#! /bin/sh /usr/share/dpatch/dpatch-run
## 06_old_2xSaImmx_asm.dpatch by Jose Carlos Medeiros <debian@psabs.com.br>
##
## All lines beginning with `## DP:' are a description of the patch.
## DP: No description.
@DPATCH@
diff -urNad visualboyadvance-1.8.0~/src/i386/2xSaImmx.asm visualboyadvance-1.8.0/src/i386/2xSaImmx.asm
--- visualboyadvance-1.8.0~/src/i386/2xSaImmx.asm 2007-05-18 16:22:58.000000000 -0300
+++ visualboyadvance-1.8.0/src/i386/2xSaImmx.asm 2007-05-18 17:19:58.000000000 -0300
@@ -1,38 +1,40 @@
;/*---------------------------------------------------------------------*
; * The following (piece of) code, (part of) the 2xSaI engine, *
-; * copyright (c) 2001 by Derek Liauw Kie Fa. *
-; * Non-Commercial use of the engine is allowed and is encouraged, *
-; * provided that appropriate credit be given and that this copyright *
-; * notice will not be removed under any circumstance. *
+; * copyright (c) 1999 - 2001 by Derek Liauw Kie Fa. *
+; * Non-Commercial use of this software is allowed and is encouraged, *
+; * provided that appropriate credit be given. *
; * You may freely modify this code, but I request *
; * that any improvements to the engine be submitted to me, so *
; * that I can implement these improvements in newer versions of *
-; * the engine. *
+; * the software. *
; * If you need more information, have any comments or suggestions, *
-; * you can e-mail me. My e-mail: DerekL666@yahoo.com *
+; * you can e-mail me. My e-mail: derek-liauw@usa.net. *
; *---------------------------------------------------------------------*/
-; modified by Spacy to compile with yasm [2006-06-20]
;----------------------
-; 2xSaI, Super2xSaI, SuperEagle .. FINAL. no versioning anymore..
+; 2xSaI version 0.59 WIP, soon to become version 0.60
;----------------------
- BITS 32
+;%define FAR_POINTER
+
+
+
+ BITS 32
%ifdef __DJGPP__
GLOBAL __2xSaILine
GLOBAL __2xSaISuperEagleLine
- GLOBAL __2xSaISuper2xSaILine
+ GLOBAL __2xSaISuper2xSaILine
GLOBAL _Init_2xSaIMMX
%else
GLOBAL _2xSaILine
GLOBAL _2xSaISuperEagleLine
- GLOBAL _2xSaISuper2xSaILine
+ GLOBAL _2xSaISuper2xSaILine
GLOBAL Init_2xSaIMMX
%endif
- SECTION .text ALIGN = 32
+ SECTION .text ALIGN = 32
-%ifdef __DJGPP__
-;EXTERN_C void __2xSaILine (uint8 *srcPtr, uint32 srcPitch, uint32 width,
+%ifdef FAR_POINTER
+;EXTERN_C void _2xSaILine (uint8 *srcPtr, uint32 srcPitch, uint32 width,
; uint8 *dstPtr, uint32 dstPitch, uint16 dstSegment);
%else
;EXTERN_C void _2xSaILine (uint8 *srcPtr, uint32 srcPitch, uint32 width,
@@ -48,6 +50,8 @@
dstSegment equ 32
+
+
colorB0 equ -2
colorB1 equ 0
