Information about individual chapters was printed during demuxer
opening phase, and total chapter count (ID_CHAPTERS) was printed
according to mpctx->demuxer->num_chapters. When playing a file with
ordered chapters, this meant that chapter information about every
source file was printed individually (even though only the chapters
from the first file would be used for playback) and the total chapter
count could be wrong. Remove the printing of chapter information from
the demuxer layer and print the chapter information and count actually
used for playback in core print_file_properties().
Also somewhat simplify the internal chapters API and remove possible
inconsistencies.
Something like the OSD menu functionality could be useful. However the
current implementation has several problems and would require a
relatively large amount of work to get into good shape. As far as I
know there are few users of the existing functionality. Nobody is
working on the existing code and keeping it compiling at all while
changing other code would require extra work. So delete the menu code
and some related code elsewhere that's used by nothing else.
Some functionality provided by the windowing backend (such as x11,
win32) is optional. The function pointers in MPGLContext are NULL if
the functionality is not implemented. Check them before calling them,
instead of crashing. This happened at least on Windows with the
"vo_ontop" command.
The register combiner color conversion is broken and seems to use a
slightly incorrect color matrix (the image looks gray-ish). Completely
remove all code related to nVidia register combiners.
Unless you have an ancient nVidia GPU, there's no reason to prefer register
combiners over fragment shaders. Users with ancient GPUs without fragment
shader support can just use -vo xv.
Passing yuv=1 (register combiners) as sub option will print a warning and
use yuv=2 (fragment shaders) instead.
Using the ati-hack and force-pbo suboptions (both can be enabled
automatically when particular ATI drivers are detected) could lead to
segfaults due to incorrect length calculation.
The problem is that width*((bpp+7)/8) == width*(bpp+7)/8 doesn't hold
true in general. The old code used bpp/8, because bpp was always
guaranteed to be divisible by 8. When commit 20256a8a64 fixed bugs
around 10 bit pixel formats, this assumption was broken, and resulted
in out of bounds memory accesses.
Additionally, the color values that were used to clear image borders
were incorrect and showed up as pink borders on the right/bottom sides
of the video when playing 10 bit content. Make the clear value
approximately correct; the borders are still filled with a single byte
value even though formats with 9-16 bits should be cleared with 2-byte
values with the low byte 0.
This commit also contains some other minor cleanups with no functional
changes.
Now modifiers such as shift, control, and alt are handled.
Unfortunately this requires a hack to deal with a SDL quirks when combining
CTRL and a letter key.
Also add some missing key syms.
Now all windowing specific code is in gl_common.c.
init_mpglcontext() used to set dummy callbacks for non-optional windowing
callbacks. Remove these, as they only lead to confusion.
The actual work is done by the existing SDL code. This commit merely
makes it possible to explicitly select the SDL backend ("gl" alone
uses SDL only if the X11 and win32 backends are not available, while
the new "gl_sdl" always forces use of SDL).
Also disable YUV conversion method autodetection when SDL is used.
This gets rid of a temporary window that appears for a moment and is
immediately closed again. SDL can't deal with the VOFLAG_HIDDEN flag,
which is needed to create an invisible GL context (when the
autodetection is run, the video size isn't yet known to the VO, and
creating a window then resizing would cause problems with window
placement). Instead always pick the fragment program method by default
(yuv=2). This change affects the normal "gl" VO too if it chooses the
SDL backend.
Apparently this is fully redundant given the global "noaspect" option.
Refuse this suboption, and output an error message suggesting the
correct option.
Reformat vo_gl.c, gl_common.c, gl_common.h.
Remove all global variables and move them into a context struct (the
Windows and SDL backends still refer to global_vo though).
Change vo_gl.c to use the "new" VO API.
Delete the vo_gl2 and vo_matrixview implementations.
vo_gl2 was barely useful anymore. It was a hack based on an old
vo_gl.c version, and all it did differently was rendering the video in
tiles instead of using a single texture. That made it work with some
crappy OpenGL implementations. These days all GPUs support textures of
at least 2048x2048 pixels, which is enough for HD playback. On the
other hand, gl2 suffered from various bugs and deficiencies, all of
which are fixed in gl. Its existence also confused users; many thought
that gl2 is the next version of gl and attempted to use it, even
though it's much worse than gl and they should have used that instead.
Should it turn out that tiling is actually useful, it should be
implemented in vo_gl, instead of keeping vo_gl2 alive.
vo_matrixview was a toy that couldn't even properly display a video.
All it did was display a screensaver-like animation that showed "a
Matrix-like running-text effect". (mplayer is not a screensaver.)
After commit 75eab4f72a ("video, options: implement better YUV->RGB
conversion control"), libvo/csputils.c symbols are required by core
code. However, the file was still only compiled if GL was enabled,
causing a build failure with GL disabled. Fix by making csputils.c
compilation unconditional.
