The semantics of BGRx8888 aren't super clear and it means different
things for different parts of the codebase. In particular, the PNG
writer still writes the x channel to the alpha channel of its output.
In BMPs, the 4th palette byte is usually 0, which means after #21412 we
started writing all .bmp files with <= 8bpp as completely transparent
to PNGs.
This works around that.
(See also #19464 for previous similar workarounds.)
The added `bitmap.bmp` is a 1bpp file I drew in Photoshop and saved
using its "Save as..." saving path.
A tile is basically a strip with a user-defined width. With that in
mind, adding support for them is quite straightforward. As a lot the
common code was named after 'strips', to avoid future confusion I
renamed everything that interact with either strips or tiles to a
global term: 'segment'.
Note that tiled images are supposed to always have a 'TileOffsets' tag
instead of 'StripOffset'. However, this doesn't seem to be enforced by
encoders, so we support having either of them indifferently.
The test case was generated with the following Python script:
import pyvips
img = pyvips.Image.new_from_file('deflate.tiff')
img.write_to_file('tiled.tiff',
compression=pyvips.ForeignTiffCompression.DEFLATE,
tile=True, tile_width=64, tile_height=64)
JPEGs can store a `restart_interval`, which controls how many
minimum coded units (MCUs) apart the stream state resets.
This can be used for error correction, decoding parts of a jpeg
in parallel, etc.
We tried to use
u32 i = vcursor * context.mblock_meta.hpadded_count + hcursor;
i % (context.dc_restart_interval *
context.sampling_factors.vertical *
context.sampling_factors.horizontal) == 0
to check if we hit a multiple of an MCU.
`hcursor` is the horizontal offset into 8x8 blocks, vcursor the
vertical offset, and hpadded_count stores how many 8x8 blocks
we have per row, padded to a multiple of the sampling factor.
This isn't quite right if hcursor isn't divisible by both
the vertical and horizontal sampling factor. Tweak things so
that they work.
Also rename `i` to `number_of_mcus_decoded_so_far` since that
what it is, at least now.
For the test case, I converted an existing image to a ppm:
Build/lagom/bin/image -o out.ppm \
Tests/LibGfx/test-inputs/jpg/12-bit.jpg
Then I resized it to 102x77px in Photoshop and saved it again.
Then I turned it into a jpeg like so:
path/to/cjpeg \
-outfile Tests/LibGfx/test-inputs/jpg/odd-restart.jpg \
-sample 2x2,1x1,1x1 -quality 5 -restart 3B out.ppm
The trick here is to:
a) Pick a size that's not divisible by the data size width (8),
and that when rounded to a block size (13) still isn't divisible
by the subsample factor -- done by picking a width of 102.
b) Pick a huffman table that doesn't happen to contain the bit
pattern for a restart marker, so that reading a restart marker
from the bitstream as data causes a failure (-quality 5 happens
to do this)
c) Pick a restart interval where we fail to skip it if our calculation
is off (-restart 3B)
Together with #22987, fixes#22780.
Non-interleaved files always have an MCU of one data unit.
(A "data unit" is an 8x8 tile of pixels, and an "MCU" is a
"minium coded unit", e.g. 2x2 data units for luminance and
1 data unit each for Cr and Cb for a YCrCb image with
4:2:0 subsampling.)
For the test case, I converted an existing image to a ppm:
Build/lagom/bin/image -o out.ppm \
Tests/LibGfx/test-inputs/jpg/12-bit.jpg
Then I converted it to grayscale and saved it as a pgm in Photoshop.
Then I turned it into a weird jpeg like so:
path/to/cjpeg \
-outfile Tests/LibGfx/test-inputs/jpg/grayscale_mcu.jpg \
-sample 2x2 -restart 3 out.pgm
Makes 3 of the 5 jpegs failing to decode at #22780 go.
When present, the alpha channel is also affected by the horizontal
differencing predictor.
The test case was generated with GIMP with the following steps:
- Open an RGB image
- Add a transparency layer
- Export as TIFF with the LZW compression scheme
This tag is required by the specification, but some encoders (at least
Krita) don't write it for images with a single strip.
The test file was generated by opening deflate.tiff in Krita and saving
it with the DEFLATE compression.
Type 2 <=> One-dimensional Group3, customized for TIFF
Type 3 <=> Two-dimensional Group3, uses the original 1D internally
Type 4 <=> Two-dimensional Group4
So let's clarify that this is not Group3 1D but the TIFF variant, which
is called `CCITTRLE` in libtiff. So let's stick with this name to avoid
confusion.