colorB2 equ 2
@@ -75,23 +79,18 @@
-
-
-
-
-
%ifdef __DJGPP__
__2xSaISuper2xSaILine:
%else
-NEWSYM _2xSaISuper2xSaILine
+_2xSaISuper2xSaILine:
%endif
; Store some stuff
- push ebp
- mov ebp, esp
+ push ebp
+ mov ebp, esp
pushad
; Prepare the destination
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
; Set the selector
mov eax, [ebp+dstSegment]
mov fs, ax
@@ -99,11 +98,11 @@
mov edx, [ebp+dstOffset] ; edx points to the screen
; Prepare the source
; eax points to colorA
- mov eax, [ebp+srcPtr] ;eax points to colorA
- mov ebx, [ebp+srcPitch] ;ebx contains the source pitch
- mov ecx, [ebp+width] ;ecx contains the number of pixels to process
+ mov eax, [ebp+srcPtr] ;eax points to colorA
+ mov ebx, [ebp+srcPitch] ;ebx contains the source pitch
+ mov ecx, [ebp+width] ;ecx contains the number of pixels to process
; eax now points to colorB1
- sub eax, ebx ;eax points to B1 which is the base
+ sub eax, ebx ;eax points to B1 which is the base
; Main Loop
.Loop: push ecx
@@ -112,7 +111,7 @@
mov ecx, [ebp+deltaPtr]
- ;load source img
+ ;load source img
movq mm0, [eax+colorB0]
movq mm1, [eax+colorB3]
movq mm2, [eax+ebx+color4]
@@ -125,7 +124,7 @@
movq mm7, [eax+ebx+ebx+colorA3]
pop eax
- ;compare to delta
+ ;compare to delta
pcmpeqw mm0, [ecx+2+colorB0]
pcmpeqw mm1, [ecx+2+colorB3]
pcmpeqw mm2, [ecx+ebx+2+color4]
@@ -138,7 +137,7 @@
sub ecx, ebx
- ;compose results
+ ;compose results
pand mm0, mm1
pand mm2, mm3
pand mm4, mm5
@@ -148,14 +147,14 @@
pxor mm7, mm7
pand mm0, mm4
movq mm6, [eax+colorB0]
- pcmpeqw mm7, mm0 ;did any compare give us a zero ?
+ pcmpeqw mm7, mm0 ;did any compare give us a zero ?
movq [ecx+2+colorB0], mm6
packsswb mm7, mm7
movd ecx, mm7
- test ecx, ecx
- jz near .SKIP_PROCESS ;no, so we can skip
+ test ecx, ecx
+ jz near .SKIP_PROCESS ;no, so we can skip
;End Delta
@@ -433,15 +432,15 @@
pcmpgtw mm0, mm1
por mm7, [Mask35]
- por mm0, [Mask26]
+ por mm0, [Mask26]
movq [Mask35], mm7
movq [Mask26], mm0
.SKIP_GUESS:
- ;Start the ASSEMBLY !!! eh... compose all the results together to form the final image...
+ ;Start the ASSEMBLY !!! eh... compose all the results together to form the final image...
-
+
movq mm0, [eax+ebx+color5]
movq mm1, [eax+ebx+ebx+color2]
movq mm2, mm0
@@ -460,7 +459,7 @@
pand mm3, mm2
paddw mm0, mm3 ;mm0 contains the interpolated values
- ;---------------------------
+ ;---------------------------
@@ -485,102 +484,102 @@
%endif
- movq mm7, [Mask26]
- movq mm6, [eax+colorB2]
- movq mm5, [eax+ebx+ebx+color2]
- movq mm4, [eax+ebx+ebx+color1]
- pcmpeqw mm4, mm5
- pcmpeqw mm6, mm5
- pxor mm5, mm5
