| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
On the relative pros and cons of various RLE schemes
So over in Native crunch/decrunch code. I just suggested encoding
abbcdddeffffg as
abb{0}cdd{1}eff{2}g
Raistlin kindly liked the idea, Krill quite sensibly didn't want us to hijack the post and suggested a new post for the new topic. so... Have at it! |
|
| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
(oh, addendum - whenever the count maxes out, there's no literal following, just another count - which could be zero. This also makes it quite easy to play with different count widths, (eg nybbles or bitpairs) without going down the huffman or arbitrary precision integer routes) |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
As RLE isn't very crunchy to begin with, i'd suggest to focus on other things affecting performance,
such as whether the scheme allows for stream encoding (entire input not known beforehand) for a local optimum within this scheme,
whether or not it's a bit- rather than bytestream,
and how big a 6502 decompressor would be. |
| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
I must admit, I came up with the scheme above partly because I got most of the way through implementing a cruncher for the usual escape byte based scheme and thought there must be a simpler way - so if anything I was (vaguely) optimising for code complexity - which presumably maps to code size to some extent. |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
Quoting ChristopherJamSo over in Native crunch/decrunch code. I just suggested encoding
abbcdddeffffg as
abb{0}cdd{1}eff{2}g
Isn't this going to be expensive for certain inputs containing lots of byte pairs? |
| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
That depends how expensive your encoding for a zero additional repeat counts is. But sure, every compression scheme is vulnerable to certain inputs getting bigger instead of smaller, otherwise you could just keep recrunching to get arbitrarily small outputs. |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
but here you'll be vulnerable to something not too uncommon. I think for a run count represented with a regular byte, the escape code method is going to be smaller than both this and the lit/run chunk with count method for almost all normal c64 input. |
| |
CyberBrain
Administrator
Posts: 392 |
Quote: So over in Native crunch/decrunch code. I just suggested encoding
abbcdddeffffg as
abb{0}cdd{1}eff{2}g
Raistlin kindly liked the idea, Krill quite sensibly didn't want us to hijack the post and suggested a new post for the new topic. so... Have at it!
That looks like the encoding scheme from the codebase64 article? (https://codebase64.org/doku.php?id=base:rle_pack_unpack)
Yes, that way of encoding it, wastes one byte for each 2-byte repeat (starts saving bytes for 4-byte repeats and longer), so could be bad if there are too many of those in the input data
(I actually used that one in my latest demo to get room for more pics in memory, and this ultra simple RLE-based packing, together with delta-encoding, worked surprisingly (to me) well - it saved multiple $100s of bytes off some pics (they did have large background areas, but still), so the 2-repeats doesn't have to be a problem)
It could be interesting to hear about the other ways of representing the bytes of RLE-encoding, which are used on the C64.
For example, what is this "the usual escape byte based scheme" you guys have mentioned in this and the other thread? And the counter representation? |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
Quoting CyberBrainIt could be interesting to hear about the other ways of representing the bytes of RLE-encoding, which are used on the C64.
For example, what is this "the usual escape byte based scheme" you guys have mentioned in this and the other thread? And the counter representation?
escape byte (typical example):
<byte> - emit byte
<esc> <count> <byte> - emit count * byte
<esc> $00 - emit 'esc' byte the esc byte is selected as the least used byte in the data (with some optimizations possible).
lit/run chunk (typical example):
$01-$7f, 1-127 * <byte> - emit 1-127 literal bytes
$81-$ff, <byte> - emit 1-127 * byte
$00 - end of stream |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting tlrbut here you'll be vulnerable to something not too uncommon. I think for a run count represented with a regular byte, the escape code method is going to be smaller than both this and the lit/run chunk with count method for almost all normal c64 input. As RLE isn't that crunchy anyways, performance should be the main regard (where crunchiness plays a rather weak role, but it still has some impact).
Some test corpus wouldn't hurt, ofc. Maybe the good old Bitfire benchmark files? |
| |
CyberBrain
Administrator
Posts: 392 |
<Post edited by CyberBrain on 22/4-2024 23:52>
Cool, thank you, tlr!
