TeslaRel410/emulator/firmware-decomp/IG-SHADING-MODEL.md

# The IG-board shading model (ground truth from libDPL)

Authoritative reference for the software rasterizer, distilled from the **shipped
Division library headers** in the reconstructed game source:

- `C:\VWE\BT411\engine\MUNGA_L4\libDPL\dpl\dpltypes.h` — the public `dpl_*` enums
- `C:\VWE\BT411\engine\MUNGA_L4\libDPL\dpl\dpl.h` — the `dpl_Set*` API
- `C:\VWE\RP411\DivLoader\{VGCDivLoader.h,tagIdents.h}` — the `.biz`/BGF asset
  format the same model is authored in (DivVertex, BIZ_IDENTITIES)
- `C:\VWE\BT411\context\rendering.md` / `docs\BGF_FORMAT.md` — how the modern
  D3D9 port reproduces this same model (the ramp/BSL/effect-card fidelity work)

These are the host-side commands the game issued to the VelociRender board; our
firmware-decomp emulates the board that consumes them. Where our renderer differs
from this model, this file is the reference.

---

## 1. Vertex type mask — the per-geometry shading selector

`dpl_VERTEX_TYPE` (dpltypes.h:189) is a bit mask; it is exactly the `DivVertex.mask`
field in the asset format (VGCDivLoader.h:65):

| bit  | token                   | meaning                                    |
|------|-------------------------|--------------------------------------------|
| 0x01 | `dpl_vertex_coord`      | has x,y,z (always)                          |
| 0x02 | `dpl_vertex_normal`     | has Nx,Ny,Nz → **LIT** by the map light    |
| 0x04 | `dpl_vertex_rgba`       | per-vertex RGBA (effect cards)             |
| 0x08 | `dpl_vertex_luminance`  | per-vertex luminance `l`                   |
| 0x10 | `dpl_vertex_texture2D`  | has u,v                                     |
| 0x20 | `dpl_vertex_texture3D`  | has u,v,w (perspective texture)            |
| 0x40 | `dpl_vertex_radius`     | sphere/point radius                        |

**The key rule (rendering.md §"IG-board shading model"):** shading is chosen PER
GEOMETRY by the vertex type:

- **No-normal geometry** (terrain / mesas / sky / buildings / mech — the WHITE
  baked verts) is **UNLIT**, and colorized by the material's 2-endpoint RAMP.
- **Normal-bearing geometry** (~150 vehicle/missile files) is **LIT** by the map
  light; the ramp is NOT applied (applying it = the "dusty white blobs" bug —
  gate ramp on `!hasNormals`).
- Exception: cockpit-frame (`*_cop`) batches are pure-emissive constant colour.

---

## 2. Material ramp — how the board colours grayscale texels

`dpl_SetMaterialRamp` + `dpl_SetRampComponents(r0,g0,b0, r1,g1,b1)` (dpl.h:326,578):
a material carries a **2-endpoint colour ramp**. The (grayscale / bit-sliced)
texture's **luminance L∈[0,1] indexes the low→high gradient**:

```
out_rgb = rampLo + L * (rampHi - rampLo)        # per channel
```

This is how the IG board coloured terrain from gray bit-slice art (rendering.md /
bgfload.cpp:341): rock `0.25,0.21,0.16 → 0.8,0.5,0.4` (warm tan); grass
`0.13,0.13,0.07 → 0.38,0.33,0.23`. Materials with no diffuse relied ENTIRELY on
the ramp.

> **Applicability to the wire-capture renderer:** the `0x1a` texmaps in the
> `netdeath` capture arrive as *already-colourised RGB* (HUD labels, emblems,
> panel/terrain art — 0–12% grayscale), i.e. the ramp is pre-baked on this path.
> So `render_final` must **not** re-apply a ramp (it would double-colour). The
> ramp model matters for a raw grayscale/BSL texmap path, not for these uploads.

