Compare revisions

Commit 9beeefcb uses too new functions of librsvg. We could just bump
the minimum required version but there are issues with Rust not being
available on every platform yet. So instead, let's add some conditional
code paths, so that it still builds with librsvg 2.40.x (which was the
last versions fully in C) while we use newer code and no warnings when
using newer versions.

This was not needed when running inside GIMP since it applies filters
to drawables and automatically crops to drawable size.

If we try to call gegl_pad_is_input() with a NULL pad, the code will try
to dereference a NULL pointer, hence crash.

Also raise a proper warning with understandable message which would help
a GEGL op developer to debug their code. For instance:

> GEGL-WARNING: gegl_node_connect: the sink node 'gegl:cell-noise 0x557884b29c40' doesn't have a pad named 'input', bailing.

See: https://gitlab.gnome.org/GNOME/gimp/-/issues/9756

Several parts of gegl updated gexiv2 function calls to use functions
only available in gexiv2 0.14.0.
This causes build failures on systems with older gexiv2 versions,
e.g. Debian old-stable (used for GIMP 2.10 CI) has 0.10.9.

Since gexiv2 is optional and all files using gexiv2 use calls only
available in 0.14.0, let's set gexiv2 required version to 0.14.0.

When enabling loggint output, e.g.
GEGL_DEBUG=all GEGL_USE_OPENCL=yes ninja test

calls to GEGL_NOTE that use %s parameters produce garbage output
instead of the strings, e.g.
GEGL-Message: 14:38:47.453: [ opencl ] ../../gegl/gegl/gegl-debug.h:145:
Platform Name: \u0018\xc1\xe3C\xff\u007f

Since this is inside a function using varargs, change the call to
make that function aware of that, so use
g_strdup_vprintf instead of g_strdup_printf

Generating gegl's gir docs gives an error:
ERROR: Could not find GIR dependency in the search paths: Babl-0.1

It seems we need to add the path to Babl's gir file to the include path
for the API doc generation.

When using poly2tri-c from the subproject, `-lm` flag will be inherited
due to some Meson impurity but this is not the case when using poly2tri-c from system:

    ld: seamless-clone/libgegl-sc-0.4.so.p/sc-sample.c.o: in function `gegl_sc_compute_sample_list_part':
    sc-sample.c:(.text+0x511): undefined reference to `sqrt'

Previously, the distributor could just pass `-lm` globally when building GEGL
but since 2348b571, it is no longer possible
because it breaks `cc.has_header_symbol()` on Clang:

    clang-11: error: -lm: 'linker' input unused [-Werror,-Wunused-command-line-argument]

Let’s declare the dependency explicitly.

https://gitlab.gnome.org/jtojnar/poly2tri-c/-/commit/f68876c2370de691c542943d39a69c6f918f0481

Fixing issue #339

The gegl_color_get_pixel() and gegl_color_set_pixel() variants already
exist but are not introspectable, hence not usable in bindings. This is
a variant using GBytes which is a GLib struct containing random data
with knowledge of its own data size.

The current get|set_rgba() functions returns or set values as linear
light non premultipled RGBA data (within the sRGB space), which is very
limited. We may want to get or set values within another space, or maybe
simply even within the space this GeglColor was set as. These new
functions allow it.

As a few other functions, it is possible to use a Babl format instead of
the space argument, and the function will take care of using the space
only to return values in "R'G'B'A float" within this space.

Adds in an absolute function to test the grow-radius magnitude, rather
than its signed value. Allows shadows to once again shrink when given a
negative grow radius.

Also fixes gimp#9778, gimp#8607

In particular, the function babl_matrix_mul_vectorff_buf4_sse2() in babl
code uses _mm_load_ps() whose documentation says explicitly:

> mem_addr must be aligned on a 16-byte boundary or a general-protection
> exception may be generated.

And indeed I get a segmentation fault on a gegl_color_get_pixel() run
on a color->priv->pixel struct alignment which was not 16-bytes aligned.

Oups! The bug went unseen as it only happened with older babl version.

Fixes #355

Adds high parameter for the opencl version of Threshold using similar
code as the non-opencl version.
This also adds the high parameter with a value of 1.0 to composition
defined in threshold.c

- Start sentence of high value description with a capital.
- Threshold has two parameters now, so reflect that in its description.

This operation optimizes an existing grayscale dithering or creates
a new one with the given numer of levels.

In particular, this fails to all our native Windows builds in CI (it
works in the cross-builds).

In commit d4e80db6 bpp was changed
from signed to unsigned int. This causes incorrect computation of
pointer offsets due to mixing of signed and unsigned variables, which
in turn causes a crash if parameter seed edges is enabled.

Let's revert this back to unsigned values, which is less invasive than
explicitly casting to int everywhere it's needed. The value of bpp
will be positive everywhere, so this should not have any negative
effects in other places.

While we're at it, also silence an uninitialized warning by setting
y0 to 0.

It seems that for unclear reasons the mantiuk filter crashes when
built with clang.
This seems to be related to using OpenMP, so for now let's disable
that when using clang.

Without this g_param_values_cmp() is unable compare GeglColor-s and tell
that 2 colors are absolutely identical.