Ah, ok, i can see how the "Escape byte" representation doesn't have to waste anything for 2-byte repeats (but requires an extra pass over the input data to find the least used byte when packing).
If understand the "lit/run chunk" representation correctly, it looks like it wastes a byte on each literal-chunk (so at least 2 bytes per 256 packed bytes, as far as i can tell). Looks like this could struggle a bit if there are many short repeats mixed in between non-repeating data. Edit: No, it couldn't - fake news.
Those are the 3 standard representations on C64? |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
Quoting CyberBrainThose are the 3 standard representations on C64?
I'd say escape byte is by far the most common. I've seldom seen the lit/run chunk, but it's used in G-Packer V1.0 and also in Amiga IFF ILBM files IIRC. I've never seen the one cjam spoke of in the wild. |
| |
Bansai
Registered: Feb 2023
Posts: 35 |
In the absence of an escape byte that needs to be consulted every input byte for driving decisions in decoding, you could still have a *mode* escape of sorts where a reserved repeat count value or signaling bit in the length indicates something like "directly copy [arbitrary number of] bytes from source to dest" or there is a some form of encoded RLE skip/copy count that follows the reserved value. This puts more complexity in the compressor, but the decode is still fairly straightforward.
In an input stream if the value XX refers to the same value:
XX XX RC
if RC is positive, emit f(RC) more repeats of XX, for example, RC+1 is the repeat count, but this could grow to include large repeat counts to handle long runs of zeros, etc.
otherwise directly copy the number of bytes that is a function of -RC. |
| |
Raistlin
Registered: Mar 2007
Posts: 575 |
Isn’t it more common to use a bit to define “constant or not”?
$00-7F .. count of non-repeats ($01-$80)
$80-FF .. count of repeats ($04-$83)
There’s no point of course in having 1-2 repeats. I’m not sure there’s much point in 3 either, definitely not if breaking from a non-repeat segment to present them?
This also makes the depack code a little simpler as you simply check with BPL/BMI. |
| |
Fungus
Registered: Sep 2002
Posts: 629 |
I did a lot of RLE work in the past, and I found that output that gave the most equal sequences was best for crunching afterwards. Of course there are many different packer implementations but almost all of them either use one or the other method.
For my own I chose to use scan codes, using the least used byte in the file as the code to have minimum expansion. Some packers used more than one code to represent hard coded lengths which were scanned for in order to shorten the encoding.
Examples like
(code)(byte) for runs of 3
or
(code)(byte) where the code was a lookup table for various lengths
or simply multiple codes, but that always added more expansion.
I chose to limit runs to no more than 256 bytes, so for a length of say 4095 bytes the output would be
(code)(byte)(00)(code)(byte)(00)(code)(byte)(00)(code)(byte)(ff)
This gave an output that crunched better than other packers which would give
(code)(01=16bit)(lengthlow)(lengthhi)(byte)
or similar...
depends on use case of course.
I never did look into super zipper v8 or similar to see what they were doing though. |
| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
Quoting CyberBrainThat looks like the encoding scheme from the codebase64 article? (https://codebase64.org/doku.php?id=base:rle_pack_unpack)
Oh! Yes, it's almost exactly that, except I didn't think to reserve one count value as an end-of-data marker.
Quote:Yes, that way of encoding it, wastes one byte for each 2-byte repeat (starts saving bytes for 4-byte repeats and longer), so could be bad if there are too many of those in the input data
Yeah it's a size loss if there are more 2-byte repeats than occurrences of a potential escape byte value (assuming counts are stored as whole bytes of course). But the implementation is nice and simple.
Quote:(I actually used that one in my latest demo to get room for more pics in memory, and this ultra simple RLE-based packing, together with delta-encoding, worked surprisingly (to me) well - it saved multiple $100s of bytes off some pics (they did have large background areas, but still), so the 2-repeats doesn't have to be a problem)
Nice!
