admin/matmul-analysis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

PPT: MergeBatch b_core=2 说明

Browse Source
This commit is contained in:
admin
2026-08-24 03:24:09 +00:00
parent c04a850b02
commit 73b9cb2b06
1 changed files with 2 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
BMM算子优化分析_Release/BMM算子优化分析_v0.7_讲解PPT.html
View File

@@ -144,8 +144,9 @@ body{font-family:"PingFang SC","Microsoft YaHei",sans-serif;background:#0f172a;c
<li>合并块输出 ≤ L0C（b₀²·MN·4B·2 ≤ 256KB）</li> <li>合并块输出 ≤ L0C（b₀²·MN·4B·2 ≤ 256KB）</li>
<li>单核搬移总量 ≥ 480KB（保证 GM 带宽利用率）</li> <li>单核搬移总量 ≥ 480KB（保证 GM 带宽利用率）</li>
<li>搬移 tile ≥ 16KB（单次搬移效率下限）</li> <li>搬移 tile ≥ 16KB（单次搬移效率下限）</li>
<li>2MN/(M+N) < R₁₆/b₀（访存 Bound，保证浪费被掩盖）</li> <li>2MN/(M+N) &lt; R₁₆/b₀（访存 Bound，保证浪费被掩盖）</li>
</ul> </ul>
<p class="sub">为何不允许 b_core=2？合并后每核仅 1 组，与 IterBatch(b_core=2) 总搬移时延相同，但多 50% 冗余计算——无收益</p>
<p class="sub">典型：B=128, M=N=64, K=512 → MergeBatch；K=256 → 单核搬移不够 480KB → IterBatch</p> <p class="sub">典型：B=128, M=N=64, K=512 → MergeBatch；K=256 → 单核搬移不够 480KB → IterBatch</p>
</div> </div>