What problem does it solve? TileLang kernels on Ascend NPUs often underperform due to serial memory transfers, unvectorized scalar loops, redundant data loads, and poor Cube/Vector core coordination. This Skill provides a systematic optimization playbook that maps operator types (Cube, Vector, CV-fused) to concrete optimization techniques with before/after code patterns. ## Core Features & Use Cases - Intra-core optimization: Split-K partitioning, Double Buffer (Ping-Pong), MTE2 prefetch, Full-Load L1 residency, small-block merged loading, instruction vectorization, AXPY instruction fusion, and sparse memory access optimization for Gather workloads. - Inter-core optimization: num_stages tuning for T.Pipelined, cross-core synchronization interval adjustment, and Fixed Core mode launching by physical core count to reduce workspace memory. - Decision guidance: Optimization priority ordering by operator type (detected via IS_ASCEND_AIC / IS_ASCEND_AIV), Developer vs Expert mode selection, and a troubleshooting table mapping symptoms like Cube bubbles or memory overflow to fixes. - Use Case: A GEMM kernel on Atlas A2 shows low MTE2 bandwidth utilization; apply Split-K with Double Buffer and Full-Load to overlap L1-to-L0 transfers with MMA computation. ## Quick Start Ask the agent to analyze my TileLang Ascend kernel source and apply the appropriate intra-core and inter-core optimizations based on its operator type.