Quote:It could be interesting to hear about the other ways of representing the bytes of RLE-encoding, which are used on the C64.
For example, what is this "the usual escape byte based scheme" you guys have mentioned in this and the other thread? And the counter representation?
Oh, things like (esc)(byte)(repeat count), where (esc) is the least commonly occurring byte in the source data. It does mean than any occurrences of the escape byte are inflated.
As for counter representation, most RLE implementations use a whole byte for the repeat count, but you can do things like read a smaller number of bits to make the repeat counts more compact, much as that slows decrunch a bit (you still keep the literals byte aligned of course). I do like that idea Fungus mentioned of having a few different escape codes for common repeat counts. |
| |
Fungus
Registered: Sep 2002
Posts: 629 |
In the case of my packer I used esc,esc to represent a literal. It still caused expansion, but it also was a common sequence that would be crunched by whatever lz compressor I was using at the time, either byteboiler/level ab2, or level crusher.
The esc,esc worked for the other encoding way too.
Lightemizer and Oneway's packers would be interesting to look into these days. Back then I was trying to understand algos and make my own rather than using other peoples code. But you don't get accused of ripping anymore so looking at stuff is OK these days. |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting FungusBut you don't get accused of ripping anymore so looking at stuff is OK these days. Ripping and looking at other people's code are different things, though. =) |
| |
The Human Code Machine
Registered: Sep 2005
Posts: 110 |
Here's an exmaple of a byte packer using different pack codes: https://codebase64.org/doku.php?id=base:mdg_bytesmasher_0.04&s[.. |
| |
JackAsser
Registered: Jun 2002
Posts: 1994 |
Quote: That depends how expensive your encoding for a zero additional repeat counts is. But sure, every compression scheme is vulnerable to certain inputs getting bigger instead of smaller, otherwise you could just keep recrunching to get arbitrarily small outputs.
Like Dinosours? |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting JackAsserLike Dinosours? Dinasours? Which magical infinite cruncher are you referring to? =) |
| |
chatGPZ
Registered: Dec 2001
Posts: 11148 |
did someone mention packers.txt yet? |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
Quoting RaistlinIsn’t it more common to use a bit to define “constant or not”?
$00-7F .. count of non-repeats ($01-$80)
$80-FF .. count of repeats ($04-$83) Do you have any other examples of packers implementing this? I haven't seen that many.
I think the reason it's uncommon (on the C64) because it doesn't pack as well, and size was probably the main driver for the development of packers/crunchers. Decompression speed came much later. |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting tlrI think the reason it's uncommon (on the C64) because it doesn't pack as well, and size was probably the main driver for the development of packers/crunchers. Decompression speed came much later. I think it's uncommon because everybody just copied everybody else's approach. :)
RLE used to be LZ-crunched anyways, so RLE compression ratio wasn't much of a concern.
How are you so sure this approach (generally?) packs worse than the escape-byte approach? |
| |
tlr
Registered: Sep 2003
Posts: 1727 |
Quoting KrillRLE used to be LZ-crunched anyways, so RLE compression ratio wasn't much of a concern. Initially RLE seemed to have been used by itself.
Quoting KrillHow are you so sure this approach (generally?) packs worse than the escape-byte approach? I'm thinking because only 128-ish runs are allowed + only 128-ish literal blocks. For a typical C64 program (demo?) there would be longish runs of empty space, interweaved with longish runs of code and data.
But maybe I'm assuming too much? Would be interesting to seem some practical measurements. |
| |
chatGPZ
Registered: Dec 2001
Posts: 11148 |
Quote:For a typical C64 program (demo?) there would be longish runs of empty space, interweaved with longish runs of code and data.
Thats why people used "linkers" before packing :) But even if not so, the difference shouldn't be more than a couple bytes on a typical program - and those bytes would form a nice sequence that can be crunched later. |
| |
Fungus
Registered: Sep 2002
Posts: 629 |
Probably the most interesting Super-Zipper V8 |
| |
WVL
Registered: Mar 2002
Posts: 886 |
Quote: Quoting RaistlinIsn’t it more common to use a bit to define “constant or not”?
