@kernaut/kernaut-resources 0.1.17-beta.1 → 0.1.18

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cannbench-public-score-measurement.md

@@ -1,99 +0,0 @@
----
-id: exp-ascend-cannbench-public-score-measurement
-title: 'CANNBench public evaluation: separate elapsed improvements, signed score effects and scoring poles'
-type: experience
-vendor: ascend
-probe_slug: cannbench-public-score-measurement
-measured_on:
-  device: Historical Ascend910B2C local runs and/or Ascend910C official runs; see the source-bound scope
-source:
-- path: sources/experience/cannbench-optimization-history-20261020.md
-architectures:
-- ascend910b
-languages:
-- ascendc
-tags:
-- ascendc
-aliases:
-- CANNBench
-- cannbench
-- 公开集
-- public
-- standard
-- HAP
-- 评分
-- AA
-- ABBA
-- pole
-- '140'
-- exp-ascend-cannbench-20261020-measurement-public-scores
-- wiki/ascend/experience/cannbench-20261020-measurement-public-scores.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: Historical official task1.1.2 and board aggregate1.1; performance anchors are frozen per case. Local910B elapsed and formal910C elapsed remain different measurements.
-knowledge_id: knowledge-ascend-measurement-public-scores
-knowledge_id_source: agent-semantic-link
-knowledge_scope: conditional-general
-knowledge_page_role: theme-candidate
-verified_operators: []
-observed_operators:
-- CANNBench measurement
-condition_ids:
-- condition-cannbench-20261020-measurement-public-scores
-sources:
-- exp-source-cannbench-optimization-20261020
-source_refs:
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/dq-report.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/dq-comparison.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/campaign-phase2.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/campaign-gelu.md
-related:
-- exp-ascend-cb04-timing-and-ablation
-- exp-ascend-dynamic-quant-short-rows
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
----
-
-# CANNBench public evaluation: separate elapsed improvements, signed score effects and scoring poles
-
-This refresh prioritizes public-set Kernaut evaluation. Historical hidden observations below are separate background; a hidden failure alone does not disqualify a public20/20 baseline. Benchmark pass, broader contract coverage and source integrity remain distinct.
-
-## Conditions and evidence status
-
-Historical official task1.1.2 and board aggregate1.1; performance anchors are frozen per case. Local910B elapsed and formal910C elapsed remain different measurements.
-
-Historical source-reported evidence, reviewed on2026-10-20; no independent device reproduction in this import. `cann_version: unknown` refers to the combined record: individual local SDKs are stated where captured, and the official runner release is not assumed. Verified-operator coverage is left empty rather than inferred from historical scores.
-
-## Before and change
-
-**Before:** Summing favorable target changes, inferring gains from board totals or accepting a score spike could obscure regressions, control drift and a non-monotonic scoring denominator.
-
-**Change:** For all-public-pass panels replay the actual evaluator: score=50+(50/N)*sum((B−H)/(E+B−2H)), N=20. Preserve failed-case logic from the real evaluator when any case fails. Compare complete same-source public panels with interleaved parent/candidate and independent AA; retain each signed target and control. Keep raw score, elapsed geometric/sum metrics and conditional ratio-transfer forecasts separate.
-
-**Mechanism:** The nonlinear score weights vary per case and can have a pole. A score is neither a uniform latency metric nor a causal attribution. AA dispersion bounds resolution for that panel, not a universal statistical interval.
-
-## Recorded results
-
-DynamicQuant c5 B176.18/H139.81 gives denominator E−103.44 us; denominator<=0 yields None in the original implementation. An unchanged DQ301 scored140.0330,82.5141,90.1880 in final interleaving. A158.83 spike remeasured93.40. C05 stable~83 still delivered short-row latency improvements. Do not treat the spike as target attainment.
-
-## Boundaries, negative evidence and reuse
-
-Do not clip HAP, change anchors, add waits or choose near-pole samples. t_hw is a task metadata reference, not universally a measured vendor kernel or physical floor. Opponents actual code/timing were unavailable; these observations do not explain their scores. Public-only Kernaut testing needs no hidden run, request or hidden-success gate.
-
-Use the public parent for a new complete public comparison, keep all20 signed rows and AA controls, and bind the final build, source, task and checker. Do not require or launch a hidden evaluation for this public-only workflow.
-
-## Evidence
-
-- [dq-report](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/dq-report.md): exact input SHA256 and original line ranges.
-- [dq-comparison](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/dq-comparison.md): exact input SHA256 and original line ranges.
-- [campaign-phase2](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/campaign-phase2.md): exact input SHA256 and original line ranges.
-- [campaign-gelu](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/campaign-gelu.md): exact input SHA256 and original line ranges.
-
-Prior knowledge: `exp-ascend-cb04-timing-and-ablation`, `exp-ascend-dynamic-quant-short-rows`.
