Otogi: Myth of Demons — technical preview v0.1

A static recompilation of the original Xbox release of Otogi: Myth of Demons (FromSoftware / Sega; Japan 2002) into native executables for Linux and Windows, built with the xboxrecomp toolkit. The game's own x86 code is translated to C and compiled natively — there is no CPU emulator, no CPU interpreter, and no hand-written substitute for the title's own logic.

This is a research preview, not a playable port. It boots the retail XBE, runs the title's own main loop, plays the menu BGM and the attract video, and reaches the New Game menu, save creation and loading, the main menu, the start-game stage and the pause menu. On the stage the bamboo grove, the ground haze, the light shafts, the HUD and the player's 3D model are drawn. The remaining rendering gap is the title's movie — see docs/KNOWN_ISSUES.md.

AI-assisted. Code, debugging, and documentation were produced with substantial AI assistance; the maintainer reviewed and verified every change and claim in this repository.

<p align="center"> <img src="docs/img/title-screen.png" alt="The Otogi title screen, rendered by otogi1" width="720"> </p>

Table of contents

• Status
• Screens
• Known limitations
• Requirements
• Step by step
• Switches
• Documentation
• How to help
• Contributing
• Licensing

Status

Status as of 2026-10-30. Runs are headless, on Linux, through the OpenGL push-buffer back end: RECOMP_PB_EXEC=1 RECOMP_USB=1 RECOMP_AUTOPRESS=1 ./build-linux/bin/otogi1 drives the title from the logos to the start-game stage with no pad and no window. A physical gamepad still works and is what the earlier screenshots used.

Component	State	Notes
Builds	Works	Native Linux binaries; MinGW-w64 cross-build and MSVC 2022 build for Windows verified. The MSVC build needs one source fix in the toolkit (docs/KNOWN_ISSUES.md). tools/smoke.sh build-linux passes.
Boot	Works	Loads the XBE, reaches [AV] SetDisplayMode, runs the main loop, shows the logos and the title screen. No crash.
Kernel	Works	Threads, KEVENTs, the NV2A fence event, and asynchronous NtReadFile (RECOMP_SYNC_IO=1 restores inline completion).
File I/O	Works	All XBE sections, assets, and the title's packed .dev/.000 archives.
Audio	Works (BGM)	RECOMP_AC97_READY=1: streamed BGM with measured non-zero PCM. The DSP is a passthrough stub, so effects are not emulated.
Video	Works	The title's own XMV decoder plays the looping attract video. The toolkit's host video player is a POSIX stub and is not required.
Input	Works	Physical gamepad with RECOMP_USB=1 (OHCI controller and emulated pad). Headless runs use RECOMP_AUTOPRESS=1, which taps START then A. Drives the title, New Game, save creation and loading, the main menu, the start-game stage and the pause menu.
Saves	Create and load work	Creating and loading a save work in the menus. Where the save file is written, and its format, are not recorded yet.
Rendering	Partial, on the GPU	The push-buffer executor hands batches to an OpenGL 3.3 back end (src/nv2a/nv2a_pb_gl.c in the toolkit): logs, the front-end menus, the intro cutscene, the stage's world and the HUD all render through it, including the title's register combiners (the pixel stage, without which menu text carried black boxes), and the frame is read back into the title's own colour surface. Remaining gap: the title's movie; the effect planes' alpha is closed in docs/KNOWN_ISSUES.md. RECOMP_PB_GL=0 keeps the CPU rasteriser.
Framerate	Playable	Measured on the RX 570 / Ryzen box: 90-120 fps in the menus and the attract loop, 25-29 fps in the intro cutscene, ~40 fps on the start-game stage (12 fps before the vertex stage was decoded once, fetched selectively and run on a thread pool).

Details and exact log lines: docs/BOOT_LOG.md. Open blockers: docs/KNOWN_ISSUES.md.

Screens

Screenshots from the maintainer's runs. Unless a caption says otherwise, the switches are RECOMP_USB=1 RECOMP_FB_WINDOW=1 RECOMP_PB_EXEC=1 RECOMP_AC97_READY=1 with a physical gamepad.

Create or load save	Main menu
<img src="docs/img/save-screen.png" alt="Save-creation screen" width="420">	<img src="docs/img/main-menu.png" alt="Main menu" width="420">
Save creation. Loading a save also works in the maintainer's runs; no load-screen capture yet.	Main menu. Reached after save creation or loading.
<img src="docs/img/start-stage.png" alt="Start-game stage" width="420">	<img src="docs/img/pause-menu.png" alt="Pause menu" width="420">
Start-game stage. The player's 3D model is drawn. The stage world is not: the only scenery visible is bamboo.	Pause menu. Opened from the start-game stage.

