This version is slightly more lenient in also allowing discarding the
target if we're rendering to a region *larger* than it, and as a side
note, avoids re-normalizing the scissors.

This sort of prevents the error in 6519460a from occurring. In theory we
could still discard the framebuffer contents when rendering
non-transparent scissored subrects, but that's a separate check (and
arguably not even a good idea).

The existence of this parameter was _completely_ forgotten about. On
GPUs that actually respect the load operation, this results in any
shaders involving blending or scissoring to generate corrupted output.

Fixes #84

This fixes a failure to skip the cmat when the bit depth is identical
but only one of the two was specified.

Absolutely no reason to loop with 1ms polls here.

1) We have to restrict all image usage features that have corresponding
   format feature flags.

2) We should check for transfer_dst/color_attachment support *after*
   restricting the format features.

This somewhat uglily requires copying around the push consatnts into
pl_buffer_vars, but whatever. It'd be even uglier to try and somehow
avoid doing that.

Anyway, apart from reducing logic duplication, this (very importantly)
allows us to benefit from the offset sorting performed by this helper.

Works around a very real bug in SPIRV-Cross.

This is only done when actually emitting GLSL, as a way to avoid random
breakages when buffer vars are emitted out-of-order. (In practice, stuff
like this can break when e.g. emitting MSL via SPIRV-Cross)

The current code doesn't skip the cmat even if the only operation that
needs to be applied is the independent alpha handling.

Instead of only returning if the entire function is a no-op, emit the
parts of the code that can always be safely emitted and only skip the
cmat if it's a no-op.

The _m suffix must have been a freak oversight.

This should technically be based off the gles_ver, not
pl_glsl_desc.gles, even though the two are equivalent in practice.

Dunno why this was left as a TODO. It's a really easy thing to check.

This was never actually done, anywhere. In theory we might also want to
PL_DEBUG out the build configuration.

Ubuntu 18.04 is too old to have it.

This was completely forgotten by accident.

Two of these functions were not correctly guarded (either too strict, or
not at all), and also, the check for the extension itself was missing.

The division needs to happen before the multiplication. The bias was
also off by one.

The pl_glsl_desc can technically be overridden.

The user can technically override this version requirement, so use
SH_GLSL instead of SH_GPU. Also, check more explicitly whether or not
the pl_gpu was actually set, and clean up the gpu usages.

For debugging

Because OpenGL is now pretty decently tested and considered to be more
or less a first class citizen GPU, I think it's time to raise the major
version to commemorate its addition.

Turns out warning level 3 was *waaaay* too warn-happy. Seems to have
-Wextra and -Wpedantic and stuff.

Picked up by the stricter default warning level accidentally introduced
in d2411924.

These are pretty pointless

As suggested by a set of new warnings.

Oops. Maybe I should have used the `ci` branch to test my changes. If
only pushing to that branch didn't spam the IRC...

We're surprisingly close to the 30s timeout for llvmpipe/swrast

In particular, this also tests the static sh_lut mechanism.

We accidentally forgot to specify the correct profile, instead always
choosing compatibility profile.

Oops.

This makes the OpenGL test rather slow to execute, but oh well. I chose
to do it with OpenGL but not Vulkan because Vulkan and older GLSL don't
really make sense and probably don't really mix well either, since
vulkan semantics requires explicit layouts and stuff that needs newer
GLSL to parse anyway.

Closes #60.

These layouer qualifiers don't parse on older GLSL. So just omit them
and rely on it magically working.

Also mirror the capability restriction feature for pl_opengl

This seems like a reasonable thing to support. Also communicate to the
user what layers were successfully enabled.

This code was a bit lopsided out of laziness.

This document was still referencing ancient locations of stuff that got
moved.

The GPU CI instance still hangs inside the new environment for some
reason, but the main buildbots can handle it fine.