Tomb Raider II demo (Venice)
Core Design's 1998 Windows playable demo, tr2_demo_02.exe from
tombraiderchronicles.com, a WinZip SFX opened with 7z without running it.
Provenance and hashes: test/binaries/candidates/tomb-raider-2-demo/provenance.json.
Registry id tomb_raider_2_demo (localhost-only).
game/TOMB2.EXE909,824 bytes, sha2560811efa8...33aa, entry0x459630.- Data:
DATA\DEMOPC.DAT(script) andDATA\BOAT.TR2(the Venice level), mounted atc:\DATA\....readme.txtandSettings.lnkare not needed. - Imports DDRAW (DirectDraw2 + IDirect3D2), DINPUT, DSOUND, WINMM and
COMCTL32, which it uses for its options property sheet. The real
comctl32.dll from
test/binaries/dllsloads throughlib/dll-registry.js. Without it (a fresh worktree is missing the gitignored DLLs) the sheet falls back to WAT stubs and looks like a wizard.
Route (headless, default 200 ms/batch clock)
node test/run.js --app=tomb_raider_2_demo --quiet-api --quiet-blocks \
--max-seconds=60 --max-batches=26001 --stuck-after=0 --no-close \
--input=700:mousedown:379:398,720:mouseup:379:398,\
21000:keydown:37,22500:keyup:37,22600:keydown:38,26000:keyup:38,26000:png:out.png
| batch | screen |
|---|---|
| ~10 | options property sheet (Graphics tab: Hardware 3D Acceleration, 640x480 High Color, Z Buffer) |
| 700 | mouse click on OK at (379,398). --input=N:dlg-cmd:1 reaches the sheet's hwnd but does not close comctl32's own property sheet. |
| ~6000 | Venice alley, Lara idle. No title menu in this demo. |
| 21000+ | Left turns her round to face the camera; Up then runs her down the other alley, where a doberman attacks and the health bar appears |
No emulator change was needed: it ran on origin/main 9c9dc59d, which already had the TR3 fixes (MOV r32,CRn, the ACM driver enumeration, the D3D ZENABLE default). Holding Up from the start position moves her only a step and then she stops facing the slope ahead (not investigated); turning first and then running is the clean movement check.
Not evaluated
Audio, FPS, browser.
Evidence: scratch/runs/20261005T224500Z-tomb-raider-2-demo-claude202b4b39-venice-fcf8f2f1.
Performance: why it is slow, and which settings help (2026-11-14)
Evidence: scratch/runs/20261114-tr-perf/ (commit 522ab4035, canonical wasm
sha256 e017b4a4..., profiled with a --names build of the same source,
wa.named.wasm). Headless M1, laptop at load ~10-12, so only same-run ratios
and large effects below are trusted. Fixed work = batches 7000..10000 of the
route above with --tick-ms-per-batch=2 (Lara idle in the alley, 91 frame
ends in every 640x480 arm). --cpu-window user+sys CPU; never fps.
Settings live in HKCU\Software\Core Design\Tomb Raider II\System
(written by the options sheet on OK; nothing in lib/apps.js seeds them):
RenderMode (2 = Hardware 3D Acceleration, 1 = Software 256 Colour),
PerspectiveCorrect, Dither, ZBuffer, BilinearFiltering,
TripleBuffering, FullScreen, FullScreenWidth/Height/BPP, WindowWidth/Height,
DontSortPrimitives, TexelAdjustMode, Disable16BitTextures, FlipBroken.
Defaults after OK: HAL, 640x480x16, Z, dither, bilinear, perspective all on.
--reg-import of an edited --reg-export applies them (generator
reg-variant.js in the run folder). The sheet's mode list is ours
($handle_IDirectDraw_EnumDisplayModes): 320x200 8-bit and 640x480 and up,
each listed twice; there is no 16-bit mode below 640x480. Forcing
FullScreenWidth/Height to 320x240 or 512x384 in the registry still runs
640x480 and draws every texture with wrong colours (m-320.png / m-512.png):
an unlisted mode is not a usable setting.
| Arm (CLI) | CPU, 3000-batch window | per frame | Note |
|---|---|---|---|
| WAT software raster (CLI default), defaults | 2.46-3.3 s | ~29-36 ms | raster ~62% of CPU |
same, bilinear/perspective/dither off (lowq) |
raster -15% (profile ratio); single toggles within noise | ||
game Software 256 Colour renderer (RenderMode=1) |
1.18 s for 8 frames | ~147 ms | game's own x86 rasterizer, 5x worse |
--d3dim-gpu --headless-gl (browser default WebGL path) |
6.0-6.3 s | ~67 ms | _storeNative16 dominates |
| same, native16 storage skipped (experiment patch) | 2.2-2.4 s | ~25 ms | picture identical by eye |
Where the time goes:
- WebGL (browser default):
lib/d3dim-gpu.js_submitImmediateon a 16-bit target (since 23e0990db, 2026-11-06) draws every primitive alone (blended, or a draw of <32 primitives) or in groups of 32, and after each runs_storeNative16:copyTexSubImage2Dof the target plus a quantizing full-target quad, with agl.getError()per submission. TR2 issues ~116 draws / ~480 triangles per frame, so that is hundreds of copy+quad passes and ~116 syncgetErrors per frame. Whole-rundrawMs15.0 s with it, 4.1 s without (TR3 title/attract: 7.1 s vs 2.1 s); window CPU 2.7x. In Chrome eachgetErroris a GPU-process round trip, so the browser cost is worse than this native-GL measurement. The 2026-11-07 browser run's 65 flips in 42.7 s (tr2-webgl-gameplay-20261107.md) is consistent with it. Every Flip also reads the frame back to the DIB and the next Clear uploads ~250 rows back. - Software raster (
?d3d-renderer=software):$viewport_draw_textured_span(src/09ab-handlers-d3dim-core.wat) is 46% of all samples / ~95% of raster time: a generic per-pixel loop with twof32.divper pixel, and calls per pixel for combine, depth compare, stencil/stipple and blend. - Guest x86 is small: th_ handlers + uop + x87 folds ~12-15% of CPU. The interpreter is not the bottleneck in either renderer.
Candidate fixes, by expected gain: (1) native16 only where it is observable:
quantize once before a blended draw that reads unquantized pixels, and at
fence/readback, and check getError once per fence (expected: the WebGL path
back to ~the no-native16 arm, ~2.7x less CPU per frame); (2) a specialised
span for TR's common state (16-bit RT, 16-bit texture, modulate, Z LE,
no stencil/stipple) with perspective divides every 8-16 pixels; (3) enumerate
512x384 and 320x240 16-bit modes so a lower render size becomes a real
option (software raster is per-pixel bound). In-game quality toggles are
worth ~15% at most on the software raster and nothing on WebGL; do not use the
game's Software 256 Colour renderer.
Fix: full colour on the GPU (2026-11-14)
lib/d3dim-gpu.js now keeps RGBA8 on the GPU for 16bpp targets and reduces to
native bits only at readback; the per-primitive native store survives only for
destination-alpha blends on A1R5G5B5/A4R4G4B4 (and exactNative16: true, used
by the bit-exact contract tests). Same route, same 91 frames, same load:
old 5.45 s CPU / 6.36 s wall, new 2.10 s / 2.05 s; FullScreenBPP=32
(true colour) 2.03 s / 1.94 s. Frames look identical by eye.
Evidence scratch/runs/20261114-tr2-full-colour/ (hw16, hw32, old16 logs+PNGs).