- AUDIO_TIMING logs voices / instances / grains per block (tools/audio-bench summary prints them).
- The render path is allocation-free in real play: subscriber lists and constructor lists read in place, call
parameters on the stack, instance memory and client lists from pools (render allocations per block 0.17-0.63 ->
<= 0.001; full world 1.825 -> 0.004); render_alloc tests cover walks, posts, calls, globals and a retail world host.
- Voices: direct PCM reads when a block provably stays inside the sample; shared scratch planes in the mixer
(full world render p50 -11 to -14 %).
- MixMap keeps its held-input list between ticks; Jitter processes without a per-frame Vec.
- Every change byte-identical in all 84 e2e outputs; two attempted changes dropped (no real gain).
- Ped speech plays: speech manager -> library -> the living world's two streams, with queue / interrupts and the
near / far lines by distance; levels and filters from the retail mechanism (filters equal to the recomp's, gain
median within 0.5 %). Silent and logged when the opt-in speech decode is absent.
- NPC skater: wheels and clothing on its own instance and a bail grunt; tricks, treatment and footsteps stay off for
NPCs as in retail.
- Per-model ped data (shoe class, voice, far threshold) and traffic engine records per model exported by setup;
traffic relative-speed input and front / rear sound points; world hosts process before the mixer tick.
- Test mod: opt-in (off by default), higher world-object limits (48 per mod, 128 total), a per-second summary log.
- Specs published in docs/hails-additions/audio-specs/; code comments point at them; test data paths behind env vars
(no .local defaults).
Review fixes:
- The audio host takes each physics step exactly once (one-step pulses latched, nothing while paused or silenced,
at most 4 steps per frame), so pushes and other one-frame events are never dropped or doubled. e2e at 60 fps and
above is byte-identical; 30 / 45 fps change only by the latch and the step pair.
- Removed the emitter PROFILES fallback, the stray trace dump and the audition script; data-gated tests are
#[ignore = "needs the private install data"] and fail loudly when run without it; per-post log lines behind
SKATE_AUDIO_TRACE; checked bank parsing, finite output, camera-cut Doppler reset, fewer render allocations.
- The 15 verdict-settled A/B switches are removed; their "on" behaviour is the only one (byte-identical).
World / NPC audio, hookable by engine systems and mods:
- Map-change fix: world and NPC hosts reset when the map's banks are unloaded, so surviving objects sound again.
- Engine surface (skate-game::world_audio): TrafficAudio / PedAudio / NpcSkaterAudio components, horn / alarm /
speech messages, read-back of who is audible; retail limits (4 cars within 40 m, 15 peds within 50 m with
footsteps for the nearest 3, one NPC skater within 30 m), measured in the recompilation.
- Mod surface (Lua API 2, `sdk.world_audio.*`); test mod `mods/world-audio-test/` (cars, peds, a ghost skater).
- Per-skater audio state builder (local player output identical); remote players take the NPC skater slot; an
opt-in non-retail "more audible" setting.
- Docs: doc 15 "Hooking up world audio", doc 11 sections, PR plan.
- tools/recomp-trace: bail_impacts.py (per-bail ragdoll impacts and sound posts) and collision_posts.py (collision-sound posts by poster and owner, with the local-rider flag); README entries.
- Doc 11: listening verdicts, the riding body-collision posts are NPC skaters' (local only on big-drop landings and bumps).
- Doc 13: playing the recomp from the couch and closing it (settings overlay, LB+RB+Back hold).
- Doc 14: the added tools.
- Retail steps the board's turn intensity (0.06) and brake (0.05) once per SkateBoard process call, with no dt,
so on the 30 fps console a full turn fades over 333 ms and the brake over 667 ms. Ours stepped per 1/60 s tick,
twice as fast. Both now step once per console frame, frame-rate independent (identical at 30-365 fps); the NPC
bed follows the same cadence. SKATE_AEMS_SLEW_CONSOLE=0 = old (byte-identical renders).
- Concrete ledges grinded as metal: retail stores every audio material as the surface tag - 1 (as our wheels,
deck and feet already did), but the grind material used the raw tag and read the next AudioSurfaceMap entry
(University ledge tag 66 -> metal). Now material_of_tag like the rest; matches the recomp's grind starts
(material 65, concrete GRINDS samples, no Skate_Metal) at the same spots.
