@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/dynamic-quant-event-lifetime.md

@@ -1,95 +0,0 @@
----
-id: exp-ascend-dynamic-quant-event-lifetime
-title: 'DynamicQuant: complete-row dependency and independently reserved pipeline events'
-type: experience
-vendor: ascend
-probe_slug: dynamic-quant-event-lifetime
-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:
-- DynamicQuant
-- R13
-- DQ301
-- AllocEventID
-- FetchEventID
-- 事件
-- 流水
-- 双缓冲
-- exp-ascend-cannbench-20261020-dynamic-quant-event-lifetime
-- wiki/ascend/experience/cannbench-20261020-dynamic-quant-event-lifetime.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: Width8192, two complete rows, CANN9.1/Ascend910B2C. Width16384 uses another path.
-knowledge_id: knowledge-ascend-dynamic-quant-event-lifetime
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- DynamicQuant
-condition_ids:
-- condition-cannbench-20261020-dynamic-quant-event-lifetime
-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-method.md
-related:
-- skill-ascend-cross-core-sync
-- skill-ascend-dynamic-quant
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
-kernel_types:
-- quantization
----
-
-# DynamicQuant: complete-row dependency and independently reserved pipeline events
-
-Historical **public** reference (2026-10-16T05:33:28.798747+00:00): **85.12330552, 20/20**. This is the selected aggregate at that capture, not a claim that an earlier checkpoint below produced the latest score. Keep the source-specific public baseline and its code identity.
-
-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
-
-Width8192, two complete rows, CANN9.1/Ascend910B2C. Width16384 uses another path.
-
-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:** DQ301 already had bit-order magnitude reduction, prefetch and two output slots. A later experimental manual-event prototype reused FetchEventID without reserving independent IDs.
-
-**Change:** Reserve independent IDs with AllocEventID. Wait MTE2→V before input use, V→MTE2 before overwriting input, V→MTE3 before output transfer and MTE3→V before reusing each output generation. Drain outstanding events before releasing IDs. Release raw input only after magnitude and FP32 copies no longer read it.
-
-**Mechanism:** Input readiness, input reuse and each output generation are separate dependencies; a complete-row maximum must exist before quantizing its chunks. Explicit event ownership avoids falsely treating two fetched IDs as independent slots.
-
-## Recorded results
-
-Corrected R13 ran the original checker. The retained width8192 buffer plan is20*width+1088=164928 B. Same-address single-row and strictly serial variants were slower in the recorded comparisons; changing event management did not establish a large stable whole-task gain.
-
-## Boundaries, negative evidence and reuse
-
-The bad event prototype was an experiment, not an established defect in DQ301 or an explanation of another team. Check SDK event semantics, actual UB capacity and tail drain in a new build. Cache-policy negatives exclude tested organizations only; cache/layout/granularity combinations were not exhausted.
-
-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-method](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/dq-method.md): exact input SHA256 and original line ranges.
-- [Original operator chronicle](../../../corpus/ascend/cannbench/optimization-history-20261020/operator-records/dynamic_quant.json): preserved source text with its original paths and dates.
-
-Prior knowledge: `skill-ascend-cross-core-sync`, `skill-ascend-dynamic-quant`.
-
-[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/dynamic-quant-short-rows.md

@@ -1,102 +0,0 @@
----
-id: exp-ascend-dynamic-quant-short-rows
-title: 'DynamicQuant: continuous short-row IO with compact reduction metadata'
-type: experience
-vendor: ascend
-probe_slug: dynamic-quant-short-rows
-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:
-- DynamicQuant
-- dynamic_quant
-- 动态量化
-- 短行
-- 连续搬运
-- C04
-- C05
-- '373'
-- '67'
-- exp-ascend-cannbench-20261020-dynamic-quant-short-rows
-- wiki/ascend/experience/cannbench-20261020-dynamic-quant-short-rows.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: FP16 width=373 and BF16 width=67; Ascend910B2C copy-5 physical4; original task/checker, final nine interleaved evaluations on 2026-10-20. Broader shape applicability needs validation.
