docs: 新增 06_落地验证计划_理论方案_vs_bmmv3源码.md

理论方案落地 batch_mat_mul_v3 真实算子源码(arch35/Ascend950PR)的
比对分析与上板验证计划:

- 源码现状架构梳理(host 策略注册表静态优先级/kernel 7 参模板分发/
  平台带宽算力公式与理论口径天然对齐)
- 五分支逐项比对(StreamK/ASW 尾轮/IterBatch 形态/MergeBatch 准入仲裁/
  转Matmul特殊广播), 每项含差异清单+host/kernel 修改重点+
  预期时延变化(理论模型测算数字)+验证重点
- 收益排序: StreamK 准入放宽(定向 case -73%~-82%, host-only) >
  ASW 尾轮重切(-25%~-43%, 方案B host-only) > MergeBatch 准入补齐
  (防误捕获) > IterBatch 形态补全(kernel 重写, 收益待 T_cmd 标定)
- 上板验证方案(三层 case 集/标定项/通过标准)与分批 PR 计划

README 文档树补登 04/06 两篇。

.gitignore

@@ -12,6 +12,15 @@ venv/
 *~
 *.swp
 
+# BMM_Theory 调试/复现用临时 csv (issue#21, 防再次误提交; examples/ 下正式样例不受影响)
+/BMM/BMM_Theory/bug*.csv
+/BMM/BMM_Theory/bug*_out.csv
+/BMM/BMM_Theory/bug*_p.csv
+/BMM/BMM_Theory/t.csv
+/BMM/BMM_Theory/t_*.csv
+/BMM/BMM_Theory/t*.csv
+/BMM/BMM_Theory/p2.csv
+
 # IDE / OS
 .idea/
 .vscode/

BMM/BMM_Theory/README.md

@@ -49,10 +49,15 @@ python -m unittest discover -s tests -v
 
 ### 时延模型要点 (读结果前必看)
 
-- **搬入 MTE2 分两段**: `t_mte2_gm` (GM→L1 直读, 按 GM 带宽 1.6TB/s; 无论是否随路驻留 L2 都不再累加 L2→L1) 与 `t_mte2_l2` (数据驻留 L2 后重复读命中, 按 L2 带宽 5.2TB/s);
+- **GM 是读写共享总线 (1.6TB/s)**: MTE2 的 GM 读与 Fixpipe 直写 GM 并发时按 **(读+写)/1.6TB/s 累加**计入 MTE2 搬移链 (issue#23);
-- **Cube 计算** `t_mmad`: 单核算力 ≈15.2 TFLOPS, MergeBatch 的冗余计算计入;
+- **搬入 MTE2 分两段、同链相加**: `t_mte2_gm` (首读按 GM 带宽; 随路驻留 L2 不重复累加) + `t_mte2_l2` (共享块驻留 L2 后被其它核重复读, 按 L2 读口 5.2TB/s) + DMA 命令开销 (issue#24; **T_cmd 默认 0**, 未标定按 0 处理, issue#36);
-- **Fixpipe 搬出** `t_fixpipe`: 数据量按 **C 矩阵 dtype** 计 (fp16/fp8 时随路转换、写出量减半); StreamK 中间部分和为防精度丢失按 4B (L0C dtype) 计;
+- **搬移效率模型 (issue#36)**: 切B 两分支 (IterBatch/MergeBatch) 的 GM→L1 数据时延按单命令 tile 效率加权 —— eff = min(1, tile/min_TileSize), tile = nValue×dValue×dtype 达 16KB 饱和; **MergeBatch 合并 b0 个 batch 使单块 tile 放大 b0 倍, 即便 T_cmd=0 也比 IterBatch 逐 batch 搬移效率高** (堆叠方向视转置: A ND 非转置沿 M(nValue), B ND 非转置沿 N(dValue)); 只影响时间列, GM 字节量不变;
-- **总时延** `t_total = max(各级) + t_drain` (双缓冲稳态取最大 + 末级排空暴露);
+- **Cube 计算** `t_mmad`: 芯片算力 = 白皮书 Cube-only 432 TFLOPS (单核 13.5 TFLOPS, issue#40), MergeBatch 冗余计算计入;
+- **Fixpipe 搬出**: 直写 GM 计入 GM 共享总线; 驻留 L2 走 5.2TB/s 写口 (独立计时); 数据量按 **C 矩阵 dtype** 计 (fp16/fp8 随路转换减半); StreamK 部分和按 4B (L0C dtype);
+- **总时延** `t_total = max(MTE2搬移链, MMAD, Fixpipe-L2) + t_drain` (稳态取最大 + 末级排空暴露; REDUCE 串行追加);
+- **GM 读取下限**: 每输入字节至少从 GM 读一次 (R1); L2 只吸收"驻留后的再次读取"; 整 case 输入+输出 ≤ L2 时 GM 恰读一次、输出全驻留 L2 (issue#29 设计文档 docs/05);
+- **Fixpipe 输出落点**: 整 case (输入+输出+workspace) 可驻留 L2 → 输出写 L2 写口 5.2TB/s、GM 写流量 0; 否则输入优先保 L2, 输出直写 GM 计入共享总线 (issue#30);
+- **算力按输入 dtype**: Cube BF16/FP16 432TFLOPS 为基准 (Cube-only 口径, 白皮书表3-1 "Cube算力" 单行, issue#40), fp8=2x/fp4=4x (白皮书), fp32=1/2 (假设待标定); AIV 逐元素通量同理 (issue#28);
 - **瓶颈交换**: 搬移瓶颈可牺牲算力换搬移效率 (MergeBatch), 计算瓶颈可牺牲搬移换计算效率 (ASW_Basic 切 M/N).
 
