Submit 8086-optimized decompressor using jump tables

Because the 8086's BIU is more efficient and can read a word in
a single operation, we can skew LZSA1 decompression faster on 8086
by using a jump table to eliminate 2 comparions.
This commit is contained in:
mobygamer 2019-07-11 17:22:29 -05:00
parent 638c33b432
commit f123a6d9df
2 changed files with 281 additions and 8 deletions

View File

@ -1,4 +1,4 @@
; lzsa1fta.asm time-efficient decompressor implementation for 8088
; lzsa1fta.asm - time-efficient LZSA1 decompressor implementation for 8088
; Turbo Assembler IDEAL mode dialect; can also be assembled with NASM.
;
; Usual DOS assembler SMALL model assumptions apply. This code:
@ -112,12 +112,12 @@ lzsa1_start:
@@get_match_length:
xchg dx,ax ;dx: match offset ax: original token
and al,0FH ;isolate match length in token (MMMM)
cmp al,0FH ;MATCH_RUN_LEN?
cmp al,0FH ;MATCH_RUN_LEN?
jne @@got_matchlen_short ;no, we have the full match length from the token, go copy
lodsb ;grab extra length byte
add al,012H ;add MIN_MATCH_SIZE + MATCH_RUN_LEN
jnc @@do_long_copy ;if no overflow, we have the entire length
jnc @@do_long_copy ;if no overflow, we have the entire length
jne @@mid_matchlen
lodsw ;grab 16-bit length
@ -125,6 +125,10 @@ lzsa1_start:
jcxz @@done_decompressing ;wait, is it the EOD marker? Exit if so
jmp @@copy_len_preset ;otherwise, do the copy
@@get_long_offset:
lodsw ;Get 2-byte match offset
jmp @@get_match_length
@@got_matchlen_short:
add al,3 ;add MIN_MATCH_SIZE
xchg cx,ax ;copy match length into cx
@ -145,11 +149,6 @@ lzsa1_start:
sub ax,di
ret ;done decompressing, exit to caller
;These are called less often; moved here to optimize the fall-through case
@@get_long_offset:
lodsw ;Get 2-byte match offset
jmp @@get_match_length
;With a confirmed longer match length, we have an opportunity to optimize for
;the case where a single byte is repeated long enough that we can benefit
;from rep movsw to perform the run (instead of rep movsb).

