src/ps5_forwarders.{h,cpp}: a tile is /data/homebrew/<id> (PPSA99170 to PPSA99269, an
id no folder has and ShadowMountPlus knows nowhere) with the tile app and payload of
step 1, libc.prx, a param.json for its id and name, an icon and forward.txt (the
arguments, which also mark the folder as a tile: nothing else is listed, changed or
removed). ShadowMountPlus's HTTP API (127.0.0.1:10101, its docs/api.md 1.7) puts it on
the home screen: manual/add at once, installed when no game runs.
Routes (src/webui_ps5.cpp): GET/POST /api/forwarders (the list with ShadowMountPlus's
state, the cores, each tile's state; create), POST /api/forwarders/edit?id=,
PUT/GET /api/forwarders/icon?id= (a 512x512 PNG, up to 2 MiB), POST
/api/forwarders/delete?id=. The library JSON carries each game's core, to prefill it.
ShadowMountPlus refuses to uninstall while a game runs, and RetroArch is one whenever
the WebUI is up: a removal of an installed tile takes it out of the manual list at
once (so no scan installs it again) and is finished by the WebUI's daemon once the
title has closed (daemon/webui_daemon.cpp -> finish_pending): uninstall, then delete
the folder. A new name or icon on an installed tile is re-registered there the same
way (uninstall, back in the manual list, scan), and every tile's app and payload are
brought up to the build's.
Found on the console and handled: ShadowMountPlus 1.7-beta5fix1 can drop a source from
its manual list as it installs it (its next scan reads app.db before the install shows
there); the tile stays installed, and a re-registration adds it back first. Its game
list can take longer than 5 s after RetroArch starts: 20 s now, and a list that does not
arrive says why instead of showing every tile as waiting.
Evidence:
- host: tests/test_webui_forwarders.py against a stand-in for ShadowMountPlus's API:
create (first free id past a plain folder and a title ShadowMountPlus knows), the
files and param.json, validation (name, content path, "..", absolute path, newline,
core), edit, icon (size and format refused, stored, served), removal while busy
(pending, out of the manual list) finished by finish_pending, at once when not
installed, re-registration after the source was dropped, binaries refreshed,
ShadowMountPlus absent - 3 tests OK.
- console (base PS5 13.40, ShadowMountPlus 1.7-beta5fix1), through the WebUI from the PC:
GET /api/forwarders: ShadowMountPlus answers, the test tile PPSA99170 installed;
POST create -> PPSA99171 (eboot.bin, launch.elf, libc.prx, icon0.png, param.json
UP9000-PPSA99171_00-FORWARDER0099171, forward.txt with five arguments), in the
manual list; RetroArch closed -> ShadowMountPlus "installed game PPSA99171";
the tile opened -> tile.log "110832 sent to 127.0.0.1:9021 (errno 0)", launch.log
five arguments, RetroArch "forwarder: /app0/content/snes/<game> with
/app0/cores/snes9x_libretro.so, closing after it"; rename and icon -> updating,
delete PPSA99170 -> pending (manual list empty); RetroArch closed -> daemon
"forwarder PPSA99170: removed", "forwarder PPSA99171: installed again with its new
name or icon"; /user/appmeta/PPSA99171 titleName "Forwarder renamed" and the new
icon (first pixel 20 60 d0), /user/appmeta/PPSA99170 and /data/homebrew/PPSA99170
gone, app.db 33 -> 31 -> 32 entries; the daemon also replaced PPSA99170's older
launch.elf with this build's.
- gates: format, unit (209 tests), build, integration (209 tests), evidence - PASS.
A home screen tile is a title of its own (PPSA99170 to PPSA99269) that starts one
game in RetroArch. Titles cannot launch titles (sceSystemServiceLaunchApp returns
0x8094000F from one, sceLncUtilLaunchApp is a null import), so the tile hands a
payload to the console's ELF loader on 127.0.0.1:9021 and exits:
- forwarder/tile.c: the tile's eboot.bin. Reads its title id from
/app0/sce_sys/param.json, stamps it into /app0/launch.elf's slot, sends it to
the loader, logs /app0/tile.log, and leaves through LoadExec("exit") (a return
from main ends in a SIGSYS error notice).
- forwarder/launch.c: the payload. Reads /data/homebrew/<tile>/forward.txt (one
argument a line: --rom, --core, --exit-after-game, src/frontend_mode_ps5.cpp),
waits for the tile to leave, and starts PPSA99169 with those arguments through
sceLncUtilLaunchApp (sceSystemServiceLaunchApp as the fallback).
- tools/build-forwarder.sh builds both; tools/build-title.sh puts them in
dist/PPSA99169/forwarder/ and counts them in the title's identity; the notices
cover both files.
