| Commit message (Collapse) | Author | Age |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The basic idea is to use dynamically generated shaders instead of a
single monolithic file + a ton of ifdefs. Instead of having to setup
every aspect of it separately (like compiling shaders, setting uniforms,
perfoming the actual rendering steps, the GLSL parts), we generate the
GLSL on the fly, and perform the rendering at the same time. The GLSL
is regenerated every frame, but the actual compiled OpenGL-level shaders
are cached, which makes it fast again. Almost all logic can be in a
single place.
The new code is significantly more flexible, which allows us to improve
the code clarity, performance and add more features easily.
This commit is incomplete. It drops almost all previous code, and
readds only the most important things (some of them actually buggy).
The next commit will complete it - it's separate to preserve authorship
information.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
This is based on pretty much the same (somewhat naive) logic right now.
I'm not convinced that the extra logic that eg. madVR includes is worth
enough to warrant heavily confusing the logic for it.
This shouldn't slow down the logic at all in any sane shader compiler,
and indeed it doesn't on any shader compiler that I tested.
Note that this currently doesn't affect cscale at all, due to the weird
implementation details of that.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Instead of rendering and upscaling each video frame on every vsync, this
version of the algorithm only draws them once and caches the result,
so the only operation that has to run on every vsync is a cheap linear
interpolation, plus CMS/dithering.
On my machine, this is a huge speedup for 24 Hz content (on a 60 Hz
monitor), up to 120% faster. (The speedup is not quite 250% because of
the overhead that the larger FBOs and CMS provides)
In terms of the implementation, this commit basically swaps
interpolation and upscaling - upscaling is moved to inter_program, and
interpolation is moved to the final_program.
Furthermore, the main bulk of the frame rendering logic (upscaling etc.)
was moved to a separete function, which is called from
gl_video_interpolate_frame only if it's actually necessarily, and
skipped otherwise.
|
| |
|
|
|
|
|
| |
GLES2 shaders do not have line continuation characters. Abuse the
HAVE_ARRAYS define to exclude code which uses arrays, and which also
happens to cover all code that defines multi-line macros. (So yes, this
is a hack.)
|
| |
|
|
|
|
|
| |
This adds a small check for candidates that could potentially be inside
the radius, but aren't necessarily. This speeds up performance by a
negligible amount on my hardware, but it's mainly a prerequisite for a
further change (using a larger true radius).
|
| |
|
|
|
| |
This should be no problem... but it _might_ help with #1536, so it's
worth a try.
|
| |
|
|
|
|
|
|
|
|
| |
This introduces a new option linear-scaling, which is now implied by
srgb, icc-profile and sigmoid-upscaling.
Notably, this means (sigmoidized) linear upscaling is now enabled by
default in opengl-hq mode. The impact should be negligible, and there
has been no observation of negative side effects of sigmoidized scaling,
so it feels safe to do so.
|
| |
|
|
|
|
|
|
|
|
| |
Make the lazy gamma initialization less weird, and make the default
value of the "gamma" sub-option 1.0. This means --vo=opengl:help will
list the actual default value.
Also change the lower bound to 0.1 - avoids a division by zero (I don't
know how shaders handle NaN, but it's probably not a good idea to give
them this value).
|
| |
|
|
|
|
|
|
|
| |
There was some code accounting for different gamma values for R/G/B.
It's inherited from an old, undocumented MPlayer feature, which was at
some point disabled for convenience by myself (meaning you couldn't
actually set separate gamma because it was removed from the property
interface - mp_csp_copy_equalizer_values() just set them to the same
value). Get rid of these meaningless leftovers.
|
| |
|
|
|
|
| |
Before this, enabling :gamma in combination with :sigmoid and probably a few
other things results in ugly artifacts because the video isn't clamped until
after the :gamma was applied (or at all, if the cms_matrix is unused).
|
| |
|
|
|
|
|
|
|
| |
Windows Intel drivers seem to reject some (AFAIK) valid GLSL. Make them
happy.
<rossy> GL_RENDERER='Intel(R) HD Graphics 4400'
<rossy> GL_VERSION='3.0.0 - Build 10.18.14.4080'
<rossy> GL_SHADING_LANGUAGE_VERSION='1.30 - Build 10.18.14.4080'
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This deals with subsampled YUV video that has odd sizes, for example a
5x5 image with 4:2:0 subsampling.