---

## 3. Texture value model — wrap and alpha (implemented: `dpl_sampler.py`)

`dpl_SetTextureProperty(tex, prop, value)` (dpl.h:344) sets `dpl_TEX_PROP` fields
to `dpl_TEX_VALUE` tokens (dpltypes.h:103,121). The two that change rasterizer
*output*:

**WRAP** (`dpl_tex_prop_wrap`/`wrapu`/`wrapv`):
- `dpl_tex_value_repeat` — texel index wraps mod size (tiling terrain) — the board default
- `dpl_tex_value_clamp`  — texel index clamps to the edge (a one-shot decal / HUD
  label / emblem — tiling a "PLAYER 1" label is the wrong-wrap tell)

**ALPHA / cutout** (`dpl_tex_prop_alpha`):
There is **no stored alpha channel** — the texel word is `[pad, B, G, R]` (see §3a).
The board realises `dpl_tex_value_cut` (**PUNCH**: the cockpit/effect-card cutout
that retires the "opaque rectangle" reading — the wreck flame that read as a
"twisted drill bit", rendering.md §effect-cards) by **near-black keying** at draw
time: a texel whose `R+G+B <= ~24` is a HOLE (poly not drawn there, no z claim).
Ground truth: `dpl3-revive/patha/vrview_gl.py` ("alpha cutout (texel sum <= 24/255)").

Filtering (`point`/`bilinear`/`trilinear`/`mipmap_*`) is a quality axis; pvision
runs the board in texmode 0 = **point (nearest)**, which is what we model.

`dpl_sampler.py` implements wrap + near-black cutout (numpy-vectorized, with a
synthetic conformance self-test). Cutout is a per-MATERIAL property the wire
capture does not yet expose per drawn surface, so `render_final` defaults it OFF
(opaque) and leaves the hook for when a `SetTextureProperty` stream is decoded.

## 3a. Texel format — SVT `[pad, B, G, R]` (the ex-CYAN bug)

The `0x1a` upload's texel word is **`[pad, B, G, R]`** ("SVT xbgr") — ground truth
from the real renderer's own decode (`dpl3-revive/patha/vrboard.py:448` `do_texels`,
`stage_assets.py:43`). **Byte 0 is a PAD (0 in every texel across the capture);
the colour is bytes 1,2,3 = B,G,R, so RGB = bytes (3,2,1).** `mode` 1 is the
board's internal 8-bit storage path — the WIRE upload is always 4 bytes/texel.

`texstore.py` had been reading RGB as bytes (0,1,2) = `(pad, B, G)` — zero red —
which rendered the whole arena **CYAN**. Fixed to (3,2,1): the arena reads as its
true grayscale metal panels + coloured decals. This, not the texid binding (§5),
was the cyan.

---

## 4. Fog — the arcade model (not yet in the software renderer)

`dpl_FOG_TYPE` (dpltypes.h:238): the arcade uses **`dpl_fog_type_pixel_lin`** —
per-PIXEL linear fog on perspective W (rendering.md §fog: table fog on the
z-buffer collapses because the z-buffer is perspective-nonlinear; the board fogs
on W). Authored per map/time/weather as `fog = <near> <far> <r> <g> <b>` in the
DPL env INI (e.g. cavern/night/clear: near 90, far 1100, dark blue 0.1,0.1,0.12).
Night fog is dark BY DESIGN. Our renderer has the per-pixel Z/W planes to do this
faithfully; the near/far/colour are map-specific (not applied blind).

