CydandClaude Opus 4.8
cff783d2ad
The complete cap7 mission -- every command, all 8,562 draws -- executes in 39 seconds at 89.6M steps/s sustained (199,857 Python hook services). The replayed dict matches the QUEUE ground truth exactly; the old 'baseline' dict (16793/8397) is exposed as a budget-truncated artifact: the historic Python regressions hit the 2e9-step budget ~96% through and silently dropped the last 497 commands. This is the first complete execution of the whole mission. + M4-LIVE-SEAM.md (the remaining path to live DOSBox) and emu860c.step1() for hook-driven single-steps. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
55 lines
2.6 KiB
Markdown
55 lines
2.6 KiB
Markdown
# M4: the live DOSBox seam — design note (2026-07-19)
|
||
|
||
The last milestone of GPU-RETARGET.md: the decoded renderer running live behind
|
||
the game. All gates are now open: geometry = emu860c (351×, checkpoint-perfect),
|
||
raster = the GPU tile path (M1/M2 conformant), spec = igc_conformance.
|
||
|
||
## The loop
|
||
|
||
```
|
||
DOSBox-X (FLYK.EXE etc., unmodified)
|
||
│ port I/O 0x150-0x161 (C012 link) [VELOCIRENDER_PROTOCOL.md §7]
|
||
▼
|
||
virtual C012 link device [NEW — small FSM]
|
||
│ byte stream ⇄ message assembler (§1-§2 framing, verified vs real EXE)
|
||
▼
|
||
protocol server (Python, exists as emu_main) [boot handshake done]
|
||
│ wire commands -> the firmware receive path
|
||
▼
|
||
emu860c: VREND.MNG geometry stage [DONE — real-time 3.5×]
|
||
│ per-draw DMA stream + payload programs (captured live from C memory)
|
||
▼
|
||
igc_gpu frame loop: tile dispatch + readout [DONE for fixtures]
|
||
│ RGBA frame
|
||
▼
|
||
present: (a) vr_readpixels reply -> host SVGA [FLYK readback path]
|
||
(b) or the render-bridge window [existing live_bridge seam]
|
||
```
|
||
|
||
## What exists vs what's new
|
||
|
||
| piece | status |
|
||
|---|---|
|
||
| Wire protocol + framing | verified against the shipped EXE (protocol doc §9) |
|
||
| Boot handshake (iserver ×3, BTL/MNG swallow) | implemented in emu_main |
|
||
| Geometry at real-time | **emu860c** — done |
|
||
| DMA-stream capture at draw time | the send_capture pattern — port the w32
|
||
tap into emu860c (a C-side write-watch range on the queue pages, cheap) |
|
||
| Tile raster on GPU | igc_gpu / igc_gpu_frame — done for fixtures |
|
||
| Texel path (TXDN) | the remaining spec item (M3) — effects/bench render
|
||
without it; texture-mapped content needs it |
|
||
| C012 DOSBox device | NEW: six ports + byte FIFO in DOSBox-X (protocol §7
|
||
calls it "bounded and testable in isolation") |
|
||
| Present path | reuse render-bridge/live_bridge (frozen renderer window) or
|
||
implement vr_readpixels for FLYK's own SVGA display (§6: "expected") |
|
||
|
||
## Sequencing
|
||
1. **M4b — offline end-to-end**: wire capture in, frames out, no DOSBox:
|
||
emu860c runs a mission, the C-side DMA tap feeds igc_gpu_frame per draw,
|
||
frames dump to disk. Validates the whole chain at speed. (next)
|
||
2. **M4c — the C012 device** in DOSBox-X + the protocol server as a socket
|
||
bridge (the emulator process stays Python+C+GPU; DOSBox talks to it over a
|
||
local pipe — same shape as the existing namedpipe serial work).
|
||
3. **M4d — present**: vr_readpixels or the bridge window; frame pacing.
|
||
4. M3 texel path lands whenever the TXDN spec closes — orthogonal.
|