wget now retries 3 times with a 60s timeout and 5s backoff. The linux
build failed once because a savannah mirror redirect (freetype)
returned HTTP 500.
Route the macOS build through ccache (CC/CXX) like linux and windows;
the workflow installs it but the script never used it, so the two-arch
build always recompiled from scratch.
- Asset names are now bare per-platform library files (not zips), with the
version suffix and no 'mpv-' infix, matching release.yml.
- Document the deterministic release download URL consumed by jlibmpv's
RemoteNativeLocator.
- Note that mpv + dependency versions live in ci/versions.env.
- Describe the new build triggers (pull_request / v* tag / manual) and the
PipeWire/Pulse/ALSA audio probe order.
- Introduce ci/versions.env as the single source of truth for the mpv tag
and every dependency pin; the three build scripts and all workflows now
source it instead of redeclaring versions inline.
- Add release.yml: on a v* tag it reuses the three per-platform build
workflows (workflow_call) and publishes the bare libmpv.so / .dll / .dylib
assets plus a combined checksums.txt as a GitHub Release.
- Rework build-*.yml: trigger on pull_request (PR check) and workflow_call
instead of push to main, and expose the artifact name/version as job
outputs so release.yml can assemble the release assets.
- Route the Linux build through ccache (CC/CXX) so the CI ccache is actually
populated and reused, fixing the 'Path Validation Error' on cache save.
* fix: link libmpv against the system libasound instead of a static build
A statically linked libasound fails at runtime with 'Unknown PCM
cards.pcm.default': the default PCM is resolved dynamically per sound card
(load_for_all_cards hook enumerating cards and loading cards/<driver>.conf),
and that machinery did not work in the vendored archive, so audio produced no
output even though the config file itself was found (the earlier
--with-configdir prefix bug is fixed).
Link libasound.so.2 dynamically from the system (libasound2-dev), exactly like
CoreAudio on macOS and WASAPI on Windows: the OS audio stack stays external,
all video dependencies remain static and self-contained.
The build script no longer builds alsa-lib (the _alsa step, ALSA_VER pin and
the autoreconf dependency it needed are gone); meson resolves the system
'alsa' dependency via pkg-config. The workflow installs libasound2-dev, drops
autoconf/automake/libtool, and verifies ldd shows libasound.so.2 instead of
asserting a baked-in config dir.
* fix: enable PipeWire/Pulse AOs so audio works on PipeWire desktops
On a desktop where PipeWire owns the sound cards, the direct ALSA/dmix
'default' path is broken: snd_pcm_open('default') resolves to a dmix slave
that cannot be opened (verified: aplay -D default fails identically, while
aplay -D pulse works). libmpv with only the ALSA AO therefore got no sound.
mpv auto-probes audio drivers in the order pipewire -> pulse -> alsa, so
enable all three (each linked against its system shared lib, consistent with
CoreAudio/WASAPI). On a PipeWire/Pulse box it now uses the working server
before ever falling back to the direct ALSA path.
The workflow installs libpipewire-0.3-dev and libpulse-dev alongside
libasound2-dev, and the artifact verification now accepts any of the three
system audio libs instead of only libasound.
---------
Co-authored-by: david <david@phoenix.ipv64.de>
Pinning fontconfig's sysconfdir to /etc made the build try to install its
own conf files into the real /etc/fonts/conf.d, which fails without root
on the runner (PermissionError). Revert to the default prefix: the
prefix-baked config paths are harmless because fontconfig falls back to a
built-in config that scans /usr/share/fonts and /usr/local/share/fonts when
no config file exists at the compiled-in path - so libass's font lookup
works on any standard box without env vars.
Also tighten the artifact verification: instead of a broad build-path grep
(which would false-positive on fontconfig's benign prefix strings), assert
the ALSA config dir is the system path and not the build prefix - the
specific regression that broke audio.
Two runtime bugs in the shipped artifact:
1. ALSA: the static libasound carried ALSA_CONFIG_DIR pointing at the
build prefix (linux_prefix/share/alsa), so snd_pcm_open("default")
failed on any machine without that path ("Unknown PCM default",
video worked, audio did not). Build alsa-lib with
--with-configdir=/usr/share/alsa so the compile-time default is the
canonical system location.
2. Size: the release .so still contained ~110 MB of DWARF debug sections
(whole ffmpeg tree); strip --strip-all the artifact in the collect
step, matching the windows build.
Also pin fontconfig's sysconfdir/localstatedir to /etc and /var so the
CONFIGDIR/FONTCONFIG_PATH/FC_CACHEDIR constants compiled into libass's
font stack resolve on any standard install (the runtime config search
falls back to the compiled-in CONFIGDIR and a built-in default config
when no FONTCONFIG_FILE/FONTCONFIG_PATH is set).