-knowledge_id: knowledge-ascend-dynamic-quant-short-rows
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- DynamicQuant
-condition_ids:
-- condition-cannbench-20261020-dynamic-quant-short-rows
-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-method.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/dq-comparison.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/dq-identity.md
-related:
-- skill-ascend-dynamic-quant
-- exp-ascend-cb06-norm-quant-domains
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
-kernel_types:
-- quantization
----
-
-# DynamicQuant: continuous short-row IO with compact reduction metadata
-
-Historical **public** reference (2026-10-16T05:33:28.798747+00:00): **85.12330552, 20/20**. This is the selected aggregate at that capture, not a claim that an earlier checkpoint below produced the latest score. Keep the source-specific public baseline and its code identity.
-
-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
-
-FP16 width=373 and BF16 width=67; Ascend910B2C copy-5 physical4; original task/checker, final nine interleaved evaluations on 2026-10-20. Broader shape applicability needs validation.
-
-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:** The short-row path copied, padded, reduced and stored individual rows, paying many small DMA operations and expanding row factors.
-
-**Change:** Load contiguous groups; reduce complete internal blocks plus explicit row-edge elements into compact per-row metadata, compute one exact row max, select the row factor by true boundaries, then store continuous int8 output. FP16 uses 16-element blocks, 64 half metadata values/row and a 12288-element tile; BF16 uses FP32 reduction, 8-element blocks and 32 FP32 metadata values/row. Preserve safe-range checks and FP32 fallback.
-
-**Mechanism:** Keep large IO contiguous while representing row boundaries in much smaller reduction/factor metadata. Blocks crossing a row must not contaminate that row maximum. Every column in a row uses the same complete-row scale.
-
-## Recorded results
-
-C05 cross-run median c10: 196.71→126.97 us (35.45% lower); c14: 121.89→82.56 us (32.27% lower). Whole20 equal-weight latency geometric speedup versus R13=1.0336x. C05 passed20/20 in three final runs. This is local evidence, without a new formal910C submission.
-
-## Boundaries, negative evidence and reuse
-
-C05 c4 regressed96.56→112.31 us; c3 regressed36.06→37.40 us. Its scores83.3245/83.0644/83.4704 did not achieve the ~124 target. C04 is the narrower short-row checkpoint; neither is universally best. BF16 bit patterns must not be interpreted as ordinary half numbers. Profiler pipe durations include warmup/overlap and cannot be summed into total time.
-
-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-method](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/dq-method.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.
-- [dq-identity](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/dq-identity.md): exact input SHA256 and original line ranges.
-- [Original operator chronicle](../../../corpus/ascend/cannbench/optimization-history-20261020/operator-records/dynamic_quant.json): preserved source text with its original paths and dates.
-
-Prior knowledge: `skill-ascend-dynamic-quant`, `exp-ascend-cb06-norm-quant-domains`.
-
-[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/engram-stride-and-grid.md

@@ -1,93 +0,0 @@
----
-id: exp-ascend-engram-stride-and-grid
-title: 'EngramGateFusion: test UB row-stride bank conflicts and prove empty-grid removal'
-type: experience
-vendor: ascend
-probe_slug: engram-stride-and-grid
-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:
-- EngramGateFusion
-- engram_gate_fusion
-- gated8
-- bankgroup
-- 128B
-- fitblocks
-- 门控
-- exp-ascend-cannbench-20261020-engram-stride-and-grid
-- wiki/ascend/experience/cannbench-20261020-engram-stride-and-grid.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: Recorded gated8 FP32 row strides, local910B2C/CANN9.1. Host grid changes are conditional on the exact partition formula and Init side effects.
