Current default is to do various funky types of UB.

This version of the headers adds support for the vulkansc API, which
pulls in a bunch of definitions that we don't care about.

To properly exclude these, we need to start blacklisting all types not
defined in the vulkan environment. Also leverage this to start
blacklisting platform-specific crap in a more principled way.

This is probably still not the correct way to do it, but it works for
now.

Fixes: videolan/libplacebo#249
Fixes: https://github.com/haasn/libplacebo/issues/155

When the RPU is available, we can infer the brightness metadata for now.

So that we can use simple Dolby Vision RPU buffers to infer whatever
metadata that can be useful. Currently used to extract the dynamic scene
brightness to be used for tone mapping.

Requires `libplacebo` to be linked against `libdovi` dependency, or no-op.

As a helper for adding DOVI metadata to a `pl_frame`.

Kacper Michajłow authored 2 years ago and Niklas Haas committed 2 years ago

This provides feedback about renderer state.

Kacper Michajłow authored 2 years ago and Niklas Haas committed 2 years ago

This allows disabling hooks more selectively instead disabling all as a
bulk.

Kacper Michajłow authored 2 years ago and Niklas Haas committed 2 years ago

This makes it work nicer with color speace hint and dynamic metadata.

- Always show nominal min/max and dynamically update it in hint mode
- When toggling color space hint reset primaries to one that makes sense
  for current swapchain configuration

Kacper Michajłow authored 2 years ago and Niklas Haas committed 2 years ago

Compositor may resize window if it doesn't like our requested values.
For example if it is bigger than display resolution.

Cosmetic change to make symbol name more descriptive in binary.

Fixes: 790e3c78

A LUT marked dynamic is *expected* to change frequently, no need to nag
PL_LOG_DEBUG about it. Drop verbosity in this case.

Inside vk_buf_barrier, we also need to make host-writes available, even
for coherent memory, and even after using FlushMappedMemoryRanges.

Also include the host_read dependency for completeness' sake, just to be
on the safe side.

This needs to be properly defined as well even after removal of the
pragma before global includes.

Fixes: 82023b9e

Mostly for stylistic reasons, and to avoid drowning in global header
includes as libplacebo grows in size. Will help reduce the amount of
time needed to re-compile libplacebo when making changes to any header,
although ironically, on my end this commit actually makes a clean
compilation *slower*.

Allows us to avoid having to include every header file here.

Simple oversight.

Oops..

To avoid constant recompilations due to mixing between frames of
different source brightness (before tone-mapping).

This is not a good solution long-term, ideally we should move
tone-mapping to happen before frame mixing. But it represents the best
compromise currently available to us.

This is designed for SDR<->HDR, so prevent using it for HDR<->HDR out of
the box.

This has no inverse curve, so don't default `auto` to it.

Infinitely useful during development, no need to constantly NIH this
patch or re-invent it ad-hoc.

Instead of having to replicate this logic at every usage site, the
number of which is about to grow.

Tests the fallback path for inverse tone-mapping on functions without
inverse tone mapping curves.

Especially for the pure BPC GLSL fallback path

It's not that hard to convert to PQ and back in pure GLSL, so we might
as well do that for pure BPC. Produces significantly better results (no
clipping) than naive/linear black point stretching.

Always do BPC in PQ space instead of the function's native space,
because this produces significantly better results (no black clipping).

The D > T branch had a number of fatal issues. One, the `p / N`
calculation was accidentally rounded due to performing integer division.
Two, the knee point was picked way too high (0.5 / 0.5). And three, even
the basis knee was possibly way too bright due to being implicitly
scaled up to the target brightness.

Solve all three problems with a combination of approaches:
1. Picking a perceptually linear knee that doesn't explode as D > T
2. Constraining the basis knee to ensure we don't raise brightness
3. Mix towards linear slightly more aggressively to reduce the effects
   of the basis OOTF on the image appearance as D -> T

Meanwhile, the D < T branch resulted in very inconsistent behavior as a
result of wild deviations in the basis knee function, and also the lack
of slope intercept. Fix both problems by a combination of approaches:

1. Make the knee more clip()-like (1:1 mapping) at low luminances, but
   constrain this to prevent the P[1] slope from exploding
2. Also mix in a linear intercept for P[1] as D -> 0