| Age | Commit message (Collapse) | Author |
|---|
|
|
|
For small images, it could happen that the output stream would not be
flushed before exit(), resulting in a lack of error-reporting on
a full device. Using fflush(), a function I first introduced in sbase,
we do the flushing before returning manually and report errors if they
occurred.
|
|
We split out the libpng-setup into a separate function, it is very very
ugly.
The code also received a general cleanup and aligns itself much better
with the general coding style and structure.
|
|
|
|
|
|
|
|
Instead of calling fwrite on each channel, we write one big chunk
of a line.
This increases performance by around 70% compared to version 1 and
the farbfeld tools are now roughly fast as imagemagick's convert.
I also refactored the code, removed unnecessary variables and unified
the variable naming and error reporting a bit.
Inside jpg2ff, the loop didn't need 3 variables.
|
|
Talking with even more people and weighing the benefits, I've made the
decision to remove color spaces from farbfeld again (as it currently
is handled in version 1) and recommend sRGB, not force it.
Thing is the following: We are not ready yet for this step. Neither the
people working with the images nor the hardware to display extended
colors. Using greater colorspaces also removes the "hackability" of
farbfeld.
As it was previously possible to easily create gradients in plain C,
you have to keep multiple things in mind with linear ProPhoto RGB.
The first thing is that not all colors are "real", and there are
imaginary colors. This doesn't happen with sRGB because all the colors
it describes are "real".
The second thing is the linear gamma curve, which makes the gradients
look perceptually inconsistent.
In the interest of creating a good and simple interchange format, we
can't only weigh in points of professionals. This little excursion has
shown that aiming for the 99% is still the way to go.
And as often as you could repeat that sRGB will meet its fate in the
future when screens become better, it is surprising how badly the
industry has caught up to it.
I also must honestly say that we can't solve the color space issue
using RGB and should look at other formats to explore (CIELUV, CIELAB),
which I will probably give a talk about at slcon3.
Before releasing version 2, my aim is to make the I/O a bit more
effective (as far as possible) so it's even easier to use.
I know people will be pissed off, but fuck it, I'm the maintainer!
Live with it.
|
|
Makes things a lot easier for image manipulation algorithms which
can be expected to be applied to farbfeld data.
|
|
No need for these shitty png-types.
|
|
For some reason, those strange png_char_pp-passes were optimized away
in some way, yielding in icc_len = 0.
I now explicitly use the pointers to indicate to the compiler that we
pass by reference. Fuck those png types...
|
|
I've literally been thinking about this for quite a while now. The
initial motivation behind defaulting to sRGB was the idea that most data
on the web was in sRGB anyway.
However, my assumption was weakened in the sense that the development
is clearly moving towards images with supplied ICC profiles and software
slowly catching up to handle this. My tests have shown that more and
more people even do that on the web, even though it's been a "tradition"
that Photoshop users "Save for Web" and convert the gamut lossy into
sRGB to bring a consistent color-"experience" even to those clients not
supporting ICC profiles and which always assume sRGB.
What made this decision so difficult is that converting to "ProPhoto
RGB" requires some advanced knowledge on this topic, however, I came to
the conclusion to implement it given the *2ff- and ff2*-tools handle it
silently and well in the background, and given the little cms library is
actually quite okay to use.
When converting from ff to png, a proper "Pro Photo RGB" ICC V4-profile is
automatically included in the exported png by ff2png. V4 is not as
widespread as V2, but in the worst case (no V4 support somewhere) the
colors will just be a little off.
As an added bonus, any input files for png2ff which include ICC profiles
are also correctly handled and the color space conversions are executed
as expected.
Accordingly, the FORMAT-specification has been changed. While at it,
I added the note on alpha-multiplication. Now the format is very exact
and shouldn't leave any ambiguities.
jpeg supports ICC profiles as well, but I hadn't had the chance to look
into it (not as trivial as PNG probably, help appreciated :)), so I'm
always assuming sRGB here.
Rationale
---------
It is not obvious why one would go through the lenghts of supporting
this big-gamut colorspace and not just stick with sRGB. In 99% of the
cases, there's no reason to do that, but with even more extreme
developments in the OLED-sector and other advances display hardware is
slowly getting more powerful than sRGB, asking for color information
which is suitable for the task and actually uses the full potential.
The decision in this regard was not difficult in farbfeld because we
always use 16-Bit anyway and won't have to fear posterization in a big-
gamut color-space.
|
|
We don't need the jumps, but rather pass a nice function pointer to
libpng. The jump in libjpg was also useless.
We check each fwrite-call so there's an early bailout in case the output
file is full.
Flushing at the end worked as well, but it took too long for complex
images. We don't want to "block" a pipe here and the approach in jpg2ff
was better.
The iHDR-read was useless. Rather use the get*-functions in libpng,
saves us 2 local variables as well.
|
|
forgot to include setjmp.h. also correct the malloc error-messageSmall
|
|
As known from sbase, we want to manually flush stdout to see if all
data has been passed on.
|
|
Somehow, it broke the depth-value on libpng 1.4.*. Should be worth a
bug-report to the libpng-guys, but I can't be arsed right now.
This fucking library is a pain in the ass.
|
|
|
|
Thanks Dimitris, Hiltjo!
|
|
It took me quite a while to figure out that libpng already gives you
big endian values.
I also tweaked the library functions a bit more to really make sure
that all PNG-types (grayscale, palette, ...) are converted to RGBA
properly.
|
|
Use the arpa/inet.h conversion functions which are standardized
in POSIX.
|
|
Thanks Heiko ("rundstutzen") for reporting this!
|
|
No need for this when the tools don't accept arguments.
|
|
|
|
- Rename the format
- Change the format specification
- Drop old tools waiting to be fixed on a later date, just keep
fixed png for now
- Simplify other stuff
This is a direct consequence of my slcon2-talk on this topic.
At first I planned to have 64 bits per channel, but this is
overkill.
|
