nulib2/library/Lesson9.htm
2015-01-09 14:02:33 -08:00

366 lines
17 KiB
HTML

<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=windows-1252">
<meta name="GENERATOR" content="Microsoft FrontPage 4.0">
<meta name="ProgId" content="FrontPage.Editor.Document">
<title>Hacking Data Compression - Lesson 9</title>
<meta content="t, default" name="Microsoft Border">
</head>
<body bgcolor="#FFFFFF" text="#000000"><!--msnavigation--><table border="0" cellpadding="0" cellspacing="0" width="100%"><tr><td>
<p align="center"><font size="6"><strong></strong></font><br>
</p>
<hr>
</td></tr><!--msnavigation--></table><!--msnavigation--><table border="0" cellpadding="0" cellspacing="0" dir="ltr" width="100%"><tr><!--msnavigation--><td valign="top"><!--msnavigation--><msnavigation border="0" cellpadding="0" cellspacing="0" width="100%"><tr><td>
<p align="center"><font size="6"><strong></strong></font><br>
</p>
<hr>
</td></tr><!--msnavigation--></table><!--msnavigation--><msnavigation border="0" cellpadding="0" cellspacing="0" dir="ltr" width="100%"><tr><!--msnavigation--><msnavigation valign="top"><!--msnavigation--><msnavigation border="0" cellpadding="0" cellspacing="0" width="100%"><tr><td>
<p align="center"><font size="6"><strong></strong></font><br>
</p>
<hr>
</td></tr><!--msnavigation--></table><!--msnavigation--><msnavigation border="0" cellpadding="0" cellspacing="0" width="100%"><tr><!--msnavigation--><msnavigation valign="top">
<h2>Hacking Data Compression - Lesson 9</h2>
<p><a href="index.htm">Back to nulib.com library</a></p>
<hr><pre>
Hacking Data Compression Lesson 9 NuFX and ShrinkIt
By Andy McFadden
GEnie A2Pro Apple II University - Copyright (C) 1992 GEnie
This week's lesson is about the de facto archiving standard for the
Apple II world: NuFX, as implemented by ShrinkIt and GS/ShrinkIt.
There are a variety of other programs, like ][+ ShrinkIt,
Auto-Unshrink, NuLib, NuPak, and YankIt, but they all handle the same
archive format.
The information covered here will be of interest to people who want to
write software that works with NuFX archives. If you're only
interested in general data compression, you can probably skip this
one.
-=O=- ShrinkIt
Back in the Olden Days, the format for storage and transmission of
data was Binary II (abbreviated &quot;BNY&quot;, and occasionally referred to as
&quot;bunny&quot; format). It arranged things on 128-byte boundaries for the
convenience of terminal programs, and is still used today for that
purpose. Actually, it was primarily meant as a way to retain the
filetype, auxtype, and other goodies associated with a file, much like
MacBinary for the Macintosh, but it also worked okay as an archive
format.
What made it Really Cool was the Huffman + RLE compressor which was,
well, almost built in to the Binary II Library Utility (BLU). It did
a reasonable job of making things smaller, and in a modest amount of
time. It was called SQueezed compression, which somehow got
abbreviated &quot;.QQ&quot;. Archives with squeezed files in them were &quot;.BQY&quot;.
However, it was really depressing to log onto a UNIX system and watch
UNIX compress crunch files into nothing. So, one fine day, a skinny
little college student named Andy Nicholas, and a taller but still
rather slim guy named Kent Dickey decided it would be a really neat
thing to do LZW on the Apple II.
Since the IIgs and System 4.0 had been released, it was apparent that
the old BNY format was missing a few pieces. Specifically, it had no
clear provision for forked files, file comments, other file systems,
and the expanded size of the stuff that GS/OS was returning in
GetFileInfo calls. So, with some suggestions from a few dozen
Apple II personalities, the NuFX format came into being.
(Sounds pretty dramatic, huh? Hmm, maybe I should write fiction...)
And so it came to pass that ShrinkIt was born. It took a little while
to work all the kinks out, but before long it was recognized that the
program worked faster and packed better than the old SQueezed format.