274
asm/8088/LZSA1JMP.ASM Normal file
View File

@ -0,0 +1,274 @@
; lzsa1jmp.asm - time-efficient LZSA1 decompressor implementation for 8086.
; Turbo Assembler IDEAL mode dialect; can also be assembled with NASM.
;
; Usual DOS assembler SMALL model assumptions apply. This code:
; - Assumes it was invoked via NEAR call (change RET to RETF for FAR calls)
; - Is interrupt-safe
; - Is not re-entrant (do not decompress while already running decompression)
; - Trashes all data and segment registers
;
; Copyright (C) 2019 Jim Leonard, Emmanuel Marty
;
; This software is provided 'as-is', without any express or implied
; warranty. In no event will the authors be held liable for any damages
; arising from the use of this software.
;
; Permission is granted to anyone to use this software for any purpose,
; including commercial applications, and to alter it and redistribute it
; freely, subject to the following restrictions:
;
; 1. The origin of this software must not be misrepresented; you must not
; claim that you wrote the original software. If you use this software
; in a product, an acknowledgment in the product documentation would be
; appreciated but is not required.
; 2. Altered source versions must be plainly marked as such, and must not be
; misrepresented as being the original software.
; 3. This notice may not be removed or altered from any source distribution.
IDEAL
P8086
SEGMENT CODE para public
ASSUME cs:CODE, ds:CODE
PUBLIC lzsa1_decompress_speed_jumptable
; ---------------------------------------------------------------------------
; Decompress raw LZSA1 block
; inputs:
; * ds:si: raw LZSA1 block
; * es:di: output buffer
; output:
; * ax: decompressed size
; ---------------------------------------------------------------------------
;Jump table for handling LLL bits in initial LZSA1 tokens.
;Previous code would SHR val,4 to get a count from 0 to 7, then rep movsb.
;We can overload the shift operation into a jump table that jumps directly
;to optimized copying routine for 0-7 bytes. Must declare in code segment.
;Note: If this looks strange for declaring a jump table, that's because it
;is a workaround for the Turbo Pascal harness that tests it. Turbo Pascal
;treats OFFSET (label) as a relocatble item and throws an error, so we fool
;it by building the table with absolute EQU/literals instead.
L0b EQU OFFSET check_offset_size
L1b EQU OFFSET copy1b
L2b EQU OFFSET copy2b
L3b EQU OFFSET copy3b
L4b EQU OFFSET copy4b
L5b EQU OFFSET copy5b
L6b EQU OFFSET copy6b
L7b EQU OFFSET need_length_byte
copytable DW L0b,L0b,L0b,L0b,L0b,L0b,L0b,L0b
DW L1b,L1b,L1b,L1b,L1b,L1b,L1b,L1b
DW L2b,L2b,L2b,L2b,L2b,L2b,L2b,L2b
DW L3b,L3b,L3b,L3b,L3b,L3b,L3b,L3b
DW L4b,L4b,L4b,L4b,L4b,L4b,L4b,L4b
DW L5b,L5b,L5b,L5b,L5b,L5b,L5b,L5b
DW L6b,L6b,L6b,L6b,L6b,L6b,L6b,L6b
DW L7b,L7b,L7b,L7b,L7b,L7b,L7b,L7b
PROC lzsa1_decompress_speed_jumptable NEAR
lzsa1_start:
push di ;remember decompression offset
cld ;ensure string ops move forward
xor cx,cx
@@decode_token:
xchg cx,ax ;clear ah (cx = 0 from match copy's rep movsb)
lodsb ;read token byte: O|LLL|MMMM
mov dx,ax ;copy our token to dl for later MMMM handling
and al,070H ;isolate literals length in token (LLL)
jz check_offset_size ;if LLL=0, we have no literals; goto match
; Jump to short copy routine for LLL=1 though 6, need_length_byte for LLL=7
mov bx,ax ;prep for table lookup (must copy, don't XCHG!)
jmp [cs:copytable+bx]
need_length_byte:
lodsb ;grab extra length byte
add al,07H ;add LITERALS_RUN_LEN
jnc @@got_literals_exact ;if no overflow, we have full count
je @@big_literals
@@mid_literals:
lodsb ;grab single extra length byte
inc ah ;add 256
xchg cx,ax ;with longer counts, we can save some time
shr cx,1 ;by doing a word copy instead of a byte copy.
rep movsw ;We don't need to account for overlap because
adc cx,0 ;source for literals isn't the output buffer.
rep movsb
jmp check_offset_size
@@big_literals:
lodsw ;grab 16-bit extra length
xchg cx,ax ;with longer counts, we can save some time
shr cx,1 ;by doing a word copy instead of a byte copy.
rep movsw
adc cx,0
rep movsb
jmp check_offset_size
; Used for counts 7-248. In test data, average value around 1Ah. YMMV.
@@got_literals_exact:
xchg cx,ax
rep movsb ;copy cx literals from ds:si to es:di
jmp check_offset_size
;Literal copy sequence for lengths 1-6:
copy6b: movsb
copy5b: movsb
copy4b: movsb
copy3b: movsb
copy2b: movsb
copy1b: movsb
;Literals done; fall through to match offset determination
check_offset_size:
test dl,dl ;check match offset size in token (O bit)
js @@get_long_offset ;load absolute 16-bit match offset
mov ah,0ffh ;set up high byte
lodsb ;load low byte
@@get_match_length:
xchg dx,ax ;dx: match offset ax: original token
and al,0FH ;isolate match length in token (MMMM)
cmp al,0FH ;MATCH_RUN_LEN?
jne @@got_matchlen_short ;no, we have the full match length from the token, go copy
lodsb ;grab extra length byte
add al,012H ;add MIN_MATCH_SIZE + MATCH_RUN_LEN
jnc @@do_long_copy ;if no overflow, we have the entire length
jne @@mid_matchlen
lodsw ;grab 16-bit length
