mirror of
https://github.com/ml-explore/mlx.git
synced 2025-09-18 18:28:12 +08:00
docs up
This commit is contained in:
Awni Hannun
committed by
CircleCI Docs
CircleCI Docs
parent
27e07888b0
commit
6a6ffb598d
270
docs/build/html/reduce__col_8h_source.html
vendored
270
docs/build/html/reduce__col_8h_source.html
vendored
@@ -89,7 +89,7 @@ $(function() { codefold.init(0); });
|
||||
<div class="line"><a id="l00004" name="l00004"></a><span class="lineno"> 4</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00005" name="l00005"></a><span class="lineno"> 5</span> <span class="keyword">typename</span> U,</div>
|
||||
<div class="line"><a id="l00006" name="l00006"></a><span class="lineno"> 6</span> <span class="keyword">typename</span> Op,</div>
|
||||
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span> <span class="keywordtype">int</span> NDIMS = 0,</div>
|
||||
<div class="line"><a id="l00007" name="l00007"></a><span class="lineno"> 7</span> <span class="keywordtype">int</span> NDIMS,</div>
|
||||
<div class="line"><a id="l00008" name="l00008"></a><span class="lineno"> 8</span> <span class="keywordtype">int</span> N_READS = <a class="code hl_variable" href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a>></div>
|
||||
<div class="foldopen" id="foldopen00009" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00009" name="l00009"></a><span class="lineno"><a class="line" href="reduce__col_8h.html#adf7aeb18cd1d5042cf6d9b46b582d8ce"> 9</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__col_8h.html#adf7aeb18cd1d5042cf6d9b46b582d8ce">col_reduce_small</a>(</div>
|
||||
@@ -274,150 +274,144 @@ $(function() { codefold.init(0); });
|
||||
<div class="line"><a id="l00188" name="l00188"></a><span class="lineno"> 188</span>}</div>
|
||||
</div>
|
||||
<div class="line"><a id="l00189" name="l00189"></a><span class="lineno"> 189</span> </div>
|
||||
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span><span class="keyword">template</span> <</div>
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"> 202</span> <span class="keyword">typename</span> T,</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> <span class="keyword">typename</span> U,</div>
|
||||
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> <span class="keyword">typename</span> Op,</div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="keywordtype">int</span> NDIMS = 0,</div>
|
||||
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> <span class="keywordtype">int</span> BM = 8,</div>
|
||||
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> <span class="keywordtype">int</span> BN = 128></div>
|
||||
<div class="foldopen" id="foldopen00208" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"><a class="line" href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385"> 208</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385">col_reduce_looped</a>(</div>
|
||||
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> <span class="keyword">const</span> device T* in [[buffer(0)]],</div>
|
||||
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> device U* out [[buffer(1)]],</div>
|
||||
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& reduction_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& reduction_stride [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* shape [[buffer(4)]],</div>
|
||||
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& ndim [[buffer(6)]],</div>
|
||||
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* reduce_shape [[buffer(7)]],</div>
|
||||
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& reduce_ndim [[buffer(9)]],</div>
|
||||
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"> 219</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& non_col_reductions [[buffer(10)]],</div>
|
||||
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> uint3 gid [[threadgroup_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> uint3 gsize [[threadgroups_per_grid]],</div>
|
||||
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
|
||||
<div class="line"><a id="l00223" name="l00223"></a><span class="lineno"> 223</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
|
||||
<div class="line"><a id="l00224" name="l00224"></a><span class="lineno"> 224</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> n_simdgroups = 4;</div>
|
||||
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> tgp_size = n_simdgroups * <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a>;</div>
|
||||
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> n_reads = (BM * BN) / tgp_size;</div>
|