-
-[Public operator index](cannbench-public-experiences.md) · [source record](../../../sources/experience/cannbench-optimization-history-20261020.md)

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-dsq-padded-row-layout.md

@@ -1,46 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-dsq-padded-row-layout",
-  "title": "DequantSwigluQuant：多行DMA的真实pitch与Gather例外",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-dsq-padded-row-layout",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "DequantSwigluQuant：多行DMA的真实pitch与Gather例外",
-    "DSQ：自然行距只约束实际多行DMA路径"
-  ]
-}
----
-
-# DequantSwigluQuant：多行DMA的真实pitch与Gather例外
-
-任务x=[rows,2H]；int32输入需要weight_scale[1,2H]和activation_scale[rows]，half/bf16两者为None。activate_left决定SiLU(A)*B或SiLU(B)*A，可选quant_scale[1,H]，两个输出分别为int8 y和fp32 scale。主要fp32计算，scale=max(abs)/127后clamp_min(1e-12)。
-
-padded_general的逐行DataCopyPad按输入dtype计算dma_stride=ceil(H*sizeof(InputType)/32)*32/sizeof(InputType)。fp32系数另有cstride，half/int8视图需32B staging。bf16 H2049的2056/rpt2错误、2064/rpt2通过，2112曾507035；逻辑H对齐不替代实际pitch。
-
-当前dsq_padded_rows用184KiB预算，coeff_rows=(int32?2:0)+(has_quant_scale?1:0)，rpt≤255；dsq_padded_ok要求int32 rpt≥4、其它≥2，dispatch还要rows≥2*blocks且前序专用路线未命中。Gather计划以初始rpt≥8且ceil(rows/blocks)≥8准入，可自行roundup(H,8)，因此“任何行距只能自然对齐”错误。容量公式、use_gather重算、LAUNCH_GENERAL先后必须读完整源码。
-
-旧单侧候选以192KiB估算并删除tiled_general，H≥4096实际213–327KiB且丢回退；不能把旧192KiB与当前184KiB混成一个常量。NaN/Inf、零最大值、qs有无和两个输出的运行验证仍需在目标域完成。
-
-## 来源与复核入口
-
-- CB03-07：[kop:cannbench/design/kernel/dequant_swiglu_quant.md:42–104](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/dequant_swiglu_quant.md#L42-L104)；[kop:cannbench/design/kernel/dequant_swiglu_quant.md:1300–1363](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/dequant_swiglu_quant.md#L1300-L1363)；[kop:cannbench/tasks_official_110/level3/dequant_swiglu_quant/golden.py:17–90](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/tasks_official_110/level3/dequant_swiglu_quant/golden.py#L17-L90)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-gs3d-transpose-channel-layout.md

@@ -1,50 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-gs3d-transpose-channel-layout",
-  "title": "GridSampler3D：repeat尾块、padding与合并路由",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-gs3d-transpose-channel-layout",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "GridSampler3D：repeat尾块、padding与合并路由",
-    "GS3D：repeat==1 的修复包含窄通道与较宽尾块",
-    "GS3D：padding、align与NaN修复须分开归因",
-    "GS3D：合并后的路由才是正确性验收对象"
-  ]
-}
----
-
-# GridSampler3D：repeat尾块、padding与合并路由
-
-历史fp16 channel-last T_TransBlockRun的run=min(nBlocks-rb,8)，b16 repeat=run。在repeat==1时同时把dstRep/srcRep置0：不仅C≤16，C129..144的8+1尾块也命中；NarrowCols另一方向的kBlocks%8==1同样要查，不能只按145..160排除。b32 repeat=2*run≥2不属于这条条件。当前快照同包仍有dstRepEff/srcRepEff guard。
-
-NaN投毒曾使C1/2/4/8/12/16错误99.1–100%；−12345投毒时unwritten=0、matching=0，支持写错而非简单漏写。退stock转置和fix1是不同对照。repeat约束是910B2C行为反推，作者未核官方API出处；不升级成所有重载/SoC/工具链规范。
-
-坐标语义另核zeros/border/reflection、bilinear/nearest、align_corners及x/grid特殊值。W4、align=true、border、c1.2时u3.3，目标34而旧值23.8=0.7×34，支持跳过border。倒数代除法的稀有reflection floor翻转是另一机制；巨大有限1e30仍有残余。4形状×2dtype×9值组合×12语义应864，原称576，缺真实生成清单不能宣称全因子覆盖。
-
-CL的合并路由把C256、oe401408送入旧oe≥2^23扫描未覆盖的区域。vector_fallback_cl=!mono_gather_available && C≥8 && oe≥2^12 && input_points≤128*output_points；完整use_channel_last还有其它分支。960配置的非确定数从fix1的0到vH67/vH3 92，不能因签名近似就归因某个lane或唯一race；缺路由×kernel完整2×2控制。普通随机/原子归约也不能套此确定性要求。
-
-## 来源与复核入口
-
-- CB03-04：[kop:cannbench/design/opt11_gs3d/lane_HA.md:82–135](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt11_gs3d/lane_HA.md#L82-L135)；[kop:cannbench/design/opt11_gs3d/lane_HA.md:214–235](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt11_gs3d/lane_HA.md#L214-L235)；[kop:cannbench/design/opt11_gs3d/lane_HA_fix1.diff:1–21](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt11_gs3d/lane_HA_fix1.diff#L1-L21)。