It would be easy to handle if we actually passed separate texture
coordinates for each plane to the shader, but as of now the luma
coordinates are implicitly rescaled to chroma one. If luma and chroma
sizes don't match up, and this is not handled, you'd get a chroma shift
by 1 pixel.
The existing hack worked, but broke separable scaling. This was exposed
by a recent commit which switched to GL_NEAREST sampling for FBOs. The
rendering was accidentally scaled by 1 pixel, because the FBO size used
the original video size, while textures_sizes[0] was set to the padded
texture size (i.e. one pixel larger).
It could be fixed by setting the padded texture size only on the first
shader. But somehow that is annoying, so do something else. Don't pad
textures anymore, and rescale the chroma coordinates in the shader
instead.
Seems like this somehow doesn't work with rectangle textures (and
introduces a chroma shift), but since it's only used when doing VDA
hardware decoding, and the bug occurs only with unaligned video sizes, I
don't care much.
Fixes #1523.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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).
|
| |
|
|
|
| |
This emphasizes the fact that scale is used for *all* image upscaling,
with cscale only serving a minor role for subsampled material.
|
| |
|
|
|
|
|
|
|
|
| |
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 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 speeds up performance by a factor of something like 10%,
since it omits unnecessary checks.
This will also make adding anti-ringing easier.
|
| |
|
|
|
|
| |
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.
|
| |
|
|
|
|
|
|
|
|
| |
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).
|
| |
|
|
|
|
|
| |
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.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
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
|
| |
|
|
|
| |
This is the polar (elliptic weighted average) version of lanczos.
This introduces a general new form of polar filters.
|
| |
|
|
|
|
|
|
|
| |
This avoids issues when upscaling directly in linear light, and is the
recommended way to upscale images according to imagemagick.
The default slope of 6.5 offers a reasonable compromise between
ringing artifacts eliminated and ringing artifacts introduced by
sigmoid-upscaling. Same goes for the default center of 0.75.
|
| |
|
|
| |
This removes an old code path that was disabled in 016bb14.
|
| | |
|
| |
|
|
|
|
|
|
|
| |
Whether we have texture_rg doesn't matter much anymore; the scaler
should be fine with this. But on ES 2.0, 1st class arrays are missing,
so even if filterable float textures should be available, it won't work.
Dithering (at least the "fruit" variant) will not work either, because
it uses floats.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Apparently GLES 2 and 3 do not support this. (The implementations I
tested with were derived from desktop OpenGL and were not overly strict
with this.)
This is no problem; just use GL_RGBA and mangle the channels in the
shader.
Also disable direct support for image formats like IMGFMT_RGB555 with
GLES; at least some of them are not supported in this form, and the
formats aren't important anyway.
|
| |
|
|
|
|
|
|
| |
Rather basic support. Almost nothing works, and even if it does, it's
bound to be inefficient (due to texture upload). This was tested with
the nVidia desktop binary drivers, which provide GLES 2 support only.
However, nVidia is not known to be very strict about OpenGL, and the
driver is very new too, so the vo_opengl code will have bugs too.
|
| |
|
|
|
|
|
|
|
|
|
| |
This was a nice trick to get the mpv colormatrix directly into OpenGL,
because the memory representation happened to match.
Unfortunately, OpenGL ES 2 doesn't have glUniformMatrix4x3fv().
Even more unfortunately, the memory representation is now incompatible.
It would be nice to change it, but that would mean getting into a big
mess.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
If GL_RED was not available, we used GL_ALPHA. But this is an
unnecessary complication, and it's easier to use GL_LUMINANCE instead.
With the latter, a texture will return the .r component set, and as long
as the shader doesn't look at the other components, the shader doesn't
need any changes.
Some of the changes added in 0e8fbdbd are now unneeeded.
Also, realign the entire gl_byte_formats_legacy table.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Tested with MESA on software emulation. Seems to work well, although the
default FBO format in opengl-hq disables most interesting features. I
have no idea how well it will work on real hardware (or if it does at
all).
Unfortunately, some features, including playback of 10 bit video, are
not supported. Not sure what to do about this.
