- Cross-build against the Steam Runtime 4 arm64 SDK (build_linux.sh,
cmake/steamrt4-arm64.cmake): the executable sits at 0x98000000 next to
the game's 32-bit window, with the OpenXR loader and jsoncpp beside it
(RPATH $ORIGIN); build_linux.sh also packs a release archive.
- OpenXR on SteamVR: EGL surfaceless through XR_MNDX_egl_enable, the
Frame controller profile (both names), a flick or a wrist turn as the
Wii Remote's shake, the D-pad as look; the refresh rate goes to 120 Hz
through vrcmd where the runtime offers only its current rate.
- Timewarp in the app: SteamVR shows frames that arrive in time as they
were rendered, so at 120 Hz every game frame stood still for two
refreshes while the head turned. The game frame's eyes now go into a
scene set, and every refresh composites the last finished set turned
to its own head pose (timewarp.h, checked by tools/timewarp_test and
by the simulator's PETARI_XRSIM_WARP); sharpening runs once per game
frame; the pacing changes only after a new display period has held.
- Fixes found on the way: libstdc++'s <cstdio>/<cwchar> against the MSL
renames (compat.h), language archives a disc carries damaged (cook.py
takes another language's), host allocations on the DVD and DSP
threads, quitting without waiting on the game's threads, case of two
include names.
- ANLEITUNG-STEAM-FRAME.md: the player's guide in German.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The settings panel beside the pause menu held nine of the settings of
petari_vr.ini; the others needed the file edited over adb. Players asked
for all of them in the game, for a choice of language, for the camera's
turns the other way round and for the room around the giant screen.
- One table of the numeric settings (vr::Setting in vr_game.cpp: key,
range, default, how to read and change it) now serves the file's loader
and the panel. Values outside a setting's range are clamped instead of
ignored.
- The panel (vr_settings.cpp) is drawn from tables of rows (a switch, a
stepper or a slider) on four tabs: Screen, Diorama, Picture, Game, with
a Defaults button for the open tab. It writes only the lines of the
settings changed on it, in one pass, so the rest keep following the
defaults. The refresh rate and the highest resolution now apply at
once; the clock levels and the language at the next start.
- language: each disc's languages (European: English, French, German,
Spanish, Italian; American: English, French, Spanish) with their
folders; SCGetLanguage answers with the one asked for when the copy has
its texts. The European disc's EuDutch folder holds the English texts
again and is not offered.
- invert_camera (on by default): the game moves its camera round Mario
the way the D-pad is pressed, so the view turns the other way. With
the setting CameraLocalUtil's two tests swap sides (and CameraInfo's
two arrows), and the diorama's snap turn takes the other sign: pushing
the right stick right turns the view to the right. Pages and menus
keep their sides.
- passthrough (off by default): XR_FB_passthrough, with
com.oculus.feature.PASSTHROUGH declared (not required) in the manifest.
The feature and a reconstruction layer are made when the setting first
asks, and paused while it does not. While they run the passthrough
layer goes out first and the eye layer blends by its alpha; around the
giant screen the eye images are cleared to transparent, fading over
0.3 s. The diorama's eye images stay opaque.
Checked headless on the headset (every tab and control, the saved file,
the camera's direction with the setting off and on, French and Spanish
texts on the two discs) and in the app unworn (the language from the
file, passthrough started, paused and started again without an error,
the refresh rate changed from the panel). Measured there with
passthrough: the compositor's GPU time goes from 1.9 to 3.1-4.2 ms a
refresh; worn, a player's heavy scene fell from 60 to 44 frames a second.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A player with the American disc (RMGE01) reported a crash at the start.
The game picks the folder of its texts from the disc's game code and the
console's language (Language.cpp), and the port had the European code
RMGP01 fixed in low memory: on American files the game asked for
/EuEnglish/MessageData/Message.arc and stopped ("File isn't exist").
- dvd.cpp reads the region off the language folders in the disc's own
file table (sys/fst.bin: EuEnglish, UsEnglish, JpJapanese, KrKorean),
boot.cpp puts the matching code at 0x80000000 and SCGetLanguage answers
with a language that disc has.
- The setup screen shows a folder whose file table names none of those
folders, or whose copy lacks the texts, as "Unknown disc" instead of
letting the game start on it.
- FileRipper names the file it did not find before it panics.
- run_headless.sh and galaxy_tour.sh take the game files and the save
folder from HEADLESS_DATA and HEADLESS_NAND, to test another disc's
files beside the app's own.
Checked on the headset with both discs' converted files: the prologue
and Good Egg Galaxy, and on the American ones the two galaxies whose
stage files differ from the European disc's (Gold Leaf, Rolling Gizmo)
plus Gateway, Honeyhive, the observatory and Battlerock.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A player reported a game that works but has no stage music. Every stage
theme is a stream (.ast), and all streams could stop for the rest of the
session:
- JASAramStream follows a stream's voice and sets up its loop (or its
end) once the voice is within 400 samples of it. A voice found beyond
the loop end is a fatal error to it: the stream's voices are paused and
no stream is loaded any more until the game is restarted.
- The console's DSP renders 80 samples between two updates of the voices.