<img src="docs/img/gl-menu-2026-10-30.png" alt="New Game / Load Game menu through the OpenGL back end" width="640">

The front end, 2026-10-30. New Game / Load Game, drawn through the OpenGL push-buffer back end with the title's own register combiners and the player model behind the text. Captured headlessly with RECOMP_FB_DUMP (RECOMP_PB_EXEC=1 RECOMP_USB=1 RECOMP_AC97_READY=1).

<img src="docs/img/gl-stage.png" alt="The bamboo grove on the start-game stage" width="640">

The start-game stage, same build. The grove, the haze and the light shafts are the title's own geometry, drawn by the GPU; the floor under the player is still missing and reads as the clear colour. That is the open blocker. The capture replaces the 2026-10-28 one, which was taken while the front-face winding was still mirrored and so showed the inside of every trunk.

The open blockers behind these limits are tracked in docs/KNOWN_ISSUES.md.

Known limitations

• The stage's floor is not drawn. The player's 3D model, the HUD, the bamboo, the haze and the light shafts are drawn on the start-game stage; the floor under the player is not, and reads as the clear colour. Both rasterisers (CPU and OpenGL) show the same thing from the same decoded batches, so the fault is in what the executor builds or in what the title submits, not in the back end. The visibility-question path (ZPASS_PIXEL_CNT) used to answer 0 to every query whenever a back end drew, which is an instruction to cull; it is answered from the GPU now. See docs/KNOWN_ISSUES.md.
• Black screen after New Game. Two runs on 2026-10-24 showed a black screen after the menu pan-out delay: one plain, one with RECOMP_GUEST_LOCK=1. Neither had RECOMP_AC97_READY=1. Repeated maintainer runs with RECOMP_AC97_READY=1 did not show it. That switch is the only difference recorded, so it is a correlation, not an established cause. Stderr from a black-screen run has not been captured. See docs/KNOWN_ISSUES.md.
• The dispatch gate is a project default, not a toolkit one. src/game/main.c sets RECOMP_DISPATCH_GATE=1 because this title's USB stack self-links its software queue on a multi-core host without it (docs/BOOT_LOG.md, 2026-10-26). RECOMP_DISPATCH_GATE=0 still overrides. The gate is opt-in in the toolkit because it changes when DPCs and interrupts run for every title.
• The OpenGL back end approximates the material path. It samples texture stage 0 and multiplies by the interpolated vertex colour; the CPU rasteriser evaluates the title's register combiners and fog. Titles or scenes whose look depends on those draw differently, and the stage world's materials are untested against a correct reference.
• The save file is not documented. Creating and loading a save work. The file's location and format are not recorded yet.
• No DSP effects. The DSP is a passthrough stub, so the mixed output is produced but the title's audio effects are not emulated.
• The MSVC 2022 build needs one source fix. With the clock_gettime guard in docs/KNOWN_ISSUES.md applied it builds, boots and produces audio; without it the build stops in the toolkit's input layer. The MinGW-w64 cross-build is the path verified without a source change.

Requirements

• Linux x86-64, or Windows x86-64. The MinGW cross-build is verified without a source change; the MSVC 2022 build is verified with the one clock_gettime guard in docs/KNOWN_ISSUES.md.
• CMake 3.20+, a C11 compiler, and Ninja. The build commands below pass -G Ninja; drop that flag to use CMake's default generator.
• On Linux only: SDL2 and libepoxy for the video, audio and window path. Windows uses Direct3D 11 and XAudio2 from the Windows SDK instead.
• Python 3.10+ with capstone — regen.sh drives the toolkit's Python tools.
• Your own legally obtained copy of the game, as an Xbox disc image (.iso). Step 3 unpacks it into game/; its default.xbe is what regen.sh reads.
• The recompiled C in src/game/recomp/gen/, produced by regen.sh. It is generated, not tracked.

The runtime toolkit is a submodule at xboxrecomp/; no separate checkout is needed.

On Windows the MSVC path needs Visual Studio 2022 with the Desktop development with C++ workload (or the standalone Build Tools with the VCTools workload), the Windows SDK, and CMake. A Visual Studio install does not imply the C++ toolset — an install without it has MSBuild and no cl.exe. Check:

& "${env:ProgramFiles(x86)}\Microsoft Visual Studio\Installer\vswhere.exe" `
    -latest -products * -require Microsoft.VisualStudio.Component.VC.Tools.x86.x64 `
    -property installationPath

Empty output means the workload is missing. Reserve about 8 GB free: with less, the workload installer's pre-check fails with exit code -2147024784 (0x80070070, ERROR_DISK_FULL) and prints nothing to the console. The standalone Build Tools plus the SDK are about 5 GB installed.

Step by step

The same steps this project is developed against, in order. Commands are for a POSIX shell — Linux, or Git Bash on Windows.

