In order to support OSD redrawing for vo_xv and vo_x11, draw_bmp.c
included an awkward "backup" mechanism to copy and restore image
regions that have been changed by OSD/subtitles.
Replace this by a much simpler mechanism: keep a reference to the
original image, and use that to restore the Xv/X framebuffers.
In the worst case, this may increase cache pressure and memory usage,
even if no OSD or subtitles are rendered. In practice, it seems to be
always faster.
It's not easy to tell whether the OSD/subs are empty, or if something is
drawn. In general you have to use osd_draw() with a custom callback. If
nothing is visible, the callback is never invoked. (The actual reason
why this is so "hard" is the implementation of osd_libass.c, which
doesn't allow separating rendering and drawing of OSD elements, because
all OSD elements share the same ASS_Renderer.)
To simplify avoiding copies, make osd_draw_on_image() instead of the
caller use mp_image_make_writeable(). Introduce osd_draw_on_image_p(),
which works like osd_draw_on_image(), but gets the new image allocation
from an image pool. This is supposed to be an optimization, because it
reduces the frequency of large allocations/deallocations for image data.
The result of this is that the frequency of copies needed in conjunction
with vf_sub, screenshots, and vo_lavc (encoding) should be reduced.
vf_sub now always does true pass-through if no subs are shown.
Drop the pts check from vf_sub. This didn't make much sense.
Remove the strange things the old mp_image_setfmt() code did to the
image format parameters. This includes setting chroma shift to 31 for
gray (Y8) formats and more.
Y8 + vo_opengl_old didn't actually work for unknown reasons (regression
in this branch). Fix this. The difference is that Y8 is now interpreted
as gray RGB (LUMINANCE texture) instead of involving YUV (and levels)
conversion.
Get rid of distinguishing RGB and BGR. There wasn't really any good
reason for this.
Remove mp_get_chroma_shift() and IMGFMT_IS_YUVP16*(). mp_imgfmt_desc
gives the same information and more.
Even though #ifdef ACCURATE is removed, the result should be about the
same. The fallback is only used by packed YUV formats (YUYV, NV12), and
doing 16 bit for them instead of 8 bit is not useful.
A side effect is that Y8 (gray) is not converted drawing subs, and for
alpha formats, the alpha plane is not removed. This means the number of
planes after upsampling can be 1-4 (1: gray, 2: gray+alpha, 3: planar,
4: planar+alpha). The code has to be adjusted accordingly to work on the
color planes only. Also remove the workaround for the chroma shift 31
hack.
mplayer's video chain traditionally used FourCCs for pixel formats. For
example, it used IMGFMT_YV12 for 4:2:0 YUV, which was defined to the
string 'YV12' interpreted as unsigned int. Additionally, it used to
encode information into the numeric values of some formats. The RGB
formats had their bit depth and endian encoded into the least
significant byte. Extended planar formats (420P10 etc.) had chroma
shift, endian, and component bit depth encoded. (This has been removed
in recent commits.)
Replace the FourCC mess with a simple enum. Remove all the redundant
formats like YV12/I420/IYUV. Replace some image format names by
something more intuitive, most importantly IMGFMT_YV12 -> IMGFMT_420P.
Add img_fourcc.h, which contains the old IDs for code that actually uses
FourCCs. Change the way demuxers, that output raw video, identify the
video format: they set either MP_FOURCC_RAWVIDEO or MP_FOURCC_IMGFMT to
request the rawvideo decoder, and sh_video->imgfmt specifies the pixel
format. Like the previous hack, this is supposed to avoid the need for
a complete codecs.cfg entry per format, or other lookup tables. (Note
that the RGB raw video FourCCs mostly rely on ffmpeg's mappings for NUT
raw video, but this is still considered better than adding a raw video
decoder - even if trivial, it would be full of annoying lookup tables.)
The TV code has not been tested.
Some corrective changes regarding endian and other image format flags
creep in.
These functions weren't specific to video filters and were misplaced
in vf.c. Move them to mp_image.c.
Fix the big endian test in vf_mpi_clear/mp_image_clear, which has been
messed up in 74df1d.