 ## 目录结构
@@ -63,7 +68,8 @@ BMM_Theory/
 ├── bmm_theory/                # 软件包
 │   ├── __main__.py            # CLI 入口 (recommend / evaluate)
 │   ├── models.py              # 数据模型: BmmCase / ImplPlan(标准结构体) / HardwareTiming
-│   ├── hardware/ascend950pr.py# 950PR 硬件参数表 (换芯片只换这份)
+│   ├── hardware/              # 硬件参数表 (NpuSpec; 950PR 32/28核 + 950DT 36/32/28核,
+│   │                          #   SPECS/get_spec 索引, 默认 ASCEND950PR; 换芯片只加这份)
 │   ├── timing.py              # 时延评估引擎 (MTE2/Cube/Fixpipe/Reduce 模型)
 │   ├── constraints.py         # 单一约束源 (生成与校验共用, L0C/L0A/L0B/L1/dValue/核数)
 │   ├── router.py              # 分支决策路由 + 重叠区仲裁 + 生成后自检
@@ -88,6 +94,9 @@ BMM_Theory/
 │   │   ├── 06_ASW_Basic分支.md      # 尾轮策略已内化为其必要环节
 │   │   └── 07_尾轮处理策略.md        # 尾轮完整推导 (参考)
 │   └── 03_测评报告/             # 外部测评报告 (v1.0/v2.0 及后续复评)
+│   ├── 04_差异对照_bmmv3_vs_BMM_Theory.md   # bmmv3 行为预测器 vs 理论推导器对照
+│   ├── 05_L2驻留GM读写与dtype算力口径_设计分析.md   # GM/L2/输出落点/dtype算力统一口径 (issue#27-#30)
+│   └── 06_落地验证计划_理论方案_vs_bmmv3源码.md     # 理论落地 bmmv3 源码的比对/修改/上板验证计划
 ├── examples/                  # 示例输入输出
 └── tests/                     # 单元测试 (固化文档边界 case + issue 回归)
 ```

BMM/BMM_Theory/bmm_theory/branches/asw_basic.py

@@ -1,8 +1,10 @@
 """ASW_Basic 分支: 核间切 M/N (或混合切) 的兜底分支, 含尾轮处理.
 
 理论依据:
-  - 《BMM算子优化分析 v0.98》§八 + docs/02_分支理论/06_ASW_Basic分支.md
+  - 《ASW_Basic分支分析 v1.91》(L0C 单缓冲方形 Base tile §5.1 + SingleCoreM/N 有界
-  - 《BMM尾轮处理策略对比分析 v1.5》+ docs/02_分支理论/07_尾轮处理策略.md
+    枚举 §5.2) + docs/02_分支理论/06_ASW_Basic分支.md
+  - 《BMM尾轮处理策略对比分析 v1.5》(单块搬入稳态口径 K(sM+sN)·dt/BW, k_L1 约掉)
+    + docs/02_分支理论/07_尾轮处理策略.md
 
 核心: 兜底分支, 核间切 M/N, 重复读交给 L2 + swizzle. 尾轮处理是必要组成环节:
 默认方案 B (工程简洁, 主流场景与 A1b 严格打平), 周长型且 rho>=rho_dv 时 A1b.
@@ -13,7 +15,7 @@ from __future__ import annotations
 import math
 
 from ..hardware import NpuSpec, ASCEND950PR
-from ..models import BmmCase, ImplPlan, HardwareTiming, ceil_div, align_down
+from ..models import BmmCase, ImplPlan, HardwareTiming, ceil_div, align_down, align_up
 from ..timing import assemble_timing
 from .base import Branch, ConditionCheck
 
@@ -60,47 +62,44 @@ class AswBasicBranch(Branch):
             return self._plan_reduced_core(case, p)
 
         # ---- 正常模式: Step 0~6 ----
-        # Step 0: BaseM/BaseN 用满 L0C (32768 元素双缓冲)
+        # Step 0: BaseM/BaseN (v1.91 §5.1): UnitFlag 单缓冲 (tile 内 16x16x16 细粒度
-        base_mn = int(math.sqrt(s.l0c_bytes // 8))  # L0C/(2*4B)
+        #   流水替代 tile 间粗粒度双缓冲), BaseM*BaseN = L0C/4B = 65536 元素, 方形
-        base_m = align_down(base_mn, s.fractal)
+        #   优先 (256x256, L0A/L0B 同时装满、baseK 加倍); 仅当 M/N < 方形边长时被迫
-        base_n = align_down(base_mn, s.fractal)
+        #   跟随 M/N (另一侧按 L0C 面积余量放大); BaseK 由 L0A/L0B (双缓冲) 反推.
-        base_k = align_down(int(min(
+        from ..constraints import clamp_base_k
-            s.l0a_bytes / (2 * base_m * dt),
+        cap_l0c = s.l0c_bytes // 4                         # 65536 元素 (单缓冲)
-            s.l0b_bytes / (2 * base_n * dt),
+        sq = align_down(int(math.sqrt(cap_l0c)), s.fractal)    # 256
-        )), s.fractal)
+        # L0A/L0B 在 base_k>=16 (fractal) 下的单边上限 (v1.91 §5.7 核内约束;
+        # 防止例外路径 (如 M=16 -> BaseN=4096) 超 L0B)
+        side_cap = s.l0a_bytes // (2 * s.fractal * dt)         # L0A=L0B=64KB
+        if m >= sq and n >= sq:
+            base_m = base_n = sq
+        elif m < sq:
+            base_m = max(align_down(m, s.fractal), s.fractal)
+            base_n = max(align_down(min(cap_l0c // base_m, n, side_cap),
+                                    s.fractal), s.fractal)
+        else:  # n < sq
+            base_n = max(align_down(n, s.fractal), s.fractal)
+            base_m = max(align_down(min(cap_l0c // base_n, m, side_cap),
+                                    s.fractal), s.fractal)
+        base_k = clamp_base_k(base_m, base_n, dt, case.k, s)   # v1.91 §5.7
 
-        # Step 1: SingleCoreM/N 尽量大 (满足并行度下限)
+        # Step 1+2+6: SingleCoreM/N + k_L1 + mCnt/nCnt 有界枚举 (v1.91 §5.2):
-        min_blocks = ceil_div(s.aic_num, b)
+        (single_m, single_n, k_l1, m_cnt, n_cnt, pick_note) = \
-        single_m, single_n = self._pick_single_core(m, n, min_blocks, base_m, base_n, dt)
+            self._pick_single_core(case, base_m, base_n)
-
-        # Step 2: mCnt/nCnt
-        m_cnt = ceil_div(m, single_m)
-        n_cnt = ceil_div(n, single_n)
 
         # Step 4: swizzle 窗口 W = max{d | d|C, d <= floor(sqrt(C))}
         swizzle_w = self._swizzle_w()
 
-        # Step 5: L2 分组判断
+        # Step 5: L2 场景判定 (issue#24/#30/#32, 设计文档 docs/05 §4, 芯片 L2 128MB)
-        s_in = b * (m * k + k * n) * dt
+        #   S_A 整case全驻留: V_in+V_out <= L2 -> 输出驻留 L2 (R4, GM 写=0);
-        s_out = b * m * n * case.dtype_out_bytes
+        #   S_B 单batch可驻留 (全驻留或单侧全驻留+对侧滑窗): GM 每字节一次,
-        if s_in + s_out <= s.l2_bytes:
+        #       共享块重复读全部命中 L2 读口, 输出直写 GM;
-            l2_scene = "A_全驻留"
+        #   S_C 双侧超L2: 最小替换 2D 分组 (组间落空 GM, 窗口 L2).
-            l2_out = "resident(输出驻留L2异步回写)"
+        scene = self._l2_scene(case, single_m, single_n, m_cnt, n_cnt)
-            r_in = 1.0
+        l2_out = ("resident(整case全驻留S_A: 输出驻留L2异步回写, GM写=0)"
-        elif s_in <= s.l2_bytes:
+                  if scene["to_l2"] else
-            l2_scene = "B_输入驻留输出直写GM"
+                  "direct_gm(输出直写GM, 输入优先驻留L2)")
-            l2_out = "direct_gm(输出直写GM不占L2)"
-            r_in = 1.0
-        else:
-            l2_scene = "C_输入超L2分组执行"
-            l2_out = "direct_gm(输出直写GM不占L2)"
-            r_in = self._l2_group_r_in(case, single_m, single_n)
-
-        # Step 6: k_l1 (GM->L1 K 向粒度, 须 >= dValue 下限; K 小于下限时整 K 一次搬入不切)
-        dv_min_elems = max(s.dvalue_hw_min // dt, 1)
-        k_l1 = min(k, s.dvalue_recommend // dt)
-        if k_l1 < dv_min_elems:
-            k_l1 = k  # K 本身小于 dValue 下限: 不切 K, 整段搬入 (K 非连续维, dValue 由 M/N 保证)
 
         # ---- 尾轮决策 (必要组成环节) ----
         n_blk = b * m_cnt * n_cnt
@@ -122,9 +121,15 @@ class AswBasicBranch(Branch):
             tail_m_cnt=tail["tail_m_cnt"], tail_n_cnt=tail["tail_n_cnt"],
             tail_k_cnt=1, tail_m_main=tail["tail_m_main"], tail_n_main=tail["tail_n_main"],
             tail_block_cnt=tail["r"], tail_wave_num=tail["n_wave"],
-            fixpipe_unitflag=True,
+            fixpipe_unitflag=True,   # UnitFlag 单缓冲: tile 内 16x16x16 细粒度流水
             out_dtype_bytes=case.dtype_out_bytes,
-            note=f"L2场景{l2_scene}, r_in={r_in:.2f}; 尾轮: {tail['reason']}",
+            note=(f"tile枚举: {pick_note}; Base tile {base_m}x{base_n} "
+                  f"(L0C 单缓冲方形用满); "
+                  f"L2场景: {scene['label']} (V_in={case.input_bytes/1048576:.1f}MB, "
+                  f"V_out={case.output_bytes/1048576:.1f}MB, "
+                  f"L2={s.l2_bytes/1048576:.0f}MB); "
+                  + (scene["note_extra"] + "; " if scene["note_extra"] else "")
+                  + f"尾轮: {tail['reason']}"),
         )
 
     # ------------------------------------------------------------------
@@ -142,6 +147,7 @@ class AswBasicBranch(Branch):
         single_m, single_n = clamp_base_mn_l0c(single_m, single_n, False, s)
         # base_k 由 L0A/L0B 容量按 dtype 反推 (issue#5: 原硬编码 min(K,64) 在 fp32 下溢出)
         base_k = clamp_base_k(single_m, single_n, dt, case.k, s)
+        out_l2 = case.input_bytes + case.output_bytes <= s.l2_bytes  # R4 (issue#30)
         return ImplPlan(
             case_id=case.case_id, branch=self.name + "_降核",
             used_core_num=used,
@@ -154,7 +160,9 @@ class AswBasicBranch(Branch):
             l1_form="标准核内流水",
             base_m=single_m, base_n=single_n,
             base_k=base_k,
-            l2_policy_in="allocate", l2_policy_out="direct_gm",
+            l2_policy_in="allocate",
+            l2_policy_out=("resident(整case全驻留S_A)"
+                           if out_l2 else "direct_gm(输出直写GM)"),
             swizzle_w=0, workspace_bytes=0,
             tail_strategy="不涉及(每核一块无尾轮)",
             fixpipe_unitflag=True, out_dtype_bytes=case.dtype_out_bytes,
@@ -162,24 +170,126 @@ class AswBasicBranch(Branch):
         )
 
     # ------------------------------------------------------------------
-    def _pick_single_core(self, m, n, min_blocks, base_m, base_n, dt):
+    def _pick_single_core(self, case: BmmCase, base_m: int, base_n: int) -> tuple:
-        """Step 1: 满足并行度下限前提下 SingleCoreM/N 尽量大, 长宽比跟随 M/N."""
+        """SingleCoreM/N + k_L1 + mCnt/nCnt 有界枚举 (v1.91 §5.2 + v1.5 §2.1 修正口径).
+
+        - P = ⌈C/B⌉;  P=1 (B>=C): 先试不切分 (mCnt=nCnt=1, tile 跟随 M/N, 情形1),
+          约束不满足则进入枚举强制切分;
+        - 枚举空间: mCnt ∈ [1, ⌈M/BaseM⌉], nCnt ∈ [1, ⌈N/BaseN⌉] (约束4 上界:
+          SingleCore 为 Base 整数倍), 且 B·mCnt·nCnt >= C (约束1 并行度);
+        - 约束 2 (L1 双缓冲): k_L1 = min(K, ⌊L1/(2(sM+sN)·dt)⌋16), 须 ≥ 一个
+          fractal 且满足 dValue 下限;
+        - 约束 3 (搬移效率, 转置感知): A 非转置 dValue = k_L1·dt / A 转置 = sM·dt;
+          B 非转置 dValue = sN·dt / B 转置 = k_L1·dt; 均须 ≥ 256B (dValue 硬件
+          突发下限); 且 sM·k_L1·dt 与 k_L1·sN·dt ≥ min_TileSize (16KB);
+        - 目标 (v1.5 §2.1 修正: 稳态流水下单块搬入 = K·(sM+sN)·dt, 与分次粒度
+          k_L1 无关, k_L1 只进约束): 每 batch 搬入量 K·dt·(nCnt·M + mCnt·N) 最小
+          (共享块重复读最少 = tile 尽可能大/少); 并列取 r = B·mCnt·nCnt mod C 最大
+          (尾轮块越多, 重切收益越大, v1.91 §5.2.d);
+        返回 (single_m, single_n, k_l1, m_cnt, n_cnt, note).
+        """
         s = self.spec
-        # 从大到小枚举 (m_cnt*n_cnt >= min_blocks), 取最大 tile
+        b = case.batch_c
-        best = (base_m, base_n)
+        m, n, k = case.m, case.n, case.k
-        for m_cnt in range(1, int(math.sqrt(min_blocks)) + 2):
+        dt = case.dtype_in_bytes
-            n_cnt = ceil_div(min_blocks, m_cnt)
+        c = s.aic_num
-            if m_cnt * n_cnt < min_blocks:
+        p_min = ceil_div(c, b)        # 最少切分块数 (约束1)
-                continue
+        dv_min = s.dvalue_hw_min      # 256B
-            sm = align_down(ceil_div(m, m_cnt), base_m) or base_m
+        min_tile = s.min_tile_size    # 16KB
-            sn = align_down(ceil_div(n, n_cnt), base_n) or base_n
+
-            # 约束 2: L1 容量 2(sM+sN)*k_l1*dt <= L1, k_l1 取 256B/dt
+        def eval_tile(sm: int, sn: int):
-            k_l1_min = s.dvalue_hw_min // dt
+            """约束 2/3 + 目标值; 返回 (ok, k_l1, traffic_per_batch)."""
-            if 2 * (sm + sn) * k_l1_min * dt > s.l1_bytes:
+            k_cap = int(s.l1_bytes / (2 * (sm + sn) * dt))
-                continue
+            k_l1 = align_down(min(k, k_cap), s.fractal)
-            if sm * sn > best[0] * best[1]:
+            if k_l1 < s.fractal:
-                best = (sm, sn)
+                return False, 0, None
-        return best
+            # dValue 下限 (转置感知连续维, 同 issue#19 口径)
+            dv_a = sm * dt if case.trans_a else k_l1 * dt
+            dv_b = k_l1 * dt if case.trans_b else sn * dt
+            if dv_a < dv_min or dv_b < dv_min:
+                return False, 0, None
+            # 单次搬移量下限
+            if sm * k_l1 * dt < min_tile or k_l1 * sn * dt < min_tile:
+                return False, 0, None
+            mc = ceil_div(m, sm)
+            nc = ceil_div(n, sn)
+            traffic = k * dt * (nc * m + mc * n)   # 每 batch 总搬入字节
+            return True, k_l1, traffic
+
+        # 情形 1: P=1 (B >= C) 先试不切分, tile 跟随 M/N
+        if p_min == 1:
+            sm = max(align_up(m, s.fractal), s.fractal)
+            sn = max(align_up(n, s.fractal), s.fractal)
+            ok, k_l1, traffic = eval_tile(sm, sn)
+            if ok:
+                return (sm, sn, k_l1, 1, 1,
+                        "P=1(B>=C) 不切分, tile 跟随 M/N (v1.91 §5.2 情形1)")
+
+        # 情形 2: 有界枚举
+        best = None
+        m_max = ceil_div(m, base_m)
+        n_max = ceil_div(n, base_n)
+        for mc in range(1, m_max + 1):
+            sm = align_up(ceil_div(m, mc), base_m)
+            for nc in range(1, n_max + 1):
+                sn = align_up(ceil_div(n, nc), base_n)
+                ok, k_l1, traffic = eval_tile(sm, sn)
+                if not ok:
+                    continue
+                mc_r = ceil_div(m, sm)
+                nc_r = ceil_div(n, sn)
+                # 约束 1: 并行度 (真实块数, 考虑 align 上取后的收缩)
+                if b * mc_r * nc_r < c:
+                    continue
+                r = (b * mc_r * nc_r) % c
+                key = (traffic, -r)   # 主键搬入量最小; 并列 r 最大
+                if best is None or key < best[0]:
+                    best = (key, sm, sn, k_l1, mc_r, nc_r, traffic, r)
+        if best is not None:
+            _, sm, sn, k_l1, mc_r, nc_r, traffic, r = best
+            return (sm, sn, k_l1, mc_r, nc_r,
+                    f"P={p_min}, 有界枚举最优 mCnt={mc_r} x nCnt={nc_r} "
+                    f"(tile {sm}x{sn}, 每batch搬入{traffic/1048576:.1f}MB, r={r})")
+
+        # 情形 3: 放开约束 4 (Base 整数倍), 16 对齐网格 + 硬下限 (dValue>=128B) 再搜
+        # —— 兜底分支恒出方案 (issue#34): 256B 偏好无解时按硬下限给最优可行 tile,
+        #    标注效率降级 (搬移效率低于模型假设, 时延可能低估)
+        dv_hard = s.dvalue_min  # 128B 硬下限
+        best2 = None
+        for sm in range(s.fractal, align_up(min(m, 1024), s.fractal) + 1, s.fractal):
+            for sn in range(s.fractal, align_up(min(n, 1024), s.fractal) + 1, s.fractal):
+                k_cap = int(s.l1_bytes / (2 * (sm + sn) * dt))
+                k_l1 = align_down(min(k, k_cap), s.fractal)
+                if k_l1 < s.fractal:
+                    continue
+                dv_a = sm * dt if case.trans_a else k_l1 * dt
+                dv_b = k_l1 * dt if case.trans_b else sn * dt
+                if dv_a < dv_hard or dv_b < dv_hard:
+                    continue
+                mc_r = ceil_div(m, sm)
+                nc_r = ceil_div(n, sn)
+                if b * mc_r * nc_r < c:
+                    continue
+                traffic = k * dt * (nc_r * m + mc_r * n)
+                r = (b * mc_r * nc_r) % c
+                key = (traffic, -r)
+                if best2 is None or key < best2[0]:
+                    best2 = (key, sm, sn, k_l1, mc_r, nc_r, traffic, r)
+        if best2 is not None:
+            _, sm, sn, k_l1, mc_r, nc_r, traffic, r = best2
+            return (sm, sn, k_l1, mc_r, nc_r,
+                    f"效率降级(放开约束4, dValue 按 128B 硬下限, 搬移效率低于模型假设, "
+                    f"时延可能低估): P={p_min}, mCnt={mc_r} x nCnt={nc_r} "
+                    f"(tile {sm}x{sn}, 每batch搬入{traffic/1048576:.1f}MB, r={r})")
+
+        # 情形 4: 极端兜底 —— 兜底分支恒出方案 (issue#34): 效率下限物理不可满足
+        # 也照常给方案 + 标注效率降级 (搬移效率崩塌), 不判违规/不产 None
+        k_l1 = max(align_down(min(k, int(s.l1_bytes /
+                                         (2 * (base_m + base_n) * dt))),
+                              s.fractal), s.fractal)
+        return (base_m, base_n, k_l1, ceil_div(m, base_m), ceil_div(n, base_n),
+                "效率降级(极端形状: 效率下限物理不可满足, 搬移效率崩塌, 时延可能低估; "
+                "方案供参考, 建议调整 dtype/布局或转置)")
 
     def _swizzle_w(self) -> int:
         s = self.spec
@@ -189,17 +299,102 @@ class AswBasicBranch(Branch):
                 w = d
         return w
 
-    def _l2_group_r_in(self, case, sm, sn) -> float:
+    def _l2_scene(self, case: BmmCase, sm: int, sn: int,
-        """场景 C: L2 分组的重复读倍率 r_in = (n_grp*M + m_grp*N)/(M+N)."""
+                  m_cnt: int | None = None, n_cnt: int | None = None) -> dict:
+        """L2 场景判定 (make_plan 与 evaluate 共用, 设计文档 docs/05 §4.1, issue#32).
+
+        判定顺序 S_A -> S_B -> S_C (L2 为整芯片 128MB; 返回"每 batch"流量):
+          S_A 整case全驻留:  V_in + V_out <= L2 -> 输出驻留 L2 (R4, GM 写=0);
+          S_B 单batch可驻留: 单 batch 输入可全驻留 (a_b+bb_b<=L2) 或**单侧全驻留 +
+             对侧滑窗** (驻留侧 <= L2 且 + 2x对侧单块 <= L2) -> GM 每字节恰读一次
+             (V_in), 共享块重复读全部命中 L2 读口 (n_cnt-1)/(m_cnt-1) 次;
+          S_C 双侧均不可全驻留: 整 L2 容量约束下搜索**最小 GM** 的 2D 分组
+             (力求最小 L2 替换), 组间共享块落空计 GM、组内窗口复用计 L2.
+        """
         s = self.spec
-        m, n, k, b = case.m, case.n, case.k, case.batch_c
+        m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
-        d = s.l2_bytes / (b * k * dt)
+        a_b = m * k * dt
-        m_grp = max(1, int(d / (2 * sm)))
+        bb_b = k * n * dt
-        n_grp = max(1, int(d / (2 * sn)))
+        if m_cnt is None:
-        m_cnt = ceil_div(m, sm)
+            m_cnt = max(ceil_div(m, max(sm, 1)), 1)
-        n_cnt = ceil_div(n, sn)
+        if n_cnt is None:
-        return (ceil_div(n_cnt, n_grp) * m + ceil_div(m_cnt, m_grp) * n) / (m + n)
+            n_cnt = max(ceil_div(n, max(sn, 1)), 1)
+        blk_a = a_b / max(m_cnt, 1)     # A 行块字节 (tile 行宽 x K)
+        blk_b = bb_b / max(n_cnt, 1)    # B 列块字节 (K x tile 列宽)
+        if case.input_bytes + case.output_bytes <= s.l2_bytes:
+            return {"code": "S_A",
+                    "label": "A_整case全驻留(输入+输出<=L2)",
+                    "to_l2": True,
+                    "gm_batch": a_b + bb_b,
+                    "l2_batch": (n_cnt - 1) * a_b + (m_cnt - 1) * bb_b,
+                    "note_extra": ""}
+        if (a_b + bb_b <= s.l2_bytes or          # 双侧全驻留
+                bb_b + 2 * blk_a <= s.l2_bytes or  # B 驻留 + A 行块滑窗
+                a_b + 2 * blk_b <= s.l2_bytes):    # A 驻留 + B 列块滑窗
+            return {"code": "S_B",
+                    "label": "B_单batch可驻留(全驻留或单侧驻留+对侧滑窗)",
+                    "to_l2": False,
+                    "gm_batch": a_b + bb_b,
+                    "l2_batch": (n_cnt - 1) * a_b + (m_cnt - 1) * bb_b,
+                    "note_extra": ""}
+        g = self._l2_group_min_gm(case, m_cnt, n_cnt, blk_a, blk_b)
+        return {"code": "S_C",
+                "label": "C_双侧超L2: 最小替换2D分组(组间落空GM, 窗口L2)",
+                "to_l2": False,
+                "gm_batch": g["gm_batch"],
+                "l2_batch": g["l2_batch"],
+                "note_extra": g["note"]}
+
+    def _l2_group_min_gm(self, case: BmmCase, m_cnt: int, n_cnt: int,
+                         blk_a: float, blk_b: float) -> dict:
+        """场景 C: 整 L2 容量约束下的最小 GM 2D 分组 (issue#32, docs/05 §4.3b).
+
+        容量约束: m_grp*blk_a + n_grp*blk_b <= L2 (组工作集占满整 L2, 不再对半);
+        目标:     GM = ceil(n_cnt/n_grp)*a_b + ceil(m_cnt/m_grp)*bb_b 最小
+                  (A 行块每列组一次 GM 首读; B 列块每行组一次 GM 首读);
+        窗口 L2:  (n_cnt - ceil(n_cnt/n_grp))*a_b + (m_cnt - ceil(m_cnt/m_grp))*bb_b
+                  (组内共享块其余次复用走 L2 读口, 与 S_B 口径一致, issue#32 失真b);
+        无可行分组 (单块对都超 L2) 时保守回落: 每共享块独立落 GM, 窗口不计.
+        """
+        s = self.spec
+        m, n, k = case.m, case.n, case.k
+        dt = case.dtype_in_bytes
+        a_b = m * k * dt
+        bb_b = k * n * dt
+        best = None
+        for mg in range(1, m_cnt + 1):
+            for ng in range(1, n_cnt + 1):
+                if mg * blk_a + ng * blk_b > s.l2_bytes:
+                    continue
+                ga = ceil_div(n_cnt, ng)
+                gb = ceil_div(m_cnt, mg)
+                gm = ga * a_b + gb * bb_b
+                if best is None or gm < best[0]:
+                    best = (gm, mg, ng, ga, gb)
+        if best is None:
+            return {"gm_batch": n_cnt * a_b + m_cnt * bb_b, "l2_batch": 0.0,
+                    "note": "分组无可行解(单块对超L2): 保守每共享块独立落GM"}
+        gm, mg, ng, ga, gb = best
+        l2 = (n_cnt - ga) * a_b + (m_cnt - gb) * bb_b
+        r = gm / (a_b + bb_b)
+        return {"gm_batch": gm, "l2_batch": l2,
+                "note": f"最小替换分组 m_grp={mg}x n_grp={ng} (GM倍率{r:.2f}, "
+                        f"窗口L2/batch={l2/1048576:.1f}MB)"}
+
+    # ------------------------------------------------------------------
+    def _block_times(self, case: BmmCase, sm: int, sn: int) -> tuple:
+        """单块三段时延 (尾轮文档 v1.5 §2.1 口径: k_L1 稳态约掉, L2 命中带宽).
+
+        返回 (t_mmad_blk, t_mte2_blk, t_fix_blk); _decide_tail 主导项判定与
+        evaluate 尾轮残余 drain (issue#37) 同源共用.
+        """
+        s = self.spec
+        qc = s.q_cube(case.dtype_a, case.dtype_b)   # issue#28
+        t_mmad = 2 * sm * sn * case.k / qc
+        t_mte2 = case.k * (sm + sn) * case.dtype_in_bytes / s.bw_l2_pc
+        t_fix = sm * sn * case.dtype_out_bytes / s.bw_pc
+        return t_mmad, t_mte2, t_fix
 
     # ------------------------------------------------------------------
     def _decide_tail(self, case, sm, sn, n_blk, k_l1) -> dict:
@@ -210,7 +405,6 @@ class AswBasicBranch(Branch):
         r = n_blk % c
         rho = r / c
         dt = case.dtype_in_bytes
-        out_b = case.dtype_out_bytes
 
         base = dict(r=r, n_wave=n_wave, tail_m_cnt=1, tail_n_cnt=1,
                     tail_m_main=0, tail_n_main=0)
@@ -218,11 +412,8 @@ class AswBasicBranch(Branch):
         if r == 0:
             return {**base, "strategy": "A0", "reason": "r=0 无尾轮"}
 
-        # 主导项判定
+        # 主导项判定 (块级三段, 与 evaluate drain 同源 _block_times)
-        bw_eff = s.bw_l2_pc  # L2 命中
+        t_mmad, t_mte2, t_fix = self._block_times(case, sm, sn)
-        t_mmad = 2 * sm * sn * case.k / s.q16
-        t_mte2 = case.k * (sm + sn) * dt / bw_eff
-        t_fix = sm * sn * out_b / s.bw_pc
         t_block = max(t_mmad, t_mte2, t_fix)
         area_dominated = t_block != t_mte2  # 面积型 = MMAD 或 FIX 主导
 
@@ -261,37 +452,82 @@ class AswBasicBranch(Branch):
 
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
+        """MTE2 两段块级模型 (issue#24 用户口径 + #28/#29/#30/#32):
+
+        - 首读走 GM (按 GM 带宽, 不叠加 L2); 共享块 (A 行块被 n_cnt 个 tile 读、
+          B 列块被 m_cnt 个 tile 读) 驻留 L2 后其余 (n_cnt-1)/(m_cnt-1) 次读走
+          L2 读口 (5.2TB/s 独享);
+        - 场景 (与 make_plan 同源 _l2_scene, 返回每 batch GM/L2 流量):
+          S_A/S_B: GM = V_in 一次 (S_B 含单侧全驻留+对侧滑窗, issue#32);
+          S_C: 最小替换 2D 分组 (整 L2 容量约束下最小 GM), 组间落空计 GM、
+               组内窗口复用计 L2;
+        - 输出落点 R4 (issue#30): 仅 S_A (整 case 输入+输出 <= L2) 驻留 L2
+          (5.2 写口, GM 写 = 0); S_B/S_C 直写 GM —— GM 读写共享总线累加由
+          assemble 的 MTE2 链处理 (issue#23);
+        - 字节列整芯片口径 (issue#29); Cube 算力按输入 dtype (issue#28);
+        - t_drain = 尾轮残余闭式 (issue#37, v1.5 §3.3/§4.3/§5): A0 (1-ρ)·T_block;
+          周长型 A1b/方案B 残余见下; 面积型与 r=0 恒 0 (与稳态严格相等).
+        """
         s = self.spec
         b = case.batch_c
         m, n, k = case.m, case.n, case.k
-        dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
+        used = max(plan.used_core_num, 1)
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
 
         flops = 2.0 * b * m * n * k
-        t_mmad = flops / (plan.used_core_num * s.q16)
+        t_mmad = flops / (used * qc)
 
-        in_bytes = b * (m * k + k * n) * dt
+        # ---- 字节量 (整芯片口径) ----
-        out_bytes = b * m * n * out_b
+        out_all = b * m * n * out_b   # 输出总字节
-        # r_in 从 note 里解析困难, 重新计算
+        m_cnt = max(plan.m_cnt, 1)
-        s_in = in_bytes
+        n_cnt = max(plan.n_cnt, 1)
-        s_out = out_bytes
-        if s_in + s_out <= s.l2_bytes or s_in <= s.l2_bytes:
-            r_in = 1.0
-        else:
-            r_in = self._l2_group_r_in(case, plan.single_core_m, plan.single_core_n)
 
-        t_mte2 = r_in * in_bytes / (plan.used_core_num * s.bw_pc)
+        # ---- L2 场景 (S_A/S_B/S_C, 与 make_plan 同源; 每 batch 流量 x B) ----
-        to_l2 = "resident" in plan.l2_policy_out
+        scene = self._l2_scene(case, plan.single_core_m, plan.single_core_n,
-        t_fix = out_bytes / (plan.used_core_num * (s.bw_l2_pc if to_l2 else s.bw_pc))
+                               m_cnt, n_cnt)
+        to_l2 = scene["to_l2"]
+        gm_read = b * scene["gm_batch"]
+        l2_read = b * scene["l2_batch"]
+        t_gm = gm_read / (used * s.bw_pc)
+        t_l2 = l2_read / (used * s.bw_l2_pc)
 
-        # drain: 尾轮暴露 (方案 B 已均匀重切, drain 小; A1b 尾轮凑满, drain 小; A0 尾轮 r 核空转)
+        # ---- Fixpipe (R4) ----
+        t_fix = out_all / (used * (s.bw_l2_pc if to_l2 else s.bw_pc))
+
+        # ---- drain: 尾轮残余时延 (issue#37, 尾轮 v1.5 §3.3/§4.3/§5 闭式) ----
+        # 稳态聚合 t_steady ≡ (n_wave-1+rho)·T_block (均匀分块下 N_blk/used 恒等),
+        # drain 只计尾轮结构相对该稳态的残余:
+        #   A0 (r>0):        (1-rho)·T_block      (T_A0 = n_wave·T_block)
+        #   面积型 A1b/方案B: 0                    (§4.3 总量守恒, 与稳态严格相等)
+        #   周长型 A1b:      (sqrt(rho)-rho)·T_load        (§5: (n_wave-1+√ρ)·T_load)
+        #   周长型 方案B:    (√(n_wave(n_wave-1+rho))-(n_wave-1+rho))·T_load
+        # 块级时延与 _decide_tail 同源 (_block_times, L2 命中口径); 未建模策略
+        # (如 A1a) 维持原口径 0。首块填充/末块排空 (v1.91 §3.1 O(T_comp+T_write))
+        # 在 UnitFlag 16-granule 细粒度流水下为 granule 级小量, 量级未标定,
+        # 不入模型 (docs/02_分支理论/06 §3 标注)。
         t_drain = 0.0
-        if plan.tail_strategy == "A0" and plan.tail_block_cnt > 0:
+        r = plan.tail_block_cnt
-            t_drain = max(t_mmad, t_mte2, t_fix)  # 尾轮空转一个整块
+        if r > 0:
+            n_wave = plan.tail_wave_num if plan.tail_wave_num > 0 \
+                else ceil_div(b * m_cnt * n_cnt, used)
+            rho = min(r / used, 1.0)
+            t_mm_b, t_mv_b, t_fx_b = self._block_times(
+                case, plan.single_core_m, plan.single_core_n)
+            t_block = max(t_mm_b, t_mv_b, t_fx_b)
+            if plan.tail_strategy == "A0":
+                t_drain = (1.0 - rho) * t_block
+            elif t_block == t_mv_b:   # 仅周长型主导有残余 (面积型严格 0)
+                if plan.tail_strategy == "A1b":
+                    t_drain = (math.sqrt(rho) - rho) * t_mv_b
+                elif plan.tail_strategy == "方案B":
+                    t_drain = (math.sqrt(n_wave * (n_wave - 1 + rho))
+                               - (n_wave - 1 + rho)) * t_mv_b
 
         return assemble_timing(
-            t_mte2_gm=t_mte2, t_mte2_l2=0.0, t_dma_cmd=0.0,
+            t_mte2_gm=t_gm, t_mte2_l2=t_l2, t_dma_cmd=0.0,
             t_mmad=t_mmad, t_fixpipe=t_fix, t_reduce=0.0, t_drain=t_drain,
-            gm_read_bytes=r_in * in_bytes, l2_read_bytes=0.0, dma_cmd_count=0.0,
+            gm_read_bytes=gm_read, l2_read_bytes=l2_read, dma_cmd_count=0.0,
-            cube_flops=flops, fixpipe_bytes=out_bytes,
+            cube_flops=flops, fixpipe_bytes=out_all,
+            fixpipe_to_gm=(not to_l2),
         )

BMM/BMM_Theory/bmm_theory/branches/iter_batch.py

@@ -12,8 +12,8 @@ batch 间流水靠 L1 双 buffer / 驻留侧预取 (c 形态半预算) 掩盖.
 from __future__ import annotations
 
 from ..hardware import NpuSpec, ASCEND950PR
-from ..models import BmmCase, ImplPlan, HardwareTiming, align_down
+from ..models import BmmCase, ImplPlan, HardwareTiming, align_down, gm_move_time
-from ..timing import assemble_timing
+from ..timing import assemble_timing, output_to_l2
 from .base import Branch, BranchResult, ConditionCheck
 
 
@@ -23,11 +23,34 @@ class IterBatchBranch(Branch):
     def __init__(self, spec: NpuSpec = ASCEND950PR):
         super().__init__(spec)
 
+    # ------------------------------------------------------------------
+    # 单命令搬移 tile (issue#36 效率模型口径, evaluate 与 MergeBatch 仲裁共用)
+    # ------------------------------------------------------------------
+    def move_tiles(self, case: BmmCase) -> tuple:
+        """返回 (A侧单命令tile字节, B侧单命令tile字节).
+
+        a/b 形态: 双侧整 K 一次搬入; c 形态: 驻留侧整 K + 对侧 k_l1 分块;
+        d 形态: 双侧 k_l1 分块。tile = nValue*dValue*dt (乘积与转置/排布无关)。
+        """
+        m, n, k = case.m, case.n, case.k
+        dt = case.dtype_in_bytes
+        form, k_l1, _, resident = self.l1_form(case)
+        if form in ("a", "b"):
+            return m * k * dt, k * n * dt
+        if form == "c":
+            return (m * k * dt, k_l1 * n * dt) if resident == "A" \
+                else (m * k_l1 * dt, k * n * dt)
+        return m * k_l1 * dt, k_l1 * n * dt   # d 或 None (兜底)
+
     # ------------------------------------------------------------------
     # L1 驻留形态判定 (v0.98 §六 条件 3, 四选一)
     # ------------------------------------------------------------------
     def l1_form(self, case: BmmCase) -> tuple:
-        """返回 (形态 'a'/'b'/'c'/'d'/None, k_l1, 说明)."""
+        """返回 (形态 'a'/'b'/'c'/'d'/None, k_l1, 说明, 驻留侧 'A'/'B'/None).
+
+        驻留侧仅 c 形态非 None (issue#36: 效率模型需要区分两侧单命令 tile —
+        c 形态驻留侧整侧一次搬入 (全 K), 对侧按 k_l1 分块).
+        """
         s = self.spec
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
@@ -36,31 +59,42 @@ class IterBatchBranch(Branch):
 
         # a) 单 batch 全驻留
         if b_core == 1 and single <= s.l1_bytes:
-            return "a", k, f"单batch全驻留: (MK+KN)*dtype={single/1024:.0f}KB <= L1"
+            return "a", k, f"单batch全驻留: (MK+KN)*dtype={single/1024:.0f}KB <= L1", None
 
         # b) 双 batch 乒乓
         if b_core > 1 and 2 * single <= s.l1_bytes:
-            return "b", k, f"双batch乒乓: 2*(MK+KN)*dtype={2*single/1024:.0f}KB <= L1"
+            return "b", k, f"双batch乒乓: 2*(MK+KN)*dtype={2*single/1024:.0f}KB <= L1", None
 
         # c) 一侧驻留 + 对侧切 K, 预算按 b_core 分档
+        # dValue 守卫与条件 4 / constraints 同源 (issue#19): 转置感知连续维判据
+        # (dv_a=M*dt 当 A 转置, 否则 k_l1*dt; dv_b=N*dt 当 B 不转置, 否则 k_l1*dt).
+        from ..models import dvalue_contig_dims
         l1_budget = s.l1_bytes / min(b_core, 2)
         for resident, side in ((m * k * dt, "A"), (k * n * dt, "B")):
             other = n * dt if side == "A" else m * dt
             if resident <= l1_budget:
                 k_l1 = min(int((l1_budget - resident) / other / 2), k)
                 k_l1 = align_down(max(k_l1, s.fractal), s.fractal)
-                if k_l1 >= s.fractal and k_l1 * dt >= s.dvalue_min:
+                dv_a, dv_b = dvalue_contig_dims(case, k_l1)
+                # 双缓冲适配: 驻留侧 + 对侧 k_l1 段 x2 必须 <= 预算
+                # (resident 恰占满预算时 k_l1 被抬到 16 会造成超预算的假方案)
+                if (k_l1 >= s.fractal and
+                        resident + 2 * k_l1 * other <= l1_budget and
+                        dv_a >= s.dvalue_min and dv_b >= s.dvalue_min):
                     return "c", k_l1, (
                         f"一侧驻留({side})+对侧切K: {side}驻留{resident/1024:.0f}KB, "
-                        f"预算L1/{min(b_core,2)}, k_L1={k_l1}")
+                        f"预算L1/{min(b_core,2)}, k_L1={k_l1}, "
+                        f"dValueA={dv_a:.0f}B/dValueB={dv_b:.0f}B"), side
                     # b_core>=2 时另一半 L1 预取下一 batch 驻留侧, 边界无气泡
 
         # d) 两侧都切 K (兜底)
         k_l1 = align_down(int(s.l1_bytes / (2 * (m + n) * dt)), s.fractal)
-        if k_l1 >= s.fractal and k_l1 * dt >= s.dvalue_min:
+        dv_a, dv_b = dvalue_contig_dims(case, k_l1)
-            return "d", k_l1, f"两侧都切K: k_L1={k_l1}, K段成对流水, batch边界天然无缝"
+        if (k_l1 >= s.fractal and dv_a >= s.dvalue_min and dv_b >= s.dvalue_min):
+            return "d", k_l1, (f"两侧都切K: k_L1={k_l1}, K段成对流水, batch边界天然无缝; "
+                               f"dValueA={dv_a:.0f}B/dValueB={dv_b:.0f}B"), None
 
-        return None, 0, "L1 四形态均不满足 (M/N 相对 L1 过大)"
+        return None, 0, "L1 四形态均不满足 (M/N 相对 L1 过大)", None
 
     # ------------------------------------------------------------------
     # 进入条件 (v0.98 §六, 四条同时满足)
@@ -84,21 +118,28 @@ class IterBatchBranch(Branch):
             "2_负载均衡: B mod C == 0 或 >= minCoreNum",
             c2, f"B mod C={rem}, minCoreNum={s.min_core_num}"))
 
-        form, k_l1, form_desc = self.l1_form(case)
+        form, k_l1, form_desc, _resident = self.l1_form(case)
         checks.append(ConditionCheck(
             "3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读)",
             form is not None, form_desc))
 
         # 条件 4: 搬移效率下限 (c/d 切分后)
+        # dValue 判定按转置调整连续维 (与 l1_form 守卫/constraints 同源, issue#19):
+        # A 不转置: K 向连续 -> k_l1*dt; A 转置: M 向连续 -> M*dt;
+        # B 不转置: N 向连续 -> N*dt;    B 转置: K 向连续 -> k_l1*dt.
+        from ..models import dvalue_contig_dims
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         if form in ("c", "d"):
             tile_ok = (k_l1 * m * dt >= s.min_tile_size) or (k_l1 * n * dt >= s.min_tile_size)
-            dv_ok = k_l1 * dt >= s.dvalue_min
+            dv_a, dv_b = dvalue_contig_dims(case, k_l1)
+            # 切 K 时 A/B 两侧分段各自搬移, 两侧连续维 dValue 均须 >= 下限 (与生成守卫/约束同源,
+            # issue#19; 非转置时两侧同为 k_l1*dt, 等价于历史口径)
+            dv_ok = dv_a >= s.dvalue_min and dv_b >= s.dvalue_min
             c4 = tile_ok and dv_ok
             checks.append(ConditionCheck(
-                "4_搬移效率: 搬移分块>=min_TileSize 且 dValue>=128B",
+                "4_搬移效率: 搬移分块>=min_TileSize 且 dValue>=128B (转置调整连续维)",
-                c4, f"tile={max(k_l1*m*dt, k_l1*n*dt)/1024:.1f}KB, dValue={k_l1*dt}B"))
+                c4, f"tile={max(k_l1*m*dt, k_l1*n*dt)/1024:.1f}KB, dValueA={dv_a}B, dValueB={dv_b}B"))
         else:
             checks.append(ConditionCheck(
                 "4_搬移效率(a/b形态不切分, 恒满足)", True, ""))
@@ -113,7 +154,7 @@ class IterBatchBranch(Branch):
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         b_core = -(-case.batch_c // s.aic_num)
-        form, k_l1, form_desc = self.l1_form(case)
+        form, k_l1, form_desc, _resident = self.l1_form(case)
         form = form or "d"
 
         # L0 级 tile (BaseM x BaseN): 核内 L1->L0 的切分, 与核间切分无关.
@@ -153,6 +194,12 @@ class IterBatchBranch(Branch):
         form_name = {"a": "a_单batch全驻留", "b": "b_双batch乒乓",
                      "c": "c_一侧驻留+对侧切K", "d": "d_两侧都切K"}[form]
 
+        # 输出落点 (issue#30, R4): 整 case 输入+输出 <= L2 才驻留 L2
+        out_l2 = output_to_l2(case, s)
+        l2_out = ("resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)"
+                  if out_l2 else
+                  "direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)")
+
         return ImplPlan(
             case_id=case.case_id, branch=self.name,
             used_core_num=s.aic_num,
@@ -163,54 +210,68 @@ class IterBatchBranch(Branch):
             k_l1=k_l1, b_l1=2 if form == "b" else 1, l1_form=form_name,
             base_m=base_m, base_n=base_n, base_k=base_k,
             l2_policy_in="allocate(GM->L1随路驻留L2)",
-            l2_policy_out="direct_gm(输出仅写一次,直写GM不占L2)",
+            l2_policy_out=l2_out,
             swizzle_w=0, workspace_bytes=0,
             tail_strategy="不涉及(核内不切M/N)",
             fixpipe_unitflag=True,
             out_dtype_bytes=case.dtype_out_bytes,
-            note=form_desc,
+            note=form_desc + f"; 输出落点: {'L2驻留' if out_l2 else '直写GM'}",
         )
 
     # ------------------------------------------------------------------
     # 时延评估 (v1.1 §4 端到端模型, IterBatch 侧)
+    #   口径: dtype 感知算力 (issue#28); GM 每字节恰读一次 = V_in (issue#31);
+    #         字节列整芯片 (issue#29); 输出落点 R4 (issue#30).
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
         s = self.spec
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
+        b = case.batch_c
         b_core, k_l1 = plan.b_core, plan.k_l1
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
+        out_l2 = output_to_l2(case, s)
+        w_fix = s.bw_l2_pc if out_l2 else s.bw_pc
 
         k_truncated = k_l1 >= k
         n_k = 1 if k_truncated else -(-k // k_l1)
 
-        t_load = min(k_l1, k) * (m + n) * dt / s.bw_pc
+        t_comp_chunk = 2.0 * m * n * min(k_l1, k) / qc
-        t_comp_chunk = 2.0 * m * n * min(k_l1, k) / s.q16
+        t_write = m * n * out_b / w_fix
-        t_write = m * n * out_b / s.bw_pc
 
-        # 搬移: 每 batch n_K 次 GM->L1, 每次含 T_cmd
+        # 搬移 (issue#31): 每 batch 的 K 段/驻留侧数据互不重叠, 每个输入字节恰好从
+        # GM 读一次 (形态 c 驻留侧每 batch 只搬一次; 切 K 末段按实际剩余计) ->
+        # GM 数据量 = V_in, 数据时延 = V_in/芯片带宽 (全核并发);
+        # n_K/k_L1 只决定 DMA 命令次数 (T_cmd) 与双缓冲调度, 不放大数据量.
+        # dma_cmds 为单核命令数 (各核并行, issue#29 口径)
         dma_cmds = b_core * n_k
-        t_mte2_data = dma_cmds * t_load
+        # 搬移效率 (issue#36): 单命令 tile = nValue*dValue*dt 决定有效带宽
+        # (达 min_TileSize 饱和, 之下线性退化), 分侧口径见 move_tiles
+        tile_a, tile_b = self.move_tiles(case)
+        t_mte2_data = gm_move_time(b * m * k * dt, b * k * n * dt,
+                                   tile_a, tile_b, s)
         t_dma_cmd = dma_cmds * s.t_cmd
         t_mte2 = t_mte2_data + t_dma_cmd
 
-        # Cube: 无冗余
+        # Cube: 无冗余; 每核 flops = b_core*2MNK (整芯片列 = b*2MNK)
         flops_pc = b_core * 2.0 * m * n * k
-        t_mmad = flops_pc / s.q16
+        flops_chip = b * 2.0 * m * n * k
+        t_mmad = flops_pc / qc
 
-        # Fixpipe: b_core 个 batch 输出, 按 C dtype
+        # Fixpipe (R4): b_core 个 batch 输出, 按 C dtype
         fix_bytes_pc = b_core * m * n * out_b
-        t_fix = fix_bytes_pc / s.bw_pc
+        fix_bytes_chip = b * m * n * out_b
+        t_fix = fix_bytes_pc / w_fix
 
         # drain: 末 batch 排空 = T_comp + T_write
         t_drain = t_comp_chunk + t_write
 
-        gm_bytes_pc = b_core * (m * k + k * n) * dt
-
         return assemble_timing(
             t_mte2_gm=t_mte2_data, t_mte2_l2=0.0, t_dma_cmd=t_dma_cmd,
             t_mmad=t_mmad, t_fixpipe=t_fix, t_reduce=0.0, t_drain=t_drain,
-            gm_read_bytes=gm_bytes_pc, l2_read_bytes=0.0,
+            gm_read_bytes=case.input_bytes, l2_read_bytes=0.0,
-            dma_cmd_count=dma_cmds, cube_flops=flops_pc,
+            dma_cmd_count=dma_cmds, cube_flops=flops_chip,
-            fixpipe_bytes=fix_bytes_pc,
+            fixpipe_bytes=fix_bytes_chip,
+            fixpipe_to_gm=(not out_l2),
         )

BMM/BMM_Theory/bmm_theory/branches/merge_batch.py

@@ -4,9 +4,13 @@
   - 《BMM算子优化分析 v0.98》§五 (进入条件 + 实现方案 Step1~3)
   - 《MergeBatch_vs_IterBatch分析 v1.1》§三/§四 (执行模型 + 分界条件)
 
-核心思想: 合并 b0 个 batch 的 A'[b0*M,K] @ B'[K,b0*N] 为单次 DMA 搬入,
+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),
+合并的核心收益有二 (用户澄清, issue#36): (1) 搬移命令数减少 (省 T_cmd);
-进入条件 5 保证 case 为访存 Bound, 冗余算力被搬移时延掩盖.
+(2) **单块 tile = nValue*dValue*dt 放大 b0 倍 -> 搬移效率提升** (堆叠方向视
+转置/排布: A ND 非转置沿 M(nValue) 堆叠, B ND 非转置沿 N(dValue) 堆叠) ——
+即便 T_cmd=0 (950PR 默认, 未标定按 0) 收益依然成立, 由 move_eff 效率模型刻画。
+核心劣势是交叉项冗余算力 ((b0-1)/b0 被算出但丢弃) 与 drain 暴露放大 b0 倍;
+进入条件 5 保证 case 为访存 Bound, 冗余算力被搬移时延掩盖。
 """
 
 from __future__ import annotations
@@ -14,8 +18,8 @@ from __future__ import annotations
 import math
 
 from ..hardware import NpuSpec, ASCEND950PR
-from ..models import BmmCase, ImplPlan, HardwareTiming, align_down
+from ..models import BmmCase, ImplPlan, HardwareTiming, align_down, gm_move_time
-from ..timing import MoveInPlan, assemble_timing
+from ..timing import assemble_timing, output_to_l2
 from .base import Branch, BranchResult, ConditionCheck
 
 MIN_B0 = 2  # b0: 合并搬移有收益的最小合并数
@@ -82,6 +86,20 @@ class MergeBatchBranch(Branch):
             "5_访存Bound: 2MN/(M+N) < R16/b0",
             c5, f"AI={ai:.1f} vs R16/b0={s.r16/MIN_B0:.1f}"))
 
+        # 条件 6: 转置对齐 (参考 bmmv3 源码条件 17/20)
+        # A 转置且 M>1 时, tempAlignM = b0 * alignM (M 维按 b0 对齐)
+        # A 转置或 B 不转置时, minBaseK 需按 basic_block_size 对齐
+        from ..models import align_up
+        if case.trans_a and m > 1:
+            temp_align_m = MIN_B0 * align_up(m, s.fractal)
+            l0a_need_trans = temp_align_m * s.fractal * dt * 2
+            c6 = l0a_need_trans <= s.l0a_bytes
+            checks.append(ConditionCheck(
+                "6_A转置对齐: A转置时 tempAlignM=b0*alignM 需驻留 L0A",
+                c6, f"tempAlignM={temp_align_m}, L0A需{l0a_need_trans/1024:.0f}KB(≤{s.l0a_bytes/1024:.0f})"))
+        else:
+            checks.append(ConditionCheck("6_A转置对齐", True, "A不转置或M=1, 无额外对齐要求"))
+
         return checks
 
     # ------------------------------------------------------------------
@@ -125,9 +143,17 @@ class MergeBatchBranch(Branch):
         ))
         b_l1 = max(b_l1, b0)
 
+        # Step 5: 输出落点 (issue#30): 整 case 输入+输出可驻留 L2 才写 L2 (R4)
+        out_l2 = output_to_l2(case, s)
+        l2_out = ("resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)"
+                  if out_l2 else
+                  "direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)")
+
         note = (f"b0={b0} (L0C上限{b0_l0c:.1f}/算存比上限{b0_ai:.1f}/b_core={b_core}); "
                 f"{'K截断' if k_truncated else 'L1绑定'}; "
-                f"合并后单次DMA搬入 A'[{b0*m},{k_l1}]+B'[{k_l1},{b0*n}]")
+                f"合并后单次DMA搬入 A'[{b0*m},{k_l1}]+B'[{k_l1},{b0*n}]; "
+                f"输出落点: {'L2驻留' if out_l2 else '直写GM'} (整case V_in+V_out"
+                f"={case.input_bytes/1048576:.1f}MB vs L2={s.l2_bytes/1048576:.0f}MB)")
 
         return ImplPlan(
             case_id=case.case_id, branch=self.name,
@@ -140,7 +166,7 @@ class MergeBatchBranch(Branch):
             # L0C 双缓冲约束已由进入条件 2 保证 (2*(b0*M)*(b0*N)*4B <= L0C)
             base_m=b0 * m, base_n=b0 * n, base_k=max(min(k_l0, k_l1, k), s.fractal),
             l2_policy_in="allocate(GM->L1随路驻留L2)",
-            l2_policy_out="direct_gm(输出仅写一次,直写GM不占L2)" if not case.out_nd else "direct_gm",
+            l2_policy_out=l2_out,
             swizzle_w=0, workspace_bytes=0,
             tail_strategy="不涉及(核内不切M/N)",
             fixpipe_unitflag=True,
@@ -149,55 +175,72 @@ class MergeBatchBranch(Branch):
         )
 
     # ------------------------------------------------------------------
-    # 时延评估 (v1.1 §4 端到端模型)
+    # 时延评估 (v1.1 §4 端到端模型; issue#27/#28/#29/#30 口径)
+    #   - Cube flops (issue#27 复核): 每合并步计算全网格 (b0*M)x(b0*N)xK
+    #     含交叉项冗余, flops_step = 2*(b0M)*(b0N)*K, 每核步数 = b_core/b0
+    #     -> 每核 flops = b_core*b0*2MNK (与 CSV 及文档公式一致, 无修改);
+    #   - 算力按输入 dtype (issue#28): q_cube(dtype_a, dtype_b);
+    #   - 字节列 = 整芯片口径 (issue#29); GM 首读 = V_in 一次 (合并按组搬入
+    #     各 (batch,K段) 数据互不重叠, 无 L2 重复读);
+    #   - 输出落点按整 case 驻留判定 (issue#30, R4).
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
         s = self.spec
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
+        b = case.batch_c
         b_core, b0, k_l1 = plan.b_core, plan.merge_b0, plan.k_l1
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
+        out_l2 = output_to_l2(case, s, 0.0)
+        w_fix = s.bw_l2_pc if out_l2 else s.bw_pc
 
-        k_truncated = k_l1 >= k
+        k_truncated = k_l1 >= k   # 保留语义标记 (plan.note/调试用)
-        # 每 K 分块搬移/计算时延 (未合并基准, v1.1 §4.1 符号)
+        # 每 K 分块计算时延 (未合并基准, v1.1 §4.1 符号) / 单 batch 输出写回 (R4)
-        t_load = k_l1 * (m + n) * dt / s.bw_pc
+        t_comp_chunk = 2.0 * m * n * k_l1 / qc
-        t_comp_chunk = 2.0 * m * n * k_l1 / s.q16
+        t_write = m * n * out_b / w_fix
-        t_write = m * n * out_b / s.bw_pc     # 单 batch 输出写回 (单核带宽份额)
-
-        if k_truncated:
-            # K 截断: n_K=1, 合并后单次搬移量 b0 倍
-            n_move = b_core / b0
-            t_mte2_data = n_move * b0 * t_load
-            dma_cmds = n_move
-        else:
-            # L1 绑定: k_L1^m = k_L1/b0, n_K^m = b0*n_K, 搬移次数与 IterBatch 相同
-            n_k = -(-k // k_l1)
-            n_move = b_core * n_k
-            t_mte2_data = n_move * t_load
-            dma_cmds = n_move
 
+        # 搬移 (issue#31): 合并组/切 K 各 (组, K段) 数据互不重叠, 每个输入字节恰好
+        # 从 GM 读一次 (K截断与 L1 绑定均如此; 切 K 末段按实际剩余计, 不再整段上取)
+        # -> GM 数据量 = V_in, 数据时延 = V_in/芯片带宽 (全核并发);
+        # 搬移命令数只决定 T_cmd (issue#35 修正: 真实命令数 = 合并组数 x 每组 K 段数,
+        # 不得把"命令数与 IterBatch 相同 = b_core*n_K"(v1.1 §4.4, n_K 为未合并粒度)
+        # 误代入合并后段数 —— 那样会多计 b0 倍):
+        #   K截断:  n_K^m = 1 -> cmds = ceil(b_core/b0) (比 IterBatch 省 b0 倍)
+        #   L1绑定: k_L1^m = k_L1/b0 理想情形退化为 b_core*n_K, 与 IterBatch 相同;
+        #           dValue 512B cap 截断等情形按实际 ceil(K/k_l1^m) 计
+        n_k = -(-k // k_l1)
+        dma_cmds = -(-b_core // b0) * n_k
         t_dma_cmd = dma_cmds * s.t_cmd
+        # 搬移效率 (issue#36, 用户澄清): 合并 b0 个 batch 使单命令 tile 放大 b0 倍
+        # (A': b0*M x k_l1^m, B': k_l1^m x b0*N; 堆叠方向视转置/排布, 乘积不变),
+        # 小 M/N case 下相对 IterBatch 逐 batch 搬移的效率收益显著, T_cmd=0 时仍成立
+        k_eff = min(k_l1, k)
+        t_mte2_data = gm_move_time(b * m * k * dt, b * k * n * dt,
+                                   b0 * m * k_eff * dt, b0 * n * k_eff * dt, s)
         t_mte2 = t_mte2_data + t_dma_cmd
 
-        # Cube: 合并计算含冗余 (b0^2 输出, 有效 b0) -> 计算量 = b_core*b0*2MNK
+        # Cube (每核口径): 合并计算含冗余 (b0^2 输出, 有效 b0), 每核 b_core/b0 步,
+        # 每步 flops = 2*(b0*M)*(b0*N)*K  -> 每核 = b_core*b0*2MNK (issue#27 复核一致)
         flops_pc = b_core * b0 * 2.0 * m * n * k
-        t_mmad = flops_pc / s.q16
+        flops_chip = b * b0 * 2.0 * m * n * k        # 整芯片口径列
+        t_mmad = flops_pc / qc
 
-        # Fixpipe: 只写对角块, 写出量 = b_core*MN*outB (C 矩阵 dtype, 随路转换)
+        # Fixpipe (R4): 只写对角块, 写出量 = b_core*MN*outB (C 矩阵 dtype, 随路转换)
         fix_bytes_pc = b_core * m * n * out_b
-        t_fix = fix_bytes_pc / s.bw_pc
+        fix_bytes_chip = b * m * n * out_b
+        t_fix = fix_bytes_pc / w_fix
 
         # drain: 末合并 batch 排空 = b0*(T_comp + T_write)
         t_drain = b0 * (t_comp_chunk + t_write)
 
-        gm_bytes_pc = b_core * (m * k + k * n) * dt
-
         return assemble_timing(
             t_mte2_gm=t_mte2_data, t_mte2_l2=0.0, t_dma_cmd=t_dma_cmd,
             t_mmad=t_mmad, t_fixpipe=t_fix, t_reduce=0.0, t_drain=t_drain,
-            gm_read_bytes=gm_bytes_pc, l2_read_bytes=0.0,
+            gm_read_bytes=case.input_bytes, l2_read_bytes=0.0,
-            dma_cmd_count=dma_cmds, cube_flops=flops_pc,
+            dma_cmd_count=dma_cmds, cube_flops=flops_chip,
-            fixpipe_bytes=fix_bytes_pc,
+            fixpipe_bytes=fix_bytes_chip,
+            fixpipe_to_gm=(not out_l2),
         )
 
     # ------------------------------------------------------------------
@@ -206,8 +249,29 @@ class MergeBatchBranch(Branch):
     def beats_iterbatch(self, case: BmmCase) -> tuple:
         """返回 (MergeBatch是否更优, 说明).
 
-        MergeBatch 最优  ⟺  K截断 (k_L1=K) 且 b_core > b0*(T_comp+T_write)/T_cmd
+        v1.1 §4.5 统一分界的泛化口径 (issue#35 命令数 / issue#36 搬移效率):
-        L1 绑定情形 MergeBatch 恒劣于 IterBatch (搬移次数相同, 只放大 drain).
+            cmds_iter = b_core * ⌈K/k_l1_iter⌉    (IterBatch 逐 batch 逐 K 段一条)
+            cmds_mb   = ⌈b_core/b0⌉ * ⌈K/k_l1^m⌉  (合并组数 x 每组 K 段数)
+            命令节省  = (cmds_iter - cmds_mb) * T_cmd
+            效率节省  = t_data_iter - t_data_mb  (合并放大单块 tile b0 倍 ->
+                      搬移效率提升, issue#36 用户澄清: 即便 T_cmd=0 也有收益)
+            drain 惩罚 = (b0-1) * (T_comp + T_write)   (T_comp 按未合并基线分块)
+        MergeBatch 最优 ⟺ 命令节省 + 效率节省 > drain 惩罚.
+
+        与 v1.1 §4.5 闭式的等价性 (T_cmd>0 且效率打平时):
+        - K 截断 (k_l1^m = K): cmds_mb = b_core/b0, 退化为文档闭式
+          b_core > b0*(T_comp+T_write)/T_cmd;
+        - L1 绑定理想情形 (k_l1^m = k_l1^iter/b0): 命令数相同, 命令节省=0;
+        - dValue 512B 推荐值截断等第三情形 (文档二分未覆盖): 按实际命令数比较.
+
+        截断判定用**合并后** plan.k_l1 >= K (与 make_plan/evaluate 同源,
+        issue#35): 未合并 k_l1 截断不代表合并后仍截断 (合并使 L1 占用放大 b0 倍,
+        且受 dValue 512B 推荐值截断), v1.1 line 219 的字面定义 (未合并口径) 与
+        line 161 的合并公式矛盾时以后者为准.
+
+        T_cmd=0 (950PR 默认, 未标定): 命令节省项为 0, 由效率节省与 drain 惩罚
+        决定 —— 合并 tile 放大带来的效率收益被显式建模, 不再需要"T_cmd<=0 时
+        K截断即优先"的策略覆盖 (issue#36 退役)。
         """
         s = self.spec
         m, n, k = case.m, case.n, case.k
@@ -215,21 +279,49 @@ class MergeBatchBranch(Branch):
         out_b = case.dtype_out_bytes
         b_core = case.batch_c // s.aic_num
 
-        # IterBatch 基准的 k_L1 (未合并): L1 双缓冲单 batch
+        # IterBatch 基线: 每核命令数 + 单命令 tile (与 iter_batch.evaluate 同源)
-        k_l1_iter = min(k, s.l1_bytes / (2 * (m + n) * dt))
+        from .iter_batch import IterBatchBranch
-        k_truncated = k_l1_iter >= k
+        ib = IterBatchBranch(s)
+        _, k_l1_iter, _, _ = ib.l1_form(case)
+        if not k_l1_iter:
+            k_l1_iter = k
+        n_k_iter = -(-k // min(k_l1_iter, k))
+        cmds_iter = b_core * n_k_iter
+        tile_a_ib, tile_b_ib = ib.move_tiles(case)
+        bb = case.batch_c
+        t_data_ib = gm_move_time(bb * m * k * dt, bb * k * n * dt,
+                                 tile_a_ib, tile_b_ib, s)  # 整芯片口径 (量纲一致)
 
+        # MergeBatch 实际每核命令数 (issue#35) + 合并 tile (issue#36)
         plan = self.make_plan(case)
         b0 = plan.merge_b0
-        t_comp = 2.0 * m * n * min(k_l1_iter, k) / s.q16
+        n_k_mb = -(-k // min(plan.k_l1, k))
-        t_write = m * n * out_b / s.bw_pc
+        cmds_mb = -(-b_core // b0) * n_k_mb
-        threshold = b0 * (t_comp + t_write) / s.t_cmd
+        k_truncated = plan.k_l1 >= k   # 合并口径截断判定 (issue#35)
+        regime = f"K截断(k_l1^m={plan.k_l1}>=K)" if k_truncated else \
+                 f"L1绑定(k_l1^m={plan.k_l1}<K={k})"
+        k_eff = min(plan.k_l1, k)
+        t_data_mb = gm_move_time(bb * m * k * dt, bb * k * n * dt,
+                                 b0 * m * k_eff * dt, b0 * n * k_eff * dt, s)
 
-        win = k_truncated and (b_core > threshold)
+        # T_comp 按输入 dtype 算力 (issue#28), 未合并基线分块 (v1.1 §4.1 符号);
-        detail = (f"k_L1={'K(截断)' if k_truncated else f'{k_l1_iter:.0f}<K(L1绑定)'}; "
+        # T_write 保持 v1.1 直写 GM 语义
-                  f"b_core={b_core} vs 阈值 b0*(T_comp+T_write)/T_cmd={threshold:.1f}; "
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
-                  f"drain惩罚=(b0-1)*(T_comp+T_write)={((b0-1)*(t_comp+t_write))*1e6:.2f}us, "
+        t_comp = 2.0 * m * n * min(k_l1_iter, k) / qc
-                  f"搬移节省=b_core*(1-1/b0)*T_cmd={b_core*(1-1/b0)*s.t_cmd*1e6:.2f}us")
+        t_write = m * n * out_b / s.bw_pc
+        drain_pen = (b0 - 1) * (t_comp + t_write)
+        save_cmds = cmds_iter - cmds_mb
+        savings_cmd = save_cmds * s.t_cmd
+        savings_eff = t_data_ib - t_data_mb
+
+        win = (savings_cmd + savings_eff) > drain_pen
+        detail = (f"{regime}; 每核命令数 MergeBatch={cmds_mb} vs IterBatch={cmds_iter} "
+                  f"(命令节省={savings_cmd*1e6:.2f}us) + 搬移效率节省="
+                  f"{savings_eff*1e6:.2f}us vs drain惩罚=(b0-1)*(T_comp+T_write)="
+                  f"{drain_pen*1e6:.2f}us -> {'MergeBatch优' if win else 'IterBatch优'}")
+        if k_truncated and s.t_cmd > 0:
+            detail += (f" (闭式等价: b_core={b_core} vs 阈值 "
+                       f"b0*(T_comp+T_write)/T_cmd={b0 * (t_comp + t_write) / s.t_cmd:.1f})")
         return win, detail
 
     # ------------------------------------------------------------------

BMM/BMM_Theory/bmm_theory/branches/special.py

@@ -10,7 +10,7 @@ from __future__ import annotations
 
 from ..hardware import NpuSpec, ASCEND950PR
 from ..models import BmmCase, ImplPlan, HardwareTiming
-from ..timing import assemble_timing
+from ..timing import assemble_timing, output_to_l2
 from .base import Branch, ConditionCheck
 
 
@@ -24,7 +24,7 @@ class SpecialBranch(Branch):
         c1 = case.k <= 1
         checks = [ConditionCheck("1_K<=1 (Cube 无用)", c1, f"K={case.k}")]
         if case.k == 1:
-            # K=1 的 AIV 通路恒可用 (issue#12): B>=2*AIV 开 UB 乒乓; B<128 退化为
+            # K=1 的 AIV 通路恒可用 (issue#12/#17): B>=2*AIV 开 UB 乒乓; B<128 退化为
             # AIV 单缓冲 (无乒乓, 逐 batch 串行搬入), 不再是无方案空洞.
             b = case.batch_c
             pingpong = b >= 2 * self.spec.aiv_num
@@ -68,32 +68,42 @@ class SpecialBranch(Branch):
 
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
-        """AIV 通路时延: 瓶颈在搬移 (AIV 算力远剩)."""
+        """AIV 通路时延: 瓶颈在搬移 (AIV 算力远剩).
+
+        口径: AIV 逐元素通量按输入 dtype (issue#28); 输出落点 R4 (issue#30):
+        整 case 输入+输出 <= L2 -> 输出写 L2 (异步回写不占算子时延), 否则直写
+        GM 与读共享总线 (assemble 的 MTE2 链累加, issue#23).
+        """
         s = self.spec
         b = case.batch_c
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
+        out_l2 = output_to_l2(case, s)
+        w_fix = s.bw_l2 if out_l2 else s.bw_gm
+        cube_flops = 0.0
+        t_compute = 0.0
 
         if k == 0:
-            # 纯写值: 仅写出
+            # 纯写值: 仅写出 (R1 下 GM 读 = 0, 输入本身为空)
-            t_in, t_compute, t_out = 0.0, 0.0, b * m * n * out_b / s.bw_gm
+            t_in = 0.0
-            in_bytes, cube_flops = 0.0, 0.0
+            t_out = b * m * n * out_b / w_fix
+            in_bytes = 0.0
         else:
-            # 逐元素乘: 搬入 A+B, 搬出 C, AIV 算力远剩
+            # 逐元素乘: 搬入 A+B (GM 每字节一次), 搬出 C
             in_bytes = b * (m * k + k * n) * dt
             out_bytes = b * m * n * out_b
             t_in = in_bytes / s.bw_gm
-            t_out = out_bytes / s.bw_gm
+            t_out = out_bytes / w_fix
-            # AIV 求积吞吐 (近似按 Q_AIV)
+            # AIV 逐元素吞吐按输入 dtype 取通量 (issue#28: bf16/fp16 x2, int8 x4...)
-            t_compute = b * m * n / s.q_aiv
+            t_compute = b * m * n / s.aiv_elem_rate(case.dtype_in)
             cube_flops = float(b * m * n)
 
-        t_total = max(t_in, t_out, t_compute)
         return assemble_timing(
             t_mte2_gm=t_in, t_mte2_l2=0.0, t_dma_cmd=0.0,
             t_mmad=t_compute, t_fixpipe=t_out, t_reduce=0.0, t_drain=0.0,
             gm_read_bytes=in_bytes, l2_read_bytes=0.0, dma_cmd_count=0.0,
             cube_flops=cube_flops,
             fixpipe_bytes=b * m * n * out_b,
+            fixpipe_to_gm=(not out_l2),
         )

BMM/BMM_Theory/bmm_theory/branches/stream_k.py

@@ -12,7 +12,7 @@ import math
 
 from ..hardware import NpuSpec, ASCEND950PR
 from ..models import BmmCase, ImplPlan, HardwareTiming, ceil_div
-from ..timing import assemble_timing, eval_streamk_reduce
+from ..timing import assemble_timing, eval_streamk_reduce, output_to_l2
 from .base import Branch, ConditionCheck
 
 
@@ -32,10 +32,14 @@ class StreamKBranch(Branch):
         p = self._p_value(case)
         return int(self.spec.aic_num // max(1, math.ceil(p)))
 
-    def _theta_c(self) -> float:
+    def _theta_c(self, case: BmmCase) -> float:
-        """归约代价系数 theta_c = Q16/2 * (8B/W_L2 + 1/Q_AIV) ≈ 12."""
+        """归约代价系数 theta_c = Q_cube(dtype)/2 * (8B/W_L2 + 1/Q_AIV_fp32).
+
+        Q_AIV 用 fp32 档 (部分和为 fp32, issue#28); Cube 侧按输入 dtype.
+        """
         s = self.spec
-        return s.q16 / 2 * (8 / s.bw_l2 + 1 / s.q_aiv)
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
+        return qc / 2 * (8 / s.bw_l2 + 1 / s.q_aiv)
 
     # ------------------------------------------------------------------
     def check_conditions(self, case: BmmCase) -> list:
@@ -61,7 +65,7 @@ class StreamKBranch(Branch):
 
         # 条件 3: K > grid_K^2/(grid_K-1) * theta_c (归约代价可接受)
         if grid_k >= 2:
-            theta_c = self._theta_c()
+            theta_c = self._theta_c(case)
             k_thresh = grid_k * grid_k / (grid_k - 1) * theta_c
             c3 = case.k > k_thresh
             checks.append(ConditionCheck(
@@ -129,39 +133,64 @@ class StreamKBranch(Branch):
 
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
+        """芯片口径评估, 按 plan 实际 tile 布局 (issue#25).
+
+        切 K 组 (tile = single_m x single_n) 由 grid_k 核协作: 组内 K 段零重复读
+        (GM 一次); 组间共享 (A 行块被 n_cnt 个 tile 用) 在单 batch 工作集可驻留时
+        走 L2 读口 (与 ASW 同规则). 全部 tile-组在 C 核上分波执行:
+          waves = ceil(b*m_cnt*n_cnt*grid_k / C).
+        - GM 读: 首读总量 = b*(m*k + k*n)*dt (组间共享命中 L2 时), 或按 r_gm 放大
+          (工作集超 L2, 保守同 ASW 场景 C 公式);
+        - MMAD: 芯片总量 2*b*m*n*k / (C核) —— 全核忙稳态;
+        - 归约 (部分和 4B 写 L2/AIV 读回求和/最终写回): 每 tile-组一次、组间串行
+          追加为 drain: t_drain = waves * eval_streamk_reduce(tile, grid_k, out);
+          最终输出写回落点按 R4 (issue#30): 整 case 输入+输出+workspace <= L2
+          时写 L2, 否则直写 GM (drain 内按 GM 带宽).
+        """
         s = self.spec
         b = case.batch_c
         m, n, k = case.m, case.n, case.k
         dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
-        grid_k = plan.grid_k
+        grid_k = max(plan.grid_k, 1)
+        m_cnt = max(plan.m_cnt, 1)
+        n_cnt = max(plan.n_cnt, 1)
+        tile_m = max(plan.single_core_m, 1)
+        tile_n = max(plan.single_core_n, 1)
+        tile = tile_m * tile_n                       # 每核实际 tile 元素数
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
+        out_l2 = output_to_l2(case, s, float(plan.workspace_bytes))
 
-        # 单 tile (L0C 满载基本块)
+        # ---- MTE2: GM 段 + L2 段 (同 ASW 块级规则) ----
-        tile_elems = s.l0c_elems  # 65536
+        a_b = m * k * dt
-        t_mmad_tile = 2.0 * k * s.l0c_bytes / 4 / s.q16  # 2K/Q16 * L0C/4B
+        bb_b = k * n * dt
-        t_mte2_tile = k * (2 * math.sqrt(tile_elems)) * dt / s.bw_pc  # 近似方形 tile
+        if a_b + bb_b <= s.l2_bytes:
+            gm_read = b * (a_b + bb_b)               # 组间共享命中 L2: GM 每字节一次
+            l2_read = b * ((n_cnt - 1) * a_b + (m_cnt - 1) * bb_b)
+        else:
+            gm_read = b * (a_b + bb_b)               # 保守: 共享跨组回落 GM 未另计
+            l2_read = 0.0
+        t_gm = gm_read / s.bw_gm                     # 芯片口径 (全核并发)
+        t_l2 = l2_read / s.bw_l2
+        t_mte2 = t_gm + t_l2
 
-        # 切 K 后流水时延缩 grid_k 倍
+        # ---- MMAD: 芯片总量, dtype 感知算力 (issue#28) ----
-        t_mmad = t_mmad_tile / grid_k
+        flops = 2.0 * b * m * n * k
-        t_mte2 = t_mte2_tile / grid_k
+        t_mmad = flops / (s.aic_num * qc)
 
-        # 归约: 部分和 4B 驻留 L2, AIV 归约 (含部分和写/读回/求和/最终按 C dtype 写回)
+        # ---- 归约: 每 tile-组一次, 组间分波串行追加 ----
-        # 口径 (issue#11): 归约整体为串行追加 (t_drain=t_reduce, reduce_serial=True),
+        waves = ceil_div(b * m_cnt * n_cnt * grid_k, s.aic_num)
-        # 部分和写出已计入 eval_streamk_reduce 的 t_write_partial —— 稳态 Fixpipe 不再
+        t_reduce_group = eval_streamk_reduce(tile, grid_k, out_b, s,
-        # 重复计账. 此前按 grid_k*tile*4B/单核带宽份额另计一次, 既重复计账又把整组
+                                             out_to_gm=not out_l2)
-        # 部分和串行压到单核写口, 高估 grid_k 倍 (streamk_demo 曾虚高到 55us/FIXPIPE).
+        t_reduce = waves * t_reduce_group
-        t_reduce = eval_streamk_reduce(tile_elems, grid_k, out_b, s)
         fix_bytes = 0.0
         t_fix = 0.0
 
-        flops_pc = 2.0 * tile_elems * k / grid_k
-        gm_bytes = k * (2 * math.sqrt(tile_elems)) * dt / grid_k
-
         return assemble_timing(
-            t_mte2_gm=t_mte2, t_mte2_l2=0.0, t_dma_cmd=0.0,
+            t_mte2_gm=t_gm, t_mte2_l2=t_l2, t_dma_cmd=0.0,
             t_mmad=t_mmad, t_fixpipe=t_fix, t_reduce=t_reduce,
             t_drain=t_reduce,  # 归约串行追加 (reduce_serial 默认 True, 不进稳态 max)
-            gm_read_bytes=gm_bytes, l2_read_bytes=0.0, dma_cmd_count=0.0,
+            gm_read_bytes=gm_read, l2_read_bytes=l2_read, dma_cmd_count=0.0,
-            cube_flops=flops_pc, fixpipe_bytes=fix_bytes,
+            cube_flops=flops, fixpipe_bytes=fix_bytes,
-            reduce_serial=True,
+            reduce_serial=True, fixpipe_to_gm=False,
         )

BMM/BMM_Theory/bmm_theory/branches/to_matmul.py

@@ -12,7 +12,7 @@ from __future__ import annotations
 
 from ..hardware import NpuSpec, ASCEND950PR
 from ..models import BmmCase, ImplPlan, HardwareTiming
-from ..timing import assemble_timing
+from ..timing import assemble_timing, output_to_l2
 from .base import Branch, ConditionCheck
 
 
@@ -71,7 +71,12 @@ class ToMatmulBranch(Branch):
 
     # ------------------------------------------------------------------
     def evaluate(self, case: BmmCase, plan: ImplPlan) -> HardwareTiming:
-        """折叠后按 Matmul 粗估 (详细切分待 MM 理论体系打通后接入)."""
+        """折叠后按 Matmul 粗估 (详细切分待 MM 理论体系打通后接入).
+
+        口径: Cube 算力按输入 dtype (issue#28); GM 首读 = V_in 一次 (R1);
+        输出落点 R4 (issue#30): 整 case 输入+输出 <= L2 时写 L2 写口, 否则
+        直写 GM 与读共享总线 (assemble 累加).
+        """
         s = self.spec
         if case.batch_b == 1:
             fold_m, fold_n, kk = case.batch_a * case.m, case.n, case.k
@@ -79,13 +84,15 @@ class ToMatmulBranch(Branch):
             fold_m, fold_n, kk = case.m, case.batch_b * case.n, case.k
         dt = case.dtype_in_bytes
         out_b = case.dtype_out_bytes
+        qc = s.q_cube(case.dtype_a, case.dtype_b)
+        out_l2 = output_to_l2(case, s)
 
         flops = 2.0 * fold_m * fold_n * kk
-        t_mmad = flops / (s.aic_num * s.q16)
+        t_mmad = flops / (s.aic_num * qc)
         in_bytes = (fold_m * kk + kk * fold_n) * dt
         out_bytes = fold_m * fold_n * out_b
         t_mte2 = in_bytes / s.bw_gm
-        t_fix = out_bytes / s.bw_gm
+        t_fix = out_bytes / (s.bw_l2 if out_l2 else s.bw_gm)
 
         return assemble_timing(
             t_mte2_gm=t_mte2, t_mte2_l2=0.0, t_dma_cmd=0.0,
@@ -93,4 +100,5 @@ class ToMatmulBranch(Branch):
             t_drain=0.0,
             gm_read_bytes=in_bytes, l2_read_bytes=0.0, dma_cmd_count=0.0,
             cube_flops=flops, fixpipe_bytes=out_bytes,
+            fixpipe_to_gm=(not out_l2),
         )

BMM/BMM_Theory/bmm_theory/constraints.py

@@ -36,6 +36,13 @@ def check_plan_constraints(case: BmmCase, plan: ImplPlan,
     s = spec
     v = []
 
+    # 占位/无效方案: used_core_num<1 说明没有真实方案 (如 router._no_plan 的占位),
+    # recommend 侧 advice 已标注"[无方案]", evaluate 侧必须判不可行 (issue#18),
+    # 不得当作可行方案给正常时延.
+    if plan.used_core_num < 1:
+        v.append("used_core_num=0: 占位/未生成方案, 不可评估")
+        return v
+
     # AIV 通路: 只校验 AIV 核数
     if plan.branch in AIV_BRANCHES:
         if plan.used_core_num > s.aiv_num:
@@ -50,8 +57,9 @@ def check_plan_constraints(case: BmmCase, plan: ImplPlan,
     if plan.used_core_num > s.aic_num:
         v.append(f"used_core_num={plan.used_core_num} 超 AIC 核数 {s.aic_num}")
 
-    # --- L0C --- (ASW 降核每核单份, 其他分支双缓冲)
+    # --- L0C --- (ASW 两分支 UnitFlag 单缓冲: tile 内 16x16x16 细粒度流水
-    l0c_factor = 1 if plan.branch == "ASW_Basic_降核" else 2
+    #             (ASW_Basic分支分析 v1.91 §5.1); IterBatch/MergeBatch 双缓冲)
+    l0c_factor = 1 if plan.branch in ("ASW_Basic", "ASW_Basic_降核") else 2
     l0c_need = plan.base_m * plan.base_n * 4 * l0c_factor
     if l0c_need > s.l0c_bytes:
         v.append(f"L0C tile 超容量: BaseM*BaseN*4B*{l0c_factor}={l0c_need}B > {s.l0c_bytes}B")
@@ -77,9 +85,24 @@ def check_plan_constraints(case: BmmCase, plan: ImplPlan,
 
     # --- dValue --- (issue#6 口径裁定: 只对"以 K 段为连续维"的方案生效)
     if _k_segment_is_contiguous(plan, case) and plan.k_l1 > 0:
-        dv = plan.k_l1 * case.dtype_in_bytes
+        if plan.branch in ("ASW_Basic", "ASW_Basic_降核"):
-        if dv < s.dvalue_min:
+            # issue#34: ASW 是兜底分支, 恒出方案; dValue 效率下限不满足只算**效率降级**
-            v.append(f"dValue={dv}B < 下限 {s.dvalue_min}B, K 段连续维搬移效率崩塌")
+            # (plan.note 含"效率降级"标注, DMA 仍能工作只是效率低), 不判违规.
+            # 真正不可行的只有容量/核数硬约束 (L0C/L0A/L0B/L1 超容, 核数超限).
+            pass
+        elif plan.branch == "IterBatch" and plan.l1_form.startswith(("c_", "d_")):
+            # 转置感知判据与生成守卫/条件 4 同源 (issue#19):
+            # dv_a = M*dt (A 转置) 或 k_l1*dt; dv_b = N*dt (B 不转置) 或 k_l1*dt;
+            # 两侧连续维 dValue 均低于下限才算违规.
+            from .models import dvalue_contig_dims
+            dv_a, dv_b = dvalue_contig_dims(case, plan.k_l1)
+            if dv_a < s.dvalue_min and dv_b < s.dvalue_min:
+                v.append(f"dValueA={dv_a:.0f}B 与 dValueB={dv_b:.0f}B 均 < 下限 "
+                         f"{s.dvalue_min}B, 搬移连续维效率崩塌")
+        else:
+            dv = plan.k_l1 * case.dtype_in_bytes
+            if dv < s.dvalue_min:
+                v.append(f"dValue={dv}B < 下限 {s.dvalue_min}B, K 段连续维搬移效率崩塌")
 
     # --- 写出 dtype --- (StreamK 部分和 4B 是正确行为)
     if plan.branch not in PARTIAL_SUM_4B_BRANCHES:

BMM/BMM_Theory/bmm_theory/evaluator.py

@@ -63,14 +63,16 @@ class PlanEvaluator:
     def _advice(self, case: BmmCase, plan: ImplPlan, res: EvalResult) -> str:
         t = res.timing
         tips = []
+        if "效率降级" in plan.note:
+            tips.append("效率降级标注 (plan.note): 搬移效率下限不满足 —— 方案照常给出 "
+                        "(兜底), 但实际效率低于模型假设, 时延可能低估; 建议调整 dtype/布局")
         if not res.feasible:
             tips.append("方案违反硬件约束, 需先修正: " + res.violations)
         bn = t.bottleneck
-        if bn == "MTE2_GM":
+        if bn == "MTE2":
-            tips.append("瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, "
+            tips.append("瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile "
-                        "或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)")
+                        "提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/"
-        elif bn == "MTE2_L2":
+                        "分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争")
-            tips.append("瓶颈在 L2 重复读: 优化核间分配/swizzle 窗口压低活跃工作集")
         elif bn == "MMAD":
             tips.append("瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 "
                         "(MergeBatch 交叉项) 可消除")

BMM/BMM_Theory/bmm_theory/hardware/__init__.py

@@ -1,3 +1,24 @@
-from .ascend950pr import ASCEND950PR, NpuSpec
+from .ascend950pr import ASCEND950PR, ASCEND950PR_C28, NpuSpec
+from .ascend950dt import ASCEND950DT, ASCEND950DT_C32, ASCEND950DT_C28
 
-__all__ = ["ASCEND950PR", "NpuSpec"]
+# 芯片规格注册表 (issue#38): name -> NpuSpec; 默认主 bin = ASCEND950PR
+SPECS = {s.name: s for s in (
+    ASCEND950PR, ASCEND950PR_C28,
+    ASCEND950DT, ASCEND950DT_C32, ASCEND950DT_C28,
+)}
+
+
+def get_spec(name: str = "Ascend950PR") -> NpuSpec:
+    """按名称取芯片规格 (默认 Ascend950PR 32 核主 bin); 未知名称抛 KeyError."""
+    try:
+        return SPECS[name]
+    except KeyError:
+        raise KeyError(f"未知芯片规格 {name!r}, 可选: {sorted(SPECS)}") from None
+
+
+__all__ = [
+    "NpuSpec",
+    "ASCEND950PR", "ASCEND950PR_C28",
+    "ASCEND950DT", "ASCEND950DT_C32", "ASCEND950DT_C28",
+    "SPECS", "get_spec",
+]

BMM/BMM_Theory/bmm_theory/hardware/ascend950dt.py

@@ -0,0 +1,57 @@
+"""Ascend950DT 硬件规格参数 (issue#38).
+
+数值来源: 昇腾950 NPU架构白皮书 表3-1 (系列 SKU 规格) / 表4-2 (Memory 层次);
+产品形态: Atlas 850E 超节点 / Atlas 650E 服务器 / Atlas 950 SuperPoD 超节点.
+
+与 950PR **共架构** (第三代 DaVinci), 仅以下维度不同:
+  - AI 子系统 36/32/28 个 (每子系统 = 1 Cube Core + 2 Vector Core);
+  - 高速片上内存 4TB/s (950PR 为 1.6/1.4TB/s), 容量 144/96GB;
+  - L2 128MB 全档一致 (950PR 28 核降 112MB).
+
+共架构的交叉验证 (表3-1/表4-2 反推, 与 950PR 单核速率一致):
+  - Cube: 432T/32核 = 486T/36核 = 13.5T/核 (= 16^3 MAC/拍 @ ~1.65GHz),
+    cube_peak_tflops 取白皮书表3-1 "Cube算力" 单行 (Cube-only 口径,
+    issue#40 用户裁决), 单核 13.5T 全系列精确一致;
+  - Vector fp32: 27T/64核 ≈ 30T/72核 -> 128 lane/拍 @ 1.65GHz (乘加计 2 次),
+    即 aiv_freq_ghz / aiv_fp32_per_cycle 默认值直接适用;
+  - L1 512KB / L0A/L0B 64KB / L0C 256KB / UB 512KB 每核一致 (表4-2);
+  - L2 带宽 7.5TB/s, 读写各自独享 (issue#39 用户澄清; 白皮书未分档,
+    DT 三档统一); 搬移效率经验值 (dValue/minTile/minDatamountPerCore/
+    T_cmd=0) 同架构沿用 950PR 口径.
+"""
+
+from __future__ import annotations
+
+from .ascend950pr import NpuSpec
+
+# 950DT 36 核主 bin (白皮书表3-1: 36 AIC / 72 AIV, Cube BF16 486 TFLOPS,
+# HBM 4TB/s 144GB, L2 128MB; L2 带宽 7.5TB/s 读写各自独享, issue#39)
+ASCEND950DT = NpuSpec(
+    name="Ascend950DT",
+    aic_num=36, aiv_num=72,
+    cube_peak_tflops=486.0,
+    bw_gm=4.0e12,
+    bw_l2=7.5e12,
+    gm_capacity_gb=144.0,
+)
+
+# 950DT 32 核档 (Cube BF16 432 TFLOPS; HBM 4TB/s 不变;
+# 容量表3-1 仅给 144/96 两档, 32 核档映射待确认)
+ASCEND950DT_C32 = NpuSpec(
+    name="Ascend950DT_C32",
+    aic_num=32, aiv_num=64,
+    cube_peak_tflops=432.0,
+    bw_gm=4.0e12,
+    bw_l2=7.5e12,
+    gm_capacity_gb=144.0,
+)
+
+# 950DT 28 核档 (Cube BF16 378 TFLOPS; HBM 4TB/s 不变, 容量 96GB)
+ASCEND950DT_C28 = NpuSpec(
+    name="Ascend950DT_C28",
+    aic_num=28, aiv_num=56,
+    cube_peak_tflops=378.0,
+    bw_gm=4.0e12,
+    bw_l2=7.5e12,
+    gm_capacity_gb=96.0,
+)

BMM/BMM_Theory/bmm_theory/hardware/ascend950pr.py

@@ -1,15 +1,47 @@
 """Ascend950PR (DAV_3510) 硬件规格参数.
 
-数值来源: 《BMM算子优化分析 v0.98》§二 + 昇腾950 NPU架构白皮书.
+数值来源: 《BMM算子优化分析 v0.98》§二 + 昇腾950 NPU架构白皮书 (表3-1 SKU 规格 /
-换芯片时逻辑结构不变, 只需新增一份同结构参数表.
+表4-2 Memory 层次). 换芯片时逻辑结构不变, 只需新增一份同结构参数表
+(系列其他 SKU 见文末 ASCEND950PR_C28 与 hardware/ascend950dt.py, issue#38).
 
 单位约定: 算力 FLOP/s, 带宽 Byte/s, 容量 Byte, 时延 秒.
+
+dtype 感知算力 (issue#28, 设计文档 docs/05 §2):
+  Cube 算力按输入 dtype 分档, 基准 = BF16 (fp16 同速); 白皮书: FP8/MXFP8/HiF8
+  提供 2x FP16 张量 TFLOPS, MXFP4 提供 4x FP16; FP32/TF32 同代比值为假设值
+  (按 DaVinci 惯例 = 1/2, 白皮书未给同代比值), int8 假设同 FP8, 均待实测标定.
+  AIV 逐元素通量按 lane 位宽等比假设 (16bit x2 / 8bit x4 / 4bit x8, 待标定).
+
+口径 (issue#40 用户裁决): cube_peak_tflops 采用白皮书表3-1 **"Cube算力"单行**
+  (Cube-only, 不含 Vector): 950PR 32核 = 432, 单核 13.5T = 16³ MAC/拍 × 1.65GHz;
+  全系列 SKU 单核精确统一 13.5T (432/32 = 378/28 = 486/36)。2026-09 前曾沿用
+  v0.98 的 "Cube+Vector 总算力" 口径 (486, 单核 15.1875T), 切换后全模型 MMAD
+  时延 ×486/432 (+12.5%), R16 607.5→540, StreamK θ_c 12.24→10.9,
+  ASW 面积/周长分界 93.5→83.1。
 """
 
 from __future__ import annotations
 
 from dataclasses import dataclass
 
+# Cube 精度因子 (相对 BF16/FP16 基准档; A/B 不一致时取较慢一侧 = min 因子)
+CUBE_DTYPE_FACTOR = {
+    "fp32": 0.5, "f32": 0.5, "tf32": 0.5,   # 假设 = 1/2, 待实测标定
+    "fp16": 1.0, "f16": 1.0, "bf16": 1.0,
+    "fp8": 2.0, "fp8_e4m3": 2.0, "fp8_e5m2": 2.0, "int8": 2.0,  # 白皮书 FP8=2xFP16; int8 假设同 FP8
+    "fp4": 4.0, "fp4_e2m1": 4.0,             # 白皮书 MXFP4=4xFP16; 普通 fp4 假设同 MXFP4
+}
+
+# AIV 逐元素通量因子 (相对 fp32 128 lane/拍/核; 位宽等比假设, 待实测标定)
+AIV_DTYPE_FACTOR = {
+    "fp32": 1.0, "f32": 1.0, "tf32": 1.0,
+    "fp16": 2.0, "f16": 2.0, "bf16": 2.0,
+    "fp8": 4.0, "fp8_e4m3": 4.0, "fp8_e5m2": 4.0, "int8": 4.0,
+    "fp4": 8.0, "fp4_e2m1": 8.0,
+}
+
+_RATE_FALLBACK = 1.0  # 未知 dtype 按基准档 (models.dtype_bytes 已先行校验, 正常不会到达)
+
 
 @dataclass(frozen=True)
 class NpuSpec:
@@ -18,7 +50,8 @@ class NpuSpec:
     # ---- 核数与算力 ----
     aic_num: int = 32                      # C: AIC (Cube) 核数
     aiv_num: int = 64                      # AIV (Vector) 核数
-    cube_peak_tflops: float = 486.0        # 全芯片 Cube BF16 标称算力 (乘加各计一次)
+    cube_peak_tflops: float = 432.0        # 全芯片 Cube BF16 标称算力 (乘加各计一次,
+                                           # 白皮书表3-1 Cube 单行, issue#40)
     aiv_freq_ghz: float = 1.65
     aiv_fp32_per_cycle: int = 128          # 单 AIV 每拍 fp32 求和吞吐
 
@@ -30,8 +63,9 @@ class NpuSpec:
 
     # ---- L2 / GM ----
     l2_bytes: int = 128 * 1024 * 1024      # L2 Cache 128MB
-    bw_l2: float = 5.2e12                  # L2 读写带宽 5.2TB/s
+    bw_l2: float = 5.2e12                  # L2 读写带宽 5.2TB/s (白皮书未给, 用户澄清口径)
     bw_gm: float = 1.6e12                  # GM(HBM) 带宽 1.6TB/s (读写共享)
+    gm_capacity_gb: float = 128.0          # GM(HBM) 容量 GB (信息项, 模型暂不消费)
 
     # ---- 搬移效率经验约束 (950PR 实测总结) ----
     dvalue_recommend: int = 512            # dValue 推荐值 (Byte), 更大无额外收益
@@ -39,10 +73,19 @@ class NpuSpec:
     dvalue_hw_min: int = 256               # DMA 硬件突发下限 (Byte) —— 尾轮文档 §2.3
     min_tile_size: int = 16 * 1024         # min_TileSize: 单块搬移最小量 16KB
     min_datamount_per_core: int = 480 * 1024   # min_DatamountPerCore: 单核搬移总量下限 480KB
-    min_core_num_ratio: float = 0.8        # minCoreNum ≈ 0.8 * C
+    min_core_num_ratio: float = 0.8        # minCoreNum ≈ 0.8 * C (经验: 约 3/4 核并发才达 90%+ 带宽利用率)
 
     # ---- DMA 固定开销 ----
-    t_cmd_ns: float = 50.0                 # T_cmd: 单次 GM->L1 DMA 命令固定开销 (ns, 估计值, 需实测标定)
+    t_cmd_ns: float = 0.0                  # T_cmd: 单次 GM->L1 DMA 命令固定开销 (ns)。
+                                           # 未标定, 按 0 处理 (issue#36 用户澄清):
+                                           # MergeBatch 的合并收益由搬移效率模型
+                                           # (move_eff: 单块 tile=nValue*dValue*dt
+                                           # 放大 b0 倍) 刻画, 不依赖 T_cmd 估计值.
+
+    # ---- 带宽模型假设 (issue#26) ----
+    # GM 1.6TB/s 为读写共享总线 (读+写累加计时); L2 带宽读写各自独享 5.2TB/s;
+    # active_cores < C 时按"每核份额线性配平"为模型假设 (如 1 核也给 50GB/s),
+    # 真实低核数带宽利用率低于线性, 需 msProf 实测曲线标定后替换该假设.
 
     # ---- Cube 计算粒度 ----
     fractal: int = 16                      # 16x16x16 基本块
@@ -50,13 +93,50 @@ class NpuSpec:
     # ---- 派生量 (属性) ----
     @property
     def q16(self) -> float:
-        """单核 Cube BF16 峰值算力 (FLOP/s)."""
+        """单核 Cube BF16/FP16 峰值算力 (FLOP/s) = 13.5T (issue#40 Cube-only)."""
         return self.cube_peak_tflops * 1e12 / self.aic_num
 
+    @staticmethod
+    def _dtype_key(dtype) -> str:
+        return str(dtype).strip().lower()
+
+    def cube_factor(self, dtype_a, dtype_b=None) -> float:
+        """按输入 dtype 取 Cube 精度因子 (issue#28).
+
+        A/B 不一致时取**较慢一侧** (因子较小者, 等价字节较大者);
+        Cube 乘法两侧的实际吞吐受较慢精度限制.
+        """
+        keys = [self._dtype_key(dtype_a)]
+        if dtype_b is not None:
+            keys.append(self._dtype_key(dtype_b))
+        factors = [CUBE_DTYPE_FACTOR.get(k, _RATE_FALLBACK) for k in keys]
+        return min(factors)
+
+    def q_cube(self, dtype_a, dtype_b=None) -> float:
+        """单核 Cube 峰值算力 (FLOP/s), 按输入 dtype 分档 (issue#28)."""
+        return self.q16 * self.cube_factor(dtype_a, dtype_b)
+
+    def aiv_elem_factor(self, dtype) -> float:
+        """按 dtype 取 AIV 逐元素通量因子 (相对 fp32 基准)."""
+        return AIV_DTYPE_FACTOR.get(self._dtype_key(dtype), _RATE_FALLBACK)
+
+    @property
+    def aiv_elem_rate_fp32(self) -> float:
+        """AIV fp32 逐元素吞吐 (元素/s), 64 核合计."""
+        return self.aiv_num * self.aiv_fp32_per_cycle * self.aiv_freq_ghz * 1e9
+
+    def aiv_elem_rate(self, dtype) -> float:
+        """AIV 逐元素吞吐 (元素/s, 64 核合计), 按 dtype 分档 (issue#28)."""
+        return self.aiv_elem_rate_fp32 * self.aiv_elem_factor(dtype)
+
     @property
     def q_aiv(self) -> float:
-        """AIV 向量求和吞吐 (元素/s), 64 核合计."""
+        """AIV fp32 逐元素吞吐 (元素/s, 64 核合计) = aiv_elem_rate_fp32.
-        return self.aiv_num * self.aiv_fp32_per_cycle * self.aiv_freq_ghz * 1e9
+
+        注意: 该值只适用于 fp32 数据 (如 StreamK 对 fp32 部分和求和);
+        逐元素运算按输入 dtype 用 aiv_elem_rate(dtype) (issue#28).
+        """
+        return self.aiv_elem_rate_fp32
 
     @property
     def bw_pc(self) -> float:
@@ -70,7 +150,11 @@ class NpuSpec:
 
     @property
     def r16(self) -> float:
-        """16bit 位宽平衡点算存比 R_16 = Cube峰值 / (GM带宽/2B) ≈ 607.5 FLOP/元素."""
+        """16bit 位宽平衡点算存比 R_16 = Cube峰值 / (GM带宽/2B) = 540 FLOP/元素.
+
+        按 issue#28 速率表该比值对全 dtype 不变 (Cube 因子与元素字节数互成反比),
+        故入口条件沿用单一 R_16.  (issue#40 Cube-only 口径: 607.5 -> 540)
+        """
         return self.cube_peak_tflops * 1e12 / (self.bw_gm / 2)
 
     @property
@@ -90,3 +174,14 @@ class NpuSpec:
 
 
 ASCEND950PR = NpuSpec()
+
+# 950PR 28 核衍生版 (白皮书表3-1: 28 AIC / 56 AIV, HBM 1.4TB/s 112GB,
+# L2 112MB; L2 带宽未分档, 沿用 5.2TB/s 待标定)
+ASCEND950PR_C28 = NpuSpec(
+    name="Ascend950PR_C28",
+    aic_num=28, aiv_num=56,
+    cube_peak_tflops=378.0,
+    l2_bytes=112 * 1024 * 1024,
+    bw_gm=1.4e12,
+    gm_capacity_gb=112.0,
+)

BMM/BMM_Theory/bmm_theory/models.py

@@ -28,13 +28,15 @@ DTYPE_BYTES = {
     "fp8_e4m3": 1,
     "fp8_e5m2": 1,
     "int8": 1,
+    "fp4": 0.5,      # 半字节 (对齐 bmmv3, 2026-09-03)
+    "fp4_e2m1": 0.5,
 }
 
-# Cube 累加器 (L0C) 中元素字节数: 16bit 输入 -> fp32 累加; fp8 输入 -> fp32 累加
+# Cube 累加器 (L0C) 中元素字节数: 16bit 输入 -> fp32 累加; fp8/fp4 输入 -> fp32 累加
 L0C_DTYPE_BYTES = 4
 
 
-def dtype_bytes(dtype: str) -> int:
+def dtype_bytes(dtype: str) -> float:
     key = dtype.strip().lower()
     if key not in DTYPE_BYTES:
         raise ValueError(f"不支持的 dtype: {dtype!r}, 支持 {sorted(DTYPE_BYTES)}")
@@ -53,6 +55,53 @@ def align_down(x: int, align: int) -> int:
     return (x // align) * align
 
 
+def dvalue_contig_dims(case: "BmmCase", k_l1: float) -> tuple:
+    """A/B 两侧 GM->L1 搬移的连续维 dValue (Byte) —— 依转置定连续维 (ND 排布).
+
+    A 不转置存储 [M,K]: 行内 K 连续, 切 K 后每行连续段 k_l1 -> k_l1*dt;
+    A 转置存储 [K,M]:   M 连续                     -> m*dt;
+    B 不转置存储 [K,N]: N 连续                     -> n*dt;
+    B 转置存储 [N,K]:   行内 K 连续, 切 K 后连续段 k_l1 -> k_l1*dt.
+
+    供三处共用 (issue#19): l1_form 生成守卫 / IterBatch 条件 4 / constraints 校验,
+    保证"生成说可行、条件说可行、校验说可行"口径一致.
+    """
+    dt = case.dtype_in_bytes
+    dv_a = case.m * dt if case.trans_a else k_l1 * dt
+    dv_b = k_l1 * dt if case.trans_b else case.n * dt
+    return dv_a, dv_b
+
+
+def move_eff(tile_bytes: float, min_tile_size: int) -> float:
+    """GM->L1 单命令搬移效率 (issue#36, 用户澄清口径).
+
+    单命令搬移效率由单块 tile = nValue × dValue × input_dtype 决定: tile 越大,
+    有效带宽越高; 达到 min_TileSize (16KB, 即进入条件 4 的效率下限常数) 饱和,
+    之下线性退化 —— 与"max(MK,KN)·dt >= min_TileSize 才保证搬移效率"的语义同源
+    (该条件只要求一侧达标, 另一侧小 tile 的低效由此函数量化)。
+
+    MergeBatch 合并 b0 个 batch 后单块 tile 放大 b0 倍 (A 侧: b0·M·k_l1·dt,
+    B 侧: b0·N·k_l1·dt; 堆叠方向视转置/排布而定 —— A ND 非转置时沿 M(nValue)
+    堆叠, B ND 非转置时沿 N(dValue) 堆叠, 乘积口径不变), 因此即便 T_cmd=0
+    合并仍有搬移效率收益。
+    """
+    if tile_bytes <= 0:
+        return 1.0
+    return min(1.0, tile_bytes / min_tile_size)
+
+
+def gm_move_time(v_a: float, v_b: float, tile_a: float, tile_b: float,
+                 spec) -> float:
+    """效率加权的 GM->L1 数据时延 (秒, 整芯片口径, issue#36).
+
+    t = (V_A/eff_A + V_B/eff_B) / BW_gm; 只影响**时间**列, GM 字节量不变
+    (仍按每输入字节恰读一次 = V_in 计, issue#31)。
+    """
+    ea = move_eff(tile_a, spec.min_tile_size)
+    eb = move_eff(tile_b, spec.min_tile_size)
+    return (v_a / ea + v_b / eb) / spec.bw_gm
+
+
 # ---------------------------------------------------------------------------
 # Case 输入
 # ---------------------------------------------------------------------------
@@ -109,12 +158,19 @@ class BmmCase:
         return max(self.batch_a, self.batch_b)
 
     @property
-    def dtype_in_bytes(self) -> int:
+    def dtype_in_bytes(self) -> float:
         # A/B 输入元素字节数 (要求 A/B 同 dtype, 不一致时取较大者并在校验中报 warning)
+        # 返回 float 以兼容 fp4 (0.5B)
         return max(dtype_bytes(self.dtype_a), dtype_bytes(self.dtype_b))
 
     @property
-    def dtype_out_bytes(self) -> int:
+    def dtype_in(self) -> str:
+        """输入侧"较慢"dtype (元素字节较大者) — Cube/AIV 速率取慢侧 (issue#28)."""
+        return (self.dtype_a if dtype_bytes(self.dtype_a) >= dtype_bytes(self.dtype_b)
+                else self.dtype_b)
+
+    @property
+    def dtype_out_bytes(self) -> float:
         return dtype_bytes(self.dtype_c)
 
     @property
@@ -241,23 +297,30 @@ class ImplPlan:
 
 @dataclass
 class HardwareTiming:
-    """各硬件流水级时延 (秒) 与数据量明细."""
+    """各硬件流水级时延 (秒) 与数据量明细.
+
+    数据量列口径 (issue#29, 设计文档 docs/05 §4.4):
+      gm_read_bytes / l2_read_bytes / fixpipe_bytes / cube_flops = **整芯片**总量
+      (跨分支可比, GM 首读下限断言: gm_read_bytes >= case.input_bytes);
+      dma_cmd_count = **单核** GM->L1 DMA 命令数 (各核引擎并行, 墙钟 T_cmd =
+      单核命令数 x T_cmd), 与字节列口径不同属.
+    """
 
     # 搬入 (MTE2)
-    gm_read_bytes: float = 0.0     # GM->L1 直读数据量 (不驻留/未命中 L2 的部分)
+    gm_read_bytes: float = 0.0     # GM->L1 直读数据量 (整芯片, 首读/落空重读)
-    l2_read_bytes: float = 0.0     # L2->L1 数据量 (驻留 L2 后重复读命中部分)
+    l2_read_bytes: float = 0.0     # L2->L1 数据量 (整芯片, 驻留 L2 后重复读命中部分)
     t_mte2_gm: float = 0.0         # GM->L1 时延 (按 GM 带宽, 不累加 L2->L1)
     t_mte2_l2: float = 0.0         # L2->L1 时延 (按 L2 带宽)
     t_mte2: float = 0.0            # 搬入合计 = t_mte2_gm + t_mte2_l2 (两者发生在不同数据上)
-    dma_cmd_count: float = 0.0     # GM->L1 DMA 命令次数 (T_cmd 分析用)
+    dma_cmd_count: float = 0.0     # 单核 GM->L1 DMA 命令次数 (T_cmd 分析用)
-    t_dma_cmd: float = 0.0         # DMA 命令固定开销合计
+    t_dma_cmd: float = 0.0         # DMA 命令固定开销合计 (墙钟)
 
     # 计算 (Cube MMAD)
-    cube_flops: float = 0.0        # Cube 实际计算量 (MergeBatch 含冗余)
+    cube_flops: float = 0.0        # Cube 实际计算量 (整芯片, MergeBatch 含冗余)
     t_mmad: float = 0.0
 
     # 搬出 (Fixpipe)
-    fixpipe_bytes: float = 0.0     # 写出数据量 (按 C 矩阵 dtype / StreamK 临时矩阵按 4B)
+    fixpipe_bytes: float = 0.0     # 写出数据量 (整芯片, 按 C 矩阵 dtype / StreamK 临时矩阵按 4B)
     t_fixpipe: float = 0.0
 
     # 归约 (StreamK 专用)

BMM/BMM_Theory/bmm_theory/router.py

@@ -4,10 +4,12 @@
   1. 前置归约: BatchA=1 或 BatchB=1 -> 转Matmul
                 K=0 / K=1 -> 特殊分支 (AIV 向量通路)
   2. B >= C 且 BatchA==BatchB: 切B -> IterBatch 与 MergeBatch 仲裁
-       仲裁规则 (v1.1 §4.5 统一分界):
+       仲裁规则 (v1.1 §4.5 统一分界 + issue#35/#36 泛化):
-         MergeBatch 最优 <=> K截断(k_L1=K) 且 b_core > b0*(T_comp+T_write)/T_cmd
+         净收益 = 命令节省(cmds 差 × T_cmd) + 搬移效率节省(合并 tile 放大 b0 倍,
-         L1 绑定时 MergeBatch 恒劣于 IterBatch;
+                  move_eff 模型) − drain 惩罚; K截断且效率打平时退化为文档闭式
-       两分支同时合法时用端到端时延模型 T_total 仲裁.
+                  b_core > b0*(T_comp+T_write)/T_cmd;
+       两分支同时合法时用端到端时延模型 T_total 终审 (T_cmd 默认 0, 未标定;
+       合并效率收益已建模, 不再需要策略覆盖).
   3. StreamK 检查: P <= C/2 且满足切K条件 -> StreamK (B/M/N 买不满时买 K)
   4. 兜底: ASW_Basic 切 M/N (含降核模式)
 """
@@ -84,7 +86,9 @@ class BranchRouter:
 
         arbitration = ""
         if mb.capable and ib.capable:
-            # 统一分界条件 + 端到端时延仲裁双保险
+            # 统一分界条件 + 端到端时延仲裁双保险; 时延模型为最终裁决.
+            # (issue#36: 合并的搬移效率收益已由 move_eff 模型量化进时延模型,
+            #  T_cmd=0 默认下不再需要"T_cmd<=0 策略优先 MergeBatch"的覆盖逻辑)
             mb_win, detail = self.merge_batch.beats_iterbatch(case)
             t_mb = mb.timing.t_total
             t_ib = ib.timing.t_total
@@ -94,10 +98,10 @@ class BranchRouter:
                 f"两分支均合法, 仲裁: "
                 f"[分界条件] MergeBatch最优={mb_win} ({detail}); "
                 f"[时延模型] T_MergeBatch={t_mb*1e6:.2f}us vs T_IterBatch={t_ib*1e6:.2f}us -> {lat_win}更优; "
-                f"[裁决] {win}" + ("" if win == lat_win else f" (分界条件与时延模型不一致, 以时延模型为准: {lat_win})")
+                f"[裁决] {lat_win}" + ("" if win == lat_win else
+                    f" (分界条件判{win}, 与时延模型不一致, 以时延模型为准)")
             )
-            if win != lat_win:
+            win = lat_win  # 时延模型为最终裁决
-                win = lat_win  # 时延模型为最终裁决
         elif any(capable.values()):
             win = next(n for n, v in capable.items() if v)
             arbitration = f"仅 {win} 条件满足"
@@ -158,6 +162,9 @@ class BranchRouter:
         if violations:
             note = (note + " [自检违规: " + "; ".join(violations) +
                     "] —— 方案生成存在缺陷, 需人工复核")
+        # issue#34: 兜底分支 (ASW) 效率下限不满足时降级标注 (warning), 不判违规
+        if result.plan is not None and "效率降级" in result.plan.note:
+            note += " [效率降级标注: 搬移效率低于模型假设, 时延可能低估, 见 plan.note]"
         return {
             "branch": result.plan.branch if result.plan else "未知",
             "plan": result.plan,

BMM/BMM_Theory/bmm_theory/timing.py

@@ -24,7 +24,19 @@ from __future__ import annotations
 from dataclasses import dataclass
 
 from .hardware import NpuSpec, ASCEND950PR
-from .models import HardwareTiming
+from .models import BmmCase, HardwareTiming
+
+
+def output_to_l2(case: BmmCase, spec: NpuSpec = ASCEND950PR,
+                 workspace_bytes: float = 0.0) -> bool:
+    """Fixpipe 输出落点决策 (issue#30, 设计文档 docs/05 §4.2 R4).
+
+    to_l2 (输出写 L2 写口 5.2TB/s, GM 写流量 = 0, 异步回写不占算子时延)
+      ⟺  整 case 输入 V_in + 输出 V_out [+ StreamK workspace] ≤ L2
+    否则输出**直写 GM**: 输入优先驻留 L2 (输入存在重复读), 输出计入 GM
+    读写共享总线 (与读累加, issue#23).
+    """
+    return (case.input_bytes + case.output_bytes + workspace_bytes) <= spec.l2_bytes
 
 
 @dataclass
@@ -56,9 +68,11 @@ def eval_mte2(move: MoveInPlan, spec: NpuSpec = ASCEND950PR,
     return t_gm, t_l2, t_gm + t_l2 + t_cmd, t_cmd
 
 
-def eval_mmad(flops_per_core: float, spec: NpuSpec = ASCEND950PR) -> float:
+def eval_mmad(flops_per_core: float, spec: NpuSpec = ASCEND950PR,
-    """Cube 计算时延: 单核计算量 / 单核算力."""
+              dtype_a=None, dtype_b=None) -> float:
-    return flops_per_core / spec.q16 if flops_per_core > 0 else 0.0
+    """Cube 计算时延: 单核计算量 / 单核 dtype 感知算力 (issue#28)."""
+    rate = spec.q_cube(dtype_a, dtype_b)
+    return flops_per_core / rate if flops_per_core > 0 else 0.0
 
 
 def eval_fixpipe(bytes_per_core: float, to_l2: bool,
@@ -75,21 +89,24 @@ def eval_fixpipe(bytes_per_core: float, to_l2: bool,
 
 
 def eval_streamk_reduce(tile_elems: float, grid_k: int, out_dtype_bytes: int,
-                        spec: NpuSpec = ASCEND950PR) -> float:
+                        spec: NpuSpec = ASCEND950PR, out_to_gm: bool = False) -> float:
     """StreamK 单 tile 归约时延 (v0.98 §七).
 
-    部分和 dtype = L0C dtype (4B, 防精度丢失), 驻留 L2, AIV 归约:
+    部分和 dtype = L0C dtype (4B, 防精度丢失), 驻留 L2, AIV 归约 (fp32 求和,
+    AIV 按 fp32 通量, 不随输入 dtype 变 — issue#28):
       AIC 写部分和 grid_k x tile x 4B / W_L2
       AIV 读回     grid_k x tile x 4B / W_L2
-      AIV 求和     grid_k x tile / Q_AIV
+      AIV 求和     grid_k x tile / Q_AIV(fp32)
-      写回         tile x outB / W_L2
+      写回         tile x outB / W (最终输出落点, issue#30):
+                   整 case 可驻留 (S_A) 时 W = W_L2; 否则直写 GM (W = W_GM)
     """
     b4 = 4
     w_l2 = spec.bw_l2
+    w_out = spec.bw_gm if out_to_gm else w_l2
     t_write_partial = grid_k * tile_elems * b4 / w_l2
     t_read_back = grid_k * tile_elems * b4 / w_l2
     t_sum = grid_k * tile_elems / spec.q_aiv
-    t_write_out = tile_elems * out_dtype_bytes / w_l2
+    t_write_out = tile_elems * out_dtype_bytes / w_out
     return t_write_partial + t_read_back + t_sum + t_write_out
 
 
@@ -99,19 +116,31 @@ def assemble_timing(t_mte2_gm: float, t_mte2_l2: float, t_dma_cmd: float,
                     gm_read_bytes: float, l2_read_bytes: float,
                     dma_cmd_count: float, cube_flops: float,
                     fixpipe_bytes: float,
-                    reduce_serial: bool = True) -> HardwareTiming:
+                    reduce_serial: bool = True,
+                    fixpipe_to_gm: bool = True) -> HardwareTiming:
     """汇总各级时延, 判定瓶颈.
 
+    带宽端口口径 (issue#23, 用户澄清 + KB):
+    - GM 1.6TB/s 为**读写共享总线**: MTE2 的 GM 读与 Fixpipe 直写 GM 并发时无法
+      拆分读写占用, 时延累加 (读+写)/1.6TB/s, 并入 MTE2 搬移链;
+    - L2 带宽读写各自独享 5.2TB/s: L2 重复读段(读口) 与 Fixpipe→L2 写(写口)
+      互不竞争; MTE2 引擎顺序服务 GM/L2 装载, 两段相加 (官方 T≈HBM/1.6+L2/5.2);
+    - Fixpipe→L2 (resident) 写出独立为 FIXPIPE 级, 不进 GM 总线.
+
     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
+    fix_gm = t_fixpipe if fixpipe_to_gm else 0.0   # Fixpipe 直写 GM 的时延
+    fix_l2 = 0.0 if fixpipe_to_gm else t_fixpipe   # Fixpipe→L2 (5.2TB/s 写口)
+    # MTE2 搬移链 = GM 总线(读写共享, 读+直写累加) + L2 重复读段 + DMA 命令开销
+    # (同一 MTE2 引擎顺序服务 GM/L2 两类装载; GM 写由 Fixpipe 并发发起, 共享 GM 总线)
+    mte2_chain = t_mte2_gm + t_dma_cmd + fix_gm + t_mte2_l2
     stages = {
-        "MTE2_GM": t_mte2_gm + t_dma_cmd,
+        "MTE2": mte2_chain,
-        "MTE2_L2": t_mte2_l2,
         "MMAD": t_mmad,
-        "FIXPIPE": t_fixpipe,
+        "FIXPIPE": fix_l2,   # 仅 Fixpipe→L2 (5.2 写口, 与 L2 读口互不竞争)
     }
     if not reduce_serial:
         stages["REDUCE"] = t_reduce
@@ -135,8 +164,7 @@ def assemble_timing(t_mte2_gm: float, t_mte2_l2: float, t_dma_cmd: float,
 def bound_type_of(bottleneck: str) -> str:
     return {
         "MMAD": "计算Bound",
-        "MTE2_GM": "访存Bound(GM)",
+        "MTE2": "访存Bound(GM读写共享+L2重复读)",
-        "MTE2_L2": "访存Bound(L2)",
+        "FIXPIPE": "写出Bound(L2写口)",
-        "FIXPIPE": "写出Bound",
         "REDUCE": "归约Bound",
     }.get(bottleneck, "")

BMM/BMM_Theory/docs/01_软件架构.md

@@ -25,7 +25,9 @@
 │  └─ constraints.py  单一约束源 (生成与校验共用)            │
 ├─────────────────────────────────────────────────────────┤
 │ 硬件层 (hardware/)                                       │
-│  └─ ascend950pr.py  NpuSpec 参数表 (换芯片只换这份)        │
+│  ├─ ascend950pr.py  NpuSpec + 950PR 32/28核 SKU          │
+│  ├─ ascend950dt.py  950DT 36/32/28核 SKU (issue#38)      │
+│  └─ __init__.py     SPECS 注册表 + get_spec(name)        │
 ├─────────────────────────────────────────────────────────┤
 │ IO 层 (io_csv.py)  case/plan/result 的 csv 读写           │
 └─────────────────────────────────────────────────────────┘
@@ -86,7 +88,7 @@ class Branch:
 
 ### 3.3 硬件参数与逻辑分离
 
-`hardware/ascend950pr.py` 的 `NpuSpec` 集中所有芯片常数（核数/算力/各级容量/带宽/搬移效率经验值/T_cmd）。所有分支通过 `self.spec` 取参数——换芯片时新增一份参数表即可，分支逻辑零改动。
+`hardware/` 的 `NpuSpec` 集中所有芯片常数（核数/算力/各级容量/带宽/搬移效率经验值/T_cmd）。所有分支通过 `self.spec` 取参数——换芯片时新增一份参数表即可，分支逻辑零改动。已注册 SKU（issue#38，昇腾950 白皮书表3-1/表4-2）：`Ascend950PR`(32核主bin)/`Ascend950PR_C28` 与 `Ascend950DT`(36核主bin)/`Ascend950DT_C32`/`Ascend950DT_C28`，经 `hardware.SPECS` / `get_spec(name)` 索引，默认仍为 `ASCEND950PR`。
 
 ### 3.4 时延引擎统一在 timing.py
 
@@ -110,6 +112,6 @@ class Branch:
 当前六分支已全部实现（转Matmul / 特殊 / MergeBatch / IterBatch / StreamK / ASW_Basic）。后续方向：
 
 1. **转Matmul 精切**: 当前折叠后只粗估（`to_matmul.py::evaluate` 按 Matmul 总量），接入 MM 理论体系做精确切分。
-2. **换芯片**: 复制 `hardware/ascend950pr.py` 改常数，`NpuSpec` 接口不变。
+2. **换芯片**: `NpuSpec(name=..., ...)` 派生实例并注册进 `hardware.SPECS`（参考 `ascend950dt.py`），接口不变；默认规格恒为 `ASCEND950PR`。
-3. **标定 T_cmd**: 当前取 50ns 估计值，实测后改 `t_cmd_ns` 一处即可。
+3. **标定 T_cmd**: 当前按 0 处理 (未标定, issue#36; MergeBatch 合并收益已由 move_eff 搬移效率模型刻画, 不依赖 T_cmd 估计值), 实测后改 `t_cmd_ns` 一处即可。
 4. **新分支**: 在 `branches/` 下加一个文件实现三接口 + 在 `router.py` 注册 + 在 `evaluator._BRANCH_EVAL` 注册；约束一律走 `constraints.py`，不在分支内另造规则。

BMM/BMM_Theory/docs/02_分支理论/00_总纲_分支决策树.md

@@ -18,7 +18,9 @@ $$T_{total} = \max\big(T_{MMAD},\; T_{MTE2},\; T_{MTE1},\; T_{Fixpipe}\;[,\;T_{R
 
 case 固有算存比与 16bit 位宽平衡点：
 
-$$AI = \frac{2MN}{M+N},\qquad R_{16} = \frac{486\ \text{TFLOPS}}{1.6\ \text{TB/s}\,/\,2\ \text{B}} \approx 607.5$$
+$$AI = \frac{2MN}{M+N},\qquad R_{16} = \frac{432\ \text{TFLOPS}}{1.6\ \text{TB/s}\,/2\ \text{B}} = 540$$
+
+（Cube 算力为白皮书 Cube-only 口径 432 TFLOPS，issue#40；此前 v0.98 沿用总算力 486 → R₁₆=607.5。）
 
 AI < R₁₆ → 访存 Bound；反之计算 Bound。
 
@@ -50,7 +52,9 @@ case (B, M, N, K, dtype)
       │
       ├─ 单 batch M×N 太小, 合并搬移 ──▶ MergeBatch (冗余算力换搬移效率)
       │      进入: 5 条件 (L0C/搬移量/tile/算存比/b_core≥2b0)
-      │      胜出: K 截断 (k_L1=K) 且 b_core > b0(T_comp+T_write)/T_cmd
+      │      胜出: 命令节省(cmds差×T_cmd) + 搬移效率节省(合并 tile 放大 b0 倍,
+      │            issue#36) > drain 惩罚; K截断且效率打平时退化为
+      │            b_core > b0(T_comp+T_write)/T_cmd (v1.1 §4.5)
       │
       └─ 逐 batch 计算 ──────────────▶ IterBatch (无浪费, 最朴素)
              进入: b_core≥1 + 负载均衡 + L1 四形态之一 + 搬移效率
@@ -64,7 +68,7 @@ case (B, M, N, K, dtype)
       │
       ▼ P ≤ C/2 (B/M/N 都买不满)
    StreamK: 核间切 K + 归约
-      进入: P ≤ C/2, K/grid_K ≥ 256B/dtype, K > grid_K²/(grid_K-1)·θ_c (θ_c≈12)
+      进入: P ≤ C/2, K/grid_K ≥ 256B/dtype, K > grid_K²/(grid_K-1)·θ_c (θ_c≈10.9, issue#40 Cube-only 口径)
 ```
 
