Files

184 lines
10 KiB
Markdown

# The VDB — VWE Video Display/Splitter Board
Reverse-engineered from the shared MUNGA_L4 driver (`CODE/RP/MUNGA_L4/
L4SVGA16.ASM` + `L4VB16.CPP`/`.HPP`) and the pod emulator's HLE
(`emulator/vpx-device/vpxlog.cpp`). **A physical VDB sample + its companion
Cirrus Logic SVGA card exist** (operator, 2026-07-10) — see the hardware
verification checklist at the end. This board is IDENTICAL between BattleTech
and Red Planet (shared driver); only the game-side display layout differs.
## What it is
An ISA card that **taps the companion Cirrus Logic SVGA card's pixel output**
off the VGA feature connector and fans that ONE framebuffer out to the six
secondary cockpit displays (5 mono MFDs + 1 color radar) via three RAMDAC-like
palette groups. It is a DUMB splitter: no framebuffer of its own, no
compute — it re-clocks the SVGA pixel stream and routes byte-lanes through
palette CLUTs to separate VGA heads. "Adam's port decoder design" (Adam G.,
VWE hardware) per the driver comments.
The game renders cockpit gauges to the Cirrus SVGA in a **640x480x16bpp**
mode (VBE mode 0x111; `L4GAUGE=640x480x16`). The Division/VPX card is separate
(the out-the-window 3D). So the SVGA framebuffer the VDB taps = GAUGES ONLY.
## Register map (definitive, from L4SVGA16.ASM)
I/O 0x300-0x31A. Three palette groups, each a 6-bit VGA-DAC-style CLUT. The
driver's port table (`L4VB16.CPP` ~3685) passes the group base as **base+2**:
```
Group base(+2) mask read-addr write-addr data
NativePalette 0x3C6 (the Cirrus's own VGA DAC — 0x3C6/0x3C8/0x3C9)
SecondaryPalette 0x302 0x302 0x303 0x304 0x305
AuxiliaryPal1 0x30A 0x30A 0x30B 0x30C 0x30D
AuxiliaryPal2 0x312 0x312 0x313 0x314 0x315
Clock divider: 0x319 write = high-color divider OFF (VWE_HC_OFF)
0x31A write = high-color divider ON (VWE_HC_ON) (any value)
```
Per-group offsets (VGA-DAC layout): **+0 pixel-mask, +1 read-address,
+2 write-address, +3 data** (auto-incrementing R,G,B triplets). DAC is
**6-bit**: `SVGAWriteFullPalette` does `shr al,2` (8→6) on write,
`SVGAReadFullPalette` does `shl al,2` (6→8) on read. Palette load =
set write-addr to 0, then `rep outsb` 256*3 bytes. Mask write = one `out`
to base+0. The game only WRITES the VDB (fire-and-forget, no status/ACK);
reads return DAC read-back.
## How the split works (the dumb part)
**High-color clock divider (0x31A on):** in 16bpp mode each pixel is two
bytes. The divider clocks the LOW byte and HIGH byte of each pixel into
SEPARATE palette lanes → separate heads. So:
- **LOW byte (bits 0-7) -> Secondary palette (0x302 group) -> one head**
- **HIGH byte (bits 8-15) -> Aux1/Aux2 palettes -> the other heads**
Each head is a physical VGA output off the splitter; an "octopus" cable fans
the (up to) three head signals to the six cockpit displays. Decode confirmed
live for BT: radar = low byte via Secondary; the 5 mono MFDs = high byte via
Aux1/Aux2 color channels. [[vdb-three-vga-head-decode]]
## How the game packs multiple displays into one framebuffer
The card is dumb, but the DRIVER is clever about loading the CLUTs so several
displays share the same pixels. Each cockpit display (`L4GraphicsPort`) has:
- **`bitMask`** — which framebuffer pixel BITS belong to this display.
- **`channelEnable`** — which DAC color channel it drives: Red / Green / Blue
/ AllChannels (+ ...TransparentZero variants that leave index 0 undefined).
