Commit Graph

124 Commits

Author SHA1 Message Date
18f59599e7 Fix #31/#32: 切B类GM每字节恰一次=V_in(去K切分整段上取) / ASW场景升级(单侧全驻留+对侧滑窗->S_B, S_C最小替换2D分组+窗口L2计账) / 06文档§3+Step5与docs/05同步
#31 IterBatch/MergeBatch: K切分各(batch,K段)互不重叠+驻留侧每batch一次+末段按实际剩余
    -> GM读取量=V_in(与L2容量无关), GM数据时延=V_in/W_GM; n_K仅决定DMA命令数(T_cmd)
    b64_m16_n256_k512 形态c: GM 25.07MB->17.83MB=V_in; 回归: 形态c/d非整除+L1绑定三类断言
#32 ASW: (1)S_B扩展单侧全驻留+对侧滑窗(a_b/b_b+2*对侧单块<=L2) -> GM=V_in, 6个场景C行回落S_B;
    (2)S_C在整L2容量约束下搜索最小GM=ceil(n_cnt/n_grp)a_b+ceil(m_cnt/m_grp)b_b(取代L2/2对半预算),
    并计组内窗口L2流量((n_cnt-ceil)a_b+(m_cnt-ceil)b_b), 与S_B'驻留命中走L2'口径一致;
    b8_m131072_n8192_k8192 GM倍率4.76x->3.88x(物理下界~3.9x, 双侧均超L2)
    大方形K行(如b128_m8192_n8192_k7168)由MMAD 253ms->MTE2(L2口)295ms: 共享块重复读1.38TB
    经L2读口5.2TB/s, 如实计账(原C窗口流量零计低估)
- docs/06 §3与Step5重写为S_A/S_B/S_C+两段链口径(旧r_in单段/除B/对半预算口径废弃)
- docs/05 R5/R6/§3.1/§4.1/§4.3/§5与docs/01、02(01_MergeBatch/02_IterBatch GM口径附注)同步
- tests 54/54; 压力seed7/6000+seed2024/4000: 0违规/0占位/0NaN/0GM<V_in; examples重生成0diff
2026-09-04 17:12:47 +08:00
b9e07edc1d Fix #27-#30: dtype感知算力(Cube/AIV速率表) / GM首读下限与整芯片字节列口径+设计文档 / Fixpipe输出落点R4(整case驻留L2否则直写GM) / MergeBatch Cube公式复核注释
- docs/05_L2驻留GM读写与dtype算力口径_设计分析.md: R1-R6公理、S_A/S_B/S_C场景、输出落点R4、dtype速率表(白皮书出处+待标定假设)、逐分支GM/L2归属表 (issue#29/#30 先文档后代码)
- hardware: CUBE_DTYPE_FACTOR(f16/bf16=1, fp8=2x, fp4=4x, fp32=1/2假设) + AIV_DTYPE_FACTOR + q_cube/aiv_elem_rate (issue#28)
- 全分支 t_mmad/t_comp/drain/尾轮主导项/θ_c/R16 语义按输入dtype取算力; 混精度取慢侧; StreamK归约保持fp32(AIV fp32部分和)
- Fixpipe输出落点R4: to_l2 <=> V_in+V_out(+workspace)<=L2; 否则直写GM计入共享总线; ASW场景改S_A整case全驻留(原单batch驻留判定漏计整case输出累积逐出)
- 字节列统一整芯片口径(gm/l2/fix/cube_flops), dma_cmd_count注明单核; GM>=V_in不变量入测试; MergeBatch每步flops=2(b0M)(b0N)K公式注释显式化(#27复核与CSV一致无数值改动)
- tests 39->49 全过; 双压力seed7/6000+seed2024/4000: 0违规/0占位/0NaN/0GM<输入; examples三件套重生成且可复现0diff
2026-09-04 16:26:10 +08:00
4843053ad3 Fix issues #23-#25 (+#26 标注): GM读写共享总线累加计时 / ASW-切M/N 共享块 GM首读1次+L2重复读(n-1)次、场景按单batch判定、分组预算不再除B / StreamK 按plan实际tile芯片口径评估 / 降核线性带宽假设文档标注 2026-09-04 11:43:23 +08:00
f4b23d9f05 Expand cases_demo to 44 cases (B 4..128 x M 1..131072 x N/K 128..8192 incl 7168) + regenerate examples outputs 2026-09-04 11:09:49 +08:00
5051e00fa0 T_cmd=0 仲裁改为策略优先 MergeBatch: K截断即优先(指令/命令次数省b0倍,未量化收益), L1绑定恒劣; 路由层以策略覆盖时延模型差额并显式标注 2026-09-04 10:47:02 +08:00
a071d87831 Fix T_cmd=0 robustness: beats_iterbatch 阈值除零防护 (T_cmd<=0 判 MergeBatch 不胜出, 无命令节省) + 3 条回归测试 2026-09-04 10:31:53 +08:00
9afe6eec02 Fix review issues #17-#22: 恢复 #11/#12/#14 (StreamK fixpipe 单次计账/K=1 AIV单缓冲/advice) + 占位方案不可评估 + 转置 dValue 判据三处同源(form c 双缓冲适配修复) + 恢复 #13/#15 回归测试 + 清理临时 csv/.gitignore + 文档同步 2026-09-03 21:10:05 +08:00
f5a0b6fe81 Update BMM_Theory: tests/test_branches.py 2026-09-03 12:44:22 +00:00
08bf74d8b8 Update BMM_Theory: examples/result_recommend.csv 2026-09-03 12:44:20 +00:00
2d9d9f3116 Update BMM_Theory: examples/result_evaluate.csv 2026-09-03 12:44:18 +00:00
6d5e452fbc Update BMM_Theory: docs/02_分支理论/04_特殊分支.md 2026-09-03 12:44:10 +00:00
