This affects only the "new" VO API. The config() title argument was barely
used, and it's hardcoded to "MPlayer" in vf_vo.c. The X11 and the Cocoa
GUI backends, which are the only ones properly supporting window titles,
ignored this argument. Remove the title argument.
Add the vo_get_window_title function. All GUI VOs are supposed to use it
for the window title.
Remove code refreshing window contents after events such as resize
from vo_vdpau, vo_gl and vo_xv. Instead have them simply set a flag
indicating that a refresh is needed, and have the player core perform
that refresh by doing an OSD redraw. Also add support for updating the
OSD contents over existing frames during slow-but-not-paused playback.
The VOs now also request a refresh if parameters affecting the picture
change (equalizer settings, colormatrix, VDPAU deinterlacing setting).
Even previously the picture was typically redrawn with the new
settings while paused because new OSD messages associated with setting
changes triggered a redraw, but this did not happen if OSD was turned
off.
A minor imperfection is that now window system events can trigger a
single one-frame step forward when using vo_xv after pausing so that
vo_xv does not yet have a copy of the current image. This could be
fixed but I think it's not important enough to bother.
Previously the core sent VFCTRL_REDRAW_OSD to change OSD contents over
the current frame. Change this to VFCTRL_REDRAW_FRAME followed by
normal EOSD and OSD drawing calls, then vo_flip_page(). The new
version supports changing EOSD contents for libass-rendered subtitles
and simplifies the redraw support code needed per VO. vo_xv doesn't
support EOSD changes because it relies on vf_ass to render EOSD
contents earlier in the filter chain.
vo_xv logic is additionally simplified because the previous commit
removed the need to track the status of current and next images
separately (now each frame is guaranteed to become "visible" soon
after we receive it as "next", with no VO code running in the interval
between).
Separate passing a new frame to VOs using the new API into two steps.
The first, vo_draw_image(), happens after a new frame is available
from the filter chain. In constrast to old behavior, now the frame is
not actually rendered yet at this point (though possible slice draw
calls can already reach the VO before). The second step,
vo_new_frame_imminent(), happens when we're close enough to the
display time of the new frame that we'll commit to flipping it as the
next action and will not change the OSD over the previous frame any
more.
This new behavior fixes a previous problem with vo_vdpau and vo_gl in
the situation where the player is paused after decoding a new frame
but before flipping it; previously changing OSD in that state would
switch to the new frame as a side effect. It would also allow an easy
way to fix extra output files produced with something like "--vo=png
--frames=1" with precise seeking, but this is not done yet.
The code now relies on a new mp_image from the filter chain staying
valid even after the vf_vo put_image() call providing it returns. In
other words decoders/filters must not deallocate or otherwise
invalidate their output frame between passing it forward and returning
from the decode/filter call.
Rewrite control of the colorspace and input/output level parameters
used in YUV-RGB conversions, replacing VO-specific suboptions with new
common options and adding configuration support to more cases.
Add new option --colormatrix which selects the colorspace the original
video is assumed to have in YUV->RGB conversions. The default
behavior changes from assuming BT.601 to colorspace autoselection
between BT.601 and BT.709 using a simple heuristic based on video
size. Add new options --colormatrix-input-range and
--colormatrix-output-range which select input YUV and output RGB range.
Disable the previously existing VO-specific colorspace and level
conversion suboptions in vo_gl and vo_vdpau. Remove the
"yuv_colorspace" property and replace it with one named "colormatrix"
and semantics matching the new option. Add new properties matching the
options for level conversion.
Colorspace selection is currently supported by vo_gl, vo_vdpau, vo_xv
and vf_scale, and all can change it at runtime (previously only
vo_vdpau and vo_xv could). vo_vdpau now uses the same conversion
matrix generation as vo_gl instead of libvdpau functionality; the main
functional difference is that the "contrast" equalizer control behaves
somewhat differently (it scales the Y component around 1/2 instead of
around 0, so that contrast 0 makes the image gray rather than black).
vo_xv does not support level conversion. vf_scale supports range
setting for input, but always outputs full-range RGB.