 ## 4. 重叠区仲裁

BMM/BMM_Theory/docs/02_分支理论/01_MergeBatch分支.md

@@ -43,24 +43,51 @@ k_L1 被 512B 截断省出的 L1 空间容纳更多 batch，提升 batch 间流
 
 ## 4. 时延模型（v1.1 §4）
 
-符号：T_load = k_L1(M+N)·dt/BW_pc（每 K 分块搬移）、T_comp = 2MN·k_L1/Q₁₆（每 K 分块计算）、T_write = MN·outB/W_GM（单 batch 写回）、T_cmd（单次 GM→L1 DMA 固定开销，~50ns）。
+符号：T_load = k_L1(M+N)·dt/(BW_pc·eff)（每 K 分块搬移，eff=move_eff(单命令tile)，见下）、T_comp = 2MN·k_L1/Q₁₆（每 K 分块计算）、T_write = MN·outB/W_GM（单 batch 写回）、T_cmd（单次 GM→L1 DMA 固定开销，**默认按 0**——未标定；issue#36 起合并收益由搬移效率模型刻画，不依赖 T_cmd 估计值）。
+
+> **搬移效率模型 (issue#36, 用户澄清)**: 单命令搬移效率由单块 tile = nValue×dValue×dt
+> 决定 —— eff(tile) = min(1, tile/min_TileSize)，达 16KB 饱和、之下线性退化 (与进入
+> 条件 4 的效率下限语义同源)。**合并 b0 个 batch 使单块 tile 放大 b0 倍** (A 侧
+> b0·M·k_L1^m·dt, B 侧 b0·N·k_L1^m·dt; 堆叠方向视转置/排布: A ND 非转置沿 M(nValue)
+> 堆叠, B ND 非转置沿 N(dValue) 堆叠, 乘积不变), 相对 IterBatch 逐 batch 搬移
+> (A 侧 tile = M·k_L1^iter·dt) 效率更高 —— **即便 T_cmd=0 也有收益**。
+> GM→L1 数据时延按两侧字节加权: t_data = (V_A/eff_A + V_B/eff_B)/W_GM;
+> GM 字节量不变 (仍 = V_in)。效率模型只接入切B 两分支 (逐命令 tile 小、效率差显著);
+> ASW/StreamK 单命令 tile 通常 ≥ 数百 KB 已饱和 (极端小 tile 形状有"效率降级"标注通道,
+> issue#34)。
 
 $$T_{mb} = \underbrace{\frac{b_{core}}{b_0}\cdot n_K^m\cdot(T_{load}^m + T_{cmd})}_{\text{搬移(合并)}} + \underbrace{b_0(T_{comp}+T_{write})}_{\text{末合并 batch drain}}$$
 
-两种情形：
+两种经典情形 (v1.1 §4.3/§4.4) 与第三情形 (issue#35):
 
-| 情形 | k_L1^m | n_K^m | 搬移命令数 | 结论 |
+| 情形 | k_L1^m | n_K^m | 每核搬移命令数 (⌈b_core/b₀⌉·n_K^m) | 结论 |
 |---|---|---|---|---|
-| **K 截断** (k_L1=K) | K 不减半 | 1 | b_core/b₀（少 b₀ 倍） | MergeBatch 可胜 |
+| **K 截断** (k_L1^m=K) | K | 1 | ⌈b_core/b₀⌉（比 IterBatch 少 b₀ 倍） | MergeBatch 可胜 |
-| **L1 绑定** (k_L1<K) | k_L1/b₀ | b₀·n_K | b_core·n_K（与 IterBatch 相同） | **MergeBatch 恒劣** |
+| **L1 绑定** (k_L1^m=k_L1^iter/b₀) | k_L1^iter/b₀ | b₀·n_K | b_core·n_K（与 IterBatch 相同） | **MergeBatch 恒劣** |
+| **dValue cap 截断** (k_L1^m=512B/dt, 文档二分未覆盖) | 512B/dt | ⌈K/k_L1^m⌉ | 按实际比较 (可与 IterBatch 打平或少) | 按泛化分界判定 |
 
-L1 绑定情形搬移次数、单次搬移量都与 IterBatch 相同，只放大 drain——证明见 v1.1 §4.4。
+L1 绑定理想情形搬移次数、单次搬移量都与 IterBatch 相同，只放大 drain——证明见 v1.1 §4.4。
+第三情形例: IterBatch 走 a/b 形态 (k_L1=K) 而合并侧被 512B 推荐值截断时, n_K^m=n_K,
+命令数打平 (既不省 b₀ 倍也不放大), 合并仅剩 drain 惩罚 -> 恒劣; 反之若 IterBatch 走
+c/d 形态切 K 而合并侧每核命令数更少, 则按实际节省判定。
+
+> **每核命令数口径 (issue#35)**: 真实命令数 = **合并组数 × 每组 K 段数** =
+> ⌈b_core/b₀⌉·⌈K/k_L1^m⌉。注意 v1.1 §4.4 "命令数与 IterBatch 相同 = b_core·n_K" 中的
+> n_K 是**未合并**粒度 (⌈K/k_L1^iter⌉); 误代入合并后段数 ⌈K/k_L1^m⌉ 会多计 b₀ 倍
+> (修复前 evaluate 即此错, L1 绑定情形命令时延虚高 b₀ 倍, 可把仲裁方向翻错)。
+
+> **GM 数据量口径 (issue#31)**: 各 (合并组, K段) 的数据互不重叠, 每个输入字节恰好从 GM
+> 读一次 —— K 截断与 L1 绑定两种情形的**芯片 GM 读取量都 = V_in** (末段按实际剩余计,
+> 无 padding 上取)。上式的 T_load 级联只用于刻画命令/双缓冲调度结构: 数据时延按
+> V_in/W_GM 计, n_K 只放大 DMA 命令项 ⌈b_core/b₀⌉·n_K^m × T_cmd (issue#35 口径)。
 
 ## 5. MergeBatch vs IterBatch 净收益
 
-$$\text{净收益} = \underbrace{b_{core}\Big(1-\frac{1}{b_0}\Big)T_{cmd}}_{\text{搬移命令节省}} - \underbrace{(b_0-1)(T_{comp}+T_{write})}_{\text{drain 惩罚}}$$
+泛化分界 (issue#35 命令数 / issue#36 搬移效率, 覆盖三种情形): 直接比较两分支**实际每核 DMA 命令数**与**效率加权搬移时延** ——
 
-大 B（b_core 大）且小 MN（T_comp 小）时 MergeBatch 最优。T_cmd 的物理成因：Nd2Nz 描述符配置（7 字段写 DMA 寄存器）+ 地址生成 + 突发启动 + L1 同步握手。
+$$\text{净收益} = \underbrace{(cmds_{iter}-cmds_{mb})\,T_{cmd}}_{\text{命令节省}} + \underbrace{(t_{data}^{iter}-t_{data}^{mb})}_{\text{效率节省}} - \underbrace{(b_0-1)(T_{comp}+T_{write})}_{\text{drain 惩罚}},\qquad \begin{array}{l}cmds_{iter}=b_{core}\lceil K/k_{L1}^{iter}\rceil\\ cmds_{mb}=\lceil b_{core}/b_0\rceil\lceil K/k_{L1}^m\rceil\end{array}$$
+
+其中效率节省来自合并 tile 放大 (t_data 按 eff(tile) 加权, 见 §4)。**T_cmd=0 时命令节省为 0, 效率节省依然在项** —— MergeBatch 凭搬移效率胜出 (用户澄清口径)。K 截断且效率打平 (两侧 tile 均 ≥16KB 饱和) 且 T_cmd>0 时严格退化为 v1.1 §4.5 闭式 b_core > b₀(T_comp+T_write)/T_cmd；L1 绑定理想情形 (k_L1^m=k_L1^iter/b₀ 且 tile 均饱和) 命令数与效率均打平、净收益恒负。大 B（b_core 大）且小 MN（T_comp 小、IterBatch 单 batch tile 小未饱和）时 MergeBatch 最优。T_cmd 的物理成因：Nd2Nz 描述符配置（7 字段写 DMA 寄存器）+ 地址生成 + 突发启动 + L1 同步握手 (待 msProf 标定后恢复非零取值)。
 
 ## 6. 与源码的差异（v1.1 §5.1）
 

BMM/BMM_Theory/docs/02_分支理论/02_IterBatch分支.md

@@ -40,9 +40,17 @@ $$T_{iter} = \underbrace{b_{core}\cdot n_K\cdot(T_{load} + T_{cmd})}_{\text{搬
 
 每 batch 的 A[M,K]+B[K,N] 作为独立 DMA 操作搬入 L1（源码 `ndNum = curIterBatchL1`，多块独立寻址）。
 
+> **GM 数据量口径 (issue#31)**: 形态 a/b/c/d 的 K 段/驻留侧数据互不重叠、每个输入字节
+> 恰好从 GM 读一次 (形态 c 的驻留侧每 batch 只搬一次; 切 K 末段按实际剩余计, 无 padding
+> 上取) —— **芯片 GM 读取量 = V_in**, 与 L2 容量无关 (IterBatch 无 L2 级重复读)。
+> n_K 只决定 DMA 命令数 (T_cmd 项, 950PR 默认 t_cmd=0) 与双缓冲调度, 不放大数据量;
+> 数据时延按**效率加权** (V_A/eff_A + V_B/eff_B)/W_GM 计 (issue#36: 单命令 tile =
+> nValue×dValue×dt 越小有效带宽越低, 达 min_TileSize=16KB 饱和; 分侧口径 —— a/b 形态
+> 双侧整 K、c 形态驻留侧整 K + 对侧 k_L1 分块、d 形态双侧 k_L1 分块)。
+
 ## 6. 与 MergeBatch 的分界
 
-见 [01_MergeBatch分支.md](01_MergeBatch分支.md) §4~5。IterBatch 在 **L1 绑定情形恒优**；MergeBatch 仅在 K 截断且 b_core 足够大时胜。
+见 [01_MergeBatch分支.md](01_MergeBatch分支.md) §4~5。IterBatch 在 **L1 绑定理想情形恒优**；MergeBatch 凭 命令节省 (T_cmd>0 时) 与 **搬移效率节省** (合并 tile 放大 b0 倍, T_cmd=0 时仍成立, issue#36) 胜出 —— 净收益 > 0 即 MergeBatch 优。
 
 ## 7. 与源码的差异（v1.1 §5.2）
 

BMM/BMM_Theory/docs/02_分支理论/04_特殊分支.md

@@ -21,13 +21,13 @@ K 维度是 Cube（MMAD）存在的意义——`C = Σ_k A[..,k]·B[k,..]` 的
 - 数据流：GM→UB（读 A、B）→ Mul → GM（写 C），全程 AIV；
 - 时延：`T = max(搬入, 搬出)`，AIV 算力远剩，瓶颈在搬移：`T ≈ B·(MK + KN + MN)·dt / W_GM`；
 - **触发条件**：`B ≥ 2×AIV核数 = 128`（开 UB 乒乓需要每核至少 2 个 batch 块）且单 batch 输入输出能驻留 UB；
-- **B < 128 时并不无解**（issue#12）：退化为 **AIV 单缓冲**——无乒乓、逐 batch 串行搬入计算，仍远优于 Cube 通路（K=1 时 Cube 16×16×16 浪费 15/16），只是流水掩盖能力下降。软件 `special.py` 按 `B ≥ 128` 自动选择"UB乒乓 / AIV单缓冲"模式。
+- **B < 128 时并不无解**（issue#12/#17）：退化为 **AIV 单缓冲**——无乒乓、逐 batch 串行搬入计算，仍远优于 Cube 通路（K=1 时 Cube 16×16×16 浪费 15/16），只是流水掩盖能力下降。软件 `special.py` 按 `B ≥ 128` 自动选择"UB乒乓 / AIV单缓冲"模式。
 
 ## 4. 软件处理
 
 `router.py` 前置归约中，`k ≤ 1` 直接路由到特殊分支：
 
 - `K=0` → 标注"纯 AIV 写值"，评估时延 = 写出时延；
-- `K=1` → 标注"AIV 逐元素乘"，评估时延 = 搬入/搬出较大者。
+- `K=1` → 标注"AIV 逐元素乘"，评估时延 = 搬入/搬出较大者；plan 按 `B ≥ 128` 自动选择"UB乒乓 / AIV单缓冲"模式（`l1_form` 与 `note` 中可见）。
 
 不进入 Cube 切分体系，无 ImplPlan 的 tile 字段（除 used_core_num = AIV 核数外均不适用）。

BMM/BMM_Theory/docs/02_分支理论/05_StreamK分支.md

@@ -16,7 +16,7 @@
 |---|---|---|
 | 1 | `P ≤ C/2` | **并行缺口**：不切 K 时至多一半核有事做，K 是唯一剩余的并行维度。阈值取 C/2 而非 C：StreamK 定义即 grid_K≥2（至少 2 路切 K），grid_K=2 时每块需 2 核，总核数需求 = ⌈P⌉×2 ≤ C |
 | 2 | `K/grid_K ≥ 256B/dtype` | **单核 K 段下限**：每核 K 段内轴连续长度不小于 dValue 推荐值，保证段内搬移效率不崩。grid_K = ⌊C/⌈P⌉⌋ |
-| 3 | `K > grid_K²/(grid_K-1)·θ_c`，θ_c≈12 | **归约代价可接受**：切 K 后总时延须小于不切 K（降核 ASW），推导见 §3 |
+| 3 | `K > grid_K²/(grid_K-1)·θ_c`，θ_c≈10.9 | **归约代价可接受**：切 K 后总时延须小于不切 K（降核 ASW），推导见 §3 |
 | 4 | 确定性等级 ≤ 1 且 ND 格式 | 归约顺序不定引入浮点非确定性，确定性等级 2/3 的业务禁用 |
 
 ## 3. 归约代价与 θ_c 推导（条件 3 的来源）
@@ -34,13 +34,13 @@ $$T_{Reduce}^t = \underbrace{\frac{grid_K\cdot M^t N^t\cdot 4B}{W_{L2}}}_{\text{
 
 **计算 Bound 情形**（T_pipe = T_MMAD^t）代入判据，两边除以 L0C/4B 后 **tile 尺寸消去**——K 阈值不依赖 M/N 具体值：
 
-$$K > \frac{grid_K^2}{grid_K-1}\cdot\theta_c,\qquad \theta_c = \frac{Q_{16}}{2}\Big(\frac{8B}{W_{L2}}+\frac{1}{Q_{AIV}}\Big) \approx 12$$
+$$K > \frac{grid_K^2}{grid_K-1}\cdot\theta_c,\qquad \theta_c = \frac{Q_{16}}{2}\Big(\frac{8B}{W_{L2}}+\frac{1}{Q_{AIV}}\Big) \approx 10.9$$
 
-代入数值：grid_K=2 → K>49；grid_K=4 → K>66；grid_K=8 → K>112。L2 读写（1.54 ps/元素）是主导项，AIV 求和仅占 5%。
+（Q16 为 Cube-only 口径 13.5T，issue#40；旧总算力口径 15.1875T 时 θ_c≈12。）代入数值：grid_K=2 → K>44；grid_K=4 → K>58；grid_K=8 → K>100。L2 读写（1.54 ps/元素）是主导项，AIV 求和仅占 5%。
 
-访存 Bound 情形阈值恒低于计算 Bound（θ_m·[M^tN^t/(M^t+N^t)]/θ_c ≤ 128/304 ≈ 0.42 < 1），故**汇总条件取计算 Bound 阈值**（保守，覆盖两种情形）。
+访存 Bound 情形阈值恒低于计算 Bound（θ_m·[M^tN^t/(M^t+N^t)]/θ_c ≤ 128/270 ≈ 0.47 < 1），故**汇总条件取计算 Bound 阈值**（保守，覆盖两种情形）。
 
-> **注意**：θ_c 对 workspace 落点敏感——部分和落 GM 时读写带宽从 5.2TB/s 降到 ~0.64TB/s，θ_c 升至约 97。**设计时应优先保证 workspace 驻留 L2**。
+> **注意**：θ_c 对 workspace 落点敏感——部分和落 GM 时读写带宽从 5.2TB/s 降到 ~0.64TB/s，θ_c 升至约 86。**设计时应优先保证 workspace 驻留 L2**。
 
 ## 4. 实现方案
 
@@ -70,4 +70,4 @@ $$T_{SK} = \max\Big(\frac{T_{MMAD}}{grid_K},\; \frac{T_{MTE2}}{grid_K}\Big) + T_
 - `8192` = C×512B = dValue 推荐值在最大 grid_K=C 下的保障（同为条件 2，取推荐值）；
 - max 取更严的 8192。
 
-修正后的归约阈值（θ_c≈12，grid_K=32 时 K>396）远低于 8192，说明 **8192 的绑定约束是 dValue（条件 2），不是归约代价（条件 3）**。源码不动态算 grid_K，用固定阈值保守合并两条条件。本软件按理论动态计算 grid_K 与 θ_c，进入条件更精细。
+修正后的归约阈值（θ_c≈10.9，grid_K=32 时 K>360）远低于 8192，说明 **8192 的绑定约束是 dValue（条件 2），不是归约代价（条件 3）**。源码不动态算 grid_K，用固定阈值保守合并两条条件。本软件按理论动态计算 grid_K 与 θ_c，进入条件更精细。