---

## 5. The remaining blocker — texid → texmap-handle binding

`render_final` still binds each drawn quad's 6-bit texid to an uploaded handle by
`texlist[tid % len]` — a placeholder. Investigated 2026-07-19/20: the draw's
texture-select word (op `0xf7`, ad 141/142) is **not** a handle — it is the
board's *resolved texture-descriptor reference* (per-texmap: high byte differs,
`0xf26a`/`0xf081`/`0xf1ea`/`0xcafa`…), computed at geometry-CREATE time from the
texmap (traced: the select-word value is written to scratch ~`0xbfc98` during the
`create` cmds, not the opaque texel store), and `battle_prog.pkl` is the
firmware's already-compiled program, so the game's material/texture *names* are
gone by then. Cracking it needs the deferred **board texture-RAM model**: RE the
firmware texmap→descriptor allocation and map the select word to an upload handle.

**Note (2026-07-20):** the monochromatic frame was NOT this binding — it was the
texel-format misread (§3a). With SVT decode fixed, most arena surfaces are
grayscale metal, so the frame reads correctly *despite* the placeholder binding;
the binding now only mis-tints the handful of coloured decals. Much lower priority.

**Escalation (2026-07-20, live-pod session):** with the reconstructed renderer
running live behind the actual game (not just replaying a static capture), this
binding gap became visually severe on **vehicles** ("trucks look like they were
painted with skittles" — user report). Vehicles are exactly the ~150
normal-bearing (LIT) BGF files that carry authored PAINT/decal materials
(rendering.md §"Per-pilot mech PAINT" — warning-stripe/badge/patch layers), so a
wrong-texture assignment here is far more visible than on a flat metal wall.
**No longer low priority — this is the top open item for texture correctness.**
Path forward (not yet done): live-instrument the firmware (we run the actual
i860 code via emu860c, so we can WATCH memory writes during `set_texmap_texels`
processing rather than guess from wire order) to find the real texid→handle
table the firmware builds, instead of the `tid % len(texlist)` placeholder.

---

## 6. Purpose of this pipeline — hardware ground truth for BT411 (2026-07-20)

This renderer's job is **not** to look good — it is to give the BT411
modernization-port developers a hardware-grounded, independently-verified
answer wherever the software renderer (`dpl3-revive`, the D3D9 port) has open
or hard-to-fix questions. Where this doc and BT411's `context/rendering.md`
overlap, treat agreement as cross-validation (two independent
reconstructions — firmware-level here, decomp-level there — agreeing) and
disagreement as a lead worth chasing on BOTH sides.

**Cross-reference index (BT411 `context/`):**
- `rendering.md` — the authoritative D3D9-port model: vertex-type shading
  select (§1 above), material ramp (§2), fog (per-map/time/weather via
  `BTDPL.INI`, per-pixel linear on W — NOT yet implemented in this renderer),
  weapon beams (`ermlaser.bgf` TUBE geometry, muzzle-anchored, NOT a special
  ray-trace primitive), mech skins (BSL bit-slice), per-pilot PAINT.
- `open-questions.md` — living list of everything BT411 has NOT yet resolved;
  check before re-deriving something they've already nailed down (or vice
  versa: this firmware pipeline may resolve one of THEIR opens).
- `combat-damage.md` — target acquisition (torso-boresight pick ray, auto-lock)
  is RECONSTRUCTED and working in BT411's port; if our live session shows "no
  weapon fire," suspect input/session state (fire button not reaching the
  emulated RIO, or simply not fired yet this session) before suspecting a
  missing render mechanism — the port's evidence is that firing itself works.