- pand mm7, mm4
- pcmpeqw mm6, mm5
- pand mm7, mm6
+ movq mm7, [Mask26]
+ movq mm6, [eax+colorB2]
+ movq mm5, [eax+ebx+ebx+color2]
+ movq mm4, [eax+ebx+ebx+color1]
+ pcmpeqw mm4, mm5
+ pcmpeqw mm6, mm5
+ pxor mm5, mm5
+ pand mm7, mm4
+ pcmpeqw mm6, mm5
+ pand mm7, mm6
- movq mm6, [eax+ebx+ebx+color3]
- movq mm5, [eax+ebx+ebx+color2]
- movq mm4, [eax+ebx+ebx+color1]
- movq mm2, [eax+ebx+color5]
- movq mm1, [eax+ebx+color4]
- movq mm3, [eax+colorB0]
+ movq mm6, [eax+ebx+ebx+color3]
+ movq mm5, [eax+ebx+ebx+color2]
+ movq mm4, [eax+ebx+ebx+color1]
+ movq mm2, [eax+ebx+color5]
+ movq mm1, [eax+ebx+color4]
+ movq mm3, [eax+colorB0]
- pcmpeqw mm2, mm4
- pcmpeqw mm6, mm5
- pcmpeqw mm1, mm5
- pcmpeqw mm3, mm5
- pxor mm5, mm5
- pcmpeqw mm2, mm5
- pcmpeqw mm3, mm5
- pand mm6, mm1
- pand mm2, mm3
- pand mm6, mm2
- por mm7, mm6
+ pcmpeqw mm2, mm4
+ pcmpeqw mm6, mm5
+ pcmpeqw mm1, mm5
+ pcmpeqw mm3, mm5
+ pxor mm5, mm5
+ pcmpeqw mm2, mm5
+ pcmpeqw mm3, mm5
+ pand mm6, mm1
+ pand mm2, mm3
+ pand mm6, mm2
+ por mm7, mm6
-
- movq mm6, mm7
- pcmpeqw mm6, mm5
- pand mm7, mm0
+
+ movq mm6, mm7
+ pcmpeqw mm6, mm5
+ pand mm7, mm0
- movq mm1, [eax+ebx+color5]
- pand mm6, mm1
- por mm7, mm6
- movq [final1a], mm7 ;finished 1a
+ movq mm1, [eax+ebx+color5]
+ pand mm6, mm1
+ por mm7, mm6
+ movq [final1a], mm7 ;finished 1a
-
- ;--------------------------------
+
+ ;--------------------------------
- movq mm7, [Mask35]
- push eax
- add eax, ebx
- movq mm6, [eax+ebx+ebx+colorA2]
- pop eax
- movq mm5, [eax+ebx+color5]
- movq mm4, [eax+ebx+color4]
- pcmpeqw mm4, mm5
- pcmpeqw mm6, mm5
- pxor mm5, mm5
- pand mm7, mm4
- pcmpeqw mm6, mm5
- pand mm7, mm6
+ movq mm7, [Mask35]
+ push eax
+ add eax, ebx
+ movq mm6, [eax+ebx+ebx+colorA2]
+ pop eax
+ movq mm5, [eax+ebx+color5]
+ movq mm4, [eax+ebx+color4]
+ pcmpeqw mm4, mm5
+ pcmpeqw mm6, mm5
+ pxor mm5, mm5
+ pand mm7, mm4
+ pcmpeqw mm6, mm5
+ pand mm7, mm6
- movq mm6, [eax+ebx+color6]
- movq mm5, [eax+ebx+color5]
- movq mm4, [eax+ebx+color4]
- movq mm2, [eax+ebx+ebx+color2]
- movq mm1, [eax+ebx+ebx+color1]
- push eax
- add eax, ebx
- movq mm3, [eax+ebx+ebx+colorA0]
- pop eax
+ movq mm6, [eax+ebx+color6]
+ movq mm5, [eax+ebx+color5]
+ movq mm4, [eax+ebx+color4]
+ movq mm2, [eax+ebx+ebx+color2]
+ movq mm1, [eax+ebx+ebx+color1]
+ push eax
+ add eax, ebx
+ movq mm3, [eax+ebx+ebx+colorA0]
+ pop eax
- pcmpeqw mm2, mm4
- pcmpeqw mm6, mm5
- pcmpeqw mm1, mm5
- pcmpeqw mm3, mm5
- pxor mm5, mm5
- pcmpeqw mm2, mm5
- pcmpeqw mm3, mm5
- pand mm6, mm1
- pand mm2, mm3
- pand mm6, mm2
- por mm7, mm6
+ pcmpeqw mm2, mm4
