chiptune_player: implemented stepwise lz4
better, but still not acceptable
This commit is contained in:
Vince Weaver
parent
2ebf6c37e8
commit
62391aa514
|
|
@ -0,0 +1,219 @@
|
|||
; Stepwise LZ4 data decompressor for Apple II
|
||||
; Only decodes X tokens before returning
|
||||
|
||||
; Code by Peter Ferrie (qkumba) (peter.ferrie@gmail.com)
|
||||
; "LZ4 unpacker in 143 bytes (6502 version) (2013)"
|
||||
; http://pferrie.host22.com/misc/appleii.htm
|
||||
; This is that code, but with comments and labels added for clarity.
|
||||
; I also found a bug when decoding with runs of multiples of 256
|
||||
; which has since been fixed upstream.
|
||||
|
||||
; For LZ4 reference see
|
||||
; https://github.com/lz4/lz4/wiki/lz4_Frame_format.md
|
||||
|
||||
; LZ4 summary:
|
||||
;
|
||||
; HEADER:
|
||||
; Should: check for magic number 04 22 4d 18
|
||||
; FLG: 64 in our case (01=version, block.index=1, block.checksum=0
|
||||
; size=0, checksum=1, reserved
|
||||
; MAX Blocksize: 40 (64kB)
|
||||
; HEADER CHECKSUM: a7
|
||||
; BLOCK HEADER: 4 bytes (le) If highest bit set, uncompressed!
|
||||
; BLOCKS:
|
||||
; Token byte. High 4-bits literal length, low 4-bits copy length
|
||||
; + If literal length==15, then following byte gets added to length
|
||||
; If that byte was 255, then keep adding bytes until not 255
|
||||
; + The literal bytes follow. There may be zero of them
|
||||
; + Next is block copy info. little-endian 2-byte offset to
|
||||
; be subtracted from current read position indicating source
|
||||
; + The low 4-bits of the token are the copy length, which needs
|
||||
; 4 added to it. As with the literal length, if it is 15 then
|
||||
; you read a byte and add (and if that byte is 255, keep adding)
|
||||
|
||||
;LZ4_SRC EQU $00
|
||||
;LZ4_DST EQU $02
|
||||
;LZ4_END EQU $04
|
||||
;COUNT EQU $06
|
||||
;DELTA EQU $08
|
||||
|
||||
;UNPACK_BUFFER EQU $5E00 ; offset of first unpacked byte
|
||||
|
||||
|
||||
;======================
|
||||
; LZ4 decode
|
||||
;======================
|
||||
; input buffer in LZ4_SRC
|
||||
; output buffer hardcoded still
|
||||
; size in ENDH:ENDL
|
||||
|
||||
lz4_decode_setup:
|
||||
|
||||
lda LZ4_SRC ; packed data offset
|
||||
clc
|
||||
adc LZ4_END
|
||||
sta LZ4_END
|
||||
lda LZ4_SRC+1
|
||||
adc LZ4_END+1
|
||||
sta LZ4_END+1
|
||||
|
||||
lda #>UNPACK_BUFFER ; original unpacked data offset
|
||||
sta LZ4_DST+1
|
||||
lda #<UNPACK_BUFFER
|
||||
sta LZ4_DST
|
||||
|
||||
rts
|
||||
|
||||
lz4_decode_step:
|
||||
ldy #0 ; used to index, always zero
|
||||
|
||||
parsetoken:
|
||||
jsr getsrc ; get next token
|
||||
pha ; save for later (need bottom 4 bits)
|
||||
|
||||
lsr ; number of literals in top 4 bits