-- CB03-05：[kop:cannbench/design/opt11_gs3d/lane_HA.md:39–81](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt11_gs3d/lane_HA.md#L39-L81)；[kop:cannbench/design/opt11_gs3d/lane_HA.md:137–233](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt11_gs3d/lane_HA.md#L137-L233)；[kop:cannbench/design/kernel/grid_sampler_3d.md:2387–2474](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/grid_sampler_3d.md#L2387-L2474)。
-- CB03-06：[kop:cannbench/design/kernel/grid_sampler_3d.md:2476–2553](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/grid_sampler_3d.md#L2476-L2553)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-matmul-rounding-nonfinite.md

@@ -1,56 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-matmul-rounding-nonfinite",
-  "title": "GMM/WQBM：补偿、窄域舍入与非有限运算顺序",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-matmul-rounding-nonfinite",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "GMM/WQBM：补偿、窄域舍入与非有限运算顺序",
-    "GMM：更高精度重算的有效范围与收益归因",
-    "GMM：bf16 bias的融合回归不是所有bias类型都不支持",
-    "WQBM：两次窄域舍入与末尾一次舍入并不等价",
-    "WQBM：非有限repair保留原运算顺序"
-  ]
-}
----
-
-# GMM/WQBM：补偿、窄域舍入与非有限运算顺序
-
-GMM要保留group_list累积边界、空组、split_item、transpose与bias。补偿求和须把comp和乘积残差纳入结果；它不是无条件精确、顺序无关，也不免于溢出/下溢或编译器重结合。新任务带独立_oracle，不能凭cpu_error_count=0强行退化oracle。历史新增通过中5例46/52/55/93/100只走未改AllBiasCube，不能归到refine；refine中位慢1.292×、最坏2.170×，不是无代价。
-
-AllBias融合回归限于当次bf16输入+bf16 bias；FIX1退出该路线、FIX2先将bias升fp32再融合都通过原252样本。fp16+fp16 bias对照通过，不能泛化所有Cube bias只支持fp32。E>8的plugin按8专家分chunk，必须覆盖空组/多chunk；64失败轴摘要不是总80失败的完整展开。
-
-WQBM新plain参考先weight.to(T)，offset/scale各to(T)，有offset时Add后物化T，再Mul物化T，之后fp32 matmul/bias与最终T；无offset仅Mul。旧tasks_official_110反量化全fp32，先固定真正调用的oracle/bench。若在fp32模拟bf16，参考要求的每个点都需显式舍入。两字节N_TILE的repeat stride为N_TILE/16个32B块，fp32为N_TILE/8，并保留尾块处理。
-
-CPU bf16 w36、offset−0.6171875、scale1.609375：两次舍入57.25、末尾一次57.0；half另有−103.9375与−104.0。当前hidden包half走Add<T>/Mul<T>再Cast，bf16仍fp32 Add/Mul后一次CAST_RINT。这是源码差异和CPU表达式反例，尚非命中实际NPU refine路径的失败证明，standard又是另一包。
-
-非有限时s*(Σxw+oΣx)不等于Σ[x*((w+o)*s)]。repair保留逐k反量化后乘x和普通Add；重新加Kahan可能以Inf−Inf生成NaN。当前rowBad=nonfinite(rowMax)||nonfinite(rowSum)，或tile任意scale/offset非有限触发；rowBad为false时非colBad列恢复yRaw。
-
-旧dev_nf.py只比isnan.sum/isinf.sum，不比位置、符号或finite值；3个CPU反例都能假通过。其12配置M8/N64各512元素，无法直接生成原报告816/4096计数，脚本/报告版本未绑定。不能将“element for element”当验收。
-
-## 来源与复核入口
-
-- CB03-08：[kop:cannbench/design/kernel/grouped_matmul.md:1850–1958](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/grouped_matmul.md#L1850-L1958)；[kop:cannbench/design/kernel/grouped_matmul.md:2781–2853](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/grouped_matmul.md#L2781-L2853)；[bench:tasks/level3/grouped_matmul/golden.py:65–154](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/tasks/level3/grouped_matmul/golden.py#L65-L154)。
-- CB03-09：[kop:cannbench/design/kernel/grouped_matmul.md:1760–1809](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/grouped_matmul.md#L1760-L1809)；[kop:cannbench/design/opt14_gmm/lane_a_sweep_gen.py:1–98](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/opt14_gmm/lane_a_sweep_gen.py#L1-L98)。
-- CB03-10：[kop:cannbench/design/kernel/weight_quant_batch_matmul.md:1106–1178](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/weight_quant_batch_matmul.md#L1106-L1178)；[kop:cannbench/tasks_official_110/level3/weight_quant_batch_matmul/golden.py:47–79](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/tasks_official_110/level3/weight_quant_batch_matmul/golden.py#L47-L79)；[bench:tasks/level3/weight_quant_batch_matmul/golden.py:28–88](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/tasks/level3/weight_quant_batch_matmul/golden.py#L28-L88)。