Evidence:
- host: python3 -m unittest tests.test_forwarder (the tile stamps and sends the
payload byte for byte, refuses a lowercase id, a missing or a doubled slot; the
payload passes the five arguments, falls back to SystemService, and an
unstamped payload starts nothing) - OK.
- console (base PS5, 13.40, ShadowMountPlus 1.7-beta5fix1), the test tile
PPSA99170 registered through ShadowMountPlus's manual list:
tile.log "tile PPSA99170: launch payload 110832 bytes, 110832 sent to
127.0.0.1:9021 (errno 0)"; launch.log five arguments, sceLncUtilLaunchApp
00002018; RetroArch's trace "forwarder: /app0/content/nes/<game> with
/app0/cores/fceumm_libretro.so, closing after it"; the tile exited cleanly
(no error notice), back to the home screen after the game.
- gates: format, unit (206 tests), build (tile 11,123 B, payload 110,832 B),
integration (206 tests; a first run timed out starting two scraper test
servers under load, both passed alone and on the rerun), evidence - PASS.
The workflow parked from mihawk-99/PS5_RetroArch#34 gave the title build
PS5_SCREENSCRAPER_DEVID and PS5_SCREENSCRAPER_DEVPASSWORD from repository
secrets. A public repository's workflow artifacts can be downloaded by
anyone, and the account is never published, so the parked copy no longer
passes them and its header says why. Nothing runs it; no value was ever
in the file.
Evidence: grep finds ScreenScraper only in that header line; format PASS.
Build 838fd9a3 (20dd7fe with both frontends and both optional cores) on
the owner's base PS5 (1440p, 120 Hz, VRR), offline:
- framegen-off-presents-once (0132): frame generation off with a 2 left
in retroarch.cfg logs "Off: each game frame is presented once (2
presents were kept)" and runs 602 frames in 10 s; turned off from its
XMB tab mid-game, the video output reinitializes (a 2.5 s gap) and the
same happens, 602 frames in the next 10 s.
- chd-flac-hash (0133): a CHD chdman made from one of the owner's
PlayStation backups (cdlz, cdzl, cdfl). Build 22b5d122: "Could not open
track" for nine consoles, "hash generation failed". This build: the
data track opens and RetroAchievements generates the hash, stopping
only at the sign-in.
- swanstation-console: the core loads, negotiates Vulkan with
RetroArch's context and boots the owner's BIOS dump; 574 frames in
10 s while the disc boots.
- melondsds-microphone (0134): the ps5 driver opens the first user's
AudioIn port at 48 kHz, captures while melonDS DS has its microphone
on (866,816 samples), reports the mute state and closes cleanly; the
game never read its microphone, so nothing was delivered (reads are
host-tested). melonDS DS at 60 frames a second.
- forwarder-launch (#34): the relaunch test hands the title a
forwarder's --rom, --core and --exit-after-game; the game runs in
RetroArch through game mode, the title closes when it quits, and the
next home launch opens the picker.
- pr-series-frontends: the picker and ES-DE of this build start; ES-DE's
library counts 35 cores with no failures. Once, the launch right after
a self-closing forwarder run ended before main(); not reproduced.
Game file names, the disc's serial and hash, and the console user id are
replaced in the records; pictures stay on this machine. The pad scripts
the commands name are in tests/pad/. Test files, the test CHD and its
memory card, framegen.cfg and melonDS DS.opt (both written by the runs)
were removed afterwards; retroarch.cfg is byte for byte as before.
Gates on this tree: format, unit (204), build (staged cores reused),
integration (204), evidence (137 captures, 0 failed) PASS.
.gitignore keeps every Markdown file but README.md and the release notes
on this machine, and c1ed6e5's cherry-pick added
tooling/swanstation/README.md past that rule. It is untracked again (the
file stays on disk), and README.md's build section now says how to
include SwanStation, as it does for melonDS DS. AngelXex's credit stays
where it was: native.patch's header, tooling/swanstation/notices.py and
the notices table.
Evidence: git ls-files '*.md' lists README.md and releases/ only;
format PASS; tests/test_notices.py PASS; no README line removed.
From mihawk-99/PS5_RetroArch#34 (Martin Pham). A forwarder, a small app
with a home screen tile of its own, starts the title with --rom <file>
[--core <core>] [--exit-after-game] (each also as --x=<value>). On a
launch that names no mode, src/frontend_mode_ps5.cpp takes it as game
mode takes a frontend's request: the content (absolute, or inside
/app0/content/ and never above it) and the core are checked by
ps5_game_check, RetroArch starts with -L, and afterwards the title
opens as from the home screen, or closes with --exit-after-game.