According to DOCS/OUTDATED-tech/colorspaces.txt, the following formats
are supposed to be palettized:
IMGFMT_BGR8
IMGFMT_RGB8,
IMGFMT_BGR4_CHAR
IMGFMT_RGB4_CHAR
IMGFMT_BGR4
IMGFMT_RGB4
Of these, only BGR8 and RGB8 are actually treated as palettized in some
way. ffmpeg has only one palettized format (AV_PIX_FMT_PAL8), and
IMGFMT_BGR8 was inconsistently mapped to packed non-palettized RGB
formats too (AV_PIX_FMT_BGR8). Moreover, vf_scale.c contained messy
hacks to generate a palette when AV_PIX_FMT_BGR8 is output. (libswscale
does not support AV_PIX_FMT_PAL8 output in the first place.)
Get rid of all of this, and introduce IMGFMT_PAL8, which directly maps
to AV_PIX_FMT_PAL8. Remove the palette creation code from vf_scale.c.
IMGFMT_BGR8 maps to AV_PIX_FMT_RGB8 (don't ask me why it's swapped),
without any palette use. Enabling it in vo_x11 or using it as vf_scale
input seems to give correct results.
mp_image_alloc_planes() allocated images with minimal stride, even if
the resulting stride was unaligned. It was the responsibility of
vf_get_image() to set an image's width to something larger than
required to get an aligned stride, and then crop it. Always allocate
with aligned strides instead.
Get rid of IMGFMT_IF09 special handling. This format is not used
anymore. (IF09 has 4x4 chroma sub-sampling, and that is what it was
mainly used for - this is still supported.) Get rid of swapped chroma
plane allocation. This is not used anywhere, and VOs like vo_xv,
vo_direct3d and vo_sdl do their own swapping.
Always round chroma width/height up instead of down. Consider 4:2:0 and
an uneven image size. For luma, the size was left uneven, and the chroma
size was rounded down. This doesn't make sense, because chroma would be
missing for the bottom/right border.
Remove mp_image_new_empty() and mp_image_alloc_planes(), they were not
used anymore, except in draw_bmp.c. (It's still allowed to setup
mp_images manually, you just can't allocate image data with them
anymore - this is also done in draw_bmp.c.)
Replace the internal pixel format stuff with code that queries the
libavutil list of pixel format descriptors.
Trying to map IMGFMT_IS_RGB() etc. turned out extremely hacky.
Instead of using a callback to "capture" the image next time the filter
function is called, do it the other way around: on every filter
invocation, create a reference to the image, and return it if a
screenshot is requested. This also fixes the 1-frame delay when taking
screenshots with the filter.
This also allows simplifying screenshot.c.
VFCAP_TIMER disables any additional waiting done by mpv in the
playloop. Remove VFCAP_TIMER, but re-use the idea for vo_image and
vo_lavc.
This means --untimed doesn't have to be passed when using --vo=image.
This allows avoiding a copy.
The logic about using the AVFrame.reference field if buffer hints are
not available is similar to mplayer-svn's direct rendering code. This is
needed because h264 decoding doesn't set buffer hints.
<wm4> michaelni: couldn't it set buffer hints from AVFrame.reference then?
<michaelni> i see no reason ATM why that wouldnt be possible
OK...
Remove mp_image.width/height. The w/h members are the ones to use.
width/height were used internally by vf_get_image(), and sometimes for
other purposes.
Remove some image flags, most of which are now useless or completely
unused. This includes VFCAP_ACCEPT_STRIDE: the vf_expand insertion in
vf.c does nothing.
Remove some other unused mp_image fields.
Some rather messy changes in vo_opengl[_old] to get rid of legacy
mp_image flags and fields. This is left from when vo_gl supported DR.
Change the entire filter API to use reference counted images instead
of vf_get_image().
Remove filter "direct rendering". This was useful for vf_expand and (in
rare cases) vf_sub: DR allowed these filters to pass a cropped image to
the filters before them. Then, on filtering, the image was "uncropped",
so that black bars could be added around the image without copying. This
means that in some cases, vf_expand will be slower (-vf gradfun,expand
for example).
Note that another form of DR used for in-place filters has been replaced
by simpler logic. Instead of trying to do DR, filters can check if the
image is writeable (with mp_image_is_writeable()), and do true in-place
if that's the case. This affects filters like vf_gradfun and vf_sub.
Everything has to support strides now. If something doesn't, making a
copy of the image data is required.
The code was entirely correct, as the VO doesn't report VFCAP_ACCEPT_STRIDE
in query_format. Add stride capability in preparation for changing the
video chain: soon all VOs will have to support arbitrary strides.