The value of the slave properties is the policy setting used for
conversions. This means they can be set to any value regardless of
whether the current VO supports that value or whether there currently
even is any video. Possibly separate properties could be added to
query the conversion actually used at the moment, if any.
Because the colorspace and level settings are now set with a single
VF/VO control call, the return value of that is no longer used to
signal whether all the settings are actually supported. Instead code
should set all the details it can support, and ignore the rest. The
core will use GET_YUV_COLORSPACE to check which colorspace details
have been set and which not. In other words, the return value for
SET_YUV_COLORSPACE only signals whether any kind of YUV colorspace
conversion handling exists at all, and VOs have to take care to return
the actual state with GET_YUV_COLORSPACE instead.
To be changed in later commits: add missing option documentation.
Commit decec7f2a3 ("vo_vdpau: skip resize code if not fully
initialized") broke preemption recovery because the resize code stayed
incorrectly disabled when it would have been used to reinitialize
things during recovery. Revert that commit and add different checks
to avoid running various code when not in a fully functional state.
Window events or user commands could trigger a call to resize() before
config() had been called. This caused at least error messages. Add a
check to make resize() exit without doing anything in this case. It
will be called again later when the output is configured.
Allow negative values of the "deint" parameter and make them select
the same deinterlacing mode as the corresponding positive value, but
without enabling deinterlace on startup. This is useful for
configuration files. Also tweak the overall manpage description of
vo_vdpau a bit to make it sound less as if the VO would be for
hardware acceleration only.
The wrong variable was used as a function argument, and as a result
the code modified the usage_count field of non-refcounted mp_image
types. This error did not have any effect on visible behavior as no
code cares about the field value in the affected case.
Add a default case for unknown image format in create_vdp_decoder. I
think this condition shouldn't currently happen, but it's worth a
sanity check. Avoids a compiler warning about vdp_decoder_profile
being possibly used uninitialized.
Also remove an obsolete #ifdef (should always be true now).
The loop initializing handles in the output surface table to
VDP_INVALID_HANDLE ran over indices from 0 to vc->num_output_surfaces.
However it is first called before that variable is initialized. As a
result later code could try to destroy the handles which still had the
"non-invalid" value 0. Most likely this caused no visible effects; at
least on my machine no valid surface gets handle 0, and libvdpau just
returns an error for the resulting invalid calls. Change the code to
loop over the whole table. Also add code to print visible warnings if
libvdpau rejects a surface destroy call (some other places already had
checks but not all).
Make the number of output surfaces allocated by vo_vdpau configurable
and allow queuing multiple future frames with the VDPAU frame timing
functionality. There are two known issues that could be polished in
the future for particularly long queue time cases. First, toplevel
code should possibly wait for the queued changes to finish displaying
in some situations like switching files. Second, the VO can block
waiting for a surface to become available, and in the worst case this
wait can be almost as long as the maximum queuing time (potentially
causing audio underruns). However those issues are not high priority.
Add -vo vdpau suboption "studio" to produce output in RGB range
16-235. Man page description mostly taken from a patch by Lauri
Mylläri (but not code). Also slightly tweak the description of two
other suboptions on the man page.
Since I dropped the hack to support older libvdpau versions (without
MPEG4 / hqscaling features) from the FFmpeg version used in the build
repo it's better to disable those libvdpau versions in MPlayer
too. Change the configure check to require a newer libvdpau version,
and drop compatibility code in vo_vdpau.c.
After a recent library update on my system MPlayer started crashing on
exit in XCloseDisplay() if -vo vdpau had been used. With normal
linking it works OK. Since there should be little need for the dynamic
loading functionality any more, I'll just drop it rather than try
investigating the exact problem with dynamic loading and whether it
would be fixable. VDPAU is now linked normally with -lvdpau if
available.
