Certain types of OBUs are stripped away before muxing into Matroska and
ISOBMFF; there are two functions to do this: One that outputs by
directly writing in an AVIOContext and one that returns a freshly
allocated buffer with the units not stripped away copied into it.
The latter option is bad for performance, especially when the input
does already not contain any of the units intended to be stripped away
(this covers typical remuxing scenarios). Therefore this commit changes
this by avoiding allocating and copying when possible; it is possible if
the OBUs to be retained are consecutively in the input buffer (without
an OBU to be discarded between them). In this case, the caller receives
the offset as well as the length of the part of the buffer that contains
the units to be kept. This also avoids copying when e.g. the only unit
to be discarded is a temporal delimiter at the front.
For a 22.7mb/s file with average framesize 113 kB this improved the time
for the calls to ff_av1_filter_obus_buf() when writing Matroska from
313319 decicycles to 2368 decicycles; for another file with 1.5mb/s
(average framesize 7.3 kB) it improved from 34539 decicycles to 1922
decicyles. For these files the only units that needed to be stripped
away were temporal unit delimiters at the front.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
ff_av1_filter_obus_buf() and ff_avc_parse_nal_units_buf() both have a
pointer-to-pointer parameter which they use to pass a newly allocated
buffer to the caller. And both functions freed what this pointer points to
before overwriting it. But no caller of these functions used this feature,
but some had to initialize the pointer just because of this. So remove
it and update the documentation of ff_av1_filter_obus_buf() wrt this fact.
ff_hevc_annexb2mp4_buf in contrast did not free the pointer. This has been
documented, too.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
Both ISOBMFF as well as Matroska require certain OBUs to be stripped
before muxing them. There are two functions for this purpose; one writes
directly into an AVIOContext, the other returns a freshly allocated
buffer with the undesired units stripped away.
The latter one actually relies on the former by means of a dynamic
buffer. This has several drawbacks: The underlying buffer might have to
be reallocated multiple times; the buffer will eventually be
overallocated; the data will not be directly copied into the final
buffer, but rather first in the write buffer (in chunks of 1024 byte)
and then written in these chunks. Moreover, the API for dynamic buffers
is defective wrt error checking and as a consequence, the earlier code
would indicate a length of -AV_INPUT_BUFFER_PADDING_SIZE on allocation
failure, but it would not return an error; there would also be no error
in case the arbitrary limit of INT_MAX/2 that is currently imposed on
dynamic buffers is hit.
This commit changes this: The buffer is now parsed twice, once to get
the precise length which will then be allocated; and once to actually
write the data.
For a 22.7mb/s file with average framesize 113 kB this improved the time
for the calls to ff_av1_filter_obus_buf() when writing Matroska from
753662 decicycles to 313319 decicycles (based upon 50 runs a 2048 frames
each); for another 1.5mb/s file (with average framesize of 7.3 kB) it
improved from 79270 decicycles to 34539 decicycles (based upon 50 runs a
4096 frames).
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
The output size is already returned via a pointer argument, so there is
no need to return it via the ordinary return value as well. The
rationale behind this is to not poison the return value on success.
It also unifies the behaviour of the *_buf-functions for AVC, AV1 and
HEVC.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
