This commit is contained in:
CircleCI Docs
2024-11-22 20:24:16 +00:00
parent 6433a22733
commit 8ba8a39e43
905 changed files with 30035 additions and 16934 deletions

View File

@@ -97,9 +97,9 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00004" name="l00004"></a><span class="lineno"> 4</span> </div>
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span><span class="preprocessor">#include &quot;<a class="code" href="kernels_2indexing_8h.html">mlx/backend/metal/kernels/indexing.h</a>&quot;</span></div>
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span> </div>
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> IdxT, <span class="keywordtype">int</span> NIDX, <span class="keywordtype">int</span> IDX_NDIM&gt;</div>
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span><span class="keyword">template</span> &lt;<span class="keyword">typename</span> T, <span class="keyword">typename</span> IdxT, <span class="keywordtype">int</span> NIDX, <span class="keywordtype">int</span> IDX_NDIM, <span class="keyword">typename</span> LocT&gt;</div>
<div class="foldopen" id="foldopen00008" data-start="{" data-end="}">
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"><a class="line" href="gather_8h.html#abdec470e1af0109563ddae3e85e6526c"> 8</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="gather_8h.html#abdec470e1af0109563ddae3e85e6526c">gather_impl</a>(</div>
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"><a class="line" href="gather_8h.html#a767d7c5be6f2f649101f581449af5599"> 8</a></span>METAL_FUNC <span class="keywordtype">void</span> <a class="code hl_function" href="gather_8h.html#a767d7c5be6f2f649101f581449af5599">gather_impl</a>(</div>
<div class="line"><a id="l00009" name="l00009"></a><span class="lineno"> 9</span> <span class="keyword">const</span> device T* src [[buffer(0)]],</div>
<div class="line"><a id="l00010" name="l00010"></a><span class="lineno"> 10</span> device T* out [[buffer(1)]],</div>
<div class="line"><a id="l00011" name="l00011"></a><span class="lineno"> 11</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* src_shape [[buffer(2)]],</div>
@@ -110,44 +110,44 @@ $(function(){ initResizable(false); });
<div class="line"><a id="l00016" name="l00016"></a><span class="lineno"> 16</span> <span class="keyword">const</span> thread <a class="code hl_struct" href="struct_indices.html">Indices&lt;IdxT, NIDX&gt;</a>&amp; indices,</div>
<div class="line"><a id="l00017" name="l00017"></a><span class="lineno"> 17</span> uint3 index [[thread_position_in_grid]],</div>
<div class="line"><a id="l00018" name="l00018"></a><span class="lineno"> 18</span> uint3 grid_dim [[threads_per_grid]]) {</div>
<div class="line"><a id="l00019" name="l00019"></a><span class="lineno"> 19</span> <span class="keywordtype">size_t</span> src_idx = 0;</div>
<div class="line"><a id="l00019" name="l00019"></a><span class="lineno"> 19</span> LocT src_idx = 0;</div>
<div class="line"><a id="l00020" name="l00020"></a><span class="lineno"> 20</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i &lt; NIDX; ++i) {</div>
<div class="line"><a id="l00021" name="l00021"></a><span class="lineno"> 21</span> <span class="keywordtype">size_t</span> idx_loc;</div>
<div class="line"><a id="l00021" name="l00021"></a><span class="lineno"> 21</span> LocT idx_loc;</div>
<div class="line"><a id="l00022" name="l00022"></a><span class="lineno"> 22</span> <span class="keywordflow">if</span> (IDX_NDIM == 0) {</div>
<div class="line"><a id="l00023" name="l00023"></a><span class="lineno"> 23</span> idx_loc = 0;</div>
<div class="line"><a id="l00024" name="l00024"></a><span class="lineno"> 24</span> } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (IDX_NDIM == 1) {</div>
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> idx_loc = index.x * indices.strides[indices.ndim * i];</div>
<div class="line"><a id="l00025" name="l00025"></a><span class="lineno"> 25</span> idx_loc = index.x * <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(indices.strides[indices.ndim * i]);</div>
<div class="line"><a id="l00026" name="l00026"></a><span class="lineno"> 26</span> } <span class="keywordflow">else</span> {</div>
