mirror of
https://github.com/ml-explore/mlx.git
synced 2025-11-06 12:09:43 +08:00
rebase
This commit is contained in:
CircleCI Docs
505
docs/build/html/steel__attention_8h_source.html
vendored
505
docs/build/html/steel__attention_8h_source.html
vendored
@@ -177,7 +177,7 @@ $(function(){initNavTree('steel__attention_8h_source.html',''); initResizable(tr
|
||||
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span> <span class="keyword">template</span> <<span class="keyword">typename</span> T></div>
|
||||
<div class="foldopen" id="foldopen00052" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00052" name="l00052"></a><span class="lineno"><a class="line" href="struct_exp_sub_op.html#a00e457a01cb38f959dfd789455e7f334"> 52</a></span> METAL_FUNC <span class="keyword">static</span> <span class="keyword">constexpr</span> T <a class="code hl_function" href="struct_exp_sub_op.html#a00e457a01cb38f959dfd789455e7f334">apply</a>(T x, T y) {</div>
|
||||
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> <span class="keywordflow">return</span> fast::exp(x - y);</div>
|
||||
<div class="line"><a id="l00053" name="l00053"></a><span class="lineno"> 53</span> <span class="keywordflow">return</span> fast::exp2(x - y);</div>
|
||||
<div class="line"><a id="l00054" name="l00054"></a><span class="lineno"> 54</span> }</div>
|
||||
</div>
|
||||
<div class="line"><a id="l00055" name="l00055"></a><span class="lineno"> 55</span>};</div>
|
||||
@@ -237,254 +237,285 @@ $(function(){initNavTree('steel__attention_8h_source.html',''); initResizable(tr
|
||||
<div class="line"><a id="l00103" name="l00103"></a><span class="lineno"> 103</span> tidl.x * BQ * params->O_strides[2]; <span class="comment">// Seqeunce</span></div>
|
||||
<div class="line"><a id="l00104" name="l00104"></a><span class="lineno"> 104</span> </div>
|
||||
<div class="line"><a id="l00105" name="l00105"></a><span class="lineno"> 105</span> <span class="comment">// Prepare threadgroup memory</span></div>
|
||||
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padQ = 0; <span class="comment">// 16 / sizeof(T);</span></div>
|
||||
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padK = 0; <span class="comment">// 16 / sizeof(T);</span></div>
|
||||
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padV = 0; <span class="comment">// 16 / sizeof(T);</span></div>
|
||||
<div class="line"><a id="l00106" name="l00106"></a><span class="lineno"> 106</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padQ = 16 / <span class="keyword">sizeof</span>(T);</div>
|
||||
<div class="line"><a id="l00107" name="l00107"></a><span class="lineno"> 107</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padK = 16 / <span class="keyword">sizeof</span>(T);</div>
|
||||
<div class="line"><a id="l00108" name="l00108"></a><span class="lineno"> 108</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> padV = 16 / <span class="keyword">sizeof</span>(T);</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="keyword">constexpr</span> <span class="keywordtype">short</span> LDQ_tgp = BD + padQ;</div>
|
||||
<div class="line"><a id="l00111" name="l00111"></a><span class="lineno"> 111</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> LDK_tgp = BK + padK;</div>
|
||||
<div class="line"><a id="l00112" name="l00112"></a><span class="lineno"> 112</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> LDV_tgp = BD + padV;</div>
|
||||
<div class="line"><a id="l00113" name="l00113"></a><span class="lineno"> 113</span> </div>
|
||||
<div class="line"><a id="l00114" name="l00114"></a><span class="lineno"> 114</span> threadgroup T Qs[BQ * (BD + padQ)];</div>
|
||||
<div class="line"><a id="l00115" name="l00115"></a><span class="lineno"> 115</span> threadgroup T Ks[(BK + padK) * BD];</div>
|
||||
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> threadgroup T Vs[BK * (BD + padV)];</div>
|
||||
<div class="line"><a id="l00114" name="l00114"></a><span class="lineno"> 114</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> tgp_mem_0 = (BK + padK) * (BD);</div>
|
||||
<div class="line"><a id="l00115" name="l00115"></a><span class="lineno"> 115</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> tgp_mem_1 = BK * (BD + padV);</div>
|
||||
<div class="line"><a id="l00116" name="l00116"></a><span class="lineno"> 116</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> tgp_mem_s = tgp_mem_0 > tgp_mem_1 ? tgp_mem_0 : tgp_mem_1;</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="comment">// Prepare block loaders</span></div>
|
||||
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> <span class="keyword">using </span>QBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> <span class="comment">/* short BROWS = */</span> BQ,</div>
|
||||
<div class="line"><a id="l00122" name="l00122"></a><span class="lineno"> 122</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span> <span class="comment">/* short kDstStrRow = */</span> LDQ_tgp,</div>
|
||||
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> <span class="comment">/* short kDstStrCol = */</span> 1,</div>
|
||||
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> <span class="comment">/* short reduction_dim = */</span> 1,</div>
|
||||
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</div>
|
||||
<div class="line"><a id="l00127" name="l00127"></a><span class="lineno"> 127</span> </div>
|
||||
<div class="line"><a id="l00128" name="l00128"></a><span class="lineno"> 128</span> <span class="comment">// K is loaded in transposed</span></div>
|
||||
<div class="line"><a id="l00129" name="l00129"></a><span class="lineno"> 129</span> <span class="keyword">using </span>KBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00130" name="l00130"></a><span class="lineno"> 130</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00131" name="l00131"></a><span class="lineno"> 131</span> <span class="comment">/* short BROWS = */</span> BK,</div>
|
||||
<div class="line"><a id="l00132" name="l00132"></a><span class="lineno"> 132</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> <span class="comment">/* short kDstStrRow = */</span> 1,</div>
|
||||
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> <span class="comment">/* short kDstStrCol = */</span> LDK_tgp,</div>
|
||||
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> <span class="comment">/* short reduction_dim = */</span> 0,</div>
|
||||
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</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="keyword">using </span>VBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> <span class="comment">/* short BROWS = */</span> BK,</div>
|
||||
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"> 142</span> <span class="comment">/* short kDstStrRow = */</span> LDV_tgp,</div>
|
||||
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> <span class="comment">/* short kDstStrCol = */</span> 1,</div>
|
||||
<div class="line"><a id="l00144" name="l00144"></a><span class="lineno"> 144</span> <span class="comment">/* short reduction_dim = */</span> 0,</div>
|
||||
<div class="line"><a id="l00145" name="l00145"></a><span class="lineno"> 145</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</div>
|
||||
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> </div>
|
||||
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> QBlockLoader loader_q(</div>
|
||||
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> Q, params->Q_strides[2], Qs, simd_group_id, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> KBlockLoader loader_k(</div>
|
||||
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> K, params->K_strides[2], Ks, simd_group_id, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> VBlockLoader loader_v(</div>
|
||||
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> V, params->V_strides[2], Vs, simd_group_id, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00118" name="l00118"></a><span class="lineno"> 118</span> threadgroup T Q_smem[BQ * (BD + padQ)];</div>
