perf-optimization-audit

name	perf-optimization-audit
description	Audit Tensix/SFPU LLK compute kernels for PERFORMANCE — unfilled latency shadows/bubbles and redundant NOPs, redundant Dst/LReg store-load traffic, loop-invariant work, predication that should be branchless arithmetic (min/max/abs/setsgn), un-fused mul+add, ignored APPROXIMATION_MODE, and unroll/register-pressure mistakes. Use after touching any ckernel_sfpu_*.h, hand-written TTI_SFP*/TTI_* sequence, or the compute inner loop. This is a PERF audit (wasted cycles), NOT a correctness/race audit — pair it with instruction-latency-audit.
user_invocable	true

/perf-optimization-audit — Tensix/SFPU compute-kernel performance review

Ground-truth precedence: the live sources — the pinned sfpi-gcc (latencies / throughput / xtt_delay / what the compiler already schedules) and the tt-isa-docs MCP VectorUnit.md (latency + reciprocal-throughput — WH/BH only; Quasar is NOT in tt-isa-docs, ground QSR HW timing via race-audit-all's Quasar ladder) — outrank every rule, table, and example baked into this skill (treat those as dated illustrations). If a live source contradicts a baked rule here, do NOT silently proceed: surface the conflict to the user and ask whether the baked rule should be overwritten, discarded, or kept. Default to the live source.

MANDATORY — before any verdict, read the shared grounding policy. The per-architecture source ladder (which docs to consult), the ground-or-abstain rule, and the Source preflight (list the sources you'll consult with their reachability + hierarchy, then PAUSE for the user) are defined once in race-audit-all → .claude/skills/race-audit-all/SKILL.md. Your FIRST action is to Read that file and follow its "Ground-truth source ladder", "Ground-or-abstain", and "Source preflight" sections — a perf verdict on a stale/guessed latency or throughput number is worthless. Timing authority = pinned sfpi-gcc (all archs) + VectorUnit (WH/BH) — reuse the Fetch recipe in instruction-latency-audit's freshness contract. If race-audit-all can't be read, say so and abstain.

Coverage — floor, not ceiling. The grep patterns and site lists here are a seed, not an exhaustive enumeration — widen with full reasoning (semantic search by effect, resolving macros/wrappers, call-graph, and diffing WH/BH/QSR variants — often byte-identical copies, so a win in one applies to all three; a divergence is itself a signal). State residual gaps explicitly.

Execution — parallel by default. For more than a few kernels, fan out concurrent Agent calls (one per file/kernel-family, ~10–16 concurrency); synthesis stays sequential. Agents only return findings; the orchestrator is the sole writer and appends each wave incrementally. The heavyweight Workflow tool remains explicit-opt-in.

Reference map — read the file for the phase you're in

Split across references/ (each file self-contained, <4000 chars). Read as you reach each phase; never skip the provenance lens or the equivalence gate.

• references/overview-and-method.md — what this audit is / is not, the provenance lens (run FIRST — decides which findings are valid), and the 6-step method. Read before starting.
• references/checks-traffic-loops-shadows.md — catalogue A (Dst/LReg traffic), C (loop & template), D (latency shadows — raw TTI_* only), F (TT_→TTI_ encoding).
• references/checks-selection-and-fusion.md — catalogue B (instruction selection / strength reduction — the biggest wins) and E (fusion / reconfig above the loop).
• references/equivalence-guards-verdict.md — the semantic-equivalence gate, the false-positive guards, and the verdict vocabulary.
• references/validation-and-output.md — disassembly proof (sfpi + objdump), perf-counter validation, and the output format.