This new filter is slightly sharper, and significantly faster, than
mitchell. It also tends to preserve detail better. All in all, there is
no reason not to use it by default, especially from a performance PoV.
(In vo_gpu_next, hermite is implemented efficiently using hardware
accelerated bilinear interpolation)
See-Also: 75b3947b2c
It can hurt people's feelings to refer to scalers as "high quality" and
"low quality", when it is so subjective.
I decided to preserve the lanczos sections at least because it's mostly
talking about the difference between EWA Lanczos and Lanczos, which is
less controversial than the difference between, say, Lanczos and
Catmull.
Avoid generating too much audio after EOF.
Note: This often has no effect, because less audio is produced than
required.
Usually this comes to effect with the userspeed filter at high speed
(4x) and going back to 1x speed to remove the filter.
After the final input packet, the filter padded with silence to allow
one more iteration. That was not enough to process the final frames.
Continue padding the end of `input_buffer` with silence until the final
frames have been processed.
Implementation: Instead of padding when adding final samples, pad before
running WSOLA iteration. Count number of added silent frames and
remaining input frames for time keeping.
This changes the emitted pts values from the start of the search block
to the center of the search block. Change initial `output_time`
accordingly. Initial `search_block_index` is irrelevant, because it's
overwritten before the first iteration.
Using the `output_time` removes the rounding of `search_block_index`,
which also fixes the <20 microsecond gaps in timestamps between output
packets.
Rationale:
The variance in audio position was in the range `0..search-interval`.
With this change, the range is
(- search-interval / 2)..(search-interval / 2)`
which ensures lower maximum offset.
Target block can be anywhere in the previous search-block, varying by
`search-interval` while the filter is active. This resulted in constant
audio offset when returning to 1x playback speed.
- Move the search block to the target block to sync up exactly.
- Drop old frames to minimize input_buffer usage.
The internal time update function involved multiple problems:
- Time was updated after WSOLA iteration. The means speed was updated
one iteration later than it could be.
- The update functions caused spikes of too many or too few samples
advanced, leading to audio glitches on speed changes.
- The inconsistent updates made it very difficult to produce gapless
audio packets.
- The `output_time` update function involved complicated feedback:
`search_block_index` influenced how many frames from `input_buffer`
are retained, which influenced how much `output_time` is changed,
which influenced `search_block_index`.
With these changes:
- Time is updated before WSOLA iterations. Speed changes are effective
instantly.
- There are no spikes in playback speed during speed changes.
- No significant gaps are introduced in output packets.
- The time update function becomes (function calls omitted for brevity)
output_time += ola_hop_size * playback_rate
Functions received a `playback_rate` parameter to check how many samples
are needed before iteration. Internal state is only updated when the
iteration is actually run, so the speed is allowed to change until
enough data is received.
The first WSOLA iteration overlapped audio with whatever was in the
`wsola_output` buffer. This was either silence (if not run before), or
old frames (if switching to 1x and back to a different speed).
Track the state of the output buffer and memcpy the whole window for the
first iteration instead.
`output_time` is used to set the center of the search block. Init of
both `search_block_index` and `output_time` with 0 caused inconsistent
search block movement for the first iterations.
Initialize with `search_block_center_offset` for consistency with initial
`search_block_index`.
Fixes#12028
There was an additional issue that audio was always delayed by half the
configured search-interval. This was caused by the `out` buffer length
not being included in the delay calculation.
Notes:
- Every WSOLA iteration advances the input buffer by _some amount_, and
produces data in the output buffer always of size `ola_hop_size`.
- `mp_scaletempo2_fill_buffer` is always called with `ola_hop_size`
- Thus, the rendered frames are always cleared immediately after
processing, and `num_complete_frames` is 0 in the delay calculation.
- The factors contributing to delay are:
- the pending samples in the input buffer according to the search
block position, and
- the pending rendered samples in the output buffer (always empty in
practice).
The frame_delay code looked like that of the rubberband filter, which
might not work for scaletempo2. Sometimes a different amount of input
audio was consumed by scaletempo2 than expected. It may have been caused
by speed changes being a more dynamic process in scaletempo2. This can
be seen by where `playback_rate` is used in `run_one_wsola_iteration`:
`playback_rate` is only referenced after the iteration, when updating
the time and removing old data from buffers.
In scaletempo2, the playback speed is applied by changing the amount the
search block is moved. That apparently averages out correctly at
constant playback speed, but when the speed changes, the error in this
assumption probably spikes. This error accumulated across all speed
changes because of the persistent `frame_delay` value.
With the removal of the persistent `frame_delay`, there should be no way
for the audio to drift off. By deriving the delay from filter buffer
positions, and the buffers are filled only as much as needed, the delay
always stays within buffer bounds.
Very amusingly, the manual previously said "The rpi VO will be removed
in mpv 0.23.0." Obviously, this didn't happen. It looks like wm4 changed
his mind on this with 51fd8f6fe1, but
never updated the docs. What was the problem? Who knows but allegedly
there are people that use --vo=rpi (it's supposed to work at least).