Without --core the playlists choose, else the one core whose info file
lists the extension (Genesis Plus GX for the Mega Drive extensions both
Sega cores list) or, for a folder, the one core that opens folders. A
refusal goes to trace.txt and the launch goes on as from the home
screen; a handover's --ps5-mode wins over --rom.
Changed from the pull request: its README examples and test fixtures no
longer name games and use the ES-DE folder names content/ has now; one
header comment is moved to its declaration. Its line-ending rewrites of
the PowerShell scripts are dropped (they are gone, 8f14452), and its
GitHub workflow is parked in parked/github-build-workflow, with what
would finish it at the top of build.yml: it cannot run here, and it
does not fetch PS5_MihawkGen, which the frontend compiles now.
To prove it on a console, whose control payload launches without
arguments: the relaunch test takes its restarts' arguments from the
lines after the arm line, and tools/run-title.sh --relaunch-arg writes
them.
Evidence: tests/frontend_mode_ps5_test.cpp (parsing, paths, cores by
playlist, extension and folder, refusals, after the game);
tests/relaunch_ps5_test.cpp (arguments carried; a control character or
a ninth argument arms nothing). Gates: format, unit (204), build
(staged cores reused), integration, evidence PASS. Console: with this
series' deploy.
From mihawk-99/PS5_RetroArch#53 and #52 (Jorge Feitosa), renumbered
(0131 to 0133 were taken), taken together: #52 fixes the read loop #53
turns on.
- src/microphone_ps5.cpp (0134): a native AudioIn microphone driver,
the default one, for the first controller's user. It captures only
while a core has its microphone on, sends silence under the DualSense
mute, and keeps PCM only in a bounded queue in memory. RetroArch is
configured with the microphone; the title links a host stub for
libSceAudioIn's imports.
- 0135: a microphone read that gets no frames returns silence instead of
looping forever (a paused core, an empty nonblocking device, a failed
capture).
- tools/build-melondsds.sh: melonDS DS at pinned revisions (git objects
by SHA), software renderer and a native-memory JIT, staged only with
PS5_WITH_MELONDSDS=1; DeSmuME stays the DS default. The DualSense
touchpad drives its stylus as it drives DeSmuME's.
Changed from the pull requests: the README and the 0134 entry describe
what happens, not an opt-in the code never had. RetroArch's microphone is
on by default, as upstream; melonDS DS listens only while its microphone
button is held (its default), and Settings > Audio > Microphone turns it
off for every core. tests/test_microphone_read.py applies the port's
edits to vendor/retroarch itself rather than reading build/ra-conf, which
a tree configured before 0135 would fail. The SwanStation and melonDS DS
title changes are merged into one licence-file rule.
Evidence: tests/test_microphone_ps5.py (the driver against a simulated
AudioIn: lifecycle, mute, input failure); tests/test_microphone_read.py
(it times out with upstream's loop); tests/test_touchpad_stylus.py
(melonDS DS's relative stylus); tools/build-melondsds.sh, core ABI PASS
(25 exports, kernel_web, sha256 b03d4c93...). A title built with both
optional cores: 61 notice parts, check-notices covers every shipped
executable, check-manifest PASS, 78 melonDS DS options in the catalog.
Gates (default build): format, unit (204), build (staged cores reused),
integration, evidence PASS. Console: with this series' deploy.
From mihawk-99/PS5_RetroArch#54 (Jorge Feitosa). tools/build-swanstation.sh
builds libretro/swanstation at a pinned revision (archive SHA-256 checked)
as a native shared core: Vulkan, the x64 recompiler on title executable
memory (ps5_exec_allocate) with LUT fastmem instead of the mmap path.
The platform changes are adapted from AngelXex's SwanStationPS5 1.3.4,
credited in the patch, the README and the notices. A title build stages it
only with PS5_WITH_SWANSTATION=1, with its licence file beside the info
file; Beetle PSX HW stays the PlayStation default, and the WebUI catalog
lists its options when it is staged.
Changed from the pull request: tooling/swanstation/README.md no longer
names the game it was tested with.
Evidence: tools/build-swanstation.sh, core ABI PASS (25 exports,
kernel_web, sha256 dfb1521f...); tests/test_optional_core_metadata.py
(catalog with and without the core). A title built with
PS5_WITH_SWANSTATION=1: 60 notice parts, check-notices covers every
shipped executable, check-manifest PASS, 113 SwanStation options in the
catalog. Gates (default build): format, unit (201), build (staged cores
reused), integration, evidence PASS. Console: with this series' deploy.