+ pcmpeqw mm6, mm5
+ pcmpeqw mm1, mm5
+ pcmpeqw mm3, mm5
+ pxor mm5, mm5
+ pcmpeqw mm2, mm5
+ pcmpeqw mm3, mm5
+ pand mm6, mm1
+ pand mm2, mm3
+ pand mm6, mm2
+ por mm7, mm6
-
- movq mm6, mm7
- pcmpeqw mm6, mm5
- pand mm7, mm0
+
+ movq mm6, mm7
+ pcmpeqw mm6, mm5
+ pand mm7, mm0
- movq mm1, [eax+ebx+ebx+color2]
- pand mm6, mm1
- por mm7, mm6
- movq [final2a], mm7 ;finished 2a
+ movq mm1, [eax+ebx+ebx+color2]
+ pand mm6, mm1
+ por mm7, mm6
+ movq [final2a], mm7 ;finished 2a
- ;--------------------------------------------
+ ;--------------------------------------------
%ifdef dfhsdfhsdahdsfhdsfh
@@ -601,133 +600,168 @@
product1b = INTERPOLATE (color5, color6);
%endif
- push eax
- add eax, ebx
- pxor mm7, mm7
- movq mm0, [eax+ebx+ebx+colorA0]
- movq mm1, [eax+ebx+ebx+colorA1]
- movq mm2, [eax+ebx+ebx+colorA2]
- movq mm3, [eax+ebx+ebx+colorA3]
- pop eax
- movq mm4, [eax+ebx+ebx+color2]
- movq mm5, [eax+ebx+ebx+color3]
- movq mm6, [eax+ebx+color6]
+ push eax
+ add eax, ebx
+ pxor mm7, mm7
+ movq mm0, [eax+ebx+ebx+colorA0]
+ movq mm1, [eax+ebx+ebx+colorA1]
+ movq mm2, [eax+ebx+ebx+colorA2]
+ movq mm3, [eax+ebx+ebx+colorA3]
+ pop eax
+ movq mm4, [eax+ebx+ebx+color2]
+ movq mm5, [eax+ebx+ebx+color3]
+ movq mm6, [eax+ebx+color6]
- pcmpeqw mm6, mm5
- pcmpeqw mm1, mm5
- pcmpeqw mm4, mm2
- pcmpeqw mm0, mm5
- pcmpeqw mm4, mm7
- pcmpeqw mm0, mm7
- pand mm0, mm4
- pand mm6, mm1
- pand mm0, mm6
+ pcmpeqw mm6, mm5
+ pcmpeqw mm1, mm5
+ pcmpeqw mm4, mm2
+ pcmpeqw mm0, mm5
+ pcmpeqw mm4, mm7
+ pcmpeqw mm0, mm7
+ pand mm0, mm4
+ pand mm6, mm1
+ pand mm0, mm6
- movq mm4, [eax+ebx+color2]
- movq mm5, [eax+ebx+ebx+color5]
- movq mm6, [eax+ebx+ebx+color3]
- pcmpeqw mm5, mm4
- pcmpeqw mm2, mm4
- pcmpeqw mm1, mm6
- pcmpeqw mm3, mm4
- pcmpeqw mm1, mm7
- pcmpeqw mm3, mm7
- pand mm2, mm5
- pand mm1, mm3
- pand mm1, mm2
+ push eax
+ add eax, ebx
+ movq mm1, [eax+ebx+ebx+colorA1]
+ pop eax
+ movq mm4, [eax+ebx+ebx+color2]
+ movq mm5, [eax+ebx+color5]
+ movq mm6, [eax+ebx+ebx+color3]
- movq mm2, mm0
- movq mm7, [I2333Pixel]
- movq mm6, [I2223Pixel]
- movq mm5, [I23Pixel]
- movq mm4, [Mask35]
- movq mm3, [Mask26]
+ pcmpeqw mm5, mm4
+ pcmpeqw mm2, mm4
+ pcmpeqw mm1, mm6
+ pcmpeqw mm3, mm4
+ pcmpeqw mm1, mm7
+ pcmpeqw mm3, mm7
+ pand mm2, mm5
+ pand mm1, mm3
+ pand mm1, mm2
- por mm2, mm4
- pand mm4, [eax+ebx+ebx+color3]
- por mm2, mm3
- pand mm3, [eax+ebx+ebx+color2]
- por mm2, mm1
- pand mm0, mm7
- pand mm1, mm6
- pxor mm7, mm7
- pcmpeqw mm2, mm7
- por mm0, mm1
- por mm3, mm4
- pand mm2, mm5
- por mm0, mm3
- por mm0, mm2
- movq [final2b], mm0
