What problem does it solve? When writing or optimizing AscendC kernels for Ascend NPU backends, choosing the wrong initial kernel structure wastes iteration cycles. This Skill provides ready-made implementation templates for common operator families so an agent or developer can pick a correct skeleton before tuning performance. ## Core Features & Use Cases - Operator Family Templates: Provides skeletons for elementwise, broadcast, reduction, softmax-like, gather/scatter indexed, and matmul-with-epilogue operators. - Shape-Based Fast-Path Classification: Includes a host-side pattern classification scheme (same-shape, scalar broadcast, last-dim broadcast, small reduction, single-tile reduction) with a generic fallback, plus guidance on keeping mode branches outside the inner loop. - Numerical and Tiling Rules: Documents UB capacity limits, tail masking, fp32 accumulation for fp16/bf16 reductions, softmax numerical stability, and Cube/Vector load balancing for matmul epilogues. - Use Case: During a batch kernel-optimization run for a new reduction operator, use this Skill to decide between per-core row reduction, segmented multi-core reduction with workspace merge, or a single-tile UB path before writing any device code. ## Quick Start Ask the agent to use the ascendc-op-patterns skill to choose an initial kernel skeleton for your operator's shape and semantics before editing the kernel math.