||||
<div class="line"><a id="l00228" name="l00228"></a><span class="lineno"> 228</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> n_read_blocks = BN / n_reads;</div>
|
||||
<div class="line"><a id="l00229" name="l00229"></a><span class="lineno"> 229</span> </div>
|
||||
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> threadgroup U shared_vals[BN * BM];</div>
|
||||
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> U totals[n_reads];</div>
|
||||
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> <a class="code hl_struct" href="structlooped__elem__to__loc.html">looped_elem_to_loc<NDIMS></a> loop;</div>
|
||||
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> <span class="keyword">const</span> device T* row;</div>
|
||||
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> </div>
|
||||
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> totals[i] = Op::init;</div>
|
||||
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> }</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> <span class="keywordtype">short</span> lid = simd_group_id * <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a> + simd_lane_id;</div>
|
||||
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> short2 offset((lid % n_read_blocks) * n_reads, lid / n_read_blocks);</div>
|
||||
<div class="line"><a id="l00241" name="l00241"></a><span class="lineno"> 241</span> <span class="keywordtype">size_t</span> column = BN * gid.x + offset.x;</div>
|
||||
<div class="line"><a id="l00242" name="l00242"></a><span class="lineno"> 242</span> <span class="keywordtype">bool</span> safe = column + n_reads <= reduction_stride;</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> <span class="keywordtype">size_t</span> out_idx = gid.y + gsize.y * size_t(gid.z);</div>
|
||||
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> <span class="keywordtype">size_t</span> in_idx = <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a2e49fa7ab8f6348543455c6c45d7e2a9">elem_to_loc</a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> in += in_idx + column;</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> <span class="keywordtype">size_t</span> total = non_col_reductions * reduction_size;</div>
|
||||
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#a05558dabba889ee0d80ed4b567d901ca">next</a>(offset.y, reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> <span class="keywordflow">for</span> (<span class="keywordtype">size_t</span> r = offset.y; r < total; r += BM) {</div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> row = in + loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#accc6d4957a8aeb38f5062754793b74d2">location</a>(r, reduce_shape, reduce_strides, reduce_ndim);</div>
|
||||
<div class="line"><a id="l00252" name="l00252"></a><span class="lineno"> 252</span> </div>
|
||||
<div class="line"><a id="l00253" name="l00253"></a><span class="lineno"> 253</span> <span class="keywordflow">if</span> (safe) {</div>
|
||||
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(row[i]), totals[i]);</div>
|
||||
<div class="line"><a id="l00201" name="l00201"></a><span class="lineno"> 201</span><span class="keyword">template</span> <<span class="keyword">typename</span> T, <span class="keyword">typename</span> U, <span class="keyword">typename</span> Op, <span class="keywordtype">int</span> NDIMS, <span class="keywordtype">int</span> BM, <span class="keywordtype">int</span> BN></div>
|
||||
<div class="foldopen" id="foldopen00202" data-start="{" data-end="}">
|
||||
<div class="line"><a id="l00202" name="l00202"></a><span class="lineno"><a class="line" href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385"> 202</a></span>[[kernel]] <span class="keywordtype">void</span> <a class="code hl_function" href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385">col_reduce_looped</a>(</div>
|
||||
<div class="line"><a id="l00203" name="l00203"></a><span class="lineno"> 203</span> <span class="keyword">const</span> device T* in [[buffer(0)]],</div>
|
||||
<div class="line"><a id="l00204" name="l00204"></a><span class="lineno"> 204</span> device U* out [[buffer(1)]],</div>
|
||||
<div class="line"><a id="l00205" name="l00205"></a><span class="lineno"> 205</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& reduction_size [[buffer(2)]],</div>
|
||||
<div class="line"><a id="l00206" name="l00206"></a><span class="lineno"> 206</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& reduction_stride [[buffer(3)]],</div>
|
||||
<div class="line"><a id="l00207" name="l00207"></a><span class="lineno"> 207</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* shape [[buffer(4)]],</div>
