="keyword">typename</span> Op,</div>
|
||||
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> <span class="keyword">typename</span> IdxT = size_t,</div>
|
||||
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> <span class="keyword">typename</span> IdxT = int64_t,</div>
|
||||
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> <span class="keywordtype">int</span> N_READS = <a class="code hl_variable" href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a>,</div>
|
||||
<div class="line"><a id="l00265" name="l00265"></a><span class="lineno"> 265</span> <span class="keywordtype">int</span> N_WRITES = <a class="code hl_variable" href="defines_8h.html#a68c33274e15a2f163f7631a36280d82f">REDUCE_N_WRITES</a>></div>
|
||||
<div class="foldopen" id="foldopen00266" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#aef628dfccdb1361da5546f8b17c510bf"> 266</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#aef628dfccdb1361da5546f8b17c510bf">row_reduce_simple</a>(</div>
|
||||
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#a7dbd7cc81b1ce5a4271d56b99ce595d4"> 266</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#a7dbd7cc81b1ce5a4271d56b99ce595d4">row_reduce_simple</a>(</div>
|
||||
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> <span class="keyword">const</span> device T* in [[buffer(0)]],</div>
|
||||
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> device U* out [[buffer(1)]],</div>
|
||||
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& reduction_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& out_size [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> <span class="keyword">const</span> constant int64_t& out_size [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span> uint3 gid [[threadgroup_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> uint3 gsize [[threadgroups_per_grid]],</div>
|
||||
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> uint3 lid [[thread_position_in_threadgroup]],</div>
|
||||
@@ -406,16 +420,16 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> <span class="keywordtype">int</span> NDIMS,</div>
|
||||
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> <span class="keywordtype">int</span> N_READS = <a class="code hl_variable" href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a>></div>
|
||||
<div class="foldopen" id="foldopen00314" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#afba85f5a1c935c124ef52e986d4b2c49"> 314</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#afba85f5a1c935c124ef52e986d4b2c49">row_reduce_looped</a>(</div>
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"><a class="line" href="reduce__row_8h.html#a72611b8006ae5642b69f4d250d69865b"> 314</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__row_8h.html#a72611b8006ae5642b69f4d250d69865b">row_reduce_looped</a>(</div>
|
||||
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> <span class="keyword">const</span> device T* in [[buffer(0)]],</div>
|
||||
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> device U* out [[buffer(1)]],</div>
|
||||
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& row_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& non_row_reductions [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> <span class="keyword">const</span> constant int64_t& row_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> <span class="keyword">const</span> constant int64_t& non_row_reductions [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* shape [[buffer(4)]],</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> <span class="keyword">const</span> constant int64_t* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& ndim [[buffer(6)]],</div>
|
||||
<div class="line"><a id="l00322" name="l00322"></a><span class="lineno"> 322</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* reduce_shape [[buffer(7)]],</div>
|
||||
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> <span class="keyword">const</span> constant int64_t* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& reduce_ndim [[buffer(9)]],</div>
|
||||
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> uint3 gid [[threadgroup_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> uint3 gsize [[threadgroups_per_grid]],</div>
|
||||
