Initial D Arcade Stage 3 Recompiled

Initial D Arcade Stage 3 — Native SH-4 Recompilation (THIS VERSION IS OBSOLETE JOIN THE DISCORD FOR THE MUCH BETTER VERSION https://discord.gg/6uANCeZnG)

(https://discord.gg/U3vVMbd4t) JOIN THE DISCORD FOR NEW UPDATES An experimental static ahead-of-time recompilation of the **NAOMI 2** release of *Initial D Arcade Stage 3* (GDS-0033) for modern Windows PCs.

Public early demo available: this is unfinished playtest software, not a finished port. Start with the default 60 FPS mode. The public download does not include game data, a BIOS, card saves, custom music, logs, or extracted assets. You must provide your own legally obtained matching game files.

Download Public Early Demo v2489 for Windows x64

What this project is

• A native static-AOT recompilation of the original SH-4 program.
• A native NAOMI 2 runtime with Flycast-derived hardware knowledge where noted.
• A Vulkan renderer for the game's submitted ELAN/PVR work.
• A BIOS-free executable path with no SH-4 interpreter or JIT fallback.

It is not a Dreamcast port and the released product does not run the game through a NAOMI 2 BIOS. The setup utility uses a matching CHD and security PIC only to verify and locally extract the data required by the recompilation.

Quick start

1. Download and extract Initial.D.Arcade.Stage.3.Recompiled.v2489.zip from the v2489 prerelease.
2. Place your own matching files in the included game files folder:
   • gds-0033.chd
   • 317-0384-com.pic
3. Run Initial D Arcade Stage 3 Recompiled.exe.

The Windows launcher checks each input, explains which file is missing or incorrect, extracts the required files, and repairs an incomplete setup. Python is not required. Allow roughly 1.3 GB of temporary free space during first-run extraction. A NAOMI 2 BIOS is neither required nor accepted.

Public ZIP SHA-256: 6AF00678196E9E54AC1C625866D08D2CD9855D0F43866B5DE1EDCBD7DCE0EF93

Current integration progress (v2489 WIP and public early demo)

• BIOS-free translated SH-4 boot, menu, race, result, and tested save flows.
• Vulkan rendering with an authentic 60 Hz mode plus experimental 90/120 Hz presentation that keeps gameplay, physics, timers, input, and audio on the original 60 Hz guest cadence.
• Output resolution, widescreen, antialiasing, texture filtering, fullscreen, V-sync, FPS counter, and live HUD-position settings.
• An F1 menu with Video, Controls, HUD, Card, and Music pages.
• Keyboard, Xbox/XInput, and DirectInput wheel paths with persistent remapping; Xbox triggers are independent accelerator/brake axes.
• The F1 Controls page now accepts D-pad/left-stick navigation and A/B actions, and deliberate 35% axis travel is sufficient for remapping. The same shared XInput loader used by the product passed a no-window probe against a real attached Xbox controller on Windows.
• Adjustable force-feedback strength for supported DirectInput wheels.
• Live 0–100% Steering Smoothing for controllers and wheels; 0% preserves the original unfiltered response and the chosen value is saved.
• Native ARM7/AICA audio and exact-slot optional MP3/WAV/FLAC race-music replacement. Clearing a replacement restores the original game track and no original asset is overwritten. v2462 scans native Windows Unicode filenames, including Japanese track titles, without lossy ANSI path conversion.
• Card creation/selection in the visible card data folder. The last selected card is loaded on the next start, and changing cards displays an unsaved-data restart warning. Managed selections now remain portable when the extracted demo folder is moved or renamed.
• Post-race ejection now reports the required empty-reader/removal state before queuing the selected saved card for the next prompt. A bounded Bunta result run wrote exactly 207 bytes and crossed this reinsert transition with zero recognized runtime faults.
• The Windows launcher checks GitHub releases before starting the game. Newer versions are shown with their release date and Yes/No choices; an optional checkbox enables verified automatic installs on future launches. Updates preserve game files, cards, music, logs, and user settings and roll back if installation fails.
• v2489 archives extract into one Initial D Arcade Stage 3 Recompiled folder. Updates merge into the existing installation: new product files are added, versioned product files are replaced, and destination-only files plus user data remain untouched. The main entry point is Initial D Arcade Stage 3 Recompiled.exe; the native helper is now Initial D Arcade Stage 3 Recompiled Runtime.exe.
• A self-contained, Python-free setup and integrity checker for user-owned inputs.
• Cooperative renderer shutdown: the program stops new Vulkan presentation, joins the raster worker, and only then destroys its window.
• Single-instance enforcement prevents two recomp graphics contexts from running concurrently. Windows QA/watchdog exits use the same cooperative renderer drain as a normal close.
• Bounded Vulkan acquire, graphics-fence, presentation-fence, and startup upload waits; no queue-idle, device-idle, or infinite fence wait remains.
• Direct Vulkan presentation bypasses the CPU/GDI display copy and is the factory default for fresh v2489 installations. A saved Direct/Safe choice remains authoritative, and Safe remains a manual troubleshooting option.
• GPU projection, ELAN lighting, depth normalization, static-geometry reuse, packed topology, and persistent mapped geometry/uniform streams remove the previous per-frame CPU staging copies.
• Exact-matched static course/car batches retain their immutable topology eligibility summaries, avoiding repeated full-array scans on each 120 Hz presentation phase.
• The FPS overlay now reports actual swapchain output as OUT and genuinely distinct authentic/interpolated frames as NEW; exact repeats are excluded.
• Authored attract-mode cinematic mattes now expand to the real widescreen edges. Recognition requires an exact four-vertex mesh/material signature and proof that the projected quad spans the native 0..640 viewport, so HUD and world geometry retain their existing aspect behavior.

