Very hard to get precise controls here.

Based on two-pass design where the main interpolation kernel is combined
with a smaller, second ...

Normalizes the two branches by simply assigning to `c` in the `in`
branch.

To help with debugging this stuff a bit more.

This can force the program into certain framerates on VRR displays, for
testing purposes. (Or it...

This effectively makes the render loop time itself to the system clock
instead of the display/vs...

Based on nanosleep(). No point in using clock_nanosleep() since
nanosleep() already implies CLOC...

Rather than always adding the measured delta from the previous frame,
use a single clock as the ...

Sorely needed for being able to programmaticaly debug or reason about
the contents of a pl_queue.

There are 240 Hz monitors on the market now. 200 Hz is definitely not a
valid cut-off point.

Instead of floats. Floating precision is around 1e-7f, so after ~10000
seconds of playback (a fe...

When source FPS exceeds display FPS (and we know this fact), we should
increase the number of fr...

In VFR sources, delaying for 8 frames unnecessarily prolongs the time it
takes to react to chang...

To mirror the D3D11 option, and to work around driver-related dithering
bugs.

See-Also: https:/...

No inherent need to use the sync2 command here because we don't require
the extra granularity pr...

Notably missing on MoltenVK. Work around it for now.

Fixes: https://github.com/haasn/libplacebo...

Not very efficient (does a bunch of allocations per submit), but such is
the price of being on o...

Requires some annoying boilerplate to map the flags and barriers, but
fortunately we can get awa...

After this became available inside vk_cmdpool. Saves a bunch of
redundant boilerplate struct pas...

And even if we need this in the future, we can get it from the
vk_cmdpool now.

As a convenience.

The cutoff radius was being computed with the normalized radius, rather
than the scaled radius, ...

Instead of forcing everything but linear/saturation to clip, force
everything *but* clip to line...

This reverts commit d03db4df6c4242daf47ee40cdd426533ccfa66c5.

Fixes: https://github.com/mpv-pla...

Needed to ensure the extended capabilities required by vf_libplacebo
(i.e. rendering) are actual...

Fixes: https://github.com/mpv-player/mpv/issues/11905

If expressions would end up with the same slugify'd name, we need to
properly deduplicate them, ...

After a long deliberation period, I settled on this design to improve
both the workflow of writi...

More of a test / proof of concept than anything else.

This is one of the main use cases where it actually provides a benefit,
as this significantly re...

Including the new #pragma GLSL macro

Required for the GLSL pragma refactor step.

For consistency and better syntax highlighting.

To better show the size of the 'protected region' in a various color
mapping scheme, as well as ...

For overlays, which are probably saturated synthetic graphics, this
probably leads to a subjecti...

Much better than clipping here, if the target brightness is less than
the monitor's brightness (...

This can be done for free by just multiplying the appropriate conversion
matrix into the colorsp...

No reason to forbid this, because it's implemented in pure GLSL.
Override the force-tone-mapping...

Ideally we'd somehow auto-detect this based on detected GPU performance
or type (iGPU vs discret...

Splits peak detection and tone-mapping params into standard and high
quality. The former disable...

This is not a performance/quality trade-off option.

It really is impossible to stress how much peak detection improves
out-of-the-box HDR tone-mappi...

We hard-code gaussian for this so there's no reason to disable it here.
I'm pretty sure it was a...

Enables histogram estimation and contrast recovery on high quality
systems out of the box.

To match the existing logic of silently disabling them if the FBO is
non-storable. Probably bett...

Always enables this feature if requested by the user, period.

Except when upscaling, in which case we update the peak before
upscaling.

Note that there is no...

But not PRE/POST-kernel, since no kernel was applied.

Image overlays haven't been drawn directly onto the indirect image in
quite a while, making this...

Substantially improves performance on systems that are constrained by
global atomic pressure, li...

Makes user-applied modifications visible.

With the new tone-mapping in place, this now produces a substantially
better image out-of-the-box.

More descriptive name.

But not 100%, just somewhere in the middle. Constant tuned to preserve a
nice balance between ex...

So I can actually understand what's going on a bit more easily.

Requires indirection via pl_buf_copy.

Fixes: https://github.com/mpv-player/mpv/issues/11864

The reason this new curve works so well is because it incorporates a bit
of hard clipping in RGB...

To allow re-using it in other places.

Instead of applying this complicated soft-knee along the gamma axis, as
inspired by the gamut ma...

This reverts commit f016363d589a1ffd744be65a14455e4e79750bbd.

This change had unintended conseq...

After rotating hue vectors, strongly saturated colors may exceed the
source chromaticity bounds ...

Check for unguarded availability to avoid calling possiblly not
available weak symbols.

Make this less confusing re: HLG, now that we no longer allow users to
change the OOTF.

As advertised/documented. Also clamp to the upper bound of 10k nits
(sane PQ value range), and h...

The old logic incorrectly inferred SDR min_luma from the static
max_luma, even if maximum lumina...

Add a series of explicit tests to ensure not regressing this logic
during refactors. Note that t...

min_luma missing should not count against HDR10 metadata validity,
especially given that this va...

In particular, how values of 0 are treated as "unknown / missing".

To be more informative about what values we're setting and why. I also
decided to drop the HDR b...

When requesting a dynamic metadata type, we currently fall back to the
fixed 10k from PQ. It mak...

It's better to filter those illegal values out closer to the source
(i.e. when parsing the struct).

Sanity clamp here, to avoid regressing badly broken metadata samples.

This is in sync with mpv minimum value. If anything just for test we can
go so low.

This is the minimum value in mpv and users expect it to work. So let them
set it so low, for bet...

This is a documentation bug only, the actual sover is and continues to
be 290.

Makes this transition a lot softer, especially around blue, to preserve
gradients better. We sti...

Significantly improves cache hit rate by allowing us to have a separate
per-hue cache before and...

Costs us basically nothing and improves the result slightly.

Absolutely massive speedup on typical systems, for very little overhead.

Probably not the smart...

No longer needed after 8ceb7a544, at least in theory.

Move this to a dedicated struct so we can more easily replace the cache
by a different one. Stil...

Dramatically improves appearance on test samples and other badly
out-of-range content. In such c...

This helps hide issues with badly out-of-range values in certain test
patterns, Mad Max, etc.

And re-organize slightly.

After the recent series of improvements, LMS hybrid mixing is no longer
needed and actually hurt...

Hue is the most important axis, and even 45 is not enough by far.
Increase it considerably. Idea...

For my own experimentation.

Substantially higher quality than the default trilinear interpolation,
but also significantly sl...

Could probably be improved further by avoiding some redundant mix()
calls on the texture coordin...

Do the calculation with 64-bit floating points, converting the final
matrix back to 32-bit float...

4e-3 was an extremely optimistic, and in practice too low, cutoff value,
resulting in very bad 3...

Documentation fix only.

To be odd, instead of even, for accuracy reasons.