Actually context_rpi is probably broken since wm4 changed fullscreen
handling in 59cdfe50b2 and no one ever
bothered to fix it after that commit. In fact, that's the last commit
that touched that particular file. Someone did actually fix vo_rpi after
that so that one probably works at least. Anyways, I wouldn't mind just
deleting all rpi stuff, but there's probably some people out there that
use it so whatever. Remove the outdated deprecated language in the
documentation and just accept that we support this unfortunately.
Defining MPV_CPLUGIN_DYNAMIC_SYM during plugin compilation will replace mpv_*
functions with function pointers. Those pointer will be initialized when
loading the plugin.
It is recommended to use this symbol table when targeting Windows. The loader
does not have notion of global symbols. Loading cplugin into mpv process will
not allow this plugin to call any of the symbols that may be available in
other modules. Instead cplugin has to link explicitly to specific PE binary,
libmpv-2.dll/mpv.exe or any other binary that may have linked mpv statically.
This limits portability of cplugin as it would need to be compiled separately
for each of target PE binary that includes mpv's symbols. Which in practice
is unrealictis, as we want one cplugin to be loaded without those restrictions.
Instead of linking to any PE binary, we create function pointer for all mpv's
exported symbols. For convinience names of entrypoints are redefined to those
pointer so no changes are required in cplugin source code, except defining
MPV_CPLUGIN_DYNAMIC_SYM. Those function pointer are exported to make them
available for mpv to init with correct values during runtime, before calling
`mpv_open_cplugin`.
Note that those pointer are decorated with `selectany` attribute, so no need
to worry about multiple definitions, linker will keep only single instance.
This fixes cplugin usability on Windows. Without any API changes, only
recompilation with -DMPV_CPLUGIN_DYNAMIC_SYM is needed.
The goal is to provide simple to understand quality/performance level
profiles for the users.
Instead of default and gpu-hq profile. There main profiles were added:
- fast: can run on any hardware
- default: balanced profile between quality and performance
- high-quality: out of the box high quality experience. Intended
mostly for dGPU.
Summary of three profiles, including default one:
[fast]
scale=bilinear
cscale=bilinear (implicit)
dscale=bilinear
dither=no
correct-downscaling=no
linear-downscaling=no
sigmoid-upscaling=no
hdr-compute-peak=no
[default] (implicit mpv defaults)
scale=lanczos
cscale=lanczos
dscale=mitchell
dither-depth=auto
correct-downscaling=yes
linear-downscaling=yes
sigmoid-upscaling=yes
hdr-compute-peak=yes
[high-quality] (inherits default options)
scale=ewa_lanczossharp
cscale=ewa_lanczossharp (implicit)
hdr-peak-percentile=99.995
hdr-contrast-recovery=0.30
allow-delayed-peak-detect=no
deband=yes
scaler-lut-size=8
296d40dc6f changed how the vo handled
redraw requests in order to fix a race condition that can occur with
pausing. However, there was a slight oversight because a redraw request
that occurred while the core was unlocked and the video was still
playing would still be kept true (previously, this was always cleared).
That redraw is essential if mpv is paused otherwise the old issue comes
back, but if the video is playing it's unnecessary since the next loop
around will simply draw whatever we needed. The extra redraw could cause
a frame drop for some people in certain instances, so the solution is to
simply always clear redraw requests if !in->paused. This eliminates the
extra redraw but still keeps it when pausing.
Fixes#12426 and fixes#11579.
After fixing the B and C params for bcspline, it ended up being the same
thing as bicubic. There's no reason to have two names for the same
filter, so remove bcspline and keep bicubic to match libplacebo.
All we have to do is set VO_CAP_ROTATE90 and then use buffer transform.
Maybe ideally we would rotate with vavpp instead if possible, but this
would be a convoluted mess that I'm not really interested in dealing
with right now.
No need for this since it's entirely redundant with just changing the
filter radius directly. In fact, that's the whole *point* of the filter
radius - it does not modify the filter, it modifies the scaling of the
window.
Of course, this does not work for non-resizable kernels. But, really,
who cares?
Replaced by mathematically and numerically precise constants, the result
of exact computation. Also updates the description to provide more
explanation and motivation.
We currently always scaled the window to the size of the configured
radius. However, this is wrong - we should instead be scaling it to the
size of the sharpened/blurred kernel. Since the window is always
stretched to the configured size of the filter, we can accomplish this
easily by just multiplying the blur value into the filter radius
directly, and then using that adjusted radius in place of `f.radius`
everywhere.
On a side note, this gives a very minor performance boost to
ewa_lanczossharp for no downside.
Upstream finally caved in to peer pressure and added this filter. Of
course, this also removes the fallback for people on older versions of
libplacebo, but people using mpv git master are probably using
libplacebo git master anyway. It's time to debloat this code.
Kacper Michajłow
Niklas Haas
Alex Mitzsch
ferreum
Dudemanguy
chuck-
llyyr