It may happen that target format during merge has different sampling
order than the source format, for example rg8 merged with r8 into bgrx8.
In this case merge shader will implicitly reorder components and since
component_mapping is not connected to sample_order, it has to be updated
to match the new reality. Note that we don't apply component_mapping
during merge, it will be applied at the end of read_image pass.

Fixes: https://github.com/mpv-player/mpv/issues/16208

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

To avoid confusion, especially that it was printed as %x without 0x
prefix.

Needed to use LIBAVFORMAT_VERSION_INT. This used to always use the #else
branch up until now without any compiler warnings...

Fixes build failure after https://github.com/FFmpeg/FFmpeg/commit/ec8e796b42bccf999fdf31c110385e2cc119b7f2

When blending onto e.g. yuv420p10le, the alpha channel being scaled down
will result in an overlay that is almost invisible.

oops...

Fixes: 53768aff

This enabled validation of vkCreateInstance.

Remove use of deprecated VK_LAYER_LUNARG_standard_validation and
VK_EXT_validation_features.

This also gives more fain-grained control on VVL options.

This is still unfortunately too new for any current Ubuntu release, so the
CI does not cover it, but it just enables GPU-AV so I'm not too worried about
missing the validation here as long as I can still run it locally.

See-Also: https://github.com/KhronosGroup/Vulkan-ValidationLayers/pull/8559

This fixes a bug where the layout transition would happen too early when
commands were scheduled with external dependencies tied to specific execution
stages, but the layout transitions could happen as soon as the command buffer
started execution.

Fixes: #333

The driver is unloaded we can't really add suppressions for them. So we
are kinda forced to just ignore all leaks.

See:
https://gitlab.freedesktop.org/mesa/mesa/-/issues/12805
https://gitlab.freedesktop.org/mesa/mesa/-/issues/12811

For some reason debug layer thinkgs that maxPushDescriptors is 0, while
it clearly is not.

See: https://github.com/KhronosGroup/Vulkan-ValidationLayers/issues/9542

lcov 2.0+ is more strict about the inconsistencies in the coverage data.

Some of the issues may be related to our source code preprocessing and
line numbering, but even when I tried with smaler example without
shaders, it still doesn't work.

We get errors like:

lcov: ERROR: (inconsistent) "/usr/include/c++/13/bit":368: line is hit
but no branches on line have been evaluated.

or

genhtml: ERROR: duplicate merge record src

and few other, which looks bogus and would need deeper dive into what is
actually happening here and if this is our fault at all.

I tried making it work, and so far only gcovr or lcov < 2 works. With
lcov > 2 it fails in one way or another, I tried some things but it
never full passes.

Additionally Meson doesn't allow to select gcovr instead of lcov easily.
It always prefers lcov.

So for now downgrade lcov to 1.16 in the job, it's small download. And
maybe one day someone will figure this out.

This fixes errors with coverage intrumentation enabled:

error: function starts on a higher line number than it ends
[-Werror=coverage-invalid-line-number]

It's still not perfect, as we can have colisions of the line numbers
between call site and function site. Probably would be better to
generate functions in separate file and #include it. Then the line
numbering will refer to different file name.

But for now this fixes the build with GCC 13.

They are false positives and safe to ignore.

meson since 1.3.0 export UBSAN_OPTIONS and ASAN_OPTIONS for tests when
needed.

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

Merging planes will change format, use this.

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

Revert this change as it is required for interpreting bgr planes
correctly.

Fixes sample generated with: ffmpeg -i <img> -vf format=bgr0 -c:v rawvideo -y test.nut

Also fixes: https://github.com/mpv-player/mpv/issues/13398

This reverts commit 6e15daee.

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

This was incomplete fix and only worked for uploading case. Better fix
to follow in next commit.

This reverts commit a018ab04.

Fixes an issue where unscaled planes with pre-existing shaders would
not be correctly multiplied by the output scale factor because the
scaler was never dispatched.

Sponsored-by: nxtedition

dispatch_sampler() will lose the alpha channel unless explicitly requested.

This could index the array out of bounds if the source struct contains
invalid data. Switch to a switch/case with an explicit pl_unreachable()
default case.

Instead of leaving it as implicitly either premul or independent as a
result of whatever the last operation happened to be.

We should also skip clearing the image background to be consistent with
the previous behavior.

PL_CLEAR_AUTO does not exist. This was clearly left behind from a previous
iteration of the design. The current default is to treat both as
PL_CLEAR_COLOR.

In case demos are compiled, but the nuklear headers are not available,
compilation fails, because the implementation of parse_args function
is protected by the HAVE_NUKLEAR preprocessor macro.

To avoids this issue, move this implementation of this function outside
of this ifdef.

Signed-off-by: Gyorgy Sarvari <skandigraun@gmail.com>

Glslang 15.0.0 moved some code around, add also linking to glslang,
while this is not needed for older versions, it will still work.

Unfortunately CMake config embedded in distributions is not usable
without installing additional package like glslang-tools, because at
least Ubuntu splits it. On Arch it would work, but generally the CMake
config require cmake binary to work also, so let's keep it as-is for
now.

May be useful for debugging.

The colors are encoded in XYZ with EE white point, but this is only the
encoding part. Displayed white point is at code poit X'=3794 Y'=3960
Z'=3890. Reference projector is expected to produce DCI white for those
values.

Although SMPTE standard does not enforce specific white point. For
example ACES uses D60 instead. But the workflow for reference projector
to be DCI compliant and it is on encoders to ensure the content looks
good.

So this is what we do. Decode DCDM X'Y'Z into DCI-P3 RGB.

Now that GLES can do readback this test fails under llvmpipe
but not on real hardware.
This isn't too concerning, so fix the test by raising the tolerance.