BMM/BMM_Theory/docs/02_分支理论/06_ASW_Basic分支.md

@@ -1,8 +1,9 @@
 # ASW_Basic 分支理论（含尾轮处理）
 
-> 整理自《BMM算子优化分析 v0.98》§八 + 《BMM尾轮处理策略对比分析 v1.5》.
+> 整理自《BMM算子优化分析 v0.98》§八 + 《ASW_Basic分支分析 v1.91》 + 《BMM尾轮处理策略对比分析 v1.5》.
 > 尾轮策略的完整推导见 [07_尾轮处理策略.md](07_尾轮处理策略.md), 本文将其结论**内化为 ASW_Basic 最优实现的必要环节**.
 > 对应软件实现: `bmm_theory/branches/asw_basic.py`.
+> 2026-09 更新: Step0/Step1 按 v1.91 §5.1/§5.2 重写为"UnitFlag 单缓冲方形 256 + 有界枚举最小搬入" (issue#33)。
 
 ## 1. 定位：兜底分支，实践中最常命中
 
@@ -20,15 +21,50 @@ ASW_Basic 是 BMM 的**兜底分支**——核间切 M/N（或混合切），不
 
 ## 3. 时延建模
 
-核间切 M/N（不切 K），每核处理若干 [singleCoreM, singleCoreN] 输出块：
+核间切 M/N（不切 K），每核处理若干 [singleCoreM, singleCoreN] 输出块；单 batch 内
+A 行块被 n_cnt 个列 tile 共享、B 列块被 m_cnt 个行 tile 共享（issue#24 用户口径，
+最新场景口径见 issue#32 与 docs/05 §4）：
 
 $$T = \max(T_{MTE2},\; T_{MMAD},\; T_{FIX}) + T_{drain}$$
 
-其中 `T_MMAD = 2BMNK/(C·Q16)`、`T_FIX = B·MN·outB/(C·W_pc)` 与分配策略无关——**分配策略只影响 T_MTE2**（通过 L2 命中率影响有效 GM 带宽）：
+- `T_MMAD = 2BMNK / (used·Q16(dtype))` —— 与分配策略无关；Q16 按输入 dtype 分档 (issue#28)；
+- **MTE2 两段搬移链**（issue#23/#31/#32，GM 为读写共享总线）:
+  - `t_gm = S_in^GM / W_GM`：GM 直读量（首读 + 落空重读；与输出直写 GM 并发时
+    读写累加，输出落点见 docs/05 §4.2 R4）；
+  - `t_l2 = S_in^L2 / W_L2`：L2 重复读量（共享块驻留后其余次读取，5.2TB/s 读口独享）；
+  - MTE2 链 = t_gm + t_l2 + DMA 命令开销。
+- `T_FIX`：R4 输出落点——整 case 输入+输出 ≤ L2 时写 L2 写口（异步回写不占算子时延），
+  否则直写 GM 计入共享总线（issue#30）。
 
-$$T_{MTE2} = \frac{S_{in}^{GM}}{C\cdot BW_{pc}},\qquad S_{in}^{GM} = r_{in}\cdot S_{in}$$
+**GM/L2 流量归属 —— L2 场景判定**（与代码 `_l2_scene` 同源，docs/05 §4.1/§4.3）:
+S_A 整case全驻留 / S_B 单batch可驻留（全驻留**或单侧全驻留 + 对侧滑窗**）→
+GM = V_in 一次（r_in=1），共享块重复读全部命中 L2；S_C 双侧均超 L2 → 最小替换 2D 分组
+（整 L2 容量约束下最小 GM），组间共享块落空回 GM、组内窗口复用按 L2 计账。
 
-`r_in` = 重复读倍率（GM 输入流量/输入总量），r_in≥1，下界 1 表示每字节只从 GM 读一次（后续复用全命中 L2）。**分配策略的全部目标就是让 r_in 尽量接近 1**。
+`r_in` = GM 输入流量/输入总量，r_in≥1，下界 1 表示每字节只从 GM 读一次（后续复用
+全命中 L2）。**分配策略的全部目标就是让 r_in 尽量接近 1**：S_B 已识别"单侧可全驻留"
+调度（不再把可驻留 case 错误地按分组放大 GM）；只有双侧都放不下 L2 时（S_C）r_in 才 > 1。
+
+**尾轮残余 T_drain（闭式, issue#37）**：均匀分块下稳态聚合 ≡ `(n_wave−1+ρ)·T_block`
+（`N_blk/used = n_wave−1+ρ` 恒等），`T_drain` 只计尾轮结构相对稳态的残余；块级三段
+时延与 `_decide_tail` 主导项判定同源（L2 命中口径）：
+
+| 策略 / 主导项 | T_drain | 出处 |
+|---|---|---|
+| r = 0 | 0 | — |
+| A0（r>0，尾轮 r 核各 1 整块、C−r 核空转） | `(1−ρ)·T_block` | v1.5 §3.3 `T_A0 = n_wave·T_block` |
+| 面积型 A1b / 方案B | 0 | v1.5 §4.3 总量守恒严格相等 |
+| 周长型 A1b | `(√ρ−ρ)·T_load` | v1.5 §5 `T = (n_wave−1+√ρ)·T_load` |
+| 周长型 方案B | `(√(n_wave(n_wave−1+ρ)) − (n_wave−1+ρ))·T_load` | v1.5 §5 `T = √(n_wave(n_wave−1+ρ))·T_load` |
+
+边界说明（**不计入模型**的两项, issue#37 决议仅标注）:
+
+- **首块填充/末块排空**：v1.91 §3.1 的 `T_drain = O(T_comp+T_write)` 在 UnitFlag
+  16×16×16 细粒度流水下真实暴露为 granule 级小量（首 k 段搬入 + 末 granule 排空），
+  量级未标定，不入模型——大 n_wave 时相对误差 O(1/n_wave)；
+- **尾轮重切的流量放大**：方案B（全局 tile 缩 1/√g）与 A1b（尾轮区缩 √ρ）会真实
+  增加搬移总量（周长和 ×√g / ×1/√ρ），字节列仍按主 tile 几何计账；仅影响周长型
+  主导角区的精度（面积型 MTE2 非瓶颈，无影响）。
 
 ## 4. 核间分配策略：B→M→N 线性映射
 
@@ -42,22 +78,50 @@ $$T_{MTE2} = \frac{S_{in}^{GM}}{C\cdot BW_{pc}},\qquad S_{in}^{GM} = r_{in}\cdot
 
 ## 5. 实现方案
 
-**Step 0: BaseM/BaseN**（L0 级 tile，先把 L0C 用满）
+**Step 0: BaseM/BaseN**（L0 级 tile, 先按 UnitFlag 单缓冲把 L0C 用满——v1.91 §5.1）
 
-$$\text{BaseM}\times\text{BaseN} = \frac{L0C}{2\times 4B} = 32768\ \text{元素}$$
+Fixpipe 的 UnitFlag 提供 **tile 内部 16×16×16 细粒度流水**, 取代 tile 间粗粒度双缓冲 →
+L0C 只需一份 buffer:
 
-双缓冲两份、FP32 4B/元素。长宽比跟随 SingleCoreM/N，对齐 16。baseK = min(L0A/(2·BaseM·dt), L0B/(2·BaseN·dt)) 向下 16 对齐（L1→L0 无 dValue 要求）。
+$$\text{BaseM}\times\text{BaseN} = \frac{L0C}{4\text{B}} = 65536\ \text{元素（UnitFlag 单缓冲）}$$
 
-**Step 1: SingleCoreM/N**（每核输出 tile，≥ BaseM/N）
+长宽比**方形优先**（而非跟随 M/N——跟随不会改善 GM→L1 搬移, 只会使 L0A/L0B 容量利用率
+失衡、baseK 减半）: 默认 BaseM = BaseN = 256（L0A/L0B 同时装满）; **例外**: M 或 N 小于
+方形边长时被迫跟随: BaseM = min(256, ⌊M⌋₁₆), 另一维 = min(65536/BaseM, N) 向下 16 对齐
+（例：M=128, N=4096 → BaseM=128, BaseN=512）。
 
-不受 L0 容量直接约束（内部由若干 BaseM×BaseN L0 tile 组成）。核心影响 **GM→L1 搬移效率和 L2 重复读率**：
+$$\text{baseK} = \min\Big(\frac{L0A}{2 \cdot \text{BaseM} \cdot dtype},\; \frac{L0B}{2 \cdot \text{BaseN} \cdot dtype}\Big) \text{ 向下 16 对齐}$$
 
-- 约束 1 并行度：`mCnt×nCnt ≥ ⌈C/B⌉`；
+（L0A/L0B 仍开双缓冲; 16 对齐是 Cube K 向粒度要求, L1→L0 搬移无 dValue 要求。）
-- 约束 2 L1 容量：`2(sM+sN)·k_L1·dt ≤ L1`，`k_L1·dt ≥ 256B`；
-- 约束 3 搬移效率：`sM·k_L1·dt ≥ 16KB`，`k_L1·sN·dt ≥ 16KB`；
-- 约束 4：SingleCoreM/N 是 BaseM/N 的整数倍。
 
-选取策略：满足约束 1 前提下 SingleCoreM/N 尽量大，长宽比跟随 M/N。
+**Step 1: SingleCoreM/N + k_L1 + mCnt/nCnt**（有界枚举——v1.91 §5.2, 与尾轮 v1.5 §2.1
+修正口径一致）
+
+SingleCoreM/N **不受 L0 容量直接约束**（内部由若干 BaseM×BaseN L0 tile 组成）, 核心影响
+GM→L1 搬移效率与 L2 重复读率:
+
+- **P = ⌈C/B⌉**（最少切分块数）;
+- **P=1 (B ≥ C)**: 先试不切分 (mCnt=nCnt=1, tile 跟随 M/N); 约束不满足则强制切分, 进入枚举;
+- **枚举空间**（有界, host 端遍历）: mCnt ∈ [1, ⌈M/BaseM⌉], nCnt ∈ [1, ⌈N/BaseN⌉],
+  且 B·mCnt·nCnt ≥ C（约束 1 并行度）;
+- 每候选: sM = align_up(⌈M/mCnt⌉, BaseM)（约束 4: BaseM 整数倍）, sN 同理;
+- **约束 2**（L1 双缓冲）: k_L1 = min(K, ⌊L1/(2(sM+sN)·dtype)⌋₁₆)（k_L1 与形状耦合）;
+- **约束 3**（搬移效率, 转置感知）: dValue ≥ 256B——A 非转置 k_L1·dt / A 转置 sM·dt;
+  B 非转置 sN·dt / B 转置 k_L1·dt; 且单次搬移量 sM·k_L1·dt 与 k_L1·sN·dt ≥ 16KB;
+- **目标**: 每 batch 总搬入 `K·dt·(nCnt·M + mCnt·N)` 最小（v1.5 §2.1 修正: 稳态流水下
+  k_L1 约掉——单块搬入 = K(sM+sN)·dt 与分次粒度无关, k_L1 只进约束）;
+  **并列时取 r = B·mCnt·nCnt mod C 最大**（尾轮块越多, 重切越省）。
+
+mCnt = ⌈M/singleCoreM⌉, nCnt = ⌈N/singleCoreN⌉（按实际 tile 反推）。
+
+> **兜底分支恒出方案 (issue#34)**：枚举无可行候选时，不再产出"违规/不可行"方案——
+> 先放开约束 4（Base 整数倍）按 16 对齐网格 + dValue 128B 硬下限再搜；仍无解
+> （极端形状，如 N=8 int8 大 K 时 B 侧 dValue = sN·dt 物理上不可能 ≥ 下限）时退回
+> Base tile 并标注**效率降级**（warning，搬移效率低于模型假设、时延可能低估，
+> 建议调 dtype/布局），**不是违规/不可行**：dValue/minTile 是效率下限（DMA 仍能工作），
+> 真正不可行的只有容量/核数硬约束（L0C/L0A/L0B/L1 超容、used_core_num 超核数）。
+> 对比：IterBatch/MergeBatch/StreamK 等**有替代分支**的分支仍按违规处理
+> （不满足条件不该进，由路由另择或落 ASW）。
 
 **Step 2: mCnt/nCnt 与核间分配**：`mCnt=⌈M/sM⌉`，`nCnt=⌈N/sN⌉`，总块数 B·mCnt·nCnt，B→M→N 线性映射。
 
@@ -67,19 +131,21 @@ $$\text{BaseM}\times\text{BaseN} = \frac{L0C}{2\times 4B} = 32768\ \text{元素}
 
 同一波 C 个核所需的 A 行块 + B 列块集合是"活跃工作集"，超 L2 就回 GM 读。swizzle 编排输出块执行顺序，把每一波的活跃工作集压到最小。窗口宽度 `W = max{d | d|C, d ≤ ⌊√C⌋}`（C=32 时 W=4）；窗口内先扫 M、扫满 W 行进下一列 N，奇数窗口行 N 向反向（蛇形）。
 
-**Step 5: L2 分组（工作集超 L2 时）**
+**Step 5: L2 场景判定与分组（make_plan/evaluate 共用 `_l2_scene`，issue#24/#32）**
 
-记 S_in = B(MK+KN)·dt，S_out = B·MN·outB。两个不变量：GM 流量下界 = S_in + S_out；L2 读入可用空间 L2_read = L2 − S_out^resident。
+记单 batch 输入 a_b = MK·dt、b_b = KN·dt，整 case V_in = S_in、V_out = S_out；
+A 行块 block_a = a_b/m_cnt、B 列块 block_b = b_b/n_cnt。
 
-| 场景 | 条件 | 策略 |
+| 场景 | 条件 | GM / L2 计账（每 batch） |
 |---|---|---|
-| A 全驻留 | S_in + S_out ≤ L2 | 输入读一遍（r_in=1），输出驻留 L2 异步回写，无需切分 |
+| S_A 整case全驻留 | V_in + V_out ≤ L2 | GM = a_b+b_b；重复读 (n_cnt−1)a_b+(m_cnt−1)b_b 走 L2；输出驻留 L2（GM 写=0） |
-| B 输入能驻留，加输出超了 | S_in ≤ L2 < S_in+S_out | **输入驻留、输出直写 GM**，保住 r_in=1；校验总线 (S_in+S_out)/T_MMAD ≤ W_GM |
+| S_B 单batch可驻留 | a_b+b_b ≤ L2，**或**单侧全驻留+对侧滑窗：b_b+2·block_a ≤ L2 或 a_b+2·block_b ≤ L2 | GM = a_b+b_b（r_in=1）；重复读 (n_cnt−1)a_b+(m_cnt−1)b_b 全部命中 L2；输出直写 GM |
-| C 输入本身超 | S_in > L2 | 分组执行，每组输入工作集 ≤ L2，输出直写 GM |
+| S_C 双侧超 L2 | 以上均不满足 | 容量 m_grp·block_a+n_grp·block_b ≤ L2 下最小化 GM = ⌈n_cnt/n_grp⌉·a_b + ⌈m_cnt/m_grp⌉·b_b；组间落空计 GM，窗口 (n_cnt−⌈n_cnt/n_grp⌉)a_b+(m_cnt−⌈m_cnt/m_grp⌉)b_b 计 L2 |
 
-场景 C 分组：每组覆盖 M 向 m_grp、N 向 n_grp 个基本块，`m_grp = ⌊D/(2·sM)⌋`，`n_grp = ⌊D/(2·sN)⌋`，D = L2/(B·K·dt)；重复读倍率 `r_in = (n_grp·M + m_grp·N)/(M+N)`。块内错位分核（对角线分配）避免同地址并发读串行化。
+> 注：早期口径（预算 D = L2/(B·K·dt) 且对半切、组内窗口流量零计）已废弃
+> （issue#24 除总 B 的矛盾、issue#32 对半预算与窗口零计失真）；分组不再对半、窗口复用如实计 L2。
 
-**Step 6: 核内 tiling**：BaseM×BaseN×4B×DB ≤ L0C；BaseM×k_L0×dt×2 ≤ L0A；k_L0×BaseN×dt×2 ≤ L0B；内轴按 dValue 256B/512B 对齐。
+**Step 6: 核内 tiling**：BaseM×BaseN×4B ≤ L0C（UnitFlag 单缓冲, 见 Step 0）；BaseM×k_L0×dt×2 ≤ L0A；k_L0×BaseN×dt×2 ≤ L0B；内轴按 dValue 256B/512B 对齐。
 
 **Step 7: 内部特化**：单边无 batch 且该侧矩阵小时小侧整个常驻 L1 只搬一次。
 
@@ -106,7 +172,7 @@ $$\text{BaseM}\times\text{BaseN} = \frac{L0C}{2\times 4B} = 32768\ \text{元素}
 
 - A1b 与方案 B **理论时延严格相等**（总量守恒 `T = c·B·M·N/C`）；
 - **默认选方案 B（工程简洁：一套 tile、无尾轮分支）**；追求搬移下限选 A1b（周长和恒 ≤ 方案 B，均值不等式）；
-- r 小（ρ < (187/s)²）时 A1b 尾轮翻出为周长型，**方案 B 微优 ~6%**；
+- r 小（ρ < (166/s)²，Cube-only 口径下由 187 修正，issue#40）时 A1b 尾轮翻出为周长型，**方案 B 微优 ~6%**；
 - A1a 是 A1b 真子集，仅在 r | C 且 s*=⌊C/r⌋ 完美时打平，工程上可不单列。
 
 **边角场景（周长型主导，小 tile 或工作集超 L2 的 GM 直读）**：

BMM/BMM_Theory/docs/02_分支理论/07_尾轮处理策略.md

@@ -17,12 +17,14 @@ $$T_{MMAD} = \frac{2\cdot sM\cdot sN\cdot K}{Q_{16}},\qquad T_{MTE2} = \frac{K(s
 | 对比 | 判据 | 决定因素 |
 |---|---|---|
 | MMAD vs MTE2 | `sM·sN/(sM+sN)` ⋛ `dt·Q16/(2·BW_eff)` | **tile 尺寸**（K 约掉） |
-| MMAD vs FIX | `K` ⋛ `outB·Q16/(2·BW_pc)` ≈ 304 | **K** |
+| MMAD vs FIX | `K` ⋛ `outB·Q16/(2·BW_pc)` ≈ 270 | **K** |
+
+（Q16 为 Cube-only 口径 13.5T（issue#40）：MMAD/MTE2 分界 93.5→**83.1**，MMAD/FIX 分界 304→**270**。）
 
 **缩放类型二分**：MMAD 与 FIX 都 ∝ 面积 sM·sN，MTE2 ∝ 周长 (sM+sN)。
 
-- **面积型主导**（MMAD 或 FIX 最大）：tile 大（sM·sN/(sM+sN)≥93.5，L2 命中）且（K≥304 时 MMAD、K<304 时 FIX）——**主流 prefill/decode case 均属此类**；
+- **面积型主导**（MMAD 或 FIX 最大）：tile 大（sM·sN/(sM+sN)≥83.1，L2 命中）且（K≥270 时 MMAD、K<270 时 FIX）——**主流 prefill/decode case 均属此类**；
-- **周长型主导**（MTE2 最大）：tile 小（<93.5，L2 命中）或工作集超 L2 的 GM 直读（<304）——边角 case。
+- **周长型主导**（MTE2 最大）：tile 小（<83.1，L2 命中）或工作集超 L2 的 GM 直读（<270）——边角 case。
 
 ## 2. 四种策略定义
 
@@ -54,7 +56,7 @@ $$\big(n_{wave}-1+\sqrt{\rho}\big)^2 \le n_{wave}(n_{wave}-1+\rho) \iff (\sqrt{\
 
 **A1b 周长和恒 ≤ 方案 B**（搬入总量少、L2 重复读少、掩盖余量大）。离散 16 对齐后时延互有胜负（<3%，数值依赖、无系统性方向）。
 
-**尾轮翻出修正**（§7.2.4）：面积型主导但 r 小（ρ < (187/s)²）时 A1b 尾轮 tile 缩得太小，周长/面积比上升使主导项翻转为周长型（搬入翻出），方案 B 的全局 tile 缩得温和（1/√g > √ρ 恒成立）不翻出——**r 小的面积型 case 方案 B 可微优 ~6%**。
+**尾轮翻出修正**（§7.2.4）：面积型主导但 r 小（ρ < (166/s)²，Cube-only 口径下 187→166，issue#40）时 A1b 尾轮 tile 缩得太小，周长/面积比上升使主导项翻转为周长型（搬入翻出），方案 B 的全局 tile 缩得温和（1/√g > √ρ 恒成立）不翻出——**r 小的面积型 case 方案 B 可微优 ~6%**。
 
 ## 4. 周长型主导（边角场景）
 
@@ -79,7 +81,7 @@ $$\big(n_{wave}-1+\sqrt{\rho}\big)^2 \le n_{wave}(n_{wave}-1+\rho) \iff (\sqrt{\
 | 1 | r = 0 | A0 | n_wave·T_block |
 | 2 | 面积型，0<r≤C/2，无翻出 | A1a，s*=⌊C/r⌋ | T_block(n_wave−1+1/s*) |
 | 3 | 面积型，r>C/2，无翻出 | A1b 或方案 B（严格打平） | T_block(n_wave−1+ρ) |
-| 4 | 面积型，ρ<(187/s)²（翻出） | 方案 B | T_block·n_wave/g |
+| 4 | 面积型，ρ<(166/s)²（翻出，issue#40 口径） | 方案 B | T_block·n_wave/g |
 | 5 | 周长型，ρ≥ρ_dv，B 可行 | A1b | T_load(n_wave−1+√ρ) |
 | 6 | 周长型，ρ<ρ_dv 且分界成立且 B 可行 | 方案 B | T_load·n_wave/√g |
 | 7 | 周长型，广义 B 损失>0 | A1b（恒不劣） | 广义枚举 min w×T_block |

BMM/BMM_Theory/docs/03_测评报告/BMM_Theory软件测评报告_v2.0.md

@@ -0,0 +1,213 @@
+# BMM/BMM_Theory 理论最优实现分析软件 测评报告（第二轮 · 整改后复评）
+
+- 测评对象：`git.magicnetworld.com/admin/matmul-analysis` 仓库 `BMM/BMM_Theory`（软件包 `bmm_theory`）
+- 本轮基准：HEAD `99a25a6b42efa63689f21ae53a7cc467c2fca79f`（main，2026-09-03），相对上轮基准 `d36e224` 共 **33 个整改提交**
+- 测评方式：整改 diff 逐文件审查 + issue 逐条复测 + 全量运行/单测 + 同分布压力回归（与上轮同种子同分布对比）+ 口径深查
+- 说明：上轮报告为 `docs/03_测评报告/BMM_Theory软件测评报告_v1.0.md`（2026-09-03 入库，commit d36e224）；本轮为整改后的复评。
+
+---
+
+## 0. 结论摘要（TL;DR）
+
+**整改诚意与质量都很好：上轮提出的 7 个 issue（#4–#10）中 6 个实质闭环，1 个（#9 StreamK 计账口径）修复方向正确但引入新疑点，需二次复核。** 软件从"可跑但自相矛盾"升级为"可自检、自洽率大幅提升"的可信工具雏形。
+
+关键数字对比：
+
+| 指标 | v1（上轮） | v2（本轮） |
+|---|---|---|
+| 单元测试 | 17/17 | **23/23**（新增 6 条 issue 回归） |
+| P0 崩溃（K=1 且 B<128） | 必崩 AttributeError | **不崩溃**，占位标注"[无方案]" |
+| 典型 LLM 形状自检不可行率（6000 例） | **18.0%**（1079 例） | **0.0%** |
+| 宽范围随机自检不可行率（8000 例） | **6.0%**（480 例） | **0.2%**（16 例，全部为极端瘦长 MergeBatch 且自检已显式标注） |
+| 生成结果可复现性（vs 仓库固化示例） | 0 差异 | 0 差异 |
+| 推荐→评估回灌 feasible | 94% | **99.8%**（16/8000 标注违规但仍推荐） |
+| NaN/负时延/崩溃 | 无（P0 区除外） | 无 |
+
+**遗留问题（详见 §5）**：
+1. 【P2，建议跟进】issue#9 的修复引入新口径问题：StreamK 部分和写出被**双重计账且按单核串行假设高估 32 倍**——demo case 总时延从 9.96us 变为 55.03us、瓶颈变成 FIXPIPE（51.6us），而同一批字节在 `t_reduce` 内仅计 1.61us。该数字已被作者固化进 examples，建议复核。
+2. 【P3】K=1 且 B<128 区域仍**无任何理论方案**（占 K=1 域的约一半），现以"无方案占位"标注兜底（崩溃已解决、可接受），建议后续补 AIV 单缓冲/Cube 兜底方案。
+3. 【P3】MergeBatch 极端瘦长 case（M 或 N 很小 × 大 B）仍有 0.2% 生成不可行方案——`b0` 选择未把 L0A/L0B 容量纳入上限；现已被生成自检显式标注（不再静默矛盾）。
+4. 【P3】负/零矩阵维度无输入校验，静默产出无意义方案；FIXPIPE 建议文案对 StreamK（部分和 4B）给出 fp16/fp8 减半的误导提示。
+5. 【P3】过程噪音：整改过程中 11 个 commit 误提交临时文件（bug1.csv/t.csv 等）后删除，建议加 .gitignore。
+
+**总体评价：6.2 → 约 7.6/10。** 架构与理论映射维持高水准；内部一致性与鲁棒性是本轮最大提升（单一约束源 + 生成自检是正确架构决策）；剩余扣分集中在 StreamK 时延口径复核、输入校验与若干边界覆盖。
+
+---
+
+## 1. 本轮整改概览（33 commits）
+
+| 主题 | 主要提交 | 对应 issue |
+|---|---|---|
+| 新增 `bmm_theory/constraints.py`（170 行，单一约束源 + 生成侧收敛辅助） | 16a84ac | #5 #6 共性根因 |
+| `router.py`：生成后自检 `_wrap_checked` + 无方案兜底 `_no_plan` | dad4585 | #4 #5 |
+| `evaluator.py`：约束校验委托单一约束源 | 0c19d99 | #5 |
+| `__main__.py`：plan=None / timing=None / 自检违规打印防护 | b57ba57 | #4 |
+| `io_csv.py`：out_nd 解析 + A/B dtype 不一致告警 + 建模边界说明 | 9ef240f | #7 #8 |
+| `timing.py`：REDUCE 移出稳态 max()（`reduce_serial` 约定） | 933b428 | #9 |
+| `branches/stream_k.py`：fixpipe 改按部分和 4B、显式 reduce_serial | f756729 | #9 |
+| `branches/asw_basic.py`：降核 base_k 按 dtype 容量反推（clamp 同源）；k_l1 小 K 不切 | e6aa792 | #5 |
+| `branches/iter_batch.py`：L0 tile 长宽比跟随 + L0C/L0A/L0B 同源收敛 | c0b358f | #5 |
+| `branches/merge_batch.py`：新增条件 2b（合并后最小 K 粒度下 L0A/L0B 可驻留） | 1385d01 | #5 |
+| README / docs/01_软件架构.md 同步 6 分支现状 + 约束模块/归约口径文档 | 7bcacef / 593ffcf | #10 |
+| examples 两个结果 csv 按新模型重生成 | bbef360 / cff5665 | — |
+| tests 新增 `TestIssueRegression` 6 条 | 87b7be5 | #4–#9 |
+| 误提交临时 csv 后清理 | 0e644ac…99a25a6（11 commits） | — |
+
+净改动：15 个文件，+408/−92 行。无第三方依赖新增。
+
+---
+
+## 2. issue 逐条闭环验证
+
+### #4 【P0】K=1 且 batch<128 崩溃 → ✅ 已闭环（标注式兜底）
+
+复测：`B=64 M=8192 N=32 K=1 int8` 走 `recommend`（-v / 不带 -v / 输出 plans.csv 三条路径）均**不再崩溃**：
+
+```
+=== bug_k1_smallB: ... K=1 int8 -> [特殊分支]
+    仲裁: [无方案] K=1逐元素乘...; 但进入条件不满足 (2_K=1的AIV触发: B >= 2*AIV核数), 该区域暂无理论方案...
+    方案: 核数=0 ...
+```
+
+实现：`router.py:52-58` capable=False 或 plan=None 时走 `_no_plan`（占位 ImplPlan，used_core_num=0，timing=None）；`__main__.py:85-90` 增加空值防护。
+
+**残留（P3）**：该区域仍未给出任何实现方案。扫描 B=1..255 × K=1 × 若干 M/N：**49.4%（378/765）落"无方案占位"**；K=0 全部有方案。原 issue 允许"标注无方案"作为最低要求（已达成），但建议后续补 AIV 单缓冲或 Cube 兜底方案消除空洞，或至少在文档（00_总纲 / 04_特殊分支）中明确该限制。
+
+### #5 【P1】推荐模式不做约束自检 → ✅ 基本闭环（6.0%/18.0% → 0.2%/0.0%）
+
+架构修复正确：新增 `constraints.py` 为**唯一事实来源**，`router._wrap_checked` 在每次生成后调用同一校验函数，违规时在 arbitration/advice 中显式标注"[自检违规: …]"；`evaluator._check_constraints` 改为委托同一函数。同时生成侧收敛辅助（`clamp_base_k` / `clamp_base_mn_l0c`）消除降核 base_k 硬编码等缺陷。
+
+复测（与 v1 同种子同分布）：
+- 上轮全部最小违规样例（IterBatch K=8/K=2、MergeBatch fp8 K=32、ASW_降核 fp16/fp32、fp32 大 MN 等 12 个）→ **全部 violations=OK**；
+- 宽范围 8000 例：480 → **16**（0.2%）；典型 6000 例：1079 → **0**。
+
+**残留（P3）**：16/8000 全部为极端瘦长 MergeBatch（如 `B=811 M=33 N=1 K=2459 fp32`、`B=1024 M=1 N=1024 K=1024 int8`）：`make_plan` 的 b0 上限只取 L0C/算存比/b_core 三者的 min，**未纳入 L0A/L0B 容量**，导致 `b0*M`（或 `b0*N`）过大时 `base_k≥16`（fractal 下限）也放不进 L0A/L0B。新条件 2b（merge_batch.py:55-62）只按 MIN_B0=2 检查，未覆盖实际选中 b0 的情形。因生成自检已显式标注，属于"可感知的残余"，建议把 L0A/L0B 容量加入 b0_max 求解（如 b0 ≤ L0A/(2·m·16·dt) 与 L0B 侧取 min 后取不超过 b_core 的因子）。
+
+### #6 【P1】dValue 口径矛盾 → ✅ 闭环
+
+理论裁定已落地并文档化（constraints.py:7-15、docs/01_软件架构.md §3.5）：**dValue 128B 下限只对"K 被切分、K 段成为搬移连续维"（k_l1 < K）的方案生效；K 整驻留（k_l1 ≥ K，形态 a/b、小 K 整搬）时连续维是 M/N，豁免 K 向检查**。实现于 `_k_segment_is_contiguous`（constraints.py:118-134）。复测：IterBatch K=8/K=2 整驻留样例 → 无违规；形态 c/d 与 MergeBatch/ASW/StreamK 的 K 切分路径仍受约束。新增回归测试 test_issue6。
+
+### #7 【P2】out_nd 无法经 CLI 传入 → ✅ 闭环
+
+`io_csv.load_cases` 现解析 `out_nd`（缺省 True，大小写不敏感）。CLI 复测：同一 StreamK case `out_nd=0` → ASW_Basic_降核（StreamK 条件 4 生效）；`out_nd=1` → StreamK。csv 头注释补充说明。新增回归测试 test_out_nd_disables_streamk。
+
+### #8 【P2】A/B dtype 不一致无告警 → ✅ 闭环
+
+`io_csv.load_cases` 加载后对 `dtype_a ≠ dtype_b` 打印 `[warn]`（stderr，含 case_id 与建模口径说明）。CLI 复测：`fp8 vs bf16` 输入正确输出告警。
+
+### #9 【P2】StreamK 归约计账口径 → ⚠️ 部分闭环，引入新疑点（重点跟进）
+
+修复部分正确：REDUCE 不再进稳态 max()（`assemble_timing` 增加 `reduce_serial=True` 约定，timing.py:96-132），归约只作为 drain 串行追加一次；最终写回只留在 `eval_streamk_reduce` 内；回归测试 test_issue9 验证 `t_total = t_steady + t_drain` 且 `drain = t_reduce`（代数恒等）。
+
+**但 fixpipe 的新口径有实质问题**（复算证据，demo case：B=4 M=N=128 K=10240 bf16，grid_K=32）：
+
+| 账目 | 字节量 | 带宽分母 | 结果 |
+|---|---|---|---|
+| `t_reduce` 内 t_write_partial（部分和写 L2） | grid_K×tile×4B = 8.39MB | 整片 5.2TB/s（并发写） | **1.61us** |
+| 新 `t_fixpipe`（stream_k.py:152-153） | 同一批 8.39MB | **单核份额 162.5GB/s** | **51.62us** |
+
+1. **同一批部分和字节被计两次账**（一次在 reduce 内按整片带宽、一次在 fixpipe 按单核带宽）——issue#9 原指控的"重复计账"并未消除，只是换了个位置与分母；
+2. **单核串行化假设与软件自身的带宽模型矛盾**：timing.py 文档与 00_总纲均声明"每核独立 DMA 引擎、带宽按核数配平"，即 grid_K 个核的部分和是**并发写出**的，墙钟时间应为 字节量/整片带宽 ≈1.61us；51.62us 相当于把整组部分和压到单核写口串行写，**高估 grid_K=32 倍**；
+3. 后果已固化进 examples：streamk_demo 总时延 9.96us → **55.03us**，瓶颈从 MTE2_GM → FIXPIPE，且 advice 还给出"fp16/fp8 可减半写出量"的提示——该提示对按 4B 防精度丢失的部分和不成立（advice 生成器未感知 StreamK 例外，evaluator.py 的 _advice FIXPIPE 分支）；
+4. 修复后的回归测试只验证汇总代数式，**验证不了字节量与带宽口径的物理正确性**。
+
+建议：fixpipe 部分和按"每核自己的部分和（tile×4B）÷ 单核份额"计（=1.61us，与 reduce 内并发口径一致），或与 `t_reduce` 的 t_write_partial **二选一**计账并在文档写明；再校验 8.39MB 是否应为每核 256KB 的并发总和语义。此条建议**重新开启 issue#9 或另开 issue**。
+
+### #10 【P3】文档-代码漂移 → ✅ 闭环
+
+README 目录树与分支注释更新为 6 分支全实现；docs/01_软件架构.md 分层图、§3.5（单一约束源）、§4 扩展指南（转Matmul 精切为后续方向、新分支须走 constraints.py）均已同步；Python 版本声明改为"实测 3.12/3.14，建议 3.10+"（3.14 本机复验通过）。轻微残留：`router._wrap`（旧包装函数）成为死代码未删除；README 目录树仍缺 `docs/03_测评报告` 条目（小事）。
+
+---
+
+## 3. 回归测试与可复现性
+
+- 单测：`python -m unittest discover -s tests -v` → **Ran 23 tests, OK**（0.028s）。
+- recommend 输出 vs 仓库 `examples/result_recommend.csv`：10 行全列 **0 差异**；evaluate 输出 vs `examples/result_evaluate.csv`：**0 差异**（含新的 streamk 55.03us 数值——即作者已用新模型重生成示例）。
+- 压力回归（与 v1 完全同种子同分布）见 §0 表；两组均 0 崩溃、0 NaN、0 负时延。
+
+---
+
+## 4. 代码质量复查（针对改动）
+
+- `constraints.py` 抽象正确：校验与生成收敛共用；注释完整记录 dValue 裁定缘由。
+- 三处发现的小问题：
+  1. `constraints._l1_need_bytes` 对 IterBatch 形态 c 的预算推断（resident = min 侧 + 对侧 k_l1 双缓冲）与生成侧 iter_batch.l1_form 逻辑手写两遍，仍属"双源"（本次未合并），好在口径一致、且有校验兜底；建议未来直接让 make_plan 调用同一 helper。
+  2. `merge_batch.py` 条件 2b 用 MIN_B0 预检但实际 b0 更大时仍可能溢出（§2 #5 残留）——校验兜底已标注，但生成侧仍可产生违规方案（0.2%）。
+  3. 输入范围校验仍缺失：`M=-5` / `M=0` 静默产出 IterBatch 伪方案与无意义时延；非法 dtype 抛裸 ValueError。建议 load_cases 后加维度正数校验。
+
+---
+
+## 5. 遗留问题清单（按优先级）
+
+| # | 级别 | 问题 | 证据 | 建议 |
+|---|---|---|---|---|
+| N1 | P2 | StreamK fixpipe 部分和写双重计账 + 单核串行化高估 ~32 倍；examples 已固化 55.03us 结果 | stream_k.py:152-153 vs timing.py:77-93；demo 复算 51.62us vs 1.61us | reopen #9 或新开 issue；fixpipe 与 reduce 内部分和写二选一，按并发口径计 |
+| N2 | P3 | K=1 且 B<128 无方案空洞（占 K=1 域 49.4%） | B=1..255 × K=1 扫描 378/765 占位 | 后续补 AIV 单缓冲/Cube 兜底方案或文档明示 |
+| N3 | P3 | MergeBatch 瘦长 case b0 未含 L0A/L0B 上限 → 0.2% 自检违规（已标注不静默） | 8000 例中 16 例，样例见 §2 #5 | b0_max 纳入 L0A/L0B 容量约束 |
+| N4 | P3 | FIXPIPE 建议文案对 StreamK 误导（fp16/fp8 减半对 4B 部分和不成立） | evaluator.py _advice | advice 生成感知分支例外 |
+| N5 | P3 | M≤0 等非法输入无校验，静默输出伪方案 | M=-5/M=0 探针 | load_cases/route 前参数校验 |
+| N6 | P3 | 无 .gitignore、整改期误提交 11 个临时文件 commit；死代码 router._wrap；README 目录树缺 03_测评报告 | git log 0e644ac..99a25a6 | 补 .gitignore、删死代码 |
+
+---
+
+## 6. 总体评分（第二轮）
+
+| 维度 | v1 | v2 | 说明 |
+|---|---|---|---|
+| 设计架构 | 9.0 | 9.0 | 单一约束源使架构更完整 |
+| 理论-代码映射 | 8.0 | 8.5 | dValue 裁定、归约约定均有文档 |
+| 功能完成度 | 8.0 | 8.5 | 双模式 + 自检 + 告警 |
+| 内部一致性 | 5.0 | 7.0 | 0.2%/0% 自洽率；StreamK 口径遗留 |
+| 鲁棒性/错误处理 | 4.0 | 7.5 | 崩溃消除、占位兜底、告警落地；输入校验仍缺 |
+| 测试 | 5.0 | 7.0 | 23 条含 6 条 issue 回归；仍无 io_csv/CLI 单测 |
+| 工程化/交付 | 5.0 | 6.5 | 文档同步、模块化；无 CI/打包/.gitignore、误提交噪音 |
+| **综合** | **6.2** | **7.6** | POC → 可用工具雏形 |
+
+**定位更新**：已从"理论文档的代码化 POC"升级为"分支归属与方案生成可自检、结果自洽率 99.8% 的辅助分析工具"。仍不建议把绝对时延数值直接当硬件实测使用（T_cmd=50ns 等参数未标定、StreamK 口径待复核），但**方案归属、tile 取值与相对瓶颈分析已经具备可信度**。
+
+**优先建议**：① 复核 StreamK fixpipe 口径（N1）；② 补输入校验（N5）；③ 清理工程噪音并考虑 CI（N6）；④ 中远期做硬件标定与 ops-nn 源码对照验证。
+
+---
+
+## 附录 A：本轮复测命令
+
+```bash
+git fetch origin && git log --oneline d36e224..origin/main   # 整改提交
+python -m unittest discover -s tests -v                      # 23/23
+python -m bmm_theory recommend examples/cases_demo.csv -o r.csv --plans p.csv
+python -m bmm_theory evaluate examples/cases_demo.csv p.csv -o e.csv
+# 与 examples/result_*.csv 比对: 0 差异
+```
+
+## 附录 B：残余 MergeBatch 违规样例（16/8000，0.2%）
+
+```
+B=811/811 M=33 N=1 K=2459 fp32 -> L0A tile 超容量 (b0=25: base tile 825×…×16×4B×2)
+B=1024/1024 M=1 N=1024 K=1024 int8 -> L0B 超容量 + dValue=48B
+B=2048/2048 M=512 N=1 K=16 bf16 -> L0A 超容量 (b0=8)
+...
+```
+特征：M 或 N 极小 × B 巨大 × b0 因子选择偏大 → 合并后 base tile 一维过长，L0A/L0B 在 base_k≥16 下无法驻留。已在 advice 中标注"[自检违规]"，不静默。
+
+## 附录 C：K=1 无方案空洞扫描
+
+B=1..255 × K=1 × M/N∈{64×64, 256×256, 4096×64}：765 例中 378 例（49.4%）返回 used_core_num=0 占位"[无方案]"；K=0 同域 0 例占位。
+
+## 附录 D：遗留问题闭环确认（2026-09-06，commit 4bedf0a / HEAD 3a6c952）
+
+本报告 §5 的遗留问题（N1–N6）已全部转为 issue #11–#16 并修复闭环：
+
+| Issue | 修复内容 | 验证 |
+|---|---|---|
+| #11 N1（P2）StreamK fixpipe 双重计账/32×高估 | 部分和写/读回/求和/最终写回全部单次计入串行 drain（稳态 Fixpipe 账目清零） | streamk_demo 55.03→**9.96us**，瓶颈回到 MTE2_GM |
+| #12 N2（P3）K=1 且 B<128 无方案空洞 | 新增 **AIV 单缓冲**模式（B≥128 仍乒乓），04_特殊分支.md 同步 | B=1..255×K=1 全扫 **765/765 全有方案** |
+| #13 N3（P3）MergeBatch 瘦长不可行（0.2%） | b0_max 纳入 L0A/L0B 容量上限；路由对胜出方案自检，违规时回退（另一切B候选→StreamK→ASW） | 8000 例宽范围违规 **16→0** |
+| #14 N4（P3）FIXPIPE advice 对 StreamK 误导 | StreamK 跳过 dtype 减半提示，新增 REDUCE 瓶颈建议 | 文案验证 |
+| #15 N5（P3）输入校验缺失 | BmmCase.__post_init__ 维度/dtype 校验，非法输入明确抛错 | M≤0/K<1/dtype 非法均正确报错 |
+| #16 N6（P3）工程噪音 | 根目录 .gitignore、删除 router._wrap 死代码、README 目录树补 03_测评报告 | — |
+
+**修复后全量验证**：单测 **28/28**（新增 8 条回归）；宽范围 8000 例 + 典型形状 6000 例压力测试 **违规 0 / 异常 0 / 无方案 0 / NaN 0**；examples 已按新模型重新生成并与代码 0 差异提交。Gitea issue #4–#16 已全部关闭（附修复注释）。
+
+---
+
+*第二轮复评完成：本地动态测试 + 与仓库 HEAD 99a25a6 比对；全部问题均有可复现证据。修复确认见附录 D（HEAD 3a6c952）。*

BMM/BMM_Theory/docs/04_差异对照_bmmv3_vs_BMM_Theory.md

@@ -0,0 +1,176 @@
+# bmmv3 vs BMM_Theory 差异对照
+
+> 对照对象：`bmmv3/bmmv3/bmmv3_arch35_branch_checker.py`（下称 **bmmv3**）与本仓库 `bmm_theory/branches/merge_batch.py` + `iter_batch.py`（下称 **BMM_Theory**）。
+> 写作目的：说清两版"是什么关系、差在哪、各自什么时候用、该互相吸收什么"。
+
+---
+
+## 0. 一句话结论
+
+**两版不是竞争关系，是互补关系。**
+
+- **bmmv3** = 现有算子源码行为的**预测器/对拍器**（source-of-truth 是 `batch_matmul_v3_mergebatch_basicapi_tiling.cpp` / `iterbatch_tiling.cpp`）；
+- **BMM_Theory** = 理论最优实现的**推导器**（source-of-truth 是《BMM算子优化分析 v0.98》《MergeBatch_vs_IterBatch分析 v1.1》）。
+
+要"知道现有 kernel 实际会怎么跑"→ 用 bmmv3；要"知道理论上应该怎么跑最优"→ 用 BMM_Theory。
+
+---
+
+## 1. 根本定位差异
+
+| 维度 | bmmv3 | BMM_Theory |
+|---|---|---|
+| 目的 | 判断 case 能否进**现有源码**的 `MERGE_BATCH_BASICAPI` / `ITER_BATCH` | 推导 case 的**理论最优**实现方案 |
+| 依据 | ops-nn 源码逐条件还原 | 理论文档推导 + 硬件容量/带宽约束 |
+| tile 语义 | 按源码硬编码（ping_pong 开关、step 幂次搜索） | 按理论容量极限 + 生成后自检 |
+| 时延模型 | 三行带宽公式（CUBE/FIXP/MTE2 各一行） | 分阶段 max(MMAD,MTE2,FIXPIPE)+drain，对齐 v1.1 |
+| 分支覆盖 | 仅 MergeBatch + IterBatch | 六分支全（转Matmul/特殊/MergeBatch/IterBatch/StreamK/ASW_Basic） |
+| 输入校验 | 无（静默接受非法输入） | `__post_init__` 明确报错（issue#15） |
+| 生成后自检 | 无 | constraints.py 同源校验（issue#5） |
+
+---
+
+## 2. MergeBatch 分支逐项对照
+
+### 2.1 准入条件
+
+| 条件 | bmmv3 | BMM_Theory | 差异说明 |
+|---|---|---|---|
+| batch 关系 | A/B 4 个 batch 轴完全相等 | BatchA==BatchB | bmmv3 更严（源码实现细节） |
+| 每核 batch 下限 | `min_batch_l0=4`（每核≥4 batch） | `b_core ≥ 2*b0`（b0≥2 即每核≥4） | 等价，bmmv3 硬编码 4，我们参数化 |
+| L0C 容量 | `l0cSize = tempAlignM×tempAlignN×4×l0c_factor ≤ L0C` | `2×(b0·M)×(b0·N)×4 ≤ L0C` | 一致（l0c_factor=2 即双缓冲） |
+| L0A/L0B 容量 | 准入条件 21/22 查（`al0Size/bl0Size`） | **b0 计算时显式收敛（issue#13）** | **我们更完整**——bmmv3 只在准入查，没在 b0 计算里收敛 |
+| K 对齐下限 | `alignK ≥ merge_min_align_k=64` | 无 | bmmv3 源码特有 |
+| M≤N 约束 | 有（条件 13） | 无 | bmmv3 源码特有（实现限制，非理论必要） |
+| 转置 view | 非连续转置 view 排除 | 不建模 | bmmv3 源码特有 |
+| bias | `has_bias=false` | 不建模 | bmmv3 源码特有 |
+| 访存 Bound | `2MN/(M+N) < 607/2`（经验常数） | `2MN/(M+N) < R16/b0`（R16=算存比） | **我们更通用**——bmmv3 的 607 对应旧总算力口径 R16=607.5；现行 Cube-only 口径 R16=540（issue#40），fp32 时 R16 再减半但 bmmv3 不会 |
+
+### 2.2 b0 计算
+
+```
+bmmv3:  b0 = min(L0C/(2·M·N·4), ceil(B/C), 607.5·(M+N)/(2·M·N))
+我们:   b0 = min(√(L0C/(8MN)), R16·(M+N)/(2MN), √(L0A/(2·M·16·dt)), √(L0B/(2·N·16·dt)), b_core) 取因子
+```
+
+**关键差异**：我们多了 L0A/L0B 上限（issue#13），bmmv3 靠 `min_batch_l0=4` 间接保证 L0A/L0B 不溢出，但不如显式约束通用（b0=2 时 bmmv3 的间接保证失效）。
+
+### 2.3 KL0/KL1 计算
+
+| | bmmv3 | BMM_Theory |
+|---|---|---|
+| KL0 | `floor_align16(min(L0A/(2·b0·M·dt), L0B/(2·b0·N·dt)))` | 同（`align_down(min(...), fractal)`） |
+| KL1 | `ceil_align16(min(K, 512/dt, L1/(2·b0·(M+N)·dt)))`，不满足 L1 则减 16 循环 | `align_down(min(k_l1_star, K, dvalue_cap), fractal)` |
+| 实现风格 | 源码试探式（ceil 后逐步减） | 理论反推式（直接 align_down） |
+
+**差异说明**：bmmv3 的"ceil 后减 16 循环"是源码实现细节（为了向上对齐后再试探），理论反推直接用 `align_down` 即可，结果等价但我们更简洁。
+
+---
+
+## 3. IterBatch 分支逐项对照
+
+### 3.1 准入条件
+
+| 条件 | bmmv3 | BMM_Theory |
+|---|---|---|
+| 每核 batch | `floor(B/C) ≥ 1` | `b_core ≥ 1`（ceil） | 一致 |
+| 负载均衡 | `B mod C == 0` 或 `≥ 26`（硬编码） | `B mod C == 0` 或 `≥ min_core_num`（参数化） | **我们更通用** |
+| L1 形态 | a/b/c1/c2/d 五种 | a/b/c/d 四种 | **bmmv3 更细**（c 拆 c1/c2） |
+
+### 3.2 c 形态处理
+
+**bmmv3**：
+- c1：MK 整体驻留 + KN/step 分块
+- c2：KN 整体驻留 + MK/step 分块
+- step 按 2/4/8/16… 幂次递增遍历取最大
+
+**BMM_Theory**：
+- c 形态统一："驻留较小侧 + 对侧切 K"
+- k_l1 连续反推：`k_l1 = (budget - resident) / (other·dt) / 2`
+
+**差异说明**：bmmv3 的 step 幂次搜索是源码为了 tiling 生成方便做的离散化；理论极限下 k_l1 连续取最优即可。**做理论推导用我们的，做源码对拍用 bmmv3 的。**
+
+### 3.3 d 形态处理
+
+**bmmv3**：`step_d` 由 16KB 和 128B 两条规则推导，`KL1 = ceil_align16(K/step_d)`，不满足 L1 则减 16 循环。
+
+**BMM_Theory**：`k_l1 = align_down(L1/(2·(M+N)·dt), fractal)`，直接反推。
+
+**差异说明**：同上，bmmv3 是源码试探式实现，我们是理论反推。
+
+---
+
+## 4. 时延模型对照
+
+| | bmmv3 | BMM_Theory |
+|---|---|---|
+| CUBE | `2·B·M·K·N / (MAC_TFLOPS·1e6)` | `2·b_core·M·N·K / q16`（单核） |
+| MTE2 | `(2·M·K + 2·K·N)·B / (MTE2_BW·1e6)` | 分 GM/L2 两段 + T_cmd + drain |
+| FIXP | `2·M·N·B / (FIXP_BW·1e6)` | `b_core·M·N·out_dtype / bw_pc` + unitflag |
+| 流水掩盖 | 无 | max(MMAD,MTE2,FIXPIPE) + drain |
+| 分支仲裁 | 无 | MergeBatch vs IterBatch 谁优 |
+
+**差异说明**：bmmv3 的时延是纯带宽粗估（三行公式），我们做分阶段 + drain + 仲裁，能定位瓶颈。
+
+---
+
+## 5. 各自适用场景
+
+| 场景 | 推荐 |
+|---|---|
+| 预测现有 kernel 实际行为（对拍） | **bmmv3** |
+| 推导理论最优方案（设计） | **BMM_Theory** |
+| 分析瓶颈（MTE2/Cube/Fixpipe 谁卡） | **BMM_Theory**（bmmv3 无瓶颈分析） |
+| 验证源码准入条件是否正确 | **bmmv3**（BMM_Theory 不管源码现状） |
+| 大批量 case 筛选（XLSX 原表追加） | **bmmv3**（我们暂无 XLSX） |
+
+---
+
+## 6. 可吸收点清单
+
+### 6.1 bmmv3 → BMM_Theory（我们该补的）
+
+| 项 | 优先级 | 状态 |
+|---|---|---|
+| fp4（0.5B）dtype 支持 | 高 | ✅ **已补（本期）**——`DTYPE_BYTES` 加 `"fp4": 0.5` / `"fp4_e2m1": 0.5`，`dtype_bytes()` 返回 float |
+| 转置对 dValue/base_k 的影响建模 | 高 | ✅ **已补（本期）**——MergeBatch 加条件 6（A 转置对齐），IterBatch dValue 判定按转置调整连续维 |
+| 转置判据三处同源 | 中 | ✅ **已补（issue#19）**——`models.dvalue_contig_dims` 统一连续维 dValue，`l1_form` c/d 生成守卫 / IterBatch 条件 4 / `constraints` 校验共用同一判据，非转置行为不变 |
+| XLSX 原表追加输出 | 中 | 待做 |
+| c1/c2 的 step 幂次搜索 | 低 | 不做（理论极限不需要离散化） |
+
+**fp4 已补细节**：
+- `models.py`: `DTYPE_BYTES` 加 `"fp4": 0.5` / `"fp4_e2m1": 0.5`
+- `dtype_bytes()` 返回类型改为 `float`（兼容 0.5）
+- `dtype_in_bytes`/`dtype_out_bytes` 属性返回类型改为 `float`
+- 测试：`TestFp4Support` 3 个用例
+
+**转置建模已补细节**：
+- **MergeBatch**（`merge_batch.py`）：新增条件 6——A 转置且 M>1 时 `tempAlignM = b0 * alignM`，校验 L0A 容量
+- **IterBatch**（`iter_batch.py`）：dValue 判定按转置调整连续维——A 不转置判 `K*dt`，A 转置判 `M*dt`；B 不转置判 `N*dt`，B 转置判 `K*dt`
+- 测试：`TestTransposeModeling` 3 个用例（含 A 转置大 M 小 K 通过、A 不转置小 K 失败的边界 case）
+
+### 6.2 BMM_Theory → bmmv3（他们可参考的）
+
+| 项 | 说明 |
+|---|---|
+| 分支仲裁 | MergeBatch vs IterBatch 谁优的判断 |
+| 时延 drain/流水掩盖 | 三行公式 → 分阶段 max + drain |
+| 生成后自检 | 生成即输出 → 生成后跑约束校验 |
+| L0A/L0B 的 b0 显式收敛 | issue#13：b0 计算时直接纳入 L0A/L0B 上限 |
+| 输入校验 | issue#15：非法维度/dtype 明确报错 |
+
+---
+
+## 7. 结论
+
+两版**互补不替代**。bmmv3 是"源码行为预测器"（对拍用），BMM_Theory 是"理论最优推导器"（设计用）。
+
+**已吸收**：fp4 支持 + 转置对 dValue 的建模（本期补入）。
+
+**不建议吸收**：c1/c2 的 step 幂次搜索（源码离散化细节，理论推导不需要）。
+
+**待考虑**：XLSX 输出（如用户需要）。
+
+---
+
+*版本：v1.1 | 2026-09-06 | 基于 BMM_Theory main（含 issue#11-#22 修复：含 #17 恢复 StreamK 单次计账/K=1 单缓冲/#14 advice，#19 转置判据三处同源）与 bmmv3 最新版对照*