-- CB03-11：[kop:cannbench/design/kernel/weight_quant_batch_matmul.md:260–276](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/weight_quant_batch_matmul.md#L260-L276)；[kop:cannbench/design/kernel/weight_quant_batch_matmul.md:334–350](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/weight_quant_batch_matmul.md#L334-L350)；[kop:cannbench/design/kernel/weight_quant_batch_matmul.md:680–742](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/weight_quant_batch_matmul.md#L680-L742)；[kop:cannbench/probes/README.md:1–22](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/probes/README.md#L1-L22)；[kop:cannbench/probes/weight_quant_batch_matmul/dev_nf.py:1–58](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/probes/weight_quant_batch_matmul/dev_nf.py#L1-L58)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-recurrent-rounding-tail-copy.md

@@ -1,56 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-recurrent-rounding-tail-copy",
-  "title": "LSTM/GRU：CPU执行身份、成本门和拷贝尾部",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-recurrent-rounding-tail-copy",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "LSTM/GRU：CPU执行身份、成本门和拷贝尾部",
-    "LSTM：no_grad与CPU实现身份优先于算术模型",
-    "LSTM：exact成本门、对齐与容量要分开",
-    "GRU：fp32 LocalTensor拷贝遗漏尾部",
-    "GRU：门前额外舍入是被否决的候选"
-  ]
-}
----
-
-# LSTM/GRU：CPU执行身份、成本门和拷贝尾部
-
-LSTM保留y/hn/cn、proj、layer/direction、batchFirst、布局及输入/中间dtype。历史x86 AVX512无proj brgemm的FUSED条件为slc==sic且非(mb==1且T≥2)，否则SPLIT；逐k FMA、bias/cell FMA和激活实现必须绑定no_grad、ISA/oneDNN。直调训练/ref路径slc384曾3719NaN，brgemm为0；aarch64的BN48/BK256/unroll6和glibc模型是另一来源。39seed逐位是有限样本，不是所有K/ISA定律。
-
-LSTM任务三dtype阈值均0.05，对应固定比较器一阶段MERE<.05、MARE<.5；CPU 1.001对1非逐位仍通过。近零分母1e-7可很严格，但不逻辑推出每个上游ulp均以原幅度传播或永远只能bitwise。
-
-当前hidden exact门：无proj，D∈{1,2}，L≥1，L*D≤8，16≤H≤512，H%(bf16?32:16)==0，max(I,D*H)≤1024；S*B*H≥20000，waves=ceil(B/48)，S*H*(I+H)*waves*L*D≤150000000。成本门不是正确性证明。六级改善还含veccast、多层/双向kernel和launch变化，不全是plugin；旧注释I≤512与实际≤1024不一致。
-
-fp32 BinaryRepeatParams的uint8 stride以32B计，H1024时128合法、H2048时256截0；这不是所有DMA字段都8bit，也不是当前容量/路由已触发活跃bug。WholeReduceMax mask64/repeat≤255等容量分别核。
-
-GRU ConvertToFloat/FromFloat的sizeof(T)==sizeof(float)分支用LocalTensor DataCopy，count=H且H%8≠0：H13只覆盖8，尾5/13约.3846。vec及旧kernel都要修，低精度走Cast是不同路径；H13/I26或64、幅度.1–2.0与H16/32、halfH13对照保留。Adds(dst,src,0,count)覆盖尾部，但数值加零不能直接叫逐位copy，±0/NaN、mask/count和别名另验。
-
-GRU门前激活在sigmoid/tanh之前额外舍到T，使MERE .0225→2.310；不能类比WQBM给所有GRU中间点加Cast。第61个权重±5样本normal4/1156/native1不等于线上327/native0；gru_32的y错256而hn零误差仍分别保存。gru_101本身fp32，不能按half存储丢位归因。
-
-## 来源与复核入口
-
-- CB03-12：[kop:cannbench/design/kernel/lstm.md:42–130](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/lstm.md#L42-L130)；[kop:cannbench/design/kernel/lstm.md:239–339](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/lstm.md#L239-L339)；[kop:cannbench/tasks_official_110/level4/lstm/proto.yaml:1–34](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/tasks_official_110/level4/lstm/proto.yaml#L1-L34)；[bench:src/kernel_eval/eval/op_runner.py:285–312](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/eval/op_runner.py#L285-L312)。
-- CB03-13：[kop:cannbench/design/kernel/lstm.md:239–339](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/lstm.md#L239-L339)；[kop:cannbench/design/OPT_ROUND_CHECKLIST.md:1039–1059](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/OPT_ROUND_CHECKLIST.md#L1039-L1059)；[kop:cannbench/design/OPT_ROUND_CHECKLIST.md:3495–3633](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/OPT_ROUND_CHECKLIST.md#L3495-L3633)。
-- CB03-14：[kop:cannbench/design/kernel/gru.md:530–591](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gru.md#L530-L591)。
-- CB03-15：[kop:cannbench/design/kernel/gru.md:592–669](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gru.md#L592-L669)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-reference-and-special-values.md