When Binary II support was added, the fate of BLU and its ilk was
sealed.
One interesting bit of history is that the LZW routines were initially
meant for a program like DDD (Dalton's Disk Disintegrator), i.e. a
*disk* packer which happened to pack files as well. This caused some
interesting twists in the design.
-=O=- Details, Details
The first implementation of LZW was designed around packing a 5.25&quot;
disk using as little memory as possible. The strategy was to grab 4K
chunks (the size of a single track on a 5.25&quot; floppy), put it through
an RLE pass (since tracks were often filled with zeros, this was very
effective), and then an LZW pass. If the RLE or LZW pass failed, it
would be omitted. If both failed, the block was stored without
compression. A few bits at the start of each chunk in the compressed
output held the information.
It turns out that putting a typical file through an RLE pass first
doesn't really do you much good, though it isn't likely to do much
harm. Mostly it just slows things down, unless the file is filled
with a single character. In some cases it can make things slightly
worse, because LZW is encoding the three-byte RLE codes instead of
encoding long strings of zeros with a single 9 to 12-bit symbol. If
more or all of a *disk* track is empty, RLE is a win, but it just
doesn't happen all that often in text files.
The LZW encoder was reset after every 4K of *input*, which meant that
it never had a chance to get completely full, and some of the 12-bit
code space was wasted. However, this limited the maximum height of
the tree considerably. For example, 4K of zeros can be compressed by
a degenerate tree with 91 nodes in it. Since RLE would compress such
a chunk, to get the worst case you have to alternate characters, so
the worst you can get is around 64. This meant that the LZW stack
required well under 100 bytes instead of 4096, and the loss in
compression wasn't too unpleasant.
The other nice feature is that, since the LZW table can't get
completely full, there's no need for explicit (or implicit) table
clears. The whole thing just gets wiped on 4K boundaries.
To find the end of the data, the size of the output was stored.
(Recall there are three ways of knowing when to stop: an explicit stop
character, knowing the size of the input, and knowing the size of the
output.) The LZW code given in lesson 8 used an explicit &quot;end of
file&quot; code because I thought it was more interesting to do it that
way, and because it works on a stream of data (you can't seek back and
write the length bytes on a continuous stream... but since ShrinkIt
knows the data will be &lt; 4K, it can just stuff the bytes in before it
writes to disk).
This scheme was eventually referred to as &quot;LZW-I&quot;, and is still used
by the ][+ and //e versions of ShrinkIt.
-=*=-
When writing GS/ShrinkIt, Andy Nicholas decided to hunker down and do
LZW with table clears and all the fancy nonsense. However, he kept
the 4K boundaries, which made things a bit confusing (and led to a bug
which hung around in GSHK, NuLib, and YankIt until rather recently).
Basically, the input was still read in 4K chunks, and the RLE pass was
still done. As before, the LZW pass was done on either the original
4K or the RLE output, depending on whether or not RLE helped. The
difference was that the table clears were no longer done after the end
of the input. Instead, an explicit table clear code ($100) was sent
whenever the table filled. To avoid having to make a copy of the
table on every chunk, the table was also cleared whenever LZW failed
to compress a piece of the file.
Of course, now the decoder stack had to be 4K, and all the fun with
the table clears had to be added in. The long-lasting bug mentioned
earlier occurred whenever a chunk ended after a table clear plus one
character - if you look at the LZW decoder, you'll see that it grabs
the first character, sets the prefix to that, and then plows onward.
So the decoder was grabbing the first character, reaching the end of
input, and then restarting with the next chunk by grabbing the first
character again. This isn't an issue if you don't chop things into 4K
pieces, which is why you don't see it mentioned in the class stuff.
So, LZW-II has all the advantages of an encoder which doesn't break
the input into pieces, and retains its heritage as a disk compressor.
One nice feature of the &quot;chunking&quot; is that it enables files with tough
spots in them to be compressed better. Suppose you have a file with
three pieces, call them A, B, and C. Let A and C be squishy text-like
substance, while B is essentially random digitized sound. Even if we
get 50% compression in A and C, B could expand by some amount (say,
25%), reducing or even nullifying the compression of A and C. By
storing the tough parts without compression, the only increase in B's
size is the few bits of overhead needed to identify it as
uncompressed.