### 6a. The lit-color plane (r24/g24/b24) — REAL, mechanism confirmed, exact use UNRESOLVED
Verified live from the actual wire + firmware source (`sda4/DPL3/VRENDER/PXPL5SUP/`):
- `DIVPXMAP.H` defines `dvpx_r24/g24/b24` (addr 117/125/133, 8 bits each) as
  scan-conversion-time fields; `EOF.C` (lines ~1456, 1742) copies them
  straight into `dvpx_eofr/g/b` (184/192/200 — the final screen color) via a
  24-bit `IGC_CPY` in the simple (unfogged) case.
- Confirmed present in the live per-draw coefficient program: `TREEclmpintoMEM`
  (op 0x5a) at addr 118/126/134 (one bit off from the header's 117/125/133 —
  this bytecode's own addressing, taken empirically), a 3-word instruction
  whose 3rd word is a clamped float — REAL per-polygon values observed
  (~49..395, not the same everywhere, roughly B > R > G — a genuine signal,
  not noise).
- **What's still open:** applying this as a flat-poly color (§ current
  gpu_raster.py) is a reasonable, low-risk use since flat polys had no real
  data before (a placeholder gray). Applying it as a TEXTURE TINT is WRONG
  (§5 caveat: textures arrive pre-colourised) — tried live 2026-07-20,
  reverted same session (commit 0527022e) after the user reported "more
  colors but neither correct." The exact relationship between this
  BOARD-INTERNAL scan-conversion value and BT411's GAME-LEVEL ramp/lighting
  model (§1-2) is NOT yet established — they may be the SAME quantity
  (the board-side result of the ramp/lighting the game already computed and
  is now shipping per-polygon) or a DIFFERENT one (e.g. a fog term). Next
  step: cross-check a KNOWN lit vs unlit polygon (from BT411's vertex-type
  rule) against its r24/g24/b24 value to see if unlit polys show the
  material ramp's actual endpoint colors.

### 6b. Fog — IMPLEMENTED (2026-07-20), depth formula source-derived
BT411 (`rendering.md` §Fog) has the complete, verified model: per-map/time/
weather leaf in `content/BTDPL.INI` (`fog=<near> <far> <r> <g> <b>`), per-pixel
LINEAR fog on distance. For THIS mission (testarn.egg = arena1/day/clear) the
leaf is `[ardayclear]`: `clip=0.25 1300.0`, `fog=200.0 1250.0 0.32 0.3 0.5` —
verified against the live game tree, exact match.

**The depth term was initially attempted with `tz` (texz) directly, which is
WRONG** — measured live (a vertical scan up a receding ceiling, row 0 to row
160): `tz` DECREASES with distance (324 -> 121), i.e. it behaves as an
inverse-depth quantity, not the "linear on W" BT411 describes. Re-derived from
actual firmware SOURCE instead of guessing again: `sda4/DPL3/VRENDER/AS860/
XFPROJ.SS`'s own comment on the projection code says `invz=eye_d/rz` and
`pz=wz*invz (where pz=1 at near clip, 0 at infinity)` — i.e. the z-buffer value
(`dvpx_zbuf`, 20-bit) is a NORMALIZED `hither/distance` quantity. So:

```
distance = hither * 2^20 / raw_zbuf_value
```

Implemented in `gpu_raster.py` (`FOG_HITHER/FOG_ZSCALE/FOG_NEAR/FOG_FAR/FOG_RGB`
+ the shader's new 10th output channel carrying the raw z-buffer float). Sanity
check (frame 11, netdeath capture, a receding ceiling): distance climbs
394 -> 1051 smoothly from the near edge to the horizon vanishing point,
landing squarely inside the authored 200-1250 fog range with no re-guessing
needed after the tz correction. Visually: the horizon now shows the authored
blue-purple haze (0.32,0.3,0.5), fading correctly toward the near floor.

**Not yet independently cross-validated:** the exact 2^20 scale (vs. 2^20-1 or
a slightly different fixed-point convention) and treating `hither` as a fixed
per-mission constant (0.25) rather than reading it per-view from the wire's
view-flush payload (the OTHER renderer already parses `hither +44, yon +48` —
see [[dpl3-revive-integration]] memory) are reasonable, source-grounded
choices, not exact proof. Good enough to trust qualitatively; treat exact
near/far crossover distances as approximate until cross-checked against a
reference screenshot or the view-flush hither/yon.

### 6c. HUD — not rendering at all, cause NOT YET INVESTIGATED
No wire-level investigation done yet. BT411's `gauges-hud.md` /
`cockpit-view.md` should be checked first for the mechanism (a separate
draw pass? a different coordinate space/orthographic projection? a distinct
wire action type?) before searching the firmware blind.

