The benefit of the color indexing transform is to have only one
varying channel after it (the green channel, which after the
transform serves as index into the color table).
If there is only one varying channel before the transform, it's
not beneficial. (...except if there are <= 16 colors, then the
pixel bundling presumably still works.)
Storing a number n needs floor(log2(n) + 1) bits, not ceil(log2(n)).
(The two expressions are identical except for when n is a power of 2.)
Serendipitously covered by the indexed color transform tests in this PR.
If an image has <= 16 colors, WebP lossless files pack multiple
color table indexes into a single pixel's green channel, further
reducing file size. This adds support for that.
My current test files all have more than 16 colors. For a 16x16
black-and-white bitmap that contains a little smiley face in the
middle, this reduces the output size from 128B to 54B.
If an image has 256 or fewer colors, WebP/Lossless allows storing
the colors in a helper image, and then storing just indexes into that
helper image in the main image's green channel, while setting
r, b, and a of the main image to 0.
Since constant-color channels need to space to store in WebP,
this reduces storage needed to 1/4th (if alpha is used) or 1/3rd
(if alpha is constant across the image).
If an image has <= 16 colors, WebP lossless files pack multiple
color table indexes into a single pixel's green channel, further
reducing file size. This pixel packing is not yet implemented in
this commit.
GIFs can store at most 256 colors per frame, so animated gifs
often have 256 or fewer colors, making this effective when
transcoding gifs.
(WebP also has a "subtract green" transform, which can be used
to need to store just a single channel for grayscale images, without
having to store a color table. That's not yet implemented -- for now,
we'll now store grayscale images using this color indexing transform
instead, which wastes to storage for the color table.)
(If an image has <= 256 colors but all these colors use only a single
channel, then storing a color table for these colors is also wasteful,
at least if the image has > 16 colors too. That's rare in practice,
but maybe we can add code for it later on.)
(WebP also has a "color cache" feature where the last few used colors
can be referenced using very few bits. This is what the webp spec says
is similar to palettes as well. We don't implement color cache writing
support yet either; maybe it's better than using a color indexing
transform for some inputs.)
Some numbers on my test files:
sunset-retro.png: No performance or binary size impact. The input
quickly uses more than 256 colors.
giphy.gif (184k): 4.1M -> 3.9M, 95.5 ms ± 4.9 ms -> 106.4 ms ± 5.3 ms
Most frames use more than 256 colors, but just barely. So fairly
expensive runtime wise, with just a small win.
(See comment on #24454 for the previous 4.9 MiB -> 4.1 MiB drop.)
7z7c.gif (11K): 118K -> 40K
Every frame has less than 256 colors (but more than 16, so no packing),
and so we can cut filesize roughly to 1/3rd: We only need to store an
index per channel. From 10.7x as large as the input to 3.6x as large.
No behavior change, but this makes it easy to correctly set this
flag when adding an indexing transform: Opacity then needs to be
determined based on if colors in the color table have opacity,
not if the indexes into the color table do.
With this struct, only the first time something sets opacity is
honored, giving us those semantics.
In an early version of the huffman writing code, we always used 8 bits
here, and the comments still reflected that. Since we're now always
writing only as many bits as we need (in practice, still almost always
8), the comments are misleading.
Deflate and WebP can store at most 15 bits per symbol, meaning their
huffman trees can be at most 15 levels deep.
During construction, when we hit this level, we used to try again
with an ever lower frequency cap per symbol. This had the effect
of giving the symbols with the highest frequency lower frequencies
first, causing the most-frequent symbols to be merged. For example,
maybe the most-frequent symbol had 1 bit, and the 2nd-frequent
two bits (and everything else at least 3). With the cap, the two
most frequent symbols might both have 2 symbols, freeing up bits
for the lower levels of the tree.
This has the effect of making the most-frequent symbols longer at
first, which isn't great for file size.
Instead of using a frequency cap, ignore ever more of the low
bits of the frequency. This sacrifices resolution where it hurts
the lower levels of the tree first, and those are stored less
frequently.
For deflate, the 64 kiB block size means this doesn't have a big
effect, but for WebP it can have a big effect:
sunset-retro.png (876K): 2.02M -> 1.73M -- now (very slightly) smaller
than twice the input size! Maybe we'll be competitive one day.
(For wow.webp and 7z7c.webp, it has no effect, since we don't hit
the "tree too deep" case there, since those have relatively few
colors.)
No behavior change other than smaller file size. (No performance
cost either, and it's less code too.)
This implements some of basic webp compression: Huffman coding.
(The other parts of the basics are backreferences, and color cache
entries; and after that there are the four transforms -- predictor,
subtract green, color indexing, color.)
How much huffman coding helps depends on the input's entropy.
Constant-color channels are now encoded in constant space, but
otherwise a huffman code always needs at least one bit per symbol.
This means just huffman coding can at the very best reduce output
size to 1/8th of input size.
For three test input files:
sunset-retro.png (876K): 2.25M -> 2.02M
(helps fairly little; from 2.6x as big as the png input to 2.36x)
giphy.gif (184k): 11M -> 4.9M
(pretty decent, from 61x as big as the gif input to 27x as big)
7z7c.gif (11K): 775K -> 118K
(almost as good as possible an improvement for just huffman coding,
from 70x as big as the gif input to 10.7x as big)
No measurable encoding perf impact for encoding.
The code is pretty similar to Deflate.cpp in LibCompress, with just
enough differences that sharing code doesn't look like it's worth
it to me. I left comments outlining similarities.
We still construct the code length codes manually, and now we also
construct a PrefixCodeGroup manually that assigns 8 bits to all
symbols (except for fully-opaque alpha channels, and for the
unused distance codes, like before). But now we use the CanonicalCodes
from that PrefixCodeGroup for writing.
No behavior change at all, the output is bit-for-bit identical to
before. But this is a step towards actually huffman-coding symbols.
This is however a pretty big perf regression. For
`image -o test.webp test.bmp` (where test.bmp is retro-sunset.png
re-encoded as bmp), time goes from 23.7 ms to 33.2 ms.
`animation -o wow.webp giphy.gif` goes from 85.5 ms to 127.7 ms.
`animation -o wow.webp 7z7c.gif` goes from 12.6 ms to 16.5 ms.
* Matches how the loader is organized
* `compress_VP8L_image_data()` will grow longer when we add actual
compression
* Maybe someone wants to write a lossy compressor one day
No behavior change.
