The AdvanceCOMP recompression utilities is a suit of commands for optimizing the compression ratio for a number of zlib based file formats, such as .png, .mng, and .gz.

This is a fork of the AdvanceCOMP recompression utilities that has been modified by adding the advpngidat recompression tool that can recompress Apples iPhone optimized PNG proprietary format.

The files in the src/ directory are the unpacked contents of the advancecomp-1.15.tar.gz distribution.

See also:
    Xcode Build Setting Reference – COMPRESS_PNG_FILES (Compress .png files)
    Technical Q&A QA1681 – Viewing iPhone-Optimized PNGs
    CgBI file format

This fork of the AdvanceCOMP tools does not modify any of the original tools, but adds an additional tool– advpngidat.

The advpngidat tool has a couple of options, which you can view via --help, but there is really only one option that is used in practice: -z4. This specifies that the original .png file should be recompressed using the highest compression / maximum effort setting. For example:

shell% advpngidat -z4 FILE.png
      308008      283951  92% FILE.png

The original .png file is replaced with the recompressed version. The .png file is left unmodified if advpngidat was unable to make the .png smaller.

You can also wild-card batch process .png files too:

shell% advpngidat -z4 Images/*.png
         937         908  96% Images/[email protected]
      147935      136238  92% Images/Default.png
      523332      479938  91% Images/[email protected]
         127         127 100% Images/DottedLine.png (Bigger 133)
         148         148 100% Images/[email protected]

Helpful Hint: The AdvanceCOMP distribution also includes the advdef tool. This tool performs the same recompression optimization on .gz files, and it is used exactly the same way as advpngidat– advdef -z4 FILE.gz.

The advpngidat tool was created by taking the repng.cc and copying it to repngidat.cc. The PNG optimization code from repng.cc was then completely removed, and code that only recompresses a PNG's IDAT chunk was added.

The advpngidat tool performs only a single type of optimization: recompression of the IDAT chunk in a .png file. The contents of the decompressed IDAT chunk are not examined nor modified in anyway, it is simply recompressed using the 7z RFC 1950 / zlib compression engine. The other PNG chunks in the .png file are passed through unmodified.

RFC 1950 / zlib allows for two types of streams:

• Normal. This stream type includes a header, trailer, and Adler-32 checksum of the data.
• Raw. This stream type does not include a header, trailer, or a Adler-32 checksum.

The PNG specification mandates that the IDAT chunk be compressed using the normal, non-raw stream format.

The iPhone optimized PNG format requires the IDAT chunk to be compressed using the raw stream format.

The advpngidat tool works by recompressing the IDAT chunk using the 7z RFC 1950 / zlib compression engine, which can usually achieve an additional 3% to 7% additional compression relative to zlib / gzip -9 maximum compression setting (it's sort of like gzip that goes to gzip -11).

The iPhone optimized PNG format includes an additional, non-standard PNG chunk type: CgBI. The presence or absence of this chunk type is used to determine whether or not the IDAT chunk is a raw or normal RFC 1950 / zlib stream. The advpngidat tool works on either iPhone optimized .png files, or PNG standard conforming .png files. The presence or absence of a CgBI chunk is used to determine whether or not the IDAT chunk should be read and re-written as a raw or normal RFC 1950 / zlib stream.

Some numbers (obtained with shell% wc Images/*.png | tail -1) based on the .png files in Resources/Images at the time of this writing:

This represents a savings of 444673 bytes, or 434KB.

There is an additional option that you can use with pngcrush: -brute. This option tries a very large number of permutations of the various compression knobs, which means it can take a lot longer. Here's the results with -brute enabled:

Using a combination of both -brute and advpngidat saves 1306135 bytes, or 1.24MB, relative to the default Xcode COMPRESS_PNG_FILES optimization.

zlib is a LZ77 based compression scheme, and like all LZ77 based compressors, it uses a length, distance pair in the compressed byte stream that is used by the decompressed to copy length number of bytes from a position that is a distance number of bytes back from the bytes that have already been decompressed.

In practice, there are usually a number of possible matches for any given length, distance pair. Since the number of combinations raises very quickly, there is no practical way to find the combination that results in the smallest number of compressed bytes. A number of heuristics are used by LZ77 based compressors, and the compression ratio achieved by a particular implementation is highly dependent on the length, distance search algorithm used.

The AdvanceCOMP recompression tools use the search / match implementation from the 7-Zip / LZMA compression engine which has been specifically designed to maximize the compression ratio. The 7-Zip / LZMA matcher is typically able to get an additional 4% to 15% compression than zlib at its highest setting (i.e., gzip -9). Of course, it also takes a lot longer to perform the compression. Therefore it makes sense to spend the extra effort only when the files in question are of the "compress once, decompress many times" variety.

As a general rule, decompression speed of LZ77 based compressors is completely independent of the quality and implementation of the compressor. In fact, it usually takes less time to decompress a highly compressed LZ77 byte stream (i.e., AdvanceCOMP recompressed files) than it does to decompress a LZ77 byte stream that was created by a compressor that was optimized for compression speed (i.e., gzip -1). This is because a highly compressed LZ77 stream invariably contains less length, distance pairs that need to be processed, while the length, or the number of bytes copied, tends to be greater, and copying bytes is typically something that is highly optimized (i.e., memcpy).