Commit Graph

10 Commits

Author SHA1 Message Date
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
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
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
87b7be551a Update BMM_Theory: tests/test_branches.py (fix review issues #4-#10) 2026-09-03 11:34:36 +00:00
5ab04bab11 Update BMM_Theory: tests/test_branches.py 2026-09-03 09:25:58 +00:00
5a6bf2137b Add BMM_Theory: tests/test_branches.py 2026-09-03 08:10:38 +00:00