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> idx_loc = index.x * indices.strides[indices.ndim * i];</div>
<div class="line"><a id="l00027" name="l00027"></a><span class="lineno"> 27</span> idx_loc = index.x * <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(indices.strides[indices.ndim * i]);</div>
<div class="line"><a id="l00028" name="l00028"></a><span class="lineno"> 28</span> idx_loc += indices.row_contiguous[i]</div>
<div class="line"><a id="l00029" name="l00029"></a><span class="lineno"> 29</span> ? index.y</div>
<div class="line"><a id="l00030" name="l00030"></a><span class="lineno"> 30</span> : <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a>(</div>
<div class="line"><a id="l00030" name="l00030"></a><span class="lineno"> 30</span> : <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc&lt;size_t, LocT&gt;</a>(</div>
<div class="line"><a id="l00031" name="l00031"></a><span class="lineno"> 31</span> index.y,</div>
<div class="line"><a id="l00032" name="l00032"></a><span class="lineno"> 32</span> &amp;indices.shapes[indices.ndim * i + 1],</div>
<div class="line"><a id="l00033" name="l00033"></a><span class="lineno"> 33</span> &amp;indices.strides[indices.ndim * i + 1],</div>
<div class="line"><a id="l00034" name="l00034"></a><span class="lineno"> 34</span> indices.ndim - 1);</div>
<div class="line"><a id="l00035" name="l00035"></a><span class="lineno"> 35</span> }</div>
<div class="line"><a id="l00036" name="l00036"></a><span class="lineno"> 36</span> <span class="keyword">auto</span> ax = axes[i];</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> <span class="keyword">auto</span> idx_val = <a class="code hl_function" href="kernels_2indexing_8h.html#ab41167dc537c06fbdb4df100972393df">offset_neg_idx</a>(indices.buffers[i][idx_loc], src_shape[ax]);</div>
<div class="line"><a id="l00038" name="l00038"></a><span class="lineno"> 38</span> src_idx += idx_val * src_strides[ax];</div>
<div class="line"><a id="l00037" name="l00037"></a><span class="lineno"> 37</span> <span class="keyword">auto</span> idx_val = <a class="code hl_function" href="kernels_2indexing_8h.html#a58a65ea6215999cd4ccb4fe757cc2dc8">offset_neg_idx</a>(indices.buffers[i][idx_loc], src_shape[ax]);</div>
<div class="line"><a id="l00038" name="l00038"></a><span class="lineno"> 38</span> src_idx += <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(idx_val) * <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(src_strides[ax]);</div>
<div class="line"><a id="l00039" name="l00039"></a><span class="lineno"> 39</span> }</div>
<div class="line"><a id="l00040" name="l00040"></a><span class="lineno"> 40</span> </div>
<div class="line"><a id="l00041" name="l00041"></a><span class="lineno"> 41</span> <span class="keyword">auto</span> src_offset = <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a>(index.z, slice_sizes, src_strides, src_ndim);</div>
<div class="line"><a id="l00042" name="l00042"></a><span class="lineno"> 42</span> </div>
<div class="line"><a id="l00043" name="l00043"></a><span class="lineno"> 43</span> <span class="keywordtype">size_t</span> out_idx = index.z;</div>
<div class="line"><a id="l00044" name="l00044"></a><span class="lineno"> 44</span> <span class="keywordflow">if</span> (IDX_NDIM == 1) {</div>
<div class="line"><a id="l00045" name="l00045"></a><span class="lineno"> 45</span> out_idx += <span class="keyword">static_cast&lt;</span><span class="keywordtype">size_t</span><span class="keyword">&gt;</span>(grid_dim.z) * index.x;</div>
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (IDX_NDIM &gt;= 2) {</div>
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> out_idx +=</div>
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> grid_dim.z * (index.x * <span class="keyword">static_cast&lt;</span><span class="keywordtype">size_t</span><span class="keyword">&gt;</span>(grid_dim.y) + index.y);</div>
<div class="line"><a id="l00041" name="l00041"></a><span class="lineno"> 41</span> <span class="keyword">auto</span> src_offset =</div>