BMM/BMM_Theory/docs/05_L2驻留GM读写与dtype算力口径_设计分析.md

@@ -0,0 +1,220 @@
+# 05 L2 驻留、GM 读写与 dtype 算力口径 — 设计分析
+
+> 关联 issue: #27 (MergeBatch Cube 复核) / #28 (dtype 感知算力) /
+> #29 (GM 读取量下限 + L2 驻留工作集) / #30 (Fixpipe 输出落点)
+>
+> 状态: 设计文档先行, 代码按本文档落地 (2026-06 评审轮)。
+> 本文档是 GM/L2/输出计账与算力口径的**单一语义来源**; 各分支 evaluate/生成注释以本文档为准。
+
+---
+
+## 0. 本轮的四个问题与结论摘要
+
+| # | 问题 | 结论 | 落地 |
+|---|---|---|---|
+| #27 | MergeBatch Cube 时延 "每次 b0 个 batch, flops=2·b0M·b0N·K, 共 b_core/b0 次" | **公式与代码一致** (总 flops = b_core·b0·2MNK, 证明见 §6); 无计算改动 | 注释显式化; 暴露的"字节列每核/芯片口径混用"随 #29 统一 |
+| #28 | 估算时延恒用 BF16 Cube / fp32 AIV 算力 | **确认错误** | §2 dtype 感知速率表 + 全链路替换 |
+| #29 | GM 读取量"小于输入数据量" | 计算层面各分支 GM ≥ 输入一次 (逐分支表见 §5); **真问题**: ① 字节列每核/芯片口径混用导致比对失真; ② 缺整 case 驻留边界与统一驻留判定 | §3 公理化 + §4 场景模型 + 芯片口径列统一 + GM≥输入不变量测试 |
+| #30 | Fixpipe 输出落点 (默认写 L2, 容量不足才写 GM) | **确认错误** (Iter/Merge/特殊恒直写 GM; ASW 场景 A 按单 batch 判定, 整 case 输出累积必逐出) | §4.2 输出落点规则 |
+
+---
+
+## 1. 硬件事实 (Ascend950PR, 与 hardware/ascend950pr.py 对应)
+
+- GM (HBM): **1.6TB/s, 读写共享总线** —— 同一时刻读与写累加计时 (issue#23 已落地)。
+- L2: 128MB, **5.2TB/s, 读口/写口各自独享** —— L2 重复读(读口)与 Fixpipe→L2 写(写口)互不竞争 (issue#23 已落地)。
+- 输入首访一律走 GM 带宽计一次, 不叠加 L2 (issue#24 已落地); L2 只吸收"驻留后的再次读取"。
+- Cube 算力分精度 (白皮书: FP8/MXFP8/HiF8 = 2×FP16, MXFP4 = 4×FP16; 16bit 档 = 基准):
+  BF16 **432 TFLOPS/芯片 (=13.5 TFLOPS/核, Cube-only 口径 = 白皮书表3-1 "Cube算力" 单行,
+  issue#40 用户裁决; 486 为 Cube+Vector 总算力, 不作为 Cube 时延基准)**。
+  FP32/TF32 同代比值白皮书未给 → **假设 ½, 待实测标定** (见 §2 假设表)。
+- AIV (Vector): 64 核 × 128 fp32 lane/拍 × 1.65GHz = 13.5T 元素/s (fp32 档)。
+
+---
+
+## 2. dtype 感知算力 (issue #28)
+
+### 2.1 速率表 (因子, 相对基准档)
+
+Cube 因子 (相对 BF16, 同倍率作用于整芯片与单核):
+
+| dtype | 因子 | 依据 |
+|---|---|---|
+| bf16 / fp16 | 1.0 | 基准档 (950 同速) |
+| fp32 / tf32 | 0.5 | **假设**: 白皮书仅述 FP16/FP32 单核较上代均 +100%, 未给同代比值; 按 DaVinci 惯例 FP32 累加通量减半。待 msProf 实测标定 |
+| fp8 / fp8_e4m3 / fp8_e5m2 | 2.0 | 白皮书: FP8 = 2×FP16 |
+| int8 | 2.0 | **假设** 同 FP8 (8bit 整数张量档), 待实测 |
+| fp4 / fp4_e2m1 | 4.0 | 白皮书: MXFP4 = 4×FP16; 普通 fp4 按 MXFP4 假设, 待实测 |
+
+AIV (Vector) 逐元素通量因子 (相对 fp32 lane 基准):
+
+| dtype | 因子 | 依据 |
+|---|---|---|
+| fp32 / tf32 | 1.0 | 基准 (128 lane/拍/核) |
+| fp16 / bf16 | 2.0 | **假设** 16bit 双元素/lane, 待实测 |
+| fp8 / int8 | 4.0 | **假设** 8bit 四元素/lane, 待实测 |
+| fp4 | 8.0 | **假设** 待实测 |
+
+A/B dtype 不一致 (混精度): Cube 取**较慢一侧**的因子 (max(字节数) 侧, 即 min(因子))。
+StreamK 归约是对 **fp32 部分和**求和 → AIV 归约恒按 fp32 档 (因子 1.0), 不随输入 dtype 变。
+
+### 2.2 替换点 (全部时延估算)
+
+| 位置 | 现状 | 改为 |
+|---|---|---|
+| 各分支 t_mmad / t_comp_chunk / drain / 尾轮决策主导项 | flops / spec.q16 | flops / spec.q_cube(dtype_a, dtype_b) |
+| MergeBatch beats_iterbatch 阈值中的 T_comp | q16 | 同上 (该分支 A/B 同 dtype, 简化为 dtype_a) |
+| special K=1 t_compute (逐元素乘) | b·m·n / spec.q_aiv (fp32) | b·m·n / spec.aiv_elem_rate(dtype_in) |
+| StreamK θ_c (归约代价系数) | q16(bf16) | q_cube(输入 dtype) (AIV 侧仍 fp32) |
+| R16 平衡点 (MergeBatch 进入条件 5 / b0 算存比上限) | r16(bf16 基准) | r16_for(dtype) |
+| 入口条件/生成侧 | — | 同步按 §2.1 因子 (dtype 只影响时延, 不改变分支结构) |
+
+---
+
+## 3. 存储口径公理 (issue #29)
+
+记号:
+
+- V_in = 输入存储总量 = batch_a·M·K·dt + batch_b·K·N·dt (广播前实际存储), V_out = batch_c·M·N·out_dt;
+- a_b = M·K·dt (单 batch A 字节), b_b = K·N·dt (单 batch B 字节)。
+
+- **R1 (GM 首读下限)**: 每个输入字节最初在 GM, 至少从 GM 读一次 → **GM_read ≥ V_in**。
+  违反即模型 bug (统一用测试锁死)。
+- **R2 (L2 只吸收驻留后的重复读)**: 同一字节的第 2..n 次访问, 仅当两次访问之间该字节仍驻留 L2
+  (所在"复用窗口"的工作集 ≤ L2 减去其它占用) 才按 L2 读口计时; 否则按 GM 重读计时。
+- **R3 (整 case 全驻留边界)**: V_in + V_out (StreamK 另加部分和 workspace) ≤ L2 时,
+  全程零逐出: GM_read = V_in (每字节一次), 全部重复读走 L2, 输出驻留 L2 (见 §4.2)。
+- **R4 (输出落点)**: 见 §4.2。
+- **R5 (输入工作集场景)**: batch 内 tile 共享块 (A 行块被 n_cnt 个列 tile 复用 →
+  (n_cnt−1) 次重复; B 列块被 m_cnt 个行 tile 复用 → (m_cnt−1) 次) 的流量归属由 §4.1
+  场景判定: 可驻留时按 L2 读口计 (含**单侧全驻留+对侧滑窗**调度, issue#32);
+  双侧都放不下时按最小替换分组 (§4.3)。
+- **R6 (已知简化, 文档标注)**: ① BatchA≠BatchB (均 ≥2, 如 2 vs 64) 的广播/混合 batch 结构
+  按"每 batch 独立矩阵"计 GM = b·(a_b+b_b), 是 V_in 的**保守放大**
+  (真实为 V_in + 共享矩阵跨 batch 重复读 ≤ 该值), 不做特殊建模;
+  ② K 切分 (k_L1<K) 各段数据互不重叠: GM 按实际字节精确计 (每字节恰从 GM 读一次,
+  无 padding 上取; 切段数只影响 DMA 命令数/T_cmd 与双缓冲调度, 见 §5 与 issue#31);
+  ③ 场景按单 batch 工作集判定, 多核波次并发窗口按 batch 粒度近似。
+
+### 3.1 逐分支 GM/L2 流量归属表 (现行模型正确性核查结果)
+
+| 分支 | 结构 | GM 直读 | L2 重复读 | 核查 |
+|---|---|---|---|---|
+| IterBatch (切 B) | 逐 batch 整驻留/切 K; 每 (batch, K段) 数据互不重叠 | **V_in** (每字节恰一次, 无 padding 上取, issue#31) | 0 (L1 形态吸收核内复用) | ✓ GM=V_in |
+| MergeBatch | 合并组 (b0 batch) 一次搬入; K 段/整 K 均不跨段重读同一字节 | **V_in** (issue#31) | 0 | ✓ GM=V_in |
+| ASW (切 M/N) | batch 内 tile 共享行/列块 | S_A/S_B: **V_in** (含单侧全驻留调度, issue#32); S_C: 最小替换分组 GM (≥V_in, 仅双侧超L2 时 >1) | S_A/S_B: b·[(n_cnt−1)·a_b + (m_cnt−1)·b_b]; S_C: 组内窗口计 L2 (issue#32) | ✓ GM≥V_in |
+| StreamK | 组内 K 段零重复读; 组间共享块同 ASW | V_in (命中时) | 组间共享块 (命中时) | ✓ |
+| 特殊分支 K=1 | AIV 通路, 每元素一次 | V_in | 0 | ✓ |
+| 转Matmul | 折叠单次 GEMM | V_in | 0 (Matmul 精切未展开) | ✓ |
+
+→ **整芯片口径统一后 (issue#29): GM 均 = V_in (无重复读结构) 或 ≥ V_in (S_C 双侧超 L2 的最小替换分组)**, 不变量 GM ≥ V_in 由测试锁死 (违反即 bug)。
+
+---
+
+## 4. 场景模型 (输入侧) 与输出落点 (issue #30 / #32)
+
+### 4.1 场景判定 (对切 M/N 及切 B 分支共用; L2 为整芯片 128MB)
+
+- **S_A 整 case 全驻留**: V_in + V_out [+ workspace] ≤ L2 → 输出驻留 L2 (R4);
+- **S_B 单 batch 可驻留**: 不满足 S_A, 但单 batch 输入可全驻留 (a_b + b_b ≤ L2)
+  **或单侧全驻留 + 对侧滑窗**: b_b + 2·block_a ≤ L2 或 a_b + 2·block_b ≤ L2
+  (block_a = a_b/m_cnt = A 行块字节, block_b = b_b/n_cnt = B 列块字节)
+  → GM = V_in 一次, 共享块重复读全部命中 L2 读口; 输出直写 GM;
+- **S_C 双侧均超 L2**: a_b、b_b 及其滑窗组合都放不下 → 最小替换 2D 分组 (§4.3)。
+
+判定顺序 S_A → S_B → S_C。S_B 从原"双侧全驻留"扩展出"单侧全驻留+对侧滑窗"调度
+(issue#32): 大矩阵行块流式、小矩阵全驻留时, GM 仍 = V_in, 不再被分组模型成倍放大。
+(原 #24 场景 A 的"单 batch 输入+输出可驻留"判定被 S_A 取代: 整 case 输出跨 batch
+累积时单 batch 判定不可靠, 见 #30。)
+
+### 4.2 R4 输出落点规则 (取代一切"恒直写 GM / 单 batch 判定驻留")
+
+> to_l2(输出写 L2 写口 5.2TB/s, GM 写流量=0, 异步回写不占算子时延 — #23 口径延续)
+> ⟺ **V_in + V_out [+ StreamK workspace] ≤ L2** (整 case 判定)
+> 否则输出**直写 GM**: 计入 GM 读写共享总线, 与读累加进 MTE2 搬移链 (#23 口径)。
+
+理由: 输出字节只写一次无复用价值, L2 应优先保输入工作集 (输入存在重复读);
+只有当"整 case 输入+输出"可同时全驻留时才值得让输出占 L2 (省掉整条 GM 写时延)。
+各分支应用:
+
+| 分支 | 原行为 | 新行为 |
+|---|---|---|
+| IterBatch / MergeBatch / 特殊分支 / 转Matmul | 恒 direct_gm | S_A 时 to_l2; 否则直写 GM |
+| ASW | 场景 A (单 batch in+out 驻留) 时 to_l2 | S_A (整 case) 时 to_l2; S_B/S_C 直写 GM |
+| StreamK | 归约内最终写回恒走 L2 | 最终写回按 S_A(+workspace); 不满足时写 GM (drain 内按 GM 带宽) |
+
+### 4.3 场景 C: 最小替换 2D 分组 (issue#32)
+
+S_C (双侧均不可全驻留) 时, 在**整 L2 容量约束**下搜索最小 GM 的分组
+(取代旧版 "L2/2 对半预算 + 组内窗口流量零计" 口径; 每 batch 计算, 全 case ×b):
+
+- 容量约束: m_grp·block_a + n_grp·block_b ≤ L2 (组工作集可占满整 L2);
+- 目标: GM = ⌈n_cnt/n_grp⌉·a_b + ⌈m_cnt/m_grp⌉·b_b **最小**
+  (A 行块在其列组内一次 GM 首读、B 列块在其行组内一次 GM 首读; GM ≥ V_in);
+- 窗口 L2: (n_cnt − ⌈n_cnt/n_grp⌉)·a_b + (m_cnt − ⌈m_cnt/m_grp⌉)·b_b
+  (组内共享块其余次复用走 L2 读口, 与 S_B 口径一致);
+- 无可行分组 (单块对都超 L2) 时保守回落: 每共享块独立落 GM (⌈·⌉ = n_cnt/m_cnt), 窗口不计。
+
+分组追求"最小 L2 替换": 一次只让一组的工作集占 L2, 组间共享块落空回 GM。
+
+### 4.4 字节列与计数列语义 (修正 ①, 整芯片口径)
+
+输出 CSV 中数据量列一律为**整芯片**口径: gm_read_bytes / l2_read_bytes /
+fixpipe_bytes / cube_flops = 整芯片量; 时延列 t_* 为墙钟时延 (已含核数折算);
+dma_cmd_count 为**单核**命令数 (各核 DMA 引擎并行执行, 墙钟 T_cmd = 单核命令数 × t_cmd),
+与字节列口径不同属, 单独标注。
+
+---
+
+## 5. 各分支 GM/输出计账公式落地 (与代码对应)
+
+统一由 case 属性提供 V_in / V_out; 输出落点判断函数 `output_to_l2(case, spec, workspace)` (S_A)。
+
+| 分支 | GM 读 (芯片) | L2 读 (芯片) | 输出 |
+|---|---|---|---|
+| MergeBatch | V_in (每字节恰一次; K 截断/L1 绑定均无重复读, issue#31) | 0 | S_A→L2 写口; else GM |
+| IterBatch | V_in (a/b/c/d 各形态均每字节恰一次, issue#31) | 0 | 同 MergeBatch |
+| ASW 切M/N | S_A/S_B: V_in; S_C: b·[⌈n_cnt/n_grp⌉·a_b + ⌈m_cnt/m_grp⌉·b_b] (最小替换分组) | S_A/S_B: b·[(n_cnt−1)·a_b + (m_cnt−1)·b_b]; S_C: b·[(n_cnt−⌈n_cnt/n_grp⌉)·a_b + (m_cnt−⌈m_cnt/m_grp⌉)·b_b] | S_A→L2; else GM |
+| StreamK | V_in (组间共享命中) | b·[(n_cnt−1)·a_b + (m_cnt−1)·b_b] (命中时) | 最终写回: S_A(+ws)→L2; else GM |
+| 特殊分支 | K=1: V_in; K=0: 0 | 0 | S_A→L2 写口; else GM (K=0 只有输出) |
+| 转Matmul | V_in (折叠后) | 0 | S_A→L2; else GM |
+
+> **搬移效率 (时间列口径, issue#36)**: 上表为**字节量**口径 (不受效率模型影响);
+> 切B 两分支的 GM→L1 **时间**按单命令 tile 效率加权 t = (V_A/eff_A + V_B/eff_B)/W_GM,
+> eff = min(1, tile/min_TileSize), tile = nValue×dValue×dt。MergeBatch 合并使 tile
+> 放大 b0 倍 -> 效率项是 T_cmd=0 时 MergeBatch 的主要收益来源 (详见
+> docs/02_分支理论/01_MergeBatch分支.md §4~5)。ASW/StreamK 单命令 tile 大、效率
+> 恒饱和, 不接入 (极端小 tile 形状走 issue#34 "效率降级"标注通道)。
+
+时延计算: 全核并发时 t = 芯片字节 / 芯片带宽; 降核 (used < C) 时按 used×单核份额
+(线性假设, issue#26 标注), Cube t = 芯片 flops / (used × q_cube(dtype))。
+
+---
+
+## 6. MergeBatch Cube 复核 (issue #27) — 证明
+
+合并语义: A'[b0·M, K] @ B'[K, b0·N] 作为单次 GEMM, Cube 硬件计算**全网格**
+(b0·M)×(b0·N) 输出 (含交叉项, 冗余比例 (b0²−b0)/b0² = (b0−1)/b0), 只保留 b0 个对角块。
+
+- 每次合并计算 (步): flops_step = 2·(b0·M)·(b0·N)·K;
+- 每核合并组数: b_core/b0;
+- 总: flops = (b_core/b0)·2·(b0·M)·(b0·N)·K = **b_core·b0·2MNK** —— 与现行代码
+  `flops_pc = b_core·b0·2·m·n·k` 逐项相等, t_mmad = flops/Q16 不变。
+
+CSV 证据 (merge_demo_k_trunc, B=2048 M=N=32 K=256, b0=4, b_core=64):
+flops_step = 2·128·128·256 = 8,388,608; 步数 16; 合计 134,217,728 = 代码值 = CSV cube_flops 列;
+t_mmad = 134,217,728 / 13.5e12 = 9.942us = CSV t_mmad (Cube-only 口径 issue#40;
+旧总算力口径 15.1875T 时为 8.837us)。**结论: 无需数值修改** (指 flops 公式);
+若意图是"只算对角块"则与全网格 GEMM 语义冲突 (冗余消失, v0.98 §五 理论前提需另行裁决)。
+
+---
+
+## 7. 变更影响与验证口径
+
+1. 字节列整芯片化 (IterBatch/MergeBatch/special 数值 ×C 或按 b 直接计算);
+2. S_A 判定下输出写 L2: 小 case (整 case 可全驻留) GM 写流量清零、输出时延降为 L2 写口;
+3. S_A 之外输出 GM 写流量如实计入共享总线 → 修正单 batch 驻留判定导致的漏计;
+4. dtype 感知算力: bf16 案例数值零回退; fp32/fp8/fp4/… 案例时延按 §2.1 表修正;
+5. 新增回归: GM≥V_in 不变量 (全分支随机集)、S_A 输出落点开关、dtype 速率比例、
+   fp32 Cube 时延 = bf16 的 2 倍、字节列整芯片口径等;
+6. examples 三件套重生成, 与模型改动前 diff 隔离后入库; 压力回归 0 违规/0 NaN。

BMM/BMM_Theory/docs/06_落地验证计划_理论方案_vs_bmmv3源码.md

@@ -0,0 +1,391 @@
+# 06 理论方案落地 batch_mat_mul_v3 源码 — 比对分析与验证计划
+
+> 目的：把 BMM_Theory 的理论最优方案落到真实算子源码 `batch_mat_mul_v3`（ops-nn），
+> 上板（**Ascend950PR 主 bin，arch35 / DAV_3510**）实测验证收益，达标后向真实算子源码仓提 PR。
+> 比对基线源码：`\\HwFs\HW_WorkBuddy\昇腾NPU\昇腾NPU知识库\代码仓\ops-nn\matmul\batch_mat_mul_v3`
+> （下称"源码"，行号为该仓当前快照）；理论侧：本仓 `bmm_theory/` + `docs/02_分支理论/`。
+> 文中所有"预期收益"数字均为**理论模型测算**（含 T_cmd=0 等未标定假设），仅用于排优先级，
+> 最终以 msProf 上板实测为准。
+
+---
+
+## 0. 结论摘要（TL;DR）
+
+| 优先级 | 分支/主题 | 源码现状一句话 | 理论差异一句话 | 模型预估收益（定向 case） | 改动面 | 建议批次 |
+|---|---|---|---|---|---|---|
+| P0 | **StreamK 准入放宽** | 固定 K≥8192 门槛，拒绝后落 ≤4 核 ASW | 动态 grid_K + θ_c≈10.9 归约代价判据（K>~360 起可进） | **总时延 −73%~−82%**（K=4096~6144 细长 case） | **纯 host tiling**（kernel 已有完整 StreamK 实现） | 批次 1 |
+| P1 | **ASW 尾轮重切** | 尾轮 `index≥totalCnt` 空转（A0）；ASW_BASIC 仅做单套 tile 的均衡率枚举 | 方案B（全局一套 tile 重切）/A1b（尾轮第二套 tile），A0 在 r>0 时严格劣 | **总时延 −25%~−43%**（2 波 + 尾轮 ρ=0.125~0.5 的 case） | 方案B 纯 host；A1b 需 kernel 读第二套尾轮参数（共享库有现成范式） | 批次 2 |
+| P2 | **MergeBatch 准入补齐 + 仲裁接线** | 准入偏宽（缺搬移量/tile/算存比三条），且静态优先级恒先于 IterBatch | 理论 5 条件 + 与 IterBatch 的净收益仲裁（L1 绑定时 MergeBatch 恒劣） | 防误捕获（小搬移量 case 冗余计算/小 tile 低效）；仲裁差异 T_cmd=0 下 ≤0.5% | **纯 host tiling** | 批次 3 |
+| P3 | **IterBatch 形态补全（a/c/d）** | 仅"整 K 双侧驻留 + 双 batch 乒乓"（形态 b），放不下 L1 直接拒绝 → 落 ASW | 理论四形态（a 单 batch 驻留 / c 一侧驻留+对侧切K / d 两侧切K），k_l1 反推不受 512B 上限 | T_cmd=0 下 k_l1 差异本身 ≈0（<0.6%）；收益=命令数 4× 减少 ×T_cmd + 消除 ASW 承接时的 L2 重复读 | **kernel 重写 mmad 流水（最难）** + host 准入 | 批次 4 |
+| P4 | 转Matmul / 特殊分支 / 广播 | 已有 TO_MUL、K_EQ_ZERO、ITER_BATCH_BROADCAST、MergeBatchAndMAxis 折叠 | 理论基本一致（BatchA=1 大 A 的"重排转 Matmul"源码无，低优先） | 对齐核查即可 | 核查 | 随批附带 |
+
+**总体判断**：收益大头在 **StreamK 准入**与 **ASW 尾轮** 两处，且都可先做 **host-only** 改动拿到大部分收益；
+IterBatch 形态补全改动最大、收益依赖 T_cmd 标定结果，放最后。
+
+---
+
+## 1. 源码现状架构速览（arch35 / DAV_3510）
+
+### 1.1 host 侧：策略注册表 + 静态优先级
+
+`batch_mat_mul_v3_tiling.cpp`：`IsAdvancedSocVersion` → arch35 advanced 路径
+（`arch35/batch_matmul_v3_tiling_advanced.*`），按 `batch_matmul_v3_tiling_strategy.h`
+的优先级表**逐策略试 `IsCapable()`，首个通过即中，无回退、无时延仲裁**：
+
+```
+K_EQUAL_ZERO(0) → TO_MUL(1) → STREAM_K(2) → MERGE_BATCH_BASICAPI(3)
+→ ITER_BATCH_BROADCAST_BASICAPI(4) → ITER_BATCH_BASICAPI(5) → ITER_BATCH(6)
+→ AL1_FULL_LOAD_BASIC(7) → BL1_FULL_LOAD_BASIC(8) → ASW_BASIC(9) → BASE(999)
+```
+
+与理论决策树（`router.py`）的结构差异：
+
+| 项 | 源码 | 理论 |
+|---|---|---|
+| 路由方式 | 静态优先级 + IsCapable 一票否决/通过 | 决策树 + 重叠区端到端时延模型仲裁 |
+| StreamK vs 切B | StreamK **优先于** MergeBatch/IterBatch | 切B（B≥C）优先，StreamK 仅在 P≤C/2 |
+| Merge vs Iter | MergeBatch 恒优先（无仲裁） | 净收益仲裁 + 时延模型终审 |
+| 兜底 | ASW_BASIC（等 batch）/ BASE（广义，固定 256×256 起步） | ASW_Basic（含降核）恒出方案 |
+
+### 1.2 kernel 侧：7 参模板分发
+
+`op_kernel/arch35/batch_mat_mul_v3.cpp:171` 起：模板参数
+`<API_LEVEL, A_TRANS, B_TRANS, ITER_MODEL, BMODEL, FULL_LOAD, L0C2OUT>`，`if constexpr` 链分发：
+IterBatch（cmct `KernelMatMulIterBatch` / 旧库 `MultiBatchKernel` 两世代并存）、
+MergeBatch（cmct `KernelMatMulMergeBatch`）、ASW（`BatchMatMulAswKernel/AswBlock`）、
+StreamK（Blaze `MatMulStreamKKernel` 或内置 `MatMulStreamKActKernel`）等。
+**关键事实**：StreamK 的 kernel（动态 kCnt 调度、fp32 部分和、AIV 归约）**已完整存在**；
+ASW 尾轮重切所需的 `MatMulV3TailInfo{mCnt,nCnt,kCnt,mTailMain,nTailMain}` 字段在
+`MatMulV3TilingData` 已存在（共享 mat_mul 的 `block_scheduler_aswt.h`/`mat_mul_asw_block.h`
+有现成"尾 tile 再切子块"范式），**batch 版 AswBlock 只是没有读这些字段**。
+
+### 1.3 口径天然对齐（好消息）
+
+源码 tiling 内的平台公式与理论硬件口径一致：
+`GetHbmBW = freq×32×31/1024 ≈ 1.6TB/s`；`GetL2BW = freq×32×100/1024 ≈ 5.16TB/s ≈ 理论 5.2TB/s`；
+`singleCoreComputePower = freq×8(K) ≈ 13.2T ≈ 理论 Cube-only 13.5T`（差 2%，经验取整）。
+→ 理论模型与源码 tiling 的"带宽/算力世界观"兼容，参数级差异小，差异集中在**策略结构**。
+
+---
+
+## 2. 分支逐项比对与修改重点
+
+### 2.1 StreamK（P0，最大收益，host-only 可落地大半）
+
+**理论方案**（`docs/02_分支理论/05_StreamK分支.md`）：
+
+- 进入：P = B·M·N·4B/L0C ≤ C/2；K/grid_K ≥ 256B/dtype；**K > grid_K²/(grid_K−1)·θ_c，θ_c≈10.9**
+  （grid_K=32 → K>360）；确定性等级 ≤1 且 ND。
+- grid_K 动态：blocksPerBatch=⌊C/B⌋，grid_K=blocksPerBatch/(mCnt·nCnt)，mCnt/nCnt 收拢为因子。
+- 部分和 fp32 写 workspace（防精度丢失），AIV 归约后按 C dtype 写出；workspace 应尽量驻留 L2
+  （落 GM 时归约带宽 5.2→~0.64TB/s，θ_c 升至 ~86）。
+
+**源码现状**（`arch35/batch_matmul_v3_basic_streamk_tiling.cpp`）：
+
+- 准入 `CheckStreamKSKTiling`：**固定门槛 `CeilAlign(K,256) ≥ max(8192, aicNum×256B/dtype)`**
+  （bf16 即 8192）；`batch·mCnt·nCnt ≤ aicNum/2`（mCnt/nCnt 按 256 估算）；fp32 非 HF32 且 K>2e6 拒绝。
+- grid_K 已有动态雏形：`tailInfo.kCnt = ⌊aicNum/(batch·mCnt·nCnt)⌋`，`singleCoreK=⌈K/kCnt⌉`
+  （落盘 `skSingleCoreK`）；mCnt/nCnt 有">blocksPerBatch/3 且 <blocksPerBatch/2 则收拢到 /2"的启发式。
+- kernel 侧（`BlockSchedulerStreamKBuiltIn` / `BlockEpilogueStreamK`）完整：AIC 写 fp32 部分和到
+  **GM workspace**（host 分配 `aicNum×256×256×4B + RPC`），AIV `CrossCoreWaitFlag + SyncAll`
+  后读回 UB 级联加、Cast 写 C。**workspace 名义 GM，经 L2 缓存；容量够时实际命中 L2**（待实测确认）。
+
+**差异清单**：
+
+| # | 差异 | 影响 |
+|---|---|---|
+| S1 | 固定 K≥8192 vs 理论动态 K>360（grid_K=32） | **K∈(360, 8192) 的细长 case 全部被拒绝**，落 ASW 只用 batch·mCore·nCore ≤ C/2 核 |
+| S2 | 准入无"归约代价 vs 不切 K"比较（θ_c 判据） | 门槛同时承担了 dValue 与归约代价两职，过保守 |
+| S3 | workspace 大小固定 `C×256×256×4B`（8MB+RPC） | 与 mCnt/nCnt/kCnt 实际解耦，偏大但不阻断 |
+| S4 | mCnt/nCnt 启发式收拢 vs 理论"收拢为 blocksPerBatch 因子最大化 grid_K" | 参数级，影响 K 并行度 |
+| S5 | workspace 显式 L2 驻留无法由 op 侧表达（只有 SetL2CacheHint） | 架构级，暂不做，靠 L2 自然缓存 |
+
+**修改重点（host-only，批次 1）**：
+
+1. `CheckStreamKSKTiling`：把固定 8192 门槛替换为**理论两条判据**——
+   先按现行逻辑试算 mCnt/nCnt/kCnt，然后要求 `K/kCnt ≥ 256B/dtype`（条件 2，dValue）且
+   `K > kCnt²/(kCnt−1)·θ_c'`（条件 3，归约代价）。θ_c' 取 **GM 口径 ~86**（保守，workspace 名义 GM）
+   或按 workspace ≤ L2 余量切换 L2 口径 10.9——建议第一版用 GM 口径（grid_K=32 时 K>~2841），
+   上板测得归约段实际命中 L2 后再放宽。
+2. mCnt/nCnt 收拢规则替换为"blocksPerBatch 因子 + grid_K 最大化"（理论 §4）。
+3. 保持 kernel 不动（tiling 字段 `skSingleCoreK`/`tailInfo.kCnt` 语义不变）。
+
+**预期时延变化**：减少的是 **MTE2 稳态段**（并行度 4→32 核）与 MMAD 段同比例；
+新增有限的 **归约尾（t_reduce）**。模型测算（bf16）：
+
+| case | 源码现状（ASW 降核，核数=batch·mCore·nCore） | 理论 StreamK | 总时延节省 |
+|---|---|---|---|
+| B=4, M=N=128, K=4096 | 41.9us（4 核） | 8.65us（steady 5.24 + reduce 3.41） | **−79.4%** |
+| B=4, M=N=128, K=6144 | 62.9us（4 核） | 11.27us | **−82.1%** |
+| B=8, M=256, N=128, K=4096 | 62.9us（8 核） | 16.60us | **−73.6%** |
+
+敏感性：若 workspace 实际走 GM 带宽（不命中 L2），K=4096 例 reduce 升至 ~10us，总 ~15.4us，
+仍省 ~63%。**反向保护**：θ_c 判据保证 K 太小（归约吃不下）时不进——理论 K=2048 例即被判不可行。
+
+**验证重点**：
+- 定向 case 集：B∈{2,4,8}，M/N∈{64,128,256}，K∈{2048, 2843±, 4096, 6144, 8000, 8192±, 10240}，
+  覆盖"源码拒绝→修改后进入"的跃迁区间与 θ_c 边界两侧（K∈{2841±} 附近加边界 case，验证不错进/不漏进）；
+- msProf 拆解：MTE2/MMAD 稳态是否如预期缩短，**归约段实测时延 vs 模型 3.4us（L2 口径）/~10us（GM 口径）**
+  —— 据此裁决 workspace 口径与 θ_c' 取值；
+- 正确性：fp32 部分和 + AIV 归约的数值精度（对比 golden，deterministic_level=0/1）；
+- 回归：K≥8192 老 case 时延不回退（grid_K 推导变化的影响面）。
+
+### 2.2 ASW_Basic 尾轮重切（P1，方案B 可 host-only）
+
+**理论方案**（`docs/02_分支理论/06_ASW_Basic分支.md` §6 + `07_尾轮处理策略.md`）：
+总块数 N_blk = B·mCnt·nCnt 不能整除 C 时，尾轮 r = N_blk mod C 个核干活、C−r 核空转一个整块。
+**A0（不重切）在 r>0 时严格劣**；面积型主导下 A1b（尾轮第二套小 tile 凑满核）与方案B（全局一套
+tile 均匀重切到 n_wave·C 块）理论时延严格相等 = (n_wave−1+ρ)·T_block；周长型且 ρ≥ρ_dv 时 A1b 恒优；
+r 小翻出时方案B 微优 ~6%。软件默认输出方案B，周长型 ρ≥ρ_dv 输出 A1b。
+
+**源码现状**：
+- kernel `asw_kernel_advanced.h` Process：`if (index < totalCnt)` 才干活，**尾轮 C−r 核真空转（A0）**；
+- host `ASW_BASIC` 的 `GetRebalanceBlock`（`mat_mul_v3_tiling_helper.cpp:387`）枚举 baseM/baseN
+  最大化"均衡率"，**但 `GetBalanceRateWithTail` 里尾轮切分项在 `batchInfo != nullptr` 时被禁用**
+  （:240 早退），即 batch 场景的均衡率按 A0 模型算——枚举只隐式接近"广义方案B"，且无
+  "搬入最小化"目标与尾轮翻出概念；
+- `BASE` 策略（广播/非等 batch 兜底）连 GetRebalanceBlock 都没有，固定 baseM=baseN=256 + A0。
+
+**差异清单**：
+
+| # | 差异 | 影响 |
+|---|---|---|
+| T1 | 尾轮空转（A0） | r>0 时白丢 (1−ρ)/n_wave 比例的总时延 |
+| T2 | 首轮 tile 枚举目标=均衡率/cubeBound 启发式 vs 理论"每 batch 搬入 K·dt·(nCnt·M+mCnt·N) 最小 + 尾轮 r 最大 tie-break" | L2 重复读量、dValue 效率差异 |
+| T3 | 无 A1b（尾轮第二套 tile） | 周长型主导区少拿 A1b vs 方案B 的差 |
+| T4 | 无降核口径对齐：源码 usedCoreNum=min(batch·mCore·nCore, C) vs 理论 P=L0C 满载粒度降核 | 并行度估计口径不同，需上板对齐 |
+
+**修改重点**：
+1. **方案B（host-only，先做）**：`GetRebalanceBlock` 之后追加"整轮均匀重切"——由 N_blk、n_wave、r
+   计算 g = n_wave·C/N_blk，把 baseM/baseN 在候选集内按 1/√g 收缩重选（或直接在枚举目标函数里
+   把 A0 均衡率换成 `w×T_block` 估计），使尾轮满载。kernel 零改动（一套 tile，totalCnt 变为
+   n_wave·C，天然无空转）。
+2. **A1b（kernel 配合，后做）**：host 计算第二套尾轮 tile 参数写入 `MatMulV3TailInfo` 现有字段；
+   `asw_block_advanced.h` 的 Init/UpdateBasicIndex/UpdateBlockParams/CalcGMOffset 按 roundIdx
+   切换主/尾相位——**直接移植共享库 `block_scheduler_aswt.h` 的尾 tile 子块范式**（mTailCnt/nTailCnt
+   现成字段），保持每子片单核单写（不引入跨核归约）。纯运行期数据驱动，**不增编译变体**。
+3. 尾轮策略选择逻辑（A0/A1b/方案B 的五步闭式判定）放在 host，输出到 tiling 备注字段便于验证对照。
+
+**预期时延变化**：减少的是 **drain/尾轮段**（(1−ρ)·T_block 量级），即 max 稳态之后的空转暴露。
+模型测算（理论 ASW evaluate，强制 A0 vs 最优尾轮）：
+
+| case | N_blk / 波次 / ρ | A0（源码现状） | 最优尾轮 | 总时延节省 |
+|---|---|---|---|---|
+| B=2, M=1024, N=1536, K=1024 | 48 / 2 / 0.50 | 19.88us | 14.91us（方案B） | **−25.0%** |
+| B=2, M=1024, N=1280, K=1024 | 40 / 2 / 0.25 | 19.88us | 12.43us（方案B） | **−37.5%** |
+| B=2, M=768, N=1536, K=1024 | 36 / 2 / 0.125 | 20.04us | 11.34us（方案B） | **−43.4%** |
+
+（同为 MMAD/MTE2 主导的面积型 case；周长型边角区 A1b 另有收益，量级待 case 库扫描。）
+注意：源码 ASW_BASIC 的均衡率枚举已能自发吸收其中一部分（相当于不完整的方案B），
+**实测基线必须先跑**，预期净收益 = 表值 − 源码枚举已拿到的部分；BASE 兜底命中的 case
+（广播、非等 batch）则全额可期。
+
+**验证重点**：
+- 定向构造 N_blk mod C ≠ 0 且 n_wave ∈ {1,2,3,≥4} 的 case 矩阵（B∈{2,3,5,6} × M/N 组合扫描），
+  确认尾轮核不再空转（msProf 核间时间线/各核结束时刻拉齐）；
+- 对比维度：总时延、尾轮起止时刻、GM/L2 流量（方案B tile 变小 → 搬入周长和 √g 放大，
+  理论已标注 P4 流量放大不入模型，**需实测是否吞掉收益**——这是该改动的头号验证风险）；
+- 正确性：尾轮重切后边界块（mBaseTail/nBaseTail）数值正确性；
+- 回归：r=0 的 case 零变化（应自动保持 A0）。
+
+### 2.3 IterBatch 形态补全（P2/P3，kernel 重写流水，收益依赖 T_cmd 标定）
+
+**理论方案**（`docs/02_分支理论/02_IterBatch分支.md`）：核间切 B 后，核内按 L1 容量四形态选一——
+a) 单 batch 全驻留（b_core=1）；b) 双 batch 乒乓（2(MK+KN)·dt ≤ L1，命中最多）；
+c) 一侧驻留+对侧切 K（驻留侧 ≤ L1/min(b_core,2)，b_core≥2 时另一半 L1 预取下一 batch 驻留侧）；
+d) 两侧都切 K（兜底，K 段成对流水）。四形态共同保证**单 batch 计算核内零重复读、GM=V_in**；
+k_l1 按容量反推（不受 512B 硬上限，dValue≥128B 下限约束）。
+
+**源码现状**：
+- `iterbatch_basicapi_tiling.cpp` IsCapable：`batchC > aicNum`（严格 B>C）+
+  **`2×((M·K+K·N)·dt+bias) ≤ L1`**——即只有形态 b；放不下直接 false（或 L0 放不下单 batch 时走
+  0.8 均衡率兜底），**无 c/d 形态的退化承接**；
+- kernel（cmct `BlockMmad` IterBatch 偏特化）实证：L1 布局=整 K 的 A/B 各 2 缓冲按 batch 叠放，
+  L1→L0 才按 baseK 切 K，mmad 循环 iter1(batch)→N→M→K；**代码中不存在"一侧驻留+对侧切K"或
+  "两侧切K"的显式模式**；`mmadParams.unitFlag` 恒 0（理论假设的 UnitFlag 16-granule 细粒度流水
+  在该路径未启用）；写回 L0C→GM 直写（ON_THE_FLY）或经 UB/AIV（ND_FIXPIPE_1_2），不经 L2。
+
+**形态 c/d 的 case 在源码里的实际去向**：落 ASW_BASIC/BASE——ASW 把 M/N 切到 ≤256 的 base 块、
+K 由 L1 tiling 切，共享块重复读走 L2。**不是灾难，但偏离理论最优**：
+1. M/N>256 时被切块 → 引入 (nCnt−1)·a_b+(mCnt−1)·b_b 的 L2 重复读（理论 IterBatch 为 0）；
+2. K 粒度受 512B 内轴对齐与 stepK≤8（issue queue）限制，单命令 tile 更小、命令数更多；
+3. 无 batch 边界预取（理论 c 形态半预算预取驻留侧，边界无气泡）。
+
+**模型测算**（形态 d 例：B=64, M=N=64, K=8192, bf16）：
+
+| 口径 | k_l1 | 单核 DMA 命令数 | 总时延 |
+|---|---|---|---|
+| 理论 IterBatch（k_l1 反推=1024） | 1024 | **16** | 85.0us |
+| 源码式承接（k_l1 上限 512B/dt=256，ASW 模拟） | 256 | **64** | 84.5us |
+
+**T_cmd=0 假设下时延差 <0.6%**——形态 d 的收益几乎全部押在 **T_cmd > 0**（每核命令少 4×）
+与 L2 重复读消除上。→ **T_cmd 标定是该批次的入场券**（见 §4.3）：若标定出 T_cmd 可观
+（如 ≥100ns），form_d 例命令差 48 次 → ~5us 级收益；若 T_cmd≈0，则本批次降级为
+"框架统一 + 大 M/N case 的 L2 重复读消除"（form_c 型例：ASW 承接的 L2 重读 ~96MB ≈ 18.5us 级，
+需实测扣除其流水掩盖）。
+
+**修改重点**（若标定结果支持做）：
+1. host `iterbatch_basicapi_tiling.cpp`：IsCapable 放宽——L1 放不下双 batch 整 K 时进入形态
+   选择（a/c/d），按理论公式算驻留侧/k_l1；DoOpTiling 输出形态字段；
+2. kernel：cmct `BlockMmad`(IterBatch) 的 L1 布局与循环次序重写（c 形态：驻留侧一次 CopyIn、
+   对侧 K 分块双缓冲，batch 边界半预算预取；d 形态：两侧 K 分块成对流水）——
+   **这是全计划最难的 kernel 改动**（L1 flag 4 组 + L0 乒乓 + L0C 半区 + AIC↔AIV 握手耦合极密）；
+   tilingData `BatchMatMulV3IterBatchBasicTilingData` 加形态/分侧参数字段（pack(8) 对齐，
+   host/kernel 同步改）；尽量纯运行期字段表达以**避免新增模板组合/二进制**；
+3. 同步考虑启用 unitflag（细粒度排空）以压低 drain——理论 drain 模型（末 batch T_comp+T_write）
+   在 unitflag=0 的源码上可能**低估**真实排空，需实测校准（这也影响理论模型自身的 drain 闭式）。
+
+**验证重点**：T_cmd 标定先行；形态 c/d 定向 case（M/N>256 且 2(MK+KN)>L1、以及双侧超 L1）的
+总时延与 L2 流量对比；batch 边界气泡（核内时间线）；正确性（K 分段累加顺序变化）。
+
+### 2.4 MergeBatch 准入补齐 + 仲裁接线（P2，纯 host）
+
+**理论方案**（`01_MergeBatch分支.md`；与 bmmv3 对拍器口径的逐项对照另见
+`04_差异对照_bmmv3_vs_BMM_Theory.md` §2）：5 条件准入（含 ③ 单核搬移量 ≥480KB、④ 单 tile ≥16KB、
+⑤ AI < R16/b0 访存 Bound 守卫），b0 = min(L0C/L0A/L0B 容量上限, R16 算存比上限, b_core) 取因子；
+k_l1 = min(L1 反推, K, 512B/dt)；与 IterBatch 重叠区按净收益仲裁
+（命令节省×T_cmd + 搬移效率节省 − drain 惩罚；**L1 绑定（k_l1<K 且两侧 tile 饱和）时恒劣**）。
+
+**源码现状**（`mergebatch_basicapi_tiling.cpp` + kernel 实证）：
+- 准入：batch 四轴全等、**batchC ≥ 4·C**（对应理论 b_core≥2b0@b0=2）、K 对齐 ≥64、**M≤N**、
+  L0A/L0B/L0C 容量查（min 4 batch）——**缺理论条件 ③④⑤**；
+- b0（mergeBatchL0）= min(L0C 多项式解, L0B 上限, ceil(B/C))——无 L0A 显式收敛（靠 min4 间接保证）、
+  无算存比上限；
+- k 粒度：baseK=min(L0B 反推, 64)，L1 四缓冲 stepKa ≤ min(⌈b_core/b0⌉, L1/4/(b0·M·baseK·dt))——
+  无 512B 推荐值概念，但 L1 约束下与理论值经常重合（见下）；
+- kernel 语义与理论一致：batch 并入空间维做稠密 (b0·M)×(b0·N) MMAD，对角片写回
+  （Fixpipe ndNum=b0、对角 stride），冗余比例 (b0−1)/b0 与理论 §6 证明一致；bias 不支持
+  （wrapper 直接 return，理论同样不建模 bias）；unitflag 恒 0。
+- **仲裁缺失**：优先级表 MERGE_BATCH(3) 在 ITER_BATCH(5/6) 之前，IsCapable 通过即中，
+  L1 绑定情形无法回退 IterBatch。
+
+**差异清单与修改重点**：
+
+| # | 差异 | 修改 |
+|---|---|---|
+| M1 | 缺条件③（单核搬移量 ≥480KB） | IsCapable 追加：小搬移量 case 防误捕获（此类 case 时延绝对值小、启动开销占比高，合并收益不真实） |
+| M2 | 缺条件④（tile ≥16KB） | IsCapable 追加：tile 不足时有效带宽线性退化（eff=tile/16KB），合并红利不成立 |
+| M3 | 缺条件⑤（AI < R16/b0） | IsCapable 追加：防把计算 Bound case 抓进来让冗余计算变瓶颈 |
+| M4 | b0 无 L0A 显式收敛/无算存比上限 | DoOpTiling b0 计算并入两上限（理论 issue#13 口径） |
+| M5 | 无 Merge vs Iter 仲裁 | 两条路：(a) 简单——IsCapable 里并入"L1 绑定判据"（k_l1 反推 <K 且 tile 饱和 → false，放给 IterBatch）；(b) 完整——引入轻量时延估计择优。**建议先 (a)** |
+| M6 | M≤N 限制、K≥64 限制 | 源码实现限制（B 沿 N 合并+对齐），保留并写入文档说明 |
+
+**预期时延变化**：参数层（b0/k_l1）在已命中 case 上模型测算差异为 **0%**（两者经常收敛到同值）；
+收益全部在**防止误捕获**与**仲裁防劣化**：L1 绑定区 MergeBatch 相对 IterBatch 的劣势在
+T_cmd=0 下 ≤0.5%（drain 放大项），T_cmd>0 时放大 b0 倍——**同样依赖 T_cmd 标定**。
+条件③④⑤ 误捕获区的典型代价：tile=4KB → 有效带宽 ×0.25，或计算 Bound 下 b0 倍冗余 MMAD。
+→ 该批次定位为**健壮性/正确性加固**，收益以上板实测误捕获 case 为准。
+
+**验证重点**：构造条件③④⑤ 各自的越界 case（小搬移量、小 tile、高 AI），确认修改后不再进
+MergeBatch 且落点分支时延更优；L1 绑定仲裁区 case（merge_iter_arbitrate 类：B=128, M=N=64,
+K=512）；已命中 case 全量回归零变化。
+
+### 2.5 转Matmul / 特殊分支 / 广播（P4，对齐核查）
+
+| 理论 | 源码对应 | 核查结论 |
+|---|---|---|
+| K=0 纯 AIV 写值 | `K_EQUAL_ZERO(0)`（AIV 清输出 kernel） | ✅ 已覆盖，对齐即可 |
+| K=1 逐元素乘走 AIV | `TO_MUL(1)`（matmul2mul tiling + vector kernel） | ✅ 已覆盖；理论 B<128 的 AIV 单缓冲退化需核查源码对应行为 |
+| BatchB=1 零代价折叠 | `MergeBatchAndMAxis`（batchB==1 且 !isATrans → 折进 M） | ✅ 已覆盖 |
+| BatchA=1：A 小留 BMM 广播 / A 大比较重排 | ITER_BATCH_BROADCAST(4)（单广播轴）/ ASW 广播模型 | ⚠️ "A 大则重排转 Matmul"源码无，低优先 |
+| 交叉广播（双轴/多轴） | 落 ASW_Basic/BASE | ✅ 与理论一致（理论也由 ASW 承接） |
+
+### 2.6 路由层差异（贯穿项）
+
+1. **优先级顺序**：源码 StreamK(2) 先于 MergeBatch(3)/IterBatch(5)。重叠区（B≥C 且 P≤C/2 且
+   K≥门槛）源码必选 StreamK，理论先走切 B。模型测算该区域两者接近（切 K 并行度收益 ≈ 切 B 免
+   共享收益），**维持源码顺序不改**，用定向 case 实测确认无显著回退即可。
+2. **仲裁机制**：源码无"时延模型终审"。短期用各分支 IsCapable 的判据细化（M5(a)、S1/S2）逼近
+   理论裁决；长期若要引入端到端时延估计进 tiling，需单独评审（tiling 耗时预算、常量标定维护）。
+
+---
+
+## 3. 横切差异（时延模型与口径）
+
+| 项 | 理论模型 | 源码 tiling | 处置 |
+|---|---|---|---|
+| 算力口径 | Cube-only 13.5T/核（432T/32 核，issue#40） | freq×8K≈13.2T（经验式） | 一致（差 2%），无需改 |
+| 带宽口径 | GM 1.6TB/s 读写共享；L2 5.2TB/s 读写独享 | GetHbmBW≈1.6T、GetL2BW≈5.16T | 一致，无需改 |
+| 输出落点 | S_A（整 case ≤L2）时写 L2 写口，否则直写 GM | 全部直写 GM，C 从不显式写 L2 | **观察项**：S_A 小区 case 理论上有 GM 写时延节省；硬件 L2 写回自然缓存可能已近似，msProf 实测后再决定是否建模差异 |
+| T_cmd（DMA 命令固定开销） | 默认 0（未标定） | 源码无显式建模，但 stepK≤8、dValue 512B 等限制隐含"命令有代价"的工程经验 | **标定项**（§4.3），结果决定 IterBatch/MergeBatch 仲裁类收益是否成立 |
+| 搬移效率 | eff=min(1, tile/16KB)（issue#36） | 源码用 16KB/64KB 硬编码阈值（fullCopySize=64KB、mmadCount=8） | 同源不同形，验证 tile 饱和点 |
+| drain | 闭式（issue#37） | kernel unitflag=0，tile 级双缓冲排空 | 实测校准理论 drain |
+| fp32 大 K | 不建模 | 核内 splitK（8192 阈值串行累加） | 理论补标注即可，不动源码 |
+
+---
+
+## 4. 上板验证方案（Ascend950PR 主 bin / arch35）
+
+### 4.1 环境与工具
+
+- 芯片：Ascend950PR 主 bin（32 AIC / 64 AIV，HBM 1.6TB/s，L2 128MB）；arch=DAV_3510。
+- 入口：`aclnnBatchMatMul`（`examples/arch35/test_aclnn_batchmatmul.cpp` 为模板自建 harness），
+  dtype 以 bf16 为主测面，fp16/fp32 抽查。
+- 测量：**msProf** 算子级时延（总时延 = case 执行完的墙钟）+ 核间时间线（尾轮空转可视化）；
+  正确性用 `tests/assets/bmmv3_aclnn_golden.py` 对拍（bf16 容差按现有 ST 标准）。
+- 对照：同一 case 跑**修改前源码 build**（基线）与**修改后 build**，各 ≥20 次取中位数；
+  tiling 日志（OP_LOGI/D）落盘确认实际命中策略与理论预测分支一致。
+
+### 4.2 case 集设计（三层）
+
+1. **定向差异 case**（每批次一张表，见 §2 各节"验证重点"）：围绕每个差异点的进入/退出边界构造，
+   含边界两侧 ±1 档；
+2. **理论 demo 全量**：本仓 `examples/cases_demo.csv` 44 例全跑，与理论 recommend 分支预测比对
+   "预测分支 == 实际命中策略"命中率（目标 100%，除已知口径差异清单）；
+3. **回归面**：源码自带 ST（`tests/st/arch35/ttk_*.csv` + atk json）全量，保证零回退；
+   另加理论压力回归同分布随机 case（seed7/6000 量级采样到可编译子集）抽查时延不回退。
+
+### 4.3 标定项（先行，阻塞批次 3/4 的裁决）
+
+| 标定 | 方法 | 用途 |
+|---|---|---|
+| **T_cmd** | 固定 case，变 k_l1 分段数（命令数）扫 MTE2 时延，线性拟合斜率=带宽、截距/命令数=T_cmd | 决定 IterBatch 形态化与 Merge 仲裁的真实收益（批次 3/4 立项裁决） |
+| 归约段落点 | StreamK case 扫 K，拆 reduce 段时延，对比 L2(5.2T)/GM(0.64~1.6T) 两口径模型 | 决定 θ_c' 取值与 workspace 策略（批次 1 内完成） |
+| tile 效率曲线 | 变单命令 tile 大小（4/8/16/32KB）测有效带宽 | 验证 eff=min(1,tile/16KB) 模型与 16KB 饱和点（issue#36 口径上板确认） |
+| drain 量级 | 单波 case 变 N_blk 测尾轮/排空暴露 | 校准理论 drain 闭式（unitflag=0 下） |
+| fp32 Cube 算力 | fp32 大 K 计算 Bound case | 确认 ½ 假设（docs/05 §2 待标定项） |
+
+### 4.4 通过标准（提 PR 门槛）
+
+1. 定向差异 case：修改后总时延 **严格更短**（中位数，超过噪声带 ~2%）；
+2. 回归面：零精度问题、零 crash、时延回退 case 比例 = 0（个别 <2% 且可解释除外）；
+3. 理论预测分支与实际上板命中策略一致率 100%（不一致项必须有文档化解释）；
+4. tiling 耗时不显著增长（host 枚举量增加控制在毫秒级）。
+
+---
+
+## 5. 实施批次与 PR 计划
+
+| 批次 | 内容 | 改动文件（源码侧） | 依赖 | 风险 |
+|---|---|---|---|---|
+| 0 | 标定（§4.3）+ 基线库建立（44 例 + 定向 case 的修改前时延基线） | 无（测试基建） | 板子/harness | 低 |
+| 1 | StreamK 准入放宽 + grid_K 推导（§2.1） | `op_host/op_tiling/arch35/batch_matmul_v3_basic_streamk_tiling.cpp`（host-only） | 批次 0 归约口径 | 低（kernel 不动）；注意 θ_c 边界 |
+| 2 | ASW 方案B 尾轮（host-only）→ 视收益决定是否续作 A1b（kernel 第二套 tile，移植 aswt 范式） | `batch_matmul_v3_asw_basic_tiling.cpp` +（A1b 时）`asw_block_advanced.h` / `asw_kernel_advanced.h` | 批次 0 | 中（方案B 低；A1b 的 index 重映射需细致 review） |
+| 3 | MergeBatch 准入补齐 + L1 绑定仲裁（§2.4） | `batch_matmul_v3_mergebatch_basicapi_tiling.cpp`（host-only） | T_cmd 标定 | 低 |
+| 4 | IterBatch 形态补全（§2.3），含 unitflag/ drain 校准 | `iterbatch_basicapi_tiling.cpp` + cmct `block_mmad_iterbatch.h` / scheduler + tilingData | T_cmd 标定结果支持 | **高**（kernel 流水重写）；可拆子 PR（先 host 形态 d → 再 kernel c/d） |
+
+每个批次独立 PR：附理论文档链接（本仓 docs/02 对应分支文档 + 本文）、基线/修改后时延对照表、
+ST/UT 通过证据、定向 case 清单。PR 顺序即批次顺序（host-only 先行，kernel 改动殿后）。
+
+---
+
+## 6. 风险与开放问题
+
+1. **T_cmd 若 ≈0**：批次 3 仲裁价值缩水为"防误捕获"，批次 4 失去主要收益支撑 → 届时重新裁决
+   是否只做 host 侧准入/路由对齐。
+2. **方案B 的流量放大**（tile 缩 1/√g → 搬入周长和放大）理论标注不入模型（issue#37 P4），
+   周长型主导角区可能吞掉尾轮收益——批次 2 必须含 GM/L2 流量实测。
+3. **理论 ASW 降核口径**（P=L0C 满载粒度）与源码 usedCoreNum 口径不同，"理论落点分支预测"
+   在降核区需要对齐后再比对。
+4. **Blaze/旧库栈在仓外**：StreamK 的 TENSOR 级与 IterBatch 高级版主体实现不在当前源码包内，
+   批次 1/4 若需触及要单独申请对应源码包；优先走仓内可读的 cmct 内置栈。
+5. **理论模型自身的已知简化**（docs/05 §3 R6、issue#37 P3/P4）在上板数据回来后需要一轮
+   模型校准，再反哺后续批次的预期值。
+6. **确定性等级**：StreamK 相关改动保持 deterministic_level>1 禁用不变；批次 1 PR 需带
+   level=0/1 的数值稳定性证据。
+
+---
+
+*版本：v1.0 | 2026-09-07 | 基于 BMM_Theory main（issue#40 Cube-only 口径）与
+ops-nn batch_mat_mul_v3 arch35 源码快照比对；量化数字由 `bmm_theory` 测算脚本产出
+（ASCEND950PR 规格，T_cmd=0 未标定假设），仅供排期参考。*