|
||||
<div class="line"><a id="l00119" name="l00119"></a><span class="lineno"> 119</span> threadgroup T KV_smem[tgp_mem_s];</div>
|
||||
<div class="line"><a id="l00120" name="l00120"></a><span class="lineno"> 120</span> </div>
|
||||
<div class="line"><a id="l00121" name="l00121"></a><span class="lineno"> 121</span> threadgroup T* Qs = Q_smem;</div>
|
||||
<div class="line"><a id="l00122" name="l00122"></a><span class="lineno"> 122</span> threadgroup T* Ks = KV_smem;</div>
|
||||
<div class="line"><a id="l00123" name="l00123"></a><span class="lineno"> 123</span> threadgroup T* Vs = KV_smem;</div>
|
||||
<div class="line"><a id="l00124" name="l00124"></a><span class="lineno"> 124</span> </div>
|
||||
<div class="line"><a id="l00125" name="l00125"></a><span class="lineno"> 125</span> <span class="comment">// Prepare block loaders</span></div>
|
||||
<div class="line"><a id="l00126" name="l00126"></a><span class="lineno"> 126</span> <span class="keyword">using </span>QBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00127" name="l00127"></a><span class="lineno"> 127</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00128" name="l00128"></a><span class="lineno"> 128</span> <span class="comment">/* short BROWS = */</span> BQ,</div>
|
||||
<div class="line"><a id="l00129" name="l00129"></a><span class="lineno"> 129</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00130" name="l00130"></a><span class="lineno"> 130</span> <span class="comment">/* short kDstStrRow = */</span> LDQ_tgp,</div>
|
||||
<div class="line"><a id="l00131" name="l00131"></a><span class="lineno"> 131</span> <span class="comment">/* short kDstStrCol = */</span> 1,</div>
|
||||
<div class="line"><a id="l00132" name="l00132"></a><span class="lineno"> 132</span> <span class="comment">/* short reduction_dim = */</span> 1,</div>
|
||||
<div class="line"><a id="l00133" name="l00133"></a><span class="lineno"> 133</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</div>
|
||||
<div class="line"><a id="l00134" name="l00134"></a><span class="lineno"> 134</span> </div>
|
||||
<div class="line"><a id="l00135" name="l00135"></a><span class="lineno"> 135</span> <span class="comment">// K is loaded in transposed</span></div>
|
||||
<div class="line"><a id="l00136" name="l00136"></a><span class="lineno"> 136</span> <span class="keyword">using </span>KBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00137" name="l00137"></a><span class="lineno"> 137</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00138" name="l00138"></a><span class="lineno"> 138</span> <span class="comment">/* short BROWS = */</span> BK,</div>
|
||||
<div class="line"><a id="l00139" name="l00139"></a><span class="lineno"> 139</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00140" name="l00140"></a><span class="lineno"> 140</span> <span class="comment">/* short kDstStrRow = */</span> 1,</div>
|
||||
<div class="line"><a id="l00141" name="l00141"></a><span class="lineno"> 141</span> <span class="comment">/* short kDstStrCol = */</span> LDK_tgp,</div>
|
||||
<div class="line"><a id="l00142" name="l00142"></a><span class="lineno"> 142</span> <span class="comment">/* short reduction_dim = */</span> 0,</div>
|
||||
<div class="line"><a id="l00143" name="l00143"></a><span class="lineno"> 143</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</div>
|
||||
<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="keyword">using </span>VBlockLoader = <a class="code hl_struct" href="structmlx_1_1steel_1_1_block_loader_t.html">BlockLoaderT</a><</div>
|
||||
<div class="line"><a id="l00146" name="l00146"></a><span class="lineno"> 146</span> <span class="comment">/* typename T = */</span> T,</div>
|
||||
<div class="line"><a id="l00147" name="l00147"></a><span class="lineno"> 147</span> <span class="comment">/* short BROWS = */</span> BK,</div>
|
||||
<div class="line"><a id="l00148" name="l00148"></a><span class="lineno"> 148</span> <span class="comment">/* short BCOLS = */</span> BD,</div>
|
||||
<div class="line"><a id="l00149" name="l00149"></a><span class="lineno"> 149</span> <span class="comment">/* short kDstStrRow = */</span> LDV_tgp,</div>
|
||||
<div class="line"><a id="l00150" name="l00150"></a><span class="lineno"> 150</span> <span class="comment">/* short kDstStrCol = */</span> 1,</div>
|
||||
<div class="line"><a id="l00151" name="l00151"></a><span class="lineno"> 151</span> <span class="comment">/* short reduction_dim = */</span> 0,</div>
|
||||
<div class="line"><a id="l00152" name="l00152"></a><span class="lineno"> 152</span> <span class="comment">/* short tgp_size = */</span> WM * WN * 32>;</div>
|
||||
<div class="line"><a id="l00153" name="l00153"></a><span class="lineno"> 153</span> </div>
|
||||
<div class="line"><a id="l00154" name="l00154"></a><span class="lineno"> 154</span> <a class="code hl_struct" href="struct_transform_scale.html">TransformScale<T></a> ts(<span class="keyword">static_cast<</span>T<span class="keyword">></span>(params->scale));</div>
|
||||
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> </div>
|
||||
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> <span class="comment">// Prepare MMA tiles</span></div>
|
||||
<div class="line"><a id="l00157" name="l00157"></a><span class="lineno"> 157</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> kFragSize = 8; <span class="comment">// MMAFrag size</span></div>
|
||||
<div class="line"><a id="l00158" name="l00158"></a><span class="lineno"> 158</span> <span class="keyword">using </span>MMAFrag_acc_t = <a class="code hl_struct" href="structmlx_1_1steel_1_1_base_m_m_a_frag.html">BaseMMAFrag<AccumType, kFragSize, kFragSize></a>;</div>
|
||||
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> </div>
|
||||
<div class="line"><a id="l00160" name="l00160"></a><span class="lineno"> 160</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> kNWarps = WM * WN;</div>
|
||||
<div class="line"><a id="l00161" name="l00161"></a><span class="lineno"> 161</span> <span class="keyword">static_assert</span>(</div>
|
||||
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span> BQ >= (kNWarps * kFragSize) && BQ % (kNWarps * kFragSize) == 0,</div>
|
||||
<div class="line"><a id="l00163" name="l00163"></a><span class="lineno"> 163</span> <span class="stringliteral">"Each simdgroup must host atleast 1 simdgroup matrix along Q sequence."</span>);</div>
|
||||
<div class="line"><a id="l00164" name="l00164"></a><span class="lineno"> 164</span> </div>
|
||||
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span> <span class="comment">// Q seq frags per warp</span></div>
|
||||
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"> 166</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TQ = BQ / (kNWarps * kFragSize);</div>
|
||||
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> <span class="comment">// KV sequence frags (all warps load the same frags)</span></div>
|
||||
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TK = BK / kFragSize;</div>
|
||||
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> <span class="comment">// HeadDim frags (all warps load the same frags)</span></div>
|
||||
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TD = BD / kFragSize;</div>
|
||||
<div class="line"><a id="l00154" name="l00154"></a><span class="lineno"> 154</span> QBlockLoader loader_q(</div>
|
||||
<div class="line"><a id="l00155" name="l00155"></a><span class="lineno"> 155</span> Q, params->Q_strides[2], Qs, simd_group_id, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00156" name="l00156"></a><span class="lineno"> 156</span> KBlockLoader loader_k(</div>
|
||||
<div class="line"><a id="l00157" name="l00157"></a><span class="lineno"> 157</span> K, params->K_strides[2], Ks, simd_group_id, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00158" name="l00158"></a><span class="lineno"> 158</span> VBlockLoader loader_v(</div>
|
||||