Demux_demuxers checked a pts value against 0 to see if it was unset,
but other code uses MP_NOPTS_VALUE for that now. As a result audio and
subtitle streams could seek to a large negative position (effectively
to 0) instead of the correct target position. This broke --audiofile;
the --initial-audio-sync code could compensate for wrong demuxer seek
up to 5 minutes from the start of the file, masking the bug, but
seeking further than that audio would seek to 0 instead.
Note that the current --audiofile implementation using the
demux_demuxers wrapper is a fundamentally unsound design and still not
expected to generally work properly even after fixing this particular
problem.
Codec selection for audio and video decoding had a "dynamic plugin"
feature that tried to load a shared library for any codec that had not
been enabled at compilation (disabled by default, but could be enabled
with --enable-dynamic-plugins configure switch; for unknown reasons
some distro packages have enabled it). The implementation was buggy
and could cause normal codec selection fallback to fail if the feature
was enabled. I'm not aware of any real uses of such dynamic plugins
and the feature seems questionable anyway (there are no ABI guarantees
that would make it safe to use). Remove the buggy feature.
Libav stopped automatically filling missing codec_tag field for raw
codecs based on pix_fmt in libav commit bb416bd68c ("lavf: do not set
codec_tag for rawvideo"). This broke demux_lavf for raw video in
formats like YUV4MPEG, as the video format was not exported from
demux_lavf in any form (the information only existed in the pix_fmt
field of the struct AVCodecContext from libavformat, and that is not
exported). Add an explicit call to avcodec_pix_fmt_to_codec_tag() to
set the codec tag again so that selecting the correct raw decoder
based on the tag works.
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.
vo_gl.c hardcoded the assumption that YUV formats always use 1 byte per
component. This corrupted the output when playing 10 bit h264 video.
Only the PBO code path was affected. PBOs are normally unused, unless ATI
is detected, or the force-pbo option is used.
getch2.c read data into a "char" array, and returned values other than
escape sequences directly from there. This meant that it could return
negative values (except on platforms where "char" is unsigned) if the
input contained bytes >= 128. This would break later parsing in
input.c as the values would be interpreted as having the MP_KEY_DOWN
flag set, which would make the key binding code think a key is held
down (and never released). Fix by changing the buffer type to unsigned
char.
The bug itself was very old, but started triggering visible breakage
more easily after commit 82b8f89bae ("input: rework event reading and
command queuing"). Before that the key values would be passed through
the input.c "key read function" interface, which (mis)interpreted the
negative values as errors from the function, and in most cases
discarded them without much visible effect.
Libass has its own BiDi handling now, and preprocessing the subtitles
with FriBiDi before passing them to libass breaks things. Disable our
own FriBiDi code when libass rendering is used. This affects external
subtitle files of other formats parsed with subreader.c and converted
to ASS tracks.
Set the "Style" attribute for subtitle events created in sd_ass to
match the "default_style" attribute of the track. This is required to
make --ass-styles work with recent libass versions (otherwise the
event would use style 0, which is a fallback style added by libass
now).
Commit 378ada847c ("sub/ass: use default
style, not first style unconditionally") fixed the same issue for
external subtitles converted with mp_ass_read_subdata(); this commit
fixes it for muxed ones.
The last libpostproc major version change from 51 to 52 changed the
type of the "pp_help" symbol from a pointer to help text to the help
text itself. This made --pphelp crash. Change the option definition to
match the new type. This probably makes it crash if compiled against
older libpostproc, but the option is not important enough to try
supporting that (I've seen no reports of the crash, probably people
just don't use the option).
Add functionality to mark options that depend on features disabled at
compile time as disabled rather than not compiling the option
definitions at all. This allows printing a warning about the option
not being available because of a disabled feature, instead of just
"unknown option". Because the option definitions are still compiled
fully, this only works for definitions that do not reference symbols
which are not available if the feature is disabled. Use the new
functionality for options depending on libass.
1916b95b8 changed two function types from returning "void" to
returning "int", but was missing changes to add "return 0;" to the
functions. Fix.
The reason for the change in the original commit was that the
functions were called through a function pointer returning int anyway,
so the missing return probably made things no more likely to fail at
runtime than they were before that commit. However, it caused a
compilation failure with clang, which treats non-void function not
returning a value as a fatal error (in GCC it's just a warning).
Previously the code converting text subtitles to ASS format converted
newline characters, and only those, to ASS "new line" markup. If the
subtitles contained "\r\n", the "\r" was thus left in the text. In
previous libass versions the "\r" was not visible, but in the current
one it produces an empty box. Improve the conversion to remove the
"\r" in that case. Also treat a lone "\r" as a newline.
demux_mf allocated the "type" suboption of "--mf" with strdup if it
was not explicitly set. This caused a crash after playing an mf://
entry. Fix to use talloc instead.
When parsing the command line, map "--no-foo" to "--foo=no" if an
option named "foo" exists and is a flag option. Non-empty parameters
are not allowed with this syntax ("--no-foo=no" is invalid).
This implementation is different from the existing "--nofoo" variants
for most flag options. Those are implemented as completely separate
options; there's an option named "fs" and a separate option named
"nofs" (thus "--no-nofs" actually works after this change...). The
reason for adding the new syntax is to support the much more standard
"--no-" prefix and to allow eventually cleaning up the option handling
(though the "nofoo" variants of existing options can't be removed soon
due to backwards compatibility).