|
||||
lsr ; so shift into place
|
||||
lsr
|
||||
lsr
|
||||
beq copymatches ; if zero, then no literals
|
||||
; jump ahead and copy
|
||||
|
||||
jsr buildcount ; add up all the literal sizes
|
||||
; result is in ram[count+1]-1:A
|
||||
tax ; now in ram[count+1]-1:X
|
||||
jsr docopy ; copy the literals
|
||||
|
||||
lda LZ4_SRC ; 16-bit compare
|
||||
cmp LZ4_END ; to see if we have reached the end
|
||||
lda LZ4_SRC+1
|
||||
sbc LZ4_END+1
|
||||
bcs done
|
||||
|
||||
copymatches:
|
||||
jsr getsrc ; get 16-bit delta value
|
||||
sta DELTA
|
||||
jsr getsrc
|
||||
sta DELTA+1
|
||||
|
||||
pla ; restore token
|
||||
and #$0f ; get bottom 4 bits
|
||||
; match count. 0 means 4
|
||||
; 15 means 19+, must be calculated
|
||||
|
||||
jsr buildcount ; add up count bits, in ram[count+1]-:A
|
||||
|
||||
clc
|
||||
adc #4 ; adjust count by 4 (minmatch)
|
||||
|
||||
tax ; now in ramp[count+1]-1:X
|
||||
|
||||
beq copy_no_adjust ; BUGFIX, don't increment if
|
||||
; exactly a multiple of 0x100
|
||||
bcc copy_no_adjust
|
||||
|
||||
inc COUNT+1 ; increment if we overflowed
|
||||
copy_no_adjust:
|
||||
|
||||
lda LZ4_SRC+1 ; save src on stack
|
||||
pha
|
||||
lda LZ4_SRC
|
||||
pha
|
||||
|
||||
sec ; subtract delta
|
||||
lda LZ4_DST ; from destination, make new src
|
||||
sbc DELTA
|
||||
sta LZ4_SRC
|
||||
lda LZ4_DST+1
|
||||
sbc DELTA+1
|
||||
sta LZ4_SRC+1
|
||||
|
||||
jsr docopy ; do the copy
|
||||
|
||||
pla ; restore the src
|
||||
sta LZ4_SRC
|
||||
pla
|
||||
sta LZ4_SRC+1
|
||||
|
||||
clc ; c==0 means not done
|
||||
|
||||
rts
|
||||
|
||||
; jmp parsetoken ; back to parsing tokens
|
||||
|
||||
done:
|
||||
pla
|
||||
sec ; c=1 means done
|
||||
|
||||
rts
|
||||
|
||||
|
||||
|
||||
;=========
|
||||
; getsrc
|
||||
;=========
|
||||
; gets byte from src into A, increments pointer
|
||||
getsrc:
|
||||
lda (LZ4_SRC), Y ; get a byte from src
|
||||
inc LZ4_SRC ; increment pointer
|
||||
bne done_getsrc ; update 16-bit pointer
|
||||
inc LZ4_SRC+1 ; on 8-bit overflow
|
||||
done_getsrc:
|
||||
rts
|
||||
|
||||
;============
|
||||
; buildcount
|
||||
;============
|
||||
buildcount:
|
||||
ldx #1 ; high count starts at 1
|
||||
stx COUNT+1 ; (loops at zero?)
|
||||
cmp #$0f ; if LITERAL_COUNT < 15, we are done
|
||||
bne done_buildcount
|
||||
buildcount_loop:
|
||||
sta COUNT ; save LITERAL_COUNT (15)
|
||||
jsr getsrc ; get the next byte
|
||||
tax ; put in X
|
||||
clc
|
||||
adc COUNT ; add new byte to old value
|
||||
bcc bc_8bit_oflow ; if overflow, increment high byte
|
||||
inc COUNT+1
|
||||
bc_8bit_oflow:
|
||||
inx ; check if read value was 255
|
||||
beq buildcount_loop ; if it was, keep looping and adding
|
||||
