What problem does it solve? CATLASS matmul and fused kernels on Ascend NPU often underperform due to suboptimal tiling, full-tile C HBM round-trips, and poor AIC/AIV pipeline overlap, and diagnosing these bottlenecks requires deep knowledge of profiler signals and hardware constraints. ## Core Features & Use Cases - Parameter Tuning Guidance: Explains the four tunable levers—Kernel type (C writeback strategy), DispatchPolicy (Pingpong vs Preload), L1/L0 TileShape, and BlockScheduler Swizzle—with capacity constraints and empirical starting points for fp16 and int8. - Bottleneck Diagnosis Tables: Maps msprof profiler symptoms (HBM bandwidth, MTE2 ratio, Cube utilization, AIV stalls) to root causes and prioritized fixes, including specialized tables for Linear Attention, KDA varlen, and FlashAttention stage tuning. - Single-Variable Tuning Discipline: Enforces one-change-at-a-time attribution with immediate rollback on regression, plus code diff patterns for switching Kernel writeback modes and DispatchPolicies. - Use Case: When a fused matmul_gelu kernel underperforms the torch baseline on large N shapes, use this Skill to identify the full-tile C HBM round-trip as the bottleneck and switch to a multi-stage rotating workspace kernel with PreloadAsyncWithCallback. ## Quick Start Ask the assistant to load the catlass-op-perf-tune skill and diagnose why my CATLASS matmul kernel is slower than the torch baseline using the profiler data.
