What problem does it solve? Ascend C operator kernels often underperform without a clear reason, and blindly tweaking tile sizes wastes time. This Skill classifies the kernel's bottleneck from profiling or simulation traces (VEC bound, memory bound, scalar bound, or no bound) and loads the matching optimization playbook, plus a structured tiling parameter-space analysis method. ## Core Features & Use Cases - Bound-type routing: Maps profiling/simulation evidence to one of four bottleneck categories, each with severity grading, trace signatures, and concrete fixes such as DoubleBuffer, Cast reduction, fused instructions, and L2 reuse. - Tiling parameter-space analysis: A four-stage method (parameter inventory, type/constraint tracing, algorithm enablement, candidate space construction) that separates hardware constraints from software heuristics before any search. - Structural optimization guidance: Covers sample bucketing by dtype/rank/broadcast/reduce-axis, scalar small-D paths, batched CopyOut, and branch hoisting for cases where parameter tuning alone yields no gains. - Use Case: Given a slow elementwise kernel whose trace shows 65% VECTOR utilization with heavy Cast instructions, the Skill routes to the VEC-bound playbook and recommends eliminating fp16/fp32 round-trip casts and fusing vector operations. ## Quick Start Ask the agent to analyze this Ascend C kernel's profiling trace, determine its bound type, and apply the matching optimization strategy from the ascendc-perf-optimize knowledge base.