Images with a display mask ("stencil" as it's called in DPaint) add
an extra bitplane which acts as a mask. For now, at least skip it
properly. Later we should render masked pixels as transparent, but
this requires some refactoring.
We now allow all subsampling factors where the subsampling factors
of follow-on components evenly decode the ones of the first component.
In practice, this allows YCCK 2111, CMYK 2112, and CMYK 2111.
Some apps seem to generate malformed images that are accepted
by most readers. We now only throw if malformed data would lead to
a write outside the chunky buffer.
When using the BMP encoding, ICO images are expected to contain a 1-bit
mask for transparency. Regardless an alpha channel is already included
in the image, the mask is always required. As stated here[1], the
mask is used to provide shadow around the image.
Unfortunately, it seems that some encoder do not include that second
transparency mask. So let's read that mask only if some data is still
remaining after decoding the image.
The test case has been generated by truncating the 64 last bytes
(originally dedicated to the mask) from the `serenity.ico` file and
changing the declared size of the image in the ICO header. The size
value is stored at the offset 0x0E in the file and I changed the value
from 0x0468 to 0x0428.
[1]: https://devblogs.microsoft.com/oldnewthing/20101021-00/?p=12483
This fixes an issue where GIF images without a global color table would
have the first segment incorrectly interpreted as color table data.
Makes many more screenshots appear on https://virtuallyfun.com/ :^)
TIFF files are made in a way that make them easily extendable and over
the years people have made sure to exploit that. In other words, it's
easy to find images with non-standard tags. Instead of returning an
error for that, let's skip them.
Note that we need to make sure to realign the reading head in the file.
The test case was originally a 10x10 checkerboard image with required
tags, and also the `DocumentName` tag. Then, I modified this tag in a
hexadecimal editor and replaced its id with 30 000 (0x3075 as a LE u16)
and the type with the same value as well. This is AFAIK, never used as
a custom TIFF tag, so this should remain an invalid tag id and type.
We currently assume that the K (black) channel uses the same sampling
as the Y channel already, so this already works as long as we don't
error out on it.
TIFF images with the PhotometricInterpretation tag set to RGBPalette are
based on indexed colors instead of explicitly describing the color for
each pixel. Let's add support for them.
The test case was generated with GIMP using the Indexed image mode after
adding an alpha layer. Not all decoders are able to open this image, but
GIMP can.
UnassociatedAlpha is the one used by GIMP when generating TIFF images
with transparency. Support is added for Grayscale and RGB images as it's
the two that we support right now but managing transparency should be
really straightforward for other types as well.
This compression (tag Compression=2) is not very popular on its own, but
a base to implement CCITT3 2D and CCITT4 compressions.
As the format has no real benefits, it is quite hard to find an app that
accepts tho encode that for you. So I used the following program that
calls `libtiff` directly:
```cpp
#include <vector>
#include <cstdlib>
#include <iostream>
#include <tiffio.h>
// An array containing 0 and 1 of length width * height.
extern std::vector<uint8_t> array;
int main() {
// From: https://stackoverflow.com/a/34257789
TIFF *image = TIFFOpen("input.tif", "w");
int const width = 400;
int const height = 300;
TIFFSetField(image, TIFFTAG_IMAGEWIDTH, width);
TIFFSetField(image, TIFFTAG_IMAGELENGTH, height);
TIFFSetField(image, TIFFTAG_PHOTOMETRIC, 0);
TIFFSetField(image, TIFFTAG_COMPRESSION, COMPRESSION_CCITTRLE);
TIFFSetField(image, TIFFTAG_BITSPERSAMPLE, 1);
TIFFSetField(image, TIFFTAG_SAMPLESPERPIXEL, 1);
TIFFSetField(image, TIFFTAG_ROWSPERSTRIP, 1);
std::vector<uint8_t> scan_line(width / 8 + 8, 0);
int count = 0;
for (int i = 0; i < height; i++) {
std::fill(scan_line.begin(), scan_line.end(), 0);
for (int x = 0; x < width; ++x) {
uint8_t eight_pixels = scan_line.at(x / 8);
eight_pixels = eight_pixels << 1;
eight_pixels |= !array.at(i * width + x);
scan_line.at(x / 8) = eight_pixels;
}
int bytes = int(width / 8.0 + 0.5);
if (TIFFWriteScanline(image, scan_line.data(), i, bytes) != 1)
std::cerr << "Something went wrong\n";
}
TIFFClose(image);
}
```
This commit un-deprecates DeprecatedString, and repurposes it as a byte
string.
As the null state has already been removed, there are no other
particularly hairy blockers in repurposing this type as a byte string
(what it _really_ is).
