The hook mechanism allows arbitrary processing stages to get dispatched
whenever certain named textures have been "finalized" by the code.
This is mostly meant to serve as a change that opens up the internal
processing in pass_read_video to user scripts, but as a side benefit all
of the code dealing with offsets and plane alignment and other such
confusing things has been rewritten.
This hook mechanism is powerful enough to cover the needs of both
debanding and prescaling (and more), so as a result they can be removed
from pass_read_video entirely and implemented through hooks.
Some avenues for optimization:
- The prescale hook is currently somewhat distributed code-wise. It might be
cleaner to split it into superxbr and NNEDI3 hooks which can each be
self-contained.
- It might be possible to move a large part of the hook code out to an
external file (including the hook definitions for debanding and
prescaling), which would be very much desired.
- Currently, some stages (chroma merging, integer conversion) will
*always* run even if unnecessary. I'm planning another series of
refactors (deferred img_tex) to allow dropping unnecessary shader
stages like these, but that's probably some ways away. In the meantime
it would be doable to re-add some of the logic to skip these stages if
we know we don't need them.
- More hook locations could be added (?)
Instead of rounding down, we round to the nearest float. This reduces
the maximum possible error introduced by this rounding operation. Also
clarify the comment.
Fixes broken colors with --vf=format=0bgr (but only if deband is
disabled).
0bgr means the first byte is padding, while the following three bytes
are bgr. From the vo_opengl perspective, it has 4 physical components
with 3 logical components. copy_img_tex() simply copied 3 components
from the physical representation, which means the last component (r) was
sliced off.
Fix this by not using p->color_swizzle for packed formats, and instead
let packed formats set the per-plane swizzle in texplane.swizzle. The
latter applies the swizzle as part of operation in copy_img_tex(), which
essentially moves physical to logical representations.
Unfortunately, debanding (and thus with opengl-hq defaults) is still
broken.
Make the find_plane_format function take a bit count.
This also makes the function's comment true for the first time the
function and its comment exist. (It was commented as taking bits, but
always took bytes.)
Now that we know in advance whether an implementation should support a
specific format, we have more flexibility when determining which format
to use.
In particular, we can drop the roundabout ES logic.
I'm not sure if actually trying to create the FBO for probing still has
any value. But it might, so leave it for now.
Even if everything else is available, the need for first class arrays
breaks it. In theory we could fix this since we don't strictly need
them, but I guess it's not worth bothering.
Also give the misnamed have_mix variable a slightly better name.
This merges all knowledge about texture format into a central table.
Most of the work done here is actually identifying which formats exactly
are supported by OpenGL(ES) under which circumstances, and keeping this
information in the format table in a somewhat declarative way. (Although
only to the extend needed by mpv.) In particular, ES and float formats
are a horrible mess.
Again this is a big refactor that might cause regression on "obscure"
configurations.
In theory this was needed for the previous commit (but wasn't in
practice, since for hwdec the LUMINANCE_ALPHA mangling is not applied
anymore, and ANGLE uses RG textures in absence of GL_ARB_texture_rg for
whatever crazy reasons).
In practice this caused funky colors on OSX with the uyvy422 format,
which is also fixed in this commit.
Rename gl_hwdec_driver.map_image to map_frame, and let it fill out a
struct gl_hwdec_frame describing the exact texture layout. This gives
more flexibility to what the hwdec interop can export. In particular, it
can export strange component orders/permutations and textures with
padded size. (The latter originating from cropped video.)
The way gl_hwdec_frame works is in the spirit of the rest of the
vo_opengl video processing code, which tends to put as much information
in immediate state (as part of the dataflow), instead of declaring it
globally. To some degree this duplicates the texplane and img_tex
structs, but until we somehow unify those, it's better to give the hwdec
state its own struct. The fact that changing the hwdec struct would
require changes and testing on at least 4 platform/GPU combinations
makes duplicating it almost a requirement to avoid pain later.
Make gl_hwdec_driver.reinit set the new image format and remove the
gl_hwdec.converted_imgfmt field.
Likewise, gl_hwdec.gl_texture_target is replaced with
gl_hwdec_plane.gl_target.
