mirror of
https://github.com/bryanthaboi/gen1recomp
synced 2026-10-27 23:35:16 -04:00
Shane McGovern
7e2307f58c
* Use spare CPU cores to remove in-game hitches on low-power devices Per-frame game logic and drawing are cheap; the remaining stalls on low-power handhelds are one-off main-thread hitches from chip SFX/cry synthesis, RED++ atlas recolouring, image decodes and the require shim. Move those onto background threads, each with a synchronous fallback so behaviour is identical when threads are unavailable (POKEPORT_NO_THREAD=1, no love.thread, worker failure). Also adds an F3 frame-time overlay and POKEPORT_GAME_PROF headless profiling to measure these costs. * Address code review: shared WorkerLane, F3 gating, decode/epoch fixes - Extract src/core/WorkerLane.lua (worker lifecycle, prioritised lanes, per-lane epochs, bounded LRU) and rebuild the SFX and atlas/decode prefetchers on it - Keep decoded sprite sheets in every colour mode (the OBJ palette bake re-reads them), skip oversized decodes, and let palette changes drop only atlas bakes - Only bump the SFX render epoch when the channel mix actually changes - F3 overlay toggle no longer steals the key from rebinding screens or player bindings; profilers stop only when a quit really proceeds - Require fast path: test hook instead of a hot-path counter - Shared tests/drivers/perf_util.lua; drivers restore their patches * WorkerLane: no budget leak on duplicate sends, keep priority on requeue - Lane:pump only counts a job against the in-flight budget once per key - requeueInflight puts entries back at the head of their own queue, so background hints no longer overtake front requests after a bump - shutdown() keeps a fatally failed worker off - Hiding the F3 overlay logs one summary line instead of a full block * F3 overlay: last-120-frame stats, font-sized box, scale on large screens * Prefetch: skip useless hints, no main-thread fallback, mix check vs worker config - Don't queue decodes for sprite sheets whose Image and OBJ bake are cached - Atlas hint is dropped (sync bake at build time) when the worker can't decode the tileset itself, instead of decoding on the main thread - SFX epoch bump compares the live mix with what the worker was last configured with, replacing the load-time snapshot * Sprite decode skip: raw-path cache check, never under RED++ * Gen 2 battles prefetch both parties' cries; sprite skip keyed by raw path * Prefetch: skip cached tilesets outside RED++, log hint failures once * Profiler: bench survives F3 hide, frame total excludes vsync; safer SFX pump, ordered requeue, tileset skip respects pending variants * Profiler counts GPU stalls in present; water void-fill keeps tileset hint; README + test coverage * Profiler vsync cap: measured cadence via RefreshRate, no per-frame allocation * Use upstream's chip-worker prewarm for SFX/cries; drop our separate SFX worker Our extra coverage (Gen 1 common SFX + party cries, both parties' battle cries, lead move sounds, hit and battle SFX, Gen 2 bench cries) now goes through ChipAudio.prewarmSfx/prewarmCry and a new Sound.prewarmMove. * Skip decode hints for composed assets; share move cache key; drop dead demoteAll * Drop dev-only profiling extras: POKEPORT_JITP hook and perf drivers * README: name the worker threads POKEPORT_JIT leaves alone * Profiler: count the pre-present vblank wait as present, log on NX quit - time PresentSync.waitBeforePresent as "present" (the X11 GLX vblank wait lands there), so it is excluded like the wait inside present - NX quit now emits the PROF summary / overlay line too - love.quit's scripted check reads FrameProfiler.benchRequested, so POKEPORT_GAME_PROF=0 no longer counts as a bench run - prefetchAtlas skips composed tileset sheets (Assets.isComposed) - hit sounds defined once in EffectRegistry.HIT_SOUNDS, shared with the battle prewarm (each hit still gets its own copy) - drop redundant obpPaths, dead FrameProfiler.visible and the unused Lane:bump requeue mode; restore applyOptions' doc comment