Remove the special case for fullscreen; the same variables should have
the correct values in windowed and fullscreen modes. This fixes libass
subtitle positioning in the (non-typical) case where you have black
bands inside a window. Behavior in fullscreen and in the typical
windowed case without borders should not be affected.
Add -vo vdpau suboptions "queuetime_windowed" and "queuetime_fs" to
specify the maximum number of milliseconds how far into the future a
frame flip can be queued using the VDPAU presentation queue
functionality. The intended main use of these options is to allow
disabling use of the queuing feature on systems where using it causes
choppiness in other graphics behavior (this is an NVIDIA driver issue;
the video itself isn't affected).
Misplaced #endif broke compilation with old libvdpau versions that
lack VDP_VIDEO_MIXER_FEATURE_HIGH_QUALITY_SCALING_L1 #define.
Also add missing space to the text in related mp_msg() call.
Check the return value when setting any video mixer attribute and
print an error message if the operation failed. Also simplify code by
changing update_csc_matrix() to use the utility function added for
this.
Remove the help text explaining -vo vdpau suboptions that was printed
in case of parsing errors. It did perhaps have some value, but there
are also reasons to remove it: it was printed in an ugly manner in the
middle of output, most other MPlayer options do not have such internal
help texts either, and it was detailed enough that it required
maintaining documentation about the options in two separate places
(the man page and the help message).
Part of the code is currently under #ifdef to allow compilation with
older VDPAU library versions; that can be removed later.
Partially based on a patch by Carl Eugen Hoyos.
Add a property to select YUV colorspace. Currently implemented only in
vo_vdpau and vo_xv. Allows switching between BT.601, BT.709 and
SMPTE-240M (vdpau only).
The xv support uses the "XV_ITURBT_709" attribute. At least my NVIDIA
card supports that; I don't know whether other xv implementations do.
Bind the colorspace switch to the 'c' key by default. 'c' is currently
used by vo_sdl for some fullscreen mode change thing, but at the moment
that does not conflict and if it will in the future then vo_sdl can
change.
VDPAU part based on a patch from Lauri Mylläri <lauri.myllari@gmail.com>
Main things added are custom frame dropping for VDPAU to work around
the display FPS limit, frame timing adjustment to avoid jitter when
video frame times keep falling near vsyncs, and use of VDPAU's timing
feature to keep one future frame queued in advance.
NVIDIA's VDPAU implementation refuses to change the displayed frame
more than once per vsync. This set a limit on how much video could be
sped up, and caused problems for nearly all videos on low-FPS video
projectors (playing 24 FPS video on a 24 FPS projector would not work
reliably as MPlayer may need to slightly speed up the video for AV
sync). This commit adds a framedrop mechanism that drops some frames
so that no more than one is sent for display per vsync. The code
tries to select the dropped frames smartly, selecting the best one to
show for each vsync. Because of the timing features needed the drop
functionality currently does not work if the correct-pts option is
disabled.
The code also adjusts frame timing slightly to avoid jitter. If you
for example play 24 FPS video content on a 72 FPS display then
normally a frame would be shown for 3 vsyncs, but if the frame times
happen to fall near vsyncs and change between just before and just
after then there could be frames alternating between 2 and 4
vsyncs. The code changes frame timing by up to one quarter vsync
interval to avoid this.
The above functionality depends on having reliable vsync timing
information available. The display refresh rate is not directly
provided by the VDPAU API. The current code uses information from the
XF86VidMode extension if available; I'm not sure how common cases
where that is inaccurate are. The refresh rate can be specified
manually if necessary.
After the changes in this commit MPlayer now always tries to keep one
frame queued for future display using VDPAU's internal timing
mechanism (however no more than 50 ms to the future). This should make
video playback somewhat more robust against timing inaccuracies caused
by system load.
Clean up code related to frame buffering and generate pts information
also for the next frame in the output queue. The timing information
will be used in a following framedrop patch.