@@ -424,7 +438,7 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00329" name="l00329"></a><span class="lineno"> 329</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
|
||||
<div class="line"><a id="l00330" name="l00330"></a><span class="lineno"> 330</span> uint simd_per_group [[simdgroups_per_threadgroup]],</div>
|
||||
<div class="line"><a id="l00331" name="l00331"></a><span class="lineno"> 331</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
|
||||
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> Op op;</div>
|
||||
<div class="line"><a id="l00333" name="l00333"></a><span class="lineno"> 333</span> threadgroup U shared_vals[<a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a>];</div>
|
||||
<div class="line"><a id="l00334" name="l00334"></a><span class="lineno"> 334</span> U total = Op::init;</div>
|
||||
<div class="line"><a id="l00335" name="l00335"></a><span class="lineno"> 335</span> </div>
|
||||
@@ -432,55 +446,56 @@ $(function(){ initResizable(false); });
|
||||
<div class="line"><a id="l00337" name="l00337"></a><span class="lineno"> 337</span> </div>
|
||||
<div class="line"><a id="l00338" name="l00338"></a><span class="lineno"> 338</span> <span class="comment">// lid.x * N_READS breaks the per_thread_row_reduce interface a bit. Maybe it</span></div>
|
||||
<div class="line"><a id="l00339" name="l00339"></a><span class="lineno"> 339</span> <span class="comment">// needs a small refactor.</span></div>
|
||||
<div class="line"><a id="l00340" name="l00340"></a><span class="lineno"> 340</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc<size_t, IdxT></a>(out_idx, shape, strides, ndim) +</div>
|
||||
<div class="line"><a id="l00341" name="l00341"></a><span class="lineno"> 341</span> lid.x * N_READS;</div>
|
||||
<div class="line"><a id="l00342" name="l00342"></a><span class="lineno"> 342</span> </div>
|
||||
<div class="line"><a id="l00343" name="l00343"></a><span class="lineno"> 343</span> <a class="code hl_struct" href="struct_looped_elem_to_loc.html">LoopedElemToLoc</a><NDIMS, IdxT, (NDIMS > 2)> loop(reduce_ndim);</div>
|
||||
<div class="line"><a id="l00344" name="l00344"></a><span class="lineno"> 344</span> <span class="keyword">const</span> device T* row;</div>
|
||||
<div class="line"><a id="l00345" name="l00345"></a><span class="lineno"> 345</span> <span class="keywordtype">int</span> blocks = IdxT(row_size) / (lsize.x * N_READS);</div>
|
||||
<div class="line"><a id="l00346" name="l00346"></a><span class="lineno"> 346</span> <span class="keywordtype">int</span> extra = row_size - blocks * (lsize.x * N_READS);</div>
|
||||
<div class="line"><a id="l00347" name="l00347"></a><span class="lineno"> 347</span> </div>
|
||||
<div class="line"><a id="l00348" name="l00348"></a><span class="lineno"> 348</span> <span class="keywordflow">for</span> (IdxT i = 0; i < non_row_reductions; i++) {</div>
|
||||
<div class="line"><a id="l00349" name="l00349"></a><span class="lineno"> 349</span> row = in + loop.location();</div>
|
||||
<div class="line"><a id="l00350" name="l00350"></a><span class="lineno"> 350</span> </div>
|
||||
<div class="line"><a id="l00351" name="l00351"></a><span class="lineno"> 351</span> <span class="comment">// Each thread reduces across the row</span></div>
|
||||
<div class="line"><a id="l00352" name="l00352"></a><span class="lineno"> 352</span> U row_total;</div>
|
||||
<div class="line"><a id="l00353" name="l00353"></a><span class="lineno"> 353</span> <a class="code hl_function" href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce<T, U, Op, N_READS, 1></a>(</div>
|
||||
<div class="line"><a id="l00354" name="l00354"></a><span class="lineno"> 354</span> &row_total, &row, blocks, extra, lsize.x, lid.x);</div>
|
||||
<div class="line"><a id="l00355" name="l00355"></a><span class="lineno"> 355</span> </div>
|
||||
<div class="line"><a id="l00356" name="l00356"></a><span class="lineno"> 356</span> <span class="comment">// Aggregate across rows</span></div>
|
||||
<div class="line"><a id="l00357" name="l00357"></a><span class="lineno"> 357</span> total = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(total, row_total);</div>
|
||||
<div class="line"><a id="l00358" name="l00358"></a><span class="lineno"> 358</span> </div>
|
||||
