With the mpv backend on Android, every subtitle change cost two dropped
video frames on a Fire TV Cube (36 per minute with PGS SDH subtitles, 15
with srt, #2240) and accounts for the residual drops on the Shield in #2202,
whose sessions also had subtitles on. The diagnostic build showed the drops
in pairs: a frame handed over on time, then the video thread waking 90-100
ms late. vo_mediacodec rasterized subtitles and blitted them into the OSD
Surface on that thread, and locking the Surface waits CPU-side for the
compositor - dequeue plus the release fence, one or two vsyncs, 42-83 ms at
23.976 Hz. Kodi, VLC 4 and ExoPlayer all keep subtitle output off the
presentation thread.
Move the native pin to edde746/mpv-build@88c1b135e7: the whole OSD stage
runs on its own thread, fed with the video pts at every flip, so the video
thread never waits on the compositor or on libass. Subtitles latch on the
same vsync as the video in the common case and one later when the lock has
to wait. Shield on vo=mediacodec with ASS subtitles: no video-thread stage
over 10 ms and 0 drops across five 150 s runs, subtitles verified on
screen; DTS-HD MA 4K unchanged at 0 drops. The pin keeps the #2202
diagnostics and adds VO stage timing to them. Apple, Linux and Windows
assets are unchanged.
The smoothed passthrough clock from aeb06c2c9 (edde746/mpv-build 0105)
still let the Shield in #2202 drop a frame every 70 s or so on TrueHD and
DTS-HD MA. The average handles head jitter completely - on a Shield with
the raw head artificially held for 150 ms it went from 410 dropped frames
in 200 s to 0 - but it regresses on stalls: when the head freezes (device
underrun, rebuffer, a HAL that stops updating) while the wall clock runs,
the estimate runs past what was written, (written - head) wraps, and
getLatency reports 0 - the core sees the whole device buffer vanish in one
step and drops video. With the AO starved for 400 ms every 15 s that was 77
drops in 170 s against 8 with the raw clock, and a Box R 4K Plus hit the
wrap 361 times in four minutes of ordinary DTS-HD playback.
Move the native pin to edde746/mpv-build@a58f0c1336: the estimate is
clamped to the written position and its window restarted on overshoot, and
between reports it may deviate from "last report + elapsed" by at most 10%
of the elapsed time unless it is a second or more off - the two guards
media3's AudioTrackPositionTracker keeps over the same average. Starvation
run: 9 drops, no wrap; real HAL with pauses and seeks: 0 drops. The pin also
carries verbose-level diagnostics for this issue (per-drop timing in vo.c,
passthrough clock anomalies and a 10 s summary in ao_audiotrack) so the
next log from the reporter pins whatever remains. Apple, Linux and Windows
assets are unchanged (their content keys did not move).
With ambient lighting enabled, ASS subtitles anchored to a corner (and PGS
bitmaps) rendered against the whole screen instead of the video: the effect
stretches mpv's frame to the window with video-aspect-override and lets the
shader composite the real picture inside it, and mpv places subtitles against
that stretched frame.
The pinned libmpv now carries --sub-video-rect-aspect (edde746/mpv-build
ec08018), which makes mpv derive its subtitle margins from the picture's real
aspect inside the displayed rect. AmbientLightingService hands it the video
aspect before overriding the frame, so a libmpv without the option refuses
cleanly with the frame untouched, and resets it on disable.
Verified on a Pixel 7 (2400x1080) with a 16:9 file whose OP credits are
\pos-anchored: subtitle extents are pixel-identical with ambient lighting on
and off, and mpv reports OSD borders l=240 r=240 for the overridden frame.
close#2120
High-bitrate HEVC episodes showed smeared frames and green flashes for the
first 10-15 seconds of playback on the Fire TV Stick 4K Max 2nd gen with the
mpv backend, while ExoPlayer played the same file cleanly (#2227). The
reporter's clip failed at exactly two timestamps on every run: the two IDR
frames larger than 1 MiB.
The FFmpeg mediacodec decoder never set "max-input-size" on the codec
format, so OMX.MTK.VIDEO.DECODER.HEVC allocated 1,052,672-byte input buffers
and FFmpeg split any larger packet across several queueInputBuffer calls.
The MediaTek VPU treats each buffer as a complete access unit, rejects the
remainder ("invalid u4BufferAccumulated"), fails the slice and conceals with
lost-picture references until the next IDR. ExoPlayer is unaffected because
media3 sizes HEVC input buffers at max(2 MiB, w*h*3/4) and never splits a
sample.
Move the native pin to edde746/mpv-build@83879375c9, which sizes the
buffers the same way ExoPlayer does for H.264, HEVC, MPEG-4, VP8, VP9 and
AV1 and logs a warning if a packet still does not fit. Verified on the
AFTKRT with the reporter's clip: zero VPU errors and no lost pictures across
repeated runs, and an H.264 file with ~1 MB keyframes still plays on the
MediaTek AVC decoder; both files also play on a Pixel 7 (Exynos C2). Apple,
Linux and Windows assets are unchanged (their content keys did not move).
close#2227
Playing lossless passthrough audio (TrueHD, DTS-HD MA) on an Nvidia Shield
running Android 9 dropped a video frame every ten to twenty seconds with the
mpv backend, while lossy passthrough and PCM were unaffected and A/V sync
never moved (#2202).
IEC 61937 and raw passthrough tracks in ao_audiotrack never use
AudioTimestamp; their clock is AudioTrack.getPlaybackHeadPosition() taken at
face value, which on a direct track is the HAL's render position and
advances in HAL-period steps - 8192 frames at 192 kHz (42.7 ms) for the HBR
output on the Shield, longer than a 24p frame. The AO thread samples it once
per chunk and feeds it straight into the end time mpv schedules video
against, so a forward snap larger than a frame's remaining lead makes mpv
drop that frame; lossy passthrough runs at 48 kHz where the period stays
below a frame.
