Until now the pool only released its packets on shutdown or when the
cache size was reduced at runtime, so it kept growing across a session.
The amount of packets would be indirectly bound by demuxer cache size,
else the packets are reused. Still over long playback or switching
files, this may give impression of elevated memory usage, or even memory
leak. This was particularly visible with cache-on-disk, where users
reasonably expect RAM usage to stay low.
Pop now keeps the first POOL_HOT_RESERVE packets at the head of the
freelist for fast reuse, and frees up to POOL_GC_BATCH packets from
beyond that on each call. Bulk pushes (e.g. cache reset on file end)
remain no-work policy. The pool drains across subsequent allocations
rather than stalling on a single huge free. Self-balancing without any
size tracking.
This drains the pool in steady state. Note that on file switch we
effectively reuse only ~N / POOL_GC_BATCH packets from the old cache,
because each pop reuses just one packet while freeing up to
POOL_GC_BATCH. That's fine, both free and alloc calls are throttled to
the same per-pop rate, so allocator pressure stays bounded even though
not all packets get reused.
And make them actually async when copied from host ptr. The issue was
when we used DR buf upload, which is always async, it was not
synchronized in anyway.
Fixes: #13811
It is more expected from GUI application to stay open, even after
playback. Additionally this gives user a chance to inspect possible load
failures, while previously mpv would disappear.
Add adaptor for arrays of properties, this is not used as real option,
as those are always string list. But as a connector struct for creating
typed node arrays, in (sub)_properties.
Lavf exposes programs/renditions as a flat track list tagged with
program_id. We respect this metadata and group tracks into their
respective programs, as tracks from different programs may be completely
unrelated to each other.
This commits adds an option to flatten the editions and ignore the
grouping.
Fixes: https://github.com/mpv-player/mpv/issues/17836#issuecomment-4345930205
When the current edition supplies its own track of a given type
(e.g. yt-dlp muxed editions ship their audio paired with the video),
shared program_id == -1 tracks of that type are filtered out of
both the track list and auto-selection. Editions that don't ship
their own track (DASH video-only editions sharing the audio pool)
continue to see the program-agnostic tracks.
Extract the codec descriptor out of mp_format_track_metadata into a
shared demux_append_codec_desc() helper, and use it from a new
demux_compose_edition_title() composer that both demux_timeline.c and
demux_lavf.c call when building editions. Codec block is only added
when the program has exactly one stream of that kind, and fields
already present in the prefix (e.g. WxH from yt-dlp's format label)
are skipped to avoid duplication.
There is no need to pipe things through shell and grep/awk, python is
perfectly fine word processor. This makes it simpler to reason what is
extracted and from where and also makes the script faster as we spawn
only single process directly, not shell with multiple children.
100 is enough and avoids mistimed frames to affect long term stats. It
would be good to improve vsync estimation accuracy, but increasing
averaging size alone is not the best solution.
This reverts commit 4e44fe024c.
The old code had several interacting problems that prevented A/V from
reaching a steady state.
When the A/V difference was large, it applied the maximum allowed
compensation in one step. That was fine for catching up, but the value
was never re-evaluated, so audio speed routinely overshot the target.
Once the difference crossed zero (or re-entered the acceptable range) it
would run a linear regression over recent av_diff samples and snap the
audio speed to the estimated drift. The fatal flaw was that those
samples already included the controller's own compensation: even when
the underlying drift was zero, the controller would "detect" the drift
it had just induced and try to cancel it. Between these rare, sparse
updates audio speed stayed frozen, so in practice the controller was
introducing drift more than it was fixing it. On high fps outputs,
where the vsync range is tight, this degenerated into a ping-pong:
overshoot, snap back, overshoot again, never settling.
The new code replaces the bang-bang direction controller and one-shot
regression with a proportional controller driven by a low-pass-filtered
av_diff. The LP filter (1 s time constant) rejects per-frame vsync
jitter so the compensation no longer chases noise. The target then
decays av_diff exponentially toward zero with a 10 s recovery time.
Adaptive slew caps per-frame motion to the desired correction magnitude,
so large excursions are absorbed quickly while near-equilibrium motion
stays below audibility. The result is a continuous, gentle control loop
that actually reaches steady state and does so without audible
audio-speed modulation.
Those values are often very small, but it's still good information to
have how they are changing.
This brings those values closes to old `%+.05f%%` that were used here.
scRGB is basically intermediary format, but we don't really know what
would be the real target. And since we send floating-point buffer, we
can just expect the downstream will handle dithering and quantization at
some point.
Mark every filter window whose libplacebo counterpart is resizable as
resizable in mpv too.
This allows --tscale-radius to be used. Prior to this, vo=gpu
silently ignored the radius override while vo=gpu-next honored it,
producing different output for the same options.