<div class="line"><a id="l00159" name="l00159"></a><span class="lineno"> 159</span> V, params->V_strides[2], Vs, simd_group_id, simd_lane_id);</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> <a class="code hl_struct" href="struct_transform_scale.html">TransformScale<T></a> ts(<span class="keyword">static_cast<</span>T<span class="keyword">></span>(params->scale * 1.44269504089));</div>
|
||||
<div class="line"><a id="l00162" name="l00162"></a><span class="lineno"> 162</span> </div>
|
||||
<div class="line"><a id="l00163" name="l00163"></a><span class="lineno"> 163</span> <span class="comment">// Prepare MMA tiles</span></div>
|
||||
<div class="line"><a id="l00164" name="l00164"></a><span class="lineno"> 164</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> kFragSize = 8; <span class="comment">// MMAFrag size</span></div>
|
||||
<div class="line"><a id="l00165" name="l00165"></a><span class="lineno"> 165</span> <span class="keyword">using </span>MMAFrag_acc_t = <a class="code hl_struct" href="structmlx_1_1steel_1_1_base_m_m_a_frag.html">BaseMMAFrag<AccumType, kFragSize, kFragSize></a>;</div>
|
||||
<div class="line"><a id="l00166" name="l00166"></a><span class="lineno"> 166</span> </div>
|
||||
<div class="line"><a id="l00167" name="l00167"></a><span class="lineno"> 167</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> kNWarps = WM * WN;</div>
|
||||
<div class="line"><a id="l00168" name="l00168"></a><span class="lineno"> 168</span> <span class="keyword">static_assert</span>(</div>
|
||||
<div class="line"><a id="l00169" name="l00169"></a><span class="lineno"> 169</span> BQ >= (kNWarps * kFragSize) && BQ % (kNWarps * kFragSize) == 0,</div>
|
||||
<div class="line"><a id="l00170" name="l00170"></a><span class="lineno"> 170</span> <span class="stringliteral">"Each simdgroup must host atleast 1 simdgroup matrix along Q sequence."</span>);</div>
|
||||
<div class="line"><a id="l00171" name="l00171"></a><span class="lineno"> 171</span> </div>
|
||||
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> <span class="keyword">static_assert</span>(TQ == 1, <span class="stringliteral">"Check TQ"</span>);</div>
|
||||
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> </div>
|
||||
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, 1, MMAFrag_acc_t></a> Qtile;</div>
|
||||
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, 1, TK, MMAFrag_acc_t></a> Ktile;</div>
|
||||
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, TK, MMAFrag_acc_t></a> Stile;</div>
|
||||
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TK, TD, MMAFrag_acc_t></a> Vtile;</div>
|
||||
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, TD, MMAFrag_acc_t></a> Otile;</div>
|
||||
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> </div>
|
||||
<div class="line"><a id="l00180" name="l00180"></a><span class="lineno"> 180</span> Otile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">clear</a>();</div>
|
||||
<div class="line"><a id="l00181" name="l00181"></a><span class="lineno"> 181</span> </div>
|
||||
<div class="line"><a id="l00182" name="l00182"></a><span class="lineno"> 182</span> <span class="comment">// Prepare mma tile offsets</span></div>
|
||||
<div class="line"><a id="l00183" name="l00183"></a><span class="lineno"> 183</span> <span class="keyword">const</span> short2 simd_coord = MMAFrag_acc_t::get_coord(simd_lane_id);</div>
|
||||
<div class="line"><a id="l00184" name="l00184"></a><span class="lineno"> 184</span> <span class="keyword">const</span> <span class="keywordtype">short</span> sm = simd_coord.y;</div>
|
||||
<div class="line"><a id="l00185" name="l00185"></a><span class="lineno"> 185</span> <span class="keyword">const</span> <span class="keywordtype">short</span> sn = simd_coord.x;</div>
|
||||
<div class="line"><a id="l00186" name="l00186"></a><span class="lineno"> 186</span> <span class="keyword">const</span> <span class="keywordtype">short</span> tm = kFragSize * TQ * simd_group_id;</div>
|
||||
<div class="line"><a id="l00187" name="l00187"></a><span class="lineno"> 187</span> </div>
|
||||
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Qs_offset = (tm + sm) * LDQ_tgp + sn;</div>
|
||||
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Ks_offset = sm * LDK_tgp + sn;</div>
|
||||
<div class="line"><a id="l00190" name="l00190"></a><span class="lineno"> 190</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Vs_offset = sm * LDV_tgp + sn;</div>
|
||||
<div class="line"><a id="l00191" name="l00191"></a><span class="lineno"> 191</span> </div>
|
||||
<div class="line"><a id="l00192" name="l00192"></a><span class="lineno"> 192</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> Qs_tile_stride = kFragSize;</div>
|
||||
<div class="line"><a id="l00193" name="l00193"></a><span class="lineno"> 193</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> Ks_tile_stride = kFragSize * LDK_tgp;</div>
|
||||
<div class="line"><a id="l00172" name="l00172"></a><span class="lineno"> 172</span> <span class="comment">// Q seq frags per warp</span></div>
|
||||
<div class="line"><a id="l00173" name="l00173"></a><span class="lineno"> 173</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TQ = BQ / (kNWarps * kFragSize);</div>
|
||||
<div class="line"><a id="l00174" name="l00174"></a><span class="lineno"> 174</span> <span class="comment">// KV sequence frags (all warps load the same frags)</span></div>
|
||||
<div class="line"><a id="l00175" name="l00175"></a><span class="lineno"> 175</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TK = BK / kFragSize;</div>
|
||||
<div class="line"><a id="l00176" name="l00176"></a><span class="lineno"> 176</span> <span class="comment">// HeadDim frags (all warps load the same frags)</span></div>
|
||||
<div class="line"><a id="l00177" name="l00177"></a><span class="lineno"> 177</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> TD = BD / kFragSize;</div>
|
||||
<div class="line"><a id="l00178" name="l00178"></a><span class="lineno"> 178</span> </div>
|
||||
<div class="line"><a id="l00179" name="l00179"></a><span class="lineno"> 179</span> <span class="keyword">static_assert</span>(TQ == 1, <span class="stringliteral">"Check TQ"</span>);</div>
|
||||
<div class="line"><a id="l00180" name="l00180"></a><span class="lineno"> 180</span> </div>
|
||||
<div class="line"><a id="l00181" name="l00181"></a><span class="lineno"> 181</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, 1, MMAFrag_acc_t></a> Qtile;</div>
|
||||
<div class="line"><a id="l00182" name="l00182"></a><span class="lineno"> 182</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, 1, TK, MMAFrag_acc_t></a> Ktile;</div>
|
||||
<div class="line"><a id="l00183" name="l00183"></a><span class="lineno"> 183</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, TK, MMAFrag_acc_t></a> Stile;</div>
|
||||
<div class="line"><a id="l00184" name="l00184"></a><span class="lineno"> 184</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, 1, 1, MMAFrag_acc_t></a> Vtile;</div>
|
||||
<div class="line"><a id="l00185" name="l00185"></a><span class="lineno"> 185</span> <a class="code hl_struct" href="structmlx_1_1steel_1_1_m_m_a_tile.html">MMATile<AccumType, TQ, TD, MMAFrag_acc_t></a> Otile;</div>
|
||||
<div class="line"><a id="l00186" name="l00186"></a><span class="lineno"> 186</span> </div>
|
||||
<div class="line"><a id="l00187" name="l00187"></a><span class="lineno"> 187</span> Otile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">clear</a>();</div>
|
||||
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span> </div>
|
||||
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> <span class="comment">// Prepare mma tile offsets</span></div>
|
||||
<div class="line"><a id="l00190" name="l00190"></a><span class="lineno"> 190</span> <span class="keyword">const</span> short2 simd_coord = MMAFrag_acc_t::get_coord(simd_lane_id);</div>
|
||||
<div class="line"><a id="l00191" name="l00191"></a><span class="lineno"> 191</span> <span class="keyword">const</span> <span class="keywordtype">short</span> sm = simd_coord.y;</div>