done_buildcount:
|
||||
rts
|
||||
|
||||
;============
|
||||
; getput
|
||||
;============
|
||||
; gets a byte, then puts the byte
|
||||
getput:
|
||||
jsr getsrc
|
||||
; fallthrough to putdst
|
||||
|
||||
;=============
|
||||
; putdst
|
||||
;=============
|
||||
; store A into destination
|
||||
putdst:
|
||||
sta (LZ4_DST), Y ; store A into destination
|
||||
inc LZ4_DST ; increment 16-bit pointer
|
||||
bne putdst_end ; if overflow, increment top byte
|
||||
inc LZ4_DST+1
|
||||
putdst_end:
|
||||
rts
|
||||
|
||||
;=============================
|
||||
; docopy
|
||||
;=============================
|
||||
; copies ram[count+1]-1:X bytes
|
||||
; from src to dst
|
||||
docopy:
|
||||
|
||||
docopy_loop:
|
||||
jsr getput ; get/put byte
|
||||
dex ; decrement count
|
||||
bne docopy_loop ; if not zero, loop
|
||||
dec COUNT+1 ; if zero, decrement high byte
|
||||
bne docopy_loop ; if not zero, loop
|
||||
|
||||
rts
|
||||
|
|
@ -32,7 +32,7 @@ CHIPTUNE_PLAYER: chiptune_player.o
|
|||
chiptune_player.o: chiptune_player.s \
|
||||
../asm_routines/mockingboard.s \
|
||||
../asm_routines/dos33_routines.s \
|
||||
../asm_routines/lz4_decode.s \
|
||||
../asm_routines/lz4_decode_step.s \
|
||||
rasterbars.s volume_bars.s interrupt_handler.s \
|
||||
chip_title.inc zp.inc
|
||||
ca65 -o chiptune_player.o chiptune_player.s -l chiptune_player.lst
|
||||
|
|
|
|||
|
|
@ -2,3 +2,5 @@
|
|||
+ Right/Left arrows on screen
|
||||
+ Keyboard, Right/Left/Pause
|
||||
+ Calculate maximum decode time for songs
|
||||
|
||||
+ Put graphics update in interrupt routine, put debug in normal space
|
||||
|
|
|
|||
Binary file not shown.
|
|
@ -293,12 +293,9 @@ read_size EQU $4000
|
|||
adc #0
|
||||
sta LZ4_SRC+1
|
||||
|
||||
; Fall through to next_subsong
|
||||
|
||||
;=================
|
||||
; next sub-song
|
||||
;=================
|
||||
next_subsong:
|
||||
; Decompress first chunks
|
||||
|
||||
lda #$0
|
||||
sta COPY_OFFSET
|
||||
lda #$3
|
||||
|
|
@ -306,6 +303,27 @@ next_subsong:
|
|||
lda #$20
|
||||
sta DECODER_STATE
|
||||
|
||||
jsr setup_next_subsong
|
||||
|
||||
jsr lz4_decode_setup
|
||||
|
||||
our_lz4_loop:
|
||||
jsr lz4_decode_step
|
||||
bcc our_lz4_loop
|
||||
|
||||
rts
|
||||
|
||||
;=================
|
||||
; next sub-song
|
||||
;=================
|
||||
setup_next_subsong:
|
||||
; lda #$0
|
||||
; sta COPY_OFFSET
|
||||
; lda #$3
|
||||
; sta CHUNKSIZE
|
||||
; lda #$20
|
||||
; sta DECODER_STATE
|
||||
|
||||
ldy #0
|
||||
|
||||
lda (LZ4_SRC),Y ; get next size value
|
||||
|
|
@ -322,7 +340,7 @@ next_subsong:
|
|||
adc #0
|
||||
sta LZ4_SRC+1
|
||||
|
||||
jsr lz4_decode ; decode
|
||||
; jsr lz4_decode ; decode
|
||||
|
||||
; tail-call?