@@ -1,52 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-reference-and-special-values",
-  "title": "参考版本、特殊值和有效探针",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-reference-and-special-values",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "参考版本、特殊值和有效探针",
-    "先读返回分支与全部输出，再解释误差字段",
-    "版本标识要落到 task、oracle、bench 与 checker 文件",
-    "特殊值的位置、符号与生成分布不能被摘要抹掉",
-    "环境与probe：阴性结果只能覆盖实际运行的分支"
-  ]
-}
----
-
-# 参考版本、特殊值和有效探针
-
-先固定任务的proto/cases/golden、可选_oracle/_bench、loader、比较器及实际加载环境。相同版本字符串不能证明文件相同；相同golden.py也不能证明no_grad、CPU ISA/BLAS和所调用分支相同。当前对照是本地CANN-Bench d59d44ca，未证明等同历史服务端。
-
-NaN位置失败的提前返回不能证明有限元素正确。非有限验收要区分NaN位置、Inf位置和符号、有限值误差及全部输出。相同NaN/Inf数量不足以证明逐元素相同。历史特殊值生成器约50% NaN；Inf先写头n项−Inf再写尾n项+Inf，n=max(1,numel//20)，numel=1最终为+Inf。无generator的randn与有generator的均匀[-1,1]不是同分布。
-
-历史AIV FP32的Inf+(-Inf)在14/14输入与CPU匹配；异常观测限于当次BF16 Cube累加路径，不能写成Ascend全局NaN规则。16行特殊值表为5行finite比例非零、11行为零，正文12+4~5的说法不一致，保留缺口。
-
-头文件600→764的路径集合不是API语义超集。D4实际是20例正逆序位串，未做原计划UB投毒。Transpose探针的continue跳过尾PIPE_ALL使恒等对照也失败；双转置又不能排除no-op。Sigmoid不幂等，重复同输入的探针必须确认输入未被原地改写。新seed通过不注销旧失败。
-
-## 来源与复核入口
-
-- CB03-01：[kop:cannbench/design/WHY_WE_BURNED_THE_COMPUTE.md:513–583](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/WHY_WE_BURNED_THE_COMPUTE.md#L513-L583)；[bench:src/kernel_eval/utils/compare.py:365–690](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/utils/compare.py#L365-L690)；[bench:src/kernel_eval/utils/compare.py:692–925](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/utils/compare.py#L692-L925)。
-- CB03-02：[bench:src/kernel_eval/eval/evaluator.py:240–360](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/eval/evaluator.py#L240-L360)；[bench:src/kernel_eval/eval/evaluator.py:877–897](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/eval/evaluator.py#L877-L897)；[bench:src/kernel_eval/_version.py:16–59](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/_version.py#L16-L59)；[bench:src/kernel_eval/report/report_generator.py:278–294](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/report/report_generator.py#L278-L294)。
-- CB03-03：[kop:cannbench/design/SPECIAL_VALUE_CASES_20260826.md:8–185](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/SPECIAL_VALUE_CASES_20260826.md#L8-L185)；[kop:cannbench/design/HOW_WE_FOUND_THE_PATH.md:805–817](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/HOW_WE_FOUND_THE_PATH.md#L805-L817)；[bench:src/kernel_eval/data/data_generator.py:292–345](https://gitcode.com/cann/cann-bench/blob/d59d44ca1cf14ed20c1e90cbbcbf4a3747ff84de/src/kernel_eval/data/data_generator.py#L292-L345)。
-- CB03-20：[kop:cannbench/design/VERSION_DIVERGENCE_PROBE.md:20–188](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/VERSION_DIVERGENCE_PROBE.md#L20-L188)；[kop:cannbench/design/REVIEW_20260825.md:340–412](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/REVIEW_20260825.md#L340-L412)；[kop:cannbench/design/REVIEW_20260825.md:625–707](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/REVIEW_20260825.md#L625-L707)；[kop:cannbench/design/OPT_ROUND_CHECKLIST.md:454–465](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/OPT_ROUND_CHECKLIST.md#L454-L465)；[kop:cannbench/design/OPT_ROUND_CHECKLIST.md:1168–1181](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/OPT_ROUND_CHECKLIST.md#L1168-L1181)；[kop:cannbench/lessons/lessons.md:762–898](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/lessons/lessons.md#L762-L898)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb03-roi-gqa-layout-boundaries.md

@@ -1,54 +0,0 @@
----
-{
-  "id": "exp-ascend-cb03-roi-gqa-layout-boundaries",
-  "title": "ROIAlign/GQA：采样序、UB对齐与真实路由",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb03-roi-gqa-layout-boundaries",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本次仅固定源码与CPU证据",