|
||||
<div class="line"><a id="l00192" name="l00192"></a><span class="lineno"> 192</span> <span class="keyword">const</span> <span class="keywordtype">short</span> sn = simd_coord.x;</div>
|
||||
<div class="line"><a id="l00193" name="l00193"></a><span class="lineno"> 193</span> <span class="keyword">const</span> <span class="keywordtype">short</span> tm = kFragSize * TQ * simd_group_id;</div>
|
||||
<div class="line"><a id="l00194" name="l00194"></a><span class="lineno"> 194</span> </div>
|
||||
<div class="line"><a id="l00195" name="l00195"></a><span class="lineno"> 195</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00196" name="l00196"></a><span class="lineno"> 196</span> </div>
|
||||
<div class="line"><a id="l00197" name="l00197"></a><span class="lineno"> 197</span> <span class="comment">// Load Q blocks apply scale</span></div>
|
||||
<div class="line"><a id="l00198" name="l00198"></a><span class="lineno"> 198</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a171fdea1b23976453f5dc5e6b3161982">align_Q</a> && <span class="keywordtype">int</span>(tid.x) == (params->NQ_aligned)) {</div>
|
||||
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"> 199</span> loader_q.load_safe(short2(BD, params->qL - params->NQ_aligned * BQ));</div>
|
||||
<div class="line"><a id="l00200" name="l00200"></a><span class="lineno"> 200</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span> loader_q.load_unsafe();</div>
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> }</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> loader_q.apply_inplace_op(ts);</div>
|
||||
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> </div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="comment">// Init row reduction variables</span></div>
|
||||
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> kRowsPT = <span class="keyword">decltype</span>(Stile)::kRowsPerThread;</div>
|
||||
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> </div>
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> AccumType max_score[kRowsPT];</div>
|
||||
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> AccumType sum_score[kRowsPT] = {0};</div>
|
||||
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> </div>
|
||||
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> <span class="comment">// Init to -Inf</span></div>
|
||||
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> max_score[i] = <a class="code hl_variable" href="struct_limits.html#a6e81584ba65a4dc6ff9366b458e3a20e">Limits<AccumType>::min</a>;</div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> }</div>
|
||||
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> </div>
|
||||
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> <span class="comment">// Loop over KV seq length</span></div>
|
||||
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> kb = 0; kb < params->NK; kb++) {</div>
|
||||
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"> 219</span> <span class="comment">// Load K block and apply scale</span></div>
|
||||
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</span> loader_k.load_safe(short2(BD, params->kL - params->NK_aligned * BK));</div>
|
||||
<div class="line"><a id="l00223" name="l00223"></a><span class="lineno"> 223</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00224" name="l00224"></a><span class="lineno"> 224</span> loader_k.load_unsafe();</div>
|
||||
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> }</div>
|
||||
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> </div>
|
||||
<div class="line"><a id="l00195" name="l00195"></a><span class="lineno"> 195</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Qs_offset = (tm + sm) * LDQ_tgp + sn;</div>
|
||||
<div class="line"><a id="l00196" name="l00196"></a><span class="lineno"> 196</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Ks_offset = sm * LDK_tgp + sn;</div>
|
||||
<div class="line"><a id="l00197" name="l00197"></a><span class="lineno"> 197</span> <span class="keyword">const</span> <span class="keywordtype">short</span> Vs_offset = sm * LDV_tgp + sn;</div>
|
||||
<div class="line"><a id="l00198" name="l00198"></a><span class="lineno"> 198</span> </div>
|
||||
<div class="line"><a id="l00199" name="l00199"></a><span class="lineno"> 199</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> Qs_tile_stride = kFragSize;</div>
|
||||
<div class="line"><a id="l00200" name="l00200"></a><span class="lineno"> 200</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> Ks_tile_stride = kFragSize * LDK_tgp;</div>
|
||||
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span> </div>
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> </div>
|
||||
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> <span class="comment">// Load Q blocks apply scale</span></div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a171fdea1b23976453f5dc5e6b3161982">align_Q</a> && <span class="keywordtype">int</span>(tid.x) == (params->NQ_aligned)) {</div>
|
||||
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> loader_q.load_safe(short2(BD, params->qL - params->NQ_aligned * BQ));</div>
|
||||
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> loader_q.load_unsafe();</div>
|
||||
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> }</div>
|
||||
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> loader_q.apply_inplace_op(ts);</div>
|
||||
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> </div>
|
||||
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> <span class="comment">// Init row reduction variables</span></div>
|
||||
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> kRowsPT = <span class="keyword">decltype</span>(Stile)::kRowsPerThread;</div>
|
||||
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> </div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> AccumType max_score[kRowsPT];</div>
|
||||
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> AccumType sum_score[kRowsPT] = {0};</div>
|
||||
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> </div>
|
||||
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> <span class="comment">// Init to -Inf</span></div>
|
||||
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"> 219</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> max_score[i] = <a class="code hl_variable" href="struct_limits.html#a6e81584ba65a4dc6ff9366b458e3a20e">Limits<AccumType>::min</a>;</div>
|
||||
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</span> }</div>
|
||||
<div class="line"><a id="l00223" name="l00223"></a><span class="lineno"> 223</span> </div>
|
||||
<div class="line"><a id="l00224" name="l00224"></a><span class="lineno"> 224</span> <span class="comment">// Loop over KV seq length</span></div>
|
||||
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> kb = 0; kb < params->NK; kb++) {</div>
|
||||
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> <span class="comment">// Load K block and apply scale</span></div>
|
||||
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> </div>
|
||||
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> <span class="comment">// Do S = Q @ K.T</span></div>
|
||||
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> Stile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">clear</a>();</div>
|
||||
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> </div>
|
||||
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> dd = 0; dd < TD; dd++) {</div>
|
||||
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> simdgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> </div>
|
||||
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> Qtile.template load<T, 1, 1, LDQ_tgp, 1>(</div>
|
||||
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> &Qs[Qs_offset + dd * Qs_tile_stride]);</div>
|
||||
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> Ktile.template load<T, 1, 1, LDK_tgp, 1>(</div>
