PPT修正 StreamK 条件为 P<C/2
This commit is contained in:
@@ -169,7 +169,8 @@ body{font-family:"PingFang SC","Microsoft YaHei",sans-serif;background:#0f172a;c
|
||||
<div class="slide" id="s10">
|
||||
<h2>9. StreamK — K 维核间切分</h2>
|
||||
<p><b>思路</b>:B/M/N 都填不满核时,把 K 切给多个核各算一段,再归约</p>
|
||||
<p style="margin-top:12px"><b>进入条件</b>:P = B·MN·4B/L0C < 32;K/grid_K ≥ 128 元素(dValue 256B)</p>
|
||||
<p style="margin-top:12px"><b>进入条件</b>:P = B·MN·4B/L0C < 16(=C/2);K/grid_K ≥ 128 元素(dValue 256B)</p>
|
||||
<p class="sub">阈值取 C/2:grid_K≥2 时每块需 2 核,P×2 ≤ C 才放得下</p>
|
||||
<p><b>归约代价</b>(按实现流程推导):</p>
|
||||
<div class="code">第1步:grid_K 个 AIC 各算 K/grid_K 段,L0C 原地累加
|
||||
第2步:fixpipe 写部分和到 workspace(驻留 L2,走 5.2TB/s 写口)
|
||||
|
||||
Reference in New Issue
Block a user