|
||||
|
||||
|
|
@ -430,7 +448,7 @@ krw_file:
|
|||
.include "../asm_routines/gr_setpage.s"
|
||||
.include "../asm_routines/dos33_routines.s"
|
||||
.include "../asm_routines/gr_hlin.s"
|
||||
.include "../asm_routines/lz4_decode.s"
|
||||
.include "../asm_routines/lz4_decode_step.s"
|
||||
.include "rasterbars.s"
|
||||
.include "volume_bars.s"
|
||||
.include "interrupt_handler.s"
|
||||
|
|
|
|||
|
|
@ -166,7 +166,12 @@ handle_copy:
|
|||
jmp increment_offset ; 3
|
||||
|
||||
decompress_step:
|
||||
; TODO
|
||||
lda LZ4_DONE
|
||||
bne increment_offset
|
||||
|
||||
jsr lz4_decode_step
|
||||
bcc increment_offset
|
||||
inc LZ4_DONE
|
||||
|
||||
;==============================================
|
||||
; incremement offset. If 0 move to next chunk
|
||||
|
|
@ -193,13 +198,22 @@ increment_offset:
|
|||
start_c:
|
||||
lda #1
|
||||
sta CHUNKSIZE
|
||||
|
||||
; setup next three chunks of song
|
||||
jsr setup_next_subsong ; and decompress next ; 6
|
||||
jsr lz4_decode_setup
|
||||
lda #0
|
||||
sta LZ4_DONE
|
||||
|
||||
jmp back_to_first_reg
|
||||
|
||||
wraparound_to_a:
|
||||
; setup next three chunks of song
|
||||
|
||||
jsr next_subsong ; and decompress next ; 6
|
||||
|
||||
lda #$0
|
||||
sta COPY_OFFSET
|
||||
lda #$3
|
||||
sta CHUNKSIZE
|
||||
lda #$20
|
||||
sta DECODER_STATE
|
||||
|
||||
;==============================
|
||||
; After 14th reg, reset back to
|
||||
|
|
@ -224,12 +238,12 @@ back_to_first_reg_a:
|
|||
back_to_first_reg_c:
|
||||
lda #>(UNPACK_BUFFER+$2A00) ; in proper chunk (1 of 3) ; 2
|
||||
|
||||
update_r0_pointer:
|
||||
sta INH ; update r0 pointer ; 3
|
||||
|
||||
;============
|
||||
; 18
|
||||
|
||||
update_r0_pointer:
|
||||
sta INH ; update r0 pointer ; 3
|
||||
|
||||
;=================================
|
||||
; Finally done with this interrupt
|
||||
|
|
|
|||
|
|
@ -1,11 +1,10 @@
|
|||
.define EQU =
|
||||
|
||||
LZ4_SRC EQU $00
|
||||
LZ4_DST EQU $02
|
||||
LZ4_END EQU $04
|
||||
COUNT EQU $06
|
||||
DELTA EQU $08
|
||||
|
||||
LZ4_SRC EQU $00
|
||||
LZ4_DST EQU $02
|
||||
LZ4_END EQU $04
|
||||
COUNT EQU $06
|
||||
DELTA EQU $08
|
||||
|
||||
;; Zero page monitor routines addresses
|
||||
|
||||
|
|
@ -85,6 +84,7 @@ DECODER_STATE EQU $7F
|
|||
;NUM1L EQU $7E
|
||||
;NUM1H EQU $7F
|
||||
CHUNKSIZE EQU $80
|
||||
LZ4_DONE EQU $81
|
||||
|
||||
A_COLOR EQU $83
|
||||
B_COLOR EQU $84
|
||||
|
|
|
|||
|
|
@ -109,6 +109,7 @@ int main(int argc, char **argv) {
|
|||
FILE *fff;
|
||||
int size;
|
||||
short orgoff,paksize,pakoff;
|
||||
int token_count=0;
|
||||
|
||||
init_6502();
|
||||
|
||||
|
|
@ -186,7 +187,8 @@ int main(int argc, char **argv) {
|
|||
y=0; // used for offset //ldy #0
|
||||
|
||||
parsetoken:
|
||||
printf("LOAD TOKEN: ");
|
||||
token_count++;
|
||||
printf("LOAD TOKEN %d: ",token_count);
|
||||
getsrc(); // jsr getsrc
|
||||
// get token
|
||||
pha(); // save for later // pha
|
||||
|
|
@ -290,6 +292,8 @@ done:
|
|||
|
||||
printf("Out size=%d\n",out_size);
|
||||
|
||||
printf("Total tokens: %d\n",token_count);
|
||||
|
||||
fwrite(&ram[ORGOFFSET],1,out_size,fff);
|
||||
|
||||
fclose(fff);
|
||||
|
|
|
|||
Loading…
Reference in New Issue