From mihawk-99/PS5_RetroArch#50 (Jorge Feitosa), renumbered: 0131 and
0132 were taken. chdman lists CD FLAC among a CD's codecs by default,
and libchdr refuses a CHD that declares a codec it was built without, so
the frontend could not open such a disc to hash it for RetroAchievements
even where the core read it with its own copy of libchdr. RetroArch is
now configured with builtin FLAC; build-title.sh leaves out the one FLAC
define whose value is a string, which tools/build.sh cannot take as a
name and the title's sources do not use.
The CHD reader's FLAC codec decodes with dr_flac (deps/dr, public domain
or MIT-0). libFLAC is compiled into libretroarch.a, but nothing calls
it: the linked title holds no FLAC__ symbol and 51 drflac ones, so the
notices are unchanged.
Evidence:
- tests/test_chd_flac.py: the configure flags enable builtin FLAC; the
vendored reader, compiled as Makefile.common compiles it with builtin
FLAC, opens tests/chd/tone-cdfl.chd (synthetic: a second of a 440 Hz
tone and silence, chdman 0.289 -c cdfl, 3,783 bytes) and decodes its
75 frames to the generated PCM byte for byte; without FLAC it fails.
- Host, with a CHD chdman made from one of the owner's PlayStation
backups (cdlz, cdzl, cdfl; 2,338 of 31,002 hunks CD FLAC): the reader
from build/ra-conf answers "unsupported format" without FLAC, and with
it opens the disc and reads all 31,002 hunks with their CRCs.
- Console, before this change (build 22b5d122, offline, a test user):
Beetle PSX HW loads that CHD, and RetroArch's hash answers "Could not
open track" for each of nine candidate consoles, then "hash generation
failed". After: with this series' deploy.
Gates: format, unit (199), build (staged cores reused), integration,
evidence PASS.
The two parts of mihawk-99/PS5_RetroArch#62 (Marco Tundo) that 0131 did
not already cover. Frame generation's plan writes the presents per game
frame into video_shader_subframes, which retroarch.cfg keeps. Once frame
generation was turned off, in its menus or in ES-DE's, the 2 it left
there held every game frame for two refreshes: half speed on a 60 Hz
output. retroarch.c now sets one present per game frame whenever frame
generation is off, and logs it when it replaces a larger number.
Turning frame generation on or off in the menus reinitializes the video
output, as a Target Framerate change already did, so the plan applies
at once (src/ps5_framegen_menu.c).
A Shader Sub-Frames value set by hand no longer survives a core load
while frame generation is off; sub-frame shaders have not been tested
on this console.
Evidence: tests/test_framegen_plan.py runs the plan as the script
inserts it. Two new cases start from a 2 left in retroarch.cfg with
frame generation off, on a 60 Hz output and on a VRR one; both fail
with 0132's edit left out. Gates: format PASS; unit PASS (197 tests);
build PASS; integration PASS; evidence PASS (131 captures). The build
gate ran with PS5_REUSE_STAGED_CORES naming the 33 shipped cores: core
stamps hash absolute paths, so in a worktree every core looks changed,
and the staged cores are main's. Console run: with this series' deploy.
Co-authored-by: Marco Tundo <marco@bymar.co>
The eight tools/*.ps1 scripts came in with 4df5c68 from
ps5-native-app-boilerplate and nothing in the build, the gates or the
docs runs them: the project builds and deploys from Linux only.
tools/prepare-assets.sh only wrapped prepare-assets.ps1, so it goes
too. .gitattributes loses the CRLF rule for *.ps1 and tools/lint.sh
no longer lists .ps1 among the files whose headers it checks.
Evidence: git grep finds no reference to any removed script;
tools/verify.sh format PASS; tools/verify.sh unit PASS (283 s,
klog/unit-gate.log).
The owner's base PS5 on a 1440p display with 120 Hz and VRR (48-120 Hz):
the driver lists only the 59.94 Hz mode (its idle vblank check read 20.873
ms, VRR's floor, and restored the default mode), and RetroArch's timed
flips came 8.32-8.35 ms apart, 119.877 and 119.898 a second. Frame
generation therefore planned at 119.88 Hz (Target Framerate 120, 1.99
pictures per game frame) and, on one of the owner's SNES game backups,
presented 119.9 pictures a second with 59.9 generated and the core at 59.9
calls a second (1.65 ms a picture). A self-made NES image looping on a
black screen gave the same timing and plan; its pictures never change, so
nothing was generated.
Evidence: evidence/framegen-flip-rate-vrr, distilled from the console's
retroarch.log (build 22b5d122); python3 tools/evidence.py compare evidence:
131 replayed, 0 failed.