|
||||
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> &Ks[Ks_offset + dd * Ks_tile_stride]);</div>
|
||||
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> </div>
|
||||
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> simdgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> </div>
|
||||
<div class="line"><a id="l00242" name="l00242"></a><span class="lineno"> 242</span> <a class="code hl_function" href="namespacemlx_1_1steel.html#ad583e6038efc119542410f43b603d4ad">tile_matmad</a>(Stile, Qtile, Ktile, Stile);</div>
|
||||
<div class="line"><a id="l00243" name="l00243"></a><span class="lineno"> 243</span> }</div>
|
||||
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> </div>
|
||||
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> <span class="comment">// Mask out of length sequence</span></div>
|
||||
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00247" name="l00247"></a><span class="lineno"> 247</span> <span class="keyword">using </span>stile_t = <span class="keyword">decltype</span>(Stile);</div>
|
||||
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> <span class="keyword">using </span>selem_t = <span class="keyword">typename</span> stile_t::elem_type;</div>
|
||||
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> <span class="keyword">constexpr</span> <span class="keyword">auto</span> neg_inf = -metal::numeric_limits<selem_t>::infinity();</div>
|
||||
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> <span class="keyword">const</span> <span class="keywordtype">short</span> lim = params->kL - params->NK_aligned * BK;</div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> </div>
|
||||
<div class="line"><a id="l00252" name="l00252"></a><span class="lineno"> 252</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00253" name="l00253"></a><span class="lineno"> 253</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < stile_t::kTileRows; i++) {</div>
|
||||
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> j = 0; j < stile_t::kTileCols; j++) {</div>
|
||||
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> <span class="keywordtype">short</span> col_pos = sn + (j * stile_t::kFragCols);</div>
|
||||
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> jj = 0; jj < stile_t::MMAFrag_t::kElemCols; jj++) {</div>
|
||||
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> <span class="keywordflow">if</span> ((col_pos + jj) >= lim) {</div>
|
||||
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> Stile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(i, j)[jj] = neg_inf;</div>
|
||||
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> }</div>
|
||||
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"> 262</span> }</div>
|
||||
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> }</div>
|
||||
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> }</div>
|
||||
<div class="line"><a id="l00265" name="l00265"></a><span class="lineno"> 265</span> }</div>
|
||||
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"> 266</span> </div>
|
||||
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> simdgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> </div>
|
||||
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span> <span class="comment">// Load V blocks</span></div>
|
||||
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span> loader_v.load_safe(short2(BD, params->kL - params->NK_aligned * BK));</div>
|
||||
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> loader_v.load_unsafe();</div>
|
||||
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"> 274</span> }</div>
|
||||
<div class="line"><a id="l00275" name="l00275"></a><span class="lineno"> 275</span> </div>
|
||||
<div class="line"><a id="l00276" name="l00276"></a><span class="lineno"> 276</span> <span class="comment">// Do softmax</span></div>
|
||||
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"> 277</span> </div>
|
||||
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> <span class="comment">// Temp variables</span></div>
|
||||
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"> 279</span> AccumType new_max[kRowsPT];</div>
|
||||
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"> 280</span> AccumType factor[kRowsPT];</div>
|
||||
<div class="line"><a id="l00281" name="l00281"></a><span class="lineno"> 281</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00282" name="l00282"></a><span class="lineno"> 282</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> new_max[i] = max_score[i];</div>
|
||||
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"> 284</span> }</div>
|
||||
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> loader_k.load_safe(short2(BD, params->kL - params->NK_aligned * BK));</div>
|
||||
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> loader_k.load_unsafe();</div>
|
||||
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> }</div>
|
||||
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> </div>
|
||||
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> <span class="comment">// Do S = Q @ K.T</span></div>
|
||||
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> Stile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">clear</a>();</div>
|
||||
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> </div>
|
||||
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> </div>
|
||||
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> dd = 0; dd < TD; dd++) {</div>
|
||||
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> simdgroup_barrier(mem_flags::mem_none);</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> Qtile.template load<T, 1, 1, LDQ_tgp, 1>(</div>
|
||||
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> &Qs[Qs_offset + dd * Qs_tile_stride]);</div>
|
||||
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> Ktile.template load<T, 1, 1, LDK_tgp, 1>(</div>
|
||||
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> &Ks[Ks_offset + dd * Ks_tile_stride]);</div>
|
||||
<div class="line"><a id="l00247" name="l00247"></a><span class="lineno"> 247</span> </div>
|
||||
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> simdgroup_barrier(mem_flags::mem_none);</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> <a class="code hl_function" href="namespacemlx_1_1steel.html#ab9fdcb06fb1f639f9120ab14cfedd150">tile_matmad</a>(Stile, Qtile, Ktile, Stile);</div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> }</div>
|
||||
<div class="line"><a id="l00252" name="l00252"></a><span class="lineno"> 252</span> </div>
|
||||
<div class="line"><a id="l00253" name="l00253"></a><span class="lineno"> 253</span> <span class="comment">// Mask out of length sequence</span></div>
|
||||
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> <span class="keyword">using </span>stile_t = <span class="keyword">decltype</span>(Stile);</div>
|
||||
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> <span class="keyword">using </span>selem_t = <span class="keyword">typename</span> stile_t::elem_type;</div>
|
||||
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> <span class="keyword">constexpr</span> <span class="keyword">auto</span> neg_inf = -metal::numeric_limits<selem_t>::infinity();</div>
|
||||
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> <span class="keyword">const</span> <span class="keywordtype">short</span> lim = params->kL - params->NK_aligned * BK;</div>
|
||||
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> </div>
|
||||
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < stile_t::kTileRows; i++) {</div>
|
||||
<div class="line"><a id="l00262" name="l00262"></a><span class="lineno"> 262</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00263" name="l00263"></a><span class="lineno"> 263</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> j = 0; j < stile_t::kTileCols; j++) {</div>
|
||||
