On Tue, Dec 31, 2024, 13:42 Christoph M. Becker <cmbecke...@gmx.de> wrote:
> On 11.07.2024 at 11:35, Giovanni Giacobbi wrote:
>
> > The recent PR #14877 [1] proposes to add the imagecompare gd function
> that
> > mimics the gdImageCompare function from libgd. I always thought that a
> > pixel-by-pixel matching function for two images was a big missing feature
> > in PHP, as the corresponding userland implementation is really, REALLY
> slow.
> >
> > The problem with gdImageCompare is that it checks several aspects of the
> > two images, and fails on its core purpose. The behavior is better seen in
> > its source code [2] rather than explained. I see the following problems:
> > - The gdImageCompare is kind of buggy, as it completely disregards the
> > alpha channel. As a user, I would expect that two images that have a
> pixel
> > rgba(255, 0, 0, 100%) and one rgba(255, 0, 0, 50%) are considered
> > different, but the current implementation reports them as identical.
> > - Checking for bitmasks is cumbersome, and it requires 9 new constants in
> > the global namespace (the IMG_CMP_*), which are nearly useless. Most use
> > cases will just want to check for this: imagecompare($im1, $im2) &
> > IMG_CMP_IMAGE.
> > - The checks in gdImageCompare are also a bit arbitrary, why compare the
> > interlace flag but not the palette order? Why check the number of colors
> in
> > the palette if they are not even used in the image?
> > - If a user is interested in the other checks, they can all be
> implemented
> > in userland with existing functions: imageinterlace,
> imagecolortransparent,
> > imageistruecolor, imagesx, imagesy, imagecolorstotal.
> >
> > I believe PHP should only offer required building blocks and leave to the
> > libraries the implementation of more structured functionality. My
> proposal
> > is a completely different function called imagematch() that solves the
> > exact problem of the pixel-by-pixel matching, with the added value of
> > optionally matching portions of the images, using the following
> signature:
> >
> > function imagematch(GdImage $image1, GdImage $image2, int $x1 = 0, int
> $y1
> > = 0, int $x2 = 0, int $y2 = 0, ?int $width = null, ?int $height = null):
> > bool {}
> >
> > I already drafted PR #14914 [3] with the first two parameters. There is
> not
> > an equivalent libgd implementation, so the implementation is entirely on
> > the php side, but as you can see it's quite trivial. Even with the
> extended
> > functionality of matching portions of the images it would be just a few
> > extra lines of code, so not a big maintenance burden.
> >
> > If an RFC is required, I'm happy to redact it after the initial feedback
> > round.
> >
> > [1] https://github.com/php/php-src/pull/14877
> > [2]
> >
> https://github.com/php/php-src/blob/5586d0c7de00acbfb217e1c6321da918b96ef960/ext/gd/libgd/gd.c#L2834
> > [3] https://github.com/php/php-src/pull/14914
>
> Difficult. It seems to me we first need to consider the use-case.
> Likely useful only for testing purposes, but I may miss some other
> use-cases.
>
> And now comes the really hard part: when do we consider images to be
> equal? You have mentioned that gdImageCompare() completely ignores the
> alpha channel, and that would be a mistake. I tend to agree, but if the
> images are stored via imagejpeg(), for instance, the alpha channel would
> be ignored as well, so the resulting images would be equal.
>
For the "test-oriented" use case I agree, but it doesn't only apply to
php's internal testsuite: i maintain quite a few projects which do image
manipulation whose testsuites would benefit from this, and that already
places us in userspace use case.
For the "what to consider different" i would apply the KISS principle and
just have a pixel-by-pixel yes/no difference, not even considering
palette/truecolor (as per my sample PR implementation). If you need to do
further checks (i.e. the mentioned palette/truecolor) you can still check
in userspace for o(1).
The idea being, can we provide the most expensive operation, that is basic
pixel by pixel comparison, as native function? Otherwise we keep discussing
bikeshedding and years go by without this being available. This does not
stop future more advanced comparison algorithms from being introduced (i.e.
imagediff()).
> Then there is the question whether two fully transparent pixels (alpha
> value 127) are different if their RGB channels have different values.
> In theory, sure, but it practice is may not really matter.
>
Fully transparent black should be considered matching fully transparent
white, cause there can be no possible discerning difference in the
rendering.
Then we have the transparent pixel. Should equality comparison take
> that into account? Again, in theory, it is not the same as a pixel with
> alpha channel 127, but in practice it may very well be.
>
Yes. Overlaying a black 100% opaque pixel is different from overlaying 50%
black pixel, on a non-black surface. Again result is "rendering oriented".
> Now you claim that doing pixel-to-pixel comparisons in PHP would be
> *really* slow. I'd argue that with modern PHP the actual looping over
> the pixels isn't that slow. Even getting the pixel colors
> (imagecolorat()) and checking them for equality shouldn't be that slow.
>
I'm happy to run a quick bechmark and provider results, but my feeling is
we are talking orders of magnitude (without jit, i personally don't
consider it production ready). But I will also benchmarm it with JIT
enabled.
It only gets slow if you want to do more elaborate comparisons; e.g.
> getting the actual values a palette entry refers to, but if you move
> that out of the loop, performance might not be that bad. Implementing
> that in C is certainly faster, but the difference should be measured,
> and I wonder how much could be gained using SIMD instructions (something
> libgd still ignores completely).
>
I'm talking about a naive pixel by pixel comparison, in plain C, without
any special cpu optimization. Using simd would be nice but out of scope
(that should definitely be a libgd discussion).
I will gladly follow up with numbers in a few days.
