Turns out inverting a matrix was the largest chunk of the CoglMatrix
code. By switching to Graphene, a lot of it can go away. The inverse
is still cached in the CoglMatrix struct itself, to preserve the
optimization.
However, switching to graphene_matrix_t to calculate the inverse has
a challenge: float precision. We had to work around it here, and it
needs an explanation.
The way to detect whether a matrix is invertible or not (i.e.
whether it's not a "singular" matrix, or not) is by checking
if the determinant equals 0. So far, so good.
Both graphene_matrix_t and CoglMatrix use single-precision
floats to store their 4x4 matrices. Graphene uses vectorized
operations to optimize determinant calculation, while Cogl
tries to keep track of the matrix type and has special-purpose
determinant functions for different matrix types (the most
common one being a 3D matrix).
Cogl, however, has a fundamentally flawed check for whether
the matrix is invertible or not. Have a look:
```
float det;
…
if (det*det < 1e-25)
return FALSE;
```
Notice that 1e-25 is *way* smaller than FLT_EPSILON. This
check is fundamentally flawed.
"In practice, what does it break?", the reader might ask.
Well, in this case, the answer is opposite of that: Cogl
inverts matrices that should not be invertible. Let's see
an example: the model-view-projection of a 4K monitor. It
looks like this:
```
| +0,002693 +0,000000 +0,000000 +0,000000 |
| +0,000000 -0,002693 +0,000000 +0,000000 |
| +0,000000 +0,000000 +0,002693 +0,000000 |
| -5,169809 +2,908017 -5,036834 +1,000000 |
```
The determinant of this matrix is -0.000000019530306557.
It evidently is smaller than FLT_EPSILON. In this situation,
Cogl would happily calculate the inverse matrix, whereas
Graphene (correctly) bails out and thinks it's a singular
matrix.
This commit works around that by exploiting the maths around
it. The basis of it is:
inverse(scalar * M) = (1/scalar) * M'
which can be extrapolated to:
inverse(M) = scalar * inverse(scalar * M) = M'
In other words, scaling the to-be-inversed matrix, then
scaling the inverse matrix by the same factor, gives us
the desired inverse. In this commit, the scale is calculated
as 1 / (smallest value in the diagonal).
I'm sorry for everyone that has to read through this :(
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
Next commits, and this patchset in general, will make this patchset
obsolete, since it'll only test graphene types against each other.
If at all useful, the Euler test should be moved to graphene.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
Use dot products to simplify calculations. Because the 'w' column of
the matrix is always summed, use 1.f in the 'w' component of the point
vector.
Because CoglMatrix is column-major and graphene_matrix_t is row-major,
it is necessary to transpose the matrix before retrieving the rows.
When we switch CoglMatrix to be row-major, this transposition will
go away.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
This is pretty similar to the other conversions, except we need to
store the matrix flags before operating on it, and update it using
this old value after. That's because cogl_matrix_init_from_array()
marks the matrix as entirely dirty, and we don't want that.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
At this point, we are still only changing CoglMatrix APIs internally, and
it should still produce the same output as before.
To achieve this, using graphens matrix implementation, we need to exploit
some knowledge about conventions used in Cogl and graphene respectively.
In Cogl, transformation matrices are equivalent to those of affine
transformation matrices. The convention used by graphene, however, is to
operate on matrices that are transposed compared to their affine
counterparts.
So for example, let's say we want to multiply the affine matrices A and B,
to get C.
A × B = C
The first step is to convert A and B to graphene matrices. We do this by
importing the floating point array, importing it directly using graphene.
Cogl exports its matrix to a column major floating point array. When we
import this in graphene, being row major, we end up with the same matrix,
only transposed.
Cogl Graphene
A <===> Aᵀ
B <===> Bᵀ
We then multiply these imported matrices in reverse
Bᵀ × Aᵀ
which in turn, due to ABᵀ = BᵀAᵀ, gives us
Bᵀ × Aᵀ = (A × B)ᵀ
Our original goal was to find C, thus we know that
A × B = C
That means we can shuffle things around a bit.
A × B = C
Bᵀ × Aᵀ = (A × B)ᵀ
Bᵀ × Aᵀ = Cᵀ
With the same conversion as done when going from Cogl to graphene, only
the other way around, we still end up effectively transposing the matrix
during the conversion.
Graphene Cogl
Cᵀ <===> C
Thus when converting Cᵀ to Cogl, we in fact end up with C.
(Explanation authored by Jonas Ådahl)
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
Graphene provides skewing as part of graphene_matrix_t API, and it'll
be easier for the transition to just expose similar API surfaces.
Move the matrix skew methods to CoglMatrix.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1439
During seat initialization, we process early libinput events (adding all known
devices) before the seat gets a stage assigned. This causes warnings when trying
to handle the corresponding CLUTTER_DEVICE_ADDED events, as they are sent
stageless.
As it is definitely too soon to have those events sent meaningfully, filter
those events out and instead handle the CLUTTER_DEVICE_ADDED emission for all
known devices after the seat receives an stage. This makes the events guaranteed
to be emitted early in initialization, but not so soon that they can't be
handled yet.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1472
Mutter itself is versioned now, so passing the version information
to the plugin is redunant now: The version is already determined by
linking to a particular API version (gnome-shell) or by installing
to a versioned plugin path (external plugins).
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1473
A Meta.WaylandClient() object has a GSubprocessLauncher object
passed externally. Currently this object is kept while the
WaylandClient object exists, but is is only needed until the call
to spawn is made.
This patch frees that GSubprocessLauncher just after that call,
thus freeing those resources.
Fix https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1462
This reverts the commits 372d73e275 and 1d20045247 - the special
case for alpha-less textures could only happen on Wayland, but now
the opaque region is also set in those cases.
This commit saves us some allocations, simplifies the logic a bit and
makes sure culling uses the same opaque region as our painting paths.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1463
Wayland clients using buffers without alpha channel are not expected to
set an opaque region. However, we rely on the opaque region for the fast
painting path in `MetaShapedTexture`.
Thus, make sure to always set an opaque region internally in those cases.
For X11 clients, wo do so already.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1463
Just like we used to before 30809665d8.
Because in some cases `clip_region` is able to shave off an extra pixel
from the edge of the redraw rectangle(s). And not shaving that off was
making the background rendering inconsistent with shaped-texture, causing
occasional off-by-one artefacts. Now both shaped-texture and
background-content agree on the clip region again that doesn't happen.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1443https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1464
This is essentially a revert of
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/326. This commit
had the unintended side effect that the built sources are actually
rebuilt for every individual user of libmutter_dep. With there being more
tests and generated files, the number of targets to build is increasing
squarely.
Not doing this reduces the number of targets from 2044 to 874, thus
saving man hours and CI burnt cycles in the long run. There's the slight
risk of reintroducing the random build breaks, but mutter is essentially
doing as suggested at https://github.com/mesonbuild/meson/issues/1084
(the only difference being addressed in the previous commit), so meson
ought to behave as expected.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1458
We only update the last device from actual input interaction here,
avoid this pair of events. This is specially nasty with
CLUTTER_DEVICE_REMOVED, since the device we're notifying upon will be
disposed soon after emission.
https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/1460
Like it's done for the pointer in other places. Without a stage assigned,
some bits (like IM handling) may end up with events ignored, and misbehave.
Fixes: https://gitlab.gnome.org/GNOME/mutter/-/issues/1413
