Took me quite a while to realize what was going on, but in retrospect
it's obvious -- the rasterizer just *rounds* the sub-pixel-positioned
glyph quads to nearest pixels. Which then can cause the neighboring
glyph data to leak to it (because the texture is then sampled not
directly on the edge pixel, but slightly outside of it), or it can also
cut away the edge, when it gets rounded in the other direction.
This was a problem with the original -- laughably inefficient -- atlas
packer as well, but because that packer had excessive padding around
everything, only the second edge-cutting artifact manifested, and that
one is rather subtle unless you know what to look for.
This means the packing is now slightly worse than before and sizes that
previously worked may no longer fit anymore. But since the new atlas
packer is relatively new (well, from September, time sure flies
different here), and the improvement compared to the original packer is
still quite massive, I don't think this is a problem.
Especially given that nullptr causes an assert. All call sites basically
ended up passing a &font and all that extra annoyance just doesn't make
sense.
Given this API is still relatively recent, I'm not bothering with
backwards compatibility.
In basically all cases it's two independent operations so forcing them
to be done together doesn't really bring any potential efficiency
advantages. On the other hand, splitting them allows allows the caller
to better make use of available memory, as the new
renderGlyphQuadsInto() allows the input and output arrays to be aliased.
Bumping AbstractFont plugin interface versions as this is a breaking
change.
Empty input was asserting originally, but such strict behavior doesn't
seem to be useful in practice as all application code would need to
explicitly make sure that shape() isn't called if the input or the input
range is empty. In particular, I hit this assert as soon as I rebuilt
and ran magnum-player, and the assertion didn't really uncover any bug
or make anything better by firing.
Treating zero glyph output as a failure also isn't good -- rendering
for example a string of whitespace may also result in zero glyphs on the
output, which isn't a failure. Plus it's currently unclear what should
actually be a failure -- neither FreeType nor HarfBuzz really have a
concept of the shaping operation failing, hb_shape() returns void -- so
special-casing glyphCount() being zero as a failure just doesn't make
sense.
This thus means the doScript(), doLanguage() and doDirection()
implementations are now called always, regardless of whether shape() was
called at all, or whether it produced any glyphs. This allows the shaper
for example to give back a detected script etc even if the actual
input range to shape was empty.
Replaces the previous, grossly inefficient AbstractLayouter which was
performing one virtual call per glyph (!). It's now also reusable,
meaning it doesn't need to be allocated anew for every new shaped text,
and it no longer requires each and every font plugin to implement the
same redundant glyph data fetching from the glyph cache, scaling etc. --
all that is meant to be done by the users of AbstractShaper, i.e.
Renderer. The independency on a glyph cache theorerically also means it
can be used for a completely different, non-texture-based way to render
text (such as direct path drawing directly on the GPU), although I won't
be exploring that path now.
It also exposes an interface for specifying script, language,
direction and typographic features. Such interface will be currently
only implemented in HarfBuzz, but that's the intent -- to provide a
flexible enough interface to support all possible use cases that a font
or a font plugin may support, instead of exposing a least common
denominator and then having no easy way to shape a text in a non-Latin
script or use a fancy OpenType feature the chosen font has.
The old public interface is preserved for backwards compatibility,
marked as deprecated, however the virtual APIs are not, as supporting
that would be too nasty. I don't think any user code ever implemented a
font plugin so this should be okay.
To ensure smooth transition with no regressions, the Renderer class and
MagnumFont tests still use the old API in this commit, and their test
pass the same way as they did before (except for two removed MagnumFont
test cases which tested errors that are now an assertion in the
deprecated layout() API and thus cannot be tested from the plugin
anymore). Porting them away from the deprecated API will be done in
separate commits.
The internals are still rather ew, that's for another time, but the goal
here was to make it compile and correctly handle the new variability in
created and passed glyph cache instances. In particular:
- The MagnumFont createGlyphCache() now uploads the texture image
directly, skipping a copy through the CPU-side image which may not
have a size matching the input image. That's kind of nasty (and too
tied to OpenGL), but will be cleaned up once the GlyphCache APIs are
fixed to allow this in a nicer way.
