So far, there was one common loop for vertical flip, and six loops per
bit depth for each other (non-trivial) transform, or 19 different inner
loops in total via macro expansion.

This keeps the common loop for vertical flip and reduces the other down
to two per bit depth: transposition and horizontal flip.
Antitransposition and the three rectangular rotations are obtained by
composing the previous two transforms with vertical flips before and/or
after the main transform. Those vertical flips are achieved for "free"
by using negative strides.

This is a relic from whence the user-specified transform could be
handled solely by the transform filter whence this code originates.
Now the display modules can accept 4:2:2 chroma subsampling *and*
handle the orientation at the same time and without silently
down-sampling the Cb and Cr components to 4:2:0.

Steve Lhomme authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

After a flush the streamStarted is reset, a new call to MFT_MESSAGE_NOTIFY_START_OF_STREAM
has to be done (ie startStream). It's not necessary to do it on exit.

This is only useful for async MFTs.

Steve Lhomme authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

As it's not always done on exit, we have to release all output the decoder may
have produced.

Steve Lhomme authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

The MFT_MESSAGE_NOTIFY_END_STREAMING actually does some D3D cleaning as well.

+ don't crash if DoRelease was called due to an error.

+ don't reset the input/output types anymore.

Steve Lhomme authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

Sometimes we get that kind of format before the packetizer decides what the
format is.

Steve Lhomme authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

They are not mandatory and just help to prepare the internal buffers. This will
have to be done no matter what so it doesn't matter when. Let the MFT do it on
their own.

No matter how you look at it, this did not belong in the image adjust
filter. This could belong in a separate module, or as a submodule of
the colour threshold filter -- if somebody actually ever needs it.

This function was never implemented by the hardware-accelerated adjust
filters, as it is indeed not a standard part of HSL and HSV adjustment
functionality, or really at all a part of it. So this has actually
become a functional inconsistency.

Rémi Denis-Courmont authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

The registered callback does nothing with the given variable name.

Rémi Denis-Courmont authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

The registered callback does nothing with the given variable name.

Pierre Lamot authored 3 years ago and Hugo Beauzée-Luyssen committed 3 years ago

the animation was running and forcing the scenegraph to render even when the
seekbar (in the miniplayer) is hidden, this reduce CPU load when the application
is idle.

There is no memory content to synchronise here. The only needed
guarantee is for sine and cosine to be consistent with one another.
The atomic variable readily provides for that.

Lyndon Brown authored 3 years ago and Steve Lhomme committed 3 years ago

These options have null shorttext labels because they are not meant to show
up in GUI preferences. They were not configured in any way to be hidden
from GUI prefs though. Being available in GUI prefs yet having a null
shorttext pointer breaks the expectations of the GUI prefs code resulting
in a null-deref crash.

(The crash occurs upon simply selecting these plugin nodes within the
advanced preferences tree, or upon opening the new upcoming 'expert'
preferences interface).

The most appropriate solution is to mark these developer options as
private. Since there is actually no need for these options to be saved in
the saved settings file, I have marked them as volatile, which also implies
private.

Lyndon Brown authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

 1. Should be a value like `"any"` as with the other module selection
    options to avoid mismatch with the choice list.

 2. Per review, the default wanted for this should actually be `"none"`.

Pierre Ynard authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

Since last month, requests without a valid referrer get rejected with an
HTTP 403 error.

We might want to consider that the lua script API should track and pass
such referrers by itself.

Pierre Ynard authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

The descrambling function is now called through an intermediate array
variable. This change has also added two extra ways to recover the
function name. Add support to parse and resolve any of them.

Fixes #26574

As per the AArch64 ELF ABI, object files that support BTI need to
advertise it via a dedicated Linux ELF object file property in an ELF
note section. The linker can then check if all input object files
support the feature to advertise it at the shared object or executable
level.

In the case of BTI, this is necessary to enable the guarded page bit
in the memory mapping for BTI-capable .text sections.

This defines a macro to populate the note section and uses it.

This was superseded by fc8208ec.

Pierre Lamot authored 3 years ago and Rémi Denis-Courmont committed 3 years ago

re-initialising the context was causing severe performance degradation (hangs,
CPU consumption & RAM) in the snap version (at least) when KDE QuickStyle was
used.

Proper handling of errors and interrupts of course requires that the
target only exists on succesful completion. .DELETE_ON_ERROR cannot be
used for directories. The contribs source README indeed has required to
keep $(MOVE) for as long as the current incarnation of contribs has
existed.

This hack is no longer necessary anyway as per the previous change.