| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
| |
Reader can skip data at the beginning of the file. We were not
informing decsrt how much was skipped when pts_to_start caused the skip.
Fixes https://forum.handbrake.fr/viewtopic.php?f=11&t=36258
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This very small error snowballs into a crash in x264 :-p
If the amount of jitter on the first frame in the queue was small
(about 1 tick) then jitter would not be removed from that frame.
This extra tick of jitter can appear on different frames depending
on when frame arrives and how much has been queued. This very small
amount of randomness lead to problems in the VFR filter. A frame
duration difference as small as 1 tick can lead to an extra frame
getting duplicated when doing CFR. When doing 2 pass encoding, this
extra frame causes x264 to crash at the end of the 2nd pass.
|
|
|
|
|
|
| |
If a stream is delayed by a large amount and the first timestamp from
the stream is AV_NOPTS_VALUE, sync assumed a 0 timestamp which caused
loss of sync. Drop the buffer instead.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Delayed subtitles were causing incorrect muxing in mkv. The mkv muxer
writes chunks where all samples should be relative to a chunk's base
timestamp. When the subtitle is delayed long enough for a new chunk to
start before it gets muxed, the calculated offset to the chunk's base
time is negative (which is illegal).
Note that this is still a possibility with subtitles that must be
delayed (e.g. CC and VOBSUB) because the duration is not known until
the next subtitle's start time is known. The only fix for this would be
to add a special subtitle parsing pass that caches subtitle timestamps
before the main encoding pass is performed.
|
| |
|
|
|
|
|
|
| |
reader adjusts pts_to_start after seeking. if the adjustment makes
pts_to_start == 0, sync didn't properly search for the start point and
hung.
|
|
|
|
|
| |
Continue processing input queues until none are full after p-to-p end
time is reached.
|
| |
|
|
|
|
|
|
| |
When a stream is finished, we need to see if there were any other
streams that were pending. The other streams could be blocking on a
condition variable and need to be unstuck.
|
|
|
|
| |
Fixes https://github.com/HandBrake/HandBrake/issues/328
|
|
|
|
| |
an hb_list_t was not freed at the end of an encode
|
|
|
|
|
| |
wake up potentially waiting sync threads when p-to-p end point is
reached.
|
|
|
|
| |
The output buffer size was not set correctly after resampling
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
When doing point-to-point encoding, subtitles can cause a long
delay in finishing the job when the stop point is reached. This
is due to the sparse nature of subtitles. We may not even see
any additional subtitle till we reach the end of the file.
So when all audio and video streams have reached the end point,
force the termination of all subtitle streams by pushing an
end-of-stream buffer into each subtitles input fifo.
This will cause each subtitle sync worker to wake and return
HB_WORK_DONE.
|
|
|
|
| |
A full input queue could cause the search to stall
|
| |
|
|
|
|
|
|
|
|
|
| |
After finding the start position, some data prior to the start from
other streams could leak through causing duplicate timestamps in the
output.
Also, improves alignment of stop times of all streams when a stop time
is set.
|
|
|
|
|
|
| |
It was dropping subtitles because the "end of CC" marker buffer can have
the same time as the next valid CC which triggered the subtitle overlap
dropping code.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
* sync: correct timestamp discontinuities in sync instead of reader
This patch passes discontinuity information through the pipeline till it
reaches sync.c. The timestamps are passed through the pipeline as read
and unmodified to sync.c (instead of attempting to correct
discontinuities in reader). In sync, when we see a discontinuity,
we know where the next timestamp should be based on the timestamp
and duration of the previous buffer (before the discontinuity). So
we calculate an "SCR" offset based on the timestamp after the
discontinuity and what we calculate it should be.
The old discontinuity handling code was broken due to the following.
The MPEG STD timing model relies heavily on the decoder having an STC
that is phase lock looped to the PCRs in the stream. When decoding a
broadcast stream, the decoder can count on the time measure between PCRs
using the STC to match to a high degree of accuracy.
I.e. STC - lastSTC == PCR - lastPCR. When a discontinuity occurs, the
decoder calculates a new PCR offset = PCR - STC. I.e. the offset is the
new PCR value minus what it would have been if there had been no
discontinuity.
The above does not work without a reliable STC, which we do not have.
We have been attempting to approximate one by avereraging the duration
of received packets and extrapolating an "STC" based on the last PTS and
the average packet duration. But this is highly variable and
unreliable.
* decavcodec: fix data type of next_pts
It needs to be double so that partial ticks are not lost
* deccc608sub: clarify comment
* sync: allow queueing more audio
Audio is small, and there is often a significant amount of audio in the
stream before the first video frame.
* sync: improve handling of damaged streams
When data is missing, the audio decoder was extrapolating timestamps
from the last pts before the error caused by the missing data which
caused sync issues.
Also, missing data can cause the video decoder to output a frame out of
order with the wrong scr sequence. Drop such frames.
|
|
|
|
|
|
| |
essentially an off-by-one error. OutputBuffer had to wait for one more
buffer before any output was performed after the queue should have
already been filled to it's minimum levels.
|
| |
|
| |
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
| |
It's not strictly necessary because it gets done elsewhere as well. But
putting it here makes the code more understandable.
|
| |
|
| |
|
|
|
|
|
| |
... by allowing a deeper initial buffer when looking for the fist PTS of
each stream.
|
|
|
|
| |
... at log level 11 ;)
|
| |
|
|
|
|
|
|
|
| |
We were dropping all buffers before the start frame was found regardless
of the buffers start time. Now we keep track of the start time of the
last video frame seen and only drop buffers that start before that
frame.
|
| |
|
| |
|
| |
|
| |
|
|
|
|
|
|
|
|
| |
since sync interleaves it's output by PTS, the stream of the incoming
buffer is mostly not the same as the stream of the outgoing buffer. This
causes a delay in the data to get to it's respective fifo until the sync
work function for the output stream is called next. Writing directly to
the output fifo fixes this.
|
| |
|
| |
|
| |
|
|
|
|
|
| |
If the entire stream fits in the sync queues, the first PTS was not
detected and initial offsets were not applied.
|
|
|
|
|
| |
The way the constant is defined requires an (int64_t) cast to force it
to be signed.
|
| |
|
|
|
|
|
|
| |
When more than 2 subtitles overlapped in time, they were not merged
properly and could result in cases where the subtitle time went
backwards
|
| |
|
| |
|
|
|
|
|
| |
If there are no more subtitles in a subtitle track after the stop time
of a p-to-p encoding, the encode would never finish.
|
|
|
|
|
|
| |
subtitle stop time was getting incorrectly squashed when special
value AV_NOPTS_VALUE. it should have been set to the next subtitles
start time.
|