<div class="line"><a id="l00359" name="l00359"></a><span class="lineno"> 359</span> loop.next(reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00360" name="l00360"></a><span class="lineno"> 360</span> }</div>
|
||||
<div class="line"><a id="l00361" name="l00361"></a><span class="lineno"> 361</span> </div>
|
||||
<div class="line"><a id="l00362" name="l00362"></a><span class="lineno"> 362</span> <span class="comment">// Reduce across the threadgroup</span></div>
|
||||
<div class="line"><a id="l00363" name="l00363"></a><span class="lineno"> 363</span> <a class="code hl_function" href="reduce__row_8h.html#aa146bb611069fd2892f03714fd1cc3cf">threadgroup_reduce<T, U, Op, N_READS, 1></a>(</div>
|
||||
<div class="line"><a id="l00364" name="l00364"></a><span class="lineno"> 364</span> &total, shared_vals, lid, simd_lane_id, simd_per_group, simd_group_id);</div>
|
||||
<div class="line"><a id="l00365" name="l00365"></a><span class="lineno"> 365</span> </div>
|
||||
<div class="line"><a id="l00366" name="l00366"></a><span class="lineno"> 366</span> <span class="comment">// Write the output</span></div>
|
||||
<div class="line"><a id="l00367" name="l00367"></a><span class="lineno"> 367</span> <span class="keywordflow">if</span> (lid.x == 0) {</div>
|
||||
<div class="line"><a id="l00368" name="l00368"></a><span class="lineno"> 368</span> out[out_idx] = total;</div>
|
||||
<div class="line"><a id="l00369" name="l00369"></a><span class="lineno"> 369</span> }</div>
|
||||
<div class="line"><a id="l00370" name="l00370"></a><span class="lineno"> 370</span>}</div>
|
||||
<div class="line"><a id="l00340" name="l00340"></a><span class="lineno"> 340</span> in += <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a497dd9f1a00c8a4303d8782158a0812a">elem_to_loc<IdxT></a>(out_idx, shape, strides, ndim) + lid.x * N_READS;</div>
|
||||
<div class="line"><a id="l00341" name="l00341"></a><span class="lineno"> 341</span> </div>
|
||||
<div class="line"><a id="l00342" name="l00342"></a><span class="lineno"> 342</span> <a class="code hl_struct" href="struct_looped_elem_to_loc.html">LoopedElemToLoc</a><NDIMS, IdxT, (NDIMS > 2)> loop(reduce_ndim);</div>
|
||||
<div class="line"><a id="l00343" name="l00343"></a><span class="lineno"> 343</span> <span class="keyword">const</span> device T* row;</div>
|
||||
<div class="line"><a id="l00344" name="l00344"></a><span class="lineno"> 344</span> <span class="keywordtype">int</span> blocks = IdxT(row_size) / (lsize.x * N_READS);</div>
|
||||
<div class="line"><a id="l00345" name="l00345"></a><span class="lineno"> 345</span> <span class="keywordtype">int</span> extra = row_size - blocks * (lsize.x * N_READS);</div>
|
||||
<div class="line"><a id="l00346" name="l00346"></a><span class="lineno"> 346</span> </div>
|
||||
<div class="line"><a id="l00347" name="l00347"></a><span class="lineno"> 347</span> <span class="keywordflow">for</span> (IdxT i = 0; i < non_row_reductions; i++) {</div>
|
||||
<div class="line"><a id="l00348" name="l00348"></a><span class="lineno"> 348</span> row = in + loop.location();</div>
|
||||
<div class="line"><a id="l00349" name="l00349"></a><span class="lineno"> 349</span> </div>
|
||||
<div class="line"><a id="l00350" name="l00350"></a><span class="lineno"> 350</span> <span class="comment">// Each thread reduces across the row</span></div>
|
||||
<div class="line"><a id="l00351" name="l00351"></a><span class="lineno"> 351</span> U row_total;</div>
|
||||
<div class="line"><a id="l00352" name="l00352"></a><span class="lineno"> 352</span> <a class="code hl_function" href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce<T, U, Op, N_READS, 1></a>(</div>
|
||||
<div class="line"><a id="l00353" name="l00353"></a><span class="lineno"> 353</span> &row_total, &row, blocks, extra, lsize.x, lid.x);</div>
|
||||
<div class="line"><a id="l00354" name="l00354"></a><span class="lineno"> 354</span> </div>
|
||||
<div class="line"><a id="l00355" name="l00355"></a><span class="lineno"> 355</span> <span class="comment">// Aggregate across rows</span></div>
|
||||
<div class="line"><a id="l00356" name="l00356"></a><span class="lineno"> 356</span> total = op(total, row_total);</div>
|
||||
<div class="line"><a id="l00357" name="l00357"></a><span class="lineno"> 357</span> </div>
|
||||
<div class="line"><a id="l00358" name="l00358"></a><span class="lineno"> 358</span> loop.next(reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00359" name="l00359"></a><span class="lineno"> 359</span> }</div>
