ARE ambient-animation detection (memory-primary, disk fallback) - #21
Draft
TheForgotten69 wants to merge 20 commits into
Draft
TheForgotten69 wants to merge 20 commits into
TheForgotten69 wants to merge 20 commits into
Conversation
TheForgotten69
force-pushed
the
feature/are-animation-detection
branch
8 times, most recently
from
July 17, 2026 21:14
dc44bcf to
5606827
Compare
…cale heuristic At the LoadArea boundary, walk CGameObjectArray for the active area's live CGameStatic objects (the raw authored ARE record is retained at +0x60) and classify authored point sources into an AppContext snapshot (graphics roadmap §10.3 classification channel). Resolution rides EEex's version-independent GetShare pattern with RIP-operand decoding of the m_maxArrayIndex/entry-table globals; any ambiguity fails closed and disables the scan. Offline-verified on both supported builds (2.6.6.0: RVA 0x276490, globals 0x68D434/0x68D450; 2.7.3.0: RVA 0x276700, globals 0x68F8F4/0x68F910) and validated in-game on 2.7.3. Unclassified resrefs are logged per area to grow the classification table. Water bodies stay on the WED overlay path. Classification table calibrated against the full BG1EE (96%+) and BG2EE ARE exports plus visual frame inspection of the ambiguous BAMs: FLM*/FLSM*/FIM*/ YSFL* flame families, candle-named flames as light, steam as smoke, bubbles as water, new water/lava/wildlife kinds, and a built-in exact-resref table for the identified per-area AM* overlays. AR<area>W[DN]* overlays are day/night shadow scenery, not lit windows. First consumer: the fpSEAM override renders point effects from the snapshot — warm flickering fire glow with heat shimmer above flames, drifting chimney- smoke haze columns, and steady candle/glow halos. Shown fire/light/smoke points are packed into a 32-slot uIeePoints[] uniform array (fire prioritized under capacity pressure) and fed through the uniform bridge under its own revision. Gated by [Shaders] EnablePointEffects and the F10 master toggle; background layer only, with vanilla output untouched when off. Also removes the tile-scale UV/texture-id heuristic tier: raw UV magnitudes are atlas-origin noise and GL texture ids grow with session allocations, so the tier produced false 4x detections on vanilla areas (observed in-game on 2.7.3). Detection is now TIS header -> PVR-table grid GCD -> fail closed to standard 1x delegation via the existing sampling path.
TheForgotten69
force-pushed
the
feature/are-animation-detection
branch
from
July 17, 2026 21:34
5606827 to
b42c659
Compare
…kers The struct-returning helper was the remaining suspect for the effects not rendering despite a verified uniform feed (locations resolved, count fed) — old GLSL frontends miscompile struct returns. The loop now uses flat locals inside main. ALIGN mode (F10 x2) draws a dot and 40px ring at every fed point (orange fire, grey smoke, yellow light), so point placement is verifiable in-game independently of the effect math.
… metadata-free tilesets In-game ALIGN markers confirmed exact point placement, but the first-pass constants were too timid, especially at night where multiplicative glow cannot lift dark pixels. Fire glow gains a night-adaptive additive cast-light term, a wider radius (140px) with a hot core, and an up-biased center onto the visible flame; smoke columns are taller, denser, and read as moonlit haze over dark roofs; candle glow strengthened. Also skips the 10-draw sampling window when a tileset is resident and readable but simply has no deterministic scale metadata (classic paletted tilesets): more draws cannot change that verdict, so it latches standard 1x immediately. Sampling remains for resources still streaming in.
…nder suppression Radial-gradient halos read as artificial airbrush blobs (and painted the unexplored void). This pass makes fire and smoke real replacements, the same approach as the water pillar: - New tileable FBM asset iee_effects_noise.rgba (generated in-repo, IRGB blob, unit 6/uIeeNoiseMap) riding the existing texture registry with the same DDS-override path. - fpSEAM draws shaped flame bodies eroded by rising noise with a blackbody ramp (per-resref scale: sconces small, fire pits full), night-adaptive cast light that never paints the void, and noise-billowed smoke plumes that wander and dissipate with altitude. - CGameStatic::Render hook (prologue pattern offline-verified unique on both builds: 0x1F2570 / 0x1F27D0) skips the engine's authored fire/smoke BAM draws while the effect is active; WBM/PVRZ setpieces and all other kinds stay vanilla, and failure to resolve keeps engine draws.
…e; fix haze mottling Calibrated against the exported BAM frame dimensions: FIRE* torch flames are 12x45 px, FLAM* sconces ~10x30, YSFL* small flames 7x14 — the v2 flame body was 42x76 for everything, hence the traffic-cone look. Flame strength now derives from authored height per family with the authored ~1:3.5 aspect, noise sampling scales with flame size so small flames keep structure, and the hot-white core is confined to flicker peaks. FLAMBLU*-family flames keep their authored blue (palette id in the kind fraction drives both ramp and cast-light color). Smoke plumes are narrower, wispier, and capped at 62% opacity so the roof stays readable, and the per-texel luma gate is removed from the haze mix — it mottled dark stonework; the gate remains only on the fire cast-light additive term, which is what protects the unexplored void. Also archives the owner-observed water follow-up (some cells render more transparent than the opaque water body) in the phase-2 water plan backlog.
