Martin Storsjö authored 2 years ago and Niklas Haas committed 2 years ago

Clang 15 made "incompatible pointer to integer conversion" an error
instead of a plain warning. This fixes errors like these:

../src/vulkan/gpu_pass.c:208:59: error: incompatible integer to pointer conversion passing 'VkPipeline' (aka 'unsigned long long') to parameter of type 'const void *' [-Wint-conversion]
        vk_dev_callback(vk, (vk_cb) destroy_pipeline, vk, *out_pipe);
                                                          ^~~~~~~~~
../src/vulkan/command.h:36:52: note: passing argument to parameter 'arg' here
                     const void *priv, const void *arg);
                                                   ^
../src/vulkan/gpu_pass.c:209:19: error: incompatible pointer to integer conversion assigning to 'VkPipeline' (aka 'unsigned long long') from 'void *' [-Wint-conversion]
        *out_pipe = NULL;
                  ^ ~~~~

The current ICC profile handling had a number of limitations:
1. Parts of the 3DLUT range were wasted in unrepresentable sub-blacks.
2. V2 profiles with black points not matching the V4 perceptual black
   target completely broke when using perceptual intent, causing
   black distortions.
3. We (intentionally) did not do any black point compensation.
4. Our approximation profiles were a poor fit for BT.1886-style
   technical gamma functions.

Improve the status quo by making two important changes:
1. Tune the approximation profile to use a BT.1886-style technical gamma
   curve tuned to the exact measured black point of the display.
2. Enable black point compensation

This should allow owners of broken colorimetric V2 profiles to rely on
perceptual mapping tables with black point compensation instead,

Closes: videolan/libplacebo#210
See-Also: https://github.com/mpv-player/mpv/issues/10442
See-Also: https://github.com/mm2/Little-CMS/issues/328

cmsDetectDestinationBlackPoint simply does not work in absolute
colorimetric intent. Work around by using relative colorimetric.

This is fine because this function only needs to produce approximate
results, only meant to inform upstream tone-mapping code - we generate
the 3DLUT with black point compensation either way.

Fixes https://github.com/mpv-player/mpv/issues/10532

Fixes: https://github.com/haasn/libplacebo/issues/136

Also bump the API version. Do this after updating all of the
public-facing headers, so users can put a dependency on it. (For reverse
dependencies, just use <5)

If only there was an actual way to mark macros as deprecated..

This turned into quite the accidental epic. But oh well, not worth the
effort to try and split it up again now.

Building fails if the user only has glad1 installed, but not glad2, with
this error: argument --api: Invalid api-string: "gl:core,gles2,egl"

And if the user has neither glad1 nor glad2, it just fails with the
unhelpful "module not found". Improve both by explicitly testing for
this in the build system.

Doing things the right way (tm)

This finally gives users complete control over the GL loading mechanism,
and also makes libdl optional.

Mirror the vulkan-link option by making the opengl-loader option
non-mandatory.

We need to interface with the glad loader manually for this test.

This re-enables all EGL functionality.

It's worth pointing out that this commit temporarily disables GLES2/EGL
support (and thus also the test framework) - to be re-added in the next
commit.

Closes videolan/libplacebo#208

Also introduces a new helper to facilitate accessing this dispatch table
in a "standardized" way from any sort of pl_gpu-based code.

For now, just use the global loader. I will add a `get_proc_addr`
callback to `pl_opengl_params` in a later commit.

Note: We don't really need to care about cleanup, because by default,
the `--mx` glad loader will simply dlclose() after returning from the
loader. (Though, confusingly, it still does so in a quite race-y way,
since it temporarily stores it inside a static variable for no reason
whatsoever)

In preparation for the migration to an implementation entirely based on
glad2, build against both glad and epoxy.

To facilitate the transition, import both headers, and suppress the
check designed to prevent this.

No integration yet, but this links glad (+ dependencies) into the
repository.

This design basically mirrors Vulkan's. This is pretty much the best
place to stick the dispatch table.

In theory we could just make the dispatch table itself a separate type
of internally exposed object (e.g. by overloading it on `struct priv`
somehow`, but there's really little point in hiding these internal
struct members.

This adds the compilation and link flags for things such as Vulkan, enabling
building with not everything being under a single sysroot.

This should fix the sporadic CI failures.

This argument is normally a void*, intended for things such as an X11
display handle, and so on. But EGL_MESA_platform_surfaceless
specifically requires it to be set to EGL_DEFAULT_DISPLAY, an integer
constant.

Make this cast to void* explicit to suppress errors on some platforms.

This needs a GLAPIENTRY to get the right calling convention on win32.