What problem does it solve? When a Triton Ascend kernel fails verification, raw verifier output (precision metrics, per-dimension error ranges, sample coordinates) is hard to interpret. This Skill turns that output into a concrete root-cause hypothesis—precision, mask, index, boundary, NaN/Inf, or store errors—so you fix the right code instead of guessing. ## Core Features & Use Cases - Precision line parsing: Interprets [precision] fields (hard, outlier, strict, mere, mare) and the strict/relaxed tolerance formulas to classify failure type before touching thresholds. - Per-dimension error analysis: Reads Error location per dimension ranges and coverage to distinguish boundary/tail bugs, tile mapping errors, and global formula or store mistakes. - Sample-value heuristics: Maps patterns like impl=0, magnitude mismatches, sign flips, and repeated values to likely causes such as mask over-restriction, stride errors, or buffer overwrite. - Use Case: A kernel fails with hard=254 and errors only in dim2: [4:5]; the Skill directs you to check that dimension's offset, stride, and tail mask rather than rewriting accumulation logic. ## Quick Start Ask the agent to diagnose the failed Triton Ascend kernel verification using the verifier log's precision line, per-dimension error distribution, and sample coordinates.
