The complex vertical low-pass filter slightly over-sharpens the picture. This becomes visible when several transcodings are cascaded and the error potentises, e.g. some generations of HD->SD SD->HD.
To prevent this behaviour the destination pixel must not exceed the source pixel when the average of the pixels above and below is less than the source pixel. And the other way around.
Tested and approved in a visual transcoding cascade test by video professionals.
SSIM/PSNR test with the first generation of an HD->SD file as a reference against the 6th generation(3 x SD->HD HD->SD):
Results without the patch:
SSIM Y:0.956508 (13.615881) U:0.991601 (20.757750) V:0.993004 (21.551382) All:0.974405 (15.918463)
PSNR y:31.838009 u:48.424280 v:48.962711 average:34.759466 min:31.699297 max:40.857847
Results with the patch:
SSIM Y:0.970051 (15.236232) U:0.991883 (20.905857) V:0.993174 (21.658049) All:0.981290 (17.279202)
PSNR y:34.412108 u:48.504454 v:48.969496 average:37.264644 min:34.310637 max:42.373392
Signed-off-by: Thomas Mundt <tmundt75@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
On ARM platforms, accessing the PMU registers requires special user
access permissions. Since there is no other way to get accurate timers,
the current implementation of timers in FFmpeg rely on these registers.
Unfortunately, enabling user access to these registers on Linux is not
trivial, and generally involve compiling a random and unreliable github
kernel module, or patching somehow your kernel.
Such module is very unlikely to reach the upstream anytime soon. Quoting
Robin Murphin from ARM:
> Say you do give userspace direct access to the PMU; now run two or more
> programs at once that believe they can use the counters for their own
> "minimal-overhead" profiling. Have fun interpreting those results...
>
> And that's not even getting into the implications of scheduling across
> different CPUs, CPUidle, etc. where the PMU state is completely beyond
> userspace's control. In general, the plan to provide userspace with
> something which might happen to just about work in a few corner cases,
> but is meaningless, misleading or downright broken in all others, is to
> never do so.
As a result, the alternative is to use the Performance Monitoring Linux
API which makes use of these registers internally (assuming the PMU of
your ARM board is supported in the kernel, which is definitely not a
given...).
While the Linux API is obviously cross platform, it does have a
significant overhead which needs to be taken into account. As a result,
that mode is only weakly enabled on ARM platforms exclusively.
Note on the non flexibility of the implementation: the timers (native
FFmpeg vs Linux API) are selected at compilation time to prevent the
need of function calls, which would result in a negative impact on the
cycle counters.
Adds another test for asetnsamples filter where padding of the last
frame is switched off. Renames the existing test to make the difference
obvious.
Tested-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Tobias Rapp <t.rapp@noa-archive.com>
Makes the handling of unspecified/unknown color_range values on stream
level consistent to the value used on frame level.
Signed-off-by: Tobias Rapp <t.rapp@noa-archive.com>
Adds FATE tests for the previously untested allrgb, allyuv, rgbtestsrc,
smptebars, smptehdbars and yuvtestsrc filters.
Also adds a test for testsrc2 filter with rgb+alpha.
Tested-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Tobias Rapp <t.rapp@noa-archive.com>
The -map option allows for a trailing ? so that an error is not thrown if
the input stream does not exist.
This capability is extended to the map_channel option.
This allows a ffmpeg command not to break if an input channel does not
exist, which can be of use (for instance, scripts processing audio
channels with sources having unset number of audio channels).
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
When sidx box support is enabled, the code will skip reading all
trun boxes (each containing ctts entries for samples inthat box).
If seeks are attempted before all ctts values are known, the old
code would dump ctts entries into the wrong location. These are
then used to compute pts values which leads to out of order and
incorrectly timestamped packets.
This patch fixes ctts processing by always using the index returned
by av_add_index_entry() as the ctts_data index. When the index gains
new entries old values are reshuffled as appropriate.
This approach makes sense since the mov demuxer is already relying
on the mapping of AVIndex entries to samples for correct demuxing.
As a result of this all ctts entries are now 1-count. A followup
change will be submitted to remove support for > 1 count entries
which will simplify seeking.
Notes for future improvement:
Probably there are other boxes (stts, stsc, etc) that are impacted
by this issue... this patch only attempts to fix ctts since it
completely breaks packet timestamping.
This patch continues using an array for the ctts data, which is not
the most ideal given the rearrangement that needs to happen (via
memmove as new entries are read in). Ideally AVIndex and the ctts
data would be set-type structures so addition is always worst case
O(lg(n)) instead of the O(n^2) that exists now; this slowdown is
noticeable during seeks.
Signed-off-by: Dale Curtis <dalecurtis@chromium.org>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Fixes filter-pixfmts-scale test failing on big-endian systems due to
alpSrc not being cast to (const int32_t**).
Also fixes distortions in the output alpha channel values by copying the
alpha channel code from the rgba64 case found elsewhere in output.c.
Fixes ticket 6555.
Signed-off-by: James Cowgill <James.Cowgill@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This commit switches off forced correct nesting of tags and only keeps
it for font tags. See long explanations in the code for the rationale.
This results in various FATE changes which I'll explain here:
- various swapping in font attributes, this is mostly noise due to the
old reverse stack way of printing them. The new one is more correct as
the last attribute takes over the previous ones.
- unrecognized tags disappears
- invalid tags that were previously displayed aren't anymore (instead,
we have a warning). This is better for the end user
The main benefit of this commit is to be more tolerant to error, leading
to a better handling of badly nested tags or random wrong formatting for
the end user.
This reverts commit 04aa09c4bc
and reintroduces 0ff5567a30 that
was temporarily reverted due to minor regressions.
It also reverts e5bce8b4ce that fixed FATE refs.
The fate-ffm change is caused by field_order now being set
on the output format because the first frame arrives earlier.
The fate-mxf change is assumed to be the same.
The scale2ref filter will now maintain the DAR of the main input and
not the DAR of the reference input. This previous behavior was deemed
counterintuitive for most (all?) use-cases.
Before:
scale2ref=iw/4:ow/mdar
in w:320 h:240 fmt:rgb24 sar:1/1
ref w:640 h:360 fmt:rgb24 sar:1/1
out w:160 h:120 fmt:rgb24 sar:4/3 flags:0x2
SAR: ((120 * 640) / (160 * 360)) * (1 / 1) = 4 / 3
DAR: (160 / 120) * (4 / 3) = 16 / 9
(main out now same DAR as ref)
Now:
scale2ref=iw/4:ow/mdar
in w:320 h:240 fmt:rgb24 sar:1/1
ref w:640 h:360 fmt:rgb24 sar:1/1
out w:160 h:120 fmt:rgb24 sar:1/1 flags:0x2
SAR: ((120 * 320) / (160 * 240)) * (1 / 1) = 1 / 1
DAR: (160 / 120) * (1 / 1) = 4 / 3
(main out same DAR as main in)
The scale2ref FATE test has also been updated.
Signed-off-by: Kevin Mark <kmark937@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