This commit adds one frame of buffering delay in vo_vdpau and
increases the number of buffered vdpau video surfaces from 3 to 4. The
delay increase makes it more important to fix remaining code in
MPlayer that doesn't deal well with filter/VO delay; OTOH it should
help any decoding/filtering parallelism in the underlying VDPAU
implementation as now filtering a frame for display can happen while
the next one is being decoded.
check_events() first checked for a RESIZE event and called resize() if
needed, and then queued a frame to be reshown if in pause state and
the event was either RESIZE or EXPOSE. The most obvious problems with
the code were:
- resize() already called flip_page() internally, so the code in
check_events could lead to _two_ frames being queued.
- The call in resize() didn't depend on pause status, so the
behavior was inconsistent.
- The code in check_events() actually queued the wrong output
surface. It showed the same surface as flip_page() would show
_next_, while it should have shown the previously shown one. This
typically led to the screen contents changing to a previous
state, as the new surface had not been initialized and had
contents from a previous use.
Fix the double update. Make resize() also only immediately update the
video if paused (this also affects changing to/from fullscreen) and
otherwise leave the old window contents be until the next frame. Queue
the right frame in check_events(). Also make resize() a bit more
careful to only show contents if they were successfully updated
(though a case where we're paused without content to show shouldn't
normally happen).
Add interfaces to allow VO drivers to add or remove frames from the
video stream and to alter timestamps. Currently this functionality
only works with in correct-pts mode. Use the new functionality in
vo_vdpau to properly support frame-adding deinterlace modes.
Frames added by the VDPAU deinterlacing code are now properly timed.
Before every second frame was always shown immediately (probably next
monitor refresh) after the previous one, even if you were watching
things in slow motion, and framestepping didn't stop at them at all.
When seeking the deinterlace algorithm is no longer fed a mix of
frames from old and new positions.
As a side effect of the changes a problem with resize events was also
fixed. Resizing calls video_to_output_surface() to render the frame at
the new resolution, but before this function also changed the list of
history frames, so resizing could give an image different from the
original one, and also corrupt next frames due to them seeing the
wrong history. Now the function has no such side effects. There are
more resize-related problems though that will be fixed in a later
commit.
The deint_mpi[] list of reserved frames is increased from 2 to 3
entries for reasons related to the above. Having 2 entries is enough
when you initially get a new frame in draw_image() because then you'll
have those two entries plus the new one for a total of 3 (the code
relied on the oldest mpi implicitly staying reserved for the duration
of the call even after usage count was decreased). However if you want
to be able to reproduce the rendering outside draw_image(), relying on
the explicitly reserved list only, then it needs to store 3 entries.
Add code to reinitialize all VDPAU objects if a display preemption
condition occurs. Reinitializing them in the middle of playback will
cause video corruption at least until the next keyframe when using
hardware decoding, but decoding does seem to recover after a keyframe.
Create a single large bitmap surface for EOSD objects and pack all the
bitmap rectangles inside that. The old code created a separate bitmap
surface for every bitmap and then resized the cached surfaces when
drawing later frames. The number of surfaces could be large (at least
about 2000 for one sample subtitle script) so this was very
inefficient. The old code also used a very simple strategy for pairing
existing surfaces to new bitmaps; it could resize tiny surfaces to
hold large glyphs while using existing large surfaces to hold tiny
glyphs and as a result allocate arbitrarily much more total surface
area than was necessary.
The new code only supports using a single surface, freeing it and
allocating a larger one if necessary. It would be possible to support
multiple surfaces in case of hitting the maximum bitmap surface size,
but I'll wait to see if that is actually needed before implementing
it. NVIDIA seems to support bitmap surface sizes up to 8192x8192, so
it would take either a really pathological subtitle script rendered at
a high resolution or an implementation with lower limits before
multiple surfaces would be necessary.
The packing algorithm should successfully pack the bitmaps into a
surface of size w*h as long as the total area of the bitmaps does not
exceed 16/17 (w-max_bitmap_width)*(h-max_bitmap_height), so there
should be no totally catastrophic failure cases. The 16/17 factor
comes from approximate sorting used in the algorithm. On average
performance should be better than this minimum guaranteed level.
