Kosinski compression

Kosinski compression is the name given to a compression format used in Sonic games for the Sega Genesis/Mega Drive. The name first showed up on SHaC and was to pay tribute to the person who cracked the format, Brett Kosinski. It is a variation of the LZSS algorithm.

Kosinski compression is used to compress the following data types:

• Sonic the Hedgehog - 256x256 block mappings.
• Sonic the Hedgehog 2 - level graphics, block mappings and level layouts.
• Mega CD - SUB-CPU BIOS code and data.
• Phantasy Star IV - Text.

Compression Theory

Basic Format

The Kosinski compression scheme is very powerful and includes support for both uncompressed data and two types of run-length encoding to optimally compress repeating patterns.

The compressed data follows the following format:

AA AA BB BB .. AA AA BB BB ..

The A field is referred to as the description field. This is always 2 bytes in length, starting at the beginning of the compressed data, and is followed by the B field, which is referred to as the data field. This goes on for as long as is necessary to cover all the data described in the description field. After that, the pattern repeats until the end of compression sequence is is encountered.

Description Field

The description field is made up of 16 bits in little endian bit (not byte) order, which means that the bytes are in the correct order, but each byte is read backwards. So for example, if the description field is:

FF FE

then the description field, in bits, in the correct order for interpretation, will be:

[1111 1111] [0111 1111]

Now, going from left to right, each bit is interpreted in the following way:

• If the bit is a 1, it indicates uncompressed data.
• If the bit is a 0, followed by a 1, it indicates separate run-length encoding.
• If the bit is a 0, followed by a 0, it indicates inline run-length encoding. The two bits following this give the copy count - 2.

If the end of the description field occurs, a new description field is read from the next 2 bytes of the data. If the end of compression sequence occurs and there are still description field bits left, they are ignored.

Uncompressed Data

If a 1 is read in the description field, the corresponding byte in the data field is already uncompressed and simply copied to the destination as-is. So for example, the following block of Kosinski-compressed data:

FF FF 00 01 02 03 04 05 06 07 08 09 0A 0B 0C 0D 0E 0F

simply produces the following output:

00 01 02 03 04 05 06 07 08 09 0A 0B 0C 0D 0E 0F

because all bits in the description field are 1, indicating that each byte in the data field (00 through 0F) is already uncompressed.

Inline Run-Length Encoding

Inline run-length encoding is indicated by 00 XX in the description field. The XX is incremented by 2 to get the number of bytes to copy, and the corresponding byte in the data field is added to -256 to get the offset from the current position in the uncompressed stream of the byte to start copying from. So for example (including the description field):

F1 FF 25 FF 01 02 03 04 05 06 07 08 09 0A 0B 0C

outputs to:

25 25 25 25 01 02 03 04 05 06 07 08 09 0A 0B 0C

because the description field

F1 FF = [1000 1111] [1111 1111]

starts with an uncompressed [1] value (25), then an inline RLE value repeated for 3 bytes [00 01] with an offset of -$100 + $FF = -1, which means the 1 previous byte (which is 25) is repeated 3 times, followed by 11 more uncompressed [1] bytes (01 through 0C).

Note: Keep in mind that the repeat count in the description field is not how many times the pattern is repeated, but for how many bytes the pattern is repeated. Using a repeat count of 11 (3 + 2 = 5 bytes) and an offset of $FB (-$100 + $FB = -5) only repeats the pattern once. Note that anything farther back than the last 5 bytes cannot be repeated entirely using this method because the largest possible repeat count is 5. Anything that needs to be repeated for longer than 5 bytes needs to use separate run-length encoding.

Separate Run-Length Encoding

Separate run-length encoding is indicated in the description field with a 01 and has 2 or 3 corresponding bytes in the data field, depending on the second byte. The format of these bytes is:

[LLLL LLLL] [HHHH HCCC]

if there are two bytes or

[LLLL LLLL] [HHHH H000] [CCCC CCCC]

if there are three bytes. The three byte format is used when the C bits of the second byte are all unset.

In the two byte format, -8192 + HHHHH * 256 + LLLLLLLL gives the offset from the current position in the uncompressed stream of the byte to start copying from, and CCC + 2 gives the number of bytes to copy. In the three byte format, the offset is calculated in the same way but the copy count is given by CCCCCCCC + 1. Therefore the maximum number of bytes that can be copied is 9 in the 2 byte format and 256 in the 3 byte format.

So for example, in:

F0 F8 40 = [1111 0000] [1111 1000] [0100 0000]

the offset would be -16, or 16 bytes backward since -8192 + 31 * 256 + 240 = -16. Also in this example, 65 bytes would be copied (CCCCCCCC + 1 where the value of the C bits is 64 [40]).

Two special cases occur when the three byte format is used and the third byte is either 0 or 1, and are described below.

End of Compression sequence

The end of the compressed data is indicated by a three byte data field corresponding to separate run-length encoding in which the third byte is 0.

Read Next Description sequence

Alternatively, if the third byte is 1 rather than 0 it indicates that the current description is to be discarded and a new one to be read.

One Final Example

Compressed Data:

  FF 3F 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5
  FC 15 FE C3 44 78 88 98 44 30 FF FF 00 F8 00

Decompression process:

  FF 3F translates to [1111 1111] [1111 1100]
  First 14 data bytes are uncompressed (1111 1111 1111 11)

      Output: 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5

  The next description (00) indicates inline RLE but more description data is needed, so the next description field is read:
  FC 15 translates to [0011 1111] [1010 1000]

  2 bytes (00) are to be copied and the offset is -$100 + $FE = -2:

      Output: 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5 C4 C5

  The next 7 data bytes are uncompressed (11 1111 1)

      Output: 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5 C4 C5 C3 44 78 88 98 44 30

  The next description (01) indicates separate RLE.  The corresponding data bytes (FF FF) translate to an offset of 
  -8192 + 31 * 256 + 255 = -1 and a copy count of 7 + 2 = 9 (i.e. the last 1 output byte repeated for 9 output bytes).