…engine draw Two field-observed defects, both from ignoring the BAM frame tables: - Suppression scope: AM*/AR* per-area overlay images (AM5204C lit-hearth art in AR2616) were suppressed like standalone flames, blacking out the hearth while a synthetic flame burned on the floor anchor. Replacement is now limited to standalone flame/smoke BAM families (should_replace_animation_draw); overlay fires keep the engine draw and contribute cast light only (palette id .2), overlay smoke packs no point. - Placement/size: points now carry the exact authored draw box, extracted from the game BAM frame tables (frame w/h and center offsets; e.g. FLAMBLU2 draws down-right of its anchor at 8x15). Each point uses two vec4 slots (base-center position + height, half-width); flame bodies, plumes, and glows derive from authored pixels instead of guessed scales.
…appled cast light The blue-flame palette rode the kind's fractional digit through the uniform feed; field evidence says it never survived to the shader on the target driver. The palette is now a plain float in the point's second slot (0 warm, 1 blue, 2 glow-only) — no fract decoding anywhere — and every packed point is logged at DEBUG (kind/palette/pos/geometry) so the feed is verifiable from the game log alone. Cast light no longer reads as an airbrushed disc: the footprint is a 1.9:1 ellipse (isometric floor projection), modulated by drifting noise dapple, with reduced multiplicative and additive weights. Also archives the liquid-type sweep findings in the water backlog: lava's opaque art never engages the alpha-hole mask (root cause for "lava not working"), washed-out sewage channels, and the saturated harbor sea — all scoped to the dedicated water pass.
…centered on the art Field log confirmed the blue-flame bug: live resrefs keep their authored casing and this install stores "flamblu2" lowercase — the exact-match geometry table and the "BLU" palette check both missed, falling back to default geometry with a warm palette (classification survived because it uppercases internally). Both lookups now run on the uppercased resref, with a lowercase regression test. Glow-only overlay points (hearths) anchor at the art's floor edge; their cast light is now centered well up onto the coals instead of pooling on the floor in front.
…rtion
Placement evidence across rounds (ALIGN markers land on the sconce bowls;
frame-box-derived offsets rendered flames below them) says the ARE anchor is
already the flame's base point — the per-family dx/dy shifts are gone and
only the authored footprint varies.
The heat shimmer offset the ATLAS texture coordinate; crossing a tile
boundary sampled unrelated atlas content, which showed as a vertical
smear-seam ("compressed map"). Shimmer is removed from the tile pass —
world-space distortion belongs to a future post pass.
…-box offsets CGameStatic::RenderBam ground truth (decompilation): the engine draws the current frame's top-left at (CGameObject::m_pos - frameCenter) — the BAM's embedded per-frame centers ARE applied, and the render base is the live object position, not the ARE header coordinates (which can diverge; a chimney showed header y 163px off its authored value). The walk now reads m_pos, packing places the flame at the frame box's bottom-center relative to it, and the per-animation debug line logs both positions so any divergence stays visible in the field log.
The per-resref geometry table is now a fallback only. The walk mirrors the engine's own cached current-frame entry (CVidCell::m_pFrame at CGameStatic +0xB0+0x128, the exact frameTableEntry RenderBam draws with), so any flame BAM — including mods — places and sizes itself from the same numbers the engine uses: frame top-left at (m_pos - center), flame at the box's bottom-center. The fallback covers only the first refresh before an object has rendered once and the disk-parse path. The per-animation debug line now prints the live frame geometry, and CVidCell/CGameStatic layouts carry static asserts for the new fields.
The ARE object position is already the flame's on-screen base — ALIGN markers at raw objPos land on the sconce bowls, so every per-family dx/dy offset was the source of the persistent shift. Placement now uses objPos directly; only the authored size varies per family. Also records the finding that CVidCell::m_pFrame stays null for replaced flames: the Render-suppression hook returns before RenderBam populates the frame cache (chicken-and-egg), so the size table is the normal path and the live-frame read is an opportunistic upgrade only.
CGameStatic::RenderBam draws at y = m_pos.y - m_posZ; mounted flames (wall sconces) carry authored Z, so anchoring at raw m_pos.y left them hanging below the art by exactly z pixels. The walk now reads the live m_posZ (+0x14, already modeled), the disk mirror loads the ARE height field, and point packing applies the same subtraction. objZ joins the per-animation debug log so field logs show the elevation directly.
Enables building InfinityEngine-Enhancer.dll from Linux via a mingw-w64 toolchain, for local Proton-based iteration without a Windows host. Fixes three MSVC-only assumptions the cross build exposed: - release_bundle unconditionally copied $<TARGET_PDB_FILE:...>, which hard-errors at CMake generate time for non-MSVC linkers (mingw has no PDB). Gated behind if(MSVC); the MSVC/CI path is unaffected. - opengl_types.cpp implicitly converted FARPROC to void*, an MSVC extension GCC rejects under -fpermissive-off. Made explicit. - shader_probe.cpp called _ReturnAddress(), which mingw-w64 declares for source compatibility but never defines, causing an unresolved symbol at link time. Added a caller_return_address() wrapper that uses __builtin_return_address(0) on GCC/Clang. Adds cmake/toolchains/mingw-w64.cmake for the cross build itself.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Detection data path for the graphics roadmap §10.3 classification channel.
CGameObjectArrayfor the area's liveCGameStaticobjects (raw authored ARE record retained at +0x60). Resolution via EEex's version-independentGetSharepattern + RIP-operand decode; any ambiguity fails closed.override/thenchitin.key/BIFF V1 reader, cached by resref.0x276490on 2.6.6.0 with operands decoding to0x68D434/0x68D450; parser validated against real BG2EE data (AR0406/AR0700). Full trail indocs/are-animation-detection.md.🤖 Generated with Claude Code