ES-DE's library adapter writes custom_systems/es_systems.xml and the game
lists under /app0/es-de/ES-DE before ES-DE runs, but ES-DE makes that folder
itself only once it starts. On a fresh install's first start every mkdir
below it failed, all 29 writes failed (28 game lists and the systems file)
and ES-DE showed no games; from the second start on it worked. The earlier
ES-DE evidence never met this: those consoles had ES-DE's folder already.
The adapter makes the folder first.
Evidence: tests/esde_library_test.cpp's first start now begins without the
folder and checks 0 failures and a written systems file (it fails without
this change): python3 -m unittest tests.test_esde_library OK. On the
console, evidence/esde-folder-names: the first start before the change,
'28 systems, 255 games ... 29 failures', 'systems written -1'; the second,
0 failures; the first start after it (ES-DE's folder moved aside), '34
systems, 317 games ... 0 failures', 'systems written 34', every system's
path /app0/content/<name>; a home launch of a fresh install started the
picker (picker 1, es-de 1, remembered ask -> picker). python3
tools/evidence.py compare evidence: 131 replayed, 0 failed. tools/verify.sh:
all five gates PASS.
The content scan counted a system's .zip and .7z as games only when none of
its cores took a .cue or .chd, so a core that also runs a disc system hid a
cartridge system's archives: Genesis Plus GX (Sega CD) for the Mega Drive,
Master System, SG-1000 and Game Gear, PicoDrive (Sega CD) for the 32X,
Beetle PCE (the PC Engine CD) for the PC Engine and SuperGrafx. Their
archived games were in neither ES-DE nor the WebUI's library. Whether a
system's archives are games now goes by the system: the disc systems (psx,
ps2, psp, saturn, dreamcast, segacd, pcfx, 3do, neogeocd, gc, wii) keep
theirs out, as before, since RetroArch extracts only the file it loads and a
.cue's tracks would be left behind.
Evidence: tests/ps5_library_test.c, an archived game in an ES-DE megadrive
folder whose core also takes cue and chd (counted; it was not before this
change: the test failed on the game count) beside an archived PlayStation
disc (still left out): python3 -m unittest tests.test_ps5_library OK. On
the console (evidence/esde-folder-names): ES-DE's library went from 28
systems and 255 games to 34 and 317 with the same 325 game backups, the six
systems above among the new ones. tools/verify.sh: all five gates PASS.
Testers kept reporting folders EmulationStation did not detect. The title
made a folder in content/ for each system named as RetroArch's databases
are ("Sony - PlayStation 2"); it now names each as ES-DE names the system
("ps2", "snes", "genesis", the 50 platform ids, all of them ES-DE system
names checked against ES-DE 3.5.0's own systems file), so content/ is
ES-DE's own layout, <ROM folder>/<system>.
- ps5_library_make_system_folders makes content/ first when it is missing:
a copy of the title without its empty folders had none at all, because
the frontend made content/ only later and the step gave up.
- ps5_library_rename_system_folders, at each start before any frontend:
a folder an earlier build made under RetroArch's name is renamed to the
ES-DE one, or merged into it when none of its entries' names is taken
there (else left as it is, still its system's). The paths in RetroArch's
playlists (and their scan folders) and the per-folder overrides, core
options, shader presets and remaps under config/ follow; a core whose
folder bears the content folder's own name (MAME, ScummVM) keeps its
files, which are that core's. Folders the user named (PS1) stay as they
are and are still recognised.
- The WebUI's Add Game puts a game in the system's ES-DE folder.
- README: content/<system>/ with ES-DE's names (content/Saturn/ was out of
date: only system/Saturn is made).
Evidence: tests/ps5_library_test.c (system folders by ES-DE name, a missing
content folder made, a rename, a merge, a clash left in place, a folder
whose name only begins the same, playlists, per-folder settings, a core
folder named as the content folder): python3 -m unittest
tests.test_ps5_library OK. On the console (evidence/esde-folder-names,
with the commits after this one): a home launch renamed three seeded
folders (a rename and two merges), the playlist and options file followed,
and every ES-DE system path is /app0/content/<name>. tools/verify.sh: all
five gates PASS.
build-title.sh copied Mozilla's CA bundle from .deps/es-de, which only
tools/build-esde.sh clones, and it does so after this step and only for a
build with frontends. With .deps emptied, neither the development build
(the build gate) nor the all-target build could pass: both stopped at the
missing bundle. The bundle is now fetched by itself from ES-DE at the commit
build-esde.sh pins (one pin, read from that script), checked against its
sha256 (f66dff1b..., the same file ES-DE's checkout holds), and kept in
.deps/downloads.