Verified progress

• All 48 defined route/branch targets have loaded and all 32 rival profiles have produced real movement through normal guest physics/input.
• All 16 unique Time Attack layouts and all 32 retained rival profiles (30 Legend plus two Bunta profiles) have separate natural, game-owned result evidence across the historical accepted ledger.
• The complete 70-row Time Attack/Bunta condition matrix now passes on the exact final v2449 executable: 62 Time Attack direction/weather/time rows and all eight Bunta courses. Every row matched its requested course, condition, state, and movement requirements; Time Attack direction was independently proven by three agreeing selector fields. All runs reported zero recognized faults and exited cooperatively.
• All 16 retained Time Attack layout routes were also renewed to a natural, game-owned result on that exact v2449 executable using normal 2560x1080 Direct Vulkan at 60 Hz. Every route showed authentic movement before the result, no forced outcome, zero recognized errors/runtime faults, a 59.8 FPS or better Vulkan minimum, and cooperative clean-session shutdown.
• Exact-v2453 rival-result renewal is now tracked separately from the historical mixed-checkpoint ledger. Evo 5, Evo 6, the Keisuke rematch, Kyoko, the Ryosuke rematch, Sakamoto, Smiley, the Sudo rematch, and Wataru each reached positive player travel before a natural, game-owned result. After each target armed, the log recorded zero requested or applied helper outcomes. All nine strict runs averaged 60.0 FPS for display and Direct Vulkan presentation, reported zero fault signals, and shut down cooperatively. This is 9/32 profiles renewed on the exact v2453 SHA-256, not a whole-ledger same-build claim.
• The fixed k_ez regression route produced 120 distinct motion samples in each of 23 complete two-second race intervals, with no repeated moving-race endpoints. A four-course priority matrix measured 119.8–120.0 FPS minimum presentation in its accepted moving windows.
• v2415 offline checks cover controller discovery/remapping policy, axis/button normalization and capture while the F1 menu is paused, exact custom music routing, presentation interpolation, ELAN/card/AICA behavior, offscreen Vulkan, card-eject classification, translation integrity, link freshness, the no-firmware product boundary, cooperative QA shutdown, and single-instance enforcement.
• The v2441 integration build passed the same complete offline suite, including a new BGR24/BGRA32 presenter channel-accuracy test. Four fresh v2441 Direct-Vulkan route runs then completed cooperative shutdown with exit code 0 and zero moving-race repeats. The exact 2560x1080 Akagi target sustained 120.0 FPS minimum/average for both display and Vulkan across 35 steady two-second samples, with 120 distinct generated motion samples minimum per bucket. Myogi and Akina/Akagi 640x480 controls measured 119.0–120.0 FPS minimum and also recorded zero steady-race repeats.
• The v2442 controller checkpoint preserved that renderer/runtime base and passed the complete offline suite. Its product-shared XInput probe loaded xinput1_4.dll, found the attached controller in slot 0, converted its neutral state to centered steering and released pedals, and left no game or Vulkan process running. Axis/button capture, lower-travel axis remapping, controller menu navigation, music, card, interpolation, lifecycle, link freshness, and the no-firmware boundary all passed their focused checks.
• The v2443 audio checkpoint opened the exact product 44.1 kHz stereo PCM format on the preferred Windows endpoint without emitting sound, and its null-sink cold run produced 7,080,718 post-mix signal frames with zero AICA drops before cooperative exit.
• The v2444 card-portability checkpoint rebuilt every presenter-dependent owner and passed the complete controller, physical XInput, AICA mailbox/SGC, Windows endpoint, exact custom-music, interpolation, Direct-session, ELAN, card, offscreen Vulkan, timing, lifecycle, translation, freshness, and standalone no-firmware suite. The clean public ZIP contains 14 audited files and no game data, BIOS, cards, music, logs, captures, or private paths.
• The v2445 launcher checkpoint maps the saved F1 Direct choice to the native swapchain path and Safe to the bounded GPU-readback path through one tested helper. Direct and Safe no-launch validation both passed, all maintained launchers parsed cleanly, and the native v2444 executable was unchanged.
• The v2454 card/updater checkpoint passed 27 exact card-device transactions, the complete offline product suite, updater selection/digest/extraction/ preservation/rollback tests, and the moved-launcher test. Its 17-file public ZIP was then extracted and re-audited with no game data, BIOS, card, music, log, capture, or private path. Bunta win and loss post-result runs plus a three-minute no-input attract run all exited cooperatively without a recognized crash or untranslated-runtime fault.