-knowledge_id: knowledge-ascend-engram-stride-and-grid
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- EngramGateFusion
-condition_ids:
-- condition-cannbench-20261020-engram-stride-and-grid
-sources:
-- exp-source-cannbench-optimization-20261020
-source_refs:
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/engram-banks.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/campaign-phase2.md
-related:
-- exp-ascend-cb06-elementwise-and-fused-precision
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
-kernel_types:
-- fused-kernel
----
-
-# EngramGateFusion: test UB row-stride bank conflicts and prove empty-grid removal
-
-Historical **public** reference (2026-10-16T05:33:28.798747+00:00): **64.73859366, 20/20**. This is the selected aggregate at that capture, not a claim that an earlier checkpoint below produced the latest score. Keep the source-specific public baseline and its code identity.
-
-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
-
-Recorded gated8 FP32 row strides, local910B2C/CANN9.1. Host grid changes are conditional on the exact partition formula and Init side effects.
-
-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:** Power-of-two UB row strides produced repeated bank-group conflicts; the host Prepare grid also launched empty blocks. These are distinct mechanisms.
-
-**Change:** For the bank experiment insert128B row skew within a fixed1024 tile/UB bound and retain unchanged non-power-of-two rows as controls. For grid pruning separately prove ceil(T/A)=r after A=ceil(T/ceil(T/B0)); every surviving block must retain its exact row range and removed blocks must have no writes.
-
-**Mechanism:** Change address congruence without shrinking useful work, or remove provably empty launches without changing ownership. Each needs its own final-source measurement.
-
-## Recorded results
-
-Bank dose series0/32/64/128/384/512B recorded conflict119.1/94.7/79.2/65.4/65.5/119.0 us; local treatment score+.2876 with controls−.0066, all32 evaluations20/20. The source explicitly calls this local evidence, not a formal point gain.
-
-## Boundaries, negative evidence and reuse
-
-PMU ratio denominator was task_cyc, not vector cycles. Do not multiply by the wrong time or pool away hot-loop exposure. An independent EF04 empty-convolution-grid experiment had locally positive means but formal64.738594→64.662932,−.075661; preserve it. Actual40/48 AIV changes eligibility, and source-level equality does not prove build or host-text equality.
-
-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
-
-- [engram-banks](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/engram-banks.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.
-- [Original operator chronicle](../../../corpus/ascend/cannbench/optimization-history-20261020/operator-records/engram_gate_fusion.json): preserved source text with its original paths and dates.
-
-Prior knowledge: `exp-ascend-cb06-elementwise-and-fused-precision`.
-
-[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/foreach-addcdiv-single-tile-queue.md

@@ -1,91 +0,0 @@
----
-id: exp-ascend-foreach-addcdiv-single-tile-queue
-title: 'ForeachAddcdivScalar: remove unused queue state only for one-tile launches'
-type: experience
-vendor: ascend
-probe_slug: foreach-addcdiv-single-tile-queue
-measured_on:
-  device: Historical Ascend910B2C local measurements; see source-bound scope
-  recorded_on: '2026-10-15'
-source:
-- path: sources/experience/cannbench-branch-review.md
-architectures:
-- ascend910b
-languages:
-- ascendc
-tags:
-- ascendc
-aliases:
-- ForeachAddcdivScalar
-- FS802
-- FS801
-- single tile
-- single shot
-- queue bookkeeping
-- TensorList
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: FS802-v1 non-fused per-tensor launches with 0<blockLength<=8192; local 910B2C/CANN9.1.
-  The measured legal family has fifteen members; current public20 has zero new-route
-  hits.
