The logic disabled framedropping if the frame was interpolated (i.e. the
render call is only done to interpolate between the previous frame, and
the frame before that).
It seems doing this wasn't even necessary, and broke framedrop in
smoothmotion mode. In fact, this code did nothing for display with video
fps below display fps. It did prevent the framedrop counter from going
up, though. So change it so that dropped interpolated frames are never
reported. (Doing so can give confusing results, such as dropping 1000s
of frames on slow operations like video start or changing filters.)
SmoothMotion is a way to time and blend frames made popular by MadVR. It's
intended behaviour is to remove stuttering caused by mismatches between the
display refresh rate and the video fps, while preserving the video's original
artistic qualities (no soap opera effect). It's supposed to make 24fps video
playback on 60hz monitors as close as possible to a 24hz monitor.
Instead of drawing a frame once once it's pts has passed the vsync time, we
redraw at the display refresh rate, and if we detect the vsync is between two
frames we interpolated them (depending on their position relative to the vsync).
We actually interpolate as few frames as possible to avoid a blur effect as
much as possible. For example, if we were to play back a 1fps video on a 60hz
monitor, we would blend at most on 1 vsync for each frame (while the other 59
vsyncs would be rendered as is).
Frame interpolation is always done before scaling and in linear light when
possible (an ICC profile is used, or :srgb is used).
These are now auto-detected sanely; and enabled whenever it would be a
performance or quality gain (which is pretty much everything except
bilinear/bilinear scaling).
Perhaps notably, with the absence of scale_sep, there's no more way to
use convolution filters on hardware without FBOs, but I don't think
there's hardware in existence that doesn't have FBOs but is still fast
enough to run the fallback (slow) 2D convolution filters, so I don't
think it's a net loss.
This is significantly faster for FBOs on most modern GPUs, although it
did not result in a huge difference for the video source texture on the
sizes I tested. It might be more significant for 1080p or 4K content, so
it's worth revisiting this in the future.
It also renames SAMPLE_BILINEAR to SAMPLE_TRIVIAL to match the
semantics.
This is better even for non-separable. The only exception is when using
bilinear for both lscale and cscale. I've fixed the
documentation/comments to make more sense.
This is not quite the same thing as madVR's antiringing algorithm, but
it essentially does something similar.
Porting madVR's approach to elliptic coordinates will take some amount
of thought.
This fixes compatibility with GLES 2.0 and makes the code a bit neater
in general. It also properly forces indirect scaling for subsampled
video regardless of the lscale setting.
At least the scale_sep_fbo could have been uninitialized or initialized
incorrectly when switching between scalers (e.g. from bilinear to
lanczos). Calling check_resize() should take care of this.
Instead of reading back the image from textures, keep a reference to the
original image, and return that.
The main reason this was done this way was that originally, images
weren't refcounted, and would be deallocated or overwritten as soon as
the VO's draw call returned. But now there isn't really a good reason
for this anymore. One possibly _could_ argue that it was better because
other code could reuse the image sooner (e.g. for the cache), but on the
other hand, the VO runs already on a different thread, and filtering and
decoding each run on other threads too, so this argument probably
wouldn't hold up.
There was a case when we could have rendered to an output surface while
it's still used for display. Not sure why the API doesn't do this
automatically.
Instead of converting the hw surface to an image in the VO, provide a
generic way to convet hw surfaces, and use this in the screenshot code.
It's all relatively straightforward, except vdpau is being terrible. It
needs a huge chunk of new code, because copying back is not simple.
Before this commit, each hw backend had their own specific struct types
for context, and some, like VDA, had none at all. Add a context struct
(mp_hwdec_ctx) that provides a somewhat generic way to pass the hwdec
context around. Some things get slightly better, some slightly more
verbose.
mp_hwdec_info is still around; it's still needed, but is reduced to its
role of handling delayed loading of the hwdec backend.
This also fixes the maximum range to 16.0, which was previously set to
32.0 and incorrectly documented as 8.0. 16 taps should be more than
anybody will ever need, but it's the highest radius that's supported by
all affected filters.
And remove all uses of the VFCAP_CSP_SUPPORTED* constants. This is
supposed to reduce conversions if many filters are used (with many
incompatible pixel formats), and also for preferring the VO's natively
supported pixel formats (as opposed to conversion).
This is worthless by now. Not only do the main VOs not use software
conversion, but also the way vf_lavfi and libavfilter work mostly break
the way the old MPlayer mechanism worked. Other important filters like
vf_vapoursynth do not support "proper" format negotation either.
Part of this was already removed with the vf_scale cleanup from today.
While I'm touching every single VO, also fix the query_format argument
(it's not a FourCC anymore).