Split out a init_image_desc function from init_format. The latter is not
called in the hwdec case at all anymore. Setting up most of struct
texplane is also completely separate in the hwdec and normal cases.
video.c does not check whether the hwdec "mapped" image format is
supported. This should not really happen anyway, and if it does, the
hwdec interop backend must fail at creation time, so this is not an
issue.
This gives us 16 bit fixed-point integer texture formats, including
ability to sample from them with linear filtering, and using them as FBO
attachments.
The integer texture format path is still there for the sake of ANGLE,
which does not support GL_EXT_texture_norm16 yet.
The change to pass_dither() is needed, because the code path using
GL_R16 for the dither texture relies on glTexImage2D being able to
convert from GL_FLOAT to GL_R16. GLES does not allow this. This could be
trivially fixed by doing the conversion ourselves, but I'm too lazy to
do this now.
The active texture and some pixelstore parameters are now always reset
to defaults when entering and leaving the renderer. Could be important
for libmpv.
Apply basic transformations like rotation by 90° and mirroring when
sampling from the source textures. The original idea was making this
part of img_tex.transform, but this didn't work: lots of code plays
tricks on the transform, so manipulating it is not necessarily
transparent, especially when width/height are switched. So add a new
pre_transform field, which is strictly applied before the normal
transform.
This fixes most glitches involved with rotating the image.
Cropping and rotation are now weirdly separated, even though they could
be done in the same step. I think this is not much of a problem, and
has the advantage that changing panscan does not trigger FBO
reallocations (I think...).
This commit refactors the 3DLUT loading mechanism to build the 3DLUT
against the original source characteristics of the file. This allows us,
among other things, to use a real BT.1886 profile for the source. This
also allows us to actually use perceptual mappings. Finally, this
reduces errors on standard gamut displays (where the previous 3DLUT
target of BT.2020 was unreasonably wide).
This also improves the overall accuracy of the 3DLUT due to eliminating
rounding errors where possible, and allows for more accurate use of
LUT-based ICC profiles.
The current code is somewhat more ugly than necessary, because the idea
was to implement this commit in a working state first, and then maybe
refactor the profile loading mechanism in a later commit.
Fixes#2815.
This also draws it after color management etc. In a nutshell, this
change makes the transparency checkerboard independent of upscaling,
panning, cropping etc. It will always be the same apparent size and
position (relative to the window).
It will also be independent of the video colorspace and such things.
(Note: This might cause white imbalance issues if playing a file with a
white point that does not match the display, in absolute colorimetric
mode. But that's uncommon, especially in conjunction with transparent
image files, so it's not a primary concern here)
Until now, we've let the windowing backend decide. But since they
usually require premultiplied alpha, and premultiplied alpha is easier
to handle, hardcode it.
The recent changes fixed rotation handling, but reversed the rotation
direction. The direction is expected to be counter-clockwise, because
demuxers export video rotation metadata as such.
This has been completely broken since commit 93546f0c. But even before,
rotation handling did not make too much sense. In particular, it rotated
the contents of the cropped image, instead of adjusting the crop
rectangle as well. The result was that things like panscan or zooming
did not behave as expected with rotation applied.
The same is true for vertical flipping. Flipping is triggered by
negative image stride. OpenGL does not support flipping the image on
upload, so it's done as part of the rendering. It can be triggered with
--vf=flip, but other filters and even decoders could setup negative
stride to flip the image.
Fix these issues by applying transforms to texture coordinates properly,
and by making rotation and flipping part of these transforms.
This still doesn't work properly for separated scaling. The issue is
that we'd have to adjust how the passes are done. For now, pick a very
stupid solution by rotating the image to a FBO, and then scaling from
that. This has the avantage that the scale logic doesn't have to be
complicated for such a rare case. It could be improved later.
Prescaling is apparently still broken. I don't know if chroma
positioning works properly either. None of this should affect the case
with no rotation.
If the texture count is lower than 4, entries in va.textcoord[] will
remain uninitialized. While this is unlikely to be a problem (since
these values are unused on the shader side too), it's not nice and might
explain some things which have shown up in valgrind.
Fix by always initializing the whole thing.
Glitches when resizing are still possible, but are reduced. Other VOs
could support this too, but don't need to do so.
(Totally avoiding glitches would be much more effort, and probably not
worth the trouble. How about you just watch the video the player is
playing, instead of spending your time resizing the window.)
This also gets rid of the kind of hard to read texture swizzle setup and
turns it into something dumber.
