On macOS a newer FFmpeg's headers (Homebrew's) came first through another
library's include directory, and AVIDump.cpp failed to build against them.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The ffmpeg submodule held prebuilt libraries for every platform next to the
sources (about 1.5 GB checked out), and nothing uses those any more. Only the
needed sources are copied in, from the submodule's commit, along with every
header and template they include. The path is new, so checking out commits on
either side of this works.
It's also cut down to what PPSSPP uses: the H.264, AAC, MP3, ATRAC3 and ATRAC3+
decoders, the H.264 parser, the demuxers as before, and the FFV1 encoder, AVI
muxer and file protocol for AVI dumping. config.h takes its CONFIG_ values from
FFmpeg 3.0.2's own configure run with just those enabled. The raw H.264 demuxer
stays even though nothing opens raw H.264: probing PSMF data uses it to identify
the video stream's codec, which otherwise comes out as none.
On top of that:
- The files nothing links (configure would build them, but they're never pulled
out of the static library, going by the GCC link map) are left out: mostly
libavutil's crypto and hashing and the alternative DCT/IDCT/FFT versions.
- libswresample isn't built, now that nothing uses it.
- The H.264 decoder only does 8-bit video, which is all the PSP has. The SPS
parser rejects any other bit depth, and the DSP files no longer build their 9-
to 16-bit variants, which the linker kept because the init functions pick one
by bit depth at run time.
That's 160 files to compile instead of 277, and 391 of FFmpeg's vendored, about
5.8 MB. The README in ext/ffmpeg has what was done, the PPSSPP patches in the
copy, how to add a component, and the tradeoffs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The lossy MPEG-4 encoder (with the MPEG-video encoding code it needs) is most
of what FFmpeg's encoders weigh, and lossy footage is better recorded from the
output with a screen recorder. Dumps are now always lossless FFV1, which also
makes for much larger files than the old default.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
FFmpegAudioDecoder only used swresample to turn the decoders' output into
interleaved stereo s16, at the same sample rate. A small conversion does that
now, with swresample's rounding: floats go through lrintf and are clamped, and
mono goes to both channels at -3 dB, the coefficient swresample's mixing matrix
uses. Its output is sample-identical to swresample's for the MP3, AAC and
ATRAC3+ decoders, mono and stereo.
With "Use FFMPEG for all compressed audio", ATRAC3 got closed on its first
frame: its decoder sets no channel layout, and swresample refused to set up
without one. The conversion goes by the channel count, so that no longer fails.
The system-FFmpeg build no longer looks for libswresample.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every build compiles FFmpeg now, so the per-platform library paths are gone, and
the Alpine job no longer builds it with the old script.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The ndk-build builds (android/jni and libretro/jni) get a static library module
with the files in ext/ffmpeg-build/sources.mk, in place of the prebuilt
libraries. The libretro Makefile compiles them itself, with FFmpeg's include
directories first so its config.h wins over the others on the include path, and
puts the objects in obj-ffmpeg/ rather than next to the sources, where a
configure build may have left its own.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ext/ffmpeg-build/ffmpeg.vcxproj (and UWP/ffmpeg_UWP for UWP) builds the same
sources as CMake, copied from sources.mk by update_vcxproj.py. It's optimized in
Debug too, and without /Zc:inline, since MSVC otherwise keeps references to
disabled components that FFmpeg expects the compiler to drop.
The projects that linked the prebuilt libraries reference it instead, and take
the headers from the source tree.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ext/ffmpeg-build now has a committed config.h in place of the configure checks.
Assembly and arch-specific code are off, as in the --disable-asm builds of the
old prebuilt libraries, so only the platform has to be told apart, with a few
preprocessor checks. The sources are listed once in sources.mk, for every build
system to use, and build one static library. ATRAC3 and ATRAC3+ are enabled to
match the prebuilt libraries, and HAVE_AV_CONFIG_H is now defined, as FFmpeg's
own build does.
CMake builds this unless USE_SYSTEM_FFMPEG or FFMPEG_DIR is set, and no longer
looks for the prebuilt libraries in ffmpeg/. MinGW now builds it too, instead
of falling back to the system's.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The screenshot readback ran in ScreenManager's post-render callback, and
on Vulkan it ends the backbuffer render step. renderImDebugger() draws
after that, so taking a screenshot with the debugger open made its first
SetViewport() write through a null render step (an assert in debug
builds, a Play crash report in release). Do the readback at the end of
the frame in NativeFrame instead, which also puts the debugger in the
screenshot, and remove the now unused callback.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
SetFileLogPath() held logFileLock_ while creating the folder and opening
the file, and both log (CreateDir at INFO, OpenCFile for content URIs and
on errors). Logging to the file takes logFileLock_, so the thread
deadlocked on itself, and every thread that logged after it blocked too.