-    "cpu": "x86_64, torch 2.13.0+cpu（仅标明的CPU反例）"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/29-cb03-correctness-and-environment-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "固定快照2026-09-11；历史测量、源码观察与CPU反例分开；未新跑NPU，不声明完整输入域正确。",
-  "aliases": [
-    "ROIAlign/GQA：采样序、UB对齐与真实路由",
-    "ROIAlign：运算顺序、布局与参考环境绑定",
-    "ROIAlign：核数空闸与RowWalk采样序耦合",
-    "GQA：先判真实pack/restore与Nd2Nz字段",
-    "GQA：死分支缺陷与特殊值防护各保留负结果"
-  ]
-}
----
-
-# ROIAlign/GQA：采样序、UB对齐与真实路由
-
-ROIAlign插值轴为[n,C,oh,ow,gh,gw]，归约gh/gw，不是旧注释[n,C,oh,gh,ow,gw]。历史fallback需确认torchvision是否可导入、ATen/ISA/no_grad。左结合hy*hx*v00+hy*lx*v01+ly*hx*v10+ly*lx*v11、两次RoundOnce、sum/count除法及阻FMA路径按当时参考核，不能泛化固定8lane顺序。fp32阈值.01、fp16 .1并非完全无效；2^-16对2.4e-7约64倍，不是五个数量级。
-
-sr4/grid4的pairSamples顺序0,2,1,3与RowWalk自然DMA0,1,2,3相冲突，旧负UB偏移转uint32形成巨大地址，507035。修法按自然顺序预存ubRowTab[4]，配对时查表，保留(!pairSamples || gridHeight==4)与compactSlab准入。当前blocks=clamp(max(B*C,rois),1,cores)，没有ROI_BLOCKS环境入口。历史720配置只有6形状、blocks32/48；实际success281/failed4/cascade435，不能称716通过或全核数覆盖。
-
-GQA只有实际pack/restore且G=Nq/Nkv≥2时检查rowElements=G*D、rowBytes=rowElements*sizeof(T)及32B落点。G1的stridedGroupLayout绕过同一pack；headMajor/direct也先排除。逐行fallback条件rowsPerTile<1或rowBytes%32!=0；Nd2NzParams::srcDValue为uint16，Nq*D65536截断与行距是独立风险，49152通过不越过边界。
-
-当前standard三处有rowBytes guard，hidden相应只有rowsPerTile<1。hidden的fp16、causal、qSeq>1、kvHeads>1、G2/D12静态可到pack，但尚未设备复现，也未证明就是不可见hidden67/68的根因。
-
-旧未实例化GqaBatchedSoftmax因果路径Duplicate偏移localRow*ubRow+validKeys不一定8float对齐。强制路线修后通过，但case6/20/19分别慢97%/71%/40%，不能用跨case比较宣称收益。score×scale非有限保护需逐行数据判据；仅bf16 host scale>1曾有1.953e-3位移。数据条件与形状条件可同时满足，不能称互斥或保证任意合并顺序。
-
-## 来源与复核入口
-
-- CB03-16：[kop:cannbench/design/kernel/roi_align.md:36–105](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/roi_align.md#L36-L105)；[kop:cannbench/tasks_official_110/level3/roi_align/golden.py:1–248](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/tasks_official_110/level3/roi_align/golden.py#L1-L248)；[kop:cannbench/tasks_official_110/level3/roi_align/proto.yaml:1–13](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/tasks_official_110/level3/roi_align/proto.yaml#L1-L13)。
-- CB03-17：[kop:cannbench/design/kernel/roi_align.md:1286–1415](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/roi_align.md#L1286-L1415)；[kop:cannbench/harness/gen_roi_blockcov.py:1–52](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/harness/gen_roi_blockcov.py#L1-L52)。
-- CB03-18：[kop:cannbench/design/kernel/gqa.md:1222–1318](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gqa.md#L1222-L1318)；[kop:cannbench/design/kernel/gqa.md:1483–1507](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gqa.md#L1483-L1507)。
-- CB03-19：[kop:cannbench/design/kernel/gqa.md:1497–1507](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gqa.md#L1497-L1507)；[kop:cannbench/design/kernel/gqa.md:1556–1610](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/gqa.md#L1556-L1610)。
-
-完整包成员/哈希、两侧源码、输入生成差异和CPU原始结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/ba7842cfb7e01dcaae4c85cc97b6ee44f88009ff/docs/ascendc-migration/evidence/cb03-20260911/README.md)。完整包身份不证明未审路径正确。榜单分数、平台配额与提交规则由CANNBench skill维护。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb04-aap3d-packing.md

@@ -1,47 +0,0 @@
----
-{
-  "id": "exp-ascend-cb04-aap3d-packing",
-  "title": "AdaptiveAvgPool3D：packed plane 与分组 store 地址",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb04-aap3d-packing",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本轮固定源码与CPU算术/parser复核"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/30-cb04-data-movement-pipelining-and-fixed-cost-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "2026-09-11固定快照；历史测量/源码/CPU反例分开，未新增NPU评测。性能值不是API单价或完整合法域保证。",
-  "aliases": [
-    "AdaptiveAvgPool3D：packed plane 与分组 store 地址",
