Update BMM_Theory: bmm_theory/timing.py (fix review issues #4-#10)

BMM/BMM_Theory/bmm_theory/timing.py

@@ -98,18 +98,28 @@ def assemble_timing(t_mte2_gm: float, t_mte2_l2: float, t_dma_cmd: float,
                     t_drain: float,
                     gm_read_bytes: float, l2_read_bytes: float,
                     dma_cmd_count: float, cube_flops: float,
-                    fixpipe_bytes: float) -> HardwareTiming:
-    """汇总各级时延, 判定瓶颈."""
+                    fixpipe_bytes: float,
+                    reduce_serial: bool = True) -> HardwareTiming:
+    """汇总各级时延, 判定瓶颈.
+
+    归约计账约定 (issue#9): REDUCE 默认**串行追加** (reduce_serial=True, 归约不可
+    掩盖, 体现在 t_drain 中), 不进稳态 max(); 仅当调用方显式声明归约可流水掩盖
+    (reduce_serial=False) 时才进稳态 max(). 避免"既取最大又串行追加"的双倍计账.
+    """
     t_mte2 = t_mte2_gm + t_mte2_l2 + t_dma_cmd
     stages = {
         "MTE2_GM": t_mte2_gm + t_dma_cmd,
         "MTE2_L2": t_mte2_l2,
         "MMAD": t_mmad,
         "FIXPIPE": t_fixpipe,
-        "REDUCE": t_reduce,
     }
+    if not reduce_serial:
+        stages["REDUCE"] = t_reduce
     bottleneck = max(stages, key=stages.get)
     t_steady = max(stages.values())
+    # 归约串行追加时, 若它是全链路最大项则瓶颈标注为 REDUCE (但时延只计一次)
+    if reduce_serial and t_reduce > t_steady:
+        bottleneck = "REDUCE"
     return HardwareTiming(
         gm_read_bytes=gm_read_bytes, l2_read_bytes=l2_read_bytes,
         t_mte2_gm=t_mte2_gm, t_mte2_l2=t_mte2_l2, t_mte2=t_mte2,