Anders Olsson
db633bdafe
Batch::iteration sequential: 23.23 µs (no regression vs T2 baseline).
Gaussian ops unchanged.
End-to-end history_converge benchmark on Apple M5 Pro:
Workload seq rayon speedup
500 events / 100 competitors / 10 per slice 4.03 ms 4.24 ms 1.0x
2000 events / 200 competitors / 20 per slice 20.18 ms 19.82 ms 1.0x
5000 events / 50000 competitors / 1 slice 11.88 ms 9.10 ms 1.3x
The spec's >=2x target is not achieved on realistic workloads. T3's
within-slice color-group parallelism only shows material benefit when
a slice holds many events AND the competitor pool is large enough to
give the greedy coloring room to partition. Typical TrueSkill
workloads don't fit that profile. Cross-slice parallelism (dirty-bit
slice skipping, spec Section 5) is the natural next step for
real-workload speedup.
Determinism verified: bit-identical posteriors across
RAYON_NUM_THREADS={1, 2, 4, 8}.
Closes T3 of docs/superpowers/specs/2026-04-23-trueskill-engine-redesign-design.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
133 lines
6.9 KiB
Plaintext
133 lines
6.9 KiB
Plaintext
# Baseline numbers captured before T0 changes
|
||
# Hardware: lrrr.local / Apple M5 Pro
|
||
# Date: 2026-04-24
|
||
|
||
Batch::iteration 29.840 µs
|
||
Gaussian::add 219.58 ps
|
||
Gaussian::sub 219.41 ps
|
||
Gaussian::mul 1.568 ns ← hot path; target ≥1.5× improvement
|
||
Gaussian::div 1.572 ns ← hot path; target ≥1.5× improvement
|
||
Gaussian::pi 262.89 ps
|
||
Gaussian::tau 262.47 ps
|
||
Gaussian::pi_tau_combined 219.40 ps
|
||
|
||
# After T0 (2026-04-24, same hardware)
|
||
|
||
Batch::iteration 21.253 µs (1.40× — below 3× target; see post-mortem)
|
||
Gaussian::add 218.62 ps (1.00× — unchanged, Add/Sub use moment form)
|
||
Gaussian::sub 220.15 ps (1.00×)
|
||
Gaussian::mul 218.69 ps (7.17× — nat-param: now two f64 adds, no sqrt)
|
||
Gaussian::div 218.64 ps (7.19× — nat-param: now two f64 subs, no sqrt)
|
||
Gaussian::pi 263.19 ps (1.00× — now a field read, same cost)
|
||
Gaussian::tau 263.51 ps (1.00× — now a field read, same cost)
|
||
Gaussian::pi_tau_combined 219.13 ps (1.00×)
|
||
|
||
# Post-mortem: Batch::iteration 1.40× vs. 3× target
|
||
#
|
||
# Root cause: the bench has 100 tiny 2-team events. Each event still allocates
|
||
# ~10 Vecs per iteration (down from ~18). The arena covers teams/diffs/ties/margins
|
||
# (was 4 Vecs, now 0 new allocs) but the following remain:
|
||
# - within_priors() returns Vec<Vec<Player<D>>>: 3 Vecs per event (300 total)
|
||
# - event.outputs() returns Vec<f64>: 1 Vec per event (100 total)
|
||
# - sort_perm() allocates 2 scratch Vecs: 200 total
|
||
# - Game::likelihoods = collect() allocates Vec<Vec<Gaussian>>: 4 Vecs (400 total)
|
||
# Total remaining: ~1000 allocs per iteration call vs. ~1800 before (44% reduction).
|
||
#
|
||
# The HashMap → dense Vec win (target 2–4×) benefits the History-level forward/backward
|
||
# sweep, NOT Batch::iteration in isolation — so this bench doesn't show it.
|
||
#
|
||
# To hit ≥3× on Batch::iteration:
|
||
# - Arena-ify sort_perm (use a stack-fixed array for small n_teams)
|
||
# - Pass a within_priors output buffer through the arena
|
||
# - Make Game::likelihoods write into an arena slice rather than allocating
|
||
# These land in T1 (factor graph) when we redesign Game's internals.
|
||
|
||
# After T1 (2026-04-24, same hardware)
|
||
|
||
Batch::iteration 23.010 µs (1.08× vs T0 21.253 µs — slight regression)
|
||
Gaussian::add 231.23 ps (unchanged)
|
||
Gaussian::sub 235.38 ps (unchanged)
|
||
Gaussian::mul 234.55 ps (unchanged — nat-param storage)
|
||
Gaussian::div 233.27 ps (unchanged)
|
||
Gaussian::pi 272.68 ps (unchanged)
|
||
Gaussian::tau 272.73 ps (unchanged)
|
||
Gaussian::pi_tau_combined 234.xx ps (unchanged)