### 6d. Weapon-fire rendering / board-side pick-range mechanism — PARTIALLY INVESTIGATED, PAUSED
Per BT411 `rendering.md` §"Weapon beams": beams are ordinary TUBE geometry
(`ermlaser.bgf`), muzzle-anchored every frame — i.e. on the wire this should
look like ordinary create/list_add/draw_scene traffic for a thin, elongated
object, NOT a special ray-trace primitive. Per `combat-damage.md`, the
original game's fire/target-lock mechanism is fully reconstructed and working
in BT411's port, so a live session showing no fire is more likely an input/
session-state gap (RIO fire button not reaching this DOSBox session) than a
missing render path.

**User's counter-point (2026-07-21), correcting an oversimplification above:**
the renderer is NOT a purely passive tap in the strict protocol sense — there
IS a real query/response shape in the wire protocol (VELOCIRENDER_PROTOCOL.md
`vr_sect_pixel`(4)/`vr_sect_vector`(5): synchronous, blocking, the board runs a
REAL Möller–Trumbore ray-triangle test over its own live scene and replies
with an instance handle — "the exact range to the pixel in the center of the
target and what that pixel is a part of"). Whether the SHIPPED BT4.10/MUNGA
build actually uses this path (vs. computing it host-side, as BT411's own
`HudSimulation`/pick-ray reconstruction does) was the open question.

**Investigated, using this pipeline's unique capability (running the REAL
firmware, not a reimplementation):**
- `sect_pixel`(4)/`sect_vector`(5) never appear on this capture's wire at all
  (0 occurrences) — this build doesn't use the 1994-enum pick-query actions.
- Action `0x23` (the notes' "fire/pick" candidate) DOES appear (23 times), and
  checked directly against the real firmware's OWN reply mechanism (`h_reply`
  hook, which fires on ANY firmware-issued reply regardless of action): NO
  reply follows `0x23` in this build either — consistent with the old
  fire-and-forget finding, now confirmed against the actual firmware code
  rather than just the software reimplementation.
- BUT the `0x23` payload itself does NOT match the documented FLYK format
  (`[u:f][v:f][host ptr][-1]`, screen coords) — this build's payload is
  `[1][incrementing event ID | 0x30000000][0]`, no screen coordinates. Wire
  context around it (`0x23`×2 → `list_remove`×2 → `0x1b`×2 → a 264-byte
  `0x1d`, matching the documented PSFX/particle-burst branch ≥160B) reads
  like a WEAPON-DISCHARGE NOTIFICATION with visual consequences (a projectile
  resolving + effect spawn), NOT a board-side "what's under the crosshair"
  query. So `0x23` in BT4.10/MUNGA is very likely NOT the same mechanism the
  1994-era FLYK notes describe under the same action number.

**Net conclusion: open.** Either this build's real pick/range mechanism lives
under a still-unidentified action, or BT4.10/MUNGA moved this computation
entirely host-side (matching BT411's own architecture) and `0x23` (whatever
it really is) is unrelated to targeting. Paused here per the user
(2026-07-21) rather than open a broader search without a specific pointer —
resume by identifying what wire traffic accompanies a KNOWN weapon fire in a
fresh, targeted capture (fire once, deliberately, with minimal other action
noise) rather than mining an existing capture for a guessed action number.

### 6e. Cockpit cage — intermittent, NOT YET CROSS-CHECKED
See `emulator/render-bridge/COCKPIT-CAGE-NOTES.md` (the OTHER renderer's
findings: PUNCH-texel windows, CAGE_TWIST_SIGN calibration, GLANCE_DCS
opt-in) and BT411's `cockpit-view.md` (the canopy stencil-cut kit, solved
there). Not yet cross-checked against what this firmware pipeline receives
on the wire for cockpit-frame geometry.