This commit is auto-generated:
$ xs=$(ack -l \bDeprecatedString\b\|deprecated_string AK Userland \
Meta Ports Ladybird Tests Kernel)
$ perl -pie 's/\bDeprecatedString\b/ByteString/g;
s/deprecated_string/byte_string/g' $xs
$ clang-format --style=file -i \
$(git diff --name-only | grep \.cpp\|\.h)
$ gn format $(git ls-files '*.gn' '*.gni')
This change limits the amount of memory that is initially allocated for
the color table. This prevents an OOM condition if the file contains an
incorrect color table size.
Previously, the regression tests for OSS-Fuzz issues 62033 and 63296
used test case files directly from OSS-Fuzz. These files are invalid
in multiple ways because they have been generated by a fuzzer. This
commit replaces these files with ones that only expose the issue being
tested.
The first implementation of this property was just plain wrong. Looks
like this property isn't used a lot as I found the issue by reviewing
the code and not because of a specific image.
The test image is a 32x32 mosaic of alternating black and yellow pixels,
it was generated using this code:
Bitdepth 8
RCT 1
Width 32
Height 32
if W-WW-NW+NWW > -300
- Set -1000
- Set 900
This image is exactly the same as the previous one, excepted the RCT
transformation. It has been generated with:
Width 64
Height 64
RCT 29
Upsample 2
Bitdepth 10
if N > 300
- NE -6
- W 6
This image uses the modular encoding with a very simple prediction tree.
It also makes use of two features: upsampling (x2 factor) and a
non-standard bit depth (10 bits). The file has been generated on
https://jxl-art.surma.technology/ , with the following input:
Width 64
Height 64
Upsample 2
Bitdepth 10
if N > 300
- NE -6
- W 6
Reordering these calls allow us to ensure that all encoders are able to
return the size of the image before they are requested to decode the
whole bitmap.
This tests that we can successfully parse the "everything" TVG files,
which make use of every feature in TinyVG.
Test files taken from https://github.com/TinyVG/examples (MIT).
This just works at the moment after e19892a099, but if we ever do
the FIXME in ColorIndexingTransform::transform(), this test will
remind us to think of this case there too.
catdog-alert-13-alpha-used-false.webp is identical to
catdog-alert-13.web but with the byte at offset 0x2a changed from
0x10 to 0x00 -- that is, the bit in the VP8L header that stores
`is_alpha_used` is cleared.
See the commit message of e19892a099 for more information.
simple-vp8l-alpha-used-false.webp is a copy of simple-vp8l.webp,
with the byte at offset 0x18 changed from 0x10 to 0x00 -- that
is, the bit in the VP8L header that stores `is_alpha_used` is cleared.
We would already allocated a BGRx8888 instead of a BGRA8888 bitmap,
but keep actual alpha data in the `x` channel.
That lead to at least `image` still writing a PNG with an alpha channel.
So explicitly set the alpha channel to 0xff when is_alpha_used is false,
to make sure all consumers of decoded lossless webp data have behavior
consistent with other webp readers.
In practice, webp encoders usually don't write files that have
`is_alpha_used` set to false and then write actual alpha data to their
output. So this is rarely observable. However, for example for
lossy+ALPH webp files, the lossless webp used to store the ALPH channel
has `is_alpha_used` set to false and all channels but green are 0
(since the lossless green channel stores the alpha channel of a
lossy+ALPH webp). So if we dump such a bitmap to a standalone webp
file (e.g. with the temporary debugging code in fc3249a1ca),
then without this commit here, `image` would convert that webp to
a fully transparent webp, while other webp software would correctly
display the green image with opaque alpha.
Each secondary partition has an independent BooleanDecoder.
Their bitstreams interleave per macroblock row, that is the first
macroblock row is read from the first decoder, the second from the
second, ..., until it wraps around again.
All partitions share a single prediction state though: The second
macroblock row (which reads coefficients off the second decoder) is
predicted using the result of decoding the frist macroblock row (which
reads coefficients off the first decoder).
So if I understand things right, in theory the coefficient reading could
be parallelized, but prediction can't be. (IDCT can also be
parallelized, but that's true with just a single partition too.)
I created the test image by running
examples/cwebp -low_memory -partitions 3 -o foo.webp \
~/src/serenity/Tests/LibGfx/test-inputs/4.webp
using a cwebp hacked up as described in #19149. Since creating
multi-partition lossy webps requires hacking up `cwebp`, they're likely
very rare in practice. (But maybe other programs using the libwebp API
create them.)
Fixes#19149.
