Currently, both bsfs used the same CodedBitstreamContext for reading and
writing; as a consequence, the state of the writer's context at the
beginning of writing a fragment is exactly the state of the reader after
having read the fragment; in particular, the writer might not have
encountered one of its active parameter sets yet.
This is not nice and may lead to invalid output even when the input
is completely spec-compliant: Think of an access unit containing
a primary coded picture referencing a PPS with id id (that is known from
an earlier access unit/from extradata), then a new version of the PPS
with id id and then a redundant coded picture that is also referencing
the PPS with id id. This is spec-compliant, as the standard allows to
overwrite a PPS with a different PPS in between coded pictures and not
only at the beginning of an access unit. In this scenario, the reader
would read the primary coded picture with the old PPS and the redundant
coded picture with the new PPS (as it should); yet the writer would
write both with the new PPS as extradata which might lead to errors or
to invalid data being output without any error (e.g. if the two PPS
differed in redundant_pic_cnt_present_flag).
The above scenario does not directly translate to HEVC as long as one
restricts oneself to input with nuh_layer_id == 0 only (as cbs_h265
does: it currently strips away any NAL unit with nuh_layer_id > 0 when
decomposing); if one doesn't the same issue as above can happen.
If one also allowed input packets to contain more than one access unit,
issues like the above can happen even without redundant coded
pictures/multiple layers.
Therefore this commit uses separate contexts for reader and writer.
Reviewed-by: Mark Thompson <sw@jkqxz.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Several cbs-functions had an unused CodedBitstreamContext parameter.
This commit removes these.
Reviewed-by: Mark Thompson <sw@jkqxz.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
This commit changes h265_metadata to (a) use ff_bsf_get_packet_ref
instead of ff_bsf_get_packet (thereby avoiding one malloc and free per
filtered packet) and (b) to use only one packet structure at all,
thereby avoiding a call to av_packet_copy_props.
(b) has been made possible by the recent changes to ff_cbs_write_packet.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@gmail.com>
Currently, a fragment's unit array is constantly reallocated during
splitting of a packet. This commit changes this: One can keep the units
array by distinguishing between the number of allocated and the number
of valid units in the units array.
The more units a packet is split into, the bigger the benefit.
So MPEG-2 benefits the most; for a video coming from an NTSC-DVD
(usually 32 units per frame) the average cost of cbs_insert_unit (for a
single unit) went down from 6717 decicycles to 450 decicycles (based
upon 10 runs with 4194304 runs each); if each packet consists of only
one unit, it went down from 2425 to 448; for a H.264 video where most
packets contain nine units, it went from 4431 to 450.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@googlemail.com>
* commit 'ce5870a3a8f2b10668ee4f04c2ae0287f66f31b2':
cbs: Refcount all the things!
Some changes for bitstream API.
Merged-by: Mark Thompson <sw@jkqxz.net>
This makes it easier for users of the CBS API to get alloc/free right -
all subelements use the buffer API so that it's clear how to free them.
It also allows eliding some redundant copies: the packet -> fragment copy
disappears after this change if the input packet is refcounted, and more
codec-specific cases are now possible (but not included in this patch).