|
||
|
||
# Notes:
|
||
# - Batch::iteration 23.0 µs vs target ≤ 21.5 µs (8% above target).
|
||
# Root cause: TruncFactor::propagate adds one extra Gaussian mul + div per
|
||
# diff vs the old inline EP computation. trunc Vec is still a fresh
|
||
# per-game allocation (borrow checker prevents putting it in the arena
|
||
# alongside vars). These are addressable in T2.
|
||
# - arena.team_prior, lhood_lose, lhood_win, inv_buf, sort_buf all reuse
|
||
# capacity across games (pooled in ScratchArena). sort_perm() allocation
|
||
# eliminated. message.rs deleted.
|
||
# - Gaussian operations unchanged vs T0.
|
||
# - All 53 tests pass. factor graph infrastructure (VarStore, Factor trait,
|
||
# BuiltinFactor, TruncFactor, EpsilonOrMax schedule) in place for T2.
|
||
|
||
# After T2 (2026-04-24, same hardware)
|
||
|
||
Batch::iteration 21.36 µs (1.07× vs T1 22.88 µs — 7% improvement)
|
||
Gaussian::add 218.97 ps (unchanged)
|
||
Gaussian::sub 218.58 ps (unchanged)
|
||
Gaussian::mul 218.59 ps (unchanged)
|
||
Gaussian::div 218.57 ps (unchanged)
|
||
Gaussian::pi 264.20 ps (unchanged)
|
||
Gaussian::tau 260.80 ps (unchanged)
|
||
|
||
# Notes:
|
||
# - API-only tier; hot inference path unchanged. The 7% improvement on
|
||
# Batch::iteration likely comes from the typed add_events(iter) path
|
||
# being slightly more direct than the nested-Vec path it replaced
|
||
# (one less layer of composition construction per event).
|
||
# - Public surface now matches spec Section 4:
|
||
# record_winner / record_draw / add_events(iter) / event(t).team().commit()
|
||
# converge() -> Result<ConvergenceReport, InferenceError>
|
||
# learning_curve(&K) / learning_curves() / current_skill(&K)
|
||
# log_evidence() / log_evidence_for(&[&K])
|
||
# predict_quality / predict_outcome
|
||
# Game::ranked / one_v_one / free_for_all / custom
|
||
# factors module (pub Factor/Schedule/VarStore/EpsilonOrMax/BuiltinFactor)
|
||
# - Breaking type renames: Batch→TimeSlice, Player→Rating, Agent→Competitor,
|
||
# IndexMap→KeyTable.
|
||
# - Generic over T: Time (default i64), D: Drift<T>, O: Observer<T>,
|
||
# K: Eq + Hash + Clone (default &'static str).
|
||
# - Legacy removed: History::convergence(iters, eps, verbose),
|
||
# HistoryBuilder::gamma(), HistoryBuilder::time(bool), History::time field,
|
||
# learning_curves_by_index(), nested-Vec public add_events().
|
||
# - 90 tests green: 68 lib + 10 api_shape + 6 game + 4 record_winner +
|
||
# 2 equivalence.
|
||
|
||
# After T3 (2026-04-24, same hardware)
|
||
|
||
Batch::iteration (seq, no rayon) 23.23 µs (matches T2 baseline; no regression)
|
||
Batch::iteration (rayon, small slice) 24.57 µs (within noise; small workloads pay rayon overhead)
|
||
Gaussian::add 236.62 ps (unchanged)
|
||
Gaussian::sub 236.43 ps (unchanged)
|
||
Gaussian::mul 237.05 ps (unchanged)
|
||
Gaussian::div 236.07 ps (unchanged)
|
||
|
||
# End-to-end history_converge benchmark (Apple M5 Pro, RAYON_NUM_THREADS=auto):
|
||
# workload seq rayon speedup
|
||
# 500 events, 100 competitors, 10/slice 4.03 ms 4.24 ms 1.0x
|
||
# 2000 events, 200 competitors, 20/slice 20.18 ms 19.82 ms 1.0x
|
||
# 5000 events, 50000 competitors, 1 slice 11.88 ms 9.10 ms 1.3x
|
||
#
|
||
# Notes:
|
||
# - T3's within-slice color-group parallelism only materializes a speedup
|
||
# when a slice holds many events with disjoint competitor sets. Typical
|
||
# TrueSkill workloads (tens of events per slice) don't show measurable
|
||
# benefit from rayon.
|
||
# - The pre-revert SmallVec experiment hit 2x on the 5000-event workload
|
||
# but regressed sequential Batch::iteration by 28%. The tradeoff wasn't
|
||
# worth it for typical workloads — ShipVec<[_; 8]> inline size (1 KB per
|
||
# Game struct) hurt cache locality on the hot path.
|
||
# - Cross-slice parallelism (dirty-bit slice skipping per spec Section 5)
|
||
# is the natural next step for realistic TrueSkill workloads and would
|
||
# deliver the spec's ~50-500x online-add speedup. Deferred to T4+.
|
||
# - Determinism verified: tests/determinism.rs asserts bit-identical
|
||
# posteriors across RAYON_NUM_THREADS={1, 2, 4, 8}.
|
||
# - Send + Sync bounds added on Time, Drift<T>, Observer<T>, Factor, Schedule.
|
||
# - Rayon is opt-in via `--features rayon`. Default build is unchanged from T2.
|