Assumes that we don't create any FBOs with 2 channel formats. (Only the
video source textures are handled by this commit.)
This is a fresh implementation from scratch that carries with it
significantly less baggage and verbosity from the previous (ported)
version.
The actual values for the masks and such were copied from the
current code. Behavior and performance should be unaffected.
An important difference between the old code and the new code is that
the new code always explicitly samples from the first component, rather
than being able to process multiple planes at once.
Since prescale-luma only affects luma, I deemed this unnecessary. May
change in the future, if prescale-chroma ever gets implemented. But
prescaling multiple planes would be slow to do this way. (Better would
be to generalize it to differently-sized vectors)
Instead of hard-coding the logic and planes to skip, factor this out
to a reusible function, and instead add the number of relevant
coordinates to the texture state.
Since prescale now literally only affects the luma plane (and the
filters are all designed for luma-only operation either way), the option
has been renamed and the documentation updated to clarify this.
This is a pretty major rewrite of the internal texture binding
mechanic, which makes it more flexible.
In general, the difference between the old and current approaches is
that now, all texture description is held in a struct img_tex and only
explicitly bound with pass_bind. (Once bound, a texture unit is assumed
to be set in stone and no longer tied to the img_tex)
This approach makes the code inside pass_read_video significantly more
flexible and cuts down on the number of weird special cases and
spaghetti logic.
It also has some improvements, e.g. cutting down greatly on the number
of unnecessary conversion passes inside pass_read_video (which was
previously mostly done to cope with the fact that the alternative would
have resulted in a combinatorial explosion of code complexity).
Some other notable changes (and potential improvements):
- texture expansion is now *always* handled in pass_read_video, and the
colormatrix never does this anymore. (Which means the code could
probably be removed from the colormatrix generation logic, modulo some
other VOs)
- struct fbo_tex now stores both its "physical" and "logical"
(configured) size, which cuts down on the amount of width/height
baggage on some function calls
- vo_opengl can now technically support textures with different bit
depths (e.g. 10 bit luma, 8 bit chroma) - but the APIs it queries
inside img_format.c doesn't export this (nor does ffmpeg support it,
really) so the status quo of using the same tex_mul for all planes is
kept.
- dumb_mode is now only needed because of the indirect_fbo being in the
main rendering pipeline. If we reintroduce p->use_indirect and thread
a transform through the entire program this could be skipped where
unnecessary, allowing for the removal of dumb_mode. But I'm not sure
how to do this in a clean way. (Which is part of why it got introduced
to begin with)
- It would be trivial to resurrect source-shader now (it would just be
one extra 'if' inside pass_read_video).
Why was this done so stupidly, with so many complicated special cases,
before? Declare it once so the shader bits don't have to figure out where
and when to do so themselves.
Apple crap (namely hardware decoding interop) forces us to use rectangle
textures for input. But after that we continue with normal textures.
This was not considered for debanding, and the sampler type used for it
can be different depending on the exact render chain. Simply use the
target type of the input texture.
It thinks that integer_conv_fbo[index] is implied to be accessed with up
to index=5. Although that is theoretical only, it has a point that this
makes no sense. Use the same constant for the array allocation, to make
it more uniform and robust.
Fixes CID 1350060.
GLES does not support high bit depth fixed point textures for unknown
reasons, so direct 10 bit input is not possible. But we can still use
integer textures, which are supported by GLES 3.0. These store integer
data just like the standard fixed point textures, except they are not
normalized on sampling. They also don't support bilinear filtering, and
require a special sampler ("usampler2D").
While these texture formats enable us to shuffle the data to the GPU,
they're rather impractical with the requirements mentioned above and our
current architecture. One problem is that most code assumes it can
always use bilinear scaling (even if bilinear is never used when using
appropriate scale/cscale options). Another is that we don't have any
concept of running a function on a texture in an uniform way.
So for now, run a simple conversion step through a FBO. The FBO will use
the rgba16f format normally, which gives enough bits for 10 bit, and
will at least gracefully degrade with higher depth input.
This is bound to be much slower than a more "direct" method, but at
least it works and is simple to implement.
The odd change of function call order in init_video() is to properly
disable "dumb mode" (no FBO use) if these texture formats are in use.