- Docs: doc 11 sections for both, gotchas.
- Body region impacts are scaled by retail's centre-of-mass speed curve (x1 at rest up to x5 from ~0.95 m/s,
read from the vault; no clamp after the multiply), so bails reach the tier-1 / tier-2 body and concrete hits.
Setup exports the curve; older installs get the identical default. SKATE_AEMS_BODY_CURVE=0 = old.
- Region impacts take |Δv·n| as retail does (matches 12,054 of 12,054 recomp region lines).
- The body and deck posters' cooldowns count console frames (0.5 s / 0.2 s at any frame rate).
SKATE_AEMS_BODY_CONSOLE=0 / SKATE_AEMS_DECK_CONSOLE=0 = old.
- Second grain bed for the NPC skater (retail's per-instance SkateBoard update; inert until the engine has
NPC skaters). The local player's renders are byte-identical.
- The state log records the COM speed; the e2e harness replays it.
- Docs: doc 11 (bail density: the missing speed curve, not the ragdoll; the hooks' local test), PR plan,
docs index; the published recomp trace reader knows the new line kinds.
- The interim cue tables and their opt-outs are gone; the native runtime is the only audio path. Old
settings/audio.json files still load ("interim" / "native" ignored). Missing native data logs an error
instead of falling back.
- Water splash from retail's mechanism (SFXObj_OffBoard start / update): Skate_Collisions 1187 / 1197 once
per water stay, 1198 when the board is in water.
- The random interim riding bed is removed; riding collision and cloth sounds compared with the recomp
(doc 11). Native output otherwise byte-identical (13 scenarios + 5 sessions).
- Published tools under tools/ (one folder per tool with a README; the recomp ones labelled for use with the
research hooks), doc 14 and PR plan bucket L.
- Docs: doc 11 (retirement, bail density measured against the recomp, collision posts), doc 13 (usage guide).
- Native AEMS runtime is the default (`SKATE_AEMS=0` or `"interim": true` for the interim tables, kept).
- MixMap, eEQChain clear and Jitter on the console's 30 Hz cadence, frame-rate independent.
- Optimisation pass (render p50 210 -> 130 us, 0 allocations per block, byte-identical) and hardware FMA
dispatch in the new skate-audio-fma crate (p50 -> ~55 us, byte-identical; skate-audio stays forbid(unsafe)).
- World emitter banks decoded ahead on a worker thread (no game-thread decode spikes).
- Session-review fixes: push sounds at the foot plant, push-foot plant/lift, bail body impacts, grind on/off.
- Foundations, inert until the engine has the systems: traffic and ped sounds, ped speech rules, NPC skaters'
board sounds.
- Docs: doc 11 sections, doc 13 (the recompilation research hooks, reference only), PR plan.
Native port of Skate 3's retail audio runtime (AEMS) behind SKATE_AEMS=1, plus the world audio work:
- skate-audio crate: formats, patch-program evaluator, voice graph, mixer, MixMap, granular rolling bed and its
bus chain, Splice contacts, the buses (env reverb network, eEQChain, flanger returns, collision and footstep
submixes, reverb zones) and the player components (rolling layers, seams on the console's 30 fps cadence,
grinds, skids, squeaks, tricks, treatment, footsteps, clothing, hands on deck, wind, wheels, collision manager).
- game_audio host: native runtime, player audio state from our physics, world emitters from the maps' .ems files,
random distant one-shots, zone ambience, state log, e2e render harness, timing.
- Setup: audio export of the banks, grains, MixMap, vault tuning and reverb zones.
- skate-core: per-wheel / per-part audio surfaces and seam patterns (additive).
- Docs: docs/hails-additions/11-audio.md and the PR plan.
Replaces our own water placement (water.rs: water collision grouped into
bodies, a random 1-2.5 s piece every 0.7-1.3 s that piled up 2-3 deep, which
the user heard as overlapping water at the University plaza).
Setup (audio manifest v4):
- audio_formats: .ems reader (72-byte records) and name_id, Bob Jenkins'
lookup8 64-bit hash with the seed the sound ids use;
- audio_export: every .ems file with each sound's attributes from the disc's
skatercollections emitter class 0xF0CEF367088EFFF8 (volume, bank, patch
selector, volume type, falloff), resolved through parents; decodes every
bank the sfx_*/skateschool emitters place (100 banks; all 486 sound
emitters resolve to a bank on the disc).
