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

This commit is contained in:
2026-09-03 21:10:05 +08:00
parent f5a0b6fe81
commit 9afe6eec02
22 changed files with 219 additions and 76 deletions

View File

@@ -24,11 +24,14 @@ 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 触发 AIV 通路需 B >= 2*AIV核数 且单 batch 输入输出能驻留 UB
c2 = case.batch_c >= 2 * self.spec.aiv_num
# K=1 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
mode = "UB乒乓" if pingpong else "AIV单缓冲(逐batch串行, B<2*AIV)"
checks.append(ConditionCheck(
"2_K=1的AIV触发: B >= 2*AIV核数 (开UB乒乓)",
c2, f"B={case.batch_c} vs {2*self.spec.aiv_num}"))
"2_K=1的AIV通路: 恒可用 (B>=128 开UB乒乓, 否则单缓冲)",
True, f"B={b}, 模式={mode}"))
return checks
# ------------------------------------------------------------------
@@ -36,11 +39,15 @@ class SpecialBranch(Branch):
s = self.spec
if case.k == 0:
sub = "K=0纯写值"
mode = ""
note = "无任何计算, C=bias 或 0, 纯 AIV 写值; 按行均分到 AIV 核"
else:
sub = "K=1逐元素乘"
note = ("退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; "
"AIV 通路 GM->UB->Mul->GM, UB 乒乓")
pingpong = case.batch_c >= 2 * s.aiv_num
mode = "UB乒乓" if pingpong else "AIV单缓冲"
note = (f"退化为 C=A⊙B 无累加深度, Cube 16x16x16 粒度浪费 15/16; "
f"走 AIV 通路 GM->UB->Mul->GM, {mode} "
f"({'B>=2*AIV 双batch乒乓流水' if pingpong else 'B<2*AIV 逐batch单缓冲串行'})")
return ImplPlan(
case_id=case.case_id, branch=self.name,
used_core_num=s.aiv_num, # 用 AIV 核
@@ -48,7 +55,8 @@ class SpecialBranch(Branch):
core_map="AIV 核间按行均分 (无 Cube tile 概念)",
b_core=0, merge_b0=1,
single_core_m=0, single_core_n=0, single_core_k=case.k,
k_l1=0, b_l1=1, l1_form="UB驻留(AIV)",
k_l1=0, b_l1=1,
l1_form="UB驻留(AIV)" if case.k == 0 else "UB驻留(AIV) " + mode,
base_m=0, base_n=0, base_k=0,
l2_policy_in="allocate", l2_policy_out="direct_gm",
swizzle_w=0, workspace_bytes=0,