`BuildSecondaryPalette` (low byte = `bitMask & 0xFF`) and
`BuildAuxiliaryPalette` (high byte = `(bitMask>>8) & 0xFF`) walk the display's
bits with a `BitWrangler` and write that display's color/ramp into its channel
of the shared 256-entry CLUT. So N mono displays coexist in one byte-lane by
owning disjoint bit groups and different color channels; the CLUT decodes each
combination back to the right per-display brightness. `BuildAuxiliaryPalette`
generates a linear brightness RAMP across the display's bit values (mono
MFD = beam intensity); `BuildSecondaryPalette` copies source colors (color
radar).
## Palette flash + fade (the pixel-mask cycling)
`SVGA16::UpdatePalette` (`L4VB16.CPP` ~4196): each palette can FLASH by
cycling its **pixel-mask** through `mask[]` states at `flashRate` (this is the
`0x302` mask writes seen cycling e.g. 0x3F/0x7F/0xBF/0xFF — a blink, NOT index
decode). **Only the Secondary palette is allowed to FADE** (its RGB scaled
over time). So the pixel-mask register is used as a per-display blink control,
not (usually) as a decode gate.
## Our emulation (vpxlog.cpp)
`vdb_write`/`vdb_read` mirror the register map (`VDB_BASE=0x300`,
`vdb_group_of`/`vdb_group_base`); `vdb_pal[0/1/2]` = Secondary/Aux1/Aux2.
`pal_draw` reads the live Cirrus framebuffer (`vga.mem.linear` at
`vga.config.real_start`, 640x480, stride 1280) and for each pixel looks up
the low byte via pal0 and high byte via pal1/pal2 (one color channel per mono
head, pentapus split). 6-bit DAC expanded to 8-bit. `VDB_PALDUMP=<prefix>`
dumps each group's 768-byte CLUT; pixel-mask writes now log their value; masks
default 0xFF; `VDB_APPLYMASK=1` ANDs the index with the mask (correct DAC
emulation, default off = BT-identical).
## RP vs BT — the decode problem, SOLVED (2026-07-10)
**ROOT CAUSE: an SVGA bank-switch GRANULARITY mismatch, NOT a VDB decode bug.**
RP's `REL410/RP/GAUGE/L4GAUGE.INI` `[640x480x16]` mode was byte-identical to
BT's EXCEPT `granularityInKB=4` (RP) vs `=64` (BT) -- same card (STB Horizon+
= CL-GD5434). The MUNGA gauge renderer (L4GREND.CPP reads L4GAUGE ->
L4GAUGE.INI -> SVGA16 ctor with that granularity) uses it to compute VESA
bank/page numbers for its paged writes down the gauge framebuffer. DOSBox-X's
CL-GD5434 bank emulation handles 64KB granularity correctly (BT gauges full)
but MISHANDLES the 4KB-granular paging, so RP's gauge writes scattered/landed
off-screen -- symptom: the visible page held only a top ~51-row strip while a
copy sat at VRAM 0x100000, and the heads decoded near-empty. Everything else
(low/high byte split, 3 palettes, stride, the whole VDB model) was ALREADY
CORRECT -- proven by BT and by the fix.
**FIX (config, no rebuild): set RP's L4GAUGE.INI `[640x480x16]`
`granularityInKB=64` to match BT.** Safe because it's the same card BT drives
at 64KB. VERIFIED live 2026-07-10: VRAM content scan went from `0x0 + 0x100000`
(top-strip + off-screen) to the FULL framebuffer `0x0..0x90000` all populated;
**operator confirmed the radar AND all MFDs render correctly** (checked vs
reference). RP cockpit displays now decode fully like BT's — RP VDB DONE.
The tooling that found it (all in vpxlog.cpp, default-off diagnostics, kept):
`VDB_PALDUMP` (dump the 3 CLUTs), pixel-mask value logging, `VDB_REALSTRIDE`
+ live scan_len/addr_add/real_start/mode logging, `VDB_SCAN` (VRAM
content-region locator -- THE decisive tool), `VDB_BASE` (read-base override),
`VDB_APPLYMASK` (honor the DAC pixel-mask). Note: proper long-term fix could
instead be DOSBox-X honoring 4KB CL-GD5434 bank granularity, but the INI match
is correct and sufficient.