The code assumed mp_image_alloc() would allocate an image large enough
for corevideo's stride, which doesn't have to be the case. If
corevideo's stride was larger than the stride of mp_image, the memcpy()
would write beyond the mp_image allocation.
This probably didn't actually happen, but fix the code to be more
correct anyway.
Setting the size of a mp_image must be done with mp_image_set_size()
now. Do this to guarantee that the redundant fields (like chroma_width)
are updated consistently. Replacing the redundant fields by function
calls would probably be better, but there are too many uses of them,
and is a bit less convenient.
Most code actually called mp_image_setfmt(), which did this as well.
This commit just makes things a bit more explicit.
Warning: the video filter chain still sets up mp_images manually,
and vf_get_image() is not updated.
Refcounting will conceptually allocate and free images all the time
when using the filter chain. Add a pool that makes these reallocations
cheap.
This only affects the image data, not mp_image structs and similar small
allocations. Small allocations are always fast with reasonable memory
managers, while large image data will trigger mmap/munmap calls each
time.
Note that if the codec doesn't support DR1, the image has to be copied.
There is no other way to guarantee that the image will be valid after
decoding the next image.
The only important codec that doesn't support DR1 yet is rawvideo. It's
likely that ffmpeg/Libav will fix this at some time. For now, this
decoder uses an evil hack and puts pointers to the packet data into the
returned frame. This means the image will actually get invalid as soon
as the corresponding video packet is free'd, so copying the image is the
only reasonable thing anyway.
Replace libavcodec's native buffer allocation with code taken from
ffplay/ffmpeg's libavfilter support. The code in lavc_dr1.c is directly
copied from cmdutils.c. Note that this is quite arcane code, which
contains some workarounds for decoder bugs and the like. This is not
really a maintainance burden, since fixes from ffmpeg can be directly
applied to the code in lavc_dr1.c.
It's unknown why libavcodec doesn't provide such a function directly.
avcodec_default_get_buffer() can't be reused for various reasons.
There's some hope that the work known as The Evil Plan [1] will make
custom get_buffer implementations unneeded.
The DR1 support as of this commit does nothing. A future commit will
use it to implement ref-counting for mp_image (similar to how AVFrame
will be ref-counted with The Evil Plan.)
[1] http://lists.libav.org/pipermail/libav-devel/2012-December/039781.html
Deprecate the hardware specific video codec entries (like ffh264vdpau).
Replace them with the --hwdec switch, which requests that a specific
hardware decoding API should be used. The codecs.conf entries will be
removed at a later time, but for now they are useful for testing and
compatibility.
Instead of --vc=ffh264vdpau, --hwdec=vdpau should be used.
Add a fallback if hardware decoding fails. Most hardware decoders
(including vdpau) support only a subset of h264, and having such a
fallback is supposed to enable a better user experience.
This was buggy and didn't even work in the simplest cases. It was
disabled when multithreading was used, and always disabled for h264.
A better alternative (reference counting) will be added later.
Hardware decoding still uses the ffmpeg DR mechanism, but has been
decoupled from mpv's DR in the previous commit.
vdpau hardware decoding used the DR (direct rendering) path to let the
decoder query a surface from the VO. Special-case the HW decoding path
instead, to make it separate from DR.
This mutated the variable for the thread count option
(lavc_param->threads) on decoder initialization. This didn't have any
practical relevance, unless formats supporting hardware video decoding
and other formats were played in the same mpv instance. In this case,
hardware decoding would set threads to 1, and all files played after
that would use only one thread as well even with software decoding.
Remove XvMC leftover (CODEC_CAP_HWACCEL).
Simplify the decoder pixel format handling by making it handle only
the case vd_lavc needs: a video stream always decodes to a single
pixel format.
Remove the handling for multiple pixel formats, and remove the
codecs.conf pixel format declarations that are left.
Remove the handling of "ambiguous" pixel formats like YV12 vs. I420 (via
VDCTRL_QUERY_FORMAT etc.). This is only a problem if the video chain
supports I420, but not YV12, which doesn't seem to be the case anywhere,
and in fact would not have any advantage.
Make the "flip" flag a global per-codec flag, rather than a pixel format
specific flag. (Some ffmpeg decoders still return a flipped image, so
this has to be done manually.) Also fix handling of the flip operation:
do not overwrite the global flip option, and make the --flip option
invert the codec flip option rather than overriding it.