A Play ANR shows the loading thread stuck in CPU_Init's
SetFileLogPath, and the main thread waiting on the same lock to log a
key press.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
sigaltstack() is per thread. InstallExceptionHandler() runs in CPU_Init()
on the loading thread, which exits after boot, while the faults come from
the CPU thread and UninstallExceptionHandler() runs on the emu thread. So
our stack was never used, and the uninstall "restored" the loading
thread's previous alt stack onto the emu thread. On Android that's the
one bionic gives each thread and frees when the thread exits, so every
later fault on the emu thread ran its handler on freed memory. A Play
crash report shows scudo finding a corrupted chunk header when freeing our
stack in UninstallExceptionHandler().
The alt stack came with the Dolphin code, which needs it to handle its own
stack overflows. Ours only handles guest memory faults, so let the handler
run on the faulting thread's stack, or the alt stack that thread already
has. This also leaves a libretro frontend's alt stack alone.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A size that makes addr + size wrap around 32 bits had Split() build a slab
that ends before it starts. Lookups then skipped it, and a later
Merge() ran FillHeads() over it with a huge count, writing far outside
heads_. Sizes come from games through HLE calls, and the large ones are
recorded even without detailed mem info. Seen in a Play crash report, on
the FlushMemInfo thread.
FillHeads() also took the address one past the end of heads_ for a slab
in the last slice, which Debug builds assert on.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The NDK's zlib is too old for uncompress2(), which the PSAR unpacker now uses,
so link ext/zlib like the CMake build already does on Android. It's linked
whole, since the prebuilt FFmpeg libraries come after it on the link line and
libavformat uses it too.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
3.60's index.dat carries tag 0x0B2B10F0, which sceResmgr_9DC14891 had no key
for, so the shell couldn't read its item index and stopped on "A serious
error has occurred". The key follows the tag in 3.60's mesg_led_02g.prx, laid
out like the other versions' (3.52's mesg_led.prx has the 3.5x key the same
way). 3.60 now boots to the XMB.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The 1.00 updater's DATA.PSAR is a version 1 archive, the only one there is,
and its blocks carry tag 0x06000000, which the PRX decrypter had no key for.
1.50's mesg_led.prx keeps its keys raw in a table after their tags, the
0x0E000000 one we already use among them, and the 0x06000000 one there
decrypts all of 1.00 (with KIRK code 0x49, following the other tags). Its
first record has no version text either, so the version now comes from the
firmware's own version.txt when that's missing.
Fix a zlib usage bug.
Headless --unpack-updater also exited 0 when every file failed, so now it
fails if any did.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Like the other tabs, except that the screen keeps what the page edits that
isn't a config setting (the compatibility reports flag, which is written
back when the screen closes, and the Windows memstick checkboxes' state) in
a SystemPageState it passes in. The page's handlers move with it, as
functions taking the page host or that state.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Nothing calls CallbackMemstickFolder or sets pendingMemstickFolder_, which
are left over from an older way of changing the folder on Android.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Family Guy and The Dark Knight are coded as separate field pictures,
bottom field first. ffmpeg hands each pair back as one frame, and we
described it with the first field's SEI record only.
The UMD Video player (video_main_plugin.prx) shows such a picture a field
per vsync, and takes the field order from comparing the output times of
the picture's two records. With the second left empty it always chose
top first, so every pair ran backwards in time and the picture jittered
between frames. Now the pair carries both fields' timing, the top field's
record first.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The XMB (vshmain.prx) was kept out of Recents because it had no icon. Now
the game list shows it as "XMB 6.61" (the installed firmware's version, from
its version.txt), with an abstract XMB icon in white on blue, and the window
title says the same instead of "Unidentified PSP title".