BMM/BMM_Theory/examples/cases_demo.csv

@@ -1,11 +1,45 @@
-case_id,batch_a,batch_b,m,n,k,dtype_a,dtype_b,dtype_c,trans_a,trans_b,has_bias,deterministic_level
+case_id,batch_a,batch_b,m,n,k,dtype_a,dtype_b,dtype_c,trans_a,trans_b,has_bias,out_nd,deterministic_level
-to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,0,0,0,0
+to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,0,0,0,,0
-special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,0,0,0,0
+special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,0,0,0,,0
-special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,0,0,0,0
+special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,0,0,0,,0
-merge_demo_k_trunc,2048,2048,32,32,256,bf16,bf16,bf16,0,0,0,0
+merge_demo_k_trunc,2048,2048,16,64,128,bf16,bf16,bf16,0,0,0,,0
-merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,0,0,0,0
+merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,0,0,0,,0
-iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,0,0,0,0
+iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,0,0,0,,0
-iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,0,0,0,0
+iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,0,0,0,,0
-streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,0,0,0,0
+streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,0,0,0,,0
-asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,0,0,0,0
+asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,0,0,0,,0
-asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,0,0,0,0
+asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,0,0,0,,0
+b4_m1_n128_k256,4,4,1,128,256,bf16,bf16,bf16,0,0,0,1,0
+b8_m2_n192_k128,8,8,2,192,128,bf16,bf16,bf16,0,0,0,1,0
+b16_m4_n256_k192,16,16,4,256,192,bf16,bf16,bf16,0,0,0,1,0
+b32_m8_n128_k256,32,32,8,128,256,bf16,bf16,bf16,0,0,0,1,0
+b64_m16_n256_k512,64,64,16,256,512,bf16,bf16,bf16,0,0,0,1,0
+b128_m8_n192_k256,128,128,8,192,256,bf16,bf16,bf16,0,0,0,1,0
+b32_m16_n8192_k7168,32,32,16,8192,7168,bf16,bf16,bf16,0,0,0,1,0
+b4_m1_n8192_k8192,4,4,1,8192,8192,bf16,bf16,bf16,0,0,0,1,0
+b16_m2_n4096_k7168,16,16,2,4096,7168,bf16,bf16,bf16,0,0,0,1,0
+b32_m64_n64_k7168,32,32,64,64,7168,bf16,bf16,bf16,0,0,0,1,0
+b64_m1024_n1024_k7168,64,64,1024,1024,7168,bf16,bf16,bf16,0,0,0,1,0
+b128_m2048_n2048_k1536,128,128,2048,2048,1536,bf16,bf16,bf16,0,0,0,1,0
+b64_m1024_n8192_k2048,64,64,1024,8192,2048,bf16,bf16,bf16,0,0,0,1,0
+b32_m1024_n1024_k512,32,32,1024,1024,512,bf16,bf16,bf16,0,0,0,1,0
+b128_m4096_n4096_k8192,128,128,4096,4096,8192,bf16,bf16,bf16,0,0,0,1,0
+b32_m8192_n4096_k7168,32,32,8192,4096,7168,bf16,bf16,bf16,0,0,0,1,0
+b32_m2048_n2048_k8192,32,32,2048,2048,8192,bf16,bf16,bf16,0,0,0,1,0
+b64_m4096_n2048_k128,64,64,4096,2048,128,bf16,bf16,bf16,0,0,0,1,0
+b16_m1024_n1024_k8192,16,16,1024,1024,8192,bf16,bf16,bf16,0,0,0,1,0
+b4_m32768_n128_k128,4,4,32768,128,128,bf16,bf16,bf16,0,0,0,1,0
+b8_m32768_n2048_k512,8,8,32768,2048,512,bf16,bf16,bf16,0,0,0,1,0
+b4_m131072_n128_k128,4,4,131072,128,128,bf16,bf16,bf16,0,0,0,1,0
+b8_m131072_n1024_k256,8,8,131072,1024,256,bf16,bf16,bf16,0,0,0,1,0
+b16_m32768_n8192_k7168,16,16,32768,8192,7168,bf16,bf16,bf16,0,0,0,1,0
+b4_m32768_n128_k8192,4,4,32768,128,8192,bf16,bf16,bf16,0,0,0,1,0
+b32_m131072_n8192_k128,32,32,131072,8192,128,bf16,bf16,bf16,0,0,0,1,0
+b64_m32768_n8192_k1536,64,64,32768,8192,1536,bf16,bf16,bf16,0,0,0,1,0
+b8_m131072_n8192_k8192,8,8,131072,8192,8192,bf16,bf16,bf16,0,0,0,1,0
+b8_m16_n7168_k1536,8,8,16,7168,1536,bf16,bf16,bf16,0,0,0,1,0
+b64_m1024_n7168_k7168,64,64,1024,7168,7168,bf16,bf16,bf16,0,0,0,1,0
+b32_m8192_n8192_k7168,32,32,8192,8192,7168,bf16,bf16,bf16,0,0,0,1,0
+b8_m4096_n4096_k128,8,8,4096,4096,128,bf16,bf16,bf16,0,0,0,1,0
+b128_m8192_n8192_k7168,128,128,8192,8192,7168,bf16,bf16,bf16,0,0,0,1,0
+special_k1_b64,64,64,8192,512,1,bf16,bf16,bf16,0,0,0,1,0

BMM/BMM_Theory/examples/plans.csv

@@ -1,11 +1,45 @@
 case_id,branch,npu,op,used_core_num,split_b,m_cnt,n_cnt,grid_k,core_map,b_core,merge_b0,single_core_m,single_core_n,single_core_k,k_l1,b_l1,l1_form,base_m,base_n,base_k,l2_policy_in,l2_policy_out,swizzle_w,workspace_bytes,tail_strategy,tail_m_cnt,tail_n_cnt,tail_k_cnt,tail_m_main,tail_n_main,tail_block_cnt,tail_wave_num,fixpipe_unitflag,out_dtype_bytes,note
 to_matmul_demo,转Matmul,Ascend950PR,batch_mat_mul_v3,32,1,0,0,1,"折叠为 Matmul [2048,2048]x[2048,2048], 复用 Matmul 切分体系",0,1,0,0,2048,0,1,,0,0,0,,,0,0,转Matmul后由 Matmul 体系决定,1,1,1,0,0,0,0,True,2,"BatchB=1免费折叠: 左矩阵 [1,2048,2048] 视图折叠为 [2048,2048], 零重排零 split"
 special_k0_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,0,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=0纯写值: 无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核"
-special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB 乒乓"
+special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) UB乒乓,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB乒乓 (B>=2*AIV 双batch乒乓流水)"
-merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,128,128,256,256,8,合并驻留,128,128,128,allocate(GM->L1随路驻留L2),direct_gm,0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限19.0/b_core=64); K截断; 合并后单次DMA搬入 A'[128,256]+B'[256,128]"
+merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,64,256,128,128,12,合并驻留,64,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限21.1/b_core=64); K截断; 合并后单次DMA搬入 A'[64,128]+B'[128,256]; 输出落点: L2驻留 (整case V_in+V_out=40.0MB vs L2=128MB)"
-merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1
+merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1; 输出落点: L2驻留
-iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1
+iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1; 输出落点: L2驻留
-iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝"
+iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: 直写GM"
 streamk_demo,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,128,128,320,256,1,K段标准分块流水,128,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=1.00, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终"
-asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,47,47,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,176,176,1024,256,1,双缓冲驻留当前tile输入,176,176,80,allocate(输入驻留L2吸收重复读),direct_gm(输出直写GM不占L2),4,0,方案B,52,52,1,52,52,2,139,True,2,"L2场景B_输入驻留输出直写GM, r_in=1.00; 尾轮: 周长型主导, rho=0.06<rho_dv=0.53, A1b被dValue卡死, 方案B反超"
+asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1024,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=16, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入512.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=64.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮"
-asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,direct_gm,0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)"
+asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)"
+b4_m1_n128_k256,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,1,128,256,256,1,标准核内流水,1,128,128,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.01<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)"
+b8_m2_n192_k128,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,2,192,128,128,1,标准核内流水,2,192,80,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.05<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)"
+b16_m4_n256_k192,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,4,256,192,192,1,标准核内流水,4,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.25<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)"
+b32_m8_n128_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,8,128,256,256,1,a_单batch全驻留,8,128,128,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,单batch全驻留: (MK+KN)*dtype=68KB <= L1; 输出落点: L2驻留
+b64_m16_n256_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,16,256,512,240,1,c_一侧驻留+对侧切K,16,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"一侧驻留(A)+对侧切K: A驻留16KB, 预算L1/2, k_L1=240, dValueA=480B/dValueB=512B; 输出落点: L2驻留"
+b128_m8_n192_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,8,192,256,256,2,b_双batch乒乓,8,192,80,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=200KB <= L1; 输出落点: L2驻留
+b32_m16_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,1,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,16,1024,7168,112,1,双缓冲驻留当前tile输入,16,1024,16,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: 效率降级(放开约束4, dValue 按 128B 硬下限, 搬移效率低于模型假设, 时延可能低估): P=1, mCnt=1 x nCnt=8 (tile 16x1024, 每batch搬入113.8MB, r=0); Base tile 16x1024 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=3591.0MB, V_out=8.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b4_m1_n8192_k8192,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,1,8192,256,256,1,K段标准分块流水,1,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,4194304,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=0.50, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终"
+b16_m2_n4096_k7168,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,16,"B/M/N切出16块, 每块16核切K归约 (归约组内核c负责K段[c*K/16,(c+1)*K/16))",1,1,2,4096,448,256,1,K段标准分块流水,2,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=16路切K+归约,1,1,16,0,0,0,0,True,4,"P=2.00, grid_K=16, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终"
+b32_m64_n64_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,64,64,7168,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: L2驻留"
+b64_m1024_n1024_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,1024,1024,7168,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: 直写GM"
+b128_m2048_n2048_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入48.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1536.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b64_m1024_n8192_k2048,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,2048,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=16 (tile 512x512, 每batch搬入128.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2304.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b32_m1024_n1024_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,1024,1024,512,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: L2驻留"
+b128_m4096_n4096_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16384.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b32_m8192_n4096_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=8 (tile 512x512, 每batch搬入1792.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=5376.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b32_m2048_n2048_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2048.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b64_m4096_n2048_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=4 (tile 512x512, 每batch搬入8.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=96.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b16_m1024_n1024_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=2 x nCnt=2 (tile 512x512, 每batch搬入64.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=512.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b4_m32768_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"resident(整case全驻留S_A: 输出驻留L2异步回写, GM写=0)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入10.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: A_整case全驻留(输入+输出<=L2) (V_in=32.1MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b8_m32768_n2048_k512,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,512,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=64 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=272.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b4_m131072_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,32,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=256 x nCnt=1 (tile 512x128, 每batch搬入40.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=128.1MB, V_out=128.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b8_m131072_n1024_k256,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,256,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=2 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=516.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b16_m32768_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,512,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入14336.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=8960.0MB, V_out=8192.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b4_m32768_n128_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,8192,192,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入640.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2056.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b32_m131072_n8192_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,4096,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1088.0MB, V_out=65536.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b64_m32768_n8192_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2048,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入3072.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7680.0MB, V_out=32768.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b8_m131072_n8192_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入65536.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: C_双侧超L2: 最小替换2D分组(组间落空GM, 窗口L2) (V_in=17408.0MB, V_out=16384.0MB, L2=128MB); 最小替换分组 m_grp=8x n_grp=8 (GM倍率3.76, 窗口L2/batch=57344.0MB); 尾轮: r=0 无尾轮"
+b8_m16_n7168_k1536,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,2,2,"B/M/N切出16块, 每块2核切K归约 (归约组内核c负责K段[c*K/2,(c+1)*K/2))",1,1,16,3584,768,256,1,K段标准分块流水,16,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,3670016,grid_K=2路切K+归约,1,1,2,0,0,0,0,True,4,"P=14.00, grid_K=2, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终"
+b64_m1024_n7168_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,14,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,56,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=14 (tile 512x512, 每batch搬入392.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=896.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b32_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b8_m4096_n4096_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入16.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+b128_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=28672.0MB, V_out=16384.0MB, L2=128MB); 尾轮: r=0 无尾轮"
+special_k1_b64,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) AIV单缓冲,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, AIV单缓冲 (B<2*AIV 逐batch单缓冲串行)"

BMM/BMM_Theory/examples/result_evaluate.csv

@@ -1,11 +1,45 @@
 case_id,batch_a,batch_b,m,n,k,dtype_a,dtype_b,dtype_c,trans_a,trans_b,has_bias,out_nd,deterministic_level,plan_case_id,plan_branch,plan_npu,plan_op,plan_used_core_num,plan_split_b,plan_m_cnt,plan_n_cnt,plan_grid_k,plan_core_map,plan_b_core,plan_merge_b0,plan_single_core_m,plan_single_core_n,plan_single_core_k,plan_k_l1,plan_b_l1,plan_l1_form,plan_base_m,plan_base_n,plan_base_k,plan_l2_policy_in,plan_l2_policy_out,plan_swizzle_w,plan_workspace_bytes,plan_tail_strategy,plan_tail_m_cnt,plan_tail_n_cnt,plan_tail_k_cnt,plan_tail_m_main,plan_tail_n_main,plan_tail_block_cnt,plan_tail_wave_num,plan_fixpipe_unitflag,plan_out_dtype_bytes,plan_note,gm_read_bytes,l2_read_bytes,t_mte2_gm,t_mte2_l2,t_mte2,dma_cmd_count,t_dma_cmd,cube_flops,t_mmad,fixpipe_bytes,t_fixpipe,t_reduce,t_steady,t_drain,t_total,bottleneck,feasible,violations,bound_type,advice
-to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,False,False,False,True,0,to_matmul_demo,转Matmul,Ascend950PR,batch_mat_mul_v3,32,1,0,0,1,"折叠为 Matmul [2048,2048]x[2048,2048], 复用 Matmul 切分体系",0,1,0,0,2048,0,1,,0,0,0,,,0,0,转Matmul后由 Matmul 体系决定,1,1,1,0,0,0,0,True,2,"BatchB=1免费折叠: 左矩阵 [1,2048,2048] 视图折叠为 [2048,2048], 零重排零 split",16777216,0.0,1.048576e-05,0.0,1.048576e-05,0.0,0.0,17179869184.0,3.534952506995885e-05,8388608,5.24288e-06,0.0,3.534952506995885e-05,0.0,3.534952506995885e-05,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,False,False,False,True,0,to_matmul_demo,转Matmul,Ascend950PR,batch_mat_mul_v3,32,1,0,0,1,"折叠为 Matmul [2048,2048]x[2048,2048], 复用 Matmul 切分体系",0,1,0,0,2048,0,1,,0,0,0,,,0,0,转Matmul后由 Matmul 体系决定,1,1,1,0,0,0,0,True,2,"BatchB=1免费折叠: 左矩阵 [1,2048,2048] 视图折叠为 [2048,2048], 零重排零 split",16777216,0.0,1.048576e-05,0.0,1.048576e-05,0.0,0.0,17179869184.0,3.97682157037037e-05,8388608,1.6131938461538462e-06,0.0,3.97682157037037e-05,0.0,3.97682157037037e-05,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
-special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,False,False,False,True,0,special_k0_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,0,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=0纯写值: 无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核",0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,16777216,1.048576e-05,0.0,1.048576e-05,0.0,1.048576e-05,FIXPIPE,True,,写出Bound,"瓶颈在 Fixpipe 写出: 检查输出 dtype (fp16/fp8 可减半写出量), 或评估输出驻留 L2 异步回写策略"
+special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,False,False,False,True,0,special_k0_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,0,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=0纯写值: 无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核",0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,16777216,3.2263876923076923e-06,0.0,3.2263876923076923e-06,0.0,3.2263876923076923e-06,FIXPIPE,True,,写出Bound(L2写口),"瓶颈在 Fixpipe 写出: 检查输出 dtype (fp16/fp8 可减半写出量), 或评估输出驻留 L2 异步回写策略"
-special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) UB乒乓,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB乒乓 (B>=2*AIV 双batch乒乓流水)",131072,0.0,8.192e-08,0.0,8.192e-08,0.0,0.0,8388608.0,6.206060606060606e-07,16777216,1.048576e-05,0.0,1.048576e-05,0.0,1.048576e-05,FIXPIPE,True,,写出Bound,"瓶颈在 Fixpipe 写出: 检查输出 dtype (fp16/fp8 可减半写出量), 或评估输出驻留 L2 异步回写策略"
+special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) UB乒乓,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB乒乓 (B>=2*AIV 双batch乒乓流水)",131072,0.0,8.192e-08,0.0,8.192e-08,0.0,0.0,8388608.0,3.103030303030303e-07,16777216,3.2263876923076923e-06,0.0,3.2263876923076923e-06,0.0,3.2263876923076923e-06,FIXPIPE,True,,写出Bound(L2写口),"瓶颈在 Fixpipe 写出: 检查输出 dtype (fp16/fp8 可减半写出量), 或评估输出驻留 L2 异步回写策略"
-merge_demo_k_trunc,2048,2048,32,32,256,bf16,bf16,bf16,False,False,False,True,0,merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,128,128,256,256,8,合并驻留,128,128,128,allocate(GM->L1随路驻留L2),direct_gm,0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限19.0/b_core=64); K截断; 合并后单次DMA搬入 A'[128,256]+B'[256,128]",2097152,0.0,4.194304e-05,0.0,4.2743039999999997e-05,16.0,8.000000000000001e-07,134217728.0,8.837381267489712e-06,131072,2.62144e-06,0.0,4.2743039999999997e-05,3.0192408230452674e-07,4.3044964082304525e-05,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+merge_demo_k_trunc,2048,2048,16,64,128,bf16,bf16,bf16,False,False,False,True,0,merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,64,256,128,128,12,合并驻留,64,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限21.1/b_core=64); K截断; 合并后单次DMA搬入 A'[64,128]+B'[128,256]; 输出落点: L2驻留 (整case V_in+V_out=40.0MB vs L2=128MB)",41943040,0.0,2.62144e-05,0.0,2.62144e-05,16,0.0,2147483648.0,4.971026962962963e-06,4194304,8.065969230769231e-07,0.0,2.62144e-05,1.2808460398860398e-07,2.6342484603988603e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
-merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,False,False,False,True,0,merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1,524288,0.0,1.048576e-05,0.0,1.0685759999999999e-05,4,2.0000000000000002e-07,16777216.0,1.104672658436214e-06,32768,6.5536e-07,0.0,1.0685759999999999e-05,4.400081646090535e-07,1.1125768164609053e-05,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,False,False,False,True,0,merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1; 输出落点: L2驻留,16777216,0.0,1.048576e-05,0.0,1.048576e-05,4,0.0,536870912.0,1.2427567407407407e-06,1048576,2.0164923076923077e-07,0.0,1.048576e-05,3.6110149287749287e-07,1.0846861492877492e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
-iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1,262144,0.0,5.24288e-06,0.0,5.4428799999999995e-06,4,2.0000000000000002e-07,8388608.0,5.52336329218107e-07,32768,6.5536e-07,0.0,5.4428799999999995e-06,3.0192408230452674e-07,5.744804082304526e-06,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1; 输出落点: L2驻留,8388608,0.0,5.24288e-06,0.0,5.24288e-06,4,0.0,268435456.0,6.213783703703703e-07,1048576,2.0164923076923077e-07,0.0,5.24288e-06,2.0575690028490026e-07,5.4486369002849e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
-iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝",4194304,0.0,8.388608e-05,0.0,8.468608e-05,16,8.000000000000001e-07,134217728.0,8.837381267489712e-06,16384,3.2768e-07,0.0,8.468608e-05,7.16176329218107e-07,8.540225632921811e-05,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: 直写GM",134217728,0.0,8.388608e-05,0.0,8.388608e-05,16,0.0,4294967296.0,9.942053925925925e-06,524288,3.2768e-07,0.0,8.421376e-05,7.852183703703703e-07,8.499897837037037e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
-streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,False,False,False,True,0,streamk_demo,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,128,128,320,256,1,K段标准分块流水,128,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=1.00, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",327680.0,0.0,6.5536e-06,0.0,6.5536e-06,0.0,0.0,41943040.0,2.761681646090535e-06,0.0,0.0,3.4067453613053613e-06,6.5536e-06,3.4067453613053613e-06,9.960345361305361e-06,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,False,False,False,True,0,streamk_demo,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,128,128,320,256,1,K段标准分块流水,128,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=1.00, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",20971520,0,1.31072e-05,0.0,1.31072e-05,0.0,0.0,1342177280.0,3.1068918518518518e-06,0.0,0.0,3.4067453613053613e-06,1.31072e-05,3.4067453613053613e-06,1.651394536130536e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
-asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,False,False,False,True,0,asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,47,47,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,176,176,1024,256,1,双缓冲驻留当前tile输入,176,176,80,allocate(输入驻留L2吸收重复读),direct_gm(输出直写GM不占L2),4,0,方案B,52,52,1,52,52,2,139,True,2,"L2场景B_输入驻留输出直写GM, r_in=1.00; 尾轮: 周长型主导, rho=0.06<rho_dv=0.53, A1b被dValue卡死, 方案B反超",67108864.0,0.0,4.194304e-05,0.0,4.194304e-05,0.0,0.0,274877906944.0,0.0005655924011193416,268435456,0.00016777216,0.0,0.0005655924011193416,0.0,0.0005655924011193416,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,False,False,False,True,0,asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1024,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=16, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入512.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=64.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",67108864,1006632960,4.194304e-05,0.00019358326153846154,0.00023552630153846153,0.0,0.0,274877906944.0,0.0006362914512592592,268435456,0.00016777216,0.0,0.0006362914512592592,0.0,0.0006362914512592592,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
-asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,False,False,False,True,0,asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,direct_gm,0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",2097152.0,0.0,2.62144e-06,0.0,2.62144e-06,0.0,0.0,268435456.0,1.104672658436214e-06,2097152,2.62144e-06,0.0,2.62144e-06,0.0,2.62144e-06,MTE2_GM,True,,访存Bound(GM),"瓶颈在 GM 搬入: 可考虑增大 tile 提升 dValue/单核搬移量, 或利用 L2 驻留吸收重复读 (MergeBatch/ASW swizzle 方向)"
+asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,False,False,False,True,0,asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",2097152,0,2.62144e-06,0.0,2.62144e-06,0.0,0.0,268435456.0,1.2427567407407407e-06,2097152,8.065969230769231e-07,0.0,2.62144e-06,0.0,2.62144e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b4_m1_n128_k256,4,4,1,128,256,bf16,bf16,bf16,False,False,False,True,0,b4_m1_n128_k256,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,1,128,256,256,1,标准核内流水,1,128,128,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.01<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",264192,0,5.28384e-06,0.0,5.28384e-06,0.0,0.0,262144.0,1.9418074074074073e-08,1024,6.3015384615384615e-09,0.0,5.28384e-06,0.0,5.28384e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b8_m2_n192_k128,8,8,2,192,128,bf16,bf16,bf16,False,False,False,True,0,b8_m2_n192_k128,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,2,192,128,128,1,标准核内流水,2,192,80,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.05<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",397312,0,7.94624e-06,0.0,7.94624e-06,0.0,0.0,786432.0,5.825422222222222e-08,6144,3.7809230769230766e-08,0.0,7.94624e-06,0.0,7.94624e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b16_m4_n256_k192,16,16,4,256,192,bf16,bf16,bf16,False,False,False,True,0,b16_m4_n256_k192,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,4,256,192,192,1,标准核内流水,4,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.25<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",1597440,0,3.19488e-05,0.0,3.19488e-05,0.0,0.0,6291456.0,4.660337777777778e-07,32768,2.0164923076923077e-07,0.0,3.19488e-05,0.0,3.19488e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b32_m8_n128_k256,32,32,8,128,256,bf16,bf16,bf16,False,False,False,True,0,b32_m8_n128_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,8,128,256,256,1,a_单batch全驻留,8,128,128,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,单batch全驻留: (MK+KN)*dtype=68KB <= L1; 输出落点: L2驻留,2228224,0.0,1.6384e-06,0.0,1.6384e-06,1,0.0,16777216.0,3.8836148148148146e-08,65536,1.2603076923076923e-08,0.0,1.6384e-06,5.1439225071225065e-08,1.689839225071225e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b64_m16_n256_k512,64,64,16,256,512,bf16,bf16,bf16,False,False,False,True,0,b64_m16_n256_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,16,256,512,240,1,c_一侧驻留+对侧切K,16,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"一侧驻留(A)+对侧切K: A驻留16KB, 预算L1/2, k_L1=240, dValueA=480B/dValueB=512B; 输出落点: L2驻留",17825792,0.0,1.114112e-05,0.0,1.114112e-05,6,0.0,268435456.0,6.213783703703703e-07,524288,1.0082461538461538e-07,0.0,1.114112e-05,1.9604786324786327e-07,1.1337167863247863e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b128_m8_n192_k256,128,128,8,192,256,bf16,bf16,bf16,False,False,False,True,0,b128_m8_n192_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,8,192,256,256,2,b_双batch乒乓,8,192,80,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=200KB <= L1; 输出落点: L2驻留,13107200,0.0,9.17504e-06,0.0,9.17504e-06,4,0.0,100663296.0,2.330168888888889e-07,393216,7.561846153846153e-08,0.0,9.17504e-06,7.71588376068376e-08,9.252198837606838e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b32_m16_n8192_k7168,32,32,16,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m16_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,1,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,16,1024,7168,112,1,双缓冲驻留当前tile输入,16,1024,16,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: 效率降级(放开约束4, dValue 按 128B 硬下限, 搬移效率低于模型假设, 时延可能低估): P=1, mCnt=1 x nCnt=8 (tile 16x1024, 每batch搬入113.8MB, r=0); Base tile 16x1024 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=3591.0MB, V_out=8.0MB, L2=128MB); 尾轮: r=0 无尾轮",3765436416,51380224,0.00235339776,9.880812307692307e-06,0.0023632785723076925,0.0,0.0,60129542144.0,0.00013918875496296296,8388608,5.24288e-06,0.0,0.0023685214523076923,0.0,0.0023685214523076923,MTE2,True,,访存Bound(GM读写共享+L2重复读),"效率降级标注 (plan.note): 搬移效率下限不满足 —— 方案照常给出 (兜底), 但实际效率低于模型假设, 时延可能低估; 建议调整 dtype/布局 | 瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b4_m1_n8192_k8192,4,4,1,8192,8192,bf16,bf16,bf16,False,False,False,True,0,b4_m1_n8192_k8192,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,1,8192,256,256,1,K段标准分块流水,1,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,4194304,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=0.50, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",536936448,0.0,0.00033558528,0.0,0.00033558528,0.0,0.0,536870912.0,1.2427567407407407e-06,0.0,0.0,1.731729603729604e-06,0.00033558528,1.731729603729604e-06,0.0003373170096037296,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b16_m2_n4096_k7168,16,16,2,4096,7168,bf16,bf16,bf16,False,False,False,True,0,b16_m2_n4096_k7168,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,16,"B/M/N切出16块, 每块16核切K归约 (归约组内核c负责K段[c*K/16,(c+1)*K/16))",1,1,2,4096,448,256,1,K段标准分块流水,2,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=16路切K+归约,1,1,16,0,0,0,0,True,4,"P=2.00, grid_K=16, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",939982848,0,0.00058748928,0.0,0.00058748928,0.0,0.0,1879048192.0,4.349648592592593e-06,0.0,0.0,1.7726896037296038e-06,0.00058748928,1.7726896037296038e-06,0.0005892619696037297,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b32_m64_n64_k7168,32,32,64,64,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m64_n64_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,64,64,7168,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: L2驻留",58720256,0.0,3.670016e-05,0.0,3.670016e-05,7,0.0,1879048192.0,4.349648592592593e-06,262144,5.041230769230769e-08,0.0,3.670016e-05,6.71790678062678e-07,3.737195067806268e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b64_m1024_n1024_k7168,64,64,1024,1024,7168,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n1024_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,1024,1024,7168,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: 直写GM",1879048192,0.0,0.00117440512,0.0,0.00117440512,224,0.0,962072674304.0,0.0022270200794074074,134217728,8.388608e-05,0.0,0.0022270200794074074,5.1885093925925924e-05,0.0022789051733333333,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b128_m2048_n2048_k1536,128,128,2048,2048,1536,bf16,bf16,bf16,False,False,False,True,0,b128_m2048_n2048_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入48.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1536.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",1610612736,4831838208,0.00100663296,0.0009291996553846154,0.0019358326153846154,0.0,0.0,1649267441664.0,0.0038177487075555555,1073741824,0.00067108864,0.0,0.0038177487075555555,0.0,0.0038177487075555555,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b64_m1024_n8192_k2048,64,64,1024,8192,2048,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n8192_k2048,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,2048,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=16 (tile 512x512, 每batch搬入128.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2304.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",2415919104,6174015488,0.00150994944,0.0011873106707692308,0.0026972601107692305,0.0,0.0,2199023255552.0,0.005090331610074074,1073741824,0.00067108864,0.0,0.005090331610074074,0.0,0.005090331610074074,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b32_m1024_n1024_k512,32,32,1024,1024,512,bf16,bf16,bf16,False,False,False,True,0,b32_m1024_n1024_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,1024,1024,512,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: L2驻留",67108864,0.0,4.194304e-05,0.0,4.194304e-05,8,0.0,34359738368.0,7.95364314074074e-05,67108864,1.290555076923077e-05,0.0,7.95364314074074e-05,2.2847604695156693e-05,0.0001023840361025641,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b128_m4096_n4096_k8192,128,128,4096,4096,8192,bf16,bf16,bf16,False,False,False,True,0,b128_m4096_n4096_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16384.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮",17179869184,120259084288,0.01073741824,0.023126746978461538,0.033864165218461535,0.0,0.0,35184372088832.0,0.08144530576118518,4294967296,0.00268435456,0.0,0.08144530576118518,0.0,0.08144530576118518,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b32_m8192_n4096_k7168,32,32,8192,4096,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m8192_n4096_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=8 (tile 512x512, 每batch搬入1792.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=5376.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮",5637144576,54492397568,0.00352321536,0.010479307224615384,0.014002522584615384,0.0,0.0,15393162788864.0,0.03563232127051852,2147483648,0.00134217728,0.0,0.03563232127051852,0.0,0.03563232127051852,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b32_m2048_n2048_k8192,32,32,2048,2048,8192,bf16,bf16,bf16,False,False,False,True,0,b32_m2048_n2048_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2048.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",2147483648,6442450944,0.00134217728,0.0012389328738461537,0.002581110153846154,0.0,0.0,2199023255552.0,0.005090331610074074,268435456,0.00016777216,0.0,0.005090331610074074,0.0,0.005090331610074074,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b64_m4096_n2048_k128,64,64,4096,2048,128,bf16,bf16,bf16,False,False,False,True,0,b64_m4096_n2048_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=4 (tile 512x512, 每batch搬入8.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=96.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",100663296,436207616,6.291456e-05,8.388608e-05,0.00014680063999999998,0.0,0.0,137438953472.0,0.0003181457256296296,1073741824,0.00067108864,0.0,0.00081788928,0.0,0.00081788928,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b16_m1024_n1024_k8192,16,16,1024,1024,8192,bf16,bf16,bf16,False,False,False,True,0,b16_m1024_n1024_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=2 x nCnt=2 (tile 512x512, 每batch搬入64.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=512.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",536870912,536870912,0.00033554432,0.00010324440615384615,0.0004387887261538461,0.0,0.0,274877906944.0,0.0006362914512592592,33554432,2.097152e-05,0.0,0.0006362914512592592,0.0,0.0006362914512592592,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b4_m32768_n128_k128,4,4,32768,128,128,bf16,bf16,bf16,False,False,False,True,0,b4_m32768_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"resident(整case全驻留S_A: 输出驻留L2异步回写, GM写=0)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入10.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: A_整case全驻留(输入+输出<=L2) (V_in=32.1MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",33685504,8257536,2.105344e-05,1.5879876923076922e-06,2.2641427692307692e-05,0.0,0.0,4294967296.0,9.942053925925925e-06,33554432,6.452775384615385e-06,0.0,2.2641427692307692e-05,0.0,2.2641427692307692e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b8_m32768_n2048_k512,8,8,32768,2048,512,bf16,bf16,bf16,False,False,False,True,0,b8_m32768_n2048_k512,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,512,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=64 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=272.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",285212672,1862270976,0.00017825792,0.00035812903384615385,0.0005363869538461539,0.0,0.0,549755813888.0,0.0012725829025185184,1073741824,0.00067108864,0.0,0.0012725829025185184,0.0,0.0012725829025185184,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b4_m131072_n128_k128,4,4,131072,128,128,bf16,bf16,bf16,False,False,False,True,0,b4_m131072_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,32,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=256 x nCnt=1 (tile 512x128, 每batch搬入40.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=128.1MB, V_out=128.0MB, L2=128MB); 尾轮: r=0 无尾轮",134348800,33423360,8.3968e-05,6.427569230769231e-06,9.039556923076924e-05,0.0,0.0,17179869184.0,3.97682157037037e-05,134217728,8.388608e-05,0.0,0.00017428164923076922,0.0,0.00017428164923076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b8_m131072_n1024_k256,8,8,131072,1024,256,bf16,bf16,bf16,False,False,False,True,0,b8_m131072_n1024_k256,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,256,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=2 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=516.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮",541065216,1606418432,0.00033816576,0.00030892662153846154,0.0006470923815384616,0.0,0.0,549755813888.0,0.0012725829025185184,2147483648,0.00134217728,0.0,0.0019892696615384617,0.0,0.0019892696615384617,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b16_m32768_n8192_k7168,16,16,32768,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b16_m32768_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,512,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入14336.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=8960.0MB, V_out=8192.0MB, L2=128MB); 尾轮: r=0 无尾轮",9395240960,231122927616,0.0058720256,0.044446716849230766,0.05031874244923076,0.0,0.0,61572651155456.0,0.14252928508207408,8589934592,0.00536870912,0.0,0.14252928508207408,0.0,0.14252928508207408,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b4_m32768_n128_k8192,4,4,32768,128,8192,bf16,bf16,bf16,False,False,False,True,0,b4_m32768_n128_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,8192,192,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入640.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2056.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",2155872256,528482304,0.00134742016,0.0001016312123076923,0.0014490513723076923,0.0,0.0,274877906944.0,0.0006362914512592592,33554432,2.097152e-05,0.0,0.0014700228923076922,0.0,0.0014700228923076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b32_m131072_n8192_k128,32,32,131072,8192,128,bf16,bf16,bf16,False,False,False,True,0,b32_m131072_n8192_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,4096,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1088.0MB, V_out=65536.0MB, L2=128MB); 尾轮: r=0 无尾轮",1140850688,33218887680,0.00071303168,0.006388247630769231,0.007101279310769231,0.0,0.0,8796093022208.0,0.020361326440296295,68719476736,0.04294967296,0.0,0.05005095227076922,0.0,0.05005095227076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b64_m32768_n8192_k1536,64,64,32768,8192,1536,bf16,bf16,bf16,False,False,False,True,0,b64_m32768_n8192_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2048,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入3072.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7680.0MB, V_out=32768.0MB, L2=128MB); 尾轮: r=0 无尾轮",8053063680,198105366528,0.0050331648,0.03809718587076923,0.04313035067076923,0.0,0.0,52776558133248.0,0.12216795864177778,34359738368,0.02147483648,0.0,0.12216795864177778,0.0,0.12216795864177778,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b8_m131072_n8192_k8192,8,8,131072,8192,8192,bf16,bf16,bf16,False,False,False,True,0,b8_m131072_n8192_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入65536.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: C_双侧超L2: 最小替换2D分组(组间落空GM, 窗口L2) (V_in=17408.0MB, V_out=16384.0MB, L2=128MB); 最小替换分组 m_grp=8x n_grp=8 (GM倍率3.76, 窗口L2/batch=57344.0MB); 尾轮: r=0 无尾轮",68719476736,481036337152,0.04294967296,0.09250698791384615,0.13545666087384614,0.0,0.0,140737488355328.0,0.3257812230447407,17179869184,0.01073741824,0.0,0.3257812230447407,0.0,0.3257812230447407,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b8_m16_n7168_k1536,8,8,16,7168,1536,bf16,bf16,bf16,False,False,False,True,0,b8_m16_n7168_k1536,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,2,2,"B/M/N切出16块, 每块2核切K归约 (归约组内核c负责K段[c*K/2,(c+1)*K/2))",1,1,16,3584,768,256,1,K段标准分块流水,16,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,3670016,grid_K=2路切K+归约,1,1,2,0,0,0,0,True,4,"P=14.00, grid_K=2, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",176553984,393216,0.00011034624,7.561846153846153e-08,0.00011042185846153846,0.0,0.0,2818572288.0,6.524472888888889e-06,0.0,0.0,2.566079254079254e-07,0.00011042185846153846,2.566079254079254e-07,0.00011067846638694638,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b64_m1024_n7168_k7168,64,64,1024,7168,7168,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n7168_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,14,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,56,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=14 (tile 512x512, 每batch搬入392.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=896.0MB, L2=128MB); 尾轮: r=0 无尾轮",7516192768,18790481920,0.00469762048,0.0036135542153846152,0.008311174695384616,0.0,0.0,6734508720128.0,0.015589140555851852,939524096,0.00058720256,0.0,0.015589140555851852,0.0,0.015589140555851852,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b32_m8192_n8192_k7168,32,32,8192,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮",7516192768,112742891520,0.00469762048,0.021681325292307693,0.026378945772307694,0.0,0.0,30786325577728.0,0.07126464254103704,4294967296,0.00268435456,0.0,0.07126464254103704,0.0,0.07126464254103704,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+b8_m4096_n4096_k128,8,8,4096,4096,128,bf16,bf16,bf16,False,False,False,True,0,b8_m4096_n4096_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入16.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",16777216,117440512,1.048576e-05,2.2584713846153845e-05,3.307047384615384e-05,0.0,0.0,34359738368.0,7.95364314074074e-05,268435456,0.00016777216,0.0,0.00020084263384615383,0.0,0.00020084263384615383,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"
+b128_m8192_n8192_k7168,128,128,8192,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b128_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=28672.0MB, V_out=16384.0MB, L2=128MB); 尾轮: r=0 无尾轮",30064771072,450971566080,0.01879048192,0.08672530116923077,0.10551578308923078,0.0,0.0,123145302310912.0,0.28505857016414815,17179869184,0.01073741824,0.0,0.28505857016414815,0.0,0.28505857016414815,MMAD,True,,计算Bound,"瓶颈在 Cube 计算: 已接近理论算力上限, 检查是否有冗余计算 (MergeBatch 交叉项) 可消除"
+special_k1_b64,64,64,8192,512,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_b64,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) AIV单缓冲,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, AIV单缓冲 (B<2*AIV 逐batch单缓冲串行)",1114112,0.0,6.9632e-07,0.0,6.9632e-07,0.0,0.0,268435456.0,9.92969696969697e-06,536870912,0.00033554432,0.0,0.00033624063999999996,0.0,0.00033624063999999996,MTE2,True,,访存Bound(GM读写共享+L2重复读),"瓶颈在 MTE2 搬移链 (GM 读写共享总线 + L2 重复读): 可增大 tile 提升 dValue/单核搬移量、利用 L2 驻留吸收重复读 (ASW swizzle/分组方向), 或评估输出驻留 L2 以减少 GM 直写与读竞争"

