Perf Regression Check

Perf compare (one test, two commits)

perf_compare_commits.sh does the whole job: resolve both refs, sweep each, compare median-vs-median, write the report. Do not hand-roll the checkout/sweep loop — the script already handles worktrees, caching, interleaving and cleanup.

SCRIPT="$(git rev-parse --show-toplevel)/tt_metal/tt-llk/.claude/scripts/perf_compare_commits.sh"

# branch HEAD vs the commit the branch was cut from (the default)
"$SCRIPT" <arch> <test>

# any two commits
"$SCRIPT" <arch> <test> --baseline <ref> --current <ref>

It prints the report, writes it under ~/.cache/tt-llk-perf-compare/reports/…, and exits non-zero when a regression is found. --dry-run shows what it would measure (and what is already cached) without running a sweep — cheap, use it to confirm the refs first.

Inputs to collect from the user (ask only what you cannot infer)

• test (required): the perf test module, e.g. perf_math_matmul. One test per run.
• arch (required): wormhole or blackhole. Infer from the machine (tt-smi) if possible.
• baseline / current: any refs. Default: baseline = git merge-base origin/main HEAD, current = HEAD. If the user names commits ("compare abc123 and def456"), pass both.
• threshold: --threshold, default 0.05 (5%).
• iterations: --iterations, default 3 per side.
• speed of light: --speed-of-light. The script applies it to both sides.

What the script guarantees (so you do not have to)

• Your checkout is never touched. Each commit is a separate sparse git worktree with its own build tree — no branch switch, no stash, a dirty tree is fine, and an interrupted run cannot leave the user on a detached HEAD. Worktrees are removed on exit.
• Only committed code is measured. Uncommitted edits are invisible to a ref; the script says so. If the user wants their working tree measured, they must commit it first.
• Runs are cached per (arch, test, variant, commit), so a commit already measured on this machine costs nothing to reuse — comparing A-vs-B then B-vs-C only sweeps C. --refresh re-measures. Cached runs from another host are refused (cycles are not comparable).
• Iterations are interleaved baseline, current, baseline, current…, so machine drift hits both sides equally instead of biasing one.

Present the result

Show the verdict, the top regressions, and the improvements if any. Point at regression_report.md plus its companions .regressions.csv (every regression) and .points.csv (every compared point, full config).

Interpreting results (tell the user)

• Comparison is per (marker, sweep-config) on mean(<run_type>) cycles, median across iterations. A point exists on both sides or it is a "new point" (no baseline).
• Cleanest when only kernel/LLK code differs between the commits. If the test's sweep changed, expect "new points" — those are reported, never counted as regressions.
• Noise: 3 iterations + median reduces it, but deltas near the threshold can still be noise. Suggest more iterations (cheap — cached runs are reused, so --iterations 5 after a 3-iteration run only measures 2 more per side) or a higher threshold.
• The further apart the two commits are, the more unrelated change is in the delta. For "did this work regress perf", the merge-base default is the honest baseline.
• TILE_LOOP is the steady-state per-tile cost; INIT/KERNEL include one-time overheads.

Failure handling

• TENSIX TIMED OUT / device hang: tt-smi -r, then re-run the same command — completed iterations are cached, so it resumes instead of starting over.
• Missing header at compile time on an older commit: that commit's layout needs more than the default sparse checkout — re-run with SPARSE_PATHS= (empty) for a full checkout.
• "no perf CSV": the test may have been deselected — check the module name and that the test carries the perf marker.
• "does not exist at ": the test was added after that commit; pick a baseline that has it.