|
||||
<div class="line"><a id="l00208" name="l00208"></a><span class="lineno"> 208</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* strides [[buffer(5)]],</div>
|
||||
<div class="line"><a id="l00209" name="l00209"></a><span class="lineno"> 209</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& ndim [[buffer(6)]],</div>
|
||||
<div class="line"><a id="l00210" name="l00210"></a><span class="lineno"> 210</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>* reduce_shape [[buffer(7)]],</div>
|
||||
<div class="line"><a id="l00211" name="l00211"></a><span class="lineno"> 211</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>* reduce_strides [[buffer(8)]],</div>
|
||||
<div class="line"><a id="l00212" name="l00212"></a><span class="lineno"> 212</span> <span class="keyword">const</span> constant <span class="keywordtype">int</span>& reduce_ndim [[buffer(9)]],</div>
|
||||
<div class="line"><a id="l00213" name="l00213"></a><span class="lineno"> 213</span> <span class="keyword">const</span> constant <span class="keywordtype">size_t</span>& non_col_reductions [[buffer(10)]],</div>
|
||||
<div class="line"><a id="l00214" name="l00214"></a><span class="lineno"> 214</span> uint3 gid [[threadgroup_position_in_grid]],</div>
|
||||
<div class="line"><a id="l00215" name="l00215"></a><span class="lineno"> 215</span> uint3 gsize [[threadgroups_per_grid]],</div>
|
||||
<div class="line"><a id="l00216" name="l00216"></a><span class="lineno"> 216</span> uint simd_lane_id [[thread_index_in_simdgroup]],</div>
|
||||
<div class="line"><a id="l00217" name="l00217"></a><span class="lineno"> 217</span> uint simd_group_id [[simdgroup_index_in_threadgroup]]) {</div>
|
||||
<div class="line"><a id="l00218" name="l00218"></a><span class="lineno"> 218</span> Op <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>;</div>
|
||||
<div class="line"><a id="l00219" name="l00219"></a><span class="lineno"> 219</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> n_simdgroups = 4;</div>
|
||||
<div class="line"><a id="l00220" name="l00220"></a><span class="lineno"> 220</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> tgp_size = n_simdgroups * <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a>;</div>
|
||||
<div class="line"><a id="l00221" name="l00221"></a><span class="lineno"> 221</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> n_reads = (BM * BN) / tgp_size;</div>
|
||||
<div class="line"><a id="l00222" name="l00222"></a><span class="lineno"> 222</span> <span class="keyword">constexpr</span> <span class="keywordtype">short</span> n_read_blocks = BN / n_reads;</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> threadgroup U shared_vals[BN * BM];</div>
|
||||
<div class="line"><a id="l00225" name="l00225"></a><span class="lineno"> 225</span> U totals[n_reads];</div>
|
||||
<div class="line"><a id="l00226" name="l00226"></a><span class="lineno"> 226</span> <a class="code hl_struct" href="structlooped__elem__to__loc.html">looped_elem_to_loc<NDIMS></a> loop;</div>
|
||||
<div class="line"><a id="l00227" name="l00227"></a><span class="lineno"> 227</span> <span class="keyword">const</span> device T* row;</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="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00230" name="l00230"></a><span class="lineno"> 230</span> totals[i] = Op::init;</div>
|
||||
<div class="line"><a id="l00231" name="l00231"></a><span class="lineno"> 231</span> }</div>
|
||||
<div class="line"><a id="l00232" name="l00232"></a><span class="lineno"> 232</span> </div>
|
||||
<div class="line"><a id="l00233" name="l00233"></a><span class="lineno"> 233</span> <span class="keywordtype">short</span> lid = simd_group_id * <a class="code hl_variable" href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a> + simd_lane_id;</div>
|
||||
<div class="line"><a id="l00234" name="l00234"></a><span class="lineno"> 234</span> short2 offset((lid % n_read_blocks) * n_reads, lid / n_read_blocks);</div>
|
||||
<div class="line"><a id="l00235" name="l00235"></a><span class="lineno"> 235</span> <span class="keywordtype">size_t</span> column = BN * gid.x + offset.x;</div>
|
||||
<div class="line"><a id="l00236" name="l00236"></a><span class="lineno"> 236</span> <span class="keywordtype">bool</span> safe = column + n_reads <= reduction_stride;</div>
|
||||
<div class="line"><a id="l00237" name="l00237"></a><span class="lineno"> 237</span> </div>
|
||||
<div class="line"><a id="l00238" name="l00238"></a><span class="lineno"> 238</span> <span class="keywordtype">size_t</span> out_idx = gid.y + gsize.y * size_t(gid.z);</div>