Move the native pin to edde746/mpv-build@af74e1906e, which estimates the
passthrough position the way media3 does: a ten-sample average of
(head - clock) offsets taken at least 30 ms apart, so a period-sized step is
spread over the window instead of landing on one frame. PCM tracks keep
their AudioTimestamp path. Apple, Linux and Windows assets are unchanged
(their content keys did not move).
Dolby Vision profile 7 titles (UHD Blu-ray remuxes, typically TrueHD or
DTS-HD MA) never showed a frame on the Fire TV Stick 4K Max 2nd gen with the
mpv backend: the spinner stayed forever while audio came up. Profile 8.1
titles carrying HDR10+ on the same device went fully black, controls
included, with audio playing.
Both are the FFmpeg mediacodec DV packet filter. It emitted every rewritten
access unit with a five-byte "00 00 00 00 01" start code - the leading zero
was copied once as head bytes and again as part of the first NAL region -
and MediaTek's VPU rejects such units ("H265_decode invalid NALU"), so the
profile 7 conversion and strip paths both produced nothing. And the filter
never dropped in-band HDR10+ SEI, which these chipsets cannot take next to
the RPU on the native DV decoder (#1296 on the ExoPlayer path).
Move the native pin to edde746/mpv-build@0ad6388c, which counts that zero
once and strips the ST 2094-40 SEI whenever a video/dolby-vision decoder is
selected. Verified on an AFTKRT with the published armeabi-v7a asset: a
P7.6 remux sample decodes on both the HEVC (strip) and DVHE.STH (convert)
paths with no VPU errors, and a P8.1 + HDR10+ remux plays on the DV decoder
with a single output-format change. Apple, Linux and Windows assets are
unchanged (their content keys did not move).
close#2176
Enabling Deinterlacing had no effect on Android and Android TV for
interlaced MPEG-2 (and any other software-decoded stream): mpv asks
libavfilter for bwdif, and the pinned Android and Linux FFmpeg builds were
configured with --disable-filters, so the request failed with
"filter 'bwdif' not found" and the frame passed through combed. Windows
and Apple ship FFmpeg with the filter, which is why the same file
deinterlaced on Windows.
Move the native pin to edde746/mpv-build@9bfbcd69, which builds bwdif into
the Android and Linux FFmpeg. Apple and Windows assets are unchanged
(content-addressed keys did not move). Verified on a Shield TV Pro with the
published arm64 asset: 480i MPEG-2 now deinterlaces cleanly.
close#2224
AAC/AC3/EAC3 5.1 tracks at 44.1 kHz played as stereo through mpv on Android:
ao_audiotrack sized its buffer as 75 ms aligned to the sample size, which at
44.1 kHz is not a whole number of 6-channel frames, so AudioTrack refused it
("Invalid audio buffer size") and mpv fell back to the stereo-only OpenSL ES
output. 48 kHz sources (DTS, or audio-samplerate=48000) happened to align.
Pin mpv-build 1779135, which frame-aligns the audiotrack PCM buffer
(edde746/mpv-build@d89cfbb). Only the Android libmpv key moves.
close#1445
Plezy consumed mpv through four unrelated supply chains: an MPVKit fork
via SwiftPM for the Apple platforms, a libmpv-android fork's AAR for
Android, an in-CI from-source build for Linux, and an unpinned
sourceforge mpv-dev 7z for Windows. All four now consume the same
per-commit, content-addressed binaries from
https://github.com/edde746/mpv-build, pinned to one commit and built
from one set of pinned sources (mpv v0.41.0 on Apple/Android/Windows,
ffmpeg n8.0.1, our libass fork).
- Apple: the SwiftPM package moves from edde746/MPVKit to
edde746/mpv-build across the ios/macos/tvos projects;
scripts/set_mpvkit_revision.sh becomes set_native_revision.sh, writes
every pin site plus the new repo-root mpv-build.lock.json, and
tvos/scripts/wire_mpv.rb derives the package repo from the locks.
- Android: the mpv Kotlin API and JNI glue move in-app under
android/libmpv (repackaged com.edde746.plezy.libmpv, exports renamed,
shrinker rules covered), and the module downloads per-ABI native
tarballs (lib/*.so incl. libc++_shared.so + include/) driven entirely
by mpv-build.lock.json, with a PLEZY_LOCAL_MPV_DIR escape hatch for
locally built artifacts. The fork AAR and its Maven coordinates are
gone.
- Linux: CI downloads the prebuilt self-relocating libmpv prefix
(lock-driven, sha256-verified) instead of compiling mpv/ffmpeg/dav1d/
libplacebo/shaderc from source; linux/packaging/build-libmpv.sh and
its test are deleted and native-inputs.json shrinks to the simdutf
entry the CMake builds still fetch.
- Windows: both arches FetchContent the mpv-build dev zips with
URL_HASH enforcement, replacing the checksum-less sourceforge
download and its ARM64 7-Zip special case; guard scripts updated.
The lock plus the Apple pin sites all point at mpv-build commit
d7c3d559, whose manifest was verified asset-by-asset against the
published release digests (17/17 match). Verified locally: wire and
pin-script suites green, runtime-input checks green, all four Android
ABI tarballs downloaded/verified/extracted through the real Gradle
tasks, the Windows FetchContent block exercised end to end through
cmake, the Linux asset hash and layout checked against the workflow
contract, and xcodebuild resolved the flipped SwiftPM graph with
SwiftPM validating every binary checksum.
edde746