233 to 105 cycles on Arrandale and Win64.
Replacing the multiplication by s_m[m] by a pand and a pxor with
appropriate vectors is slower. Unrolling is a 15 cycles win.
A SSE version was 4 cycles slower.
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
lavf/gif: correct the delay after the first frame.
To define accurately the delay between two frames, it is necessary to
have both available. Before this commit, the first frame had a delay of
0; while in practice the problem is not visible in most situation, it is
problematic with low frame rate and large scene change.
This commit notably fixes output generated with commands such as:
gif: use only one graphic control extension block per image.
The encoder now doesn't produce any extra graphic control extension
block anymore. Only the image is encoded, and the muxer writing
its own GCE containing notably the timing information now includes the
optional palette transmitted through packet side data.
This commit avoid setting clashes between the two GCE, and reduce the
size of the generated file with pal8 output.
Regression since c10d498bfd246a40cc7830838b74e18a79418839.
Unfortunately ff_thread_get_format can only be called from
a separate decode thread, running it during init will fail.
Fixes for that are welcome, for now just revert back to
calling avctx->get_format directly, which is correct
but having to decide on a case-by-case basis which approach
to use is a bit messy.
Clément Bœsch [Thu, 3 Jan 2013 02:06:43 +0000 (03:06 +0100)]
subtitles: introduce ASS codec id and use it.
Currently, we have a AV_CODEC_ID_SSA, which matches the way the ASS/SSA
markup is muxed in a standalone .ass/.ssa file. This means the AVPacket
data starts with a "Dialogue:" string, followed by a timing information
(start and end of the event as string) and a trailing CRLF after each
line. One packet can contain several lines. We'll refer to this layout
as "SSA" or "SSA lines".
In matroska, this markup is not stored as such: it has no "Dialogue:"
prefix, it contains a ReadOrder field, the timing information is not in
the payload, and it doesn't contain the trailing CRLF. See [1] for more
info. We'll refer to this layout as "ASS".
Since we have only one common codec for both formats, the matroska
demuxer is constructing an AVPacket following the "SSA lines" format.
This causes several problems, so it was decided to change this into
clean ASS packets.
Some insight about what is changed or unchanged in this commit:
CODECS
------
- the decoding process still writes "SSA lines" markup inside the ass
fields of the subtitles rectangles (sub->rects[n]->ass), which is
still the current common way of representing decoded subtitles
markup. It is meant to change later.
- new ASS codec id: AV_CODEC_ID_ASS (which is different from the
legacy AV_CODEC_ID_SSA)
- lavc/assdec: the "ass" decoder is renamed into "ssa" (instead of
"ass") for consistency with the codec id and allows to add a real
ass decoder. This ass decoder receives clean ASS lines (so it starts
with a ReadOrder, is followed by the Layer, etc). We make sure this
is decoded properly in a new ass-line rectangle of the decoded
subtitles (the ssa decoder OTOH is doing a simple straightforward
copy). Using the packet timing instead of data string makes sure the
ass-line now contains the appropriate timing.
- lavc/assenc: just like the ass decoder, the "ssa" encoder is renamed
into "ssa" (instead of "ass") for consistency with the codec id, and
allows to add a real "ass" encoder.
One important thing about this encoder is that it only supports one
ass rectangle: we could have put several dialogue events in the
AVPacket (separated by a \0 for instance) but this would have cause
trouble for the muxer which needs not only the start time, but also
the duration: typically, you have merged events with the same start
time (stored in the AVPacket->pts) but a different duration. At the
moment, only the matroska do the merge with the SSA-line codec.
We will need to make sure all the decoders in the future can't add
more than one rectangle (and only one Dialogue line in it
obviously).
FORMATS
-------
- lavf/assenc: the .ass/.ssa muxer can take both SSA and ASS packets.
In the case of ASS packets as input, it adds the timing based on the
AVPacket pts and duration, and mux it with "Dialogue:", trailing
CRLF, etc.
- lavf/assdec: unchanged; it currently still only outputs SSA-lines
packets.
- lavf/mkv: the demuxer can now output ASS packets without the need of
any "SSA-lines" reconstruction hack. It will become the default at
next libavformat bump, and the SSA support will be dropped from the
demuxer. The muxer can take ASS packets since it's muxed normally,
and still supports the old SSA packets. All the SSA support and
hacks in Matroska code will be dropped at next lavf bump.
lavf/gif: trim unnecessarily long netscape ext code.
Extension description comments are now placed along the avio calls, the
always defined macro removed, and the always true loop_count check as
well (loop option is bound to 0-65535).
This commit removes the badly duplicated code between the encoder and
the muxer. That may sound surprising, but the encoder is now responsible
from the encoding of the picture when muxing to a .gif file. It also
does not require anymore a manual user intervention such as a -pix_fmt
rgb24 to work properly. To summarize, output gif are now easier to
generate, code is saner and simpler, and files are smaller (thanks to
the lzw encoding which was unused so far with the default .gif output).
We can certainly make things even better, but this is the first step.
FATE is updated because of the output being produced by the encoder and
not the muxer (no lzw in the muxer), and in the seek test only the size
mismatches.
Set select->select internally to the function, avoid lossy double->int
conversion causing spurious frame drop (e.g. in case the evaluation
result is positive and less than 1).
projects / ffmpeg / log