BMM/BMM_Theory/examples/result_recommend.csv

@@ -1,11 +1,45 @@
 case_id,batch_a,batch_b,m,n,k,dtype_a,dtype_b,dtype_c,trans_a,trans_b,has_bias,out_nd,deterministic_level,plan_case_id,plan_branch,plan_npu,plan_op,plan_used_core_num,plan_split_b,plan_m_cnt,plan_n_cnt,plan_grid_k,plan_core_map,plan_b_core,plan_merge_b0,plan_single_core_m,plan_single_core_n,plan_single_core_k,plan_k_l1,plan_b_l1,plan_l1_form,plan_base_m,plan_base_n,plan_base_k,plan_l2_policy_in,plan_l2_policy_out,plan_swizzle_w,plan_workspace_bytes,plan_tail_strategy,plan_tail_m_cnt,plan_tail_n_cnt,plan_tail_k_cnt,plan_tail_m_main,plan_tail_n_main,plan_tail_block_cnt,plan_tail_wave_num,plan_fixpipe_unitflag,plan_out_dtype_bytes,plan_note,gm_read_bytes,l2_read_bytes,t_mte2_gm,t_mte2_l2,t_mte2,dma_cmd_count,t_dma_cmd,cube_flops,t_mmad,fixpipe_bytes,t_fixpipe,t_reduce,t_steady,t_drain,t_total,bottleneck,feasible,violations,bound_type,advice
-to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,False,False,False,True,0,to_matmul_demo,转Matmul,Ascend950PR,batch_mat_mul_v3,32,1,0,0,1,"折叠为 Matmul [2048,2048]x[2048,2048], 复用 Matmul 切分体系",0,1,0,0,2048,0,1,,0,0,0,,,0,0,转Matmul后由 Matmul 体系决定,1,1,1,0,0,0,0,True,2,"BatchB=1免费折叠: 左矩阵 [1,2048,2048] 视图折叠为 [2048,2048], 零重排零 split",16777216,0.0,1.048576e-05,0.0,1.048576e-05,0.0,0.0,17179869184.0,3.534952506995885e-05,8388608,5.24288e-06,0.0,3.534952506995885e-05,0.0,3.534952506995885e-05,MMAD,True,,计算Bound,"BatchA=1或BatchB=1, 折叠转普通Matmul"
+to_matmul_demo,1,1,2048,2048,2048,bf16,bf16,bf16,False,False,False,True,0,to_matmul_demo,转Matmul,Ascend950PR,batch_mat_mul_v3,32,1,0,0,1,"折叠为 Matmul [2048,2048]x[2048,2048], 复用 Matmul 切分体系",0,1,0,0,2048,0,1,,0,0,0,,,0,0,转Matmul后由 Matmul 体系决定,1,1,1,0,0,0,0,True,2,"BatchB=1免费折叠: 左矩阵 [1,2048,2048] 视图折叠为 [2048,2048], 零重排零 split",16777216,0.0,1.048576e-05,0.0,1.048576e-05,0.0,0.0,17179869184.0,3.97682157037037e-05,8388608,1.6131938461538462e-06,0.0,3.97682157037037e-05,0.0,3.97682157037037e-05,MMAD,True,,计算Bound,"BatchA=1或BatchB=1, 折叠转普通Matmul"
-special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,False,False,False,True,0,special_k0_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,0,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=0纯写值: 无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核",0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,16777216,1.048576e-05,0.0,1.048576e-05,0.0,1.048576e-05,FIXPIPE,True,,写出Bound,K=0纯写值
+special_k0_demo,128,128,256,256,0,bf16,bf16,bf16,False,False,False,True,0,special_k0_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,0,0,1,UB驻留(AIV),0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=0纯写值: 无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核",0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,16777216,3.2263876923076923e-06,0.0,3.2263876923076923e-06,0.0,3.2263876923076923e-06,FIXPIPE,True,,写出Bound(L2写口),K=0纯写值
-special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) UB乒乓,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB乒乓 (B>=2*AIV 双batch乒乓流水)",131072,0.0,8.192e-08,0.0,8.192e-08,0.0,0.0,8388608.0,6.206060606060606e-07,16777216,1.048576e-05,0.0,1.048576e-05,0.0,1.048576e-05,FIXPIPE,True,,写出Bound,"K=1逐元素乘, 走AIV向量通路"
+special_k1_demo,128,128,256,256,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_demo,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) UB乒乓,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, UB乒乓 (B>=2*AIV 双batch乒乓流水)",131072,0.0,8.192e-08,0.0,8.192e-08,0.0,0.0,8388608.0,3.103030303030303e-07,16777216,3.2263876923076923e-06,0.0,3.2263876923076923e-06,0.0,3.2263876923076923e-06,FIXPIPE,True,,写出Bound(L2写口),"K=1逐元素乘, 走AIV向量通路"
-merge_demo_k_trunc,2048,2048,32,32,256,bf16,bf16,bf16,False,False,False,True,0,merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,128,128,256,256,8,合并驻留,128,128,128,allocate(GM->L1随路驻留L2),direct_gm,0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限19.0/b_core=64); K截断; 合并后单次DMA搬入 A'[128,256]+B'[256,128]",2097152,0.0,4.194304e-05,0.0,4.2743039999999997e-05,16.0,8.000000000000001e-07,134217728.0,8.837381267489712e-06,131072,2.62144e-06,0.0,4.2743039999999997e-05,3.0192408230452674e-07,4.3044964082304525e-05,MTE2_GM,True,,访存Bound(GM),"两分支均合法, 仲裁: [分界条件] MergeBatch最优=True (k_L1=K(截断); b_core=64 vs 阈值 b0*(T_comp+T_write)/T_cmd=6.0; drain惩罚=(b0-1)*(T_comp+T_write)=0.23us, 搬移节省=b_core*(1-1/b0)*T_cmd=2.40us); [时延模型] T_MergeBatch=43.04us vs T_IterBatch=45.22us -> MergeBatch更优; [裁决] MergeBatch"
+merge_demo_k_trunc,2048,2048,16,64,128,bf16,bf16,bf16,False,False,False,True,0,merge_demo_k_trunc,MergeBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B均分(核间零重复读零依赖),64,4,64,256,128,128,12,合并驻留,64,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"b0=4 (L0C上限5.7/算存比上限21.1/b_core=64); K截断; 合并后单次DMA搬入 A'[64,128]+B'[128,256]; 输出落点: L2驻留 (整case V_in+V_out=40.0MB vs L2=128MB)",41943040,0.0,2.62144e-05,0.0,2.62144e-05,16,0.0,2147483648.0,4.971026962962963e-06,4194304,8.065969230769231e-07,0.0,2.62144e-05,1.2808460398860398e-07,2.6342484603988603e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"两分支均合法, 仲裁: [分界条件] MergeBatch最优=True (K截断(k_l1^m=128>=K); 每核命令数 MergeBatch=16 vs IterBatch=64 (命令节省=0.00us) + 搬移效率节省=15.73us vs drain惩罚=(b0-1)*(T_comp+T_write)=0.18us -> MergeBatch优); [时延模型] T_MergeBatch=26.34us vs T_IterBatch=41.98us -> MergeBatch更优; [裁决] MergeBatch"
-merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,False,False,False,True,0,merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1,524288,0.0,1.048576e-05,0.0,1.0685759999999999e-05,4,2.0000000000000002e-07,16777216.0,1.104672658436214e-06,32768,6.5536e-07,0.0,1.0685759999999999e-05,4.400081646090535e-07,1.1125768164609053e-05,MTE2_GM,True,,访存Bound(GM),"两分支均合法, 仲裁: [分界条件] MergeBatch最优=False (k_L1=K(截断); b_core=4 vs 阈值 b0*(T_comp+T_write)/T_cmd=17.6; drain惩罚=(b0-1)*(T_comp+T_write)=0.44us, 搬移节省=b_core*(1-1/b0)*T_cmd=0.10us); [时延模型] T_MergeBatch=11.49us vs T_IterBatch=11.13us -> IterBatch更优; [裁决] IterBatch"
+merge_iter_arbitrate,128,128,64,64,512,bf16,bf16,bf16,False,False,False,True,0,merge_iter_arbitrate,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,512,512,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=256KB <= L1; 输出落点: L2驻留,16777216,0.0,1.048576e-05,0.0,1.048576e-05,4,0.0,536870912.0,1.2427567407407407e-06,1048576,2.0164923076923077e-07,0.0,1.048576e-05,3.6110149287749287e-07,1.0846861492877492e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"两分支均合法, 仲裁: [分界条件] MergeBatch最优=False (L1绑定(k_l1^m=256<K=512); 每核命令数 MergeBatch=4 vs IterBatch=4 (命令节省=0.00us) + 搬移效率节省=0.00us vs drain惩罚=(b0-1)*(T_comp+T_write)=0.47us -> IterBatch优); [时延模型] T_MergeBatch=10.90us vs T_IterBatch=10.85us -> IterBatch更优; [裁决] IterBatch"
-iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1,262144,0.0,5.24288e-06,0.0,5.4428799999999995e-06,4,2.0000000000000002e-07,8388608.0,5.52336329218107e-07,32768,6.5536e-07,0.0,5.4428799999999995e-06,3.0192408230452674e-07,5.744804082304526e-06,MTE2_GM,True,,访存Bound(GM),仅 IterBatch 条件满足
+iter_demo_form_b,128,128,64,64,256,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_b,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,64,64,256,256,2,b_双batch乒乓,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=128KB <= L1; 输出落点: L2驻留,8388608,0.0,5.24288e-06,0.0,5.24288e-06,4,0.0,268435456.0,6.213783703703703e-07,1048576,2.0164923076923077e-07,0.0,5.24288e-06,2.0575690028490026e-07,5.4486369002849e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
-iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(输出仅写一次,直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝",4194304,0.0,8.388608e-05,0.0,8.468608e-05,16,8.000000000000001e-07,134217728.0,8.837381267489712e-06,16384,3.2768e-07,0.0,8.468608e-05,7.16176329218107e-07,8.540225632921811e-05,MTE2_GM,True,,访存Bound(GM),仅 IterBatch 条件满足
+iter_demo_form_d,64,64,64,64,8192,bf16,bf16,bf16,False,False,False,True,0,iter_demo_form_d,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,64,64,8192,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: 直写GM",134217728,0.0,8.388608e-05,0.0,8.388608e-05,16,0.0,4294967296.0,9.942053925925925e-06,524288,3.2768e-07,0.0,8.421376e-05,7.852183703703703e-07,8.499897837037037e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
-streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,False,False,False,True,0,streamk_demo,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,128,128,320,256,1,K段标准分块流水,128,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=1.00, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",327680.0,0.0,6.5536e-06,0.0,6.5536e-06,0.0,0.0,41943040.0,2.761681646090535e-06,0.0,0.0,3.4067453613053613e-06,6.5536e-06,3.4067453613053613e-06,9.960345361305361e-06,MTE2_GM,True,,访存Bound(GM),"P<=C/2, B/M/N并行度买不满, 切K (grid_K=32)"
+streamk_demo,4,4,128,128,10240,bf16,bf16,bf16,False,False,False,True,0,streamk_demo,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,128,128,320,256,1,K段标准分块流水,128,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=1.00, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",20971520,0,1.31072e-05,0.0,1.31072e-05,0.0,0.0,1342177280.0,3.1068918518518518e-06,0.0,0.0,3.4067453613053613e-06,1.31072e-05,3.4067453613053613e-06,1.651394536130536e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),"P<=C/2, B/M/N并行度买不满, 切K (grid_K=32)"
-asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,False,False,False,True,0,asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,47,47,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,176,176,1024,256,1,双缓冲驻留当前tile输入,176,176,80,allocate(输入驻留L2吸收重复读),direct_gm(输出直写GM不占L2),4,0,方案B,52,52,1,52,52,2,139,True,2,"L2场景B_输入驻留输出直写GM, r_in=1.00; 尾轮: 周长型主导, rho=0.06<rho_dv=0.53, A1b被dValue卡死, 方案B反超",67108864.0,0.0,4.194304e-05,0.0,4.194304e-05,0.0,0.0,274877906944.0,0.0005655924011193416,268435456,0.00016777216,0.0,0.0005655924011193416,0.0,0.0005655924011193416,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+asw_demo_full,2,2,8192,8192,1024,bf16,bf16,bf16,False,False,False,True,0,asw_demo_full,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1024,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=16, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入512.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=64.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",67108864,1006632960,4.194304e-05,0.00019358326153846154,0.00023552630153846153,0.0,0.0,274877906944.0,0.0006362914512592592,268435456,0.00016777216,0.0,0.0006362914512592592,0.0,0.0006362914512592592,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
-asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,False,False,False,True,0,asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,direct_gm,0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",2097152.0,0.0,2.62144e-06,0.0,2.62144e-06,0.0,0.0,268435456.0,1.104672658436214e-06,2097152,2.62144e-06,0.0,2.62144e-06,0.0,2.62144e-06,MTE2_GM,True,,访存Bound(GM),ASW_Basic兜底 (StreamK未过: 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2)
+asw_demo_reduce_core,16,16,256,256,128,bf16,bf16,bf16,False,False,False,True,0,asw_demo_reduce_core,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,16,1,1,1,1,"降核: 只用16核, 每核一个L0C满载输出块, 其余核闲置",0,1,256,256,128,128,1,标准核内流水,256,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=16.00<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",2097152,0,2.62144e-06,0.0,2.62144e-06,0.0,0.0,268435456.0,1.2427567407407407e-06,2097152,8.065969230769231e-07,0.0,2.62144e-06,0.0,2.62144e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2)
+b4_m1_n128_k256,4,4,1,128,256,bf16,bf16,bf16,False,False,False,True,0,b4_m1_n128_k256,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,1,128,256,256,1,标准核内流水,1,128,128,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.01<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",264192,0,5.28384e-06,0.0,5.28384e-06,0.0,0.0,262144.0,1.9418074074074073e-08,1024,6.3015384615384615e-09,0.0,5.28384e-06,0.0,5.28384e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受: K > grid_K^2/(grid_K-1)*theta_c)
+b8_m2_n192_k128,8,8,2,192,128,bf16,bf16,bf16,False,False,False,True,0,b8_m2_n192_k128,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,2,192,128,128,1,标准核内流水,2,192,80,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.05<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",397312,0,7.94624e-06,0.0,7.94624e-06,0.0,0.0,786432.0,5.825422222222222e-08,6144,3.7809230769230766e-08,0.0,7.94624e-06,0.0,7.94624e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受: K > grid_K^2/(grid_K-1)*theta_c)
+b16_m4_n256_k192,16,16,4,256,192,bf16,bf16,bf16,False,False,False,True,0,b16_m4_n256_k192,ASW_Basic_降核,Ascend950PR,batch_mat_mul_v3,1,1,1,1,1,"降核: 只用1核, 每核一个L0C满载输出块, 其余核闲置",0,1,4,256,192,192,1,标准核内流水,4,256,64,allocate,resident(整case全驻留S_A),0,0,不涉及(每核一块无尾轮),1,1,1,0,0,0,0,True,2,"P=0.25<C, 降核是理性选择 (强切则 tile 跌破搬移效率下限反而更慢)",1597440,0,3.19488e-05,0.0,3.19488e-05,0.0,0.0,6291456.0,4.660337777777778e-07,32768,2.0164923076923077e-07,0.0,3.19488e-05,0.0,3.19488e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受: K > grid_K^2/(grid_K-1)*theta_c)
+b32_m8_n128_k256,32,32,8,128,256,bf16,bf16,bf16,False,False,False,True,0,b32_m8_n128_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,8,128,256,256,1,a_单batch全驻留,8,128,128,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,单batch全驻留: (MK+KN)*dtype=68KB <= L1; 输出落点: L2驻留,2228224,0.0,1.6384e-06,0.0,1.6384e-06,1,0.0,16777216.0,3.8836148148148146e-08,65536,1.2603076923076923e-08,0.0,1.6384e-06,5.1439225071225065e-08,1.689839225071225e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
+b64_m16_n256_k512,64,64,16,256,512,bf16,bf16,bf16,False,False,False,True,0,b64_m16_n256_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,16,256,512,240,1,c_一侧驻留+对侧切K,16,256,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"一侧驻留(A)+对侧切K: A驻留16KB, 预算L1/2, k_L1=240, dValueA=480B/dValueB=512B; 输出落点: L2驻留",17825792,0.0,1.114112e-05,0.0,1.114112e-05,6,0.0,268435456.0,6.213783703703703e-07,524288,1.0082461538461538e-07,0.0,1.114112e-05,1.9604786324786327e-07,1.1337167863247863e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
+b128_m8_n192_k256,128,128,8,192,256,bf16,bf16,bf16,False,False,False,True,0,b128_m8_n192_k256,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),4,1,8,192,256,256,2,b_双batch乒乓,8,192,80,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,双batch乒乓: 2*(MK+KN)*dtype=200KB <= L1; 输出落点: L2驻留,13107200,0.0,9.17504e-06,0.0,9.17504e-06,4,0.0,100663296.0,2.330168888888889e-07,393216,7.561846153846153e-08,0.0,9.17504e-06,7.71588376068376e-08,9.252198837606838e-06,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
+b32_m16_n8192_k7168,32,32,16,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m16_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,1,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,16,1024,7168,112,1,双缓冲驻留当前tile输入,16,1024,16,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: 效率降级(放开约束4, dValue 按 128B 硬下限, 搬移效率低于模型假设, 时延可能低估): P=1, mCnt=1 x nCnt=8 (tile 16x1024, 每batch搬入113.8MB, r=0); Base tile 16x1024 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=3591.0MB, V_out=8.0MB, L2=128MB); 尾轮: r=0 无尾轮",3765436416,51380224,0.00235339776,9.880812307692307e-06,0.0023632785723076925,0.0,0.0,60129542144.0,0.00013918875496296296,8388608,5.24288e-06,0.0,0.0023685214523076923,0.0,0.0023685214523076923,MTE2,True,,访存Bound(GM读写共享+L2重复读),"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B [效率降级标注: 搬移效率低于模型假设, 时延可能低估, 见 plan.note]"
+b4_m1_n8192_k8192,4,4,1,8192,8192,bf16,bf16,bf16,False,False,False,True,0,b4_m1_n8192_k8192,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,32,"B/M/N切出4块, 每块32核切K归约 (归约组内核c负责K段[c*K/32,(c+1)*K/32))",1,1,1,8192,256,256,1,K段标准分块流水,1,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,4194304,grid_K=32路切K+归约,1,1,32,0,0,0,0,True,4,"P=0.50, grid_K=32, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",536936448,0.0,0.00033558528,0.0,0.00033558528,0.0,0.0,536870912.0,1.2427567407407407e-06,0.0,0.0,1.731729603729604e-06,0.00033558528,1.731729603729604e-06,0.0003373170096037296,MTE2,True,,访存Bound(GM读写共享+L2重复读),"P<=C/2, B/M/N并行度买不满, 切K (grid_K=32)"
+b16_m2_n4096_k7168,16,16,2,4096,7168,bf16,bf16,bf16,False,False,False,True,0,b16_m2_n4096_k7168,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,1,16,"B/M/N切出16块, 每块16核切K归约 (归约组内核c负责K段[c*K/16,(c+1)*K/16))",1,1,2,4096,448,256,1,K段标准分块流水,2,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,8388608,grid_K=16路切K+归约,1,1,16,0,0,0,0,True,4,"P=2.00, grid_K=16, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",939982848,0,0.00058748928,0.0,0.00058748928,0.0,0.0,1879048192.0,4.349648592592593e-06,0.0,0.0,1.7726896037296038e-06,0.00058748928,1.7726896037296038e-06,0.0005892619696037297,MTE2,True,,访存Bound(GM读写共享+L2重复读),"P<=C/2, B/M/N并行度买不满, 切K (grid_K=16)"
+b32_m64_n64_k7168,32,32,64,64,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m64_n64_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,64,64,7168,1024,1,d_两侧都切K,64,64,256,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=1024, K段成对流水, batch边界天然无缝; dValueA=2048B/dValueB=128B; 输出落点: L2驻留",58720256,0.0,3.670016e-05,0.0,3.670016e-05,7,0.0,1879048192.0,4.349648592592593e-06,262144,5.041230769230769e-08,0.0,3.670016e-05,6.71790678062678e-07,3.737195067806268e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),仅 IterBatch 条件满足
+b64_m1024_n1024_k7168,64,64,1024,1024,7168,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n1024_k7168,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),2,1,1024,1024,7168,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"direct_gm(整case超L2: 输入优先驻留L2, 输出直写GM不占L2)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: 直写GM",1879048192,0.0,0.00117440512,0.0,0.00117440512,224,0.0,962072674304.0,0.0022270200794074074,134217728,8.388608e-05,0.0,0.0022270200794074074,5.1885093925925924e-05,0.0022789051733333333,MMAD,True,,计算Bound,仅 IterBatch 条件满足
+b128_m2048_n2048_k1536,128,128,2048,2048,1536,bf16,bf16,bf16,False,False,False,True,0,b128_m2048_n2048_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入48.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1536.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",1610612736,4831838208,0.00100663296,0.0009291996553846154,0.0019358326153846154,0.0,0.0,1649267441664.0,0.0038177487075555555,1073741824,0.00067108864,0.0,0.0038177487075555555,0.0,0.0038177487075555555,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b64_m1024_n8192_k2048,64,64,1024,8192,2048,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n8192_k2048,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,2048,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=16 (tile 512x512, 每batch搬入128.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2304.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",2415919104,6174015488,0.00150994944,0.0011873106707692308,0.0026972601107692305,0.0,0.0,2199023255552.0,0.005090331610074074,1073741824,0.00067108864,0.0,0.005090331610074074,0.0,0.005090331610074074,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b32_m1024_n1024_k512,32,32,1024,1024,512,bf16,bf16,bf16,False,False,False,True,0,b32_m1024_n1024_k512,IterBatch,Ascend950PR,batch_mat_mul_v3,32,32,1,1,1,切B轮转分配(核间零重复读零依赖),1,1,1024,1024,512,64,1,d_两侧都切K,176,176,64,allocate(GM->L1随路驻留L2),"resident(整case输入+输出<=L2: 输出驻留L2异步回写, GM写=0)",0,0,不涉及(核内不切M/N),1,1,1,0,0,0,0,True,2,"两侧都切K: k_L1=64, K段成对流水, batch边界天然无缝; dValueA=128B/dValueB=2048B; 输出落点: L2驻留",67108864,0.0,4.194304e-05,0.0,4.194304e-05,8,0.0,34359738368.0,7.95364314074074e-05,67108864,1.290555076923077e-05,0.0,7.95364314074074e-05,2.2847604695156693e-05,0.0001023840361025641,MMAD,True,,计算Bound,仅 IterBatch 条件满足
+b128_m4096_n4096_k8192,128,128,4096,4096,8192,bf16,bf16,bf16,False,False,False,True,0,b128_m4096_n4096_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16384.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮",17179869184,120259084288,0.01073741824,0.023126746978461538,0.033864165218461535,0.0,0.0,35184372088832.0,0.08144530576118518,4294967296,0.00268435456,0.0,0.08144530576118518,0.0,0.08144530576118518,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b32_m8192_n4096_k7168,32,32,8192,4096,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m8192_n4096_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=8 (tile 512x512, 每batch搬入1792.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=5376.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮",5637144576,54492397568,0.00352321536,0.010479307224615384,0.014002522584615384,0.0,0.0,15393162788864.0,0.03563232127051852,2147483648,0.00134217728,0.0,0.03563232127051852,0.0,0.03563232127051852,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b32_m2048_n2048_k8192,32,32,2048,2048,8192,bf16,bf16,bf16,False,False,False,True,0,b32_m2048_n2048_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,4,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=4 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2048.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",2147483648,6442450944,0.00134217728,0.0012389328738461537,0.002581110153846154,0.0,0.0,2199023255552.0,0.005090331610074074,268435456,0.00016777216,0.0,0.005090331610074074,0.0,0.005090331610074074,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b64_m4096_n2048_k128,64,64,4096,2048,128,bf16,bf16,bf16,False,False,False,True,0,b64_m4096_n2048_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=8 x nCnt=4 (tile 512x512, 每batch搬入8.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=96.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",100663296,436207616,6.291456e-05,8.388608e-05,0.00014680063999999998,0.0,0.0,137438953472.0,0.0003181457256296296,1073741824,0.00067108864,0.0,0.00081788928,0.0,0.00081788928,MTE2,True,,访存Bound(GM读写共享+L2重复读),"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b16_m1024_n1024_k8192,16,16,1024,1024,8192,bf16,bf16,bf16,False,False,False,True,0,b16_m1024_n1024_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=2 x nCnt=2 (tile 512x512, 每batch搬入64.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=512.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",536870912,536870912,0.00033554432,0.00010324440615384615,0.0004387887261538461,0.0,0.0,274877906944.0,0.0006362914512592592,33554432,2.097152e-05,0.0,0.0006362914512592592,0.0,0.0006362914512592592,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b4_m32768_n128_k128,4,4,32768,128,128,bf16,bf16,bf16,False,False,False,True,0,b4_m32768_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"resident(整case全驻留S_A: 输出驻留L2异步回写, GM写=0)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入10.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: A_整case全驻留(输入+输出<=L2) (V_in=32.1MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",33685504,8257536,2.105344e-05,1.5879876923076922e-06,2.2641427692307692e-05,0.0,0.0,4294967296.0,9.942053925925925e-06,33554432,6.452775384615385e-06,0.0,2.2641427692307692e-05,0.0,2.2641427692307692e-05,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b8_m32768_n2048_k512,8,8,32768,2048,512,bf16,bf16,bf16,False,False,False,True,0,b8_m32768_n2048_k512,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,4,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,512,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,64,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=64 x nCnt=4 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=272.0MB, V_out=1024.0MB, L2=128MB); 尾轮: r=0 无尾轮",285212672,1862270976,0.00017825792,0.00035812903384615385,0.0005363869538461539,0.0,0.0,549755813888.0,0.0012725829025185184,1073741824,0.00067108864,0.0,0.0012725829025185184,0.0,0.0012725829025185184,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b4_m131072_n128_k128,4,4,131072,128,128,bf16,bf16,bf16,False,False,False,True,0,b4_m131072_n128_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,128,128,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,32,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=256 x nCnt=1 (tile 512x128, 每batch搬入40.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=128.1MB, V_out=128.0MB, L2=128MB); 尾轮: r=0 无尾轮",134348800,33423360,8.3968e-05,6.427569230769231e-06,9.039556923076924e-05,0.0,0.0,17179869184.0,3.97682157037037e-05,134217728,8.388608e-05,0.0,0.00017428164923076922,0.0,0.00017428164923076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b8_m131072_n1024_k256,8,8,131072,1024,256,bf16,bf16,bf16,False,False,False,True,0,b8_m131072_n1024_k256,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,2,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,256,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,128,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=2 (tile 512x512, 每batch搬入256.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=516.0MB, V_out=2048.0MB, L2=128MB); 尾轮: r=0 无尾轮",541065216,1606418432,0.00033816576,0.00030892662153846154,0.0006470923815384616,0.0,0.0,549755813888.0,0.0012725829025185184,2147483648,0.00134217728,0.0,0.0019892696615384617,0.0,0.0019892696615384617,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b16_m32768_n8192_k7168,16,16,32768,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b16_m32768_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,512,True,2,"tile枚举: P=2, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入14336.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=8960.0MB, V_out=8192.0MB, L2=128MB); 尾轮: r=0 无尾轮",9395240960,231122927616,0.0058720256,0.044446716849230766,0.05031874244923076,0.0,0.0,61572651155456.0,0.14252928508207408,8589934592,0.00536870912,0.0,0.14252928508207408,0.0,0.14252928508207408,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b4_m32768_n128_k8192,4,4,32768,128,8192,bf16,bf16,bf16,False,False,False,True,0,b4_m32768_n128_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,1,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,128,8192,192,1,双缓冲驻留当前tile输入,512,128,32,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,8,True,2,"tile枚举: P=8, 有界枚举最优 mCnt=64 x nCnt=1 (tile 512x128, 每batch搬入640.0MB, r=0); Base tile 512x128 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=2056.0MB, V_out=32.0MB, L2=128MB); 尾轮: r=0 无尾轮",2155872256,528482304,0.00134742016,0.0001016312123076923,0.0014490513723076923,0.0,0.0,274877906944.0,0.0006362914512592592,33554432,2.097152e-05,0.0,0.0014700228923076922,0.0,0.0014700228923076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b32_m131072_n8192_k128,32,32,131072,8192,128,bf16,bf16,bf16,False,False,False,True,0,b32_m131072_n8192_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,4096,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入1024.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=1088.0MB, V_out=65536.0MB, L2=128MB); 尾轮: r=0 无尾轮",1140850688,33218887680,0.00071303168,0.006388247630769231,0.007101279310769231,0.0,0.0,8796093022208.0,0.020361326440296295,68719476736,0.04294967296,0.0,0.05005095227076922,0.0,0.05005095227076922,MTE2,True,,访存Bound(GM读写共享+L2重复读),"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b64_m32768_n8192_k1536,64,64,32768,8192,1536,bf16,bf16,bf16,False,False,False,True,0,b64_m32768_n8192_k1536,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,64,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,1536,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,2048,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=64 x nCnt=16 (tile 512x512, 每batch搬入3072.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7680.0MB, V_out=32768.0MB, L2=128MB); 尾轮: r=0 无尾轮",8053063680,198105366528,0.0050331648,0.03809718587076923,0.04313035067076923,0.0,0.0,52776558133248.0,0.12216795864177778,34359738368,0.02147483648,0.0,0.12216795864177778,0.0,0.12216795864177778,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b8_m131072_n8192_k8192,8,8,131072,8192,8192,bf16,bf16,bf16,False,False,False,True,0,b8_m131072_n8192_k8192,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,256,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,8192,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=256 x nCnt=16 (tile 512x512, 每batch搬入65536.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: C_双侧超L2: 最小替换2D分组(组间落空GM, 窗口L2) (V_in=17408.0MB, V_out=16384.0MB, L2=128MB); 最小替换分组 m_grp=8x n_grp=8 (GM倍率3.76, 窗口L2/batch=57344.0MB); 尾轮: r=0 无尾轮",68719476736,481036337152,0.04294967296,0.09250698791384615,0.13545666087384614,0.0,0.0,140737488355328.0,0.3257812230447407,17179869184,0.01073741824,0.0,0.3257812230447407,0.0,0.3257812230447407,MMAD,True,,计算Bound,ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b8_m16_n7168_k1536,8,8,16,7168,1536,bf16,bf16,bf16,False,False,False,True,0,b8_m16_n7168_k1536,StreamK,Ascend950PR,batch_mat_mul_v3,32,1,1,2,2,"B/M/N切出16块, 每块2核切K归约 (归约组内核c负责K段[c*K/2,(c+1)*K/2))",1,1,16,3584,768,256,1,K段标准分块流水,16,128,64,allocate(部分和驻留L2),"resident(部分和4B驻留L2, 防精度丢失不随C的fp16/fp8转换)",0,3670016,grid_K=2路切K+归约,1,1,2,0,0,0,0,True,4,"P=14.00, grid_K=2, 部分和驻留L2按4B写出, AIV归约后按C dtype=2B写最终",176553984,393216,0.00011034624,7.561846153846153e-08,0.00011042185846153846,0.0,0.0,2818572288.0,6.524472888888889e-06,0.0,0.0,2.566079254079254e-07,0.00011042185846153846,2.566079254079254e-07,0.00011067846638694638,MTE2,True,,访存Bound(GM读写共享+L2重复读),"P<=C/2, B/M/N并行度买不满, 切K (grid_K=2)"
+b64_m1024_n7168_k7168,64,64,1024,7168,7168,bf16,bf16,bf16,False,False,False,True,0,b64_m1024_n7168_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,2,14,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,56,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=2 x nCnt=14 (tile 512x512, 每batch搬入392.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=896.0MB, L2=128MB); 尾轮: r=0 无尾轮",7516192768,18790481920,0.00469762048,0.0036135542153846152,0.008311174695384616,0.0,0.0,6734508720128.0,0.015589140555851852,939524096,0.00058720256,0.0,0.015589140555851852,0.0,0.015589140555851852,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b32_m8192_n8192_k7168,32,32,8192,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b32_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,256,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=7168.0MB, V_out=4096.0MB, L2=128MB); 尾轮: r=0 无尾轮",7516192768,112742891520,0.00469762048,0.021681325292307693,0.026378945772307694,0.0,0.0,30786325577728.0,0.07126464254103704,4294967296,0.00268435456,0.0,0.07126464254103704,0.0,0.07126464254103704,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 1_batch关系与每核份额: BatchA==BatchB 且 b_core=B/C>=2*b0; 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+b8_m4096_n4096_k128,8,8,4096,4096,128,bf16,bf16,bf16,False,False,False,True,0,b8_m4096_n4096_k128,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,8,8,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,128,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,16,True,2,"tile枚举: P=4, 有界枚举最优 mCnt=8 x nCnt=8 (tile 512x512, 每batch搬入16.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=16.0MB, V_out=256.0MB, L2=128MB); 尾轮: r=0 无尾轮",16777216,117440512,1.048576e-05,2.2584713846153845e-05,3.307047384615384e-05,0.0,0.0,34359738368.0,7.95364314074074e-05,268435456,0.00016777216,0.0,0.00020084263384615383,0.0,0.00020084263384615383,MTE2,True,,访存Bound(GM读写共享+L2重复读),ASW_Basic兜底 (StreamK未过: 1_并行缺口: P=B*MN*4B/L0C <= C/2; 2_单核K段下限: K/grid_K >= 256B/dtype 且 grid_K>=2; 3_归约代价可接受)
+b128_m8192_n8192_k7168,128,128,8192,8192,7168,bf16,bf16,bf16,False,False,False,True,0,b128_m8192_n8192_k7168,ASW_Basic,Ascend950PR,batch_mat_mul_v3,32,1,16,16,1,B->M->N线性映射+ASW滑窗蛇形(W=4),0,1,512,512,7168,128,1,双缓冲驻留当前tile输入,256,256,64,allocate(输入驻留L2吸收重复读),"direct_gm(输出直写GM, 输入优先驻留L2)",4,0,A0,1,1,1,0,0,0,1024,True,2,"tile枚举: P=1, 有界枚举最优 mCnt=16 x nCnt=16 (tile 512x512, 每batch搬入3584.0MB, r=0); Base tile 256x256 (L0C 单缓冲方形用满); L2场景: B_单batch可驻留(全驻留或单侧驻留+对侧滑窗) (V_in=28672.0MB, V_out=16384.0MB, L2=128MB); 尾轮: r=0 无尾轮",30064771072,450971566080,0.01879048192,0.08672530116923077,0.10551578308923078,0.0,0.0,123145302310912.0,0.28505857016414815,17179869184,0.01073741824,0.0,0.28505857016414815,0.0,0.28505857016414815,MMAD,True,,计算Bound,"IterBatch/MergeBatch 进入条件均不满足, 回落 ASW_Basic; IterBatch未过: 3_L1驻留形态(四选一, 核心要求: 单batch核内零重复读); MergeBatch未过: 2_L0C容量: 2*(b0*M)*(b0*N)*4B <= L0C; 2b_L0AB容量: 合并后 b0*M/b0*N 在最小K粒度下可驻留 L0A/L0B; 5_访存Bound: 2MN/(M+N) < R16/b0"
+special_k1_b64,64,64,8192,512,1,bf16,bf16,bf16,False,False,False,True,0,special_k1_b64,特殊分支,Ascend950PR,batch_mat_mul_v3,64,1,1,1,1,AIV 核间按行均分 (无 Cube tile 概念),0,1,0,0,1,0,1,UB驻留(AIV) AIV单缓冲,0,0,0,allocate,direct_gm,0,0,不涉及(AIV逐元素),1,1,1,0,0,0,0,False,2,"K=1逐元素乘: 退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; 走 AIV 通路 GM->UB->Mul->GM, AIV单缓冲 (B<2*AIV 逐batch单缓冲串行)",1114112,0.0,6.9632e-07,0.0,6.9632e-07,0.0,0.0,268435456.0,9.92969696969697e-06,536870912,0.00033554432,0.0,0.00033624063999999996,0.0,0.00033624063999999996,MTE2,True,,访存Bound(GM读写共享+L2重复读),"K=1逐元素乘, 走AIV向量通路"

BMM/BMM_Theory/tests/test_branches.py

@@ -12,6 +12,7 @@
 import unittest
 
 from bmm_theory.models import BmmCase
+from bmm_theory.hardware import ASCEND950PR
 from bmm_theory.router import BranchRouter
 from bmm_theory.branches.merge_batch import MergeBatchBranch
 from bmm_theory.branches.iter_batch import IterBatchBranch
@@ -108,12 +109,22 @@ class TestArbitration(unittest.TestCase):
         self.assertIn("L1绑定", detail)
 
     def test_large_batch_mergebatch_wins(self):
-        # 大 B + 小 MN + K 截断: MergeBatch 应胜 (v1.1 §4.5)
+        # 大 B + 小 MN + K 截断 + 合并 tile 效率差: MergeBatch 应胜
-        case = mkcase(2048, 32, 32, 256)
+        # (issue#36: (2048,16,64,128) IterBatch A tile=4KB eff=0.25 vs 合并
+        #  16KB eff=1.0, 效率节省 15.7us >> drain 0.17us; T_cmd=0 默认下仍胜)
+        case = mkcase(2048, 16, 64, 128)
         self.assertTrue(MergeBatchBranch().analyze(case).capable)
         win, detail = self.mb.beats_iterbatch(case)
         self.assertTrue(win, detail)
 
+    def test_no_eff_diff_mergebatch_loses(self):
+        # issue#36: K 截断但 IterBatch 单命令 tile 已达 16KB 饱和 (效率打平),
+        # T_cmd=0 下命令节省为 0, 合并只剩 drain 惩罚 -> IterBatch 优
+        # (原 t_cmd=50ns 时代的 MergeBatch 胜例 (2048,32,32,256), 语义迁移)
+        case = mkcase(2048, 32, 32, 256)
+        win, detail = self.mb.beats_iterbatch(case)
+        self.assertFalse(win, detail)
+
 
 class TestTimingSanity(unittest.TestCase):
     """时延模型自洽性."""
@@ -134,14 +145,16 @@ class TestTimingSanity(unittest.TestCase):
         # MergeBatch case 必为访存 Bound (进入条件 5)
         case = mkcase(128, 64, 64, 512)
         mb = MergeBatchBranch().analyze(case)
-        self.assertIn(mb.timing.bottleneck, ("MTE2_GM", "MTE2_L2"))
+        self.assertIn(mb.timing.bottleneck, ("MTE2", "MMAD"))
 
     def test_fixpipe_dtype_conversion(self):
-        # C 指定 fp16 输出时, 写出量按 2B 而非 L0C 的 4B
+        # C 指定 fp16 输出时, 写出量按 2B 而非 L0C 的 4B;
+        # 字节列为整芯片口径 (issue#29): fixpipe_bytes = B*MN*outB
         case = mkcase(128, 64, 64, 512, dtype_c="fp16")
         ib = IterBatchBranch().analyze(case)
-        expect = ib.plan.b_core * 64 * 64 * 2
+        expect = case.batch_c * 64 * 64 * 2
         self.assertAlmostEqual(ib.timing.fixpipe_bytes, expect)
+        self.assertAlmostEqual(ib.timing.fixpipe_bytes, case.output_bytes)
 
 
 class TestIssueRegression(unittest.TestCase):
@@ -151,7 +164,7 @@ class TestIssueRegression(unittest.TestCase):
         self.router = BranchRouter()
 
     def test_issue4_k1_small_batch_real_plan(self):
-        # issue#4/#12: K=1 且 B<128 不崩溃, 且给出 AIV 单缓冲真实方案 (不再是无方案占位)
+        # issue#4/#12/#17: K=1 且 B<128 不崩溃, 且给出 AIV 单缓冲真实方案 (不再是无方案占位)
         case = mkcase(64, 8192, 32, 1, dtype_a="int8", dtype_b="int8")
         r = self.router.route(case)
         self.assertIsNotNone(r["plan"])
@@ -210,26 +223,26 @@ class TestIssueRegression(unittest.TestCase):
 
 
 class TestIssueRegression2(unittest.TestCase):
-    """第二轮复评问题 (#11-#16) 回归."""
+    """第三轮复评问题 (#17-#20) + 恢复 #11-#15 回归."""
 
     def setUp(self):
         self.router = BranchRouter()
 
     def test_issue11_streamk_fixpipe_no_double_count(self):
-        # issue#11: 部分和写出只经 t_reduce 计账一次; 稳态 fixpipe 不得再计
+        # issue#11/#17: 部分和写出只经 t_reduce 计账一次; 稳态 fixpipe 不得再计
         from bmm_theory.branches.stream_k import StreamKBranch
         case = mkcase(4, 128, 128, 10240)
         sk = StreamKBranch().analyze(case)
+        self.assertTrue(sk.capable)
         t = sk.timing
         self.assertAlmostEqual(t.t_fixpipe, 0.0)   # 归约串行口径下无稳态 fixpipe 账
         self.assertAlmostEqual(t.fixpipe_bytes, 0.0)
-        # 端到端 = max(MTE2, MMAD) + 归约, 不再虚高到 55us/FIXPIPE
+        # 端到端 = 稳态(MTE2搬移链) + 归约, fixpipe 无重复账
-        expect = max(t.t_mte2, t.t_mmad) + t.t_reduce
+        self.assertEqual(t.bottleneck, "MTE2")
-        self.assertAlmostEqual(t.t_total, expect)
+        self.assertAlmostEqual(t.t_total, t.t_steady + t.t_drain)
-        self.assertEqual(t.bottleneck, "MTE2_GM")
 
     def test_issue12_k1_pingpong_still_ok(self):
-        # issue#12: K=1 且 B>=128 仍走 UB 乒乓 (原行为不变)
+        # issue#12/#17: K=1 且 B>=128 仍走 UB 乒乓 (原行为不变)
         r = self.router.route(mkcase(128, 256, 256, 1))
         self.assertEqual(r["branch"], "特殊分支")
         self.assertIn("乒乓", r["plan"].l1_form)
@@ -248,9 +261,23 @@ class TestIssueRegression2(unittest.TestCase):
 
     def test_issue13_router_fallback_when_winner_infeasible(self):
         # issue#13: 仲裁胜出的 MergeBatch 自检违规时, 回退到可行候选 IterBatch
+        # (issue#31 后该样例 MergeBatch 时延已不再胜出, 用 mock 压低其时延以固定
+        #  "胜者违规 -> 回退" 机制路径; MergeBatch 方案 dValue=112B<128 仍违规)
+        from unittest import mock
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
         from bmm_theory.constraints import check_plan_constraints
-        case = mkcase(256, 1, 256, 4096, dtype_a="int8", dtype_b="int8")  # 原 0.2% 违规样例
+        case = mkcase(256, 1, 256, 4096, dtype_a="int8", dtype_b="int8")
-        r = self.router.route(case)
+        mb_plan = MergeBatchBranch().analyze(case).plan
+        self.assertTrue(check_plan_constraints(case, mb_plan),
+                        "该样例 MergeBatch 方案应仍自检违规 (dValue)")
+        real_eval = MergeBatchBranch.evaluate
+
+        def fake_win(self, c, p):
+            t = real_eval(self, c, p)
+            t.t_total = 1e-9  # 强制 MergeBatch 时延胜出 -> 仲裁选它 -> 走自检/回退
+            return t
+        with mock.patch.object(MergeBatchBranch, "evaluate", fake_win):
+            r = self.router.route(case)
         self.assertEqual(r["branch"], "IterBatch")       # 回退
         self.assertIn("自检违规", r["arbitration"])
         self.assertIn("回退", r["arbitration"])
@@ -265,6 +292,832 @@ class TestIssueRegression2(unittest.TestCase):
         with self.assertRaises(ValueError):
             BmmCase(case_id="bad", m=64, n=64, k=1, dtype_a="xxx")
 