The workflow now verifies the artifact: ldd shows only system libs, no
debug sections remain, and no build-machine paths leaked into the
binary.
expat is built with CMake, which does not force -fPIC for a static
library; like alsa-lib it is linked into the shared libmpv.so and
fails with the same relocation error. Enable CMAKE_POSITION_INDEPENDENT_CODE.
(The meson-built dependencies already default to -fPIC for static libs.)
The static archives are linked into the shared libmpv.so, but neither
ffmpeg (configure static build) nor alsa-lib (autotools static build)
force -fPIC in that case. Linking then fails with a TPOFF32 relocation
error from libasound. Pass --enable-pic to ffmpeg and -fPIC to alsa-lib.
Meson on a Debian/Ubuntu host defaults libdir to the multiarch path
lib/x86_64-linux-gnu/, so the static archives and .pc files never land in
the expected <prefix>/lib that the later builds look in. Force libdir=lib
for every vendored dependency build.
Add a native x86_64 Linux build (ci/build-libmpv-linux.sh,
build-linux.yml). Every dependency is static and linked into a single
libmpv.so; only glibc/pthread/dl/m and the system zlib remain external.
The Linux-only additions are the font stack (fontconfig -> expat, where
macOS/Windows use CoreText/DirectWrite) and alsa audio output; gl is the
plain render-API support (no X11/Wayland context, since jlibmpv supplies
the GL context via JOGL).
For consistency, ship a single native per platform and no public headers
(jlibmpv loads the library via JNA and never compiles against it). Remove
the accidental header/import-library copies from the Windows build and the
phantom header copy from the macOS build.
mpv enables Objective-C for the macOS audio/video output backends via
add_languages('objc'); the cross file needs the objc binary defined or
meson setup fails.
ffmpeg auto-detects the CPU architecture from uname -m, which is
aarch64 on the macOS runner, and then compiles Arm assembly for the
x86_64 slice. Pass --arch explicitly so the cross build uses the
correct SIMD code paths.
meson names the shared library libmpv.2.dylib on macOS, not
libmpv-2.dylib as on Windows. Fix the collect step in the build script
and the lipo inputs in the workflow.
harfbuzz 12.2.0's option is utilities, not utils; enable coretext as the
system font provider. libass gets coretext (fontconfig would drag the
arm64-only brew package into the x86_64 slice) and asm disabled for
determinism. freetype disables all optional third-party providers except
the universal system zlib. libplacebo disables optional xxhash for the
same reason.
The local array declaration for ffmpeg configure args mis-parsed on the
macOS runner's bash 3.2; pass the args as a plain array. Use
videotoolbox-gl instead of videotoolbox-pl: the pl variant requires
vulkan in mpv v0.41.0's meson.build, while the gl variant only needs the
gl-cocoa feature which is enabled.
GitHub's macOS runners are all arm64 now, so the two parallel per-arch
jobs are obsolete: one macos-15 runner builds the arm64 slice natively
and the x86_64 slice via clang --target=x86_64-apple-darwin, then lipo
merges both into the universal libmpv.dylib. The build script is now
parameterized by ARCH, installs every meson dependency into its own
macos_prefix-<arch> (meson's macOS default prefix is /opt/homebrew),
uses release buildtype for the static deps, and disables harfbuzz icu
(the brew icu4c on the runner is arm64-only and must not leak into the
x86_64 slice).
pip installs are blocked by PEP 668 on the macOS runners, so install meson
via brew. Enable cocoa and gl-cocoa: on macOS the OpenGL render API
backend of libmpv requires them (system frameworks only). Build no zlib
ourselves - macOS ships it in the system like iconv. Use PKG_CONFIG_PATH
additively instead of overriding PKG_CONFIG_LIBDIR so system pkg-config
files (icu-uc, ...) stay visible.
Vendor ci/build-libmpv-macos.sh, which builds every dependency statically
for the native runner architecture and links them into a single
self-contained libmpv dylib (coreaudio output, videotoolbox hardware
decode, OpenGL render API; no vulkan/cocoa player). build-macos.yml runs
it on macos-14 (x86_64) and macos-15 (arm64) in parallel and a merge job
combines the two dylibs into one universal libmpv.dylib with lipo, matching
the file name jlibmpv expects on macOS.
The release build type emits DWARF .debug_* sections even for the MinGW
cross build, inflating libmpv-2.dll to ~169 MB. Strip them with
TARGET-strip --strip-all so only code/data remain. Also hand upload-artifact
the folder directly instead of a pre-zipped file, so the artifact is a
single-level archive rather than a zip containing a zip.
Vendored ci/build-libmpv-mingw64.sh builds every dependency statically and
links the single libmpv-2.dll against those archives with a static CRT, so
the artifact depends only on Windows system DLLs. Dependency versions are
pinned to the latest stable tag on the mpv v0.41.0 release date. Vulkan,
shaderc and SPIRV-Cross are dropped (jlibmpv renders through an external
OpenGL context). The workflow no longer needs a Rust toolchain.