This was never reset - absolutely can't be right. If the renderer
somehow switches back to another codepath, it certainly has to be reset.
Maybe this was hard to hit, as the normalization is going to be
idempotent in simpler cases (like rendering RGBA input).
Also get rid of the "merged" variable.
Often requested. The main argument, that prominent scalers like sharpen
change the image even if no scaling happens, disappeared anyway.
("sharpen", unsharp masking, is neither prominent nor a scaler anymore.
This is an artifact from MPlayer, which fuses unsharp masking with
bilinear scaling in order to make it single-pass, or so.)
Some VOs had support for these - remove them.
Typically, these formats will have only some use in cases where using
RGB software conversion with libswscale is faster than letting the
VO/GPU do it (i.e. almost never). For the sake of testing this case,
keep IMGFMT_RGB565. This is the least messy format, because it has no
padding/alpha bits with unknown semantics.
Note that decoding to these formats still works. We'll let libswscale
repack the data to whatever the VO in use can take.
Do this to make the license situation less confusing.
This change should be of no consequence, since LGPL is compatible with
GPL anyway, and making it LGPL-only does not restrict the use with GPL
code.
Additionally, the wording implies that this is allowed, and that we can
just remove the GPL part.
Should take care of the planned FFmpeg AV_PIX_FMT_P010 addition. (This
will eventually be needed when doing HEVC Main 10 decoding with DXVA2
copyback.)
Add a "blend-tiles" choice to the "alpha" sub-option. This is pretty
simplistic and uses the GL raster position to derive the tiles. A weird
consequence is that using --vo=opengl and --vo=opengl-hq gives different
scaling behavior (screenspace pixel size vs. source video pixel size
16x16 tiles), but it seems we don't have easy access to the original
texture coordinates. Using the rasterpos is probably simpler.
Make this option the default.
Since alpha isn't pulled through the colormatrix (maybe it should?), we
reject alpha formats with odd sizes, such as yuva444p10.
But the awful tex_mul path in vo_opengl does this anyway (at some points
even explicitly), which means there will be a subtle difference in
handling of 16 bit yuv alpha formats. Make it consistent and always
apply the range adjustment to the alpha component. This also means odd
sizes like 10 bit are supported now.
This assumes alpha uses the same "shifted" range as the yuv color
channels for depths larger than 8 bit. I'm not sure whether this is
actually the case.
Store the determined framebuffer depth in struct GL instead of
MPGLContext. This means gl_video_set_output_depth() can be removed, and
also justifies adding new fields describing framebuffer/backend
properties to struct GL instead of having to add more functions just to
shovel the information around.
Keep in mind that mpgl_load_functions() will wipe struct GL, so the
new fields must be set before calling it.
Why not.
Also, instead of disabling hue/saturation for RGB, just don't apply
them. (They don't make sense for conversion matrixes other than YUV, but
I can't be bothered to keep the fine-grained disabling of UI controls
either.)
With default setting, the matrix for fruit dithering requires 12 bits
precision (values from 0/4096 to 4095/4096). But 16-bit float
provides only 10 bits. In addition, when `dither-size-fruit=8` is
set, 16 bits are required from the texture format.
Fix this by attempting to use 16 bit integer texture first. This is
still not precise, but should be better than using a half float.
The computation of the tex_mul variable was broken in multiple ways.
This variable is used e.g. by debanding for moving expansion of 10 bit
fixed-point input to normalized range to another stage of processing.
One obvious bug was that the rgb555 pixel format was broken. This format
has component_bits=5, but obviously it's already sampled in normalized
range, and does not need expansion. The tex_mul-free code path avoids
this by not using the colormatrix. (The code was originally designed to
work around dealing with the generally complicated pixel formats by only
using the colormatrix in the YUV case.)
Another possible bug was with 10 bit input. It expanded the input by
bringing the [0,2^10) range to [0,1], and then treating the expanded
input as 16 bit input. I didn't bother to check what this actually
computed, but it's somewhat likely it was wrong anyway. Now it uses
mp_get_csp_mul(), and disables expansion when computing the YUV matrix.
It turns out that with accurate lookup we can decrease the
default size of texture now. Do it to compensate the performance
loss introduced by the LUT_POS macro.
Define a macro to correct the coordinate for lookup texture. Cache
the corrected coordinate for 1D filter and use mix() to minimize the
performance impact.
