Skip to content

Commit

Permalink
Found kkrunchy_k7 023.a4/asm07 (most recent version) source code
Browse files Browse the repository at this point in the history
  • Loading branch information
@rygorous
rygorous committed Apr 18, 2012
1 parent fd16b03 commit 5ddf84f
Show file tree
Hide file tree
Showing 7 changed files with 130 additions and 181 deletions.
2 changes: 1 addition & 1 deletion kkrunchy_k7/_config.hpp
View file
Expand Up @@ -14,4 +14,4 @@
#define sINTRO 0 // compile for small size
#define sDEBUG 1 // include debug code even in release build
#define sAPPNAME "kkrunchy" // application name for window title
#define sVERSION "0.23 alpha 3" // version number string
#define sVERSION "0.23a4/asm07" // version number string
40 changes: 10 additions & 30 deletions kkrunchy_k7/depack2.asm
View file
Expand Up @@ -4,7 +4,7 @@

bits 32

%define NBUFFERS 22
%define NBUFFERS 20
%define BUFFER dataArea.buffer

struc dataArea
Expand Down Expand Up @@ -243,9 +243,9 @@ DisUnfilter:
cmp al, 0x04
jne .nosib
xchg esi, [ebp+BUFFER+21*4]
xchg esi, [ebp+BUFFER+19*4]
movsb
xchg esi, [ebp+BUFFER+21*4]
xchg esi, [ebp+BUFFER+19*4]
.nosib mov dl, ch
and dl, 0xc0
Expand Down Expand Up @@ -276,13 +276,8 @@ DisUnfilter:
inc ebx
.nomr5 xchg esi, [ebp+BUFFER+13*4+ebx*4]
lodsd
dec esi
xchg esi, [ebp+BUFFER+13*4+ebx*4]
xchg esi, [ebp+BUFFER+18*4]
shl eax, 8
lodsb
xchg esi, [ebp+BUFFER+18*4]
ror eax, 8
bswap eax
stosd
cmp word [ebp+dataArea.codebuf], 0x24ff
Expand All @@ -301,15 +296,8 @@ DisUnfilter:
jnz .noad
xchg esi, [ebp+BUFFER+15*4]
lodsd
dec esi
movsd
xchg esi, [ebp+BUFFER+15*4]
xchg esi, [ebp+BUFFER+19*4]
shl eax, 8
lodsb
xchg esi, [ebp+BUFFER+19*4]
ror eax, 8
stosd
jmp .main
.noad dec cl
Expand All @@ -324,12 +312,8 @@ DisUnfilter:
cmp byte [edi-1], 0xe8
je .dwcal
xchg esi, [ebp+BUFFER+20*4]
lodsw
xchg esi, [ebp+BUFFER+20*4]
xchg esi, [ebp+BUFFER+17*4]
shl eax, 16
lodsw
lodsd
xchg esi, [ebp+BUFFER+17*4]
shr eax, 1
Expand All @@ -350,14 +334,10 @@ DisUnfilter:
mov eax, [ebp+dataArea.funcTable+eax*4]
jmp short .storad
.dcesc xchg esi, [ebp+BUFFER+20*4]
lodsw
xchg esi, [ebp+BUFFER+20*4]
xchg esi, [ebp+BUFFER+16*4]
shl eax, 16
lodsw
xchg esi, [ebp+BUFFER+16*4]
.dcesc xchg esi, [ebp+BUFFER+18*4]
lodsd
xchg esi, [ebp+BUFFER+18*4]

mov ebx, [ebp+dataArea.funcTable]
mov [ebp+dataArea.funcTable+ebx*4], eax
.dcinc inc byte [ebp+dataArea.funcTable]
Expand Down
99 changes: 38 additions & 61 deletions kkrunchy_k7/depacker.asm
View file
Expand Up @@ -8,21 +8,20 @@ bits 32

section .bss

%define MEMSHIFT 18
%define MEMSHIFT 23
%define MEM (1<<MEMSHIFT)
%define NMODEL 9
%define NINPUT 40
%define NWEIGHT (256+256+16)
%define NMODEL 11
%define NINPUT 48
%define NWEIGHT (256+256+16+128)
%define MAXLEN 2047
%define USEAPM 1
%define APMSIZE 1024
%define APMSIZE 8192

struc ContextModel
.cpr resd 1
.cps resd 1
.ctx resd 1
.st resd 1
.t resb MEM*2
.size:
endstruc

Expand Down Expand Up @@ -53,7 +52,8 @@ struc Work
.APMi resd 1
%endif
.cm resb ContextModel.size * NMODEL
.match resd MEM
.match resd MEM/16
.modelMem resb MEM
.tx resw NINPUT
.wx resw NINPUT*NWEIGHT
.tx2 resw 4
Expand Down Expand Up @@ -127,9 +127,9 @@ squash:

contextHash:
mov edx, eax
shr edx, MEMSHIFT-1
and eax, ((MEM / 2) - 1) & ~1
lea eax, [esi+ContextModel.t+eax*4]
shr edx, 24
and eax, ((MEM / 4) - 1) & ~1
lea eax, [ebp+Work.modelMem+eax*4]
cmp dl, [eax]
jne .notequal1
Expand Down Expand Up @@ -159,7 +159,7 @@ contextHash:
; in: eax=err
; esi=t
; edi=w
; ecx=n/8
; ecx=n/4
; out: esi,edi advanced.
train:
movd mm0, eax
Expand All @@ -168,8 +168,6 @@ train:
pcmpeqb mm1, mm1
psrlw mm1, 15