The port's rendered a whole frame of seven sub-frames on the sync
command, so the game saw a voice move 560 samples at a time and missed
the 400-sample window on about one pass in four. Which pass is fixed
for each track: the file select music stopped after 55 s (a new player
choosing a file), Gateway Galaxy's after 64 s, Honeyhive's after 83 s,
the title's after 2.5 minutes, Good Egg's after 10.5. Sound effects
and sequenced jingles went on.
The DSP now goes as the console's does: the sync command only says how
many sub-frames to render, and each is rendered when the game has updated
the voices and released the last sixteen of them (the end of
JASDSPChannel::updateAll), then reported with its mail. Notes and
envelopes get the console's 2.5 ms timing as well, instead of 17.5 ms.
Checked on the headset, headless and in the app: a minute on the file
select screen and then a galaxy (no stage music at all before, music
now), the loops found in the recordings by correlation with the disc's
samples, 14 minutes of self-play in Good Egg Galaxy across six loops, a
cutscene skipped into its boss music, and six other stages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Stereoscopic 3D (a switch on the VR settings panel, stereo_screen; off by
default): on the giant screen, frames with a 3D camera are replayed twice
from the game's own camera, the 3D draws shifted sideways in clip space by
their depth, and each eye gets its own picture as a quad layer of its own
(the Quest's compositor shows the two as one stereo quad).
- Depth: a point is drawn stereo_far - stereo_depth * (Mario's depth / its
depth) eye distances apart. The farthest things lie 0.95 behind the
screen, the most the eyes take; with the whole scene behind a screen
4.5 m away there was hardly any depth, so Mario stands in front of it
and the foreground comes further out. "3D depth" on the panel sets
stereo_depth (1.75, 0.25 to 3).
- The HUD floats in front of the scene it covers instead of sitting on the
screen behind it; the pointer's cursor lies on what it points at (the
game reports the depth under it, port_vr_pointer_depth), and with the
HUD on menus.
- The two pictures are 0.8 of the single one's size (stereo_resolution),
shown through the layers' image rectangles: 8.6 to 9.7 ms of GPU a frame
on the headset against 10.6 to 13.2 at full size, 60 frames a second.
- PETARI_VRSHOT_MS also saves the screen's layers, which the eye images do
not contain.
The app's space is the runtime's LOCAL space moved to the head and turned
to where it faces on each session's first tracked frame, after a
recentring, and when the tracking jumps in a session's first seconds.
LOCAL alone can be stale when the game starts (2.15 m from the head on a
launch right after the headset woke up): the screen came up beside or
behind the player until they recentred.
In the diorama a comet mission's screen filter is left out: on the HUD
panel it was a black patch in the middle of the view.
On the giant screen no laser is drawn from the controller: the game's
cursor shows where it points. It stays for the VR settings panel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The emulated accelerometer had its own axes (x right, y along the
pointer, z out of the button face) where KPAD's are x left, y out of the
button face, z along the pointer ((0, -1, 0) held level, KPADReset),
which the game then remaps (WPadAcceleration::update). A level
controller read as a remote standing on end: the surfing coach's "Point
your Wii Remote at the screen" never passed, and the Star Ball would
roll forward at full speed.
- wpad.cpp adds gravity in KPAD's axes. While the game steers by tilt
(port_input_use_tilt), it follows the right controller pitched up by
the game's neutral angle, so a level controller gives the grip the
game asks for; a controller whose orientation is unknown counts as
level. PortPadState carries which way is down from the controller
(xr_app.cpp) and, as accelerations, only the motion on top (shakes).
- Star Ball: neutral 80 degrees (90 - sCoreBaseDegreeYZ), not 100. Its
tutorial reads the tilt only while it checks that the remote is
raised, so aiming near the sign never counts as a shake.
- PETARI_INPUT scripts tilt the controller with PITCH= and ROLL=.
- docs/CONTROLS.md: the in-game instructions in controller terms.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A native port built on the Petari decompilation (previous commit): an
HLE platform layer (OS, DVD, GX, audio, input, NAND) under the game's
own code, an OpenXR presentation (a diorama around the player, a giant
screen, the HUD and settings as compositor layers, SpaceWarp, dynamic
resolution), Touch Plus controls, the tools that convert your own game
files (tools/cook, which also takes the few pieces of data the game's
executable holds from your sys/main.dol), and the port's changes to the
decompilation (`git diff petari-base -- decomp`). README.md explains
installing and building, docs/TECHNICAL.md how it works.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The sources of Petari (https://github.com/SMGCommunity/Petari), the
Super Mario Galaxy decompilation by the SMGCommunity, at commit
0a5c6608dc3b4c44a9aa68c2535d1793dc565f02, unchanged but for LF line
endings: LICENSE (as LICENSE-Petari.txt; CC0 1.0), include/, libs/ and
src/. Left out: the game data Petari keeps from the original
executable, the error message archive (include/Game/System/
ErrorArchive.hpp) and two tables (src/Game/System/GalaxyID.bcsv and
StoryEvent.bcsv); GalaxyQuest's converter takes them from the player's
own disc. The port's changes follow in the next commit, so
`git diff petari-base -- decomp` shows them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