      Output: 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5 C4 C5 C3 44 78 88 98 44 30
              30 30 30 30 30 30 30 30 30

  The next description also indicates separate RLE. The three byte format is used and the third byte is 0, so
  the rest of the description field is ignored and the compression is complete.
  
      Output: 54 3B C4 44 54 33 33 5B 2D 5C 44 5C C4 C5 C4 C5 C3 44 78 88 98 44 30
              30 30 30 30 30 30 30 30 30

Kosinski Moduled compression

Kosinski Moduled compression (KosM compression) is a variant of standard Kosinski compression, used by Sonic 3 & Knuckles. KosM compressed data starts off with a 2-byte header. The upper nibble is the total number of modules minus 1, and the lower three nibbles are the uncompressed size of the last module in bytes (all other modules have a fixed uncompressed size of $1000 bytes). As a result, the entire header word can be read as the total uncompressed size of the data. The only special case occurs when the header is $A000, in which case the game reads it as $8000. After the header are the actual modules. Each module can be read as standard Kosinski-compressed data, and is padded out to a size which is a multiple of $10 bytes.

KosM can be thought of as an archival format for multiple pieces of Kosinski-compressed data, and is used as an alternative to Nemesis compression for compressing art. The maximum uncompressed module size was kept as $1000 bytes because the intermediate RAM buffer to which the data is decompressed before being DMAed to VRAM has a size of $1000 bytes.

Decompression code

An annotated version of the Kosinski decompression code is provided below for reference. The code is taken from Sonic 3 & Knuckles, but the same code is presumably used in all games which use the format.

; ---------------------------------------------------------------------------
; Kosinski decompression subroutine
; Inputs:
; a0 = compressed data location
; a1 = destination
; ---------------------------------------------------------------------------

; =============== S U B R O U T I N E =======================================


Kos_Decomp:
		subq.l	#2,sp	; make space for two bytes on the stack
		move.b	(a0)+,1(sp)
		move.b	(a0)+,(sp)
		move.w	(sp),d5	; copy first description field
		moveq	#$F,d4	; 16 bits in a byte

Kos_Decomp_Loop:
		lsr.w	#1,d5	; bit which is shifted out goes into C flag
		move	sr,d6
		dbf	d4,Kos_Decomp_ChkBit
		move.b	(a0)+,1(sp)
		move.b	(a0)+,(sp)
		move.w	(sp),d5	; get next description field if needed
		moveq	#$F,d4	; reset bit counter

Kos_Decomp_ChkBit:
		move	d6,ccr	; was the bit set?
		bcc.s	Kos_Decomp_RLE	; if not, branch (C flag clear means bit was clear)
		move.b	(a0)+,(a1)+	; otherwise, copy byte as-is
		bra.s	Kos_Decomp_Loop
; ---------------------------------------------------------------------------

Kos_Decomp_RLE:
		moveq	#0,d3
		lsr.w	#1,d5	; get next bit
		move	sr,d6
		dbf	d4,Kos_Decomp_ChkBit2
		move.b	(a0)+,1(sp)
		move.b	(a0)+,(sp)
		move.w	(sp),d5
		moveq	#$F,d4

Kos_Decomp_ChkBit2:
		move	d6,ccr	; was the bit set?
		bcs.s	Kos_Decomp_SeparateRLE	; if it was, branch
		lsr.w	#1,d5	; bit which is shifted out goes into X flag
		dbf	d4,+
		move.b	(a0)+,1(sp)
		move.b	(a0)+,(sp)
		move.w	(sp),d5
		moveq	#$F,d4
+
		roxl.w	#1,d3	; get high repeat count bit (shift X flag in)
		lsr.w	#1,d5
		dbf	d4,+
		move.b	(a0)+,1(sp)
		move.b	(a0)+,(sp)
		move.w	(sp),d5
		moveq	#$F,d4
+
		roxl.w	#1,d3	; get low repeat count bit
		addq.w	#1,d3	; increment repeat count
		moveq	#-1,d2
		move.b	(a0)+,d2	; calculate offset
		bra.s	Kos_Decomp_RLELoop
; ---------------------------------------------------------------------------

Kos_Decomp_SeparateRLE:
		move.b	(a0)+,d0	; get first byte
		move.b	(a0)+,d1	; get second byte
		moveq	#-1,d2
		move.b	d1,d2
		lsl.w	#5,d2
		move.b	d0,d2	; calculate offset
		andi.w	#7,d1	; does a third byte need to be read?
		beq.s	Kos_Decomp_SeparateRLE2	; if it does, branch
		move.b	d1,d3	; copy repeat count
		addq.w	#1,d3	; and increment it

Kos_Decomp_RLELoop:
		move.b	(a1,d2.w),d0
		move.b	d0,(a1)+	; copy appropriate byte
		dbf	d3,Kos_Decomp_RLELoop	; and repeat the copying
		bra.s	Kos_Decomp_Loop
; ---------------------------------------------------------------------------

Kos_Decomp_SeparateRLE2:
		move.b	(a0)+,d1
		beq.s	Kos_Decomp_Done	; 0 indicates end of compressed data
		cmpi.b	#1,d1
		beq.w	Kos_Decomp_Loop	; 1 indicates a new description needs to be read
		move.b	d1,d3	; otherwise, copy repeat count
		bra.s	Kos_Decomp_RLELoop
; ---------------------------------------------------------------------------

Kos_Decomp_Done:
		addq.l	#2,sp	; restore stack pointer to original state
		rts
; End of function Kos_Decomp