When we clean up the uniform allocations after a frame,
it can happen that our space requirements actually increase,
due to padding that depends on the order of allocations.
Instead of asserting that it doesn't happen, just make
it work by growing our allocation.
Fixes: #3853
We need to use __cpuid() to check for the presence of F16C instructions on
Visual Studio builds, and call the half_to_float4() or float_to_half4()
implementation accordingly, as the __builtin_cpu...() functions are strictly
for GCC or CLang only.
Also, since __m128i_u is not a standard intrisics type across the board, just
use __m128i on Visual Studio as it is safe to do so there for use for
_mm_loadl_epi64().
Like running on Darwin, we cannot use the alias __attribute__ as __attribute__
is also for GCC and CLang only.
When 9-slicing shadows, omit the center tile when it is
entirely contained in the outline (that is not always
the case, depending on corners and offsets).
gsk_rounded_rect_contains_rect was calling
gsk_rounded_rect_contains_point, which potentially
checks all four corners, for a total of up to 16
corner/point checks. But there is no need to do
more than 4 such checks to answer the question.
Opportunistically use the coloring program for
drawing underlines instead of the color program.
This avoids program changes in the middle of
text.
For the Emoji text scrolling benchmark, this reduces
the program changes per frame from > 1000 to around 100.
Use an IFUNC resolver to determine whether we can use
intrinsics for FP16 conversion. This requires the functions
to be no longer inline.
Sadly, it turns out that __builtin_cpu_supports ("f16c")
doesn't compile on the systems where we want it to prevent
us from getting a SIGILL at runtime.
We only have one shader that uses the color2 attribute,
and it doesn't use the uv attribute, so save vertex
memory by putting those in the same space.
This reduce the per vertex space from 32 to 24 bytes.
This reduces the size of our Vertex struct from
48 to 32 bytes. It would be nicer if we could store
the colors in fp16 format in the rendernodes, and
avoid conversion here. But this is still good.
Move the resources of each renderer to its subdirectory.
We've previously done that for the ngl renderer, but it
is better to be consistent and do it for all the renderers.
Arrange things so that non-child parameters
are always printed before the children. This
greatly helps with readability, which really
suffers when there's hundreds of lines of indented
children between the node start and its parameters.
Update all affected tests.
Instead of rendering the unclipped child to a texture
(and risking blowing the texture size limit, and bad
downscaling), just render the clipped region, and live
with the fact that we can't cache the rendered texture.
This avoid bad artifacts when scrolling long textviews
in rounded clips.
There was confusion here about the handling of the
modelview transform. The modelview transform we are
getting is already set up for rendering the node
we are given, so keep it - except for possible adding
an extra scale on top when the texture would otherwise
be too big.
Move some work out of the loop in visit_text_node.
This takes advantage of the fact that the yoffset
of most glyphs is zero, so yphase generally does
not change in a line of text.
Allow comparing container nodes to any other
node, by pretending the other node is a single
child container (if it isn't one already).
This fixes a glitch where we redraw the full
entry text when the blinking cursor goes to
opacity 0, since GskSnapshot then optimizes
away first the opacity node, and then the
single-child container.
Previously, we translated the uniform key (an enum) into a location within
the shader program in GskNglProgram. A number of performance improvements
were focused around having low nubers for the uniform locations. Generally
this is the case, but some drivers such as old Intel drivers on Windows
may use rather large numbers for those.
To combat this, we can push the translation of uniform keys into locations
at the GskNglUniformState level so that we work with unranslated keys
through the process until applying them.
Fixes#3780
The effectiveness of the front cache is limited by
subpixel positioning making it very likely that we
will meet the same glyph in different x phases inside
a single line of text.
Factoring the xphase into the front cache key makes things
better. For the string eeeeeeeeeeeeeeeeeee
before: 0% front cache hits
after: >90% front cache hits
We don't want to be responsible for duplicating the effort of the hash
table, we just want to speed up subsequent lookups. Otherwise, we risk
not marking glyph usage when tracking usage for compaction.
This required finishing up the begin_frame/end_frame semantics for
GskNglTextureLibraryw which was apparently overlooked.
The driver was changed to provide more information to the library when
beginning frames. We do not need to use end_frame so that was removed.
The frame age is the same as GL (60) but I do wonder if that is based
on seconds if we should be using something longer for situations where
we have higher frame rates.
Fixes#3771
If cairo is a subproject, it's not necessarily installed when gtk
is built. In the build tree, libcairo-script-interpreter is not stored
in the same directory as other cairo libraries.
Recognize a common pattern: A rounded clip with
a color node, followed by a border node, with the
same outline. This is what CSS backgrounds frequently
produce, and we can render it more efficiently with
a combined shader.