-    "AAP3D packed plane 与全局归约：保留准入顺序和反例",
-    "AAP3D 输出行距的真正条件，与位复制的计数单位"
-  ]
-}
----
-
-# AdaptiveAvgPool3D：packed plane 与分组 store 地址
-
-NCDHW 的非整除池化先检查全局1³归约；2倍上采样是独立路线，主要fp32累加。packed plane 将 case8 的217次burst降到7；准入包含 hw≤16320、Oh*Ow≤4096、span_w*out_pad*4≤2048、span_h*Oh≤1024、nc*Od≥blocks、span_d≤4095、H*aligned_w≤16320、工作集≤160KiB。低精度分组还核占用、约1.15的不平衡门和容量，不能只抄某一个面积阈值。
-
-64KiB一次与2×16×2KiB的对照8.40→8.70µs是负例。55–102cycles/burst限当次工作点；1.8ns是归一化量，原循环未恢复，不能改写成.3µs/31的API单价。观测1.37–1.47TB/s也不等物理roof。
-
-分组store的源地址 j*out_pad*sizeof(T) 必须逐行核32B。out_q=max(8,32/sizeof(T))，host和kernel同时使用对应padding。旧3×7按8元素对齐到24，half行距48B，j1不对齐并曾507035；方形9×9的88 half元素=176B也是CPU地址反例，需实际分组且gcur≥2才触发第二行，尚无本轮NPU失败复现。2×8对齐；未分组且源偏移恒零的ceil8不是同一缺陷。
-
-2倍上采样还需(2*Iw*sizeof(T))%32==0、176KiB容量和min_chunks8等门。int64/half位复制的元素计数、字节数和32B块数分开记录，不能以逻辑行长代替物理落点。
-
-## 来源与复核入口
-
-- CB04-07：[kop:cannbench/design/kernel/adaptive_avg_pool_3d.md:875–996](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/adaptive_avg_pool_3d.md#L875-L996)；[kop:cannbench/design/kernel/aap3d_lanes/LANE_B_REPORT.md:178–210](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/aap3d_lanes/LANE_B_REPORT.md#L178-L210)；[kop:cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md:87–121](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md#L87-L121)；[kop:cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md:162–193](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md#L162-L193)。
-- CB04-08：[kop:cannbench/design/kernel/adaptive_avg_pool_3d.md:997–1049](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/adaptive_avg_pool_3d.md#L997-L1049)；[kop:cannbench/design/kernel/aap3d_lanes/LANE_B_REPORT.md:347–373](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/aap3d_lanes/LANE_B_REPORT.md#L347-L373)；[kop:cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md:340–407](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/aap3d_lanes/LANE_D_REPORT.md#L340-L407)。
-
-完整包、成员哈希、连续代码窗口和19项CPU结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/evidence/cb04-20260911/README.md)。未恢复probe只作报告级来源；源码包存在不证明所有分支已设备验证。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb04-conv2d-stage-lifetime.md

@@ -1,49 +0,0 @@
----
-{
-  "id": "exp-ascend-cb04-conv2d-stage-lifetime",
-  "title": "Conv2D：跨调用 stage、容量和短循环",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb04-conv2d-stage-lifetime",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本轮固定源码与CPU算术/parser复核"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/30-cb04-data-movement-pipelining-and-fixed-cost-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "2026-09-11固定快照；历史测量/源码/CPU反例分开，未新增NPU评测。性能值不是API单价或完整合法域保证。",
-  "aliases": [
-    "Conv2D：跨调用 stage、容量和短循环",
-    "Conv2D 双 stage：保护真正的最后读者",
-    "Conv2D 容量、短循环和融合负例必须随规则迁移"
-  ]
-}
----
-
-# Conv2D：跨调用 stage、容量和短循环
-
-Shift、SmallPlane、DilatedSmallPlane、StridedPlane 与 FastFinal 的两槽流水要从完整源码核对。stage 索引 int64_t& 跨 ProcessRow 调用；每槽独立 buffer 和事件，初始化两个 free event，跳过无工作行发生在 acquire 前，输出提交后发 MTE3→MTE2，生命周期结束排空。该 WAR 保护 MTE3 最后读者与下一次 MTE2 写者；保留 MTE2→V、V→MTE3。不要每个 helper 返回就排空。
-
-StridedPlane 仍在 load 共享 inputPlane 前用 PIPE_ALL；两 stage 不覆盖这块共享输入的所有复用依赖。保留 CAST_RINT 的低精度 epilogue，核两份 stage 的单槽容量、总量、skip 和提前退出。事件池耗尽/API版本不能仅由这批包的静态存在证明。
-
-历史 v20 与 v17 本地对照中部分用例快10.7–18.4%、控制±.5%，不代表所有路径。放宽容量门曾使 hidden 73→31/80 并出现507034级联；outputColStart/rightPad 的窄化与 fallback 丢失需分开定位，不能合成一个容量常量。恢复正确性后再定价。
-
-FastFinal 长循环比率约.66，只有1–2行时约.87。保留启动/排空、每核实际迭代和所有阶段耗时；只按迭代数打折忽略了调度、访存复用和容量。两个三段融合尝试分别慢308%/640%，是该实现的负例，不是“设备计时排除launch”的证明。
-
-## 来源与复核入口
-
-- CB04-05：[kop:cannbench/design/kernel/conv_2d.md:949–1004](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L949-L1004)；[kop:cannbench/design/kernel/conv_2d.md:1065–1104](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L1065-L1104)。