Evidence: the pinned URL returns sha256 f66dff1bdf8f9606...; the
all-target build (PS5_FRONTENDS='es-de picker') then ran from a fresh
.deps through ES-DE, the picker and the notices to a verified package
(dist/PS5_RetroArch.zip, python3 tools/check-notices.py dist/PPSA99169:
the licences cover every shipped executable).
A batch stream that ends inside a file left that file's .upload- temporary
until the request was freed, after the 400 answer was sent: a client reading
the folder the moment the answer arrived could still find it. The gate's
test_batch_refuses_bad_names_and_bad_streams did, once, with the gates run on
four CPUs (taskset -c 0-3). BatchUpload::abandon() now closes and removes the
cut file before the answer; the destructor calls the same.
Evidence: JOBS=4 taskset -c 0-3 nice -n 10 bash tools/verify.sh: the unit
gate failed on that test before (a .upload- file left in content/batch-c)
and passed after (197 tests OK, 1 skipped); python3 -m unittest
tests.test_webui: 17 tests OK.
The owner's base PS5, on an output that does not offer 119.88 Hz, ran the
hotfix eboot.bin built from 5c4256b (identity 1f81446a). RetroArch timed
nine flips 16.67-16.69 ms apart, 59.943 flips a second in the 59.940 Hz
mode, and frame generation planned at 59.94 Hz (a 120 target becomes 60)
instead of the 119.88 it assumed before 0131.
Evidence: evidence/framegen-flip-rate-60hz, distilled from the console's
trace.txt and retroarch.log; python3 tools/evidence.py compare evidence:
129 replayed, 0 failed.
The 59.94 Hz mode takes 120 flips a second on a VRR display and 60 on an
output without 120 Hz. 0121 assumed 119.88 Hz for every 59.94 Hz mode, so
on a 60 Hz output each picture was presented twice and games ran at half
speed with broken audio (a tester's v2.0.0-beta-2 logs: "display 119.880
Hz ... each shown 2x", 58.8 presents/s, the core at 29.6 calls/s).
Patch 0131. khr_display_ctx.c times flips once per process, on the first
swapchain, where the mode is below 100 Hz: the median of nine acquire
intervals once every image is queued, 1 s timeouts, and on a failure a
warning and the mode is believed. vulkan_common.c keeps the rate. The plan
in retroarch.c takes 119.88 only above 100 Hz, and a Target Framerate
above the display's rate becomes the display's (120 on a 60 Hz output is
60, and a 30 fps game still gets a frame made between its frames).
Evidence: evidence/framegen-flip-rate (python3 tools/evidence.py compare
evidence: 128 replayed, 0 failed). tests/test_framegen_plan.py composes
the plan from EDITS as the script applies it and runs nine cases; with
0131's edits left out it fails on the 60 Hz case. The eboot.bin built from
this tree (identity 1f81446a..., sha256 d3dc2a31...) ran on the tester's
60 Hz output, and the owner reports it worked perfectly; no log of that run
was received. Not run yet: the owner's base PS5 with its VRR display (the
console did not answer).
format PASS; evidence PASS; unit 192 of 197 OK. The 5 errors are tests
that read .deps/{ppsspp,azahar,desmume}-src, deleted with .deps on
2026-10-15 along with every core source the build gate's catalog step
reads.
The generator (framegen/src, framegen/include) and its trained network now
come from PS5_MihawkGen (github.com/mihawk-99/PS5_MihawkGen, public),
pinned at c57232a; a release build refuses any other revision. Its sources
and the d5 network are byte-identical to what the title built before (the
network sha256 31ecc7d6..., the four generator sources and the compiled
shaders; framegen.h and nn_final.comp differ only in a comment). The
RetroArch driver shim (vulkan_framegen.cpp/.h, from PS5_FrameGen 9490b0d)
moves to tooling/framegen, so all of the frame generation code in
eboot.bin is published. tooling/notices: a mihawkgen component with its
source address.
format PASS; unit (196 tests) OK.
Every game title in tracked text (test fixtures, comments, patch notes,
console evidence, the v1.0.0-beta.1 note's alt text) is now a made-up
stand-in that keeps the original's shape: its punctuation (apostrophes,
"&", " - "), region tags and, where a test depends on it, its sort order
and how a title relates to its alternate name. Evidence folders and the
Flycast pad script named after games are renamed. Only the committed
handoff/PPSA99002/eboot.bin still holds a title, inside the binary.
releases/v2.0.0-beta-2.md: PS5SX2, frame generation, RetroAchievements,
NetPlay, the WebUI in 16 languages, the library and frontend changes,
the systems tested on a Base PS5, upgrading from v1.0.0-beta.1, source
and credits. It opens with the anti-piracy banner and points, names no
game and shows no game picture: three WebUI screenshots in
assets/releases/v2.0.0-beta-2/.