c05cbeafad Update BMM_Theory: bmm_theory/branches/stream_k.py 2026-09-03 12:43:55 +00:00
49397260f6 Update BMM_Theory: bmm_theory/branches/special.py 2026-09-03 12:43:54 +00:00
4f90f76db9 Update BMM_Theory: bmm_theory/branches/merge_batch.py 2026-09-03 12:43:53 +00:00
af0a90bc25 Update BMM_Theory: bmm_theory/branches/iter_batch.py 2026-09-03 12:43:52 +00:00
a7356fec99 Update BMM_Theory: bmm_theory/evaluator.py 2026-09-03 12:43:38 +00:00
db8399b62c Update models.py: fp4 support 2026-09-03 12:42:35 +00:00
c2028a41ed Add BMM_Theory: t_p.csv 2026-09-03 12:37:18 +00:00
f3d6bc7003 Add BMM_Theory: t_out.csv 2026-09-03 12:37:16 +00:00
ebf77d9b48 Add BMM_Theory: t.csv 2026-09-03 12:37:14 +00:00
98223e0e12 Add BMM_Theory: p2.csv 2026-09-03 12:37:12 +00:00
32febc1f62 Add BMM_Theory: bug_test.csv 2026-09-03 12:37:11 +00:00
ed084a4be0 Add BMM_Theory: bug2.csv 2026-09-03 12:37:10 +00:00
340ae2c0d4 Add BMM_Theory: bug1_p.csv 2026-09-03 12:37:09 +00:00
b4003d8246 Add BMM_Theory: bug1_out.csv 2026-09-03 12:37:08 +00:00
7309215720 Add BMM_Theory: bug1.csv 2026-09-03 12:37:06 +00:00
d8457be420 Remove mistakenly pushed temp file: t_p.csv 2026-09-03 12:36:42 +00:00
7f318199ab Remove mistakenly pushed temp file: t_out.csv 2026-09-03 12:36:41 +00:00
60d37a83d2 Remove mistakenly pushed temp file: t.csv 2026-09-03 12:36:40 +00:00
d4a249e0f8 Remove mistakenly pushed temp file: p2.csv 2026-09-03 12:36:38 +00:00
9c805659c4 Remove mistakenly pushed temp file: bug_test.csv 2026-09-03 12:36:38 +00:00
0198f60f4e Remove mistakenly pushed temp file: bug2.csv 2026-09-03 12:36:37 +00:00
5fcfb38adf Remove mistakenly pushed temp file: bug1_p.csv 2026-09-03 12:36:36 +00:00
7644ac53f0 Remove mistakenly pushed temp file: bug1_out.csv 2026-09-03 12:36:35 +00:00
0440342dd9 Remove mistakenly pushed temp file: bug1.csv 2026-09-03 12:36:34 +00:00
88b1e4946c Add BMM_Theory: docs/04_差异对照_bmmv3_vs_BMM_Theory.md 2026-09-03 12:35:49 +00:00
f4e8eff040 Add BMM_Theory: t_p.csv 2026-09-03 12:35:30 +00:00
3090d76a92 Add BMM_Theory: t_out.csv 2026-09-03 12:35:28 +00:00
b8247af9ee Add BMM_Theory: t.csv 2026-09-03 12:35:27 +00:00
72769f4b90 Add BMM_Theory: p2.csv 2026-09-03 12:35:24 +00:00
0f474b3b65 Add BMM_Theory: bug_test.csv 2026-09-03 12:35:22 +00:00
69d09e1f73 Add BMM_Theory: bug2.csv 2026-09-03 12:35:21 +00:00
23780c5c67 Add BMM_Theory: bug1_p.csv 2026-09-03 12:35:20 +00:00
ae3ef6f7e4 Add BMM_Theory: bug1_out.csv 2026-09-03 12:35:19 +00:00
d0a6df0f8e Add BMM_Theory: bug1.csv 2026-09-03 12:35:18 +00:00
623f20259e Add BMM_Theory: docs/03_测评报告/ 软件复评报告 v2.0 (整改后复评 + issue #11-#16 闭环记录) 2026-09-03 20:06:55 +08:00
3a6c952301 Remove mistakenly staged temp files (t.csv/smoke.txt) 2026-09-03 20:05:21 +08:00
4bedf0a109 Fix review issues #11-#16: StreamK fixpipe 单次计账 / K=1 AIV单缓冲方案 / MergeBatch b0 L0A/L0B 上限+路由可行回退 / advice-StreamK / 输入校验 / .gitignore+死代码清理 2026-09-03 20:05:14 +08:00
99a25a6b42 Remove mistakenly pushed temp file: t_p.csv 2026-09-03 11:35:03 +00:00
2d06aa939f Remove mistakenly pushed temp file: t_out.csv 2026-09-03 11:35:02 +00:00