This slightly anachronistic &quot;chunking&quot; behavior can help significantly
for some kinds of files. By reducing the overhead to only a few bytes
per 4K chunk, ShrinkIt will almost always do better than the simple
LZW encoder given in class. The price is greater complexity in the
compressor and decompressor, and a few bytes of overhead in the file.
If you want to see what effect the improvements had, compress some
files with ShrinkIt and then compress them with GS/ShrinkIt. The
difference tends to be small, which is counter to intuition, since LZW
does poorly at the start.
-=*=-
This section explains the format of the LZW-I and LZW-II compressed
threads. To find out what &quot;threads&quot; are, you'll need to consult the
NuFX file type note.
LZW-I looks like this:
+0 (word) 16-bit CRC for the uncompressed data
+2 (byte) disk volume # (remember, it started as a 5.25 disk packer)
+3 (byte) RLE escape char, usually 0xdb
+4 a series compressed chunks, each built from 4K of input. Each
chunk starts with the following:
++0 (word) size of data after RLE but before LZW (will be 4096
if RLE wasn't used)
++2 (byte) LZW flag - zero if LZW not used
++3 compressed data
So something packed with only RLE would have a size != 4096 and an LZW
flag of zero. LZW only would have a size == 4096 and LZW flag not
equal to zero. RLE+LZW and storing without compression can be
expressed by juggling things. The value at ++0 is used to decide when
the LZW uncompressor should stop.
The CRC is computed on the uncompressed data. If the last piece of
the file doesn't fill an entire 4K chunk, then the rest of the 4K
chunk is filled with zeros. The ENTIRE chunk is compressed, zeros and
all, and the CRC includes the extra zeros.
The actual LZW compression works like it did in class, with the
modifications described earlier. Even though no explicit table clears
are used, the codes begin at $101.
LZW-II looks like this:
+0 (byte) disk volume #
+1 (byte) RLE escape char, usually 0xdb
+2 a series compressed chunks, each built from 4K of input. Each
chunk starts with the following:
++0 (word) size of data after RLE but before LZW (will be 4096
if RLE wasn't used). The hi bit is used as the LZW flag,
i.e. $8000 is ORed in if LZW is used.
++2 (word) if and only if LZW is used, an extra length word is added
here, which gives the total length of the compressed data after
BOTH RLE and LZW.
++2 or
++4 compressed data
The reason for the extra length word is to provide a way to recover
pieces of a file even if one of the chunks has been corrupted. With
LZW-I, there's no way to identify the start of a chunk; with LZW-II,
you can seek from the current chunk header to the next. The value
isn't actually needed by the decompressor. If only RLE is used, then
the size at ++0 will be accurate, so it's omitted from the output.
Notice that there's no CRC included in the header. This is because
the CRC was moved to the thread header. It's computed with an initial
seed of 0xffff (for historical reasons), and is computed on the
original file. It does NOT include the zeros used to pad things to 4K
boundaries.
-=O=- Knowne Implementations of NuFX Programmes
- ShrinkIt (a/k/a &quot;P8 ShrinkIt&quot;). Runs on enhanced //e, //c, IIgs.
Requires 65c02. Packs with LZW-I, unpacks LZW-I and LZW-II.
- GS/ShrinkIt. Runs on IIgs only. Packs with LZW-II, unpacks LZW-I
and II. Also handles several other file formats, including .Z (UNIX
compress) and .ARC.
- ShrinkIt ][+ and UnShrinkIt ][+. Run on any Apple II, including
Apple ][+. One packs, the other unpacks.
- Auto-Unshrink. Runs on any Apple II (?). Sort of a strange
concept, but what the heck. Notable feature is a certain amount of
error recovery for damaged archives (it knows about and will use the
extra length word included for LZW-II).
- NuLib. Shell-based, runs on darn near anything. Packs with LZW-I,
unpacks LZW-I, LZW-II, and UNIX compress (yes Virginia, you can have
files packed with UNIX compress in a ShrinkIt archive... you just
can't unpack them with anything else).
