These lines get chaotic in the region outside the legal range, but we
already know that we won't ever sample outside of the source gamut for
well-behaved content after 8ceb7a54.

Ported from e67b46bc, to avoid expanding gamut out of the box with
the new perceptual gamut mapper.

Need to check for the no-op case here properly.

Fixes things like ICC test profiles directly at the source level, by
simply disabling gamut mapping altogether in this case. Also disables
e.g. out-of-gamut warnings, but I consider this a degenerate case
anyway.

(And we would probably be better off fixing it to do the gamut testing
in software, anyway)

Specifically, we need this to constrain the effects of bidirectional
gamut mappers when unintended.

Note: This cherry-picked version adds the new API to an internal header,
instead of exposing it publicly.

As a result of cherry-picking the new perceptual onto the release
branch.

Makes this function return true for near-matches.

Now that the default tone mapper is over an order of magnitude faster,
we can burn a lot more iterations on a good 3DLUT. This reduces the
interpolation error to the point where tricubic interpolation is already
overkill.

The new perceptual tone mapper is substantially better than the old one
at mapping hues correctly, so we can raise this test precision by almost
an order of magnitude.

Based on a radically new design, which is just a saturation map combined
with a simple smoothstep to protect the in-gamut region from being
warped, alongside a final RGB softclip pass to combat clipping as a
result of differences in the hue leaf shape, resulting in a nice, smooth
colorimetric mapping.

Overall represents a major improvement (YMMV).

If target is 0.0.

Prevents infinite loop here.

Make it reusable.

This margin needs to be applied when the source's hue leaf chromaticity
_exceeds_ the target's. The logic was the wrong way around, leading to
this code effectively doing nothing.

Unfortunately, regresses accuracy of the hue mapping code slightly, in
exchange for being more resilient to clipping.

This was only intended for SDR *sources* with known contrast
information, mapping onto SDR outputs. Also reword the previous comment
for clarity.

Kacper Michajłow authored 1 year ago and Niklas Haas committed 1 year ago

Got broken during code move from plplay.c to windows.c. Since
waitable timers, doesn't seems to care about per process
timeBeginPeriod, as long as global timer resolution is set, I didn't
notice missing call during testing.

Even if `acquire_image` is false. Otherwise, this will not happen on
re-draws with changed image state.

Not entirely sure if this is kosher, but it seems to work fine in
practice for me. Maybe the more principled solution would be to move the
acquision of the ref image to the first pass_init().

Whenever possible, avoids conversion between different SDR contrast
levels unless both the source and target have HDR metadata forced.

Fixes: #289

This is arguably a subjective trade-off, since it means you get the HLG
OOTF look&feel instead of our own tone-mapping. But, that's also what
HLG is supposed to do, so this is arguably more faithful to the spec.

Users can already override it anyway by forcing specific metadata for
the HLG source (e.g. mastering metadata), which is more likely than not
what already is happening in practice anyway.

See-Also: https://github.com/mpv-player/mpv/issues/11461

Works around AV_NOWARN_DEPRECATED not being correctly defined for the
version of FFmpeg in this image.

Permanent solution would be to upgrade FFmpeg.

See-Also: !510

Kacper Michajłow authored 1 year ago and Niklas Haas committed 1 year ago

Fixes compilation of glslang.cc due to more strict C++ prohibiting to
jump over variable initialization.

Also as a bonus check sleep time after converting to integer as it may
end up zero.

Fixes: https://github.com/haasn/libplacebo/issues/185

rgba32f is not available everywhere, no need to make this test fail in
such cases.

Instead of pl_opengl_create(), where it doesn't belong (it's too late to
affect feature checks).

Instead of doing this in pl_opengl_create

Triggers pl_debug_abort() otherwise.

This is the minimum version required according to the
GL_ARB_compute_shader documentation, and certainly, trying to use a
lower version results in invalid syntax errors being reported for the
workgroup size layout qualifiers.

Using r8 triggers driver bugs on at least some platforms (e.g. nvidia),
rgba8 is a simple work-around.

Fixes: #252

This export/import test was completely redundant with
pl_test_export_import.

Somehow a deviation in the previous (deband) test resulted in this test
pattern being subtly different from the version actually being tested,
and only on certain platforms (NVIDIA GL, but not AMD GL). While that
is certainly odd in its own right, this commit fixes a straightforward
logic bug in the test.

Fixes: #252 (comment 401483)

Unlike their 1 and 2 channel brethren, these are randomly lumped in with
the YCbCr sampled formats, which have a variety of restrictions that our
code is not currently equipped to handle.

Easier to just disable them for now.

v1.3.250 still has a race condition where the semaphore validation can
time-out sometimes, bump it up to v1.3.259...

Possible after e58db9d0

This was bugged until 1.3.250+, the latest SDK release.

This is technically redundant with the report above, but this happens at
higher verbosity and also here the version is more relevant (at least to
me).

No idea why this is considered a warning, even.