As an aside, we also make sh_subpass not explicitly spam/fail the parent
shader in this case.

Rather than this merely representing an "in-flight" image, with the
img->sh only living for as long as this exists between different pass
types, `img` is now conceptually persistent and either in one of two
modes: `sh`, or `tex`. This allows us to, in principle, avoid doing
redundant FBO roundtrips for cases where the previous pass thinks the
next pass needs a tex but the next pass actually needs a sh, such as is
currently the case for the AV1 grain shader.

Since `pass_hook` in particular can randomly mutate `img` to either of
the two forms, callers must now be somewhat vigilant to make sure they
always use `img_tex()` and `img_sh()` to access the "current"
shader/texture, rather than relying on local variables staying
persistent.

The use of locally initialized pl_shader is now exclusive to passes that
keep their own pl_dispatch_begin calls (for various reasons).

The current approach was to pair each FBO with its use, "semantically".
The intent was to minimize the number of "reallocations" that would be
required if the number of passes changed dynamically (e.g. as the
renderer options changed). However, this is not only an unrealistic
design goal to optimize for (users can use separate pl_renderers for
wildly different purposes, and for a single conceptual video stream it
doesn't really matter), but also, it gets in the way of a planned
refactor I have concerning `struct img`.

Change this to make all FBOs dynamically allocated. The current
implementation simply uses a counter, but a more advanced implementation
could use a pool of textures and find ones that have matching sizes
before recreating ones that don't. The API shouldn't change as a result
of this.

In doing so I finally hit the time bomb caused by assuming blacklisting
compute is only needed on that specific driver version. Turns out, the
same issue is present even on newer driver versions. Since I have no
idea how else to work around and/or debug it, just permanently disable
compute on the CI.

Unfortunately, for some reason, the `shaderc` version in this version of
the CI image hits random internal exceptions when trying to compile
pretty much anything. But using glslang directly works. Except for msan,
because we don't have msan-instrumented libc++.

Some other changes needed for whatever reason.

These don't generate any errors, but the compilation status still isn't
"success". Treat these as errors as well, in terms of logging.

There's really no reason not to. Also clarify that these functions are
not, in fact, "mandatory" instance-level function pointers.

This extension was treated as global in the past, but now that we have
a proper extension framework we should just handle it like any other
extension. Solves a very concrete issue where dependencies on e.g.
VK_EXT_hdr_metadata were not satisfied if the user did not happen to
enable this extension.

Also check if the extension is loaded when we attempt actually creating
a swapchain.

Forgot to un-mark timers as recording. Probably mostly benign, but could
hypothetically cause issues.

With ca1ebc, the mismatched shaders of the variety that 092229 was
intending to solve should no longer be a possibility.

If this is still the case, it's probably a bug. Assert instead.

This function already exists.

The current logic could end rounding the plane up in both directions,
inadvertently introducing an upscale by 1 pixel to the refplane, which
not only forced an FBO indirection but also lowered the quality due to
the effective resample. Furthermore, the code for adjusting the rect by
the `rc` was wrong, since it failed to scale the "affine" part of the
transformation down to the coordinate space of the plane.

Simplify and fix this by only rounding the offset (making sure to always
round towards 0 to have the best chance of "doing the right thing"), and
also correctly scaling this offset when calculating the offsets for the
individual planes.

This can happen if e.g. the shader's output image is rounded up relative
to the rect actually being sampled, for example when when users are
sampling fractional parts of the image.

Across the boards, "output sizes" and so forth are assumed to be
integers, but a lot of the code generating them is based on floating
point math. To make it clear what happens, make things consistent by
using roundf() instead of implicit conversion (which may end up
truncating etc.).

The order these are written in currently means the second check never
has a chance to get printed, because SAMPLER_NOOP implies
SAMPLER_DIRECT.

In theory there are various ways we could reconcile this difference, but
for now, the easiest thing to do is to simply disable it.

I'll try working on a way to bring back support for this, but I wanted
to get this fix out of the way first.

09834c52 introduced a regression here that caused an unnecessary FBO
indirection to occur for the (relatively common) case of user shaders
consulting the dimensions of OUTPUT in their RPN expressions.

