matmul-analysis

MergeBatch 合并搬移效率收益建模 (tile=nValue*dValue*dt 放大 b0 倍) + t_cmd_ns 置 0

#36 by admin was closed 2026-09-07 13:09:51 +00:00

1

MergeBatch L1 绑定情形 DMA 命令数多计 b0 倍, 且 beats_iterbatch 截断判定与方案口径不一致

#35 by admin was closed 2026-09-07 12:31:54 +00:00

1

34. ASW_Basic 兜底必须恒出方案: 搬移效率下限不满足应降级标注(warning), 不是违规/不可行

#34 by admin was closed 2026-09-07 08:24:49 +00:00

1

33. ASW_Basic tile 选择僵化恒为176: BaseM/N 应用 L0C 单缓冲方形256, SingleCoreM/N 应用 v1.91 §5.2 有界枚举最小搬入

#33 by admin was closed 2026-09-07 07:49:27 +00:00

1

32. ASW 场景C GM 超输入成倍放大: 未识别单侧全驻留调度+分组非最小替换, 且06文档已漂移需同步

#32 by admin was closed 2026-09-04 09:13:06 +00:00

1

31. IterBatch/MergeBatch GM 读取量超输入总量: K切分按整段上取+驻留侧重复计, 应每字节恰读一次=V_in

#31 by admin was closed 2026-09-04 09:13:05 +00:00

1

30. Fixpipe输出落点: 默认写L2, 整case容量不足才直写GM(输入优先驻留)

#30 by admin was closed 2026-09-04 08:26:36 +00:00

1

29. GM读取量<输入数据量(列口径)+缺L2驻留工作集统一模型: 先设计文档后代码

#29 by admin was closed 2026-09-04 08:26:36 +00:00

1

28. 时延估算未按输入dtype取算力: Cube恒用BF16/AIV恒用fp32, 需dtype感知速率表

#28 by admin was closed 2026-09-04 08:26:35 +00:00

1

27. MergeBatch Cube 时延公式复核: 与建议公式一致(附证明), 字节列口径随#29统一

#27 by admin was closed 2026-09-04 08:26:34 +00:00

1

[P3] 降核/低核数带宽线性假设需标注待标定（min_core_num=0.8C 处 90%+ 带宽利用率的 KB 锚点）

#26 by admin was closed 2026-09-04 03:43:40 +00:00

1

[P2] StreamK evaluate 使用假想 65536 元素 L0C 方形 tile，未按 plan 实际 single_core_m×n 计量（瘦长 case 搬入低估 ~16x）

#25 by admin was closed 2026-09-04 03:43:39 +00:00

1

[P1] 切M/N(ASW) 的共享块重复读未按 L2 带宽计时；场景判定与分组预算 D 错误除以总 batch 数

#24 by admin was closed 2026-09-04 03:43:39 +00:00

1

[P1] GM 读写共享总线未建模：MTE2 读与 Fixpipe 直写 GM 并发时需按 (读+写)/1.6TB/s 累加计时

#23 by admin was closed 2026-09-04 03:43:39 +00:00

1

[P3] 文档与代码状态失同步：04_差异对照版本声明称"含 issue#11-#16 修复"与回退后的代码矛盾

#22 by admin was closed 2026-09-03 13:10:26 +00:00

1

[P3] 仓库卫生复发：9 个调试临时 csv 再次被提交入库（bug1*.csv/t*.csv/p2.csv），.gitignore 未覆盖

#21 by admin was closed 2026-09-03 13:10:26 +00:00

1

[P3] #13/#15 的回归测试在第三轮被删除（修复代码仍在，失去保护）

#20 by admin was closed 2026-09-03 13:10:26 +00:00

1

[P3] 转置感知 dValue 建模当前不生效：c/d 生成守卫与条件4/约束器口径不一，且新增转置测试全部空转

#19 by admin was closed 2026-09-03 13:10:25 +00:00

1

[P2] 占位方案(used_core_num=0)可被 evaluate 判为可行并给出正常时延：recommend"无方案"与 evaluate"可行"矛盾

#18 by admin was closed 2026-09-03 13:10:25 +00:00

1

[P1] 第三轮整改回退了 #11/#12/#14 的修复：StreamK fixpipe 双重计账(55us)、K=1 空洞(49.4%)、advice 误导，需恢复或给出替代论证

#17 by admin was closed 2026-09-03 13:10:25 +00:00

1