There are claims that nnedi3.c doesn't constitute its own new
implementation, but is derived from existing HLSL or OpenCL shaders
distributed under the LGPLv3 license.
Until these are resolved, do the "correct" thing and require
--enable-gpl3 to build nnedi.
At least I hope so.
Deriving the duration from the pts was not really correct. It doesn't
include speed adjustments, and becomes completely wrong of the user e.g.
changes the playback speed by a huge amount. Pass through the accurate
duration value by adding a new vo_frame field.
The value for vsync_offset was not correct either. We don't need the
error for the next frame, but the error for the current one. This wasn't
noticed because it makes no difference in symmetric cases, like 24 fps
on 60 Hz.
I'm still not entirely confident in the correctness of this, but it sure
is an improvement.
Also, remove the MP_STATS() calls - they're not really useful to debug
anything anymore.
This was just converting back and forth between int64_t/microseconds and
double/seconds. Remove this stupidity. The pts/duration fields are still
in microseconds, but they have no meaning in the display-sync case (also
drop printing the pts field from opengl/video.c - it's always 0).
I decided that I actually can't stand how vo_opengl unnecessarily puts
the video through 3 shader stages (instead of 1). Thus, what was meant
to be a fallback for weak OpenGL implementations, the dumb-mode, now
becomes default if the user settings allow it.
The code required to check for the settings isn't so wild, so I guess
it's manageable. I still hope that one day, our rendering logic can
generate ideal shader stages for this case too.
Note that in theory, dumb-mode could be reenabled at runtime due to a
color management 3D LUT being set, so a separate dumb_mode field is
required. The dumb-mode option can't just be overwritten.
Unfortunately, color management can still not work, because no GLES
version specified so far support fixed-point 16 bit textures. Maybe
we could use integer textures, but these don't support filtering.
Using float textures would be another possibility.
Polar scalers use 1D textures, because they're slightly faster on some
GPUs than 2D textures. But 2D textures work too, so add support for
them.
Allows using these scalers with ANGLE.
This makes advanced scaling sort-of work for GLES 3.0 (on ANGLE). It's
still not very advisable, as 8 bits might not be enough to avoid
debanding. (Ironically, the debanding filter can be enabled, and does
not raise any GL errors - but probably doesn't do anything useful.)
Something goes wrong somewhere. Don't bother, it's only needed for
compatibility with our absolute baseline (GL 2.1/GLES 2).
On the other hand, we can process nv12 formats just fine.
In the display-sync, non-interpolation case, and if the display refresh
rate is higher than the video framerate, we duplicate display frames by
rendering exactly the same screen again. The redrawing is cached with a
FBO to speed up the repeat.
Use glBlitFramebuffer() instead of another shader pass. It should be
faster.
For some reason, post-process was run again on each display refresh.
Stop doing this, which should also be slightly faster. The only
disadvantage is that temporal dithering will be run only once per video
frame, but I can live with this.
One aspect is messy: clearing the background is done at the start on the
target framebuffer, so to avoid clearing twice and duplicating the code,
only copy the part of the framebuffer that contains the rendered video.
(Which also gets slightly messy - needs to compensate for coordinate
system flipping.)
The nnedi3 prescaler requires a normalized range to work properly,
but the original implementation did the range normalization after
the first step of the first pass. This could lead to severe quality
degradation when debanding is not enabled for NNEDI3.
Fix this issue by passing `tex_mul` into the shader code.
Fixes#2464
Commit 27dc834f added it as such.
Also remove the check for glUniformBlockBinding() - it's part of an
extension, and the check glGetUniformBlockIndex() already checks whether
the extension is fully available.
Implement NNEDI3, a neural network based deinterlacer.
The shader is reimplemented in GLSL and supports both 8x4 and 8x6
sampling window now. This allows the shader to be licensed
under LGPL2.1 so that it can be used in mpv.
The current implementation supports uploading the NN weights (up to
51kb with placebo setting) in two different way, via uniform buffer
object or hard coding into shader source. UBO requires OpenGL 3.1,
which only guarantee 16kb per block. But I find that 64kb seems to be
a default setting for recent card/driver (which nnedi3 is targeting),
so I think we're fine here (with default nnedi3 setting the size of
weights is 9kb). Hard-coding into shader requires OpenGL 3.3, for the
"intBitsToFloat()" built-in function. This is necessary to precisely
represent these weights in GLSL. I tried several human readable
floating point number format (with really high precision as for
single precision float), but for some reason they are not working
nicely, bad pixels (with NaN value) could be produced with some
weights set.