|
||||
<div class="line"><a id="l00360" name="l00360"></a><span class="lineno"> 360</span> </div>
|
||||
<div class="line"><a id="l00361" name="l00361"></a><span class="lineno"> 361</span> <span class="comment">// Reduce across the threadgroup</span></div>
|
||||
<div class="line"><a id="l00362" name="l00362"></a><span class="lineno"> 362</span> <a class="code hl_function" href="reduce__row_8h.html#aa146bb611069fd2892f03714fd1cc3cf">threadgroup_reduce<T, U, Op, N_READS, 1></a>(</div>
|
||||
<div class="line"><a id="l00363" name="l00363"></a><span class="lineno"> 363</span> &total, shared_vals, lid, simd_lane_id, simd_per_group, simd_group_id);</div>
|
||||
<div class="line"><a id="l00364" name="l00364"></a><span class="lineno"> 364</span> </div>
|
||||
<div class="line"><a id="l00365" name="l00365"></a><span class="lineno"> 365</span> <span class="comment">// Write the output</span></div>
|
||||
<div class="line"><a id="l00366" name="l00366"></a><span class="lineno"> 366</span> <span class="keywordflow">if</span> (lid.x == 0) {</div>
|
||||
<div class="line"><a id="l00367" name="l00367"></a><span class="lineno"> 367</span> out[out_idx] = total;</div>
|
||||
<div class="line"><a id="l00368" name="l00368"></a><span class="lineno"> 368</span> }</div>
|
||||
<div class="line"><a id="l00369" name="l00369"></a><span class="lineno"> 369</span>}</div>
|
||||
</div>
|
||||
<div class="ttc" id="abackend_2metal_2kernels_2reduction_2ops_8h_html_a515b75d563a93d3c09ee677948dc83e3"><div class="ttname"><a href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a></div><div class="ttdeci">static constant constexpr const uint8_t simd_size</div><div class="ttdef"><b>Definition</b> ops.h:22</div></div>
|
||||
<div class="ttc" id="abackend_2metal_2kernels_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="acommon_2binary_8h_html_a70228731d29946574b238d21fb4b360c"><div class="ttname"><a href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a></div><div class="ttdeci">Op op</div><div class="ttdef"><b>Definition</b> binary.h:129</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="adefines_8h_html_a2ad505864a2ab786147766900bc18c21"><div class="ttname"><a href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a></div><div class="ttdeci">static constexpr int REDUCE_N_READS</div><div class="ttdef"><b>Definition</b> defines.h:12</div></div>
|
||||
<div class="ttc" id="adefines_8h_html_a68c33274e15a2f163f7631a36280d82f"><div class="ttname"><a href="defines_8h.html#a68c33274e15a2f163f7631a36280d82f">REDUCE_N_WRITES</a></div><div class="ttdeci">static constexpr int REDUCE_N_WRITES</div><div class="ttdef"><b>Definition</b> defines.h:13</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_a1e1b59fa73d2b0be978494a759f2c6ee"><div class="ttname"><a href="reduce__row_8h.html#a1e1b59fa73d2b0be978494a759f2c6ee">row_reduce_small</a></div><div class="ttdeci">void row_reduce_small(const device T *in, device U *out, const constant int64_t &row_size, const constant int64_t &non_row_reductions, const constant int *shape, const constant int64_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant int64_t *reduce_strides, const constant int &reduce_ndim, uint simd_lane_id, uint3 gid, uint3 gsize, uint3 tid, uint3 tsize)</div><div class="ttdef"><b>Definition</b> reduce_row.h:199</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_a72611b8006ae5642b69f4d250d69865b"><div class="ttname"><a href="reduce__row_8h.html#a72611b8006ae5642b69f4d250d69865b">row_reduce_looped</a></div><div class="ttdeci">void row_reduce_looped(const device T *in, device U *out, const constant int64_t &row_size, const constant int64_t &non_row_reductions, const constant int *shape, const constant int64_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant int64_t *reduce_strides, const constant int &reduce_ndim, uint3 gid, uint3 gsize, uint3 lid, uint3 lsize, uint simd_lane_id, uint simd_per_group, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> reduce_row.h:314</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_a7dbd7cc81b1ce5a4271d56b99ce595d4"><div class="ttname"><a href="reduce__row_8h.html#a7dbd7cc81b1ce5a4271d56b99ce595d4">row_reduce_simple</a></div><div class="ttdeci">void row_reduce_simple(const device T *in, device U *out, const