The parallel batch machinery from be21d627a only ever existed to chase a
default-on multi-threaded win that never came (washer w8 still +54% with
the hybrid on). At w1 — the only configuration this opt-in flag is for —
parallel_for runs inline, so the serial path gives the identical result
for ~340 fewer lines. Wire up the previously-dead
dynamic_tree_self_pairs/cross_pairs into a serial collect_batch_candidates
(three BVTT self/cross traversals -> canonical (a,b,child) sort -> serial
filter into move_results[0]) and delete BatchWork, BatchCtx, batch_drain_*,
BatchFilterCtx, batch_filter_*, bvtt_step, dynamic_tree_bvtt_drain/expand,
and the batch_frontier/worker_* scratch fields.
Determinism preserved exactly: OFF 0x61E35C31/step314 bit-identical, ON
0xBE99C5F7/step313 identical across workers 1/2/4. The debug SET-equality
oracle and the determinism_broad_phase_hybrid_across_worker_counts test
are unchanged and still pass; zero warnings.
PGO: pgo.sh never trained -bp=1, so an off-path-only profile laid the
hybrid branch out cold and collapsed the win to ~-5%. Add one -b=8 -bp=1
training run (neutral for the default path — counts merge, the OFF branch
stays hot) and retrain.
Corrected README numbers to measured values (hybrid-trained profile,
paired -bp toggle, washer w1): pair-finding stage 7.6k -> 3.6k ms/1000
(-52%); total ~-17% (~19.0k vs ~22.6k), which beats C (20661) and narrows
Rapier's lead from ~23% to ~12% — not the "17.7k / within 5% / -19%"
be21d627a claimed. Still default-off (w8 regresses ~+54%), opt-in
single-threaded accelerator for churn-heavy scenes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
WorldDef.enable_broad_phase_hybrid (default false; -bp=0/1 bench toggle):
an adaptive batch broad phase for high-churn scenes. When
move_count*4 > proxy_count, replaces the per-moved-proxy tree queries
(8k proxies x 3 root-descents on washer) with three BVTT self/cross
traversals (dynamic self + dynamic x static + dynamic x kinematic) that
share the upper-tree descent, plus an O(n) bottom-up refit instead of
the median rebuild. Both traversal and candidate-filter are parallelized
across the task system (per-worker buffers → merge → canonical sort by
(shape_a,shape_b,child) → deterministic contact creation).
Correctness: a #[cfg(debug_assertions)] SET-equality assertion (batch
candidate set == per-mover set) runs in every test and never fires — the
proof the BVTT finds identical contacts (the hash can't prove it since
creation order legitimately re-baselines). New test
determinism_broad_phase_hybrid_across_worker_counts. OFF hash 0x61E35C31
bit-identical; ON hash 0xBE99C5F7 identical across workers 1/2/4 +
external tasks. 180/186/180/180 tests, zero warnings, profile retrained.
Single-threaded washer -18.7% (17715 vs 21780, broad phase -51%) — beats
C (20661), within ~6% of Rapier (16844). DEFAULT OFF because it regresses
multi-threaded (washer w8 +52%): the batch materializes ~40-50k
candidates/step and serially merges+sorts them (a fundamental floor the
inline per-mover path avoids by filtering in the query callback), so at
w8 the parallel per-mover queries win. Cannot be worker-gated (would
break cross-worker determinism). Correct, deterministic, zero-cost when
off — an opt-in single-threaded accelerator for churn-heavy scenes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
query_tree_for_pairs was materializing every tree-query hit into a Vec
then re-iterating; C runs try_add_pair inline in the query callback
(b3PairQueryCallback). Rewrote to filter inline like C (both world
borrows are shared, so it compiles); only the rare compound inner-query
still uses a child_hits scratch (no compounds in washer/junkyard/
pyramids/trees, so pair-discovery order is unchanged). Removed the
now-dead PairScratch.hits field.
query_tree_for_pairs is washer's single hottest symbol (8k dynamic cubes
churned by a rotating drum re-query the whole tree every step). washer
broad phase -3% (paired plain + retrained-PGO), junkyard -1.4%,
pyramids/trees neutral. Hash bit-identical (0x61E35C31) — pure
structural, same discovery order. 179/185/179/179 tests, zero warnings.
Profile retrained dual-mode.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
New middle tier between full manifold recycling and the full SAT, behind
WorldDef.enable_feature_recycling (default true; OFF path bit-identical,
hash 0x61E35C31 verified):
- Case A, separated-witness early-out: a previously-non-touching contact
revalidates only the cached winning axis; still separating beyond the
speculative distance means done in one test. Sound structurally: any
cached axis is a valid separation witness (understates only, which
falls through to the full SAT). Carried junkyard: 24k skips/step,
full SATs 25.4k -> 4.5k per step.
- Case B, touching feature rebuild: re-clips the cached winning feature
under explicit staleness bounds (SATCache::sat_pose at last full SAT,
translation < 4x recycle distance, rotation < ~4.6 deg, forced refresh
every 8 steps); degenerate rebuilds and touching<->separated
transitions fall through same-step.