%if 1
add ecx, ecx
.lp:
movq mm3, [esi]
movq mm2, [edi]
Expand All @@ -182,28 +180,6 @@ train:
add esi, byte 8
add edi, byte 8
loop .lp
%else
.lp:
movq mm3, [esi]
movq mm5, [esi+8]
movq mm2, [edi]
movq mm4, [edi+8]
paddsw mm3, mm3
paddsw mm5, mm5
pmulhw mm3, mm0
pmulhw mm5, mm0
paddsw mm3, mm1
paddsw mm5, mm1
psraw mm3, 1
psraw mm5, 1
paddsw mm2, mm3
paddsw mm4, mm5
movq [edi], mm2
movq [edi+8], mm4
add esi, byte 16
add edi, byte 16
loop .lp
%endif
ret

Expand Down Expand Up @@ -406,7 +382,7 @@ STAGE0ENTRY:
xor eax, eax
call contextHash
mov [esi+ContextModel.cps], eax
add esi, ContextModel.size
add esi, byte ContextModel.size
loop .ctxModelLoop

%if USEAPM
Expand Down Expand Up @@ -448,7 +424,7 @@ STAGE0ENTRY:
mov eax, [ebp+Work.zeroprob]
neg edx
inc eax
and edx, 4095
shr edx, 32-12 ; eqv. to and edx, 4095 here
add eax, edx
shr eax, 1
mov [ebp+Work.zeroprob], eax
Expand Down Expand Up @@ -570,7 +546,7 @@ STAGE0ENTRY:
imul edi, byte (NINPUT*2)
lea edi, [ebp+Work.wx+edi]
mov esi, WorkData+Work.tx
push byte (NINPUT/8)
push byte (NINPUT/4)
pop ecx
call train
Expand Down Expand Up @@ -625,14 +601,20 @@ STAGE0ENTRY:
; fnv hash function
; assumes edi=_bufPtr
mov eax, 0x811c9dc5
inc ecx
imul eax, ecx
dec ecx
mov bl, [masks+ecx-1]
mov dh, [bitm+ecx-1]
.hashnext:
dec edi
shr bl, 1
jnc .hashnotset
xor al, [edi]
mov dl, [edi]
and dl, dh
xor al, dl
imul eax, 0x01000193
jmp short .hashnext

Expand All @@ -643,7 +625,7 @@ STAGE0ENTRY:
; match processing
; assumes esi points to "match"
and eax, MEM-1
and eax, (MEM/16)-1
mov edi, [ebp+Work.dst]
xchg edi, [esi+eax*4] ; now edi=Match[h]
mov esi, [ebp+Work.dst]
Expand Down Expand Up @@ -735,7 +717,7 @@ STAGE0ENTRY:
cmp eax, 400
jbe .contextmodels
mov dword [ebp+Work.ctx], 526
mov dword [ebp+Work.ctx], 512+14
jmp .mix

; context models
Expand Down Expand Up @@ -821,7 +803,7 @@ STAGE0ENTRY:
cmp ebx, byte 1
sbb dword [ebp+Work.ctx+8], byte -1

add esi, ContextModel.size
add esi, byte ContextModel.size
dec ecx
jnz near .contextloop
Expand All @@ -833,6 +815,17 @@ STAGE0ENTRY:
inc ah
mov [ebp+Work.ctx+4], eax
xor ebx, ebx
mov eax, [ebp+Work.matchl]
dec eax
cmovs eax, ebx
not bl
cmp eax, ebx
cmova eax, ebx
mov eax, [ebp+Work.runTable+eax*4]
shr eax, 3
add [ebp+Work.ctx+8], eax
.mix:
; perform mixing
push byte 2
Expand All @@ -845,7 +838,6 @@ STAGE0ENTRY:
mov esi, WorkData+Work.tx
pxor mm0, mm0
%if 1
push byte (NINPUT/4)
pop ecx

Expand All @@ -857,23 +849,6 @@ STAGE0ENTRY:
add esi, byte 8
add edi, byte 8
loop .mixdploop
%else
push byte (NINPUT/8)
pop ecx
.mixdploop:
movq mm1, [esi]
movq mm2, [esi+8]
pmaddwd mm1, [edi]
pmaddwd mm2, [edi+8]
psrad mm1, 8
psrad mm2, 8
paddd mm0, mm1
paddd mm0, mm2
add esi, byte 16
add edi, byte 16
loop .mixdploop
%endif
movq mm1, mm0
psrlq mm1, 32
Expand Down Expand Up @@ -971,4 +946,6 @@ section .data
squashTab dw 1, 2, 4, 6, 10, 17, 27, 45, 74, 120, 194
dw 311, 488, 747,1102,1546,2048,2550,2994,3349,3608,3785
dw 3902,3976,4022,4051,4069,4079,4086,4090,4092,4094,4095
masks db 0x1f, 0x07, 0x0a, 0x09, 0x05, 0x03, 0x04, 0x02, 0x01
masks db 0x1f, 0x27, 0x88, 0x07, 0x0a, 0x09, 0x05, 0x03, 0x04, 0x02, 0x01
bitm db 0xff, 0xff, 0xff, 0xe0, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff

0 comments on commit 5ddf84f

Please sign in to comment.