1. Install the prerequisites.

   Tool	Version	Notes
   Python	3.10 or newer	runs the toolkit's Python tools
   capstone	5.x	python3 -m pip install --user capstone
   CMake	3.20 or newer
   C compiler	C11	GCC or Clang on Linux; MinGW-w64, or MSVC 19.4x from Visual Studio 2022 Build Tools, for the verified Windows build
   Ninja	recommended	the commands below pass -G Ninja; drop it to use the default generator
   SDL2, libepoxy	Linux only	video, audio and window path

   $ python3 -c "import capstone; print(capstone.__version__)"
   5.0.1
   $ cmake --version | head -1
   cmake version 3.31.6
   

2. Clone the repository. The toolkit is a submodule pinned to this project's fork — no second checkout, no patch to apply.

   git clone --recursive https://github.com/sigods/otogi1
   cd otogi1
   

3. Unpack your disc image into game/, using the toolkit's own XDVDFS reader. Otogi's USA image carries a video partition in front of the filesystem, so tools that assume the filesystem starts at offset 0 fail; this reader probes the known partition bases.

   $ cd xboxrecomp
   $ python3 -m tools.xiso unpack "/path/to/Otogi - Myth of Demons (USA).iso" -o ../game
   $ cd ..
   

   You should end up with game/default.xbe next to chr/, map/, sfx/, movie/ and the rest — 1,121 files, about 2.2 GB.

4. Regenerate the C.

   $ ./regen.sh
   ==> xbe_parser
   ==> disasm
   ==> func_id
   ==> abi_analysis  (never skip: silent wrong signatures otherwise)
   ==> recomp  (--split 250: prevents OOM)
   ==> done. Now: cmake --build build-linux
   

   This writes src/game/recomp/gen/, which is gitignored. It takes a while: the recompile step alone is 5–15 minutes (docs/PIPELINE.md). It is only needed once per disc. It generates C and stops — the cmake --build in its closing line is a suggestion, and step 5 does the configure that makes it work.

5. Build.

   $ cmake -S . -B build-linux -G Ninja
   $ cmake --build build-linux -j8
   

   The executable is build-linux/bin/otogi1, about 19 MB.

6. Run it from the project root.

   $ ./build-linux/bin/otogi1
   

   The binary opens game/default.xbe relative to the current directory. Within a few seconds the log shows the title's own display-mode call and the kernel event it builds for the GPU fence:

   [AV] SetDisplayMode mode=0x080F0D12 format=0x00000012 pitch=2560 fb=0x00404000
   [KERNEL] event 0x00106B6C built by the title: host event created (notification)
   

   After that it sits in the title's main loop. Ctrl+C ends it.

   ./tools/smoke.sh build-linux re-checks the claims above against an existing build. It runs the title twice. The plain run captures audio, and it asserts that the title reaches [AV] SetDisplayMode, that it is still running when the timeout stops it (no early exit, no segmentation fault), that the audio is non-silent, and that the OHCI controllers stay off. The RECOMP_USB=1 run asserts that the OHCI controllers come up and that there is no segmentation fault.

7. Turn on the implemented parts. Each is opt-in. The first line is the reference run, with a physical gamepad:

   $ RECOMP_USB=1 RECOMP_FB_WINDOW=1 RECOMP_PB_EXEC=1 RECOMP_AC97_READY=1 ./build-linux/bin/otogi1
   $ RECOMP_FB_WINDOW=1 RECOMP_AC97_READY=1 ./build-linux/bin/otogi1 # window + BGM
   $ RECOMP_PB_EXEC=1 ./build-linux/bin/otogi1 # GPU command stream
   $ RECOMP_PB_EXEC=1 RECOMP_USB=1 RECOMP_AUTOPRESS=1 ./build-linux/bin/otogi1 # headless, drives itself to the stage
   $ RECOMP_PB_EXEC=1 RECOMP_USB=1 RECOMP_AUTOPRESS=1 RECOMP_PB_GL=0 ./build-linux/bin/otogi1 # same, CPU rasteriser
   

   Drawing goes through the OpenGL back end by default and needs a display; RECOMP_PB_GL=0 is the CPU rasteriser, about ten times slower on the same scene.

   On a headless box, run the window under Xvfb. These are the switches the README relies on; AGENTS.md § 7 lists the rest, including the diagnostic ones added during bring-up.