We could also add support to upload these weights with texture, just
for compatibility reason (etc. upscaling a still image with a low end
graphics card). But as I tested, it's rather slow even with 1D
texture (we probably had to use 2D texture due to dimension size
limitation). Since there is always better choice to do NNEDI3
upscaling for still image (vapoursynth plugin), it's not implemented
in this commit. If this turns out to be a popular demand from the
user, it should be easy to add it later.
For those who wants to optimize the performance a bit further, the
bottleneck seems to be:
1. overhead to upload and access these weights, (in particular,
the shader code will be regenerated for each frame, it's on CPU
though).
2. "dot()" performance in the main loop.
3. "exp()" performance in the main loop, there are various fast
implementation with some bit tricks (probably with the help of the
intBitsToFloat function).
The code is tested with nvidia card and driver (355.11), on Linux.
Closes#2230
Add the Super-xBR filter for image doubling, and the prescaling framework
to support it.
The shader code was ported from MPDN extensions project, with
modification to process luma only.
This commit is largely inspired by code from #2266, with
`gl_transform_trans()` authored by @haasn taken directly.
next_vsync/prev_vsync was only used to retrieve the vsync duration. We
can get this in a simpler way.
This also removes the vsync duration estimation from vo_opengl_cb.c,
which is probably worthless anyway. (And once interpolation is made
display-sync only, this won't matter at all.)
vo_frame.num_vsyncs can be != 1 in some cases in normal sync mode too.
This is not a very exact fix, but in exchange it's robust. (These
vo_frame flags are way too tricky in combination with redrawing and
such.)
There were occasional shader compilation and rendering failures if FBOs
were unavailable. This is caused by the FBO caching code getting active,
even though FBOs are unavailable (i.e. dumb-mode).
Boken by commit 97fc4f.
Fixes#2432.
This speeds up redraws considerably (improving eg. <60 Hz material on a 60 Hz
monitor with display-sync active, or redraws while paused), but slightly
slows down the worst case (eg. video FPS = display FPS).
Adds support for AV_PIX_FMT_GBRP9, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12,
AV_PIX_FMT_GBRP14, AV_PIX_FMT_GBRP16, AV_PIX_FMT_GBRAP, and
AV_PIX_FMT_GBRAP16.
(Not that it matters, because nobody uses these anyway.)
If interpolation is enabled, then this causes heavy artifacts if done
while unpaused. It's preferable to allow a latency of a few frames for
the change to take full effect instead. If this is done paused, the
frame is fully redrawn anyway.
Surfaces used by hardware decoding formats can be mapped exactly like a
specific software pixel format, e.g. RGBA or NV12. p->image_params is
supposed to be set to this format, but it wasn't.
(How did this ever work?)
Also, setting params->imgfmt in the hwdec interop drivers is pointless
and redundant. (Change them to asserts, because why not.)
This turns the old scalers (inherited from MPlayer) into a pre-
processing step (after color conversion and before scaling). The code
for the "sharpen5" scaler is reused for this.
The main reason MPlayer implemented this as scalers was perhaps because
FBOs were too expensive, and making it a scaler allowed to implement
this in 1 pass. But unsharp masking is not really a scaler, and I would
guess the result is more like combining bilinear scaling and unsharp
masking.
2 things are being stupid here: Apple for requiring rectangle textures
with their IOSurface interop for no reason, and OpenGL having a
different sampler type for rectangle textures.
The removal of source-shader is a side effect, since this effectively
replaces it - and the video-reading code has been significantly
restructured to make more sense and be more readable.
This means users no longer have to constantly download and maintain a
separate deband.glsl installation alongside mpv, which was the only real
use case for source-shader that we found either way.
This is mostly to cut down somewhat on the amount of code bloat in
video.c by moving out helper functions (including scaler kernels and
color management routines) to a separate file.
It would certainly be possible to move out more functions (eg. dithering
or CMS code) with some extra effort/refactoring, but this is a start.
Signed-off-by: wm4 <wm4@nowhere>
