triton-ascend-case-vector-elemwise-bench-atlas-a2

What problem does it solve? Triton kernels on Ascend Atlas A2 hardware often use mathematically correct but slow API calls for common elementwise operations like silu, tanh, sigmoid, sqrt, and relu. This Skill provides benchmarked, semantically equivalent API replacements with measured end-to-end time gains so you can speed up kernels without changing their math. ## Core Features & Use Cases - Benchmarked API substitution tables: Covers fp32, fp16, and bf16 dtypes with per-operator end-to-end time (ms) measurements on Atlas A2, showing gains from 4% up to 47%. - Precision-aligned replacements: Every suggested replacement (e.g., tl.sqrt(x) -> tl.sqrt_rn(x), tl.exp2(x) -> tl.exp(x * LN2)) is verified as semantically equivalent and precision-aligned in single-operator unit tests. - Use Case: While optimizing a Triton kernel for Ascend, you notice tl.rsqrt(x) in an fp16 path. Consulting this Skill, you replace it with 1.0 / tl.sqrt(x) for a measured 24% speedup, then re-run precision checks to confirm alignment. ## Quick Start Ask the agent to review your Triton kernel's elementwise tl.* calls and apply the recommended Atlas A2 substitutions for the target dtype, keeping the original code if precision no longer aligns.