|
||||
<div class="line"><a id="l00239" name="l00239"></a><span class="lineno"> 239</span> <span class="keywordtype">size_t</span> in_idx = <a class="code hl_function" href="backend_2metal_2kernels_2utils_8h.html#a2e49fa7ab8f6348543455c6c45d7e2a9">elem_to_loc</a>(out_idx, shape, strides, ndim);</div>
|
||||
<div class="line"><a id="l00240" name="l00240"></a><span class="lineno"> 240</span> in += in_idx + column;</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> <span class="keywordtype">size_t</span> total = non_col_reductions * reduction_size;</div>
|
||||
<div class="line"><a id="l00243" name="l00243"></a><span class="lineno"> 243</span> loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#a05558dabba889ee0d80ed4b567d901ca">next</a>(offset.y, reduce_shape, reduce_strides);</div>
|
||||
<div class="line"><a id="l00244" name="l00244"></a><span class="lineno"> 244</span> <span class="keywordflow">for</span> (<span class="keywordtype">size_t</span> r = offset.y; r < total; r += BM) {</div>
|
||||
<div class="line"><a id="l00245" name="l00245"></a><span class="lineno"> 245</span> row = in + loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#accc6d4957a8aeb38f5062754793b74d2">location</a>(r, reduce_shape, reduce_strides, reduce_ndim);</div>
|
||||
<div class="line"><a id="l00246" name="l00246"></a><span class="lineno"> 246</span> </div>
|
||||
<div class="line"><a id="l00247" name="l00247"></a><span class="lineno"> 247</span> <span class="keywordflow">if</span> (safe) {</div>
|
||||
<div class="line"><a id="l00248" name="l00248"></a><span class="lineno"> 248</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00249" name="l00249"></a><span class="lineno"> 249</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(<span class="keyword">static_cast<</span>U<span class="keyword">></span>(row[i]), totals[i]);</div>
|
||||
<div class="line"><a id="l00250" name="l00250"></a><span class="lineno"> 250</span> }</div>
|
||||
<div class="line"><a id="l00251" name="l00251"></a><span class="lineno"> 251</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00252" name="l00252"></a><span class="lineno"> 252</span> U vals[n_reads];</div>
|
||||
<div class="line"><a id="l00253" name="l00253"></a><span class="lineno"> 253</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00254" name="l00254"></a><span class="lineno"> 254</span> vals[i] =</div>
|
||||
<div class="line"><a id="l00255" name="l00255"></a><span class="lineno"> 255</span> (column + i < reduction_stride) ? static_cast<U>(row[i]) : <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.init;</div>
|
||||
<div class="line"><a id="l00256" name="l00256"></a><span class="lineno"> 256</span> }</div>
|
||||
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> U vals[n_reads];</div>
|
||||
<div class="line"><a id="l00259" name="l00259"></a><span class="lineno"> 259</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00260" name="l00260"></a><span class="lineno"> 260</span> vals[i] =</div>
|
||||
<div class="line"><a id="l00261" name="l00261"></a><span class="lineno"> 261</span> (column + i < reduction_stride) ? static_cast<U>(row[i]) : <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.init;</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> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00264" name="l00264"></a><span class="lineno"> 264</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(vals[i], totals[i]);</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> </div>
|
||||
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#a05558dabba889ee0d80ed4b567d901ca">next</a>(BM, reduce_shape, reduce_strides);</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> <span class="comment">// We can use a simd reduction to accumulate across BM so each thread writes</span></div>
|
||||
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> <span class="comment">// the partial output to SM and then each simdgroup does BN / n_simdgroups</span></div>
|
||||
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> <span class="comment">// accumulations.</span></div>
|
||||
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"> 274</span> <span class="keywordflow">if</span> (BM == 32) {</div>
|
||||
<div class="line"><a id="l00275" name="l00275"></a><span class="lineno"> 275</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> n_outputs = BN / n_simdgroups;</div>
|
||||
<div class="line"><a id="l00276" name="l00276"></a><span class="lineno"> 276</span> <span class="keyword">static_assert</span>(</div>