Gates: format, unit (196 tests), build, evidence (127 replayed, 0
failed), check-manifest and check-notices: PASS.
- tests/input_ps5_test.cpp: ps5_fg_record_request stub (the pad script's
RECORD calls PS5_FrameGen's recorder, not linked in the host test).
- tests/test_webui_qr.py: stubs for the frame generation entries 0120 binds
beside the WebUI's (ps5_fg_menu_is_label/_is_folder, action_ok_ps5_fg,
action_ok_push_default).
- src/main.cpp, src/ps5_framegen_menu.c: tools/lint-format.sh --write
(comments re-wrapped to the line limit; no code changed).
- The frame generation code already names patches 0122-0128, which
patches/series does not list; 47879d1's TLS and password patches took
0122/0123 and are now 0129 (TLS verification) and 0130 (password).
lint-format PASS; tests test_input_ps5, test_webui_qr, test_frontend,
test_webui_secrets: OK.
- 0122: HTTPS uses MBEDTLS_SSL_VERIFY_REQUIRED (upstream verified optionally
and ignored the result), so nobody on the network path can stand in for
RetroAchievements. Its chain (GTS Root R4, cross-signed by GlobalSign Root
CA) verifies against the bundled 2017 roots.
- 0123 and src/ps5_cheevos.c: a sign-in that returns a token removes
cheevos_password from /app0/config/webui.cfg (with the empty token the
WebUI puts beside a new password) and from retroarch.cfg, which keeps its
token. The WebUI clears the old token when a new password or account is
saved, so the new account signs in with its password.
- WebUI: game, file, system and source names carry data-no-translate, and a
value inside a translated sentence is translated only on an exact match:
"Shadow of the Colossus" no longer reads as "{0} of {1}" ("Shadow sur the
Colossus" in French).
- LRPS2 leftovers: the 34 orphan option texts in each catalog, the stale
wording in ps5sx2_libretro.info and build-ps5sx2.sh. Patch count 255 -> 257.
Console (192.168.50.237, klog/retroarch-191705.log): DNS, "TLS connected",
mihawk99 logged in, game 2992 identified with verification required; after
the run webui.cfg holds no cheevos_password and retroarch.cfg an empty one
with its token kept.
Tests: test_webui_secrets (token cleared; helper removes the password, keeps
the token in retroarch.cfg), test_webui_language (a file named Shadow of the
Colossus stays as it is in French), test_frontend: OK. lint-shell PASS;
lint-format fails only in src/main.cpp and src/ps5_framegen_menu.c, not
touched here.
apply-port-patches 0122-0128: private env commands 0x25001 (the core blocks inside retro_run) and 0x25002 (emulated vsyncs a frame
covers, period in microseconds) size the generator's group from what the core says, never from a measured gap; the floor is opt-in
(PS5_FRAMEGEN_FLOOR=1); core run/cpu times in the log. The menu and its config follow. build-retroarch.sh compiles fg_splat.cpp.
main.cpp reads framegen-record-width.txt for the recorder. input_ps5.cpp: reflowed comments and calls.
The PS2 core is now PS5SX2 (Swordpdf's PCSX2 for the PS5) from the fork
../Libretro_PS5SX2, branch libretro, pinned at 308e8dc. build-ps5sx2.sh
fetches and pins its dependencies (rapidyaml 0cd25a1, glslang 275822a built
for the console, shaderc v2026.2 headers), builds the core with the fork's
ps5/libretro/Makefile.libretro, ABI-checks it and stages PCSX2's game
database in system/ps5sx2/resources. PS5SX2_DEV=1 builds the fork as it
stands. BIOS and memory cards stay in system/pcsx2/ (shared).
LRPS2 is gone: build-lrps2.sh and tooling/lrps2, its notices component,
WebUI catalog and settings-guide entries, and the title's core list.
build-title.sh empties dist/{cores,info} before staging, so a core taken
out of the list does not linger. The balanced 4K default (0102) now sets
ps5sx2_upscale to 4x; the core offers up to 8x (6x is 4K).
Also: patch 0118's tests (the block count, the XMB icon names test).
Console evidence (192.168.50.237, RADV; tools/run-title.sh --game=ps5sx2):
- Shadow of the Colossus (USA) boots and plays (klog/game-122302.png).
- tests/pad/ps5sx2-states-4k.txt: frames of 3072x2688 at 6x; a 46,562,360
byte state saved and loaded from the Quick Menu (klog/retroarch-124621.log).
- RetroAchievements: SET_SUPPORT_ACHIEVEMENTS, account logged in, disc
hashed, the core's memory map published (klog/retroarch-125231.log).
