What problem does it solve? It provides QA with an objective acceptance standard for the CP4 performance checkpoint, determining whether an Ascend C operator's measured performance meets the targets declared in the requirements document, based on data rather than subjective judgment. ## Core Features & Use Cases - Dual evaluation modes: Mode A uses evaluation-set HAP scores (default threshold ≥ 0.5) as the final verdict; Mode B compares self-collected msprof op metrics (latency, bandwidth, AI Core utilization) against quantified targets. - Bottleneck verification: Requires bottleneck-dimension decomposition (compute / move-in / move-out / scalar) and cross-checks measured bottlenecks against the development plan's predictions, flagging contradictions. - Leverage assessment gate: When targets are unmet, it enumerates untried optimization levers and excludes disproven directions before deciding whether to roll back or recommend closing as a known limitation. - Use Case: After functional acceptance of a BatchMatmulMaxSum operator, QA loads this skill to run the evaluation set, compare HAP scores per case against the threshold, verify mandatory optimization items landed, and produce a structured CP4 performance acceptance report. ## Quick Start Load the workflow-cp4 skill and run the CP4 performance acceptance for the operator submission, producing a pass/fail report with bottleneck analysis.