|
||||
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"> 277</span> BM != 32 || n_outputs == n_reads,</div>
|
||||
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> <span class="stringliteral">"The tile should be selected such that n_outputs == n_reads"</span>);</div>
|
||||
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"> 279</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"> 280</span> shared_vals[offset.y * BN + offset.x + i] = totals[i];</div>
|
||||
<div class="line"><a id="l00257" name="l00257"></a><span class="lineno"> 257</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00258" name="l00258"></a><span class="lineno"> 258</span> totals[i] = <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(vals[i], totals[i]);</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> }</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> loop.<a class="code hl_function" href="structlooped__elem__to__loc.html#a05558dabba889ee0d80ed4b567d901ca">next</a>(BM, reduce_shape, reduce_strides);</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> <span class="comment">// We can use a simd reduction to accumulate across BM so each thread writes</span></div>
|
||||
<div class="line"><a id="l00266" name="l00266"></a><span class="lineno"> 266</span> <span class="comment">// the partial output to SM and then each simdgroup does BN / n_simdgroups</span></div>
|
||||
<div class="line"><a id="l00267" name="l00267"></a><span class="lineno"> 267</span> <span class="comment">// accumulations.</span></div>
|
||||
<div class="line"><a id="l00268" name="l00268"></a><span class="lineno"> 268</span> <span class="keywordflow">if</span> (BM == 32) {</div>
|
||||
<div class="line"><a id="l00269" name="l00269"></a><span class="lineno"> 269</span> <span class="keyword">constexpr</span> <span class="keywordtype">int</span> n_outputs = BN / n_simdgroups;</div>
|
||||
<div class="line"><a id="l00270" name="l00270"></a><span class="lineno"> 270</span> <span class="keyword">static_assert</span>(</div>
|
||||
<div class="line"><a id="l00271" name="l00271"></a><span class="lineno"> 271</span> BM != 32 || n_outputs == n_reads,</div>
|
||||
<div class="line"><a id="l00272" name="l00272"></a><span class="lineno"> 272</span> <span class="stringliteral">"The tile should be selected such that n_outputs == n_reads"</span>);</div>
|
||||
<div class="line"><a id="l00273" name="l00273"></a><span class="lineno"> 273</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00274" name="l00274"></a><span class="lineno"> 274</span> shared_vals[offset.y * BN + offset.x + i] = totals[i];</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> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00277" name="l00277"></a><span class="lineno"> 277</span> short2 out_offset(simd_group_id * n_outputs, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00278" name="l00278"></a><span class="lineno"> 278</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_outputs; i++) {</div>
|
||||
<div class="line"><a id="l00279" name="l00279"></a><span class="lineno"> 279</span> totals[i] =</div>
|
||||
<div class="line"><a id="l00280" name="l00280"></a><span class="lineno"> 280</span> <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.simd_reduce(shared_vals[out_offset.y * BN + out_offset.x + i]);</div>
|
||||
<div class="line"><a id="l00281" name="l00281"></a><span class="lineno"> 281</span> }</div>
|
||||
<div class="line"><a id="l00282" name="l00282"></a><span class="lineno"> 282</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> short2 out_offset(simd_group_id * n_outputs, simd_lane_id);</div>
|
||||
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"> 284</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_outputs; i++) {</div>
|
||||
<div class="line"><a id="l00285" name="l00285"></a><span class="lineno"> 285</span> totals[i] =</div>
|
||||
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"> 286</span> <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>.simd_reduce(shared_vals[out_offset.y * BN + out_offset.x + i]);</div>
|
||||
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> }</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">// Write the output.</span></div>
|
||||
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"> 290</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
|
||||
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"> 291</span> <span class="keywordtype">size_t</span> out_column = BN * gid.x + out_offset.x;</div>
|
||||
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> out += out_idx * reduction_stride + out_column;</div>