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> <span class="keywordtype">short</span> col_pos = sn + (j * stile_t::kFragCols);</div>
|
||||
<div class="line"><a id="l00265" name="l00265"></a><span class="lineno"> 265</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"> 266</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> jj = 0; jj < stile_t::MMAFrag_t::kElemCols; jj++) {</div>
|
||||
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> <span class="keywordflow">if</span> ((col_pos + jj) >= lim) {</div>
|
||||
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> Stile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(i, j)[jj] = neg_inf;</div>
|
||||
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span> }</div>
|
||||
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> }</div>
|
||||
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span> }</div>
|
||||
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> }</div>
|
||||
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> }</div>
|
||||
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"> 274</span> </div>
|
||||
<div class="line"><a id="l00275" name="l00275"></a><span class="lineno"> 275</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00276" name="l00276"></a><span class="lineno"> 276</span> </div>
|
||||
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"> 277</span> <span class="comment">// Load V blocks</span></div>
|
||||
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a8bdd2cecf97aa5b033152b1d0f0d2416">align_K</a> && kb == (params->NK_aligned)) {</div>
|
||||
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"> 279</span> loader_v.load_safe(short2(BD, params->kL - params->NK_aligned * BK));</div>
|
||||
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"> 280</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00281" name="l00281"></a><span class="lineno"> 281</span> loader_v.load_unsafe();</div>
|
||||
<div class="line"><a id="l00282" name="l00282"></a><span class="lineno"> 282</span> }</div>
|
||||
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> </div>
|
||||
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"> 284</span> <span class="comment">// Do softmax</span></div>
|
||||
<div class="line"><a id="l00285" name="l00285"></a><span class="lineno"> 285</span> </div>
|
||||
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"> 286</span> <span class="comment">// Row max</span></div>
|
||||
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> Stile.template row_reduce<MaxOp>(new_max);</div>
|
||||
<div class="line"><a id="l00288" name="l00288"></a><span class="lineno"> 288</span> </div>
|
||||
<div class="line"><a id="l00289" name="l00289"></a><span class="lineno"> 289</span> <span class="comment">// exp(Si - rowmax(Si))</span></div>
|
||||
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"> 290</span> Stile.template row_bin_op<ExpSubOp>(new_max);</div>
|
||||
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"> 291</span> </div>
|
||||
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> <span class="comment">// Factor exp(rowmax(Si) - rowmax(Si-1))</span></div>
|
||||
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"> 293</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> factor[i] = fast::exp(max_score[i] - new_max[i]);</div>
|
||||
<div class="line"><a id="l00296" name="l00296"></a><span class="lineno"> 296</span> }</div>
|
||||
<div class="line"><a id="l00297" name="l00297"></a><span class="lineno"> 297</span> </div>
|
||||
<div class="line"><a id="l00298" name="l00298"></a><span class="lineno"> 298</span> <span class="comment">// Save max for next iteration</span></div>
|
||||
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> max_score[i] = new_max[i];</div>
|
||||
<div class="line"><a id="l00302" name="l00302"></a><span class="lineno"> 302</span> }</div>
|
||||
<div class="line"><a id="l00303" name="l00303"></a><span class="lineno"> 303</span> </div>
|
||||
<div class="line"><a id="l00304" name="l00304"></a><span class="lineno"> 304</span> <span class="comment">// Row Sum</span></div>
|
||||
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> AccumType sum_score_tmp[kRowsPT] = {0};</div>
|
||||
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> Stile.template row_reduce<SumOp>(sum_score_tmp);</div>
|
||||
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> </div>
|
||||
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <span class="comment">// Update norm</span></div>
|
||||
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> sum_score[i] = sum_score[i] * factor[i] + sum_score_tmp[i];</div>
|
||||
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> }</div>
|
||||
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> </div>
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> <span class="comment">// Update O</span></div>
|
||||
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> Otile.template row_bin_op<MulOp>(factor);</div>
|
||||
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> </div>
|
||||
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> <span class="comment">// Load V into registers</span></div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> Vtile.template load<T, 1, 1, LDV_tgp, 1>(&Vs[Vs_offset]);</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> </div>
|
||||
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> simdgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00322" name="l00322"></a><span class="lineno"> 322</span> </div>
|
||||
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> <span class="comment">// Do O = S @ V</span></div>
|
||||
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> <a class="code hl_function" href="namespacemlx_1_1steel.html#ad583e6038efc119542410f43b603d4ad">tile_matmad</a>(Otile, Stile, Vtile, Otile);</div>
|
||||
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> </div>
|
||||
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> <span class="comment">// Prepare for next iteration</span></div>
|
||||
<div class="line"><a id="l00327" name="l00327"></a><span class="lineno"> 327</span> loader_k.next();</div>
|
||||
<div class="line"><a id="l00328" name="l00328"></a><span class="lineno"> 328</span> loader_v.next();</div>
|
||||
<div class="line"><a id="l00329" name="l00329"></a><span class="lineno"> 329</span> }</div>
|
||||
<div class="line"><a id="l00330" name="l00330"></a><span class="lineno"> 330</span> </div>
|
||||
<div class="line"><a id="l00331" name="l00331"></a><span class="lineno"> 331</span> <span class="comment">// Normalize output</span></div>
|
||||
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> Otile.template row_bin_op<DivOp>(sum_score);</div>
|
||||
<div class="line"><a id="l00333" name="l00333"></a><span class="lineno"> 333</span> threadgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00334" name="l00334"></a><span class="lineno"> 334</span> </div>
|
||||
<div class="line"><a id="l00335" name="l00335"></a><span class="lineno"> 335</span> <span class="comment">// Store results</span></div>
|
||||
<div class="line"><a id="l00336" name="l00336"></a><span class="lineno"> 336</span> O += (tm + sm) * params->O_strides[2] + sn;</div>
|
||||
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"> 286</span> <span class="comment">// Temp variables</span></div>
|
||||
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> AccumType new_max[kRowsPT];</div>
|
||||
<div class="line"><a id="l00288" name="l00288"></a><span class="lineno"> 288</span> AccumType factor[kRowsPT];</div>
|
||||
<div class="line"><a id="l00289" name="l00289"></a><span class="lineno"> 289</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"> 290</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"> 291</span> new_max[i] = max_score[i];</div>
|
||||
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> }</div>
|
||||
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"> 293</span> </div>
|
||||