OUTPUT apparently behaves weirdly. mpv should probably document this
better. Even better would be to have had different names for the OUTPUT
hook stage and the OUTPUT magic constant for RPN exprs. But it's
probably too late for that now. Unless we want to rename the OUTPUT
stage, which we totally could do.

Also add a TODO for something I realized while staring at this code
again.

This shader is very definitively non-resizable. Flag it as such.

This is a bit annoying, and arguably such shaders are broken to begin
with because there's no way you can possibly rely on this logic without
spamming warnings.

But it represents a difference in functionality compared to mpv, so we
implement this just to be on the safe side.

This is `int` and not `size_t`, which may cause issues on platforms
where the two are not the same size.

To make sure these messages see the light of day before the application
gets a chance to crash.

Would have made debugging https://github.com/haasn/libplacebo/issues/74
easier.

These were never updated for 7ac80fb8

PL_VERSION can be a bit unhelpful for git commits

A simple oversight. Fixes e.g. Anime4K

Rather than tying struct img to the specifics of pass->cur_img, I wanted
to loosen this up a bit and re-use it for the plane state. To this end,
I parametrized one key function, `pass_hook`, by the actual `img` to
use, and got rid of some redundant fields.

This commit doesn't (or shouldn't) change the behavior yet, it's mostly
to disentangle the behavior-changing bits of this refactor from the
merely refactoring ones.

Large slab allocations failing will end up segfaulting here

Still not entirely sure why I renamed it if I'm just gonna add a
compatibility alias anyways.

But whatever. It's starting to get to that point where I have to be at
least slightly worried about breaking downstream.

pl_color_space_infer hard-codes this as BT.709, but we could/should be
using pl_color_primaries_guess instead.

This refactor accomplishes a number of goals:

- make the pl_shader_av1_grain sample directly from `tex` instead of
  relying on pre-sampled tex, which allows it to correctly keep track of
  the texture scale even when sampling from the luma tex, and also gets
  rid of redundant fields
- makes the grain_state per-plane rather than shared, which avoids
  thrashing the grain state cache due to changing subsampling index
- fixes the channel order when applying grain to non-YCbCr content
- fixes some bugs related to grain scaling w.r.t texture sample depth

Tested to produce correct output on a 4:2:2 10-bit grain sample.

There's still one bug but that requires a near-complete refactor to
solve, coming up next.

Helps tracking down why this state might be getting thrashed.

Add a new `bool dynamic` to hint at the sh_lut that the LUT is going to
be changed frequently. This makes it use host-writable textures rather
than immutable textures, which may set different performance hints in
the pl_gpu implementation.

Since we normalized before applying grain, the grain_params needs to
contain the un-normalized repr.

Separated scaling of e.g. 10-bit chroma content blew out the chroma
channel into bright pink due to the scale accidentally being applied for
both passes of the pl_shader_sample_ortho.

In practice the pre-grain texture is going to be a different pl_tex than
the post-grain texture, so the "order" of operations does not really
make sense in this context.

The pl_3dlut_default_params was undefined in this case, but the renderer
still referenced it.

Niklas Haas authored 4 years ago and Niklas Haas committed 4 years ago

Based on ITU-R Report BT.2408, and general recommendations within the
industry, the "SDR white level", i.e. the level at which to map SDR into
HDR signals, is not 100 cd/m^2 but a value closer to 203 cd/m^2.

For PQ signals this results in a relatively straightforward change of
the code, but for HLG signals things get more complicated. For HLG,
rather than targeting a fixed brightness in cd/m^2, the recommendation
is to map SDR white levels to the 75% point in HLG, which calculates to
a value of about roughly 3.17955 in scene-referred space (where the
nominal peak is at 12.0). To fit this into the libplacebo interpretation
of these values (where 1.0 maps to the SDR white levels), we scale
things down by this factor, giving rise to a new scene-referred signal
range of 12.0/3.17955 = 3.774, and adjust the OOTF to compensate.

This commit does put into question to what extent the default tone
mapping settings need to be altered to account for this change in
interpretation when going from HDR back to SDR.

Also add tests to ensure this stuff round-trips.

Bump the API version because, apart from the fact that we changed a
public header, this is quite a drastic change in functionality.

This reverts commit 9c0d88fb.

The xta_ref is not parented, it's attached to the ref it's counting.
Should probably improve the documentation here or something.