For user-legigibility.

Oops. I was testing locally with glslang disabled, unfortunately.

libm is required.

Vulkan 1.2 is the required minimum version, so properly check for it.

This drops support for back-compat with versions of glslang below the
new versioning schema (starting with 10.15.3847). This version of
glslang is, by now, quite ancient, and dropping it simplifies a ton of
code.

Some older platforms (particularly macOS) do not support
clock_gettime(). Fortunately, this is identical to gettimeofday in the
case of CLOCK_REALTIME.

These systems also do not have pthread_condattr_setclock, so we can
re-use this check.

Also add proper error handling to this function.

intmax_t is not enough to guarantee alignment, we need max_align_t.

Fixes https://github.com/haasn/libplacebo/issues/130

This is useful in cases where users build the vulkan-headers (and thus
place vk.xml) inside their custom prefixes, in which case it makes sense
to look there as well.

Do this with highest priority - if a user is building against custom
vulkan headers, we probably want to use them. (They might be pinned to
known-good versions)

Otherwise, overlays are upside down when there's no image available.

Fixes https://github.com/mpv-player/mpv/issues/10336

Instead of just referring to them as generic "polar" and "ortho"
scaling, describe their names and directions properly.

Closes videolan/libplacebo#204
cf. https://github.com/mpv-player/mpv/issues/9451

Adding this for logging purposes, so downstream functions of filter
configuration can still reasonably talk about which filter
function/config they're currently computing.

We don't need a dedicated setting for the plane downscaler because we
never output anything subsampled (thus making downscaling impossible).

Solving the problem by brute force. Maybe it would have been wise to
make alpha also separable somehow, or even allow choosing the sampler by
plane type. But these sound like problems that don't need solving.

Fixes videolan/libplacebo#203

This avoids a dependency on upstream cmsDetectRGBProfileGamma. It's
slightly less precise, using only 256 sample points instead of 4096, but
this difference barely seems to matter in practice, as this is only a
rough estimate anyway.

The main motivation of this change is to avoid a dependency on the
relatively high lcms2 2.13, but it also brings with it the other upside
of allowing us to tune the algorithm to our needs (e.g. not filtering
out linear sections, because we care more about estimating the overall
response to minimize LUT error, and less about finding the technical
gamma).

Technically the old code was undefined behavior (integer overflow).

Sometimes a frame mix simply contains very sparse collection of frames
that are far too far in the past and the future to actually get mixed.
In this case, fall back to ZOH semantics.

Do this by retrying the loop in order to enable the caching logic to
work. Not the prettiest work-around but simple and effective.

Also add a test case to prevent further regressions of this case.

Fixes videolan/libplacebo#202

The old runner stopped existing.

Fixes #128

Even for a coherent buffer, we technically have no guaranteed thet
'vkCmdPipelineBarrier' has completed yet until the command callbacks are
fired. To solve this issue, take the reference unconditionally, and
instead move the coherency conditional to `invalidate_buf`.

As an aside, checking for `buf_vk->mem.data` here is redundant, as this
code will only ever be hit in any of the host readable/writable
conditions.

Avoid having to update a billion URLs

This makes the vk_tex_barrier code unnecessarily buggy. In particular,
the way the code is currently written, we omit pipeline barriers here
*anyway*. And trying to reason about when the synchronization scope can
be ignored and when not gave me a headache so just get rid of this
implicit layout transition business and use the same pipeline barrier
that we use for everything else.

The overhead is negligible anyway.

Don't trigger MSAN on uninstrumented lcms2.

The feature check is insufficient to check for timeline semaphore
support because it doesn't get updated if we don't actually enable the
corresponding extension also.

Recent compilers complain about this being UB

We need to hold on to a reference to the buffer, otherwise this memory
can become undefined/released before the callback fires.

This is never an issue for allocated memory because we take the
noncoherent atom size into account when determining the allocation size.

However, it can be an issue for imports, which can have arbitrary
offsets/sizes. Fortunately, there is a simple solution here - using
VK_WHOLE_SIZE.

Fixes videolan/libplacebo#201

When vulkan is autodetected but spirv/glslang is not available, this
currently triggers an error(). This is undesirable behavior especially
because it can randomly regress existing libplacebo installations merely
due to pulling in the vulkan headers.

Jan Ekström authored 2 years ago and Niklas Haas committed 2 years ago

This enum value is also utilized in d3d11/format, since the original
commit (be32ca68), while the
definition was only in d3d11/gpu. This way both can share the same
compatibility definition.

cobertura was generalized to coverage_report

Oops.

Rewrite all of the ICC profile handling code, fundamentally getting rid
of the 'ICC->ICC transform' design and representing each profile as an
independent object that can be either decoded or encoded, at separate
stages in the pipeline.

This is only an initial version, and serves to establish the API. It is
not yet fully optimized, but optimizations can and will be applied
transparently underneath the hood of this API (e.g. skipping the full
3DLUT when not required, implementing support for fixed function
profiles, natively applying matrix+shaper profiles, etc.)

This design brings with it a number of advantages, most notably allowing
support for using libplacebo-native tone mapping together with ICC
profiles, by properly forwarding the actual contrast values detected
from the profile.

In addition to a redesign of the core logic, also changes things so that
the ICC.h header is always installed, even if non-functional -
`pl_icc_open` will simply always return NULL.

SDR<->SDR conversions are best done with simple linear light stretching,
because everything else introduces too many distortions.

Even if the image has either already been linearized somehow, e.g. if
the image came in linear light to begin with.

Ignoring the ICC profile header for now because it's about to be
refactored entirely.

And also fix the wrong signature of pl_cie_xy_equal. Decided to split
this off to prevent ballooning the following commit (and its API change
log entry).

d3d11_sw_colorspace_hint is utilized both for runtime hints as well
as the initial swap chain configuration to enable consistency
between received swap chains and newly created ones.

Thus, add an internal function which contains the argument of whether
this call was internal or not. In the colorspace hint call, it is
set to false, and during swap chain initialization it is set to true.
Thus when hints are passed, logging will talk about received hints
and when the function is called during initialization, logging will
talk about initial configuration.

The current cache invalidation logic was slightly too aggressive in some
regards, for example also triggering on changes to info_callback or
dither_params. In addition to fixing this, also overhaul the params
hashing code to also provide a best effort estimate of whether or not
the pass is worth caching at all. If not, just skip it if possible.

Fixes https://github.com/mpv-player/mpv/issues/10135