Change ao_pcm to use the new audio output driver API and clean up some
of the code. Rewrite the logic controlling how playback timing works
when using -ao pcm. Deprecate the "fast" suboption; its only effect
now is to print a warning, but it's still accepted so that specifying
it is not an error.
Before, timing with -ao pcm and video enabled had two possible
modes. In the default mode playback speed was rather arbitrary - not
realtime, but not particularly fast. -ao pcm:fast tried to play back
at maximum video playback speed - mostly succeeding, but not quite
guaranteed to work in all cases. Now the default is to play at
realtime speed. The -benchmark option can now be used to get faster
playback (same as the video-only case). In the audio-only case
playback is always maximum speed.
Add a new audio output driver API. The main change is the addition of
a context struct where data can be kept instead of using globals. The
old API remains available. This commit does not yet convert any driver
to use the new API.
Reindent audio_out.c and audio_out.h. Also remove trailing '_s' from
two struct names (which are not currently used anywhere) and make the
audio_out_drivers[] table static.
Neither fd 0 slave input (-slave) nor additional opened fds (-input
file=X) were set to nonblocking mode as they should have been. Fix.
Also rename the horribly generic USE_SELECT #define used for a
specific slave input detail.
Make af_format clamp float sample values to the range [-1, 1] before
conversion to integer types. Before any out-of-range values wrapped
around and caused nasty artifacts. This filter is used for all
automatic format conversions; thus any decoder that outputs floats
with possible out-of-range values would have been affected by the bad
conversion if its output needed to be converted to integers for AO.
dec_audio.c init_audio_codec() would in one case print
"ADecoder init failed :(\n" and return failure. Its only caller
init_best_audio_codec() printed exactly the same message if the
returned result was failure. Change the latter message to say
"Could not open audio decoder %s.\n" instead. Some of the
per-open-attempt messages are kind of value about their context; this
new message should make it more clear where the attempt to open one
specific codec ends.
The init() method in ad_ffmpeg tries to decode some audio data after
opening the libavcodec decoder; however the method returned success
even if this part failed. Change it to return failure instead,
indicating that the codec could not be successfully opened.
This improves behavior at least with some AAC files, for which the
libavcodec decoder can be successfully initialized but decoding
packets always fails. Before the audio would be decoded with
libavcodec, producing only a constant stream of errors; after this
commit audio decoder initialization falls back to FAAD (if available)
which works for these samples.
vo_svga.c used its static query_format() function before defining it,
but had no prototype before the use. Thus is depended on the otherwise
useless static declaration in a shared header, and broke after that
was removed. Fix by adding an advance declaration inside the file.
Instead of strictly limiting the number of total entries in the
internal fifo, make the overall buffer bigger and try to limit entries
based on how many bound commands they're expected to generate. Now
doubleclick and button down events aren't counted for that limit.
Normally the sequence down-doubleclick-up generates at most one
command, so this better matches the quantity we actually want to
limit. Also add a mechanism to clear the button combination state kept
by input.c when the fifo is full; this avoids "stuck button" problems
due to button release events being dropped.
The key combination state clearing is partially based on MPlayer 1
changes by Reimar Döffinger (though overall the effects of this commit
are quite different). It still doesn't make "stuck button" problems
completely impossible; at least if the VO gets closed while a button
was down then nothing will send a button up event or reset state.
The code combining button presses into multibutton commands prevented
single click commands from triggering if a doubleclick event had been
generated from the same button press. As a result using the mouse
wheel to seek worked very badly. Special-case doubleclick events in
the event interpretation code to avoid this issue. This changes the
sequence of generated "keys" for press-release-press-release from
MOUSE_BTN0 MOUSE_BTN0-MOUSE_BTN0_DBL MOUSE_BTN0_DBL to
MOUSE_BTN0 MOUSE_BTN0_DBL MOUSE_BTN0.
"Keys" like MOUSE_BTN0-MOUSE_BTN0_DBL will never be generated now; any
existing configuration files using those need to be changed.
Move the definitions of all special key codes (those not passed by
ASCII value) to input/keycodes.h. Before they were spread between
osdep/keycodes.h, input/joystick.h, input/mouse.h and input/ar.h, plus
some special values in input.h. This was especially inconvenient as
the codes had to be coordinated to not conflict between the files.
The change requires a bit of ugliness as appleir.c includes
<linux/input.h> which contains various conflicting KEY_* definitions.
Work around this by adding a special preprocessor variable which can
be used to avoid defining these in keycodes.h.