-knowledge_id: knowledge-ascend-foreach-addcdiv-single-tile-queue
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- ForeachAddcdivScalar
-condition_ids:
-- condition-foreach-addcdiv-single-tile-queue-recorded-910b2c-cann91
-sources:
-- exp-source-cannbench-branch-review
-source_refs:
-- source_id: experience/cannbench-branch-review-20261020
-  path: evidence/20261015-52b1-fs802-held/LOCAL_QUALIFICATION_HELD.json
-- source_id: experience/cannbench-branch-review-20261020
-  path: evidence/20261015-52b1-fs802-held/HELD_CLOSEOUT.md
-- source_id: experience/cannbench-branch-review-20261020
-  path: evidence/20261015-52b1-fs802-held/ROOT_FAMILY_RAW_REPLAY.json
-- source_id: experience/cannbench-branch-review-20261020
-  path: evidence/20261015-52b1-fs802-held/FS803_SCREEN_RECEIPT.json
-- source_id: experience/cannbench-branch-review-20261020
-  path: source-members/fs802/ARCHIVE.json
-related:
-- exp-ascend-cb04-foreach-softmax-boundaries
-evaluation_scope: public-regression-and-recorded-local-domain
-verification_status: historical-source-review; no independent device reproduction
-recorded_current_public_target_hits: 0
----
-
-# ForeachAddcdivScalar: remove unused queue state only for one-tile launches
-
-Historical source-reported evidence, reviewed on 2026-10-20. No independent hardware reproduction or newly earned score is claimed. Recorded local 910B2C/CANN9.1 results do not establish 910C or 950 runtime behavior.
-
-## Before, change and mechanism
-
-FS801's one-tile body removes the unused second queue slot/FIFO bookkeeping where each block actually executes one tile. Three input DMAs and the original Div/Muls/Add/cast order remain. Explicit MTE2→V, V→MTE3 and final MTE3→V ownership completion remain necessary even without a next tile; allocated event IDs must be released.
-
-FS802 adds stride/mixed-dtype normalization and empty-list/stream-lifetime handling needed by that candidate's legal inputs. Those correctness helpers are part of its source binding, not evidence that conversion should be inserted on already-compatible public routes. The exact single-shot kernel, dispatch, helper and plugin are retained.
-
-## Historical result
-
-Final ZIP `0fb97fd59beb9f1035d6f590a24cdc30ac5c8be6a54acced42c0131b614cdcf1` has 37 members. The panel passed **164/164 runnable checks**; three rank82 comparator errors outside the declared rank1..8 remain recorded errors, not passes. Thirty-two edge/stream checks, nine error-contract checks, public five-point100/100 and sixty companion cases passed. Parent/FS801/FS802 output equality does not mean native-exact arithmetic: 26/132 cancellation fixtures had small task-tolerated differences.
-
-All fifteen family members, including weak/negative members, remain in paired mean log speedups **0.0768173349 / 0.0621699352**, versus same-statistic noise **0.0157150647**. The complete-family statistic was posthoc and is labeled accordingly.
-
-FS803's separate multi-tile queue replacement gave whole20 **−0.0173475003 / +0.0063795425**, within AA/BB **0.0248831233**. The old two-slot path already overlapped transfers; do not generalize the single-shot result to removing queues from multi-tile paths.
-
-## Current public relevance
-
-**Zero current public20 new-route hits; zero credited public gain.** The baseline is the literal hidden parent, distinct from the public source. This is a scoped local-domain alternative, not a new public winner or a 910C/multi-device TensorList certification. Preserve the index's public reference **88.06305084**.
-
-## Source and comparison
-
-- [LOCAL_QUALIFICATION_HELD.json](../../../corpus/ascend/cannbench/branch-review-20261020/evidence/20261015-52b1-fs802-held/LOCAL_QUALIFICATION_HELD.json)
-- [HELD_CLOSEOUT.md](../../../corpus/ascend/cannbench/branch-review-20261020/evidence/20261015-52b1-fs802-held/HELD_CLOSEOUT.md)
-- [ROOT_FAMILY_RAW_REPLAY.json](../../../corpus/ascend/cannbench/branch-review-20261020/evidence/20261015-52b1-fs802-held/ROOT_FAMILY_RAW_REPLAY.json)
-- [FS803_SCREEN_RECEIPT.json](../../../corpus/ascend/cannbench/branch-review-20261020/evidence/20261015-52b1-fs802-held/FS803_SCREEN_RECEIPT.json)
-- [ARCHIVE.json](../../../corpus/ascend/cannbench/branch-review-20261020/source-members/fs802/ARCHIVE.json)
-
-[Public operator index](cannbench-public-experiences.md) · [branch source review](../../../sources/experience/cannbench-branch-review.md)

package/resources/kop-2026-10-23-0.1.17-beta.1/product_capabilities/kernel_optflow/skills/KernelWiki-private/wiki/ascend/experience/foreachnorm-bank-layout.md

@@ -1,94 +0,0 @@
----
-id: exp-ascend-foreachnorm-bank-layout
-title: 'ForeachNorm: combine semantic repair with PMU-guided accumulator placement'
-type: experience
-vendor: ascend
-probe_slug: foreachnorm-bank-layout
-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:
-- ForeachNorm
-- foreach_norm
-- FN716
-- FN711
-- bank conflict
-- 范数
-- p0
-- 负无穷
-- exp-ascend-cannbench-20261020-foreachnorm-bank-layout
-- wiki/ascend/experience/cannbench-20261020-foreachnorm-bank-layout.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: CANN9.1 Count API, local910B2C; FP32 arithmetic includes BF16 and non-HalfPath FP16. The mixed FP16 P2 route remains separate.
