In many cases of the switch, we do not need the vertex data. This moves
the creation of the vertex_data array into a secondary function and only
calculates it the cases for which it is required.
We were putting big glyphs in the cache, in their
own texture, but forgetting to mark the texture
as permanent, so it could be reused, leading to
occasional misrendering. Fix this by marking these
textures as permanent, and explicitly freeing them
when the cache entry gets old.
Every few frames, we do extra work for the
cache aging. Arrange for the glyph and icon
caches to not cause extra work on the same
frame, to smooth things out.
There is no need for us to be very precise about
aging the glyph entries. It is enough to check
occasionally and mark old entries. This reduces
the overhead of work we do every frame on the
caches, at the cost of letting glyphs linger
a bit longer in the cache.
Make this function more similar to the icon
cache equivalent, and simplify it a bit. We
don't use the boolean return, and we don't need
to look at the age of entry when marking it
used.
Remember which atlases were removed, and only
check those when looking for icons or glyphs
to remove. For most frames, we don't have to
check at all since no atlases were removed.
Instead of copying the (rather large) RenderOp to the GArray, we can
simply set the fields directly in the allocated space for the struct.
In most cases, there wont be any allocations to make as the array size
is kept in tact across frame renderings.
We can just use memcmp here because even in the use of lookup keys with
C99 initializers, we can rely on any space between fields added by the
compiler to be zeroed. So we might as well use wider memory cmopares.
UNDEFINED initial layouts may not preserve the contents
of the attachment after transitioning the layout. We want
them to be preserved because we do partial rendering.
Use GENERAL as the initial layout for render passes.
Multiple images in the before barrier array are defined with
VK_ACCESS_TRANSFER_WRITE_BIT and VK_ACCESS_TRANSFER_READ_BIT,
which requires passing VK_PIPELINE_STAGE_TRANSFER_BIT and
VK_PIPELINE_STAGE_HOST_BIT to vkCmdPipelineBarrier().
Pass these flags correctly.
We can't just assume that the pointer we'se using as a cache key will
stay unique forever. The texture might be freed, and a later allocated
texture might have the same addres now, causing the cache to return
incorrect results.
Pass the glyph position into the glyph caching functions,
not just the glyph index. This allows us to cache different
images for different subpixel positions.
Pass the glyph position into the glyph caching functions,
not just the glyph index. This allows us to cache different
images for different subpixel positions.
We are currently using a weird mix of pango and cairo,
but there is no need for us to go through a pango
renderer here; we can just use cairo directly.
E.g. anything involving a scale. This is important when e.g. scrolling
in the node list in the recorder, which scales every recorded node down
to fit in the list.
Use the same texture atlases to back both
the glyph and icon caches, and unify their
sizes and management. Store big glyphs
in separate textures, so all atlases have
the same size. Tweak some of the eviction
parameters.
We share the caches across all GL contexts
on a display, unless the GSK_NO_SHARED_CACHES
env var is set.
The logic here seems faulty. We want to keep
a timestamp that tells us when the glyph was
last used, so always update the timestamp.
And whenever we use a glyph, it turns 'young'
again, so remove it from the old pixels
accounting.
The (MAX_AGE, MAX_AGE+CHECK_INTERVAL) interval
is only relevant to prevent us from turning
a cached glyph old more than once, and that
is already taken care of.
If somebody does a transform like
scale(5) scale(10) translate(1,1) translate(5,0)
store it instead as
scale(50) translate(6,1)
This way, less memory is consumed and transforms are easier to read.
In particular, this simplifies the typical transforms we do in GTK,
which are just one translation after another.
We don't need to just look at the scale of the new modelview matrix, but
at the one we get when multiplying the new one with the current one.
Test case attached.
Use cairo-script-interpreter to parse the scripts that generate cairo
nodes.
This requires libcairoscriptinterpreter.so to work properly, but if
it isn't found we disable this (unimportant for normal functioning)
code and just emits a parser warning.
The testsuite requires it however or it will fail.
A new test is included that tests all of this.
CSS does not do exponents, so printing numbers close to 0 as 1.234e-15
does not work.
Also up the accuracy to 17 digits because that's what everyone else
uses.
We blindly assume everywhere that a single glyph will definitely fit on
one atlas, but that's not always the case.
For now, don't crash or produce GL errors.
Instead of only allowing for glyph indexes, allow ASCII characters as
replacements. So this glyph sequence
glyphs: 65 8, 66 8, 67 8
Can be replaced by
glyphs: "ABC"
provided that the glyph for "A", "B" and "C" are 65, 66 and 67
respectively and their advance is exactly 8.
x offset and y offset must always be 0 and every glyph must start a
cluster.
Update to the docs outlined in #1887.
In particular, the changes do:
1. Require no property, have a working default for everything
2. Be clear about what gets printed and how.
Tests ahve been adapted to still pass.
Base the rewrite on testsuite/css/parser/test-css-parser - we now
require the node file to match a reference node and track the errors it
triggers.
We also no longer use gtester.