GLES 2 or 1 do not work.
|
| |
|
|
|
|
|
|
| |
Older GLSL dialects as well as GLES3 do not support the following things
in expressions:
- implicit conversions of integer constants to float
- arithmetic of float*vecN
|
| |
|
|
|
| |
Features not supported are disabled (although with a misleading error
message).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Not all filter sizes the shaders could handle were in the filter_sizes
list. The shader can handle any multiple of 4 (the sizes 2 and 6 are
special-cased to keep it simple).
Add all possible filter sizes, up to 64. 64 is ridiculously high anyway.
Most of the larger filter sizes are completely useless for upscaling,
but help with the fancy-downscaling option. (Although it would still be
more efficient to use cascaded scalers to handle downscaling better.)
I considered doing something less stupid than the hardcoded array, but
it seems this is still the simplest solution.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Before this commit, the convolution scaler shader functions were pre-
instantiated in the shader file. For every filter size, a corresponding
function (with the filter size as suffix) had to be present.
Change this, and make the C code emit the necessary bits.
This means the shader code is much reduced. (Although hopefully it
doesn't make shader compilation faster - it would require a really dumb
compiler if it spends its time on dead code.)
It also makes it more flexible, which is the main goal.
The DEF_SCALER0 stuff is needed because the C code writes the header of
the shader, at a point where scaler macros are not defined yet.
|
| |
|
|
|
|
|
|
|
| |
This was a microoptimization for small filters which need 4 or less
weights per sample point. When I originally wrote this code, using a 1D
texture seemed to give a slight speed gain, but now I couldn't measure
any difference.
Remove this to simplify the code.
|
| |
|
|
|
|
|
|
|
| |
There's not much of a reason to have the actual convolution code in a
separate function. Merging them actually simplifies the code a bit, and
gets rid of the repetitious macro invocations to define the functions
for each filter size.
There should be no changes in behavior or output.
|
| |
|
|
|
|
| |
For better downscaling.
Maybe the list of filter sizes shouldn't be static...
|
| |
|
|
|
|
| |
Also replace the weights calculations for 8/12/16 with the generic
weight function definition macro. (The weights 2/4/6 follow slightly
different rules.)
|
| |
|
|
| |
Signed-off-by: wm4 <wm4@nowhere>
|
| |
|
|
|
| |
I didn't quite understand this comment after looking at the code again
months later, so I reworded it for better clarity.
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Apparently this is needed for correct 3D mode subtitles. In general,
it seems you need to duplicate the whole "GUI", so it's done for all
OSD elements.
This doesn't handle the "duplication" of the mouse pointer. Instead,
the mouse can be used for the top/left field only. Also, it's possible
that we should "compress" the OSD in the direction it's duplicated, but
I don't know about that.
Fixes #1124, at least partially.
|
| |
|
|
| |
Removes '##' operator from OpenGL shader code.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Regression since commit f14722a4. For some reason, this worked on
nvidia, but rightfully failed on mesa.
At least in C, the ## operator indeed needs two macro arguments, and
you can't just concatenate with non-arguments.
This change will most likely fix it.
CC: @bjin
|
| |
|
|
|
|
| |
Although cscale is rarely used, it's possible that params of cscale
are accidentally set to lparam1 and lparam2, which might cause
unexpected results.
|
| |
|
|
|
|
| |
Close #837
Signed-off-by: wm4 <wm4@nowhere>
|
| |
|
|
|
|
|
|
|
| |
This also avoids an extra matrix multiplication when using :srgb, making
that path both more efficient and also eliminating more hard-coded
values.
In addition, the previously hard-coded XYZ to RGB matrix will be
dynamically generated.
|
| |
|
|
| |
Signed-off-by: wm4 <wm4@nowhere>
|
| |
|
|
|
|
|
| |
This add support for reading primary information from lavc, categorized
into BT.601-525, BT.601-625, BT.709 and BT.2020; and passes it on to the
vo. In vo_opengl, we always generate the 3dlut against the wider BT.2020
and transform our source into this colorspace in the shader.
|
| |
|
|
| |
Source: http://www.itu.int/dms_pubrec/itu-r/rec/bt/R-REC-BT.2020-0-201208-I!!PDF-E.pdf
|
wm4
Niklas Haas
Stefano Pigozzi
Bin Jin