xchg cx,ax ;get ready to do a long copy
jcxz @@done_decompressing ;wait, is it the EOD marker? Exit if so
jmp @@copy_len_preset ;otherwise, do the copy
@@got_matchlen_short:
add al,3 ;add MIN_MATCH_SIZE
xchg cx,ax ;copy match length into cx
mov bp,ds ;save ds
mov ax,es
mov ds,ax ;ds=es
xchg ax,si ;save si
mov si,di ;ds:si now points at back reference in output data
add si,dx
rep movsb ;copy match
xchg si,ax ;restore si
mov ds,bp ;restore ds
jmp @@decode_token ;go decode another token
@@done_decompressing:
pop ax ;retrieve the original decompression offset
xchg di,ax ;compute decompressed size
sub ax,di
ret ;done decompressing, exit to caller
;These are called less often; moved here to optimize the fall-through case
@@get_long_offset:
lodsw ;Get 2-byte match offset
jmp @@get_match_length
;With a confirmed longer match length, we have an opportunity to optimize for
;the case where a single byte is repeated long enough that we can benefit
;from rep movsw to perform the run (instead of rep movsb).
@@mid_matchlen:
lodsb ;grab single extra length byte
inc ah ;add 256
@@do_long_copy:
xchg cx,ax ;copy match length into cx
@@copy_len_preset:
push ds ;save ds
mov bp,es
mov ds,bp ;ds=es
mov bp,si ;save si
mov si,di ;ds:si now points at back reference in output data
add si,dx
cmp dx,-2 ;do we have a byte/word run to optimize?
jae @@do_run ;perform a run
;You may be tempted to change "jae" to "jge" because DX is a signed number.
;Don't! The total window is 64k, so if you treat this as a signed comparison,
;you will get incorrect results for offsets over 32K.
;If we're here, we have a long copy and it isn't byte-overlapping (if it
;overlapped, we'd be in @@do_run_1) So, let's copy faster with REP MOVSW.
;This won't affect 8088 that much, but it speeds up 8086 and higher.
shr cx,1
rep movsw
adc cx,0
rep movsb
mov si,bp ;restore si
pop ds
jmp @@decode_token ;go decode another token
@@do_run:
je @@do_run_2 ;fall through to byte (common) if not word run
@@do_run_1:
lodsb ;load first byte of run into al
mov ah,al
shr cx,1
rep stosw ;perform word run
adc cx,0
rep stosb ;finish word run
mov si,bp ;restore si
pop ds
jmp @@decode_token ;go decode another token
@@do_run_2:
lodsw ;load first word of run
shr cx,1
rep stosw ;perform word run
adc cx,0 ;despite 2-byte offset, compressor might
rep stosb ;output odd length. better safe than sorry.
mov si,bp ;restore si
pop ds
jmp @@decode_token ;go decode another token
ENDP lzsa1_decompress_speed_jumptable
ENDS CODE
END
;Speed optimization history (decompression times in microseconds @ 4.77 MHz):
; original E. Marty code shuttle 123208 alice 65660 robotron 407338 ***
; table for shr al,4 shuttle 120964 alice 63230 robotron 394733 +++
; push/pop to mov/mov shuttle 118176 alice 61835 robotron 386762 +++
; movsw for literalcpys shuttle 124102 alice 64908 robotron 400220 --- rb
; stosw for byte runs shuttle 118897 alice 65040 robotron 403518 --- rb
; better stosw for runs shuttle 117712 alice 65040 robotron 403343 +--
; disable RLE by default shuttle 116924 alice 60783 robotron 381226 +++
; optimize got_matchlen shuttle 115294 alice 59588 robotron 374330 +++
; fall through to getML shuttle 113258 alice 59572 robotron 372004 +++
; fall through to midLI shuttle 113258 alice 59572 robotron 375060 ..- rb
; fall through midMaLen shuttle 113247 alice 59572 robotron 372004 +.+
; movsw for litlen > 255 shuttle 113247 alice 59572 robotron 371612 ..+
; rep stosw for long runs shuttle 113247 alice 59572 robotron 371612 ...
; rep movsw for long cpys shuttle 113247 alice 59572 robotron 371035 ..+
; xchg/dec ah -> mov ah,val shuttle 112575 alice 59272 robotron 369198 +++
; force >12h len.to longcpy shuttle 101998 alice 59266 robotron 364459 +.+
; more efficient run branch shuttle 102239 alice 59297 robotron 364716 --- rb
; even more eff. run branch shuttle 101998 alice 59266 robotron 364459 ***
; BUGFIX - bad sign compare shuttle 101955 alice 59225 robotron 364117 +++
; reverse 16-bit len compar shuttle 102000 alice 59263 robotron 364460 --- rb
; jcxz for EOD detection no change to speed, but is 1 byte shorter +++
; force movsw for literals shuttle 107183 alice 62555 robotron 379524 --- rb
; defer shr4 until necessry shuttle 102069 alice 60236 robotron 364096 ---
; skip literals if LLL=0 shuttle 98655 alice 57849 robotron 363358 ---
; fall through to mid_liter shuttle 98595 alice 57789 robotron 361998 +++
; == jumptable experiments begin ==
; jumptable for small copys shuttle 101594 alice 61078 robotron 386018 ---
; start:xchg instead of mov shuttle 100948 alice 60467 robotron 381112 +++
; use table for LLL=0 check shuttle 106972 alice 63333 robotron 388304 --- rb
; jmptbl to fallthrough mov shuttle 102532 alice 60760 robotron 383070 ---
; cpy fallthrough check_ofs shuttle 98939 alice 58917 robotron 371019 +**
; single jumptable jump shuttle 97528 alice 57264 robotron 362194 ++*
; conditional check for L=7 shuttle 98610 alice 58521 robotron 368153 --- rb
; defer add MIN_MATCH_SIZE shuttle 97207 alice 57200 robotron 362884 ++*