It depends on set rounding mode, we get what we want from truncation + 0.5

This should fix #102 - it seems -fno-unsafe-math-optimizations is
not enough to counter -Ofast ending up neutering isnan()

On commonly-sized images, the HSL conversions went from near instant to
"seeing-the-tiles-slowly-updating" after we switched some GIMP code to
babl conversion.

This is because the code is actually made to work in gamma 2.2 and the
only conversion was from and to linear RGB. So when converting
non-linear RGB, we were making far too many conversions. Let's have
faster conversion paths.

--input-profile and --output-profile options are replaced respectively
by --input-space and --output-space (but the old option names still work
for scripts compatibility).

They can still be used to specify an ICC file like we used to, but
additionally if you start with '*', you can select one of the known
spaces, such as sRGB, Rec2020, etc.

In the weird case that you wanted to select a profile in a path starting
with '*' (it's not a forbidden character by the file system even though
usually used for glob file selection), you can protect the initial
character with a double asterisk '**'.

When building using clang and AVX2 is available on a users's computer
running GIMP, then we get a crash when showing a 32-bit float image that
includes NAN values.
See e.g. GIMP issue gimp#11453

Apparently clang handles NAN in a different way than gcc and needs
testing for NAN before doing a comparison on the value.
So we add a test for isnan first and adjust the other conditions
accordingly.

gimp#12064

rgbaf_to_Labaf_sse2() was falling back to the non-SSE2 algorithm when
input and output were not 16-byte-aligned. Additionally to missing some
good optimization, the slightly different implementation makes for
creating opportunities for artifact-looking results when a single canvas
is broken into smaller parts, some of which may be processed with one
implementation, others with the other implementation.

See:
gimp#12030
gimp#2933

… codepath give different results than without SSE2.

The slightly different result is expected and the SSE2 code does not
seem bugged at all. The problem though is that in a single babl_process,
we end up with different results, which may be amplified by further
algorithm into surprising renders (cf. gimp#12030).
With this commit, we ensure than when using the SSE2 codepath, we do not
end processing the last few pixels with different code.

The detailed reasons can be read in first commit from gimp!1578
In short, that was breaking DWARF generation with no benefit right now.

Biswapriyo Nath authored Feb 20, 2024 and Øyvind "pippin" Kolås committed May 04, 2024

This fixes the following linker errors.

ld.lld: error: undefined symbol: __android_log_print
referenced by babl-internal.h:134 (../babl/babl-internal.h:134)
              babl/libbabl-0.1.so.0.208.1.p/babl-component.c.o:(babl_fatal)
referenced by babl-internal.h:142 (../babl/babl-internal.h:142)
              babl/libbabl-0.1.so.0.208.1.p/babl-component.c.o:(babl_fatal)
referenced by babl-internal.h:134 (../babl/babl-internal.h:134)
              babl/libbabl-0.1.so.0.208.1.p/babl-component.c.o:(babl_log)

Taken from gimp!1154

It is possible to use GCC on macOS.

Though the models may itself be registered with a similarly named
attribute to mean they encode alpha, babl_format_has_alpha() in
particular works by verifying if any of the components
represents/encodes alpha. As a consequence, the function was failing to
consider e.g. HSVA as a model encoding alpha.

See also: gegl!142

Again, even though it's basically the same as float or half for a CLI
tool (where all these floating point values are just converted from
string representation), it is still interesting since the tool is doing
actual conversion with a fish using the requested formats. Therefore
it's a nice way to quickly test conversions in terminal.

While apparently fixing the compilation for Windows (and still compiling
on Linux), my previous commit was in fact crashing the CLI tool when it
runs babl_init() in some infinite loop (it looks like it's crashing
inside babl logging code, hence the infinite looping).

Anyway this was annoying me that we had to build-in more than necessary
from babl-internal when all I wanted was to share 2 functions. So I
split babl-util with a new babl-shared-util. The latter is to be used
only for the few utility functions which we might want to share with the
CLI tool (so far, only the 2 half to/from float (en|de)coders).

This time, I've not only compiled but also run the CLI tool both on
Linux and with Wine for the Windows cross-build.

While I didn't have a problem for Linux builds, when cross-compiling for
Windows, these errors occur:

> babl-util.c:296: undefined reference to `babl_strcat'

(and a few similar other "undefined reference" as this one gets fixed)

This is because we include babl-internal.h in bin/babl.c but don't add
the right compilation unit sources. The build was still finding them (or
maybe ignoring them as the functions were unused in the babl CLI tool
code) on Linux, but not on Windows.

The code to encode half values into and from float values was moved from
type-half.c to babl-util.h as an internal function, to make it usable by
the CLI code.

It may not seem totally useful to have half support since anyway we
convert from/to string for the CLI tool (in this regard, half and float
are the same). But since the tool performs an actual babl conversion,
it's actually a good way to test quickly that half conversion works
properly too.

Similar to babl_space_is_cmyk|gray() functions, because right now to
know whether a space is RGB, we have to test whether it's **not** any of
the other. And such code would break as soon as we will add new values
to BablICCType.

This way, we'll be able to host the generated reference and have
automatic links created between GIMP, GEGL and babl API docs.

John Marshall authored Mar 19, 2021 and John Marshall committed Jul 25, 2023

 - pluginsdir: to set BABL_PATH
 - girdir: location of introspection gir file
 - typelibdir: location of introspection typelib file