- ;-----------------------------------
-
+ movq mm7, mm0
+ por mm7, mm1
- pxor mm7, mm7
- movq mm0, [eax+colorB0]
- movq mm1, [eax+colorB1]
- movq mm2, [eax+colorB2]
- movq mm3, [eax+colorB3]
- movq mm4, [eax+ebx+color5]
- movq mm5, [eax+ebx+color6]
- movq mm6, [eax+ebx+ebx+color3]
+ movq mm4, [Mask35]
+ movq mm3, [Mask26]
+
+ movq mm6, mm4
+ pand mm6, mm7
+ pxor mm4, mm6
- pcmpeqw mm6, mm5
- pcmpeqw mm1, mm5
- pcmpeqw mm4, mm2
- pcmpeqw mm0, mm5
- pcmpeqw mm4, mm7
- pcmpeqw mm0, mm7
- pand mm0, mm4
- pand mm6, mm1
- pand mm0, mm6
+ movq mm6, mm3
+ pand mm6, mm7
+ pxor mm3, mm6
- movq mm4, [eax+ebx+color5]
- movq mm5, [eax+ebx+ebx+color2]
- movq mm6, [eax+ebx+color6]
+ movq mm2, mm0
+ movq mm7, [I2333Pixel]
+ movq mm6, [I2223Pixel]
+ movq mm5, [I23Pixel]
- pcmpeqw mm5, mm4
- pcmpeqw mm2, mm4
- pcmpeqw mm1, mm6
- pcmpeqw mm3, mm4
- pcmpeqw mm1, mm7
- pcmpeqw mm3, mm7
- pand mm2, mm5
- pand mm1, mm3
- pand mm1, mm2
- movq mm2, mm0
- movq mm7, [I5666Pixel]
- movq mm6, [I5556Pixel]
- movq mm5, [I56Pixel]
- movq mm4, [Mask35]
- movq mm3, [Mask26]
+ por mm2, mm4
+ pand mm4, [eax+ebx+ebx+color3]
+ por mm2, mm3
+ pand mm3, [eax+ebx+ebx+color2]
+ por mm2, mm1
+ pand mm0, mm7
+ pand mm1, mm6
+ pxor mm7, mm7
+ pcmpeqw mm2, mm7
+ por mm0, mm1
+ por mm3, mm4
+ pand mm2, mm5
+ por mm0, mm3
+ por mm0, mm2
+ movq [final2b], mm0
- por mm2, mm4
- pand mm4, [eax+ebx+color5]
- por mm2, mm3
- pand mm3, [eax+ebx+color6]
- por mm2, mm1
- pand mm0, mm7
- pand mm1, mm6
- pxor mm7, mm7
- pcmpeqw mm2, mm7
- por mm0, mm1
- por mm3, mm4
- pand mm2, mm5
- por mm0, mm3
- por mm0, mm2
- movq [final1b], mm0
-
- ;---------
+ ;-----------------------------------
+
- movq mm0, [final1a]
- movq mm4, [final2a]
- movq mm2, [final1b]
- movq mm6, [final2b]
+ pxor mm7, mm7
+ movq mm0, [eax+colorB0]
+ movq mm1, [eax+colorB1]
+ movq mm2, [eax+colorB2]
+ movq mm3, [eax+colorB3]
+ movq mm4, [eax+ebx+color5]
+ movq mm5, [eax+ebx+color6]
+ movq mm6, [eax+ebx+ebx+color3]
- movq mm1, mm0
- movq mm5, mm4
+ pcmpeqw mm6, mm5
+ pcmpeqw mm1, mm5
+ pcmpeqw mm4, mm2
+ pcmpeqw mm0, mm5
+ pcmpeqw mm4, mm7
+ pcmpeqw mm0, mm7
+ pand mm0, mm4
+ pand mm6, mm1
+ pand mm0, mm6
+
+ movq mm1, [eax+colorB1]
+ movq mm4, [eax+ebx+color5]
+ movq mm5, [eax+ebx+ebx+color2]
+ movq mm6, [eax+ebx+color6]
+
+ pcmpeqw mm5, mm4
+ pcmpeqw mm2, mm4
+ pcmpeqw mm1, mm6
+ pcmpeqw mm3, mm4
+ pcmpeqw mm1, mm7
+ pcmpeqw mm3, mm7
+ pand mm2, mm5
+ pand mm1, mm3
+ pand mm1, mm2
+
+
+ movq mm7, mm0
+ por mm7, mm1
+
+ movq mm4, [Mask35]
+ movq mm3, [Mask26]
+
+ movq mm6, mm4
+ pand mm6, mm7
+ pxor mm4, mm6
+
+ movq mm6, mm3