When printing, behave the same way as when parsing:
Magically skip a container node if there is one - just like the
parser magically creates a container node to hold all the nodes
it parses.
We can't just pretend we have an identity matrix when we are actually
scaling. This fixes the node editor sometimes not drawing things when
rendering to a texture. We were mistakenly discaring render nodes
because the bounds transformation was wrong.
We stuff both gl-drawn and cairo-drawn textures into the same cache, so
we can't really assume that we need to draw any of them flipped or not.
Fix this by drawing fallback stuff upside down and then using
upside-down vertex data for everything.
Fixes#1897
We don't want to return a GFile because GFile can't handle can't deal
with data: urls.
That makes the code a bit more complicated that doesn't deal with those
URLs, but it makes the other code actually work.
GtkCssImageUrl also now decodes data urls immediately instead of only at
the first load. So don't use data urls if you care about performance.
Instead of encoding the raw data, encode the full image to a PNG.
And instead of stuffing that encoding into a string, use a full
data: url.
And then remove the width and height properties, because they're now
implicitly included in the data.
And then change the parser to match.
And because the parser now parses regular urls on top of data: urls, we
can now load any random file.
Instead of the previous approach using GVariant, this new approach uses
human-readable text files as the serialization format for render nodes.
The format is a custom one, but it is inspired by QML and conforms to
the CSS syntax. Because of that, we can use the CSS machinery from GTK
to parse it, and in particular share code to parse properties that GTK's
CSS machinery also supports, such as colors.
This commit breaks all existing usages of node files - such as the
testsuite and various test tools - they will be fixed in further
commits.
Change the way we compute border color cutoffs to the same method that
browsers use. This method does not consider the corner sizes at all and
only looks at border-width.
Previously, when borders were too big - ie when a 100x100 rect had only
one 100x100 border, like the black part of ◔ - and then shrinking this
rect by 25px on either side, we'd end up with a 50x50 rect with a 75x75
border, and that's obviously not correct.
Floating point values cannot ever be compared for equality. GLib has a
G_APPROX_VALUE macro that lets us compare two value within a provided
precision, so we should use that instead.
Artisanal, homegrown, locally sourced, vegan reference counting has been
replaced by the appropriate API in GLib, which does small things like
saturation and type checking.
Apparently genTextures and friends only "reserves names", initializing
them will actually create them. Using glObjectLabel on textures before
initializing them will throw a GL_INVALID_VALUE error.
When rendering to a texture, collecting the render ops might bind a
different framebuffer, so bind the one we want again before doing the
actual rendering.
This adds debug groups in various places, including the debug
nodes if those are in use. This makes the traces in tools like
renderdoc much easier to read.
GL keeps the unoform state per-program, but not per-frame. So, we can't
pretend that this works for us. Keep the RenderOpBuilder around for the
entire lifetime of the renderer instead.
This fixes rendering to a texture on intel hardware. The glClear calls
would throw a GL_FRAMEBUFFER_INCOMPLETE error here, because the
gsk_gl_driver_begin_frame() call in do_render() reset the framebuffer
object in use.
This fixed the reftest introduced in the previous commit.
I'm using a mesh gradient here instead of drawing 4 individual sides to
avoid artifacts when those sides overlap in rounded corners.
We don't want the new transform while drawing things on a texture.
Instead, only apply the new transform matrix when adding the final
texture drawing ops.
This fixes the stack cube rotation transition to at least look somewhat
better.
When sending render nodes from the client to the daemon we add an id,
and whenever we're about to re-send the entire tree node we instead
send the old id. We track all the nodes for the previous frame
of the surface this way.
Having the id on the daemon side will allow us do to much better deltas.
gsk/gskenums.h:181: Error: Gsk: multiple "@GSK_TRANSFORM_CATEGORY_2D" parameters for identifier "GskTransformCategory":
* @GSK_TRANSFORM_CATEGORY_2D: The matrix is a 2D matrix. This is equivalent
^
gsk/gsktransform.c:1342: Warning: Gsk: gsk_transform_to_2d: unknown parameter 'm' in documentation comment, should be 'self'
gsk/gsktransform.c:1368: Warning: Gsk: gsk_transform_to_2d: invalid return annotation
gsk/gsktransform.c:1461: Warning: Gsk: gsk_transform_to_translate: unknown parameter 'm' in documentation comment, should be 'self'
This reinstates diffing in the same way that it worked for offset nodes.
It would be possible to add diffing for affine transforms or even all
transforms, but I think this is unnecessary right now - and also quite
expensive to compute.
Make the API expect a tranform of the proper category instead of
doing the check ourselves and returning TRUE/FALSE.
The benefit is that the mai use case is switch (transform->category)
statements and in those we know the category and don't need to check
TRUE/FALSE.
Using the wrong matrix will now cause a g_warning().
... instead of computing it every time we need it.
This should be faster and we want to use it a lot more prominently.
Also, we have the struct memory available anyway.
