[shinaoka]

Part of #6 (stage 1: proposals 1 and 2). Stage 2 (the broccoli Zen2 row from tensor4all/faer#91) and the faer migration come in separate PRs.

Summary

• Linear extrapolation above the largest anchor. Above w_m, WidthCosts::at(w) is cost(w_m) * w / w_m (saturating) instead of the flat cost(6). A region that does not pay at 6 therefore does not pay at 64.
  • Widths up to the largest anchor keep the existing model: linear interpolation, and the first value below the first anchor.
  • Every existing row is bit-identical at widths 1 to 6, pinned by a test.
• Variable anchors. WidthCosts::from_anchors(&[(width, Duration)]) -> Option<Self> takes up to WidthCosts::MAX_ANCHORS (12) strictly increasing anchors with widths from 2 to u16::MAX.
  • It is const (usable in a const item), allocation-free, and never panics.
  • It returns None for empty input, too many anchors, a width below 2, or widths that are not strictly increasing.
  • new([2, 4, 6]) and flat keep their signatures.
• Storage. WidthCosts now stores its anchors (widths u16, ns u64, length) and interpolates at lookup, with a 64-bit fast path that falls back to 128 bits only when a product overflows.
  • A 64-entry table would have made DispatchCosts about 1.5 KB. TeamExec and faer copy it by value.
• AVX-512 team row gains its measured width-8 anchor (0.98 us), from the same post-fix run as its 2/4/6 values (docs/design/thresholds.md). Without it the extrapolation would charge 1.2 us at width 8 on the measured 8-thread host.
  • The in-pool and outside width-8 numbers come from the pre-fix run and are not used.
• Docs. WidthCosts, for_class, README, docs/design/placement.md and docs/design/thresholds.md now describe the new contract. The thresholds doc keeps one historical note on why the model was flat before.
• threadteam 0.1.1 for the approved patch release. The decisions above width 6 change by design, no API is removed, and the constructor is additive.

Done-when checks (from #6)

• A flat 1 us row gives 1, 2, 4 and 8 us at widths 6, 12, 24 and 48. Above w_m the row is non-decreasing and at least linear, checked on five rows.
• A region that does not pay at 6 does not pay at 64.
• Interior width: width_for_by(24 us, 1, budget 64) picks 12 on the 1-us row, and a large enough estimate picks 64.
• Anchors at 2, 4, 8, 16, 32 and 64 interpolate and extrapolate as specified.
• from_anchors rejects invalid input and works in a const.
• Equality ignores unused capacity, and a Duration::MAX row never pays.
• The 128-bit fallback gives the same values as the 64-bit path, and falling interpolation truncates toward zero.
• tests/dispatch_cost.rs now expects the extrapolated values instead of the flat ones. tests/alloc_free.rs and tests/team_alloc_free.rs pass.

Lookup cost

Measured with benches/cost_lookup.rs, extended in this PR. Same bench file on base f29c8ce and on this branch, taskset -c 56-59 on primerose, best of 9 x 2M calls, ns per call:

	base	this PR
at (widths 2..8)	0.89	4.12
pays	1.97	7.84
width_for budget 8	25.9	61.8
width_for budget 64	222	522
min_cost	6.7	4.8
copy DispatchCosts	4.5	11.0
Exec::dispatch_cost (team)	3.3	6.3
Exec::width_for (budget 8)	25.5	63.5

size_of: WidthCosts 48 -> 128 B, DispatchCosts 144 -> 384 B. The extra cost is far below one section, which the rows charge at 0.6 us or more.

Verification

The AGENTS.md pre-push checklist passes locally:

• cargo fmt --check
• cargo clippy --all-targets --all-features and --no-default-features, both with -D warnings
• cargo test, --all-features and --no-default-features
• RUSTFLAGS="--cfg loom" cargo test --release --test loom_park (the park/wake core is untouched)
• cargo deny check bans
• scripts/check-no-public-rayon.sh
• cargo doc --no-deps

Independent pre- and post-implementation reviews (gpt-6.1-sol) were done. The post-review had no BLOCKER or IMPORTANT findings, and its minor ones are fixed.

Downstream

tensor4all/faer mirrors the flat model for InPool and Outside and asserts equality with threadteam through width 12. Those assertions change when faer moves to 0.1.1, which is the stage-4 PR that also deletes team_section_cost.

🤖 Generated with Claude Code