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> <span class="comment">// Row max</span></div>
|
||||
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> Stile.template row_reduce<MaxOp>(new_max);</div>
|
||||
<div class="line"><a id="l00296" name="l00296"></a><span class="lineno"> 296</span> </div>
|
||||
<div class="line"><a id="l00297" name="l00297"></a><span class="lineno"> 297</span> <span class="comment">// exp(Si - rowmax(Si))</span></div>
|
||||
<div class="line"><a id="l00298" name="l00298"></a><span class="lineno"> 298</span> Stile.template row_bin_op<ExpSubOp>(new_max);</div>
|
||||
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> </div>
|
||||
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> <span class="comment">// Factor exp(rowmax(Si) - rowmax(Si-1))</span></div>
|
||||
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00302" name="l00302"></a><span class="lineno"> 302</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00303" name="l00303"></a><span class="lineno"> 303</span> factor[i] = fast::exp2(max_score[i] - new_max[i]);</div>
|
||||
<div class="line"><a id="l00304" name="l00304"></a><span class="lineno"> 304</span> }</div>
|
||||
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> </div>
|
||||
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> <span class="comment">// Save max for next iteration</span></div>
|
||||
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> max_score[i] = new_max[i];</div>
|
||||
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> }</div>
|
||||
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> </div>
|
||||
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> <span class="comment">// Row Sum</span></div>
|
||||
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> AccumType sum_score_tmp[kRowsPT] = {0};</div>
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> Stile.template row_reduce<SumOp>(sum_score_tmp);</div>
|
||||
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> </div>
|
||||
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> <span class="comment">// Update norm</span></div>
|
||||
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> i = 0; i < kRowsPT; ++i) {</div>
|
||||
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> sum_score[i] = sum_score[i] * factor[i] + sum_score_tmp[i];</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> }</div>
|
||||
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> </div>
|
||||
<div class="line"><a id="l00322" name="l00322"></a><span class="lineno"> 322</span> <span class="comment">// Update O</span></div>
|
||||
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> Otile.template row_bin_op<MulOp>(factor);</div>
|
||||
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> </div>
|
||||
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> <span class="comment">// Load V into registers</span></div>
|
||||
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00327" name="l00327"></a><span class="lineno"> 327</span> </div>
|
||||
<div class="line"><a id="l00328" name="l00328"></a><span class="lineno"> 328</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00329" name="l00329"></a><span class="lineno"> 329</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> iq = 0; iq < TQ; iq++) {</div>
|
||||
<div class="line"><a id="l00330" name="l00330"></a><span class="lineno"> 330</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00331" name="l00331"></a><span class="lineno"> 331</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> <span class="keywordtype">id</span> = 0; <span class="keywordtype">id</span> < TD; <span class="keywordtype">id</span>++) {</div>
|
||||
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> <a class="code hl_define" href="steel_2defines_8h.html#a5a5c3095b132a7589bc19cd5cb80e2c6">STEEL_PRAGMA_UNROLL</a></div>
|
||||
<div class="line"><a id="l00333" name="l00333"></a><span class="lineno"> 333</span> <span class="keywordflow">for</span> (<span class="keywordtype">short</span> ik = 0; ik < TK; ik++) {</div>
|
||||
<div class="line"><a id="l00334" name="l00334"></a><span class="lineno"> 334</span> <span class="keywordflow">if</span> <span class="keyword">constexpr</span> (BD == 128) {</div>
|
||||
<div class="line"><a id="l00335" name="l00335"></a><span class="lineno"> 335</span> simdgroup_barrier(mem_flags::mem_none);</div>
|
||||
<div class="line"><a id="l00336" name="l00336"></a><span class="lineno"> 336</span> }</div>
|
||||
<div class="line"><a id="l00337" name="l00337"></a><span class="lineno"> 337</span> </div>
|
||||
<div class="line"><a id="l00338" name="l00338"></a><span class="lineno"> 338</span> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a171fdea1b23976453f5dc5e6b3161982">align_Q</a> && <span class="keywordtype">int</span>(tid.x) == (params->NQ_aligned)) {</div>
|
||||
<div class="line"><a id="l00339" name="l00339"></a><span class="lineno"> 339</span> <span class="keyword">auto</span> dst_tile_dims =</div>
|
||||
<div class="line"><a id="l00340" name="l00340"></a><span class="lineno"> 340</span> short2(BD - sn, params->qL - BQ * params->NQ_aligned - (tm + sm));</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> <span class="keywordflow">if</span> (dst_tile_dims.x <= 0 || dst_tile_dims.y <= 0)</div>
|
||||
<div class="line"><a id="l00343" name="l00343"></a><span class="lineno"> 343</span> <span class="keywordflow">return</span>;</div>
|
||||
<div class="line"><a id="l00344" name="l00344"></a><span class="lineno"> 344</span> </div>
|
||||
<div class="line"><a id="l00345" name="l00345"></a><span class="lineno"> 345</span> Otile.template store_safe<T, 1, 1>(O, params->O_strides[2], dst_tile_dims);</div>
|
||||
<div class="line"><a id="l00346" name="l00346"></a><span class="lineno"> 346</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00347" name="l00347"></a><span class="lineno"> 347</span> Otile.template store<T, 1, 1>(O, params->O_strides[2]);</div>
|
||||
<div class="line"><a id="l00348" name="l00348"></a><span class="lineno"> 348</span> }</div>
|
||||
<div class="line"><a id="l00349" name="l00349"></a><span class="lineno"> 349</span>}</div>
|
||||
<div class="line"><a id="l00338" name="l00338"></a><span class="lineno"> 338</span> <span class="keyword">const</span> <span class="keywordtype">short</span> kk = ik * kFragSize;</div>
|
||||
<div class="line"><a id="l00339" name="l00339"></a><span class="lineno"> 339</span> <span class="keyword">const</span> <span class="keywordtype">short</span> dd = <span class="keywordtype">id</span> * kFragSize;</div>
|
||||
<div class="line"><a id="l00340" name="l00340"></a><span class="lineno"> 340</span> </div>
|
||||
<div class="line"><a id="l00341" name="l00341"></a><span class="lineno"> 341</span> Vtile.template load<T, 1, 1, LDV_tgp, 1>(</div>
|
||||
<div class="line"><a id="l00342" name="l00342"></a><span class="lineno"> 342</span> &Vs[Vs_offset + kk * LDV_tgp + dd]);</div>
|
||||
<div class="line"><a id="l00343" name="l00343"></a><span class="lineno"> 343</span> </div>
|
||||
<div class="line"><a id="l00344" name="l00344"></a><span class="lineno"> 344</span> <span class="keywordflow">if</span> <span class="keyword">constexpr</span> (BD == 128) {</div>
|
||||
<div class="line"><a id="l00345" name="l00345"></a><span class="lineno"> 345</span> simdgroup_barrier(mem_flags::mem_none);</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> </div>
|
||||
<div class="line"><a id="l00348" name="l00348"></a><span class="lineno"> 348</span> MMAFrag_acc_t::mma(</div>
|
||||
<div class="line"><a id="l00349" name="l00349"></a><span class="lineno"> 349</span> Otile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(iq, <span class="keywordtype">id</span>),</div>
|
||||
<div class="line"><a id="l00350" name="l00350"></a><span class="lineno"> 350</span> Stile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(iq, ik),</div>
|
||||
<div class="line"><a id="l00351" name="l00351"></a><span class="lineno"> 351</span> Vtile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(0, 0),</div>
