This allows to remove checks from ff_xvid_idct_init()
(and also the AVCodecContext* parameter).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
ff_xvid_idct_init() is now only called from ff_idctdsp_init()
and only if idct_algo is FF_IDCT_XVID.
This also implies that it is unnecessary to initalize
the permutation on our own.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is unnecessary: If the dst context is not already initialized,
then it will be initialized by memcpy(dst, src, sizeof(*dst),
which already initializes the IDCT to the desired one, potentially
followed by ff_mpv_common_init(), which does not touch the IDCT.
(This call has been added in f89d76c103;
the aforementioned copying took place back then, too, but
ff_mpv_common_init() reinitialized the IDCT to a non-xvid one,
therefore the initialization here has been added to fix
multithreaded decoding.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
These are sufficiently different to warrant their own functions;
in particular, the earlier code had two callsites for the actual
"write block" function, one for intra and one for inter, yet
the "write block" function nevertheless performed a check
(that the compiler can't optimize away) for whether the current
MB is intra or not.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
While it offers modest speedups compared to the ordinary code,
removing it completely offers even bigger speedups; and given
that these speedups also exist in the ordinary mode, they are
even more important. The no-output code has a 7.8% performance
improvement (based on benchmarking mpeg4_encode_mb()), yet
removing the no-output code resulted in a 9.4% improvement.
Furthermore, the no-output code was broken for the majority of
its existence (until 9207dc3b0d)
and no one complained, so it is likely not used at all.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The MPEG-4 decoder can now initialize ff_mpeg4_rl_intra
directly given that the MPEG-4 encoder no longer wants
it performed, too.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
There four cases for the LUT entry: An ordinary entry
or one of three escaping methods. Three of these methods
are only rarely possible --they correspond to the 102
codes of the underlying VLC and so only 102 of 16384 entries
are possible.
The earlier code would nevertheless try them all for every
LUT entry and use the best one; the new code meanwhile only
uses one method (the fallback one (i.e. the worst)) for them all
and only processes the 102 valid entries afterwards.
The implementation used also means that index_run, max_level
and max_run of the RLTable are no longer needed; the earlier
code would initialize said static tables although they are only
used for a short time to initialize something else.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Multiple PutBitContexts are used when encoding partitioned
frames. When there are multiple candidates for macroblock types,
multiple states (namely the state before encoding the current MB,
the best state among the ones already tried and the current one)
need to be kept; duplicates of the PutBitContexts are among this
state. The temporary buffers for them are kept on the stack
in encode_thread() and their size is quite generous (MAX_MB_SIZE
- 3000 bytes). This commit uses tighter bounds, bringing the
size of the pb2 buffer down to 15 bytes.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The reversible VLC tables use a simpler escaping method
than the ordinary VLCs: It does not use max_run, max_level etc.
and therefore one does not need to initialize these at all.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
It is confusing, because the AV_RL32("ASUS") already
returns an endian-independent value, so converting
it via av_le2ne32() makes no real sense: one would need
to transform the native value to le and write it as
a natie endian uint32_t for it to make sense (the current
code only works because le2ne32 and ne2le32 are the same
for both endianness that we care about). Or one can just
use AV_RL32 and create the number via MKTAG().
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Removes implicit checks for "do I need to output the buffer now?".
Codesize with Clang 19 with -O3 decreased from 7136B to 6108B
(although asv2_put_level() is now inlined).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Up until now, the encoder replicated all the border pixels
for incomplete 16x16 macroblocks. In case the available width
or height is <= 8, some of the luma blocks of the MB
do not correspond to actual input, so that we should encode
them using the least amount of bits. Zeroing the block coefficients
(as this commit does) achieves this, replicating the pixels
and performing an FDCT does not.
This commit also removes the frame copying code for insufficiently
aligned dimensions.
The vsynth3-asv[12] FATE tests use a 34x34 input file and are
therefore affected by this. As the ref updates show, the size
and checksum of the encoded changes, yet the decoded output
stays the same.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Add three missing requirements on bitstream conformance from 7.4.3.19 of
H.266 (V3). Issue found using fuzzing.
Signed-off-by: Frank Plowman <post@frankplowman.com>
When called for palette-predicted CUs, boundary_strength could cause
undefined behaviour due to accessing uninitialised motion information.
The spec doesn't include this, but in the reference software it seems
the deblock strength is always set to 0 for palette CUs due to some
implementation details: perhaps this is a spec issue?
Signed-off-by: Frank Plowman <post@frankplowman.com>
In d5dbcc00d8, it was hoped that detection
of subpicture overlaps could be performed at the tile level, so as to
avoid introducing per-CTU checks. Unfortunately since that patch,
fuzzing has indicated there are some structures involving
pps_subpic_one_or_more_tiles_slice where tile-level checking is not
sufficient. Performing the check at the CTU level should (touch wood)
be the be-all and and-all of this, as CTUs are the lowest common
denominator of the picture partitioning.
Signed-off-by: Frank Plowman <post@frankplowman.com>
The issue is that state->cur[] is 8-bits, but a+b+1 can overflow
before being clipped to 0xF in the following line, causing an incorrect
state to be saved for the next symbol.
This solves numerous bitstream desyncs, particularly when coefficients
with magnitude greater than 127 are sent.
The buffer references may not be writable at this point, as the decoder
calls get_buffer2() with the AV_GET_BUFFER_FLAG_REF flag.
Fixes races as reported by tsan, producing correct output regardless of
threading choices.
Signed-off-by: James Almer <jamrial@gmail.com>
The issue is that there is an explicit lack of synchronization as only the very
first invocation writes symbols and updates the state, which other invocations
then store.
- Query GPU caps for B-frame support and warn if unsupported.
- Make `-max_b_frames` optional
- Drop explicit `-pa_lookahead_buffer_depth` requirement in
adaptive mode.
This allows the compiler to remove the implicit "Do I need to output
the PutBitContext buffer here?" checks.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The earlier code allowed callers to use arbitrary values as
symbols_size as long as no symbols were present.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This turned out to be very beneficial: For GCC 13, the codesize
of ac3_output_frame_header went down from 4522B to 1247B and
from 10762B to 9298B for eac3_output_frame_header. For Clang 17,
the numbers went down from 3923B to 2477B and from 8338B to 6548B
(always with -O3).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This will allow the compiler to optimize the "is the cache full?"
branches away from some put_bits().
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes a Clang warning when asserts are disabled:
"variable 'quant' is used uninitialized whenever switch default is taken
[-Wsometimes-uninitialized]"
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Use av_frame_move_ref() instead of av_frame_ref().
This allows to remove the separate variable for whether
we have already returned the delayed last pic. It also
makes stream looping work when looping multiple times;
previously the delayed pic was only output the first time,
because last_frame_output was never reset.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