I'm still not sure how exactly handling of "lost" devices is supposed
to be handled. In theory, you only have to "reset" the device, instead
of recreating _everything_. But as it is, the code for proper uninit
and for handling the reset is exactly the same, so move it into a
function to reduce code duplication and the danger of potential bugs.
Apparently, extremely crappy graphics drivers don't allow you to use
shaders. Simply disable use of shaders if this happens, and use the
"old" method instead.
One unexpectedly tricky thing is that you need a d3d_device to create
a shader, which in turn requires a window, so the initialization order
changes.
Don't load all the legacy functions (including ancient extensions).
Slightly simplify function loader and context creation, now that legacy
GL doesn't need to be handled. Remove the code for drawing OSD in legacy
mode.
Remove all the header hacks, which were meant for ancient OpenGL headers
which didn't even support things like OpenGL 1.3. Instead, adjust the
GLX check to make sure we get both OpenGL 3x and 2.1 symbols. For win32
and OSX, we assume that the user has the latest headers anyway. For
wayland, we hope that things somehow go right.
Simply clamp off the U/V components in the colormatrix, instead of doing
something special in the shader.
Also, since YA8/YA16 gave a plane_bits value of 16/32, and a colormatrix
calculation overflowed with 32, add a component_bits field to the image
format descriptor, which for YA8/YA16 returns 8/16 (the wrong value had
no bad consequences otherwise).
If video output and VO don't support the same format, a conversion
filter needs to be insert. Since a VO can support multiple formats, and
the filter chain also can deal with multiple formats, you basically have
to pick from a huge matrix of possible conversions.
The old MPlayer code had a quite naive algorithm: it first checked
whether any conversion from the list of preferred conversions matched,
and if not, it was falling back on checking a hardcoded list of output
formats (more or less sorted by quality). This had some unintended side-
effects, like not using obvious "replacement" formats, selecting the
wrong colorspace, selecting a bit depth that is too high or too low, and
more.
Use avcodec_find_best_pix_fmt_of_list() provided by FFmpeg instead. This
function was made for this purpose, and should select the "best" format.
Libav provides a similar function, but with a different name - there is
a function with the same name in FFmpeg, but it has different semantics
(I'm not sure if Libav or FFmpeg fucked up here).
This also removes handling of VFCAP_CSP_SUPPORTED vs.
VFCAP_CSP_SUPPORTED_BY_HW, which has no meaning anymore, except possibly
for filter chains with multiple scale filters.
Fixes#1494.
We still need to send the VO a duration in these cases. Disabling
framedrop has logically absolutely nothing to do with these cases; it
was overlooked in commit 918b06c4.
So we always send the frame duration (or a guess for it), and check
whether framedropping is actually enabled in the VO code. (It would
be cleaner to send framedrop as a flag, but I don't care about that
right now.)
Broke operation with GLSL.
Since 1D texture usage was apparently (and mysteriously) good for speed,
it might be added back, but it's unknown how to do so in a clean way.
The "ontop" and "border" properties already used a common
mp_property_vo_flag() function, and the corresponding VOCTRLs used the
same conventions. "fullscreen" is pretty similar, but was handled
slightly similar. Change how VOCTRL_FULLSCREEN behaves, and use the same
helper function for "fullscreen" as the other flags.
Fixes#1472.
(Maybe these options should have been named --autofit-max and
--autofit-min, but since --autofit-larger already exists, use
--autofit-smaller for symmetry.)
The last video frame is another case that has a separate code path,
although it's pretty similar to the one in commit 73e5aa87. Fix this
in a different way, which also takes care of the last frame case,
although without context the code becomes slightly more tricky.
As further cleanup, move the decision about framedropping itself to
the same place, so the check in vo.c becomes much simpler. The check
for the vo->driver->encode flag, which is remvoed completely, was
redundant too.
Fixes#1480.
After finding out more about how video mastering is done in the real
world it dawned upon me why the "hack" we figured out in #534 looks so
much better.
Since mastering studios have historically been using only CRTs, the
practice adopted for backwards compatibility was to simulate CRT
responses even on modern digital monitors, a practice so ubiquitous that
the ITU-R formalized it in R-Rec BT.1886 to be precisely gamma 2.40.
As such, we finally have enough proof to get rid of the option
altogether and just always do that.
The value 1.961 is a rounded version of my experimentally obtained
approximation of the BT.709 curve, which resulted in a value of around
1.9610336. This is the closest average match to the source brightness
while preserving the nonlinear response of the BT.1886 ideal monitor.
For playback in dark environments, it's expected that the gamma shift
should be reproduced by a user controlled setting, up to a maximum of
1.224 (2.4/1.961) for a pitch black environment.
More information:
https://developer.apple.com/library/mac/technotes/tn2257/_index.html