+ pand mm6, mm7
+ pxor mm3, mm6
+
+ movq mm2, mm0
+ movq mm7, [I5666Pixel]
+ movq mm6, [I5556Pixel]
+ movq mm5, [I56Pixel]
+
+
+ por mm2, mm4
+ pand mm4, [eax+ebx+color5]
+ por mm2, mm3
+ pand mm3, [eax+ebx+color6]
+ por mm2, mm1
+ pand mm0, mm7
+ pand mm1, mm6
+ pxor mm7, mm7
+ pcmpeqw mm2, mm7
+ por mm0, mm1
+ por mm3, mm4
+ pand mm2, mm5
+ por mm0, mm3
+ por mm0, mm2
+ movq [final1b], mm0
+
+ ;---------
+
+ movq mm0, [final1a]
+ movq mm4, [final2a]
+ movq mm2, [final1b]
+ movq mm6, [final2b]
+
+
+ movq mm1, mm0
+ movq mm5, mm4
punpcklwd mm0, mm2
@@ -746,12 +780,12 @@
movq [fs:edx+8], mm5
pop edx
%else
- movq [es:edx], mm0
- movq [es:edx+8], mm1
+ movq [edx], mm0
+ movq [edx+8], mm1
push edx
add edx, [ebp+dstPitch]
- movq [es:edx], mm4
- movq [es:edx+8], mm5
+ movq [edx], mm4
+ movq [edx+8], mm5
pop edx
%endif
.SKIP_PROCESS:
@@ -784,22 +818,18 @@
-
-
-
-
%ifdef __DJGPP__
__2xSaISuperEagleLine:
%else
-NEWSYM _2xSaISuperEagleLine
+_2xSaISuperEagleLine:
%endif
; Store some stuff
- push ebp
- mov ebp, esp
+ push ebp
+ mov ebp, esp
pushad
; Prepare the destination
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
; Set the selector
mov eax, [ebp+dstSegment]
mov fs, ax
@@ -1187,9 +1217,9 @@
pcmpgtw mm0, mm1
por mm7, [Mask35]
- por mm1, [Mask26]
+ por mm0, [Mask26]
movq [Mask35], mm7
- movq [Mask26], mm1
+ movq [Mask26], mm0
.SKIP_GUESS:
;Start the ASSEMBLY !!!
@@ -1301,7 +1331,7 @@
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
movq [fs:edx], mm0
movq [fs:edx+8], mm1
push edx
@@ -1310,12 +1340,12 @@
movq [fs:edx+8], mm5
pop edx
%else
- movq [es:edx], mm0
- movq [es:edx+8], mm1
+ movq [edx], mm0
+ movq [edx+8], mm1
push edx
add edx, [ebp+dstPitch]
- movq [es:edx], mm4
- movq [es:edx+8], mm5
+ movq [edx], mm4
+ movq [edx+8], mm5
pop edx
%endif
.SKIP_PROCESS:
@@ -1347,7 +1377,7 @@
;-------------------------------------------------------------------------
-;This is 2xSaI
+;This is version 0.50
colorI equ -2
colorE equ 0
colorF equ 2
@@ -1371,15 +1401,15 @@
%ifdef __DJGPP__
__2xSaILine:
%else
-NEWSYM _2xSaILine
+_2xSaILine:
%endif
; Store some stuff
- push ebp
- mov ebp, esp
+ push ebp
+ mov ebp, esp
pushad
; Prepare the destination
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
; Set the selector
mov eax, [ebp+dstSegment]
mov fs, ax
@@ -1569,12 +1599,12 @@
punpcklwd mm5, mm0
punpckhwd mm6, mm0
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
movq [fs:edx], mm5
movq [fs:edx+8], mm6
%else
- movq [es:edx], mm5
- movq [es:edx+8], mm6
+ movq [edx], mm5
+ movq [edx+8], mm6
%endif
;------------------------------------------------