- The MagnumFontConverter will now correctly select a font if the cache
contains more than one, properly deal with glyph caches that either
do or not do image processing, and will fail for array glyph caches
as that's not supported by the current format. It also now has an
early special-case handling for glyph caches with processed images,
where the actual image may have different size (and possibly format),
to match what MagnumFont expects.
Need some regression tests for the upcoming glyph cache rework as
otherwise it'll be too miserable. It now fills the glyph cache image
with some non-trivial data and compares to them, and checks the filled
glyphs for the created glyph cache too.
This also fixes a regression introduced when porting away from STL in
47a1295ab8, where UTF-8 layouting was
reporting extra glyphs at the end. Now UTF-8 support is properly tested
both in the MagnumFontConverter plugin and MagnumFont.
It won't contain just font metrics anymore. Also don't require the
struct to be zero-initialized if opening fails -- simply allow the
plugins to return garbage in that case and save the values only if
opening actually succeeded.
Strictly speaking this isn't an ABI change as the return value isn't
part of the function signature and the struct is still the same, so the
plugin interface version isn't bumped for this change.
All std::string arguments are now a StringView, what returned a
std::pair is now a Pair. STL compatibility headers are included on
deprecated builds to ease porting, as usual.
The only *really* breaking changes are in the internals, where an
ArrayView<const char32_t> is used instead of std::u32string, which is in
line with the change done in Utility::Unicode::utf32(); and a Triple is
returned instead of a std::tuple. Behaviorally nothing changed except
that fillGlyphCache() now asserts if the input string contains invalid
UTF-8 (which is also in line with the cahnge done in Utility::Unicode).
Not that C++ STL and exceptions would be anything to take inspiration
from, but there's std::out_of_range. Python IndexError is also specified
as "index out of range", not "bounds".
Partially needed to avoid build breakages because Corrade itself
switched as well, partially because a cleanup is always good. Done
except for (STL-heavy) code that's deprecated or SceneGraph-related APIs
that are still quite full of STL as well.
Importers with multi-level mesh support are here since 2020, yet somehow
this plugin never exposed those. Another reason for proper test
coverage. The original triangle.ply was used by AnySceneConverter tests,
so it was moved there instead.
Such a hopeful test coverage, almost there, and yet it doesn't test
everything. It now uses 1D, 2D and 3D KTX2 files with levels taken
directly from KtxImporter tests, the original 1d.ktx2 and 3d.ktx2 are
moved to AnyImageConverter test where they are used to verify metadata
presence.
Originally I thought I'd save plugins some time if I just give them the
custom indices directly, without being wrapped in a SceneField /
MeshAttribute enum. But in practice that didn't really save anything,
made the interfaces more error-prone due to there being no concrete type
anymore, and all code that delegates to nested importers or converters
had to re-wrap the IDs again, which is *again* error-prone.
Bumping the interface strings because this is a breaking change for the
implementations. Not for users tho, there nothing changes.
So people new to the plugin stuff can quickly get to usage introduction
and code snippets. The plugins alone don't list anything like that and
it may be *very* confusing otherwise.
Again, similarly to what's done for custom MeshAttribute and SceneField
values already. I'm bumping the importer interface version as adding new
virtual functions is a silent ABI breakage, but it's good to do in any
case as the AnimationTrackTarget enum was extended to 16 bits and the
values got shifted.
While UfbxImporter knows to import these directly, AssimpImporter tries
to load them as FBX and fails (heh!), and ObjImporter has no idea about
materials at all, so recognizing this extension would only add more harm
than good at the moment.
Because I got tired of having to manually postprocess ground truth files
for shader tests.
A bit strange / funny that the first ever version of my code manages to
produce smaller files than both stb_image and ImageMagick (with
-compress RunTimeLengthEncoded), for some reason both pick some strange
inefficient run combinations in various scenarios, such as "copy 2" +
"repeat 3" where I pick "repeat 5". And the difference is not
insignificant -- when testing with some shader test files, it resulted
in a ~17% smaller size!
The plugin follows the stricter variant of the spec by default (i.e.,
splitting runs across scanline boundaries) -- so that's *not* the
reason for the weird differences between my code and theirs -- but
provides an option to not do that for even smaller files. Which I'm
going to use for shader ground truth test files, because there every
byte counts. This option together with the above difference causes files
to be ~25% smaller, which is quite a lot.
Vladimír Vondruš