Setup default speex modes, allows decoding of speex in flv which does
not contain a header packet.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33327 b3059339-0415-0410-9bf9-f77b7e298cf2
Simplify by using FFMAX3.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33300 b3059339-0415-0410-9bf9-f77b7e298cf2
Simplify colormap generation code, avoid some mallocs and add frees
for one error case.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33302 b3059339-0415-0410-9bf9-f77b7e298cf2
Use memcpy_pic instead of reimplementing it.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33303 b3059339-0415-0410-9bf9-f77b7e298cf2
Allow DR to work with reget_buffer when no buffer_hints are set.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33287 b3059339-0415-0410-9bf9-f77b7e298cf2
Allow reget_buffer to somewhat work after direct rendering failure.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33286 b3059339-0415-0410-9bf9-f77b7e298cf2
Do not try to set up the FS segment when using quicktime to decode,
it makes no sense to do that.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33259 b3059339-0415-0410-9bf9-f77b7e298cf2
Fix artefacts at right border of scaled bitmap subtitles by clearing
the destination to 0 first if the width is not a multiple of 8.
This is necessary because some of the optimized sub drawing functions
always draw pixels in multiples of 8, so we cannot rely on any "dirt"
in the "unused" borders being ignored.
Note that due to the way rendering works, both alpha and the image need
to be cleared.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33230 b3059339-0415-0410-9bf9-f77b7e298cf2
Add dummy -shuffle and -noshuffle options so they appear in
-list-options output. These features are not implemented as options;
rather they're features of the command line parser.
Also add the CONF_NOCFG flag to the existing playlist option, which is
a similar placeholder for a command line parser feature.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33147 b3059339-0415-0410-9bf9-f77b7e298cf2
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33148 b3059339-0415-0410-9bf9-f77b7e298cf2
Ensure we always pass a buffer of at least sector size to the read
function. This is never an issue with streams that have actual
sectors, as the reads will always return a multiple of sector size and
the cache is always used in blocks of sector size. However the rtp
protocol misuses this so it can just assume it always has a
sufficiently large buffer available and thus fails without this extra
hack.
git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@33120 b3059339-0415-0410-9bf9-f77b7e298cf2
The input code read at most one event per input or command fd. If this
event was not bound to any recognized command then no command was
returned, and higher-level code could not distinguish this case from
there being no pending events left. As a result unbound events would
cause extra latency in event processing. Change the input code to
continue reading events until it either finds one that maps to a
command or hits EOF/error.
getch2.c did not make stdin non-blocking, and relied on only being
called after select() had shown readability. Stop relying on that
assumption and set stdin to non-blocking mode. Hopefully no relevant
platform has problems with this...
Commit a2d28010cc ("cleanup: find_subfiles.c: simplify (mainly using
bstr)") was missing a bstrdup() in subtitle search code, and thus the
code erroneously modified the original filename string passed in. As a
result anything which printed or otherwise used the filename after
that would use a lowercase-converted version instead of the actual
file name. Fix by adding the bstrdup() to operate on a local copy of
the name instead.
Many video filters failed to calculate or even just pass through pts
values for their output frames. Fix this, and also make the two
remaining filters that called vf_next_put_image() twice for the same
input frame (vf_softpulldown, vf_telecine) use vf_queue_frame() so
that e.g. framestepping properly sees both frames.
Changed filters: vf_bmovl, vf_detc, vf_divtc, vf_filmdint, vf_ivtc,
vf_lavc, vf_phase, vf_pullup, vf_softpulldown, vf_telecine, vf_tile,
vf_tinterlace.
Change BSTR() from a macro producing a compound literal to an inline
function returning the same value. This works for all existing uses,
and avoids a warning from BSTR(NULL) (the macro expansion contained
strlen(NULL); this was valid code because the strlen call was never
evaluated, but still triggered a GCC warning).
The input loop select() call was only run if there was at least one
input file descriptor. However other code uses the input loop to wait;
this could result in the wait becoming a busy loop when running with
-noconsolecontrols (without that there is at least one input fd for
terminal input). Make the input loop call select() to sleep even if
there are no input file descriptors.
Commit cbeed30ae8 ("core: wake up a bit less often for audio-only
files") increased the sleep time between audio buffer fills. This
turned out to cause problems on some machines where available audio
buffer sizes are extremely limited (example cases included 85 ms for
stereo and less for multichannel audio). Change the code to check
the amount of buffered audio and shorten sleep times accordingly if
needed.
Such short buffers violate some assumptions made by video timing code,
so they may still cause visible problems in some cases. At least on
some machines using ALSA the problem seems to be caused by bad
configuration defaults (small buffer memory limit which can be
increased).
Change the behavior of the iconv, freetype, fontconfig and libass
tests when autodetection fails. They now abort instead of silently
creating a crippled build. Users who really want to build without
those features can use explicit --disable flags.
If --enable-3dfx is specified but dga is not available then 3dfx is
disabled nonetheless. However, this disabling is not done properly,
and libvo/vo_3dfx.c is still compiled (but cannot be used). Fix.
The behavior of automatically disabling vo_3dfx despite --enable-3dfx
is itself questionable, but I'm not changing that now.
Update various code to use newer alternatives instead of deprecated
functions/fields that are being dropped at libav API bump. An
exception is avcodec_thread_init() which is being dropped even though
it's still _necessary_ with fairly recent libav versions, so there's
no good alternative which would work with both those recent versions
and latest libavcodec. I think there are grounds to consider the drop
premature and revert it for now; if that doesn't happen I'll add a
version-test #if check around it later.