|
||||
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"> 293</span> <span class="keywordflow">if</span> (out_column + n_outputs <= reduction_stride) {</div>
|
||||
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_outputs; i++) {</div>
|
||||
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> out[i] = totals[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> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00298" name="l00298"></a><span class="lineno"> 298</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; out_column + i < reduction_stride; i++) {</div>
|
||||
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> out[i] = totals[i];</div>
|
||||
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> }</div>
|
||||
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> }</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> </div>
|
||||
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> <span class="comment">// Each thread holds n_reads partial results. We write them all out to shared</span></div>
|
||||
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> <span class="comment">// memory and threads with offset.y == 0 aggregate the columns and write the</span></div>
|
||||
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> <span class="comment">// outputs.</span></div>
|
||||
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> <span class="keywordtype">short</span> x_block = offset.x / n_reads;</div>
|
||||
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> shared_vals[x_block * BM * n_reads + i * BM + offset.y] = totals[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> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> <span class="keywordflow">if</span> (offset.y == 0) {</div>
|
||||
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00316" name="l00316"></a><span class="lineno"> 316</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 1; j < BM; j++) {</div>
|
||||
<div class="line"><a id="l00317" name="l00317"></a><span class="lineno"> 317</span> totals[i] =</div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(shared_vals[x_block * BM * n_reads + i * BM + j], totals[i]);</div>
|
||||
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> }</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> }</div>
|
||||
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> }</div>
|
||||
<div 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">// Write the output.</span></div>
|
||||
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> <span class="keywordflow">if</span> (offset.y == 0) {</div>
|
||||
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> out += out_idx * reduction_stride + column;</div>
|
||||
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> <span class="keywordflow">if</span> (safe) {</div>
|
||||
<div class="line"><a id="l00327" name="l00327"></a><span class="lineno"> 327</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00328" name="l00328"></a><span class="lineno"> 328</span> out[i] = totals[i];</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> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00331" name="l00331"></a><span class="lineno"> 331</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; column + i < reduction_stride; i++) {</div>
|
||||
<div class="line"><a id="l00332" name="l00332"></a><span class="lineno"> 332</span> out[i] = totals[i];</div>
|
||||
<div class="line"><a id="l00333" name="l00333"></a><span class="lineno"> 333</span> }</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> }</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="l00282" name="l00282"></a><span class="lineno"> 282</span> </div>
|
||||
<div class="line"><a id="l00283" name="l00283"></a><span class="lineno"> 283</span> <span class="comment">// Write the output.</span></div>
|
||||
<div class="line"><a id="l00284" name="l00284"></a><span class="lineno"> 284</span> <span class="keywordflow">if</span> (simd_lane_id == 0) {</div>
|
||||
<div class="line"><a id="l00285" name="l00285"></a><span class="lineno"> 285</span> <span class="keywordtype">size_t</span> out_column = BN * gid.x + out_offset.x;</div>
|
||||
<div class="line"><a id="l00286" name="l00286"></a><span class="lineno"> 286</span> out += out_idx * reduction_stride + out_column;</div>
|
||||
<div class="line"><a id="l00287" name="l00287"></a><span class="lineno"> 287</span> <span class="keywordflow">if</span> (out_column + n_outputs <= reduction_stride) {</div>
|
||||