+    def test_issue18_placeholder_plan_infeasible_in_evaluate(self):
+        # issue#18: 占位方案(used_core_num=0)在 evaluate 中必须不可行,
+        # 不得被当作可行方案给出正常时延
+        from bmm_theory.evaluator import PlanEvaluator
+        from bmm_theory.models import ImplPlan
+        # K=1 且 B<128 已恢复真实单缓冲方案 (#17), 故直接构造占位 plan 验证约束层
+        case = mkcase(64, 8192, 32, 1, dtype_a="int8", dtype_b="int8")
+        r = self.router.route(case)
+        self.assertGreater(r["plan"].used_core_num, 0)   # 真实方案
+        ph = ImplPlan(case_id="ph", branch="特殊分支", used_core_num=0,
+                      note="该区域暂无理论方案(进入条件不满足)")
+        er = PlanEvaluator().evaluate(case, ph)
+        self.assertFalse(er.feasible, "占位方案应判不可行")
+        self.assertIn("used_core_num", er.violations)
+
+    def test_issue19_transpose_dvalue_guard_effective(self):
+        # issue#19: 转置感知 dValue 判据在生成守卫/条件4/约束三处同源后真正生效.
+        # 判别形状: B=64 M=4096 N=64 K=4096 bf16 —— d 形态 k_l1=16,
+        # 均不转置时 dv_a=k_l1*2=32B <128 挡下 (B 侧 dv_b=N*2=128B 恰好达标也不放行,
+        # 因为两侧切 K 两侧都要高效);
+        # A 转置后 dv_a=M*2=8192B, 应能走 IterBatch 形态 d.
+        from bmm_theory.constraints import check_plan_constraints
+        c_not = BmmCase(case_id="x", batch_a=64, batch_b=64, m=4096, n=64, k=4096,
+                        trans_a=False, trans_b=False)
+        r_not = self.router.route(c_not)
+        self.assertNotEqual(r_not["branch"], "IterBatch")  # 非转置被 dValue 守卫挡下
+        c_tr = BmmCase(case_id="x", batch_a=64, batch_b=64, m=4096, n=64, k=4096,
+                       trans_a=True, trans_b=False)
+        r_tr = self.router.route(c_tr)
+        self.assertEqual(r_tr["branch"], "IterBatch")       # A 转置 M 向连续, 守卫放行
+        self.assertIn("d_", r_tr["plan"].l1_form)
+        self.assertEqual(check_plan_constraints(c_tr, r_tr["plan"]), [])
+
+
+class TestFp4Support(unittest.TestCase):
+    """fp4 (0.5B) dtype 支持 (对齐 bmmv3, 2026-09-03)."""
+
+    def test_fp4_dtype_bytes(self):
+        from bmm_theory.models import dtype_bytes
+        self.assertEqual(dtype_bytes("fp4"), 0.5)
+        self.assertEqual(dtype_bytes("fp4_e2m1"), 0.5)
+
+    def test_fp4_case_creation(self):
+        case = mkcase(32, 64, 64, 256, dtype_a="fp4", dtype_b="fp4", dtype_c="fp16")
+        self.assertEqual(case.dtype_in_bytes, 0.5)
+        # fp4 输入 + fp16 输出: 输入 0.5B, 输出 2B
+        self.assertEqual(case.dtype_out_bytes, 2)
+
+    def test_fp4_dvalue_threshold(self):
+        # fp4 (0.5B) 时 dValue 128B 需要 k_l1 >= 256
+        from bmm_theory.constraints import check_plan_constraints
+        case = mkcase(128, 64, 64, 128, dtype_a="fp4", dtype_b="fp4", dtype_c="fp16")
+        r = BranchRouter().route(case)
+        v = check_plan_constraints(case, r["plan"])
+        # fp4 小 K 场景应能正常路由且不报 dValue 违规 (k_l1 连续维是 M/N)
+        self.assertEqual(v, [], f"fp4 case 不应报违规: {v}")
+
+
+class TestTransposeModeling(unittest.TestCase):
+    """转置对 dValue 连续维的影响建模 (对齐 bmmv3, 2026-09-03)."""
+
+    def test_transpose_affects_dvalue_judgment(self):
+        # A 不转置: K 向连续, dValue 判 K*dt
+        # A 转置:   M 向连续, dValue 判 M*dt
+        # B 不转置: N 向连续, dValue 判 N*dt
+        # B 转置:   K 向连续, dValue 判 K*dt
+        from bmm_theory.constraints import _k_segment_is_contiguous
+        # 当前版本不建模转置, 默认按不转置处理
+        case = mkcase(128, 64, 64, 512)
+        case.trans_a = False
+        case.trans_b = False
+        # 验证 trans 字段存在且可读写 (为后续建模做准备)
+        self.assertFalse(case.trans_a)
+        self.assertFalse(case.trans_b)
+        case.trans_a = True
+        self.assertTrue(case.trans_a)
+
+    def test_transpose_a_large_m_small_k(self):
+        # A 转置 + 大 M 小 K: dValue 应判 M*dt (M 向连续), 不受 K 小影响
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        # M=1024 (M*dt=2048B >= 128B), K=8 (K*dt=16B < 128B)
+        case = mkcase(128, 1024, 64, 8, dtype_a="bf16", dtype_b="bf16")
+        case.trans_a = True
+        case.trans_b = False
+        ib = IterBatchBranch().analyze(case)
+        # A 转置时 dValue 判 M*dt=2048B >= 128B, 应通过
+        c4 = [c for c in ib.checks if "搬移效率" in c.name][0]
+        self.assertTrue(c4.passed, f"A 转置时应判 M 向连续: {c4.detail}")
+
+    def test_no_transpose_small_k_fails(self):
+        # A 不转置 + 小 K + c/d 形态: dValue 判 K*dt, K=8 时 16B < 128B 应失败
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        # 大 M/N 让 L1 放不下整 K, 走 c/d 形态
+        case = mkcase(128, 1024, 1024, 8, dtype_a="bf16", dtype_b="bf16")
+        case.trans_a = False
+        case.trans_b = False
+        ib = IterBatchBranch().analyze(case)
+        c4 = [c for c in ib.checks if "搬移效率" in c.name][0]
+        # c/d 形态下 A 不转置时 K=8 应报 dValue 违规
+        if ib.plan.l1_form.startswith(("c_", "d_")):
+            self.assertFalse(c4.passed, f"A 不转置时 K=8 应报 dValue 违规: {c4.detail}")
+
+
+class TestZeroCmdHandling(unittest.TestCase):
+    """950PR 默认 T_cmd=0 (未标定, issue#36): 整链路不得除零/崩溃; 合并收益由
+    搬移效率模型 (move_eff, 合并 tile 放大 b0 倍) 刻画, 不再设策略覆盖."""
+
+    def test_beats_iterbatch_zero_cmd(self):
+        # T_cmd=0: 命令节省项为 0, 由效率节省 vs drain 惩罚决定 (issue#36)
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
+        mb = MergeBatchBranch()  # 默认 spec 即 t_cmd_ns=0
+        # (b, m, n, k, MergeBatch应胜与否=效率节省>drain惩罚)
+        # (2048,16,64,128): iter A tile=4KB eff=0.25 vs 合并 16KB eff=1.0 -> 大胜
+        # (2048,32,32,256): iter A tile=16KB 已饱和, 效率打平 -> drain 惩罚 -> 恒劣
+        # (128,64,64,512):  issue#35 第三情形 (dValue cap 截断, 命令/效率均打平) -> 恒劣
+        # (256,128,128,4096): L1 绑定, tile 均 >=128KB 饱和 -> 恒劣
+        cases = [(2048, 16, 64, 128, True), (2048, 32, 32, 256, False),
+                 (128, 64, 64, 512, False), (256, 128, 128, 4096, False)]
+        for b, m, n, k, mb_wins in cases:
+            win, detail = mb.beats_iterbatch(mkcase(b, m, n, k))
+            self.assertEqual(win, mb_wins, f"{b},{m},{n},{k}: {detail}")
+            self.assertIn("效率节省", detail)
+
+    def test_route_with_zero_cmd_efficiency_decides(self):
+        # 默认 t_cmd=0: 有效率低下的 IterBatch 小 tile case 由 MergeBatch 胜;
+        # tile 均饱和的 case 由 IterBatch 胜 (drain 惩罚, 无策略覆盖)
+        router = BranchRouter()
+        r = router.route(mkcase(2048, 16, 64, 128))  # 效率差显著 -> MergeBatch
+        self.assertEqual(r["branch"], "MergeBatch")
+        self.assertIsNotNone(r["timing"])
+        r2 = router.route(mkcase(2048, 32, 32, 256))  # tile 均饱和 -> IterBatch
+        self.assertEqual(r2["branch"], "IterBatch")
+        shapes = [(128, 64, 64, 512), (64, 64, 64, 8192), (512, 128, 128, 128),
+                  (128, 128, 128, 1024), (32, 4096, 4096, 4096)]
+        for b, m, n, k in shapes:
+            rr = router.route(mkcase(b, m, n, k))
+            t = rr["timing"]
+            self.assertIsNotNone(t, f"{b},{m},{n},{k} 应有 timing")
+            self.assertGreater(t.t_total, 0.0)
+            self.assertEqual(t.t_total, t.t_total)  # 非 NaN
+
+    def test_zero_cmd_random_smoke(self):
+        # 随机小样本冒烟: 无异常/无 NaN
+        import random
+        from bmm_theory.hardware import NpuSpec
+        router = BranchRouter(NpuSpec(t_cmd_ns=0.0))
+        rng = random.Random(9)
+        for _ in range(300):
+            b = rng.choice([2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048])
+            m = rng.choice([8, 16, 32, 64, 128, 256, 512, 1024, 4096])
+            n = rng.choice([8, 16, 32, 64, 128, 256, 512, 1024, 4096])
+            k = rng.choice([16, 32, 64, 128, 256, 512, 1024, 4096, 16384])
+            c = BmmCase(case_id="z", batch_a=b, batch_b=b, m=m, n=n, k=k)
+            rr = router.route(c)
+            self.assertIsNotNone(rr["plan"])
+            t = rr["timing"]
+            self.assertIsNotNone(t)
+            self.assertGreater(t.t_total, 0.0)
+            self.assertEqual(t.t_total, t.t_total)
+
+
+class TestIssue27to30(unittest.TestCase):
+    """第四轮评审 (issue#27-#30, 设计文档 docs/05).
+
+    #27 MergeBatch Cube 公式复核 (每步 2*(b0M)(b0N)K x b_core/b0 步);
+    #28 dtype 感知算力 (Cube/AIV 速率表);
+    #29 GM 首读下限 GM>=V_in + 字节列整芯片口径;
+    #30 Fixpipe 输出落点 (整case驻留 -> L2, 否则直写 GM).
+    """
+
+    def setUp(self):
+        from bmm_theory.hardware import ASCEND950PR
+        self.s = ASCEND950PR
+        self.router = BranchRouter()
+
+    # ---------------- #27 ----------------
+    def test_issue27_merge_cube_flops_matches_formula(self):
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
+        case = mkcase(2048, 32, 32, 256)
+        mb = MergeBatchBranch().analyze(case)
+        self.assertTrue(mb.capable)
+        p, t = mb.plan, mb.timing
+        b0, bc = p.merge_b0, p.b_core
+        # 整芯片列 = B*b0*2MNK; 等价 (b_core/b0) 步 x 每步 2*(b0M)(b0N)K x C 核
+        step_flops = 2.0 * (b0 * 32) * (b0 * 32) * 256
+        self.assertAlmostEqual(t.cube_flops,
+                               step_flops * (bc / b0) * self.s.aic_num)
+        self.assertAlmostEqual(t.cube_flops,
+                               case.batch_c * b0 * 2.0 * 32 * 32 * 256)
+        # t_mmad = 每核 flops / 单核算力
+        self.assertAlmostEqual(t.t_mmad,
+                               (bc * b0 * 2.0 * 32 * 32 * 256) / self.s.q16)
+
+    def test_issue27_merge_gm_chip_col_equals_input(self):
+        # K 截断合并: GM 每字节一次 -> gm 整芯片列 = V_in (issue#29 口径)
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
+        case = mkcase(2048, 32, 32, 256)
+        t = MergeBatchBranch().analyze(case).timing
+        self.assertAlmostEqual(t.gm_read_bytes, case.input_bytes)
+        self.assertAlmostEqual(t.l2_read_bytes, 0.0)
+
+    # ---------------- #28 ----------------
+    def test_issue28_cube_dtype_rate(self):
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        base = mkcase(128, 64, 64, 512)  # IterBatch 形态 b, 各 dtype 均可行
+        t16 = IterBatchBranch().analyze(base).timing
+        t32 = IterBatchBranch().analyze(
+            mkcase(128, 64, 64, 512, dtype_a="fp32", dtype_b="fp32")).timing
+        t8 = IterBatchBranch().analyze(
+            mkcase(128, 64, 64, 512, dtype_a="fp8", dtype_b="fp8")).timing
+        # 同计算量, 时延反比于 dtype 算力: fp32=1/2, fp8=2x (相对 bf16)
+        self.assertAlmostEqual(t32.t_mmad / t16.t_mmad, 2.0, places=6)
+        self.assertAlmostEqual(t8.t_mmad / t16.t_mmad, 0.5, places=6)
+
+    def test_issue28_aiv_dtype_rate_k1(self):
+        # K=1 AIV 逐元素: bf16 通量 2x fp32 -> fp32 时延为 bf16 的 2 倍
+        base = mkcase(64, 8192, 512, 1)
+        t_bf16 = BranchRouter().route(base)["timing"]
+        t_fp32 = BranchRouter().route(
+            mkcase(64, 8192, 512, 1, dtype_a="fp32", dtype_b="fp32"))["timing"]
+        self.assertAlmostEqual(t_fp32.t_mmad / t_bf16.t_mmad, 2.0, places=6)
+
+    # ---------------- #29 ----------------
+    def test_issue29_gm_floor_and_chip_columns(self):
+        # 各分支代表 case: GM 整芯片列 >= V_in (R1), 且迭代/合并/转matmul/special
+        # 等于 V_in (无重复读结构)
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
+        from bmm_theory.branches.stream_k import StreamKBranch
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        checks = [
+            (IterBatchBranch().analyze(mkcase(128, 64, 64, 512)).timing,
+             mkcase(128, 64, 64, 512)),
+            (MergeBatchBranch().analyze(mkcase(2048, 32, 32, 256)).timing,
+             mkcase(2048, 32, 32, 256)),
+            (StreamKBranch().analyze(mkcase(4, 128, 128, 10240)).timing,
+             mkcase(4, 128, 128, 10240)),
+            (AswBasicBranch().analyze(mkcase(8, 4096, 4096, 1024)).timing,
+             mkcase(8, 4096, 4096, 1024)),
+        ]
+        for t, case in checks:
+            self.assertGreaterEqual(t.gm_read_bytes + 1e-6, case.input_bytes,
+                                    f"{case.case_id} GM < V_in")
+        # 转Matmul / 特殊分支 (K=1) 亦等于 V_in
+        r = self.router.route(mkcase(1, 2048, 2048, 2048))
+        t = r["timing"]
+        self.assertGreaterEqual(t.gm_read_bytes + 1e-6,
+                                BmmCase(case_id="x", batch_a=1, batch_b=1,
+                                        m=2048, n=2048, k=2048).input_bytes)
+        r = self.router.route(mkcase(128, 256, 256, 1))
+        self.assertAlmostEqual(r["timing"].gm_read_bytes, 128 * (256 + 256) * 2)
+
+    def test_issue29_gm_floor_random_smoke(self):
+        # 随机多 dtype 冒烟: 所有可行方案 GM 整芯片列 >= V_in, 无 NaN
+        import random
+        rng = random.Random(20260609)
+        dtypes = ["bf16", "fp16", "fp32", "int8", "fp8", "fp4"]
+        for _ in range(400):
+            dt = rng.choice(dtypes)
+            kw = dict(dtype_a=dt, dtype_b=dt, dtype_c=rng.choice(["bf16", "fp16"]))
+            b = rng.choice([1, 2, 4, 8, 16, 32, 64, 128, 256, 2048])
+            m = rng.choice([1, 8, 32, 64, 128, 256, 1024, 4096])
+            n = rng.choice([1, 8, 32, 64, 128, 256, 1024, 4096])
+            k = rng.choice([0, 1, 32, 64, 128, 256, 512, 1024, 4096])
+            c = BmmCase(case_id="gm", batch_a=b, batch_b=b, m=m, n=n, k=k, **kw)
+            r = self.router.route(c)
+            t = r["timing"]
+            self.assertIsNotNone(t, c.case_id)
+            self.assertEqual(t.t_total, t.t_total)  # 非 NaN
+            self.assertGreaterEqual(t.gm_read_bytes + 1e-6, c.input_bytes,
+                                    f"{dt} b={b} m={m} n={n} k={k} gm={t.gm_read_bytes}")
+
+    # ---------------- #30 ----------------
+    def test_issue30_iter_output_l2_when_whole_case_fits(self):
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        # 整 case 输入+输出 16.8MB+1.0MB <= 128MB -> 输出写 L2 写口, GM 写=0
+        case = mkcase(128, 64, 64, 512)
+        self.assertLess(case.input_bytes + case.output_bytes, self.s.l2_bytes)
+        t = IterBatchBranch().analyze(case).timing
+        self.assertAlmostEqual(t.fixpipe_bytes, case.output_bytes)
+        self.assertAlmostEqual(t.t_fixpipe, case.output_bytes / self.s.bw_l2)
+
+    def test_issue30_iter_output_gm_when_case_overflows(self):
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        # 输出 268MB > L2 余量 -> 直写 GM (输入优先驻留)
+        case = mkcase(128, 1024, 1024, 64)
+        self.assertGreater(case.input_bytes + case.output_bytes, self.s.l2_bytes)
+        t = IterBatchBranch().analyze(case).timing
+        self.assertAlmostEqual(t.fixpipe_bytes, case.output_bytes)
+        self.assertAlmostEqual(t.t_fixpipe, case.output_bytes / self.s.bw_gm)
+
+    def test_issue30_asw_scene_whole_case_labels(self):
+        # S_A (整case全驻留): l2_policy_out=resident, t_fixpipe 走 L2 写口
+        case_sa = mkcase(2, 4096, 4096, 512)
+        self.assertLess(case_sa.input_bytes + case_sa.output_bytes, self.s.l2_bytes)
+        r = self.router.route(case_sa)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        self.assertTrue(r["plan"].l2_policy_out.startswith("resident"), r["plan"].l2_policy_out)
+        self.assertIn("A_整case全驻留", r["plan"].note)
+        self.assertAlmostEqual(r["timing"].t_fixpipe,
+                               case_sa.output_bytes / self.s.bw_l2)
+        # S_B (整case超L2, 单batch输入可驻留): 输出直写 GM
+        case_sb = mkcase(64, 4096, 4096, 1024)
+        self.assertGreater(case_sb.input_bytes + case_sb.output_bytes, self.s.l2_bytes)
+        self.assertLess(4096 * 1024 * 2 * 2, self.s.l2_bytes)  # 单batch输入 16.8MB
+        r = self.router.route(case_sb)
+        self.assertIn("ASW", r["branch"])
+        self.assertTrue(r["plan"].l2_policy_out.startswith("direct_gm"),
+                        r["plan"].l2_policy_out)
+        self.assertAlmostEqual(r["timing"].t_fixpipe,
+                               case_sb.output_bytes / self.s.bw_gm)
+
+    def test_issue30_to_matmul_output_l2_toggle(self):
+        # 转Matmul 粗估: 整case可驻留 -> L2; 大 case -> GM
+        small = mkcase(1, 2048, 2048, 2048)
+        self.assertLess(small.input_bytes + small.output_bytes, self.s.l2_bytes)
+        r = self.router.route(small)
+        self.assertEqual(r["branch"], "转Matmul")
+        self.assertAlmostEqual(r["timing"].t_fixpipe,
+                               small.output_bytes / self.s.bw_l2)
+        big = mkcase(1, 8192, 8192, 4096)
+        self.assertGreater(big.input_bytes + big.output_bytes, self.s.l2_bytes)
+        r = self.router.route(big)
+        self.assertEqual(r["branch"], "转Matmul")
+        self.assertAlmostEqual(r["timing"].t_fixpipe,
+                               big.output_bytes / self.s.bw_gm)
+
+
+class TestIssue31to32(unittest.TestCase):
+    """第五轮评审: #31 切B类 GM 每字节恰读一次 = V_in; #32 ASW 场景升级
+    (单侧全驻留 -> S_B 使 GM=V_in; 双侧超 L2 -> S_C 最小替换 2D 分组 + 窗口 L2 计账)."""
+
+    def setUp(self):
+        from bmm_theory.hardware import ASCEND950PR
+        self.s = ASCEND950PR
+        self.router = BranchRouter()
+
+    # ---------------- #31 ----------------
+    def test_issue31_iter_form_c_gm_equals_vin(self):
+        # b64_m16_n256_k512: 形态 c, K=512 切 n_K=3 (k_L1=240) 非整除 -> GM 仍 = V_in
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        case = mkcase(64, 16, 256, 512)
+        ib = IterBatchBranch().analyze(case)
+        self.assertIn("c_", ib.plan.l1_form)
+        self.assertLess(ib.plan.k_l1, case.k)
+        t = ib.timing
+        self.assertAlmostEqual(t.gm_read_bytes, case.input_bytes)
+        self.assertAlmostEqual(t.l2_read_bytes, 0.0)
+        self.assertAlmostEqual(t.t_mte2_gm, case.input_bytes / self.s.bw_gm)
+
+    def test_issue31_iter_form_d_gm_equals_vin(self):
+        # K=5000 非整除切段 (k_L1=1024, n_K=5, 末段剩 904): GM 仍恰一次 = V_in
+        from bmm_theory.branches.iter_batch import IterBatchBranch
+        case = mkcase(128, 64, 64, 5000)
+        ib = IterBatchBranch().analyze(case)
+        self.assertIn("d_", ib.plan.l1_form)
+        self.assertLess(ib.plan.k_l1, case.k)
+        self.assertAlmostEqual(ib.timing.gm_read_bytes, case.input_bytes)
+        self.assertAlmostEqual(ib.timing.t_mte2_gm,
+                               case.input_bytes / self.s.bw_gm)
+
+    def test_issue31_merge_l1_bound_gm_equals_vin(self):
+        # MergeBatch L1 绑定 (k_L1<K): 切 K 段互不重叠, GM = V_in
+        from bmm_theory.branches.merge_batch import MergeBatchBranch
+        case = mkcase(811, 33, 1, 2459, dtype_a="fp32", dtype_b="fp32")
+        mb = MergeBatchBranch().analyze(case)
+        self.assertTrue(mb.capable)
+        self.assertLess(mb.plan.k_l1, case.k)   # L1 绑定
+        self.assertAlmostEqual(mb.timing.gm_read_bytes, case.input_bytes)
+        self.assertAlmostEqual(mb.timing.l2_read_bytes, 0.0)
+
+    # ---------------- #32 ----------------
+    def test_issue32_asw_single_side_resident_gm_equals_vin(self):
+        # b32_m8192_n4096_k7168: 单batch输入 176MB > L2, 但 B=58.7MB 可全驻留
+        # + A 行块滑窗 (58.7+2x2.5 <= 128) -> S_B: GM = V_in, 重复读全命中 L2
+        case = mkcase(32, 8192, 4096, 7168)
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        p, t = r["plan"], r["timing"]
+        a_b = case.m * case.k * case.dtype_in_bytes
+        bb_b = case.k * case.n * case.dtype_in_bytes
+        m_cnt, n_cnt = p.m_cnt, p.n_cnt
+        self.assertGreater(a_b + bb_b, self.s.l2_bytes)   # 单 batch 确实超 L2
+        self.assertLessEqual(bb_b + 2 * (a_b / m_cnt),
+                             self.s.l2_bytes)             # 单侧全驻留成立
+        self.assertAlmostEqual(t.gm_read_bytes, case.input_bytes)
+        exp_l2 = case.batch_c * ((n_cnt - 1) * a_b + (m_cnt - 1) * bb_b)
+        self.assertAlmostEqual(t.l2_read_bytes, exp_l2)
+
+    def test_issue32_asw_scene_c_min_gm(self):
+        # b8_m131072_n8192_k8192: 双侧均不可全驻留 -> S_C 最小替换分组;
+        # gm == 容量约束最小解 (测试内复算), 且 >= V_in
+        case = mkcase(8, 131072, 8192, 8192)
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        p, t = r["plan"], r["timing"]
+        self.assertIn("C_", p.note)
+        m_cnt, n_cnt = p.m_cnt, p.n_cnt
+        dt = case.dtype_in_bytes
+        a_b = case.m * case.k * dt
+        bb_b = case.k * case.n * dt
+        blk_a = a_b / m_cnt
+        blk_b = bb_b / n_cnt
+        best = None
+        for mg in range(1, m_cnt + 1):
+            for ng in range(1, n_cnt + 1):
+                if mg * blk_a + ng * blk_b > self.s.l2_bytes:
+                    continue
+                gm = (n_cnt + ng - 1) // ng * a_b + (m_cnt + mg - 1) // mg * bb_b
+                if best is None or gm < best:
+                    best = gm
+        self.assertIsNotNone(best)
+        self.assertAlmostEqual(t.gm_read_bytes, case.batch_c * best)
+        self.assertGreaterEqual(t.gm_read_bytes, case.input_bytes)
+        # 最小替换解应严格优于"每块独立落 GM"的保守上界
+        self.assertLessEqual(t.gm_read_bytes,
+                             case.batch_c * (n_cnt * a_b + m_cnt * bb_b))
+
+
+class TestIssue33(unittest.TestCase):
+    """ASW_Basic tile 选择 (issue#33, v1.91 §5.1/§5.2 + 尾轮 v1.5 §2.1 修正口径):
+    BaseM/N = 256 方形 (UnitFlag 单缓冲 L0C 用满 65536 元素), SingleCoreM/N 有界枚举."""
+
+    def setUp(self):
+        from bmm_theory.hardware import ASCEND950PR
+        self.s = ASCEND950PR
+        self.router = BranchRouter()
+
+    def test_base_tile_square_single_buffer(self):
+        # 默认 BaseM=BaseN=256 (非 176 双缓冲口径), L0C 恰用满 (256*256*4=256KB)
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        p = AswBasicBranch().make_plan(mkcase(32, 4096, 4096, 4096))
+        self.assertEqual((p.base_m, p.base_n), (256, 256))
+        self.assertEqual(p.base_m * p.base_n * 4, self.s.l0c_bytes)
+        self.assertEqual(p.base_k, 64)          # 64KB/(2*256*2) = 64
+        self.assertTrue(p.fixpipe_unitflag)     # UnitFlag 单缓冲
+
+    def test_base_tile_exception_small_m(self):
+        # M=128 < 256: BaseM=128 (被迫跟随), BaseN = min(65536/128=512, N)=512
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        p = AswBasicBranch().make_plan(mkcase(64, 128, 4096, 1024))
+        self.assertEqual((p.base_m, p.base_n), (128, 512))
+        self.assertEqual(p.base_k, 32)          # min(64K/(2*128*2)=128, 64K/(2*512*2)=32)
+
+    def test_v191_example_singlecore(self):
+        # v1.91 §5.2 完整实例: B=8 M=N=2048 K=1024 bf16 -> (4,4) 512x512, k_l1=128, r=0
+        case = mkcase(8, 2048, 2048, 1024)
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        p = r["plan"]
+        self.assertEqual((p.single_core_m, p.single_core_n), (512, 512))
+        self.assertEqual((p.m_cnt, p.n_cnt), (4, 4))
+        self.assertEqual(p.k_l1, 128)
+        self.assertEqual(p.tail_block_cnt, 0)   # 8*4*4=128 % 32 = 0 完美整除
+
+    def test_singlecore_not_constant_176(self):
+        # 回归原 bug: SingleCoreM/N 恒为 176 (双缓冲 Base 兜底); 现在自适应
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        p = AswBasicBranch().make_plan(mkcase(8, 4096, 4096, 1024))
+        self.assertNotEqual((p.single_core_m, p.single_core_n), (176, 176))
+        self.assertEqual((p.single_core_m, p.single_core_n), (512, 512))
+        # 且为 Base 的整数倍 (约束 4)
+        self.assertEqual(p.single_core_m % p.base_m, 0)
+        self.assertEqual(p.single_core_n % p.base_n, 0)
+
+    def test_l0c_single_buffer_constraint_pass(self):
+        # ASW 单缓冲: base 256x256 通过约束 (无违规)
+        from bmm_theory.constraints import check_plan_constraints
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        case = mkcase(8, 4096, 4096, 1024)
+        p = AswBasicBranch().make_plan(case)
+        self.assertEqual(check_plan_constraints(case, p), [])
+
+    def test_p1_fallback_to_enum_when_no_split_infeasible(self):
+        # B>=C 时不切分 k_l1 过 dValue 下限失败 -> 强制切分枚举 -> 512x512 (8,8)
+        case = mkcase(128, 4096, 4096, 8192)
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        p = r["plan"]
+        self.assertEqual((p.single_core_m, p.single_core_n), (512, 512))
+        self.assertEqual((p.m_cnt, p.n_cnt), (8, 8))
+
+    def test_issue34_k_l1_decomposition_derived(self):
+        # b32_m16_n8192_k7168: 分解 = Base 16x1024 (例外: M=16<256; N 侧受 L0B 单边
+        # 上限收敛), SingleCore 16x1024 (mCnt=1,nCnt=8), k_l1 = L1 双缓冲反推
+        # ⌊L1/(2·(sM+sN)·dt)⌋16 = ⌊524288/(2·1040·2)⌋16 = 112 (v1.91 §5.2)
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        case = mkcase(32, 16, 8192, 7168)
+        p = AswBasicBranch().make_plan(case)
+        self.assertEqual((p.base_m, p.base_n), (16, 1024))
+        self.assertEqual((p.single_core_m, p.single_core_n), (16, 1024))
+        self.assertEqual(p.k_l1, 112)
+
+    def test_pathological_shape_degraded_warning_always_plan(self):
+        # issue#34: 兜底分支恒出方案 —— 极端形状 (N=8 int8 大K, B 侧 dValue=8B
+        # 物理不可满足) 照常给方案 + 标注效率降级 (warning), 不判违规/不判不可行
+        from bmm_theory.constraints import check_plan_constraints
+        from bmm_theory.evaluator import PlanEvaluator
+        case = mkcase(4096, 2048, 8, 1024, dtype_a="int8", dtype_b="int8")
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        self.assertEqual(check_plan_constraints(case, r["plan"]), [])
+        self.assertIn("效率降级", r["plan"].note)
+        er = PlanEvaluator().evaluate(case, r["plan"])
+        self.assertTrue(er.feasible)
+        self.assertIn("效率降级", er.advice)
+        self.assertGreater(er.timing.t_total, 0.0)
+
+    def test_pathological_n8_hard_floor_fallback(self):
+        # 更极端: N=8 int8 连 128B 硬下限也不满足 -> 仍恒出方案, 效率降级标注
+        from bmm_theory.evaluator import PlanEvaluator
+        case = mkcase(512, 4096, 1, 8192, dtype_a="int8", dtype_b="int8")
+        r = self.router.route(case)
+        self.assertEqual(r["branch"], "ASW_Basic")
+        self.assertIsNotNone(r["plan"])
+        self.assertIn("效率降级", r["plan"].note)
+        er = PlanEvaluator().evaluate(case, r["plan"])
+        self.assertTrue(er.feasible)
+
+
+class TestIssue35(unittest.TestCase):
+    """issue#35: MergeBatch L1 绑定情形 DMA 命令数多计 b0 倍修复 + 分界泛化口径.
+
+    用户 case 家族: B=128, M=1~16, N=128, K=512, bf16. b_core=4, b0=4,
+    合并后 k_l1^m=240/224 < K=512 (L1 绑定区), 真实每核命令数 = 1x3=3 条
+    (< IterBatch 的 4 条), 修复前被多计为 12 条导致仲裁翻错方向.
+    """
+
+    def setUp(self):
+        self.router = BranchRouter()
+        self.mb = MergeBatchBranch()
+        self.ib = IterBatchBranch()
+
+    def test_merged_cmd_count_formula(self):
+        # 每核命令数 = ceil(b_core/b0) * ceil(K/k_l1^m) (K截断时 = b_core/b0)
+        case = mkcase(128, 16, 128, 512)
+        r = self.mb.analyze(case)
+        self.assertTrue(r.capable)
+        p = r.plan
+        expect = -(-p.b_core // p.merge_b0) * (-(-case.k // p.k_l1))
+        self.assertEqual(r.timing.dma_cmd_count, expect)
+        # 本 case: b0=4, k_l1=224 -> 1*3 = 3 条 (修复前 12 条)
+        self.assertEqual((p.merge_b0, p.k_l1), (4, 224))
+        self.assertEqual(r.timing.dma_cmd_count, 3)
+
+    def test_cmd_count_truncated_unchanged(self):
+        # K 截断情形数值不变: cmds = b_core/b0 = IterBatch 的 1/b0
+        case = mkcase(2048, 32, 32, 256)
+        mb = self.mb.analyze(case)
+        ib = self.ib.analyze(case)
+        self.assertGreaterEqual(mb.plan.k_l1, case.k)  # 合并后仍截断
+        self.assertEqual(mb.timing.dma_cmd_count,
+                         -(-mb.plan.b_core // mb.plan.merge_b0))
+        self.assertAlmostEqual(
+            mb.timing.dma_cmd_count / ib.timing.dma_cmd_count,
+            1.0 / mb.plan.merge_b0, places=6)
+
+    def test_boundary_uses_merged_k_l1(self):
+        # 截断判定与 plan.k_l1 口径一致: k_l1^m < K 时不得声称 K截断
+        case = mkcase(128, 1, 128, 512)
+        _, detail = self.mb.beats_iterbatch(case)
+        p = self.mb.make_plan(case)
+        self.assertLess(p.k_l1, case.k)
+        self.assertIn("L1绑定", detail)
+        self.assertNotIn("K截断", detail)
+
+    def test_user_case_family_routing(self):
+        # B=128,M=1~16,N=128,K=512: issue#36 效率模型后, iter A tile=m*1KB 未饱和
+        # (m<16), 合并 tile 4 倍大 -> 效率节省 ~0.61us 恒定, drain 随 M 线性增长;
+        # m<=8 效率节省 > drain -> MergeBatch; m=16 iter tile 恰达 16KB 饱和 ->
+        # 效率打平, drain 0.66us 决定 -> IterBatch
+        expect = {1: "MergeBatch", 2: "MergeBatch", 4: "MergeBatch",
+                  8: "MergeBatch", 16: "IterBatch"}
+        for m, branch in expect.items():
+            r = self.router.route(mkcase(128, m, 128, 512))
+            self.assertEqual(r["branch"], branch, f"m={m}: {r['arbitration']}")
+            self.assertEqual(r["candidates"], {"MergeBatch": True, "IterBatch": True})
+            self.assertEqual(r["self_check_violations"], [])
+
+    def test_arbitration_text_final_winner_consistent(self):
+        # 仲裁文本 [裁决] 位必须是最终胜者 (分界与时延不一致时括注说明)
+        import re
+        r = self.router.route(mkcase(128, 2, 128, 512))
+        m = re.search(r"\[裁决\] (\w+)", r["arbitration"])
+        self.assertIsNotNone(m)
+        self.assertEqual(m.group(1), r["branch"])
+
+
+class TestIssue36(unittest.TestCase):
+    """issue#36: 合并搬移效率建模 (tile=nValue*dValue*dt 放大 b0 倍) + t_cmd_ns=0.
+
+    用户澄清: MergeBatch 合并多 batch 左/右矩阵一起搬移, 单块 tile 放大 ->
+    搬移效率更高, 即便 T_cmd=0 也有效益; 堆叠方向视转置 (A ND 非转置沿
+    M(nValue), B ND 非转置沿 N(dValue)), 乘积口径不变。
+    """
+
+    def test_t_cmd_default_zero(self):
+        self.assertEqual(ASCEND950PR.t_cmd_ns, 0.0)
+        self.assertEqual(ASCEND950PR.t_cmd, 0.0)
+
+    def test_move_eff_curve(self):
+        from bmm_theory.models import move_eff
+        cap = ASCEND950PR.min_tile_size  # 16KB 饱和点
+        self.assertEqual(move_eff(cap, cap), 1.0)          # 饱和
+        self.assertEqual(move_eff(2 * cap, cap), 1.0)      # 超出仍饱和
+        self.assertAlmostEqual(move_eff(cap / 4, cap), 0.25)  # 之下线性
+        self.assertEqual(move_eff(0, cap), 1.0)            # 零值守卫
+
+    def test_merge_eff_gain_beats_iterbatch(self):
+        # (2048,16,64,128): iter A tile=4KB eff=0.25, 合并 A'=16KB eff=1.0
+        # -> 效率节省 ~15.7us >> drain 0.17us, T_cmd=0 下 MergeBatch 大胜
+        case = mkcase(2048, 16, 64, 128)
+        mb = MergeBatchBranch().analyze(case)
+        ib = IterBatchBranch().analyze(case)
+        self.assertTrue(mb.capable and ib.capable)
+        win, detail = MergeBatchBranch().beats_iterbatch(case)
+        self.assertTrue(win, detail)
+        self.assertLess(mb.timing.t_total, ib.timing.t_total)
+
+    def test_gm_bytes_unchanged_by_eff(self):
+        # 效率模型只影响时间列: GM 字节量仍 = V_in (issue#31 口径不被破坏)
+        for shp in [(128, 1, 128, 512), (2048, 16, 64, 128), (128, 64, 64, 512)]:
+            case = mkcase(*shp)
+            for br in (MergeBatchBranch(), IterBatchBranch()):
+                r = br.analyze(case)
+                if r.capable:
+                    self.assertAlmostEqual(r.timing.gm_read_bytes,
+                                           case.input_bytes, places=3)
+
+    def test_iter_small_tile_slower_than_merge(self):
+        # 用户 case m=1: IterBatch A tile=1KB eff=1/16 -> t_mte2_gm 显著高于
+        # MergeBatch (A'=1.9KB eff=0.117), 且两者 GM 字节相同
+        case = mkcase(128, 1, 128, 512)
+        mb = MergeBatchBranch().analyze(case)
+        ib = IterBatchBranch().analyze(case)
+        self.assertGreater(ib.timing.t_mte2_gm, mb.timing.t_mte2_gm)
+        self.assertEqual(mb.timing.gm_read_bytes, ib.timing.gm_read_bytes)
+
+
+class TestIssue37(unittest.TestCase):
+    """issue#37: ASW_Basic evaluate 尾轮残余 drain 闭式化 (方案甲).
+
+    - P1: A0 + r>0 的 drain 由"全 case 三级 max"(≈2x 稳态) 修为 (1-ρ)·T_block;
+    - P2: 周长型 A1b/方案B 补残余 (√ρ−ρ) / (√(n_wave(n_wave−1+ρ))−(n_wave−1+ρ))·T_load;
+    - 面积型 A1b/方案B 与 r=0 残余恒 0 (v1.5 §4.3 严格相等);
+    - 块级三段时延与 _decide_tail 同源 (_block_times, L2 命中口径).
+    """
+
+    def setUp(self):
+        from bmm_theory.branches.asw_basic import AswBasicBranch
+        self.br = AswBasicBranch()
+        self.s = ASCEND950PR
+
+    @staticmethod
+    def _plan(case, sm, sn, m_cnt, n_cnt, strategy, r, n_wave):
+        from bmm_theory.models import ImplPlan
+        return ImplPlan(case_id=case.case_id, branch="ASW_Basic",
+                        used_core_num=32, m_cnt=m_cnt, n_cnt=n_cnt,
+                        single_core_m=sm, single_core_n=sn,
+                        single_core_k=case.k, k_l1=128,
+                        base_m=256, base_n=256, base_k=64,
+                        tail_strategy=strategy, tail_block_cnt=r,
+                        tail_wave_num=n_wave, fixpipe_unitflag=True,
+                        out_dtype_bytes=case.dtype_out_bytes)
+
+    def test_block_times_formula(self):
+        # 块级三段 = v1.5 §2.1 口径 (k_L1 约掉, L2 命中带宽)
+        case = mkcase(3, 1024, 1024, 1024)
+        t_mm, t_mv, t_fx = self.br._block_times(case, 256, 256)
+        self.assertAlmostEqual(t_mm, 2 * 256 * 256 * 1024
+                               / self.s.q_cube("bf16", "bf16"))
+        self.assertAlmostEqual(t_mv, 1024 * (256 + 256) * 2 / self.s.bw_l2_pc)
+        self.assertAlmostEqual(t_fx, 256 * 256 * 2 / self.s.bw_pc)
+
+    def test_a0_drain_is_block_level_residual(self):
+        # P1 修复: A0 + r>0 的 drain = (1-ρ)·T_block (修复前误用全 case 三级
+        # max -> t_total ≈ 2x 稳态)
+        case = mkcase(3, 1024, 1024, 1024)
+        p = self._plan(case, 512, 128, 2, 8, "A0", r=16, n_wave=2)
+        t = self.br.evaluate(case, p)
+        t_block = max(self.br._block_times(case, 512, 128))
+        self.assertAlmostEqual(t.t_drain, (1 - 16 / 32) * t_block)
+        self.assertAlmostEqual(t.t_total, t.t_steady + t.t_drain)
+        self.assertLess(t.t_drain, t.t_steady)   # 不再 ~2x 稳态
+
+    def test_area_dominated_tail_residual_zero(self):
+        # 面积型 (MMAD/FIX 主导): A1b/方案B 残余恒 0 (v1.5 §4.3 严格相等)
+        case = mkcase(3, 1024, 1024, 1024)
+        for strat in ("A1b", "方案B"):
+            p = self._plan(case, 512, 128, 2, 8, strat, r=16, n_wave=2)
+            t = self.br.evaluate(case, p)
+            self.assertEqual(t.t_drain, 0.0, strat)
+            self.assertAlmostEqual(t.t_total, t.t_steady)
+
+    def test_perimeter_a1b_residual(self):
+        # P2: 周长型 (块级 MTE2 主导) + A1b: drain = (√ρ−ρ)·T_load
+        case = mkcase(1, 640, 1408, 4096)
+        t_mm, t_mv, t_fx = self.br._block_times(case, 64, 64)
+        self.assertGreater(t_mv, max(t_mm, t_fx))   # 确认为周长型前提
+        p = self._plan(case, 64, 64, 10, 22, "A1b", r=28, n_wave=7)
+        t = self.br.evaluate(case, p)
+        rho = 28 / 32
+        self.assertAlmostEqual(t.t_drain, (rho ** 0.5 - rho) * t_mv)
+        self.assertAlmostEqual(t.t_total, t.t_steady + t.t_drain)
+
+    def test_perimeter_planb_residual(self):
+        # P2: 周长型 + 方案B: drain = (√(n_wave(n_wave−1+ρ))−(n_wave−1+ρ))·T_load
+        case = mkcase(1, 640, 1408, 4096)
+        p = self._plan(case, 64, 64, 10, 22, "方案B", r=28, n_wave=7)
+        t = self.br.evaluate(case, p)
+        t_mv = self.br._block_times(case, 64, 64)[1]
+        rho, x = 28 / 32, 7 - 1 + 28 / 32
+        self.assertAlmostEqual(t.t_drain, ((7 * x) ** 0.5 - x) * t_mv)
+        self.assertGreater(t.t_drain, 0.0)
+
+    def test_r0_drain_zero(self):
+        case = mkcase(3, 1024, 1024, 1024)
+        p = self._plan(case, 512, 128, 2, 8, "A0", r=0, n_wave=2)
+        self.assertEqual(self.br.evaluate(case, p).t_drain, 0.0)
+
+    def test_make_plan_perimeter_a1b_end_to_end(self):
+        # 端到端 (make_plan 自产方案): 瘦长 case 周长型 + ρ≥ρ_dv -> A1b,
+        # drain 与闭式一致 (_decide_tail 与 evaluate 同源)
+        case = mkcase(33, 16, 8192, 7168)
+        p = self.br.make_plan(case)
+        self.assertGreater(p.tail_block_cnt, 0)
+        self.assertEqual(p.tail_strategy, "A1b", p.note)
+        t_mm, t_mv, t_fx = self.br._block_times(
+            case, p.single_core_m, p.single_core_n)
+        self.assertGreater(t_mv, max(t_mm, t_fx))   # 周长型前提
+        t = self.br.evaluate(case, p)
+        rho = p.tail_block_cnt / 32
+        self.assertAlmostEqual(t.t_drain, (rho ** 0.5 - rho) * t_mv)
+        self.assertAlmostEqual(t.t_total, t.t_steady + t.t_drain)
+
+
+class TestIssue38(unittest.TestCase):
+    """issue#38: 昇腾950 系列多 SKU 硬件规格 (950PR 32/28核 + 950DT 36/32/28核,
+    白皮书表3-1/表4-2), 默认加载调用不变."""
+
+    def test_default_spec_unchanged(self):
+        # 默认加载调用不受影响: ASCEND950PR 各项与 NpuSpec() 关键字默认构造等价
+        from bmm_theory.hardware import NpuSpec, get_spec
+        s = ASCEND950PR
+        self.assertEqual((s.aic_num, s.aiv_num), (32, 64))
+        self.assertEqual(s.bw_gm, 1.6e12)
+        self.assertEqual(s.bw_l2, 5.2e12)               # PR 保持 5.2TB/s
+        self.assertEqual(s.l2_bytes, 128 * 1024 * 1024)
+        self.assertAlmostEqual(s.q16, 432e12 / 32)      # 13.5T (issue#40 Cube-only)
+        self.assertEqual(s.cube_peak_tflops, 432.0)
+        self.assertAlmostEqual(s.r16, 540.0)            # 432e12/(1.6e12/2)
+        self.assertEqual(s.gm_capacity_gb, 128.0)
+        self.assertIs(get_spec(), ASCEND950PR)            # 默认 = 950PR 主bin
+        self.assertIs(get_spec("Ascend950PR"), ASCEND950PR)
+        self.assertEqual(NpuSpec(), ASCEND950PR)          # 无参构造 == 主bin
+
+    def test_spec_registry_five_skus(self):
+        from bmm_theory.hardware import SPECS
+        self.assertEqual(sorted(SPECS), [
+            "Ascend950DT", "Ascend950DT_C28", "Ascend950DT_C32",
+            "Ascend950PR", "Ascend950PR_C28"])
+        for s in SPECS.values():
+            # 共架构 + Cube-only 口径 (issue#40): 单核 Cube 全系列精确 13.5T
+            # (432/32=378/28=486/36); AIV 恒为 AIC 两倍
+            self.assertAlmostEqual(s.q16, 13.5e12, places=6)
+            self.assertEqual(s.aiv_num, 2 * s.aic_num)
+            self.assertEqual(s.l1_bytes, 512 * 1024)      # 表4-2 各档一致
+            self.assertEqual(s.l0c_bytes, 256 * 1024)
+
+    def test_950dt_specs(self):
+        from bmm_theory.hardware import (ASCEND950DT, ASCEND950DT_C28,
+                                         ASCEND950DT_C32)
+        self.assertEqual((ASCEND950DT.aic_num, ASCEND950DT.aiv_num), (36, 72))
+        self.assertEqual(ASCEND950DT.bw_gm, 4.0e12)       # 4TB/s HBM
+        self.assertEqual(ASCEND950DT.l2_bytes, 128 * 1024 * 1024)
+        self.assertEqual(ASCEND950DT.gm_capacity_gb, 144.0)
+        self.assertEqual(ASCEND950DT.cube_peak_tflops, 486.0)   # Cube-only (issue#40)
+        # L2 带宽: DT 三档 7.5TB/s 读写各自独享 (issue#39 用户澄清)
+        for dt_spec in (ASCEND950DT, ASCEND950DT_C32, ASCEND950DT_C28):
+            self.assertEqual(dt_spec.bw_l2, 7.5e12)
+        self.assertAlmostEqual(ASCEND950DT.bw_l2_pc, 7.5e12 / 36)
+        self.assertEqual((ASCEND950DT_C32.aic_num,
+                          ASCEND950DT_C32.cube_peak_tflops), (32, 432.0))
+        self.assertEqual((ASCEND950DT_C28.aic_num,
+                          ASCEND950DT_C28.cube_peak_tflops), (28, 378.0))
+        self.assertEqual(ASCEND950DT_C28.gm_capacity_gb, 96.0)
+        # 派生量: 单核 GM 份额 4TB/36; r16 平衡点随带宽升至 4TB/s 而减半
+        self.assertAlmostEqual(ASCEND950DT.bw_pc, 4.0e12 / 36)
+        self.assertAlmostEqual(ASCEND950DT.r16, 486e12 / (4.0e12 / 2))      # 243
+        self.assertAlmostEqual(ASCEND950DT_C32.r16, 432e12 / (4.0e12 / 2))  # 216
+        # AIV 白皮书交叉验证: fp32 通量 72 核 x 128 lane x 1.65GHz x 2 ≈ 30T
+        self.assertAlmostEqual(ASCEND950DT.aiv_elem_rate_fp32 * 2 / 1e12,
+                               30.0, places=0)
+
+    def test_950pr_c28(self):
+        from bmm_theory.hardware import ASCEND950PR_C28
+        self.assertEqual((ASCEND950PR_C28.aic_num,
+                          ASCEND950PR_C28.aiv_num), (28, 56))
+        self.assertEqual(ASCEND950PR_C28.bw_gm, 1.4e12)
+        self.assertEqual(ASCEND950PR_C28.bw_l2, 5.2e12)  # PR 系列保持 5.2TB/s
+        self.assertEqual(ASCEND950PR_C28.l2_bytes, 112 * 1024 * 1024)
+        self.assertEqual(ASCEND950PR_C28.gm_capacity_gb, 112.0)
+
+    def test_get_spec_unknown_raises(self):
+        from bmm_theory.hardware import get_spec
+        with self.assertRaises(KeyError):
+            get_spec("Ascend910")
+
+    def test_dt_router_smoke(self):
+        # 换芯片只换 spec: DT 路由/时延正常; 同一 case GM 段 4TB/s 快于 1.6TB/s
+        from bmm_theory.hardware import ASCEND950DT
+        case = mkcase(8, 4096, 4096, 1024)
+        r_dt = BranchRouter(ASCEND950DT).route(case)
+        r_pr = BranchRouter().route(case)
+        self.assertGreater(r_dt["timing"].t_total, 0.0)
+        self.assertLess(r_dt["timing"].t_mte2_gm, r_pr["timing"].t_mte2_gm)
+    def test_dt_l2_bw_effective_in_fixpipe(self):
+        # issue#39: S_A (整case驻留L2) 下 DT 的 fixpipe 走 7.5TB/s 写口
+        from bmm_theory.hardware import ASCEND950DT
+        case = mkcase(2, 4096, 4096, 512)   # 输入+输出 < 128MB -> S_A
+        self.assertLess(case.input_bytes + case.output_bytes,
+                        ASCEND950DT.l2_bytes)
+        r = BranchRouter(ASCEND950DT).route(case)
+        self.assertAlmostEqual(r["timing"].t_fixpipe,
+                               case.output_bytes / 7.5e12)
+
 
 if __name__ == "__main__":
     unittest.main()