-knowledge_id: knowledge-ascend-foreachnorm-bank-layout
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- ForeachNorm
-condition_ids:
-- condition-cannbench-20261020-foreachnorm-bank-layout
-sources:
-- exp-source-cannbench-optimization-20261020
-source_refs:
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/foreach-late.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/foreach-qualification.md
-related:
-- exp-ascend-cb04-foreach-softmax-boundaries
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
-kernel_types:
-- reduction
----
-
-# ForeachNorm: combine semantic repair with PMU-guided accumulator placement
-
-Historical **public** reference (2026-10-16T05:33:28.798747+00:00): **83.60939044, 20/20**. This is the selected aggregate at that capture, not a claim that an earlier checkpoint below produced the latest score. Keep the source-specific public baseline and its code identity.
-
-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
-
-CANN9.1 Count API, local910B2C; FP32 arithmetic includes BF16 and non-HalfPath FP16. The mixed FP16 P2 route remains separate.
-
-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:** Early wide-fold candidates passed public tests but retained p=-Inf neutral, p0/subnormal and generic-NaN defects. Fusing Mul+Add alone increased measured bank-group conflict and slowed the candidate.
-
-**Change:** Start from FN711 semantic fixes: +Inf minimum neutrals; p0 by raw-word nonzero after sign removal; literal golden for generic/singleton/nonfinite orders; dtype grouping with original output order; materialize supported views. FN716 adds256B to accBuf and offsets its effective float pointer128B around the measured MulAddDst path, preserving tile8192 and proving every count/tree access fits.
-
-**Mechanism:** Fewer arithmetic calls need not be faster if operand placement adds bank conflicts. A PMU-supported placement hypothesis plus the full legal/numeric contract is stronger than an API-count guess.
-
-## Recorded results
-
-FN716 recorded1453 fixtures/3553 outputs and whole20 pairs+.135962077862/+.488786521320/+.114903941211 versus immutable FN711,120/120. Source ZIP8353df22… and loaded SO3b8fca79… are bound in the qualification receipt. FN711 itself had negative formal deltas; its repairs alone were not a score win.
-
-## Boundaries, negative evidence and reuse
-
-FN707 handoff was explicitly withdrawn after36 supplemental failures. FN715 was slower twice. No measured physical UB addresses or fixed bank-number proof exists. Placement can move subsequent buffers; preserve ranges and PMU controls. Legacy checked messages for skipped outputs or approximate infinity sentinels do not prove semantics.
-
-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
-
-- [foreach-late](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/foreach-late.md): exact input SHA256 and original line ranges.
-- [foreach-qualification](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/foreach-qualification.md): exact input SHA256 and original line ranges.
-- [Original operator chronicle](../../../corpus/ascend/cannbench/optimization-history-20261020/operator-records/foreach_norm.json): preserved source text with its original paths and dates.
-
-Prior knowledge: `exp-ascend-cb04-foreach-softmax-boundaries`.