• The v2456 Direct-Vulkan retry checkpoint passed compilation, a 33-case restart/shutdown suite, 20 route-parser checks, 15 controller/music checks, an exact synthetic learned-path round trip, link-freshness checks, and the standalone no-firmware audit. The audited 17-file ZIP contains no game data, BIOS, PIC, cards, music, logs, captures, private racing paths, generated guest translations, or private host paths. Live acceptance of the corrected F1 Direct-retry flow remains pending and is not claimed by this checkpoint.
• The v2457 Direct-authoritative checkpoint passed compilation, a 34-case restart/shutdown suite, 20 route-parser checks, 15 controller/music checks, an exact synthetic learned-path round trip, link-freshness checks, and the standalone no-firmware audit. A live isolated updater acceptance selected v2456 from an older installation, verified GitHub's asset digest, installed it, preserved user data, and confirmed that an already-current installation is not offered the same update. No live Direct-renderer run is claimed for v2457.
• The v2459 checkpoint passed the complete private build suite plus the parallel-switchback learned-path regression. A bounded Akina Time Attack run passed the old 663.071 m stall, reached 1,488.022 m and a natural result, averaged 60.0 FPS during the race, generated zero repeated presentation frames, and exited cooperatively. The post-result/card flow reported zero recognized untranslated calls or runtime faults. The real archive also passed current-installer and exact-v2457-updater merge acceptances without launching the game.
• The v2460 checkpoint audited every retained runtime fault log against the compiled dispatcher and added four genuine missing SH-4 targets while rejecting null, odd, out-of-image, and corruption-derived addresses. The complete offline suite passed. Separate no-BIOS, muted, offscreen 60 Hz Akina and Usui runs reached 49.037 m and 53.181 m of authentic movement with zero untranslated calls, runtime faults, or repeated frames; both exited cooperatively and left no game process. The real archive passed current and exact-v2457 merge-update acceptance without launching the game.
• The v2461 checkpoint repaired 67 analyzer-proven non-NOP instructions that legacy translated functions placed after an early C++ return instead of in the SH-4 RTS delay slot. A fixed-point audit now reports zero proven omissions. The complete offline suite passed, and a strict 60/60 Hz Akina Time Attack gate followed its verified private course path for 1,488.022 m, reached one genuine result with no forced result, reported zero untranslated calls or runtime faults, and exited cooperatively. The path itself is not published.
• The earlier v2446/schema-5 checkpoint established independent three-field direction identity and stopped treating dry/wet condition meshes as road direction. The final-v2449 70/70 matrix below supersedes its mixed-build route count while retaining that fail-closed evidence policy.
• The v2448 integration build passed the complete BIOS-free product suite. A normal visible 2560x1080 Direct-Vulkan Akagi run then held 120.0 FPS minimum/average for both display and Vulkan across 21 moving-race samples, produced at least 240 distinct frames per two-second bucket, added zero repeats, advanced 129.204 m, and exited cooperatively with no unclean marker.
• The v2449 integration build corrected the centered-4:3 cinematic matte on a 2560x1080 BIOS-free attract run. Twelve captured scene points covered both the initial narrow-road matte and later changing-zoom car close-ups; the masks reached both output edges while Hor+ world rendering and unstretched UI remained intact. Six matched 640x480 scenes are byte-identical to v2448, proving native presentation is unchanged. A normal visible Direct run without capture readback held 119.5–120.5 FPS after warm-up and exited without a fault, process, or Direct-session marker. The final build also passed the complete product suite.
• Conservative BIOS-free v2449 route renewal now covers the entire defined 70-row matrix: 62 Time Attack combinations and all eight Bunta courses. A no-launch strict reanalysis also returned 70 passes, zero failures, and zero missing rows for evidence tied to the final executable SHA-256.
• In the preceding v2440 live RX 9070 XT Direct Vulkan run, a 75,000–127,000-vertex course/race sequence sustained 119.7–120.3 visible presents per second, crossed into the result/continue transition, and then completed every cooperative shutdown phase with exit code 0.
• During that live run the Direct path reduced sampled renderer/process CPU demand from roughly 1.56 host cores in Safe/GDI presentation to 0.2–1.2 cores depending on scene load. Memory and handle counts stabilized, and the Direct-session marker was removed on normal exit.
• A cross-machine renderer replay produced the same accepted frame SHA-256: 34D6B91C0550A5CC2A60D4B1F8930812908CEA6DCD6C354749A0881BE426D9E2.