When the effective filter bound (radius * blur) is smaller than the
smallest tap distance, mp_compute_weights produces all-zero weights and
the subsequent normalization divides by zero, producing NaN weights that
render as black. This can happen for resizable filters where the user
picks a radius below 0.5.
Detect sum == 0 and fall back to nearest-neighbor on the closest tap.
Alpha blending is done in gamma light of the target, when the overlay is
drawn. Depending on VO and target format, the alpha value can act quite
differently.
Lower it a little to avoid excessive transparency in some cases. This
should be fixed or workaround in better way, but for now just adjust the
value.
This adds proper support for MPEG-TS programs and HLS variants. Now
program selection is exposed as editions and tracks are grouped under
such edition. If there are external tracks added, they are available in
all editions.
track-list/N is not required to match aid/vid. They are encoded as
track-list/N/id.
This fixes changes in 1a558bf5c2 where it
was asumed that track id is the same in track-list, which is not true,
in general sense.
The dynamic allocation of this strings is completelly pointless. It adds
complexity with mp_exec_once to init aexit() to free the string, which
was leaked anyway, because the structure is thread_local and atexit only
data from the main thread was freed.
Remove all this, and just use static buffer.
Makes it nicer for speaker icon to not move around on different volume
levels. Also remove scaling, it was not needed.
This required to rewrite the exclamation mark drawing code, made it look
similar to old icon now.
LLVM even in freestanding build required memset/memcpy/memmove to be
implemented. We don't use them directly but the calls are generated from
the code, which are later optimized away. Since this suppose to be small
utility application, we can always optimize it.
Fixes: #17664
This fixes multiple overflows when passing huge values. Which causes
either invalid memory accesses when memory allocations are wrong size or
huge allocations leading to OOM.
Reported by: Trail of Bits
When force-window-position=no place window on requested position only on
init (when VO requests `force_pos`), except that keep existing
position on auto resizes.
This applies to both user provided geometry and default center
position.
Fixes: #17588
This fixes dynamic hwdec switching, which would often fails, because
hw decoder were feed data without context, which caused decoding errors,
which in turn disabled hwdec.
command.c does queue_seek() after decoder reinit, but it may already be
too late, by the time seek to keyframe finishes.
Fix this, by just discarding the data until we have safe point to start
decoding again. Limit this only for hwdec to avoid formats that may not
have ffmpeg parser and not give us keyframe information and limit to 96
frames, just as safety to not lock the player, but this shouldn't happen
in practice.
This should also fix cases, when media file is cut in a way that would
cause decoding errors at the start.
Fixes: #17484
Fixes some corner cases when resizing window with keep aspect ratio
enabled. Limit correctly window to working area, also after adjusting
aspect ratio to avoid unexpected resizes and window not staying in
position. Use integer math to have consistent rounding, to avoid window
"drifting".
Fixes: #8728
libavutil deprecated AVVulkanDeviceContext.lock/unlock_queue without
replacement [1]. This prevents us from using queue locking, as those
callback will be removed in lavu. Set those callbacks as long as they
are still present, they will be noop in libplacebo with
`VK_KHR_internally_synchronized_queues`.
It's unclear what will happen to devices that do not support
`VK_KHR_internally_synchronized_queues` when the lock/unlock_queue
callbacks are removed.
[1] https://code.ffmpeg.org/FFmpeg/FFmpeg/commit/c102e89448a06d9920920923a4008fee3db4ea37
This uses Fluent UI System Icons [1] as a source for icons. The glyps
are generated using `TOOLS/gen-fluent-glyphs.py` script, which downloads
Icon and convert for our mpv font to use. Glyphs are vendored in our
repository to avoid external dependency.
Script could be extended to add more icons/variants easily, currently
only filled version are imported.
[1] https://github.com/microsoft/fluentui-system-icons
There is concern that with certain user options or script the value of
`{sub,osd}-margin-y` value may be persisted or overwrite user option
when it's changed while OSD is visible and offset is applied.
Fix this by using dedicated option.
This allows to retain deadzone in the middle of the screen when both top
and bottom bars are enabled. We also doesn't have to show bars that
early before moving mouse down.
If `no-border` or `no-title-bar` was used, window controls would enable
also in fullscreen, this was inconsistent, as window decorations are not
visible ragardless of those options in fullscreen mode.
Those strings has to be owned by mp_codec_params, not decoder wrapper
priv.
Our assumption is that we set those strings here once on decoder init
and they will not be changes during the life time of the
mp_codec_params, hence why we need to make sure they are not deallocated
too soon, in case they happen to be in use.
Fixes, use after free in some cases, when tracks are deselected.
Kacper Michajłow