admin

admin commented on issue admin/matmul-analysis#36

2026-09-07 13:11:10 +00:00

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

已实现, commit b0b48b9 已推送 main。

落地内容

搬移效率模型 (models.move_eff / gm_move_time): eff(tile) = min(1, tile/min_TileSize), 单命令 tile = nValue×dValue×input_dtype 达 16KB…

admin closed issue admin/matmul-analysis#36

2026-09-07 13:09:51 +00:00

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

admin pushed to main at admin/matmul-analysis

2026-09-07 13:09:51 +00:00

b0b48b9073 Fix #36: MergeBatch 合并搬移效率收益建模 (move_eff) + t_cmd_ns 置 0

admin opened issue admin/matmul-analysis#36

2026-09-07 12:48:11 +00:00

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

admin commented on issue admin/matmul-analysis#35

2026-09-07 12:32:39 +00:00

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

已修复, commit fdd3c88 已推送 main。

修复内容

evaluate 命令数 (缺陷1): 每核 DMA 命令数 = ⌈b_core/b0⌉ × ⌈K/k_l1^m⌉。K 截断时退化为 b_core/b0 (数值不变); L1 绑定情形消除 b0 倍多计…

admin closed issue admin/matmul-analysis#35

2026-09-07 12:31:54 +00:00

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

admin pushed to main at admin/matmul-analysis

2026-09-07 12:31:54 +00:00

fdd3c8883c Fix #35: MergeBatch L1绑定情形 DMA 命令数多计 b0 倍修复 + 分界泛化口径

admin opened issue admin/matmul-analysis#35

2026-09-07 12:20:53 +00:00

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

admin commented on issue admin/matmul-analysis#34

2026-09-07 08:25:43 +00:00

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

修复 commit: 0cd47f9

先答第 1 问 (分解推导链): b32_m16_n8192_k7168 的 k_l1=112 分解过程

BaseM/BaseN: M=16<256 -> 例外跟随 M: BaseM=16; 另一侧 BaseN = min(L0C 面积余量 4096, N=8192, L0B…

admin closed issue admin/matmul-analysis#34

2026-09-07 08:24:49 +00:00

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

admin pushed to main at admin/matmul-analysis

2026-09-07 08:24:49 +00:00

0cd47f93cb Fix #34: ASW_Basic 兜底恒出方案; 搬移效率下限不满足降级为 warning (不判违规)

admin opened issue admin/matmul-analysis#34

2026-09-07 08:15:08 +00:00

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

admin commented on issue admin/matmul-analysis#33

2026-09-07 07:50:48 +00:00

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

修复 commit: 05ca91e

落地 (对照《ASW_Basic分支分析 v1.91》§5.1/§5.2 + 《BMM尾轮处理策略对比分析 v1.5》§2.1):

BaseM/BaseN: UnitFlag 单缓冲方形 256x256 (L0C/4B=65536 元素用满, 替代旧双缓冲…

admin closed issue admin/matmul-analysis#33

2026-09-07 07:49:27 +00:00

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

admin pushed to main at admin/matmul-analysis

2026-09-07 07:49:27 +00:00

05ca91e291 Fix #33: ASW_Basic tile 选择重写 (v1.91 §5.1/§5.2 + 尾轮 v1.5 §2.1)

admin opened issue admin/matmul-analysis#33

2026-09-07 07:11:16 +00:00

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

admin commented on issue admin/matmul-analysis#32

2026-09-04 09:13:26 +00:00

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

修复 commit: 18f5959

落地:

场景升级 (_l2_scene, make_plan/evaluate 同源): S_A 整case全驻留 -> S_B 单batch可驻留 (a_b+b_b<=L2, 或单侧全驻留+对侧滑窗: b_b+2block_a<=L2 / a_b+2block_b<=L2) ->…

admin commented on issue admin/matmul-analysis#31

2026-09-04 09:13:25 +00:00

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

修复 commit: 18f5959

落地 (issue 正文方案):

IterBatch/MergeBatch evaluate: GM 数据量 = V_in (每字节恰从 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, 不再按整段上取), GM…

admin closed issue admin/matmul-analysis#32

2026-09-04 09:13:06 +00:00

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

admin closed issue admin/matmul-analysis#31

2026-09-04 09:13:05 +00:00

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