-
-[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/gather-packed-indices.md

@@ -1,96 +0,0 @@
----
-id: exp-ascend-gather-packed-indices
-title: 'Gather: compact transpose indices and preserve raw payload words'
-type: experience
-vendor: ascend
-probe_slug: gather-packed-indices
-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:
-- Gather
-- gather
-- 收集
-- 索引
-- C2C4R
-- H13
-- G902
-- Or16
-- exp-ascend-cannbench-20261020-gather-packed-indices
-- wiki/ascend/experience/cannbench-20261020-gather-packed-indices.md
-confidence: source-reported
-evidence_level: measured
-cann_version: unknown
-scope: Recorded tsplit/int64-payload routes; source-bound standard A1+F18+C2+C4R and independently ported hidden H13. CANN9.1, local910B2C.
-knowledge_id: knowledge-ascend-gather-packed-indices
-knowledge_id_source: agent-semantic-link
-knowledge_scope: operator-local
-knowledge_page_role: operator-instance
-verified_operators: []
-observed_operators:
-- Gather
-condition_ids:
-- condition-cannbench-20261020-gather-packed-indices
-sources:
-- exp-source-cannbench-optimization-20261020
-source_refs:
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/gather-handoff.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/gather-current.md
-- source_id: experience/cannbench-optimization-20261020
-  path: evidence/latest-offline.md
-related:
-- exp-ascend-cb04-gather-transpose-layout
-- skill-ascend-discrete-memory-access
-evaluation_scope: public
-verification_status: historical-source-review; no independent rerun in this import
----
-
-# Gather: compact transpose indices and preserve raw payload words
-
-Historical **public** reference (2026-10-16T05:33:28.798747+00:00): **79.36086055, 20/20**. This is the selected aggregate at that capture, not a claim that an earlier checkpoint below produced the latest score. Keep the source-specific public baseline and its code identity.
-
-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
-
-Recorded tsplit/int64-payload routes; source-bound standard A1+F18+C2+C4R and independently ported hidden H13. CANN9.1, local910B2C.
-
-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:** Transpose staging stored duplicated int64 index words; paired payload conversion constructed a second full index representation.
-
-**Change:** C2 stores only low int64 index words in compact transpose rows without changing tile geometry. C4R removes duplicate paircast and maps both int64 payload words to the original raw index even row using idxRowScale=0. Check signed bounds and exact payload words, then port and measure each actual side parent independently.
-
-**Mechanism:** Reduce index preparation/storage and full-call conversion work while keeping both words of an int64 payload tied to the same logical source index.
-
-## Recorded results
-
-Three fresh whole20 standard pairs+.170968066276/+.106729839357/+.335978192424. C4R target about21% faster, C2 target about3–4%. Exact final domain comparison84955518 words, zero mismatches. Later formal history records standard≈+.21339208 and hidden+.127628595148083 from separate frozen sides.
-
-## Boundaries, negative evidence and reuse
-
-The standard package lacked several H13 routes and was not initially qualified for hidden. CPU floating Gather canonicalized BF16 NaN0x7fc0→0xffff while NPU preserved bits; use integer-word oracle. Later G902 Or32-prefix failed semantically; Or16(count*2) repaired it, but target≈1.5x did not exceed whole20 AA noise and was not promoted. Include transpose, cast, unpack and workspace in timing.
-
-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
-
-- [gather-handoff](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/gather-handoff.md): exact input SHA256 and original line ranges.
-- [gather-current](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/gather-current.md): exact input SHA256 and original line ranges.
-- [latest-offline](../../../corpus/ascend/cannbench/optimization-history-20261020/evidence/latest-offline.md): exact input SHA256 and original line ranges.
-- [Original operator chronicle](../../../corpus/ascend/cannbench/optimization-history-20261020/operator-records/gather.json): preserved source text with its original paths and dates.
-
-Prior knowledge: `exp-ascend-cb04-gather-transpose-layout`, `skill-ascend-discrete-memory-access`.
-
-[Public operator index](cannbench-public-experiences.md) · [source record](../../../sources/experience/cannbench-optimization-history-20261020.md)