@@ -1736,6 +1766,7 @@
movd ecx, mm7
test ecx, ecx
jz near .SKIP_GUESS
+
;---------------------------------------------
; Map of the pixels: I|E F|J
; G|A B|K
@@ -1856,9 +1887,9 @@
pcmpgtw mm0, mm1
por mm7, [Mask1]
- por mm1, [Mask2]
+ por mm0, [Mask2]
movq [Mask1], mm7
- movq [Mask2], mm1
+ movq [Mask2], mm0
.SKIP_GUESS:
;----------------------------
@@ -1922,12 +1953,12 @@
push edx
add edx, [ebp+dstPitch]
-%ifdef __DJGPP__
+%ifdef FAR_POINTER
movq [fs:edx], mm0
movq [fs:edx+8], mm1
%else
- movq [es:edx], mm0
- movq [es:edx+8], mm1
+ movq [edx], mm0
+ movq [edx+8], mm1
%endif
pop edx
@@ -1961,28 +1992,28 @@
%ifdef __DJGPP__
_Init_2xSaIMMX:
%else
-NEWSYM Init_2xSaIMMX
+Init_2xSaIMMX:
%endif
; Store some stuff
- push ebp
- mov ebp, esp
+ push ebp
+ mov ebp, esp
push edx
;Damn thing doesn't work
-; mov eax,1
-; cpuid
-; test edx, 0x00800000 ;test bit 23
-; jz end2 ;bit not set => no MMX detected
+; mov eax,1
+; cpuid
+; test edx, 0x00800000 ;test bit 23
+; jz end2 ;bit not set => no MMX detected
- mov eax, [ebp+8] ;PixelFormat
- cmp eax, 555
- jz Bits555
- cmp eax, 565
- jz Bits565
+ mov eax, [ebp+8] ;PixelFormat
+ cmp eax, 555
+ jz Bits555
+ cmp eax, 565
+ jz Bits565
end2:
- mov eax, 1
- jmp end
+ mov eax, 1
+ jmp end3
Bits555:
mov edx, 0x7BDE7BDE
mov eax, colorMask
@@ -2001,7 +2032,7 @@
mov [eax], edx
mov [eax+4], edx
mov eax, 0
- jmp end
+ jmp end3
Bits565:
mov edx, 0xF7DEF7DE
mov eax, colorMask
@@ -2020,12 +2051,12 @@
mov [eax], edx
mov [eax+4], edx
mov eax, 0
- jmp end
-end:
+ jmp end3
+end3:
pop edx
- mov esp, ebp
- pop ebp
- ret
+ mov esp, ebp
+ pop ebp
+ ret
;-------------------------------------------------------------------------
@@ -2036,17 +2067,21 @@
;-------------------------------------------------------------------------
;-------------------------------------------------------------------------
- SECTION .data ALIGN = 32
+ SECTION .data ALIGN = 32
;Some constants
-colorMask dd 0xF7DEF7DE, 0xF7DEF7DE
-lowPixelMask dd 0x08210821, 0x08210821
+colorMask dd 0xF7DEF7DE,0xF7DEF7DE
+lowPixelMask dd 0x08210821,0x08210821
-qcolorMask dd 0xE79CE79C, 0xE79CE79C
-qlowpixelMask dd 0x18631863, 0x18631863
+qcolorMask dd 0xE79CE79C,0xE79CE79C
+qlowpixelMask dd 0x18631863,0x18631863
-FALSE dd 0x00000000, 0x00000000
-TRUE dd 0xffffffff, 0xffffffff
-ONE dd 0x00010001, 0x00010001
+darkenMask dd 0xC718C718,0xC718C718
+GreenMask dd 0x07E007E0,0x07E007E0
+RedBlueMask dd 0xF81FF81F,0xF81FF81F
+
+FALSE dd 0x00000000,0x00000000
+TRUE dd 0xffffffff,0xffffffff
+ONE dd 0x00010001,0x00010001
SECTION .bss ALIGN = 32