Probe-driven (junkyard: 89k full SATs/step on 105k pairs, 13k touching;
rapier maintains 3.4x fewer pairs): paired same-binary -fr=0/1 A/Bs show
junkyard -8% in every pairing (collide phase -15%), washer neutral to
-5%, pyramid/rain guards neutral, OFF costs nothing. Same-session
cross-engine junkyard: rapier's -8% lead closes to ~-3%.
pgo.sh now trains BOTH modes (single-mode training starved the remaining
full-SAT path); checked-in profile retrained dual-mode. SATCache pose
serialized in snapshots. feature_recycled_contact_count in Counters.
README: port-extension subsection, soundness argument, updated notes.
Also: rapier bench --probe mode from the workload-probe session.
179/185/179/179 tests green with the tier ON, zero warnings,
determinism suite passes across runs/workers/task systems.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Disassembly attribution on current binaries: junkyard's +23% sits
entirely in the NON-SAT narrow phase (2.06x C) — the edge SAT is now
FASTER than C. Cause: LLVM+PGO mega-inlining (third occurrence) —
update_contact compiled to 5.5x C's instruction count, collide_hulls had
no symbol at all. inline(never) on collide_hulls /
compute_convex_manifold / query_face_directions restores C's layout:
junkyard -3.6% paired (retrained profile), washer neutral, others
untouched.
hull_at coverage extended to build_face_a_contact / build_polygon /
clip_segment_to_hull_face / find_incident_face for contract completeness
— measured NEUTRAL beyond the boundary fix, and the feature's earlier
-3.6% is now captured by the safe attribute instead. README documents
the demotion honestly (the safe fix superseded the unsafe one).
Gates: 179/179/185/179 tests (default/unchecked-hulls/dp/nosimd), hash
0x61E35C31 everywhere, zero warnings. PGO profile retrained.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A sweep of every C b3ParallelFor/enqueue dispatch against the port found
the finalize-bodies pass (per-body transforms, AABB updates, sleep
accounting, continuous/TOI) and the bullet pass were left serial when
threading was ported — C runs both under b3ParallelFor. rain's w=8 gap
was almost entirely this serial fraction (Amdahl decomposition showed
its parallel portion already at C parity).
FinalizeCtx mirrors the collide pass's pattern: taken arrays + SyncSlice
disjoint per-body access, per-worker task contexts, deterministic merges
(bitset OR, split-candidate max like C), bullet list via atomic cursor
(C's b3AtomicFetchAddInt mirror). No new unsafe primitives. Pre-solve/
custom-filter callbacks force single-worker like collide.
w=8: rain +42% -> +8%, joint_grid +36% -> +11%, large_pyramid and
many_pyramids and trees25 at parity; geomean +11%. Serial geomean +5%
(Rust wins joint_grid/large_pyramid/many_pyramids outright). PGO profile
retrained; README tables + narrative updated. Remaining known
serial-vs-C difference: the split-island enqueue overlap (documented).
Hash 0x61E35C31 bit-identical (runs, workers 1/2/4, external tasks);
179/185/179 tests, zero warnings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Training set was missing washer and trees50/25 (washer -1.5% with
coverage, large_pyramid unchanged — no dilution). Profile refreshed.
Small-stage main-only fast path (generalizing C's single-block
shortcut): swept cutoffs 32/64/256 at w=8 — only large_world benefited
(-8% of ~11ms); rain regressed at every cutoff (its small-count stages
are mesh-contact stages with heavy per-item cost — serializing them
starves real parallelism) and joint_grid has few fat stages (grid
coloring = 2-4 colors), so its w=8 gap is NOT thin-stage sync. Reverted
per the measurable-win rule; negative result documented in the README.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Disassembly census showed the joint solvers at exact FMA parity with C
but +110 loads/+54 stores per joint: the full 56-byte BodyState get/set
round trip keeps untouched fields live across the ~1000-instruction
solve bodies. StateAccess::set_velocities (same unsafe contract as set,
velocities only, like C's in-place stores) + get_ref field extraction
across all 16 warm-start/solve functions in the 8 joint types.
joint_grid: 817 vs C 801 ms (was -11%). Full fresh matrix in README:
serial geomean +7% vs C with Rust WINNING large_pyramid (-6%) and
many_pyramids (-3%); w=8 geomean +28%. Checked-in PGO profile retrained
for the new code (stale profile cost ~13% on joint scenes).
Same change was measured neutral for contact scatter and correctly
dropped there (state live ~40 instrs vs ~1000) — both verdicts in the
README as a paired case study.
Hash 0x61E35C31 bit-identical everywhere; 179/185/179 tests, zero
warnings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
cargo build --release on anything in the workspace now gets the PGO'd
box3d automatically (-Cprofile-use=libs/box3d/box3d.profdata; verified:
default build runs at the explicit-PGO binary's speed). The profile is
target-independent — x86_64 cross-build with the ARM-trained profile
compiles clean — and degrades gracefully when stale (unmatched functions
keep normal heuristics). Retrain with libs/box3d/pgo.sh.
Fresh four-way interleaved matrix in the README: default box3d is now
faster than or equal to rapier-simd on all three scenes (1118 vs 1451,
1510 vs 1690, 912 vs 914 ms) and faster than non-PGO C on two of three
(C keeps joint_grid 816 vs 912). 179/185/179 tests green with the
config active.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