Now that colors aren't uniforms anymore, we don't
win much by using the inset_shadow shader. The fragment
shaders of inset_shadow and border are identical. And
the regular border setup does nine-slicing.
Colors are not state that we carry across draw ops,
so setting the color on the render job doesn't make
much sense. Instead, pass the color to the various
draw calls. Add a few new ones for that purpose.
Also, shorten the names of some by going from
'load_vertices_from_offscreen' to 'draw_offscreen'.
This reduces how many changes we make when recording uniform state, which
increases the chances that the data offset will be the same when applying
uniforms.
Since we make full snapshots when recording uniform state of batches, we
need to perform some deduplication to avoid so many repeated uniform calls.
This uses a closed hashtable to determine if we are likely changing the
value to something new.
This does not currently compare values, it instead only compares that we
are going to point at a new offset into the uniform buffer. We could go
further if we compare upon updating values (we did that early on in the
prototype) so that offsets are less likely to be changed.
When the color passed is transparent black, use
the color from the texture as source, instead of
as mask. This lets use use the coloring program
both for regular and color glyphs, avoiding
program changes in text with Emoji.
Instead of using uniforms for color used in multiple
programs, pass it as vertex attributes. This will let
us batch more draw calls, since we don't have to change
uniforms so often. In particular, for syntax-highlighted
text.
We may want to change the interface between the
shaders and the renderer for ngl, and therefore,
sharing the shaders between gl and ngl will not
be practical, going forward.
Hook up the "Show fallback rendering" switch for Vulkan.
This brings home the sobering truth that the Vulkan renderer
is doing *all* fallback, since we switched from offset nodes
to transform nodes.
The primary goal here was to cleanup the current GL renderer to make
maintenance easier going forward. Furthermore, it tracks state to allow
us to implement more advanced renderer features going forward.
Reordering
This renderer will reorder batches by render target to reduce the number
of times render targets are changed.
In the future, we could also reorder by program within the render target
if we can determine that vertices do not overlap.
Uniform Snapshots
To allow for reordering of batches all uniforms need to be tracked for
the programs. This allows us to create the full uniform state when the
batch has been moved into a new position.
Some care was taken as it can be performance sensitive.
Attachment Snapshots
Similar to uniform snapshots, we need to know all of the texture
attachments so that we can rebind them when necessary.
Render Jobs
To help isolate the process of creating GL commands from the renderer
abstraction a render job abstraction was added. This could be extended
in the future if we decided to do tiling.
Command Queue
Render jobs create batches using the command queue. The command queue
will snapshot uniform and attachment state so that it can reorder
batches right before executing them.
Currently, the only reordering done is to ensure that we only visit
each render target once. We could extend this by tracking vertices,
attachments, and others.
This code currently uses an inline array helper to reduce overhead
from GArray which was showing up on profiles. It could be changed to
use GdkArray without too much work, but had roughly double the
instructions. Cycle counts have not yet been determined.
GLSL Programs
This was simplified to use XMACROS so that we can just extend one file
(gskglprograms.defs) instead of multiple places. The programs are added
as fields in the driver for easy access.
Driver
The driver manages textures, render targets, access to atlases,
programs, and more. There is one driver per display, by using the
shared GL context.
Some work could be done here to batch uploads so that we make fewer
calls to upload when sending icon theme data to the GPU. We'd need
to keep a copy of the atlas data for such purposes.
When we are rendering a texture node to an offscreen,
and we have a clip, we must force the offscreen rendering.
Otherwise, the code will notice: Hey, it already is a texture
node, so no need to render it to a texture again. But when
clipping is involved, that is exactly what we want to do.
Testcase included.
Fixes: #3651
There isn't any state to modify in the type so we can use const here.
Doing so allows some of the renderer code to use const across a
number of functions so that repeated calls are elided if inlined.
These do not do modify anything so they can be marked as pure to
potentially ellide calls. Since they do dereference, I do not believe
we can make them const although that is unclear since we could technically
just return a pointer + offset. Therefore it *might* be possible to also
make these G_GNUC_CONST.
This also removes the return if fail macros from these as a good portion
of them didn't have them anyway. I think it's fair to say that access to
these incorrectly is a programmer error.
It significantly reduces the amount of code generated into generally a
movss,ret.
According to OpenGL spec, a shader object will only be flagged
for deletion unless it has been detached; when a program object
is deleted, those shader objects attached to it will be detached
but not deleted unless they have already been flagged for deletion.
So we shall detach a shader object before it is deleted, and delete
it before the program object is deleted best.
Here is a command to reproduce this testcase:
GDK_DEBUG=gl-gles gtk4-demo --run gears
Without this patch, Mesa throws this compile error:
0:130(13): error: no matching function for call to `mod(error, float)'; candidates are:
This is caused by `u_rotation - 90` being of type error since
`u_rotation` is a float and it’s illegal to subtract it with an integer.