Slices allowed filtering or drawing video in horizontal bands or
blocks. This allowed working on the video in smaller units. In theory,
this could bring a performance win by lowering cache pressure, as you
didn't have to keep the whole video frame in cache while filtering,
only the slice.
In practice, the slice code path was barely used for the following
reasons:
- Multithreaded decoding with ffmpeg didn't use slices. The ffmpeg
slice callback was disabled, because it can be called from another
thread, and the mplayer video chain is not thread-safe.
- There was nothing that would turn "full" images into appropriate
slices, so slices were rarely used.
- Most filters didn't actually support slices.
On the other hand, supporting slices lead to code duplication and more
complex code in general. I made some experiments and didn't find any
actual measurable performance improvements when using slices. Even
ffmpeg removed slices based filtering from libavfilter in favor of
simpler code.
The most broken thing about the slices code path is that slices can't
be queued, like it is done for images in vo.c.
For some reason, libavcodec abuses the slices rendering code path for
hardware decoding: in that case, the only purpose of the draw callback
is to pass a vdpau video surface object to video output. (It is unclear
to me why this had to use the slices code, instead of just returning an
AVFrame with the required vdpau state.)
Make this code separate within mpv, so that the internal slices code
path is not used for hardware decoding. Pass the vdpau state with
VOCTRL_HWDEC_DECODER_RENDER instead.
Remove the mencoder specific VOCTRLs.
Remove VOCTRL_DRAW_IMAGE and always set vo_driver.draw_image in VOs.
Make draw_image mandatory: change some VOs (like vo_x11) to support it,
and remove the image-to-slices fallback in vf_vo.
Remove vo_driver.is_new. This member indicated whether draw_image is
supported unconditionally, which is now always the case.
draw_image_pts is a hack until the video filter chain is changed to
include the PTS as field in mp_image. Then vo_vdpau and vo_lavc will
be changed to use draw_image.
This allowed to move the input stream layer across the network, allowing
the user to play anything that mplayer could play remotely. For example,
playing a DVD related on a remote server (say, with the host name
"remotehost1") could be done by starting the netstream server on that
remote server, and then running:
mplayer mpst://remotehost1/dvd://
This would open the DVD on the remote host, and transfer the raw DVD
sector reads over network. It works the same for other protocols, and
all accesses to the stream layer are marshaled over network. It's
comparable to the way the cache layer (--cache) works.
It has questionable use and most likely was barely used at all. There's
lots of potential for breakage, because it doesn't translate the stream
CTRLs to network packets. Just get rid of it.
The server used to be in TOOLS/netstream.c, and was accidentally removed
earlier.
This function sucks and apparently is not very portable (at least on
mingw, the configure check fails). Also remove the emulation of that
function from osdep/strsep*, and remove the configure check.
I have no idea when or how this broke, but _wstati64() is the function
we want anyway (64 bit filesize). Possibly this was a mingw-w64 bug.
It's unknown why "wstat()" just doesn't work in this case, as it's not
defined by MSDN and could be defined by mingw as it needs.
vsscanf() is in POSIX, C99, mingw, etc. Further, the implementation in
osdep/vsscanf.c was completely broken, and if it worked, it worked only
by chance.
The check determined whether the argument for .align is in bytes, or
log2(bytes). Apparently it's always in bytes for ELF i386 systems, and
this check is used for x86 inline assembler only. Even if this
assumption should be wrong, it likely won't cause much damage: the
existing code uses it only in the form ".align 4", which means in the
worst case it will try to align to 16 bytes, which doesn't cause any
problems (unless the object file format does not support such a high
alignment).
Update the filters that used this.
Quoting the GNU as manual:
For other systems, including ppc, i386 using a.out format, arm and
strongarm, it is the number of low-order zero bits the location counter
must have after advancement. For example `.align 3' advances the
location counter until it a multiple of 8. If the location counter is
already a multiple of 8, no change is needed.
Change the only usage of HAVE_BUILTIN_EXPECT, demux.h, to use an #ifdef
instead. In theory, a configure check is better, but nobody does it this
way anyway, and we seek to reduce the configure script.
mixer_setvolume() accepts float values for volume, but used the
integer function av_clip() to limit range, losing the fractional part
as a side effect. Change the code to use av_clipf() instead. For most
uses this shouldn't make any real difference; actual AO volume
settings may not have that much precision anyway.