Windows

Steps 1–3 are the same. regen.sh is a POSIX shell script, so run step 4 once from Git Bash or WSL. Then pick a toolchain:

• MinGW-w64 cross-build — verified. It links, and under Wine it reaches the title's display-mode call without faulting:

  $ cmake -S . -B build-win -G Ninja \
      -DCMAKE_SYSTEM_NAME=Windows \
      -DCMAKE_C_COMPILER=x86_64-w64-mingw32-gcc \
      -DCMAKE_FIND_ROOT_PATH=/usr/x86_64-w64-mingw32 \
      -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER
  $ cmake --build build-win -j8
  

  The result is build-win/bin/otogi1.exe.- MSVC 2022 — verified, with one source fix. Built with Visual Studio Build Tools 2022 (MSVC 19.44, Windows SDK 10.0.26100) and CMake 3.31. Until the clock_gettime guard in docs/KNOWN_ISSUES.md lands in the toolkit, the build stops in xboxrecomp/src/input/keyboard.c. The Windows SDK supplies Direct3D 11, DXGI, d3dcompiler and XAudio2, so no SDL2 is required:

  cmake -S . -B build -G "Visual Studio 17 2022" -A x64
  cmake --build build --config Release --parallel
  

  The result is build\bin\Release\otogi1.exe, about 11 MB. CMake finds a non-default install through vswhere; when it does not, -DCMAKE_GENERATOR_INSTANCE="C:\BuildTools" names it. The build also produces the toolkit's four d3d8_* smoke executables and about a hundred warnings (C4005, C4100, C4996, C4457) that are not failures. /bigobj is already set for otogi1 and is what lets the generated chunks compile — keep it.

Run from the project root, not from inside build\: the binary opens game/default.xbe relative to the current directory.

./build-win/bin/otogi1.exe # MinGW cross-build
.\build\bin\Release\otogi1.exe # MSVC

Switches

Every switch is an environment variable, read once at startup. The table below lists the switches these instructions use; AGENTS.md § 7 has the complete set, including the diagnostics added during bring-up.

Variable	Effect
RECOMP_FB_WINDOW=1	Open the SDL framebuffer window and present frames — Linux only, off by default
RECOMP_PB_EXEC=1	Software NV2A push-buffer executor
RECOMP_USB=1	OHCI controller and emulated pad (RECOMP_AUTOPRESS=1 sends START+A)
RECOMP_AC97_READY=1	MCPX APU and host audio output (SDL2 on Linux, XAudio2 on Windows)
RECOMP_AUDIO_WAV=path	Capture the mixed output to a WAV file
RECOMP_SYNC_IO=1	Complete NtReadFile inline (pre-audio-fix behaviour)

Documentation

• docs/INDEX.md — start here: which document to read, and when.
• docs/FAQ.md — the short answers, if you want them first.
• docs/ARCHITECTURE.md — which layer owns a behaviour.
• docs/PIPELINE.md — the toolchain, step by step.
• docs/RUNNING.md — building, running and recording the executable, in long form.
• docs/TITLE_OTOGI.md — title facts and open unknowns.
• docs/KNOWN_ISSUES.md — what is blocking progress now.
• docs/BOOT_LOG.md — the append-only findings ledger.
• ROADMAP.md — milestone status.
• CHANGELOG.md — what changed, and when.

How to help

• If the black screen after New Game appears, capture stderr (./build-linux/bin/otogi1 2>stderr.txt) and send the lines around [KERNEL] ordinal 175 (docs/KNOWN_ISSUES.md).
• Name kernel ordinal 175 from the toolkit's export table under xboxrecomp/ and record it in docs/BOOT_LOG.md.
• Carry the clock_gettime guard from docs/KNOWN_ISSUES.md into the toolkit and offer it upstream — it is the only thing between the MSVC 2022 build and a clean build.
• Run ./tools/smoke.sh on another distribution and report any assertion that fails.
• Correct or extend a page in docs/.
• Pick a milestone in ROADMAP.md and open an issue naming it.

Read CONTRIBUTING.md first; it has the two rules that matter.

Contributing

Corrections, bug reports and pull requests are welcome — read CONTRIBUTING.md first. Two rules matter most: never commit game data (the repo hygiene workflow, .github/workflows/hygiene.yml, fails if game/, .iso, .xbe, .xmv or .xps files are tracked), and back every claim with the evidence behind it.

Credits, the AI assistance disclosure and upstream attribution are in CONTRIBUTORS.md. Working with an AI agent? Point it at AGENTS.md first; it lists the pipeline gates that fail silently.

Licensing

This project's own source is GPL-3.0-only — see LICENSE. No game data, disc image, XBE or generated recompiled source is distributed here; you supply your own copy (see Requirements).

xboxrecomp/ is a submodule pinned to this project's fork of the toolkit, carrying local changes; xboxrecomp/UPSTREAM.md records the upstream base. Most of the toolkit is MIT-licensed. Some files derived from xemu are LGPL-2.1-or-later, not MIT: the MCPX APU sources in xboxrecomp/src/apu/ and the NV2A register header xboxrecomp/src/nv2a/nv2a_regs.h. Not every file in those directories is xemu-derived. NOTICE lists them with their copyright holders; the per-file list is in xboxrecomp/NOTICE, and the licence text is in xboxrecomp/LICENSES/LGPL-2.1.txt.