|
||||
<div class="line"><a id="l00352" name="l00352"></a><span class="lineno"> 352</span> Otile.<a class="code hl_function" href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">frag_at</a>(iq, <span class="keywordtype">id</span>));</div>
|
||||
<div class="line"><a id="l00353" name="l00353"></a><span class="lineno"> 353</span> }</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> }</div>
|
||||
<div class="line"><a id="l00356" name="l00356"></a><span class="lineno"> 356</span> </div>
|
||||
<div class="line"><a id="l00357" name="l00357"></a><span class="lineno"> 357</span> <span class="comment">// Prepare for next iteration</span></div>
|
||||
<div class="line"><a id="l00358" name="l00358"></a><span class="lineno"> 358</span> loader_k.next();</div>
|
||||
<div class="line"><a id="l00359" name="l00359"></a><span class="lineno"> 359</span> loader_v.next();</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">// Normalize output</span></div>
|
||||
<div class="line"><a id="l00363" name="l00363"></a><span class="lineno"> 363</span> Otile.template row_bin_op<DivOp>(sum_score);</div>
|
||||
<div class="line"><a id="l00364" name="l00364"></a><span class="lineno"> 364</span> threadgroup_barrier(mem_flags::mem_none);</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">// Store results</span></div>
|
||||
<div class="line"><a id="l00367" name="l00367"></a><span class="lineno"> 367</span> O += (tm + sm) * params->O_strides[2] + sn;</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> <span class="keywordflow">if</span> (!<a class="code hl_variable" href="steel__attention_8h.html#a171fdea1b23976453f5dc5e6b3161982">align_Q</a> && <span class="keywordtype">int</span>(tid.x) == (params->NQ_aligned)) {</div>
|
||||
<div class="line"><a id="l00370" name="l00370"></a><span class="lineno"> 370</span> <span class="keyword">auto</span> dst_tile_dims =</div>
|
||||
<div class="line"><a id="l00371" name="l00371"></a><span class="lineno"> 371</span> short2(BD - sn, params->qL - BQ * params->NQ_aligned - (tm + sm));</div>
|
||||
<div class="line"><a id="l00372" name="l00372"></a><span class="lineno"> 372</span> </div>
|
||||
<div class="line"><a id="l00373" name="l00373"></a><span class="lineno"> 373</span> <span class="keywordflow">if</span> (dst_tile_dims.x <= 0 || dst_tile_dims.y <= 0)</div>
|
||||
<div class="line"><a id="l00374" name="l00374"></a><span class="lineno"> 374</span> <span class="keywordflow">return</span>;</div>
|
||||
<div class="line"><a id="l00375" name="l00375"></a><span class="lineno"> 375</span> </div>
|
||||
<div class="line"><a id="l00376" name="l00376"></a><span class="lineno"> 376</span> Otile.template store_safe<T, 1, 1>(O, params->O_strides[2], dst_tile_dims);</div>
|
||||
<div class="line"><a id="l00377" name="l00377"></a><span class="lineno"> 377</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00378" name="l00378"></a><span class="lineno"> 378</span> Otile.template store<T, 1, 1>(O, params->O_strides[2]);</div>
|
||||
<div class="line"><a id="l00379" name="l00379"></a><span class="lineno"> 379</span> }</div>
|
||||
<div class="line"><a id="l00380" name="l00380"></a><span class="lineno"> 380</span>}</div>
|
||||
</div>
|
||||
<div class="ttc" id="anamespacemetal_html_a853c80479ab2264d9c4587c7bcac767b"><div class="ttname"><a href="namespacemetal.html#a853c80479ab2264d9c4587c7bcac767b">metal::max</a></div><div class="ttdeci">METAL_FUNC bfloat16_t max(bfloat16_t x, bfloat16_t y)</div><div class="ttdef"><b>Definition</b> bf16_math.h:232</div></div>
|
||||
<div class="ttc" id="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="anamespacemlx_1_1steel_html_ad583e6038efc119542410f43b603d4ad"><div class="ttname"><a href="namespacemlx_1_1steel.html#ad583e6038efc119542410f43b603d4ad">mlx::steel::tile_matmad</a></div><div class="ttdeci">METAL_FUNC void tile_matmad(thread MMATile< T, M, N > &D, thread MMATile< U, M, K > &A, thread MMATile< U, K, N > &B, thread MMATile< T, M, N > &C)</div><div class="ttdef"><b>Definition</b> mma.h:413</div></div>
|
||||
<div class="ttc" id="anamespacemlx_1_1steel_html_ab9fdcb06fb1f639f9120ab14cfedd150"><div class="ttname"><a href="namespacemlx_1_1steel.html#ab9fdcb06fb1f639f9120ab14cfedd150">mlx::steel::tile_matmad</a></div><div class="ttdeci">METAL_FUNC void tile_matmad(thread MMATile< Dtype, M, N, MMAFragD > &D, thread MMATile< Atype, M, K, MMAFragA > &A, thread MMATile< Btype, K, N, MMAFragB > &B, thread MMATile< Ctype, M, N, MMAFragC > &C)</div><div class="ttdef"><b>Definition</b> mma.h:432</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__attention_8h_html_a171fdea1b23976453f5dc5e6b3161982"><div class="ttname"><a href="steel__attention_8h.html#a171fdea1b23976453f5dc5e6b3161982">align_Q</a></div><div class="ttdeci">constant bool align_Q</div><div class="ttdef"><b>Definition</b> steel_attention.h:9</div></div>
|
||||
<div class="ttc" id="asteel__attention_8h_html_a5423b2a414f5e3c14166d568dedfbd33"><div class="ttname"><a href="steel__attention_8h.html#a5423b2a414f5e3c14166d568dedfbd33">attention</a></div><div class="ttdeci">void attention(const device T *Q, const device T *K, const device T *V, device T *O, const constant AttnParams *params, uint simd_lane_id, uint simd_group_id, uint3 tid, uint3 lid)</div><div class="ttdef"><b>Definition</b> steel_attention.h:73</div></div>
|
||||
@@ -509,9 +540,9 @@ $(function(){initNavTree('steel__attention_8h_source.html',''); initResizable(tr
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_attn_params_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_attn_params.html">mlx::steel::AttnParams</a></div><div class="ttdef"><b>Definition</b> params.h:12</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_base_m_m_a_frag_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_base_m_m_a_frag.html">mlx::steel::BaseMMAFrag</a></div><div class="ttdef"><b>Definition</b> mma.h:37</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_block_loader_t_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_block_loader_t.html">mlx::steel::BlockLoaderT</a></div><div class="ttdef"><b>Definition</b> loader.h:153</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html">mlx::steel::MMATile</a></div><div class="ttdef"><b>Definition</b> mma.h:223</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html_a1a6b1446e8c8da46885bbaa8e8fdc7e4"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">mlx::steel::MMATile::frag_at</a></div><div class="ttdeci">METAL_FUNC constexpr thread frag_type & frag_at(const short i, const short j)</div><div class="ttdef"><b>Definition</b> mma.h:256</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html_aa97a98e423827a889c13a92217626ec7"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">mlx::steel::MMATile::clear</a></div><div class="ttdeci">METAL_FUNC constexpr void clear()</div><div class="ttdef"><b>Definition</b> mma.h:249</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html">mlx::steel::MMATile</a></div><div class="ttdef"><b>Definition</b> mma.h:231</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html_a1a6b1446e8c8da46885bbaa8e8fdc7e4"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html#a1a6b1446e8c8da46885bbaa8e8fdc7e4">mlx::steel::MMATile::frag_at</a></div><div class="ttdeci">METAL_FUNC constexpr thread frag_type & frag_at(const short i, const short j)</div><div class="ttdef"><b>Definition</b> mma.h:264</div></div>
|
||||
<div class="ttc" id="astructmlx_1_1steel_1_1_m_m_a_tile_html_aa97a98e423827a889c13a92217626ec7"><div class="ttname"><a href="structmlx_1_1steel_1_1_m_m_a_tile.html#aa97a98e423827a889c13a92217626ec7">mlx::steel::MMATile::clear</a></div><div class="ttdeci">METAL_FUNC constexpr void clear()</div><div class="ttdef"><b>Definition</b> mma.h:257</div></div>
|
||||
</div><!-- fragment --></div><!-- contents -->
|
||||
</div><!-- doc-content -->
|
||||
<!-- start footer part -->
|
||||
|
||||
Reference in New Issue
Block a user