These are targeted engineering results, not proof that every game combination is complete.

Known early-build limitations

• This is not release-quality or whole-game complete. Visual, audio, timing, transition, controller, card, and course-specific defects may remain.
• Use 60 FPS first. Higher-refresh presentation is experimental and may expose jitter, clipping, flicker, or unstable performance on unverified content and hardware.
• Unlimited presentation rate is not finished.
• The defined Time Attack/Bunta route-state matrix is current, but every car, opponent, long campaign/result/persistence permutation, and error branch has not been exhaustively completed on one integration executable.
• Physical Xbox discovery now has a product-shared hardware acceptance result; live axis/button movement, multiple controller models, wheels/force-feedback, and audible end-user audio behavior still need broader acceptance.
• The cooperative shutdown path passes source/offscreen checks and bounded live Direct-Vulkan runs. Repeated stress across more GPUs and drivers is still required after earlier host black-screen incidents. Please close the game normally and report the newest logs if a failure occurs.
• Clean-machine packaging needs broader community testing.

When reporting a problem, include the mode, course/opponent, direction, weather/time choice, display resolution, FPS mode, controller, exact point of failure, and the newest logs. Do not upload game files, extracted assets, card data, or copyrighted music.

See STATUS.md for the evidence ledger, the v2489 public checkpoint for the latest runtime, correctness, updater, and package facts, the v2462 public checkpoint for the preceding Direct-default and Unicode-music facts, the v2461 public checkpoint for the preceding return-delay-slot facts, the v2460 public checkpoint for the preceding target-closure facts, the v2459 public checkpoint for the preceding post-result facts, the v2458 public checkpoint for the preceding naming-transition facts, the v2457 public checkpoint for the preceding Direct-Vulkan/release facts, the v2456 public checkpoint for the preceding Direct-Vulkan retry facts, the v2454 public checkpoint for the preceding card/updater facts, the v2445 public checkpoint for the preceding launcher facts, [the v2449