With this, webp lossy support is complete (*) :^)
And with that, webp support is complete: Lossless, lossy, lossy with
alpha, animated lossless, animated lossy, animated lossy with alpha all
work.
(*: Loop filtering isn't implemented yet, which has a minor visual
effect on the output. But it's only visible when carefully comparing
a webp decoded without loop filtering to the same decoded with it.
But it's technically a part of the spec that's still missing.
The upsampling of UV in the YUV->RGB code is also low-quality. This
produces somewhat visible banding in practice in some images (e.g.
in the fire breather's face in 5.webp), so we should probably improve
that at some point. Our JPG decoder has the same issue.)
I somehow added the wrong image here. 4.webp is the one described
by the comment in the test. Now test actually uses the image it
claims to use.
No behavior change.
The alpha channel of a lossy webp is always stored separately from
the (lossy) RGB data. Alpha is either compressed in a lossless webp
that stores just the alpha data, or it's stored completely
uncompressed. (But again, even if it's compressed, it's losslessly
compressed.)
This adds a test for uncompressed alpha, which I hadn't tested before.
It seems to work correctly, though :^)
I generated the test image by running:
~/Downloads/libwebp-1.3.0-mac-arm64/bin/cwebp \
-alpha_method 0 \
Tests/LibGfx/test-inputs/extended-lossless.webp \
-o Tests/LibGfx/test-inputs/extended-lossy-uncompressed-alpha.webp
This image covers two things that aren't covered by the existing
tests, and I found it useful for testing locally. The image's license
allows redistributing it, so add it as a test case.
WebP lossless files that use a color indexing transform with <= 16
colors use pixel bundling to pack 2, 4, or 8 pixels into a single pixel.
If the image's width doesn't happen to be an exact multiple of the
bundling factor, we need to:
1. Use ceil_div() instead of just dividing the width by the bundling
factor
2. Remember the original width and use it instead of computing
reduced width times bundling factor
This does these changes, and adds a simple test for it -- it at least
checks that the decoded images have the right size.
(I created these images myself in Photoshop, and used the same
technique as for Tests/LibGfx/test-inputs/catdog-alert-*.webp
to create images with a certain number of colors.)
For the test files, I opened Base/res/icons/catdog/alert.png in Adobe
Photoshop 2023, used Image->Mode->Index Color...->
Palette: Local (Perceptive) to reduce the number of colors to 13, 8, and
3 with transparency, and 2 without transparency, then converted it back
to Image->Mode->RGB Color (else it can't be saved as webp), then
File->Save a Copy... to save a WebP (mode lossless) for every palette
size.
The image is https://quakewiki.org/wiki/File:Qpalette.png in lossless
webp format with a color indexing transform.
I've created Qpalette.webp by running
examples/cwebp -z 0 ~/src/serenity/tmp.ppm -o Qpalette.webp
built at libwebp webmproject/libwebp@0825faa4c1 (without
png support, so I first ran
Build/lagom/image ~/Downloads/Qpalette.png -o tmp.ppm
to convert it from png to a format my cwebp binary could read).
This file also happens to explicitly set max_symbol, so it serves
as a test for that code path as well.
Introduced in 2c98eff, support for non-interleaved scans was not working
for frames with a number of MCU per line or column that is odd. Indeed,
the decoder assumed that they have scans that include a fabricated MCU
like scans with multiple components.
This patch makes the decoder handle images with a number of MCU per line
or column that is odd. To do so, as in the current decoder state we do
not know if components are interleaved at allocation time, we skip over
falsely-created macroblocks when filling them. As stated in 2c98eff,
this is probably not a good solution and a whole refactor will be
welcome.
It also comes with a test that open a square image with a side of 600px,
meaning 75 MCUs.
You can generate one by using `cjpeg` with the -scan argument.
This image has been generated with the following scan file:
0 1 2: 0 0 0 0;
0: 1 9 0 0;
2: 1 63 0 0 ;
1: 1 63 0 0 ;
0: 10 63 0 0;
This type of image isn't common, and you can probably only find one by
generating it yourself. It can be done using `cjpeg` with the -scan
argument.
This image has been generated with the following scan file:
0: 0 63 0 0;
1: 0 63 0 0;
2: 0 63 0 0;
Nobody made use of the ErrorOr return value and it just added more
chance of confusion, since it was not clear if failing to sniff an
image should return an error or false. The answer was false, if you
returned Error you'd crash the ImageDecoder.
Turns out extended-lossless-animated.webp did have a loop count of 0.
So I opened it in Hex Fiend and changed the byte at position 42
(which is the first byte of the little-endian u16 storing the loop
count) to 0x2A, so that the test can compare the loop count to something
not 0.