Game (game_audio/emitters.rs, behaviour taken from TU3):
- sphere or ellipsoid (semi-axes = extents along forward/up/side), inner core
at full level, level = volume x falloff ((1-d)^2, 1-d or flat);
- at most 5 emitters, in discovery order, released when the listener leaves;
- each water bank plays its program's measured pattern (two-voice relay of
shuffled one-shots or one loop, with its level/pitch movement) until the
AEMS evaluator is ported.
Verified muted in game: University plaza starts water_fountain and the
channel fountains_waterlaps_left; Aletown canal fountains_waterlaps_left.
PR E merged upstream; PR A trimmed to the long-paths change (upstream fixed
the extract-xiso order itself); new row K (solver iterations); upstream sync
notes and known-failure list updated.
- Pop, flip-catch and landing cues built from the layers retail plays per
moment (bank, record, level, delay, envelope measured from TU3 traces),
with hollow/normal landing tables and per-kind chances.
- One-shots start at their final level (no fade-in eating the transient);
loops still fade in.
- Riding bed from the measured statistics.
- Master volume defaults to 75 % (100 % is unity gain; 25 % was too quiet
to judge retail parity).
The stock RWMaxIterations (25) is only the Simulation's setup value; TU3
solves with 50 on every gameplay frame. With 25 the constraints converged
less: wheels sank ~10x faster at rest, the ollie deck sat ~1 cm low and
the jump velocity the pop sound reads was too small (rolling ollie
0.21-0.29 vs retail 0.50), so pops picked the quiet row.
Maps validate identically; water traces stay within expectations. See
docs/hails-additions/12-solver-iterations.md.
- docs/hails-additions/11-audio.md (current design, formats incl. the
SPLC patch tree, cue table, how it was tuned, prior work upstream
#1/#4 and why their runtime is not adopted), README row 11,
PULL-REQUESTS row J.
- tools/audio_audition.py: dev page to audition decoded samples (local
.local/ output only).
- trace-downtown-lag.json: dev-only perf-trace output from TRACE_PLAY.bat
(header only, no capture); committed because everything pending is
committed, not needed by the build.
- 09-water.md: physics, camera, splash and look after in-play testing
- 10-sky-shader-test.md, README index
- PULL-REQUESTS.md: PR H prepared against upstream 60efdef
- water-test-teleport: F7 Aletown canal, F5 position readout, the
"Teleported" text clears after 5 s
retail_sky.wgsl reads forward_io's Vertex.position, which Bevy only
declares under VERTEX_POSITIONS; the mesh pipeline sets it for every mesh
with positions, but the test's validate() helper did not, so
sky_shader_validates failed (also on upstream 60efdef). All six shader
validation tests pass.
- PULL-REQUESTS.md: pending edits to the upstream PR plan
- mods/water-test-teleport: dev-only mod for testing water (F2/F3/F4
drop into University fountain, University reservoir, DownTown fountain;
Shift+F2/F3/F4 stand beside them). Not intended for upstream.
- water_surfaces example: view spots beside water, render materials at a
point, model materials, texture and material inspection; SKATE material
ids are 1-based (an off-by-one misreported water materials)
- 09-water.md: frame-state fix, water time, why University water looks
darker than DownTown's (authored data), splashes still open
- README: fork renamed to rusty-trucks
Water and ocean shaders need assets/private/ocean-pca.json, which setup
never created, so 52 water/ocean materials rendered static (or black).
- skate_data::xex: XEX2 unpacker (retail AES file key + CBC payload, LZX
or basic decompression), no new dependencies
- skate_data::ocean_pca: locates cPCAWaterAnimationData's table in any
build by its code addressing (disc build: 0x82FC3978; TU3: 0x830118D8)
- skate3rust --extract-ocean-pca <default.xex> <out.json>
- Setup environment group runs it (optional content); refreshes that group
In-play check of the rendered water still pending.
See docs/hails-additions/09-water.md.
One skate3rust process validates TEST_WORLD and every converted map over
stdin/stdout (SKATE_MAP_CHECK lines). Existing checks still fail a map;
new spawn-support / neutral-startup / invisible-collision checks report
warnings. Setup falls back to per-map --check-assets with an older exe.