-- CB04-06：[kop:cannbench/design/kernel/conv_2d.md:54–138](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L54-L138)；[kop:cannbench/design/kernel/conv_2d.md:751–790](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L751-L790)；[kop:cannbench/design/kernel/conv_2d.md:924–948](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L924-L948)；[kop:cannbench/design/kernel/conv_2d.md:1085–1121](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/conv_2d.md#L1085-L1121)。
-
-完整包、成员哈希、连续代码窗口和19项CPU结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/evidence/cb04-20260911/README.md)。未恢复probe只作报告级来源；源码包存在不证明所有分支已设备验证。
-
-BigK 可能使用另一事件方案，不能把这些 stage 结论覆盖整树。FastFinal 有按 channel 分区复用 bias 的分支，每核 batch*cOut/blockNum 仅近似；旧19/20本地例外不成为产品验收默认。

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/cb04-cummin-scan-and-probes.md

@@ -1,53 +0,0 @@
----
-{
-  "id": "exp-ascend-cb04-cummin-scan-and-probes",
-  "title": "Cummin：扫描状态、无效探针与失败组合",
-  "type": "experience",
-  "vendor": "ascend",
-  "probe_slug": "cb04-cummin-scan-and-probes",
-  "measured_on": {
-    "device": "历史910B2C/910C记录；本轮固定源码与CPU算术/parser复核"
-  },
-  "source": "https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/30-cb04-data-movement-pipelining-and-fixed-cost-evidence.md",
-  "architectures": [
-    "ascend910b"
-  ],
-  "languages": [
-    "ascendc"
-  ],
-  "tags": [
-    "ascendc"
-  ],
-  "evidence_level": "inferred",
-  "confidence": "experimental",
-  "scope": "2026-09-11固定快照；历史测量/源码/CPU反例分开，未新增NPU评测。性能值不是API单价或完整合法域保证。",
-  "aliases": [
-    "Cummin：扫描状态、无效探针与失败组合",
-    "Cummin 转置探针：恒等性和同步有效性要分别证明",
-    "Cummin 当前scan：跨行隔离、carry和两种输出宽度",
-    "Cummin 失败组合：正确、变快和上线是不同结论"
-  ]
-}
----
-
-# Cummin：扫描状态、无效探针与失败组合
-
-双转置恒等不能排除no-op，单转置探针的continue又漏尾PIPE_ALL，使恒等对照也失败；该旧API禁令撤回。values .043/.011与indices42.2%/33.9%保留为失败观察，不推唯一转置越界。16M×8两次6/6仅限对应布局；同址Scatter不是完整流水证明。
-
-当前axis chunk按VL对齐且≤2048，ab=min(32,max(1,4096/chunk))，slab为ab+1。仅baseTasks*2≤blocks才尝试axis split，两遍需要summary/carry与重读代价。contig MAXB8、DB1以当前代码为准，不能按旧注释称DB2。输出index int64，内部int32仍需axis范围；multirun必须行内隔离，不能跨行延续最小值/索引；跨task sticky NaN、latest tie、±0和值/索引两个输出一并验。
-
-inner1的repeat按VL/axis/outer tile组织；round005的vcgmin≤8和CreateVecIndex只在原条件内，NaN快路总目标65.80→65.77近中性不能换成删NaN语义的许可证。
-
-P11把每16步48次DMA降至3次，20/20 values和indices MD5相同，但整体近中性，未兑现case8的10%预测。P12 next707010 TOTAL BAD，首次rand35_17已有vbad12/ibad32，应停止性能判读。faithful bitview保守同步、32B gap及batchCast的恢复bad0仍近中性；私有faithful原件未恢复，int32/inner>1且≤64/axis<2^24只属报告范围。
-
-scalar float inner1、chunks1、axis≤2048且%8==0，rpt2/4、24模式虽20例通过，但case2为2575.26对130.49µs，不能把串行标量探针当vector下界。失败、正确、变快及是否提交分别记录。
-
-## 来源与复核入口
-
-- CB04-14：[kop:cannbench/design/REVIEW_20260825.md:625–836](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/REVIEW_20260825.md#L625-L836)。
-- CB04-18：[kop:cannbench/design/kernel/cummin.md:3518–3532](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/cummin.md#L3518-L3532)；[kop:cannbench/design/kernel/cummin.md:4104–4183](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/cummin.md#L4104-L4183)。
-- CB04-19：[kop:cannbench/design/kernel/cummin.md:3119–3136](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/cummin.md#L3119-L3136)；[kop:cannbench/design/kernel/cummin.md:3792–3896](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/cummin.md#L3792-L3896)；[kop:cannbench/design/kernel/cummin.md:3897–3959](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/cdabbb918bd275501f4c3fe6677b44d75743b36b/cannbench/design/kernel/cummin.md#L3897-L3959)。
-
-完整包、成员哈希、连续代码窗口和19项CPU结果见 [固定证据目录](https://github.com/sii-ai-infra/kernel-opt-pipeline/blob/143fe7201e796a387879f67771e6999dbf81d661/docs/ascendc-migration/evidence/cb04-20260911/README.md)。未恢复probe只作报告级来源；源码包存在不证明所有分支已设备验证。
-
-探针身份与版本限定共用 [CB03-20](cb03-reference-and-special-values.md)。