- pcsx2_libretro.{so,info} deleted from the console's cores/ and info/.
Gates: format PASS; unit and integration fail only in
test_input_ps5.test_joypad_bindings and test_webui_qr (the uncommitted
frame-generation and input work in src/, not part of this commit); notices
PASS (33 cores); manifest PASS; evidence 127 replayed, 0 failed. The title
was built with the other cores reused (PS5_REUSE_STAGED_CORES).
Options (on/off, output rate, method, cadence, working size, status on screen) are shared
through /app0/config/framegen.cfg, which the title appends when it starts RetroArch.
The Vulkan driver's shader subframes present the generated frame at j/N
between the previous and the current content frame (video_framegen, with
video_shader_subframes 2 to 4). The hooks are anchor edits; the generator and
its shim are built from ../PS5_FrameGen and the trained network is staged at
assets/framegen/network.fgw.
After the click on Settings it waited for the Settings heading, which is static and
translated while still hidden, then read the page title before navigation had run
(seen once in a full run: "Vue d'ensemble" for "Réglages"). It now waits for the title.
Evidence: three consecutive passes of tests.test_webui_language.LanguageBrowser.
Gates: bash tools/verify.sh -> verify: PASS (format unit build integration evidence)
tools/xmb-system-icons.py makes, at build time, one tab icon per RetroArch database the
title knows (src/ps5_library.c's platform table: 48), from the WebUI's system artwork
(webui/assets/systems, Linear ES-DE, MIT): each picture cropped to its hardware (the logo
and year under it removed; XMB prints the name), centred, 256 x 256 PNG, named
<database>.png in assets/xmb/monochrome/png, where XMB looks for a playlist's icon.
Game rows keep XMB's own content icon (no <database>-content.png). A playlist with
another name keeps the default icon.
Notices: the generated icons are their own part (MIT); retroarch-assets now names its
files exactly, so neither claims the other's (fnmatch's * crosses "/").
Evidence: python3 -m unittest tests.test_xmb_assets (one icon per database, 256 x 256
RGBA, the caption cropped off, no -content icons, each icon owned by one licence); the
build printed "[xmb] 48 system icons" and tools/check-notices.py passed. Contact sheet
checked by eye. Console screenshot not taken: the title refused to start on this
console ("startup refused: elevation or installation could not be verified").
Gates: bash tools/verify.sh -> verify: PASS (format unit build integration evidence)
Package the qualified title-specific resident service and start it through
local elfldr before workers when no service is ready. Wait for preflight,
verify SDK elevation and UID 0, and reuse the service on subsequent launches.
Keep installed paths and offline fallback; no Payload Manager boot hook.
Console evidence: evidence/lapy-autostart/capture.json and expectation.json.
Reproduce with evidence/lapy-autostart/README.txt, then
python3 tools/evidence.py compare evidence/.
Verified on the base PS5: cold service startup, two authenticated achievement
sessions, reuse of one service across three requests, and NetPlay relay join
with 3837/3361 input packets and 52 matching state checks. Normal title exits;
root counts 73/72 before each request. Resident service remains running.
Installed binary hashes and unchanged Payload Manager configuration verified.
Validation: 194 unit and 194 integration tests; format; all-target build with
PS5_FRONTENDS='es-de picker' and the existing pinned build/picker-template;
manifest/notices/source bundle checks; 127 evidence captures replayed.
No secrets, saves or raw console captures included. No push.
Qualify the console build with authenticated achievement-set loading, a fresh
badge download, FCEUmm LAN hosting and relay joining. The relay session exchanges
scripted controls and passes 55 direct emulation-state CRC comparisons. Normal
quit returns the elevation service's root counts to baseline.
Expose DNS/TLS failures and add a debug-only trace at the upstream state CRC
comparison. This adds one port hook, so the intentional patch-count assertion
changes from 253 to 254. Keep the requested Online Updater removal and repeatable
RetroArch runner arguments. The verified build uses the standard SDK socket API.
Online startup still requires the qualified external Lapy service; no autoload
is installed. Document the force-close teardown timeout and untested lobby/NAT
scope. Credentials, settings, saves and raw console captures are not committed.
Validation: PS5_FRONTENDS='es-de picker' with the pinned picker template;
tools/verify.sh format unit build integration evidence passed (193 unit,
193 integration). After moving only the debug trace to the received-CRC path,
24 focused frontend/patch tests and format/build passed again, followed by the
final console runs. Final format/evidence, manifest and notices checks passed.
Evidence: evidence/online-features/capture.json and expectation.json.
Replay: python3 tools/evidence.py compare evidence/ (126 captures, zero failures).
Console reproduction: evidence/online-features/README.txt and console_netplay.py.
Mihawk