- NuPak. Started out like NuLib and GS/ShrinkIt slammed together at
high speed, but never went anywhere. The $15 shareware fee and the
release of GSHK probably did it in.
- YankIt. Shell-based, IIgs only. Unpacks LZW-I and II, does not
pack at all. Faster (and smaller) than the others.
-=O=- PROGRAM
Being the author of both NuLib and YankIt, you'd figure I'd have some
decent sample source code to give away. However, it shapes up like
this:
- NuLib wasn't so much written as it was &quot;evolved.&quot; It started out as
an archive viewer, and expanded into a rather grotesque archiver. It
looks okay on the outside, but the innards are a mess. Fortunately,
it was written in C.
- YankIt is a cute little bare-bones archive viewer and extractor. It
fit into about 6000 lines of assembly, and took under a month to
design, implement, and test. My original intent was to release the
complete source code with this lesson, but some bugs have been found
and I haven't had the time to fix them, test it, etc, etc.
So, this means I'm going to flake and just point you at some useful
stuff. I'll get the YankIt sources out eventually.
-=*=-
First of all, the definitive reference on the NuFX format is the
filetype note for $e0/8002, which can be found in GEnie's A2Pro
library. The not-quite-so-definitive-anymore reference is in the
article AndyN wrote for the Winter 1990 Call-A.P.P.L.E. The article
also contains a bit of history about NuFX and ShrinkIt.
The NuLib source code, which contains a bug in the stack declaration
(only 100 bytes - correct for LZW-I, but not LZW-II, which should be
4096), can be found in the A2Pro library on GEnie, and on the
cco.caltech.edu anonymous FTP site.
NuLib executables for the IIgs and MS-DOS can also be found on GEnie,
in the appropriate sections. Documentation for the MS-DOS version got
renamed to &quot;NULIB.SHK&quot; or something equally meaningless. The file
comment is also wrong; use &quot;nulib xvt0 nulib.shk&quot; to convert the line
terminators from CR to CRLF. (It says to use &quot;xvt2&quot;, which would
convert CRLF to CRLF - a slightly silly thing to do.) NuLib has a
whole pile of options, and would be a great program if the source code
weren't so ugly.
(Hey, I started it a year after I started learning C, and well before
I figured out how LZW worked... be charitable.)
The YankIt executable can be found in GEnie's A2 section. This
contains the same stack bug found in NuLib, but it's smart enough to
complain about the archive being corrupted and bail without scrubbing
itself. It's got a nice &quot;integrity check&quot; feature which runs through
and uncompresses every file, checking CRCs, but doesn't actually store
them anywhere. NuLib acts like it has a similar feature, but it lies.
GSHK should have a similar feature, but I caught AndyN too close to
the shipping date of v1.1.
Since NuLib and YankIt are shell-based, and I didn't want to add the
code to correctly handle disk archives, they extract disks as a single
file. This can be useful for UNIX and MS-DOS Apple II emulators which
require disk-files. Note that ShrinkIt uses ProDOS sector mapping, so
the disks will work fine for the known UNIX emulators, but not for the
MS-DOS emulator (which expects DOS 3.3 sector ordering). There are
programs available which change the order around.
-=O=- DEEP THOUGHTS
- What are the advantages of writing an archiver which processes
archive entries individually instead of reading them all at once?
(hint: &quot;yankit xv - &lt; file.shk&quot;)
- Is 4096 the optimal size for stopping and checking compression
progress?
- How does ShrinkIt's &quot;checkpoint every 4K&quot; compare to v.42bis? (see
lesson 8 for details on v.42bis)
- Why does ShrinkIt clear the table after a chunk fails to compress
with LZW? What would you have to do to avoid doing so?
-=!=-
</pre><hr>
<address>This document is Copyright by Genie and Andy McFadden</address>
<!--msnavigation--></td></tr><!--msnavigation--></table><!--msnavigation--></td></tr><!--msnavigation--></table><!--msnavigation--></td></tr><!--msnavigation--></table></body>
</html>