constant size_t &reduction_size, const constant int64_t &out_size, uint3 gid, uint3 gsize, uint3 lid, uint3 lsize, uint simd_lane_id, uint simd_per_group, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> reduce_row.h:266</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_a9d5e0049a2276f43702fc6907e74a35f"><div class="ttname"><a href="reduce__row_8h.html#a9d5e0049a2276f43702fc6907e74a35f">per_thread_row_reduce</a></div><div class="ttdeci">METAL_FUNC void per_thread_row_reduce(thread U totals[N_WRITES], const device T *inputs[N_WRITES], int blocks, int extra, uint lsize_x, uint lid_x)</div><div class="ttdoc">The thread group collaboratively reduces across the rows with bounds checking.</div><div class="ttdef"><b>Definition</b> reduce_row.h:19</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_aa146bb611069fd2892f03714fd1cc3cf"><div class="ttname"><a href="reduce__row_8h.html#aa146bb611069fd2892f03714fd1cc3cf">threadgroup_reduce</a></div><div class="ttdeci">METAL_FUNC void threadgroup_reduce(thread U totals[N_WRITES], threadgroup U *shared_vals, uint3 lid, uint simd_lane_id, uint simd_per_group, uint simd_group_id)</div><div class="ttdoc">Reduce within the threadgroup.</div><div class="ttdef"><b>Definition</b> reduce_row.h:129</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_aeb49e89f1163cb3093770bb710df9f5e"><div class="ttname"><a href="reduce__row_8h.html#aeb49e89f1163cb3093770bb710df9f5e">row_reduce_small</a></div><div class="ttdeci">void row_reduce_small(const device T *in, device U *out, const constant size_t &row_size, const constant size_t &non_row_reductions, const constant int *shape, const constant size_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant size_t *reduce_strides, const constant int &reduce_ndim, uint simd_lane_id, uint3 gid, uint3 gsize, uint3 tid, uint3 tsize)</div><div class="ttdef"><b>Definition</b> reduce_row.h:199</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_aef628dfccdb1361da5546f8b17c510bf"><div class="ttname"><a href="reduce__row_8h.html#aef628dfccdb1361da5546f8b17c510bf">row_reduce_simple</a></div><div class="ttdeci">void row_reduce_simple(const device T *in, device U *out, const constant size_t &reduction_size, const constant size_t &out_size, uint3 gid, uint3 gsize, uint3 lid, uint3 lsize, uint simd_lane_id, uint simd_per_group, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> reduce_row.h:266</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_afba85f5a1c935c124ef52e986d4b2c49"><div class="ttname"><a href="reduce__row_8h.html#afba85f5a1c935c124ef52e986d4b2c49">row_reduce_looped</a></div><div class="ttdeci">void row_reduce_looped(const device T *in, device U *out, const constant size_t &row_size, const constant size_t &non_row_reductions, const constant int *shape, const constant size_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant size_t *reduce_strides, const constant int &reduce_ndim, uint3 gid, uint3 gsize, uint3 lid, uint3 lsize, uint simd_lane_id, uint simd_per_group, uint simd_group_id)</div><div class="ttdef"><b>Definition</b> reduce_row.h:314</div></div>
|
||||
<div class="ttc" id="areduce__row_8h_html_afd80a25fa84e6cc884dcc8698859ade1"><div class="ttname"><a href="reduce__row_8h.html#afd80a25fa84e6cc884dcc8698859ade1">thread_reduce</a></div><div class="ttdeci">METAL_FUNC void thread_reduce(thread U &total, const device T *row, int blocks, int extra)</div><div class="ttdef"><b>Definition</b> reduce_row.h:166</div></div>
|
||||
<div class="ttc" id="astruct_looped_elem_to_loc_html"><div class="ttname"><a href="struct_looped_elem_to_loc.html">LoopedElemToLoc</a></div><div class="ttdef"><b>Definition</b> utils.h:208</div></div>
|
||||
<div class="ttc" id="astruct_looped_elem_to_loc_html"><div class="ttname"><a href="struct_looped_elem_to_loc.html">LoopedElemToLoc</a></div><div class="ttdef"><b>Definition</b> utils.h:197</div></div>
|
||||
</div><!-- fragment --></div><!-- contents -->
|
||||
<!-- start footer part -->
|
||||
<hr class="footer"/><address class="footer"><small>
|
||||
Generated by <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_f60cd69d27fd3faa641c79056fff0e2d.html">reduction</a></li><li class="navelem"><a class="el" href="reduce__row_8h.html">reduce_row.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>
|
||||
|
Reference in New Issue
Block a user