Rewrite the csputils.c code generating a conversion matrix for
YUV->RGB conversions (currently used by vo_gl only). Functional
differences:
- The separate "mplayer default" colorspace is removed, and BT.601 is
used instead (the default colorspace was in fact BT.601; see below).
- The old code was missing chroma scaling. As a result the "mplayer
default" colorspace actually mapped to BT.601, and everything else
was buggy (I guess the other colorspaces were added with particular
coefficient semantics, without understanding that the original
"default colorspace" was actually BT.601 and why its coefficients
differed from the added version).
- The old code had a bug in the equalizer hue equations.
- The old code assumed that for specifying whether input and output
were limited-range or full-range YUV or RGB it would make sense to
specify "no conversion" meaning full-range YUV to full-range RGB or
limited-range YUV to limited-range RGB. This isn't true; limited-
range YUV has different ranges for luma and chroma (16-235
vs 16-240) which means you have to scale chroma for limited->limited
conversions. The new code assumes limited->limited conversions for
the levelconv parameter 2. It'd probably make sense to change the
API later to specify the ranges of input and output separately.
- The undocumented EBU and XYZ colorspaces are removed. I doubt any
videos use these. Also the EBU colorspace looks like it'd expect
a different input range - at least no input would map to full RGB
red as it was.
FFmpeg has increased FF_INPUT_BUFFER_PADDING_SIZE to 16 (unlike Libav
which still has it at 8). Raise MP_INPUT_BUFFER_PADDING_SIZE to 16 to
allow compilation against FFmpeg too (demuxer.c checks the padding
size for packets is at least as much as libavcodec wants for its
decoders, and this check failed with the previous value of 8).
Commit dde8b753e4 merged an mplayer1 change (r31328) that set
correct_pts to true if FPS was not set (on the assumption that
correct-pts mode could provide proper timing without FPS). As the
merge commit noted this change was somewhat questionable, as the
option shouldn't really change after things have already been
initialized. After recent changes it can cause an outright crash
(assert in ds_get_packet2() from 9c7c4e5b7d fails). Remove the hack.
Also only print a warning about not having FPS if correct_pts is not
set (in correct_pts mode not having FPS shouldn't be a real problem,
as everything is based on timestamps anyway).
Convert the list of VOCTLR_* defines to an enum, dropping some unused
values. This resolves a collision between VOCTRL_XOVERLAY_SET_COLORKEY
and VOCTRL_REDRAW_OSD, which had the same value (the XOVERLAY one was
only used by vo_xvr100 and vo_tdfx, so this didn't matter much in
practice).
ad_ffmpeg init() function did not free resources if opening failed.
Outside code (dec_audio.c) does not automatically call uninit() if
init() returns failure, and the uninit function would have crashed in
some cases had it been called (it did freed lavc_context->extradata,
but lavc_context could have been NULL after early init failure). Add
explicit calls to uninit() after failure and make uninit function safe
to call at any point.
At least the libavcodec WavPack decoder can return output for an audio
frame in multiple parts and return 0 bytes input consumed for the
initial parts. Timing info was not set correctly in this case:
sh_audio->pts and pts_bytes were reset each time when decoding more
from the packet, as if the packet had been new (ds_get_packet_pts()
has a check to return MP_NOPTS_VALUE if the packet has already been
partially read, but that didn't trigger since libavcodec returned
exactly 0 bytes read so the demuxer-visible packet state didn't
change).
Add a field to keep track of whether a packet has already been decoded
from, and don't reset timing info again if so. Adding the field
requires adding a decoder context to store it (there wasn't one
before).
BTW the WavPack decoder behavior and avcodec_decode_audio3()
documentation don't match - the documentation says the return value is
"zero if no frame data was decompressed (used) from the input
AVPacket", while the decoder DOES return some frame data which comes
from the input packet.
Commit 6e8d420a41 ("demux: avoid a copy of demux packets with lavf,
reduce padding") was missing an av_dup_packet() line. As a result at
least formats that use parsing on the lavf side could fail (with
parsing the packet may contain pointers to temporary fields that
will be overwritten/freed when reading further packets, and
av_dup_packet() is required to allocate permanent storage).
Use the "--print-errors" flag of pkg-config when testing for the
presence of Libav libraries. Even though the error output is a bit
messy (printed on the same "Checking for"... line), it does contain
useful information for this test which checks for several libraries at
once.
Also change the test name from "FFmpeg" to "Libav" and rename the
option from --disable-ffmpeg to --disable-libav. The change should
cause no compatibility problems as the option is very rarely used.
After 0ece360eea ("demux_mkv: skip files faster in ordered chapter
file search") some Matroska files failed to open. The problem was that
demux_mkv_read_info() returned 0 on success, but the opening code
interpreted this as a value to stop parsing further headers. Fix this
and also modify some of the other return value handling.
wm4
Uoti Urpala
Clément Bœsch
harklu
Martin Herkt