In particular, add a per-category querying API for the matrix:
- gsk_transform_to_translate()
- gsk_transform_to_affine()
- gsk_transform_to_2d()
- gsk_transform_to_matrix()
This way, code can use the relevant one for the given category.
Since we can do partial redraws, dropping every shadow that's been
unused for one frame happens too fast. This is also a problem when a
shadow gets drawn on a texture for a few frames.
This can happen for certain transform nodes. The transform node's
child's bounds are fine, but the transform node bounds are all nan.
Just ignore those bounds since we can't meaningfully render them anyway.
If the given matrix is explicitly of category IDENTITY, we don't need to
do anything, and in the 2D_TRANSLATE case, just offset the child bounds.
Those are the two most common cases.
The code didn't change, it was just shuffled around to make the
with_bounds() versions of the text rendering unnecessary and instead
pass through the generic append_node() path.
They were a neat idea while they lasted. But now, it's time for
categorized transform nodes, where matrices with
GSK_MATRIX_CATEGORY_2D_TRANSLATE are the exact replacement.
Renderers have not been adapted for this purpose, so they (continue to)
run slow paths.
The first set of glyphs is created with a timestamp of 1. Later we
subtract the glyph timestamp from the cache timestamp, meaning we end up
with numbers ending in 9, e.g. 59. Now unfortunately !(60 <= 59), so we
do not end up incrasing the old_pixels count of the cache. Later we then
call lookup() and DEcrease the old_pixels count, which makes the
unsigned int wrap and cause a huge old_pixels value, which causes us to
drop the cache.
The @filename@ directive will use the full path of the file being parsed
for enumeration types; we should use @basename@, instead, as it improves
the reproducibility of the build by using only the file name.
Some of the flags got lost in the meson transition or were demoted from
error flags to warning flags.
This commit reintroduces them.
It also includes fixes for the code that had warnings with those flags.
The big one being -Wshadow.
This broke the overlay blur demoe when resizing the window to a size
that would completely move the image below a button, causing the
GtkSnapshot code to remove the clip node below the blur node.
So we can check that the currently set clip is the first one and now
intersect with it. This first clip is always the entire viewport or the
entire render_area and we don't want to end up drawing things to a
texture because of it.
Instead of getting the translation x/y everytime we use the modelview,
get it once, when extracting the metadata. Do the same with the scale.
And save if the matrix is "simple" at all, i.e. if it only consists of a
translation and/or scale. This will be helpful later when we start
drawing transformed nodes on textures.
We do this for every single node, which is a little costly, especially
since the common case for the modelview matrix these days is a simple
translation. So, check whether the new modelview matrix is only a
translation matrix and if so, don't do a full matrix multiplication per
node.
Some of the _diff implementations did a whole bunch of work just to
throw it away afterwards and invalidate the entire union of the two
render nodes, most notably the two clip nodes. Fix this to only call
gsk_render_node_diff_impossible if the previous if-condition is FALSE
and not always.
This reverts commit 8e74eb382f.
This code is not necessary. It worked around a bug in graphene where
graphene was requiring stricter alignment than glib allocators could
guarantee.
Any data that is later fed to graphene must be
allocated with proper alignment, if graphene
uses SSE2 or GCC vector instructions.
This adds custom array code (a streamlined copy
of GArray with all unnecessary bells and whistles removed),
which is then used for the state_stack instead of GArray.
There's also a runtime check for the size of GtkSnapshotState
itself being a multiple of 16. If that is not so, any array
elements past the 0th element will lose alignment.
There are probably struct attributes that can
make GtkSnapshotState always have size that is a multiple
of 16, but we'll burn that bridge if we cross it.
The code is mostly stolen from graphene.
Allocators support any alignment, but their implementation
only calls system aligned allocator functions if malloc()
is not aligned to 16-byte boundaries. If it is aligned,
the implementation just calls malloc() regardless of which
alignment is requested by the caller.
This can be fixed by saving the result of meson malloc()
alignment check and adding a few conditions to the implementation,
but right now GSK and GTK only need 16-byte alignment either way.
1. Include the broadway renderer (so we can test it properly fails on
Wayland or X11)
2. List all potential renderers, print useful information when Vulkan
is not compiled in instea dof omitting it
3. Improve docs
As they require a draw context and the draw context is already bound to
the surface, it makes much more sense and reduces abiguity by moving
these APIs to the draw context.
As a side effect, we simplify GdkSurface APIs to a point where
GdkSurface now does not concern itself with drawing anymore at all,
apart from being the object that creates draw contexts.
Previously, we got the damage, then computed the changed area, then
started a frame with that changed area.
But starting a frame computes the damage for us.
So now we start a frame, then get the damage area from that, then
compute the change area.
This does nothing but disallow passing NULL to gdk_surface_begin_paint()
and instead require this context.
The ultimate goal is to split out Cairo drawing into its own source file
so it doesn't clutter up the generic rendering path.
And of course, gsk_render_node_get_name() is gone, too.
The replacement is of course debug nodes.
As a side effect, GskRenderNode is now *really* immutable.