<div class="line"><a id="l00288" name="l00288"></a><span class="lineno"> 288</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_outputs; i++) {</div>
|
||||
<div class="line"><a id="l00289" name="l00289"></a><span class="lineno"> 289</span> out[i] = totals[i];</div>
|
||||
<div class="line"><a id="l00290" name="l00290"></a><span class="lineno"> 290</span> }</div>
|
||||
<div class="line"><a id="l00291" name="l00291"></a><span class="lineno"> 291</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00292" name="l00292"></a><span class="lineno"> 292</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; out_column + i < reduction_stride; i++) {</div>
|
||||
<div class="line"><a id="l00293" name="l00293"></a><span class="lineno"> 293</span> out[i] = totals[i];</div>
|
||||
<div class="line"><a id="l00294" name="l00294"></a><span class="lineno"> 294</span> }</div>
|
||||
<div class="line"><a id="l00295" name="l00295"></a><span class="lineno"> 295</span> }</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> </div>
|
||||
<div class="line"><a id="l00299" name="l00299"></a><span class="lineno"> 299</span> <span class="comment">// Each thread holds n_reads partial results. We write them all out to shared</span></div>
|
||||
<div class="line"><a id="l00300" name="l00300"></a><span class="lineno"> 300</span> <span class="comment">// memory and threads with offset.y == 0 aggregate the columns and write the</span></div>
|
||||
<div class="line"><a id="l00301" name="l00301"></a><span class="lineno"> 301</span> <span class="comment">// outputs.</span></div>
|
||||
<div class="line"><a id="l00302" name="l00302"></a><span class="lineno"> 302</span> <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00303" name="l00303"></a><span class="lineno"> 303</span> <span class="keywordtype">short</span> x_block = offset.x / n_reads;</div>
|
||||
<div class="line"><a id="l00304" name="l00304"></a><span class="lineno"> 304</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00305" name="l00305"></a><span class="lineno"> 305</span> shared_vals[x_block * BM * n_reads + i * BM + offset.y] = totals[i];</div>
|
||||
<div class="line"><a id="l00306" name="l00306"></a><span class="lineno"> 306</span> }</div>
|
||||
<div class="line"><a id="l00307" name="l00307"></a><span class="lineno"> 307</span> threadgroup_barrier(mem_flags::mem_threadgroup);</div>
|
||||
<div class="line"><a id="l00308" name="l00308"></a><span class="lineno"> 308</span> <span class="keywordflow">if</span> (offset.y == 0) {</div>
|
||||
<div class="line"><a id="l00309" name="l00309"></a><span class="lineno"> 309</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00310" name="l00310"></a><span class="lineno"> 310</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> j = 1; j < BM; j++) {</div>
|
||||
<div class="line"><a id="l00311" name="l00311"></a><span class="lineno"> 311</span> totals[i] =</div>
|
||||
<div class="line"><a id="l00312" name="l00312"></a><span class="lineno"> 312</span> <a class="code hl_variable" href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a>(shared_vals[x_block * BM * n_reads + i * BM + j], totals[i]);</div>
|
||||
<div class="line"><a id="l00313" name="l00313"></a><span class="lineno"> 313</span> }</div>
|
||||
<div class="line"><a id="l00314" name="l00314"></a><span class="lineno"> 314</span> }</div>
|
||||
<div class="line"><a id="l00315" name="l00315"></a><span class="lineno"> 315</span> }</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">// Write the output.</span></div>
|
||||
<div class="line"><a id="l00318" name="l00318"></a><span class="lineno"> 318</span> <span class="keywordflow">if</span> (offset.y == 0) {</div>
|
||||
<div class="line"><a id="l00319" name="l00319"></a><span class="lineno"> 319</span> out += out_idx * reduction_stride + column;</div>
|
||||
<div class="line"><a id="l00320" name="l00320"></a><span class="lineno"> 320</span> <span class="keywordflow">if</span> (safe) {</div>
|
||||
<div class="line"><a id="l00321" name="l00321"></a><span class="lineno"> 321</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; i < n_reads; i++) {</div>
|
||||
<div class="line"><a id="l00322" name="l00322"></a><span class="lineno"> 322</span> out[i] = totals[i];</div>
|
||||
<div class="line"><a id="l00323" name="l00323"></a><span class="lineno"> 323</span> }</div>
|
||||
<div class="line"><a id="l00324" name="l00324"></a><span class="lineno"> 324</span> } <span class="keywordflow">else</span> {</div>
|
||||
<div class="line"><a id="l00325" name="l00325"></a><span class="lineno"> 325</span> <span class="keywordflow">for</span> (<span class="keywordtype">int</span> i = 0; column + i < reduction_stride; i++) {</div>