<div class="line"><a id="l00042" name="l00042"></a><span class="lineno"> 42</span> <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a22eaa505dbc7dd2a63a895f2e16712f5">elem_to_loc&lt;size_t, LocT&gt;</a>(index.z, slice_sizes, src_strides, src_ndim);</div>
<div class="line"><a id="l00043" name="l00043"></a><span class="lineno"> 43</span> </div>
<div class="line"><a id="l00044" name="l00044"></a><span class="lineno"> 44</span> LocT out_idx = index.z;</div>
<div class="line"><a id="l00045" name="l00045"></a><span class="lineno"> 45</span> <span class="keywordflow">if</span> (IDX_NDIM == 1) {</div>
<div class="line"><a id="l00046" name="l00046"></a><span class="lineno"> 46</span> out_idx += <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(grid_dim.z) * index.x;</div>
<div class="line"><a id="l00047" name="l00047"></a><span class="lineno"> 47</span> } <span class="keywordflow">else</span> <span class="keywordflow">if</span> (IDX_NDIM &gt;= 2) {</div>
<div class="line"><a id="l00048" name="l00048"></a><span class="lineno"> 48</span> out_idx += grid_dim.z * (index.x * <span class="keyword">static_cast&lt;</span>LocT<span class="keyword">&gt;</span>(grid_dim.y) + index.y);</div>
<div class="line"><a id="l00049" name="l00049"></a><span class="lineno"> 49</span> }</div>
<div class="line"><a id="l00050" name="l00050"></a><span class="lineno"> 50</span> out[out_idx] = src[src_offset + src_idx];</div>
<div class="line"><a id="l00051" name="l00051"></a><span class="lineno"> 51</span>}</div>
</div>
<div class="ttc" id="abackend_2metal_2kernels_2utils_8h_html_a8fd0c8fc6058e650fc99bca8b6acd7d1"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a8fd0c8fc6058e650fc99bca8b6acd7d1">elem_to_loc</a></div><div class="ttdeci">METAL_FUNC stride_t elem_to_loc(uint elem, constant const int *shape, constant const stride_t *strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:87</div></div>
<div class="ttc" id="agather_8h_html_abdec470e1af0109563ddae3e85e6526c"><div class="ttname"><a href="gather_8h.html#abdec470e1af0109563ddae3e85e6526c">gather_impl</a></div><div class="ttdeci">METAL_FUNC void gather_impl(const device T *src, device T *out, const constant int *src_shape, const constant size_t *src_strides, const constant size_t &amp;src_ndim, const constant int *slice_sizes, const constant int *axes, const thread Indices&lt; IdxT, NIDX &gt; &amp;indices, uint3 index, uint3 grid_dim)</div><div class="ttdef"><b>Definition</b> gather.h:8</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="agather_8h_html_a767d7c5be6f2f649101f581449af5599"><div class="ttname"><a href="gather_8h.html#a767d7c5be6f2f649101f581449af5599">gather_impl</a></div><div class="ttdeci">METAL_FUNC void gather_impl(const device T *src, device T *out, const constant int *src_shape, const constant size_t *src_strides, const constant size_t &amp;src_ndim, const constant int *slice_sizes, const constant int *axes, const thread Indices&lt; IdxT, NIDX &gt; &amp;indices, uint3 index, uint3 grid_dim)</div><div class="ttdef"><b>Definition</b> gather.h:8</div></div>
<div class="ttc" id="akernels_2indexing_8h_html"><div class="ttname"><a href="kernels_2indexing_8h.html">indexing.h</a></div></div>
<div class="ttc" id="akernels_2indexing_8h_html_ab41167dc537c06fbdb4df100972393df"><div class="ttname"><a href="kernels_2indexing_8h.html#ab41167dc537c06fbdb4df100972393df">offset_neg_idx</a></div><div class="ttdeci">METAL_FUNC size_t offset_neg_idx(IdxT idx, size_t size)</div><div class="ttdef"><b>Definition</b> indexing.h:17</div></div>
<div class="ttc" id="akernels_2indexing_8h_html_a58a65ea6215999cd4ccb4fe757cc2dc8"><div class="ttname"><a href="kernels_2indexing_8h.html#a58a65ea6215999cd4ccb4fe757cc2dc8">offset_neg_idx</a></div><div class="ttdeci">METAL_FUNC size_t offset_neg_idx(IdxT idx, int size)</div><div class="ttdef"><b>Definition</b> indexing.h:17</div></div>
<div class="ttc" id="astruct_indices_html"><div class="ttname"><a href="struct_indices.html">Indices</a></div><div class="ttdef"><b>Definition</b> indexing.h:8</div></div>
</div><!-- fragment --></div><!-- contents -->
<!-- start footer part -->