$00-7F .. count of non-repeats ($01-$80)
$80-FF .. count of repeats ($04-$83) Do you have any other examples of packers implementing this? I haven't seen that many.
I think the reason it's uncommon (on the C64) because it doesn't pack as well, and size was probably the main driver for the development of packers/crunchers. Decompression speed came much later.
Here is an example : https://csdb.dk/release/?id=34685. |
| |
JackAsser
Registered: Jun 2002
Posts: 1994 |
Quote: Quoting JackAsserLike Dinosours? Dinasours? Which magical infinite cruncher are you referring to? =)
Check any Dinosours release and you'll get the point. :D |
| |
Bitbreaker
Registered: Oct 2002
Posts: 501 |
Quoting Krill
RLE used to be LZ-crunched anyways, so RLE compression ratio wasn't much of a concern.
What is kind of hard to understand, as LZ includes RLE by design (match with offset 1) |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting JackAsserCheck any Dinosours release and you'll get the point. :D I was aware that Dinasours pack everything 10 times for comedic effect,
but i thought maybe they've released the Airwolf Fixer of crunchers or something and i've missed that. :) |
| |
Krill
Registered: Apr 2002
Posts: 2855 |
Quoting BitbreakerQuoting Krill
RLE used to be LZ-crunched anyways, so RLE compression ratio wasn't much of a concern.
What is kind of hard to understand, as LZ includes RLE by design (match with offset 1) Native LZ crunchers used to require quite a bit of working memory, so the maximum size of input programs was often smaller than entirely uncompressed programs.
Thus RLE first, then LZ. Also this two-pass approach can make for better overall compression ratio. |
| |
Jetboy
Registered: Jul 2006
Posts: 228 |
Quoting KrillThus RLE first, then LZ. Also this two-pass approach can make for better overall compression ratio.
Puls LZ ones were rather slow, so prepacking with RLE first could save substantial amounts of time. As shorter file to pack tends to pack quicker. |
| |
enthusi
Registered: May 2004
Posts: 675 |
Just to add an alternative which is a bit special but cute:
Maniac Mansion and ZMK use some RLE depacker that resides in ZP and encodes most-common bytes, which I felt works rather well for graphics. I wrote something about that here:
https://www.pagetable.com/?p=603
A byte of < $40 = count of raw bytes following
$3f to < $80 represents the length of the run of the byte to follow (+$3f of course).
$7f to < $a0 is the number of runs of most common byte 1,
$9f to < $c0 and $bf to <$e0 and $df-$ff for common byte 2, 3 and 4 respectively. |
| |
ChristopherJam
Registered: Aug 2004
Posts: 1381 |
ooh, I like that one enthusi.
Really interesting the historical divide between packers and crunchers, especially the way intros tended to be displayed between one and the other. It does indeed reduce the impact on runlength limits considerably. Not sure why I ever bothered trying to improve the ratio of my higher ratio crunchers for empty spaces now :P |
| |
Martin Piper
Registered: Nov 2007
Posts: 645 |
In the old days, dictionary based compression would be quite slow and memory hungry. The usual way of speeding it up would be to limit the distance searched and limit the dictionary.
Using a RLE compression pass before a dictionary compression stage would reduce the total input data, it would reduce the strain on the dictionary, and would pass some benefit along to the dictionary based compression.
As dictionary algorithms improve, and also as time and memory becomes less of an issue, the dictionary size can be increased and this means there is less benefit to shrinking the input data with a RLE pre-pass. |
| |
Fungus
Registered: Sep 2002
Posts: 629 |
Main reason for packing first was before crunchers had REU versions with 2mhz or 8mhz support they were slow as hell and took many hours. So smaller input files crunched way faster. Also crunching large files tended to bug out. They generally were smaller too, so it became the norm. |
