Add an optional filter_line3 to the available optimisations.
filter_line3 is equivalent to filter_line, memcpy, filter_line
filter_line shares quite a number of loads and some calculations in
common with its next iteration and testing shows that using aarch64
neon filter_line3s performance is 30% better than two filter_lines
and a memcpy.
Adds a test for vf_bwdif filter_line3 to checkasm
Rounds job start lines down to a multiple of 4. This means that if
filter_line3 exists then filter_line will not sometimes be called
once at the end of a slice depending on thread count. The final slice
may do up to 3 extra lines but filter_edge is faster than filter_line
so it is unlikely to create any noticable thread load variation.
Signed-off-by: John Cox <jc@kynesim.co.uk>
Signed-off-by: Martin Storsjö <martin@martin.st>
Exports C filter_line needed for tail fixup of neon code
Adds neon for filter_line
Signed-off-by: John Cox <jc@kynesim.co.uk>
Signed-off-by: Martin Storsjö <martin@martin.st>
Adds clip and spatial macros for aarch64 neon
Exports C filter_edge needed for tail fixup of neon code
Adds neon for filter_edge
Signed-off-by: John Cox <jc@kynesim.co.uk>
Signed-off-by: Martin Storsjö <martin@martin.st>
Adds an outline for aarch neon functions
Adds common macros and consts for aarch64 neon
Exports C filter_intra needed for tail fixup of neon code
Adds neon for filter_intra
Signed-off-by: John Cox <jc@kynesim.co.uk>
Signed-off-by: Martin Storsjö <martin@martin.st>
The HDR metadata should be removed after HDR to SDR conversion,
otherwise the output frame still has HDR side data.
Signed-off-by: Haihao Xiang <haihao.xiang@intel.com>
This makes the filter output match that of the C version.
It was left intentionally while we figured out if it was better
or not, and while it makes certain samples better, it makes static
samples jump around slightly.
The discrepancy between the definition and the declaration
in allfilters.c is actually UB.
Reviewed-by: James Almer <jamrial@gmail.com>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
The old logic was trying to be excessively clever in "deducing" that the
user wanted to stretch/scale the image when ow/oh differed from iw/ih
aspect ratio. But this is almost surely unintended except in
pathological cases, and in those cases users should simply disable
normalize_sar and do all the stretching/scaling logic themselves. This
is especially important in multi-input mode, where the canvas may be
vastly different from the input dimensions of any stream. Also, passing
through input 0 SAR in multi-input mode is arbitrary and nearly useless,
so again force output SAR to 1:1 here.
Use the gcd of all input timebases to ensure PTS accuracy. For the
framerate, just pick the highest of all the inputs, under the assumption
that we will render frames with approximately this frequency. Of course,
this is not 100% accurate, in particular if the input frames are badly
misaligned. But this field is informational to begin with.
Importantly, it covers the "common" case of combining high FPS and low
FPS streams with aligned frames.
In the event that some frame mixes are OK while others are not, the
priority goes:
1. Errors in updating any frame -> return error
2. Any input incomplete -> request frames and return
3. Any inputs OK -> ignore EOF streams and render remaining inputs
4. No inputs OK -> set output to most recent status
This logic ensures that we can continue rendering the remaining streams,
no matter which streams reach their end of life, until we have no
streams left at which point we forward the last EOF.
When combining multiple inputs, the output PTS may be less than the PTS
of the input. In this case, the current's code assumption of always
draining one value from the FIFO is incorrect. Replace by a smarter
function which drains only those PTS values that were actually consumed.
When combining multiple inputs with different PTS and durations, in
input-timed mode, we emit one output frame for every input frame PTS,
from *any* input. So when combining a low FPS stream with a high FPS
stream, the output framerate would match the higher FPS, independent of
which order they are specified in.
Subsequent inputs require frame blending to be enabled, in order to not
overwrite the existing frame contents.
For output metadata, we implicitly copy the metadata of the *first*
available stream (falling back to the second stream if the first has
already reached EOF, and so on). This is done to resolve any conflicts
between inputs with differing metadata. So when e.g. input 1 is HDR and
output 2 is SDR, the output will be HDR, and vice versa. This logic
could probablly be improved by dynamically determining some "superior"
set of metadata, but I don't want to handle that complexity in this
series.
Instead of finding the ref frame in output_frame() and then passing its
signature to update_crops(), pull out the logic and invoke it a second
time inside update_crops().
This may seem wasteful at present, but will actually become required in
the future, since update_crops() runs on *every* input, and needs values
specific to that input (which the signature isn't), while output_frame()
is only interested in a single input. It's much easier to just split the
logic cleanly.
Including the queue status, because these will need to be re-queried
inside output_frame_mix when that function is refactored to handle
multiple inputs.
In anticipation of a refactor which will enable multiple input support.
Note: the renderer is also input-specific because it maintains a frame
cache, HDR peak detection state and mixing cache, all of which are tied
to a specific input stream.
If the input queue is EOF, then the s->status check should already have
covered it, and prevented the code from getting this far.
If we still hit this case for some reason, it's probably a bug. Better
to hit the AVERROR_BUG branch.
We will postpone the vpp session initialization to when input and output
frames are ready, this copy of the sequence parameters will be used to
initialize vpp session.
Signed-off-by: Haihao Xiang <haihao.xiang@intel.com>
John Cox
Paul B Mahol
Haihao Xiang
Lynne
Andreas Rheinhardt
Fei Wang
Niklas Haas
yethie