## Physical hardware (documented from the operator's sample, 2026-07-10)
CONFIRMS the reverse-engineering below the chip level:
- **Companion video card**: **Jaton KY2-JAX-CVGA54PCI** — a legacy PCI Local
Bus VGA card, **Cirrus Logic CL-GD5434** chipset. Renders the cockpit gauges
(640x480x16, VBE 0x111) and feeds the pixel stream to the VDB via a **26-pin
ribbon off the VGA Feature Connector**.
- **The splitter brain**: **AMD/Lattice MACH130-15JC** CPLD — EE-CMOS, 64
macrocells, 15 ns max prop delay, 84-pin PLCC (2nd-sourced by Rochester
Electronics). This is "Adam's port decoder": it CONSUMES the feature-connector
video stream AND takes instructions from the **ISA bus** (the 0x300-0x31A
register file — palette loads, masks, and the 0x319/0x31A high-color clock
divider). It does the byte-lane split in hardwired logic (same for BT and RP).
- **The three DACs**: **3x Brooktree Bt477KPJ80** RAMDACs (256-entry, 6-bit/
channel, 80 MHz) = the three palette groups Secondary/Aux1/Aux2. The Bt477's
6-bit DAC is exactly why the driver does `shr al,2` (8->6) on palette writes.
Output routing (confirms the head map): **one Bt477 -> a VGA stream to the
color RADAR display** (= Secondary/0x302); the **other two Bt477s -> a DB25
connector that fans out to the 5 monochrome MFDs** (= Aux1/0x30A R,G = 2
lower MFDs, Aux2/0x312 R,G,B = 3 upper MFDs; 5 mono wires total, one channel
unused). The DB25 fan-out is the "octopus/pentapus" cable.
- **Host hard-reset circuit (remote reboot)**: an **HSSR-8060** — a 6-pin
opto-isolated solid-state relay (MOSFET output, Agilent/Broadcom) — with its
**output switch (pins 4 & 6) wired to a 2-pin header silk-screened
`RESET_OUT`**, which is in parallel with the PC's front-panel RESET switch.
So energizing the SSR's input LED = "pressing reset" = a hardware reboot of
the pod's host PC. **What drives the SSR input is still unknown (operator
researching).** This is almost certainly a REMOTE/AUTOMATED REBOOT path: an
arcade operator (or the ops console over the net) could hard-reset a wedged
pod without physically touching it. Candidate input drivers to trace: (a) the
MACH130 CPLD decoding a "reset" write in the 0x300-0x31A I/O space (a soft
command -> hard reset, e.g. a watchdog or console-commanded reboot); (b) a
discrete line from the network/console board; (c) a watchdog timer. NOTE for
the emulator: our HLE has no reset-out port yet; if the driver ever writes a
reset trigger it'd appear as an unhandled 0x300-0x31A write — worth watching
the vpxresp log for writes outside the known palette/clock offsets.
**RESOLVED (operator, 2026-07-10): the SSR INPUT is driven from a spare pair
on the RADAR display's 9-pin video cable.** So it is NOT software/CPLD-driven
— it's a COCKPIT-LOCAL reset button: the reset signal is carried UP the radar
cable's unused conductors from a button in the cockpit down to the SSR, which
closes RESET_OUT across the PC front-panel reset. Lets the operator hard-
reboot a wedged pod PC from the cockpit without opening the chassis. Pure
hardware; nothing for the emulator to model.
### Still worth capturing for the archive
- Photos of each chip + the board silkscreen; the DB25 pinout (which pin ->
which MFD) and the 26-pin feature-connector ribbon pinout.
- Whether the MACH130 JEDEC fuse-map is readable (a MACH130 can often be
read back unless secured) -- that IS the splitter logic, the ultimate
ground truth, and preservable like the RIO EPROM.
- The reset-tie circuit trace (operator's follow-up).
Related: [[vdb-three-vga-head-decode]], [[tesla-cockpit-emulator-state]].