See docs/hails-additions/06-map-validator.md (PR F).
Note: install.py also carries PR G's customiser-overlap hunks
(available_memory, overlap_customiser, Background); they are interleaved
with the validator wiring in _install and are documented in doc 08.
Collections are looked up through a lazily built OnceLock<HashMap> index
with first-match semantics, identical to the linear scan on 145,807 real
lookups. See docs/hails-additions/07-collections-index.md (PR E).
Retail collision includes clustered meshes in which no unit carries a
surface ID (unit flag 0x80 unset). Across every shipped district these
are only 12-triangle boxes: zone, trigger and level-bounds volumes. The
loader made them solid, creating invisible walls (SkateSchool spawn
area, DownTown's Kube Tower plaza, MegaPark's stadium entrance, Maloof).
Authored player positions lie inside some of them, so they are not solid
in the original game.
visit_clusters now decodes and validates each mesh fully and skips it
when no unit has a surface ID. Meshes mixing surfaced and surfaceless
units are ordinary geometry and stay whole. RetailTriangle records
has_surface.
Collision triangles loaded: SkateSchool -84, DownTown -24, Maloof -24,
MegaPark -12, all other maps unchanged. Confirmed in play.
See docs/hails-additions/05-collision-volumes.md.
Every map's startup position and default landing zone is its baked
spawn. The old rule (upward triangle nearest the origin) put the skater
over the void on MegaPark and on the roof of Industrial Skate Park and
StartPark, because park origins sit at a corner and retail floors are
sometimes wound downward.
New map_starts.py resolves each district's start from retail data: the
world row's start locator (parks), a chosen fe_locations row for
districts with several (University career start, Industrial Haystings
Park, DownTown Rosalita Skate Park, Maloof street), or the district's
only fe_locations row. Position and heading come from the district's
global_locators records; the heading is now written into the .skate
header. Districts without an authored start fall back to geometry:
compact parks use the new ground_spawn search, city-sized districts keep
the original rule (ground_spawn put Industrial under the streets).
See docs/hails-additions/04-map-spawns.md.
Implements issue #14. A dedicated thread owns SDL3 (3.4.16, statically
linked via sdl3 =0.20.0) and publishes per-slot snapshots re-expressed as
the XInput-shaped XboxState, so the TU3 conversion in skate-core is
unchanged. Conversions are exact for XInput-backed pads (stick Y via
bitwise NOT, trigger round-to-nearest). SKATE3_INPUT=xinput selects the
original transport, which is also the fallback if SDL fails to start.
Optional settings/controller.json maps paddles to XInput buttons.
Build: scripts/Ensure-CMake.ps1 locates CMake (PATH or Visual Studio's
copy) for the SDL source build. sdl3-sys is vendored with a build.rs
patch setting CMAKE_MSVC_RUNTIME_LIBRARY from crt-static, because SDL's
CMake (policy CMP0091) ignores cmake-rs' -MT and the static-CRT release
build failed with unresolved __imp_* symbols.
See docs/hails-additions/03-sdl3-gamepad-input.md and vendor/README.md.
Roster work paths under deep checkouts exceed MAX_PATH (260), leaving
pro skaters unavailable. If LongPathsEnabled is off, setup asks once for
elevation to enable it and relaunches itself, since Windows reads the
setting at process start. Declining keeps the previous behaviour.
See docs/hails-additions/02-long-paths.md.
extract-xiso stops parsing options at the first image path, so
`-x <iso> -d <dir>` extracted into the working directory and then failed
on "-d" as a second image (exit 1 -> "Conversion failed"). Use
`-x -d <dir> <iso>` in both call sites.
See docs/hails-additions/01-iso-extraction.md.
Add generalized engine access and challenge mods, repair SKATE turn handling, and improve collision prediction and physics transitions. Reduce mod snapshot and material update overhead, preserve authored GLB materials with lighting and shadows, and improve CPU profiling.
Sanitize non-finite SIMD lanes in offboard air selection, add solid dynamics and mod replication APIs, improve setup refresh without re-extracting assets, and harden Build-Release.ps1 for older PowerShell hosts.
Co-authored-by: Cursor <[email protected]>
Hails