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Only moves code. The tabs other than System were built by static functions
in GameSettingsScreen.cpp, in place of the old methods, to keep the previous
change's diff readable. Now each tab has its own file, with just the
includes it needs, and MainSettingsPages.h declares their builders.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Each tab is now a SettingsPage: a list of items (headers, settings, buttons,
hints, links to other pages) that a function builds and views are created
from. Settings are added with page->Setting(&g_Config.x), which takes
the label, the kind of view, the range or choices, units, labels for special
values and the default from the config. The pages keep what depends on the
UI (which settings to show and when, enable rules, and behavior that needs the
screen), and reach the screen through the page's SettingsPageHost.
A page can link to other pages, which open on top of it in the same tab,
under a bar showing where you are, like Tools > Developer tools > Graphics,
where the earlier pages can be clicked to go back to them. The bar has back,
search and help buttons, and the back key also goes back a page (or out of a
search) before leaving the screen. Each page remembers its scroll position.
In portrait there's no tab strip: a page links to the tabs, and the bar shows
just the current page's title. Rotating keeps showing the same page. Moving
between pages only focuses an item when navigating with a keyboard or
controller.
The developer tools are now pages linked from the Tools tab, instead of a
screen of their own. The in-game developer menu and
--start-screen=developertools open the settings on that page. The toggles of
live state (the remote debugger and the MIPS tracer) show and change it
through functions, and pretending to be in-game for the notification tests is
a flag on the OSD. The memory map and memstick tests, only reachable from
commented-out code, are removed.
Search lists the matching settings from all tabs and the pages they link to,
grouped by where they are. Each result goes to the setting on its page,
scrolled to and highlighted (outlined, when picked with a mouse or touch), or
for a page, opens it. This replaces the search tab and the old filter that
hid the views that didn't match, and with them ShowSearchControls. Changing
tabs leaves the search.
The pages, the bar and search are in SettingsPagesDialogScreen, between the
tabbed dialog screen and the settings screen, so the other tabbed screens
keep plain tabs.
Other visible changes:
- Reset buttons that went to the wrong value now go to the real default: the
touch control auto-hide delay (was 20 seconds instead of 8), the mouse
wheel delay (1 ms, below its own range, instead of 80) and the UI size (0
instead of -1 on desktop). The alternative speeds get a reset button.
- Units are shown on more sliders (frames, MHz, seconds, Hz), and popups no
longer spell out what 0 means in the units, since the zero label does.
Units of up to three characters (%, x, ms, MHz) aren't translated.
- The Android HW scaler hides the scales the display is too small for,
instead of cutting the list short.
- Choices are translated in the setting's own section.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Settings will translate everything about them (label, choices, units) in one
category, the label's, so:
- "Achievement sound volume" moves from Achievements to Audio, where the other
volumes are.
- The choice strings of the ad hoc relay mode and chat positions are copied
from Dialog to Networking, the notification position's to System, and the
leaderboard tracker position's to Achievements.
- "seconds" is added to Controls and System, for the units of the touch
control auto-hide delay and the rewind snapshot interval.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
ConfigSetting gets a ConfigSettingInfo, set with constexpr chained setters:
a label, choices, a range with step and units, labels for 0 and negative
values (like Unlimited and Disabled), and a callback for what has to happen
when the UI changes the value. Using a setter on the wrong type of setting is
a compile error. The settings in the config tables now have these, so the
settings screen can take them from there instead of repeating them, including
what changing them needs (resize and GPU config messages, fullscreen, audio
and camera notifications and so on). The settings screen starts using them
in a later commit.
Config::FindSetting finds a setting from a pointer to its value. Defaults can
now also be looked up for blocks that reset as a whole (display layouts, touch
and gesture controls), by reading them from a freshly reset copy of the block.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
In landscape, keyboards went fullscreen with an edit box of their own,
covering the dialog and forwarding what was typed to it. Now the dialog's
field asks for no fullscreen editing, and the dialog moves up above the
keyboard.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
setMMapEnabled(true) doesn't mean "allow it", it sets the process's MMAP
policy to AUTO, overriding a device whose aaudio.mmap_policy disables
MMAP. Since #22766 we did that on every audio start, and a Xiaomi Pad 5
Pro (Android 13) has been stuck on a blank screen at launch since then.
The likely cause, not yet confirmed on the device: opening a stream on a
MMAP path the vendor turned off, from the UI thread.
Now the checkbox can only turn MMAP off, as on the Bluetooth TV.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
SetHighlighted outlines the view, following its round corners, to point it
out without giving it the focus, which would turn on keyboard navigation.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
For bringing a given view into the middle of the scroll view it's in, rather
than just into sight like focusing does.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
