Fix T_cmd=0 robustness: beats_iterbatch 阈值除零防护 (T_cmd<=0 判 MergeBatch 不胜出, 无命令节省) + 3 条回归测试

BMM/BMM_Theory/bmm_theory/branches/merge_batch.py

@@ -7,6 +7,8 @@
 核心思想: 合并 b0 个 batch 的 A'[b0*M,K] @ B'[K,b0*N] 为单次 DMA 搬入,
 减少 GM->L1 搬移命令数 (省 b0 倍 T_cmd); 交叉项被算出但丢弃 (冗余比例 (b0-1)/b0),
 进入条件 5 保证 case 为访存 Bound, 冗余算力被搬移时延掩盖.
+T_cmd=0 时 (无命令固定开销) 合并失去唯一收益来源, 分界仲裁直接判 MergeBatch
+不胜出 (beats_iterbatch 对 T_cmd<=0 做 +inf 处理, 见函数内说明).
 """
 
 from __future__ import annotations
@@ -222,6 +224,8 @@ class MergeBatchBranch(Branch):
 
         MergeBatch 最优  ⟺  K截断 (k_L1=K) 且 b_core > b0*(T_comp+T_write)/T_cmd
         L1 绑定情形 MergeBatch 恒劣于 IterBatch (搬移次数相同, 只放大 drain).
+        T_cmd=0 时阈值趋于无穷: 合并的唯一收益 (省 DMA 命令开销) 消失,
+        只剩 drain 放大与冗余计算, MergeBatch 无胜出通道 (见 evaluate 对拍也一致).
         """
         s = self.spec
         m, n, k = case.m, case.n, case.k
@@ -237,12 +241,21 @@ class MergeBatchBranch(Branch):
         b0 = plan.merge_b0
         t_comp = 2.0 * m * n * min(k_l1_iter, k) / s.q16
         t_write = m * n * out_b / s.bw_pc
+        drain_pen = (b0 - 1) * (t_comp + t_write)
+
+        if s.t_cmd <= 0:
+            # T_cmd=0: 阈值 b0*(T_comp+T_write)/T_cmd -> +inf, 不可除零, 直接判负
+            detail = (f"k_L1={'K(截断)' if k_truncated else f'{k_l1_iter:.0f}<K(L1绑定)'}; "
+                      f"T_cmd=0: 合并无命令开销节省 (阈值=+inf), MergeBatch 无胜出通道; "
+                      f"仅剩 drain 惩罚=(b0-1)*(T_comp+T_write)={drain_pen*1e6:.2f}us")
+            return False, detail
+
         threshold = b0 * (t_comp + t_write) / s.t_cmd
 
         win = k_truncated and (b_core > threshold)
         detail = (f"k_L1={'K(截断)' if k_truncated else f'{k_l1_iter:.0f}<K(L1绑定)'}; "
                   f"b_core={b_core} vs 阈值 b0*(T_comp+T_write)/T_cmd={threshold:.1f}; "
-                  f"drain惩罚=(b0-1)*(T_comp+T_write)={((b0-1)*(t_comp+t_write))*1e6:.2f}us, "
+                  f"drain惩罚=(b0-1)*(T_comp+T_write)={drain_pen*1e6:.2f}us, "
                   f"搬移节省=b_core*(1-1/b0)*T_cmd={b_core*(1-1/b0)*s.t_cmd*1e6:.2f}us")
         return win, detail