|
||||
<div class="line"><a id="l00326" name="l00326"></a><span class="lineno"> 326</span> out[i] = totals[i];</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> }</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>}</div>
|
||||
</div>
|
||||
<div class="ttc" id="abackend_2metal_2kernels_2reduction_2ops_8h_html_a515b75d563a93d3c09ee677948dc83e3"><div class="ttname"><a href="backend_2metal_2kernels_2reduction_2ops_8h.html#a515b75d563a93d3c09ee677948dc83e3">simd_size</a></div><div class="ttdeci">static constant constexpr const uint8_t simd_size</div><div class="ttdef"><b>Definition</b> ops.h:22</div></div>
|
||||
<div class="ttc" id="abackend_2metal_2kernels_2utils_8h_html_a2e49fa7ab8f6348543455c6c45d7e2a9"><div class="ttname"><a href="backend_2metal_2kernels_2utils_8h.html#a2e49fa7ab8f6348543455c6c45d7e2a9">elem_to_loc</a></div><div class="ttdeci">METAL_FUNC stride_t elem_to_loc(uint elem, device const int *shape, device const stride_t *strides, int ndim)</div><div class="ttdef"><b>Definition</b> utils.h:87</div></div>
|
||||
<div class="ttc" id="acommon_2binary_8h_html_a70228731d29946574b238d21fb4b360c"><div class="ttname"><a href="common_2binary_8h.html#a70228731d29946574b238d21fb4b360c">op</a></div><div class="ttdeci">Op op</div><div class="ttdef"><b>Definition</b> binary.h:141</div></div>
|
||||
<div class="ttc" id="adefines_8h_html_a2ad505864a2ab786147766900bc18c21"><div class="ttname"><a href="defines_8h.html#a2ad505864a2ab786147766900bc18c21">REDUCE_N_READS</a></div><div class="ttdeci">static constexpr int REDUCE_N_READS</div><div class="ttdef"><b>Definition</b> defines.h:12</div></div>
|
||||
<div class="ttc" id="areduce__col_8h_html_a11bfc6112ae2386ac03f5ea7b7d93385"><div class="ttname"><a href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385">col_reduce_looped</a></div><div class="ttdeci">void col_reduce_looped(const device T *in, device U *out, const constant size_t &reduction_size, const constant size_t &reduction_stride, const constant int *shape, const constant size_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant size_t *reduce_strides, const constant int &reduce_ndim, const constant size_t &non_col_reductions, uint3 gid, uint3 gsize, uint simd_lane_id, uint simd_group_id)</div><div class="ttdoc">Our approach is the following simple looped approach:</div><div class="ttdef"><b>Definition</b> reduce_col.h:208</div></div>
|
||||
<div class="ttc" id="areduce__col_8h_html_a11bfc6112ae2386ac03f5ea7b7d93385"><div class="ttname"><a href="reduce__col_8h.html#a11bfc6112ae2386ac03f5ea7b7d93385">col_reduce_looped</a></div><div class="ttdeci">void col_reduce_looped(const device T *in, device U *out, const constant size_t &reduction_size, const constant size_t &reduction_stride, const constant int *shape, const constant size_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant size_t *reduce_strides, const constant int &reduce_ndim, const constant size_t &non_col_reductions, uint3 gid, uint3 gsize, uint simd_lane_id, uint simd_group_id)</div><div class="ttdoc">Our approach is the following simple looped approach:</div><div class="ttdef"><b>Definition</b> reduce_col.h:202</div></div>
|
||||
<div class="ttc" id="areduce__col_8h_html_adf7aeb18cd1d5042cf6d9b46b582d8ce"><div class="ttname"><a href="reduce__col_8h.html#adf7aeb18cd1d5042cf6d9b46b582d8ce">col_reduce_small</a></div><div class="ttdeci">void col_reduce_small(const device T *in, device U *out, const constant size_t &reduction_size, const constant size_t &reduction_stride, const constant int *shape, const constant size_t *strides, const constant int &ndim, const constant int *reduce_shape, const constant size_t *reduce_strides, const constant int &reduce_ndim, const constant size_t &non_col_reductions, uint3 gid, uint3 gsize, uint simd_lane_id, uint simd_group_id, uint3 tid, uint3 tsize)</div><div class="ttdef"><b>Definition</b> reduce_col.h:9</div></div>
|
||||
<div class="ttc" id="astructlooped__elem__to__loc_html"><div class="ttname"><a href="structlooped__elem__to__loc.html">looped_elem_to_loc</a></div><div class="ttdef"><b>Definition</b> utils.h:334</div></div>
|
||||
<div class="ttc" id="astructlooped__elem__to__loc_html_a05558dabba889ee0d80ed4b567d901ca"><div class="ttname"><a href="structlooped__elem__to__loc.html#a05558dabba889ee0d80ed4b567d901ca">looped_elem_to_loc::next</a></div><div class="ttdeci">void next(const constant int *shape, const constant size_t *strides)</div><div class="ttdef"><b>Definition</b> utils.h:339</div></div>
|
||||
|
||||
Reference in New Issue
Block a user