Subversion Repositories Kolibri OS

Rev

Rev 6619 | Go to most recent revision | Only display areas with differences | Regard whitespace | Details | Blame | Last modification | View Log | RSS feed

Rev 6619 Rev 6769
1 
;SIZE_X equ 350
 1 
;SIZE_X equ 350
2 
;SIZE_Y equ 350
 2 
;SIZE_Y equ 350
3 
;ROUND equ 8
 3 
;ROUND equ 8
4 
;TEX_X equ 512
 4 
;TEX_X equ 512
5 
;TEX_Y equ 512
 5 
;TEX_Y equ 512
6 
;TEXTURE_SIZE EQU (512*512)-1
 6 
;TEXTURE_SIZE EQU (512*512)-1
7 
;TEX_SHIFT EQU 9
 7 
;TEX_SHIFT EQU 9
8 
 
 8 
 
9 
;CATMULL_SHIFT equ 8
 9 
;CATMULL_SHIFT equ 8
10 
;TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
 10 
;TEXTURE_SIZE EQU (TEX_X * TEX_Y)-1
11 
;Ext = SSE
 11 
;Ext = SSE
12 
;SSE = 3
 12 
;SSE = 3
13 
;MMX = 1
 13 
;MMX = 1
14 
;NON = 0
 14 
;NON = 0
15 
;use32
 15 
;use32
16 
;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
 16 
;------- Big thanks to Majuma (www.majuma.xt.pl) for absolutely great---
17 
;------- DOS 13h mode demos --------------------------------------------
 17 
;------- DOS 13h mode demos --------------------------------------------
18 
;------- Procedure draws triangle with two overlapped textures, I use --
 18 
;------- Procedure draws triangle with two overlapped textures, I use --
19 
;--------Catmull Z-buffer algorithm- (Z coordinate interpolation)-------
 19 
;--------Catmull Z-buffer algorithm- (Z coordinate interpolation)-------
20 
;--------I calc texture pixel by this way: col1*col2/256 ---------------
 20 
;--------I calc texture pixel by this way: col1*col2/256 ---------------
21 
two_tex_triangle_z:
 21 
two_tex_triangle_z:
22 
;------------------in - eax - x1 shl 16 + y1 -----------
 22 
;------------------in - eax - x1 shl 16 + y1 -----------
23 
;---------------------- ebx - x2 shl 16 + y2 -----------
 23 
;---------------------- ebx - x2 shl 16 + y2 -----------
24 
;---------------------- ecx - x3 shl 16 + y3 -----------
 24 
;---------------------- ecx - x3 shl 16 + y3 -----------
25 
;---------------------- edx - pointer to b. texture-----
 25 
;---------------------- edx - pointer to b. texture-----
26 
;---------------------- esi - pointer to e. texture-----
 26 
;---------------------- esi - pointer to e. texture-----
27 
;---------------------- edi - pointer to screen buffer--
 27 
;---------------------- edi - pointer to screen buffer--
28 
;---------------------- stack : b. tex coordinates------
 28 
;---------------------- stack : b. tex coordinates------
29 
;----------------------         e. tex coordinates------
 29 
;----------------------         e. tex coordinates------
30 
;----------------------         Z position coordinates--
 30 
;----------------------         Z position coordinates--
31 
;----------------------         pointer io Z buffer-----
 31 
;----------------------         pointer io Z buffer-----
32 
;-- Z-buffer - filled with coordinates as dword --------
 32 
;-- Z-buffer - filled with coordinates as dword --------
33 
;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
 33 
;-- (Z coor. as word) shl CATMULL_SHIFT ----------------
34 
.b_x1	equ ebp+4   ; procedure don't save registers !!!
 34 
.b_x1	equ ebp+4   ; procedure don't save registers !!!
35 
.b_y1	equ ebp+6   ; each coordinate as word
 35 
.b_y1	equ ebp+6   ; each coordinate as word
36 
.b_x2	equ ebp+8
 36 
.b_x2	equ ebp+8
37 
.b_y2	equ ebp+10	 ; b - first texture
 37 
.b_y2	equ ebp+10	 ; b - first texture
38 
.b_x3	equ ebp+12
 38 
.b_x3	equ ebp+12
39 
.b_y3	equ ebp+14	 ; e - second texture
 39 
.b_y3	equ ebp+14	 ; e - second texture
40 
.e_x1	equ ebp+16
 40 
.e_x1	equ ebp+16
41 
.e_y1	equ ebp+18
 41 
.e_y1	equ ebp+18
42 
.e_x2	equ ebp+20
 42 
.e_x2	equ ebp+20
43 
.e_y2	equ ebp+22
 43 
.e_y2	equ ebp+22
44 
.e_x3	equ ebp+24
 44 
.e_x3	equ ebp+24
45 
.e_y3	equ ebp+26
 45 
.e_y3	equ ebp+26
46 
.z1	equ word[ebp+28]
 46 
.z1	equ word[ebp+28]
47 
.z2	equ word[ebp+30]
 47 
.z2	equ word[ebp+30]
48 
.z3	equ word[ebp+32]
 48 
.z3	equ word[ebp+32]
49 
.z_buff equ dword[ebp+34]	; pointer to Z-buffer
 49 
.z_buff equ dword[ebp+34]	; pointer to Z-buffer
50 
 
 50 
 
51 
 
 51 
 
52 
.t_bmap equ dword[ebp-4]	; pointer to b. texture
 52 
.t_bmap equ dword[ebp-4]	; pointer to b. texture
53 
.t_emap equ dword[ebp-8]	; pointer to e. texture
 53 
.t_emap equ dword[ebp-8]	; pointer to e. texture
54 
.x1	equ word[ebp-10]
 54 
.x1	equ word[ebp-10]
55 
.y1	equ word[ebp-12]
 55 
.y1	equ word[ebp-12]
56 
.x2	equ word[ebp-14]
 56 
.x2	equ word[ebp-14]
57 
.y2	equ word[ebp-16]
 57 
.y2	equ word[ebp-16]
58 
.x3	equ word[ebp-18]
 58 
.x3	equ word[ebp-18]
59 
.y3	equ word[ebp-20]
 59 
.y3	equ word[ebp-20]
60 
 
 60 
 
61 
.dx12  equ dword[ebp-24]
 61 
.dx12  equ dword[ebp-24]
62 
.dbx12 equ dword[ebp-28]
 62 
.dbx12 equ dword[ebp-28]
63 
.dby12 equ dword[ebp-32]
 63 
.dby12 equ dword[ebp-32]
64 
.dby12q equ	[ebp-32]
 64 
.dby12q equ	[ebp-32]
65 
.dex12 equ dword[ebp-36]
 65 
.dex12 equ dword[ebp-36]
66 
.dey12 equ dword[ebp-40]
 66 
.dey12 equ dword[ebp-40]
67 
.dey12q equ	[ebp-40]
 67 
.dey12q equ	[ebp-40]
68 
.dz12  equ dword[ebp-44]
 68 
.dz12  equ dword[ebp-44]
69 
 
 69 
 
70 
.dx13  equ dword[ebp-48]
 70 
.dx13  equ dword[ebp-48]
71 
.dbx13 equ dword[ebp-52]
 71 
.dbx13 equ dword[ebp-52]
72 
.dby13 equ dword[ebp-56]
 72 
.dby13 equ dword[ebp-56]
73 
.dby13q equ	[ebp-56]
 73 
.dby13q equ	[ebp-56]
74 
.dex13 equ dword[ebp-60]
 74 
.dex13 equ dword[ebp-60]
75 
.dey13 equ dword[ebp-64]
 75 
.dey13 equ dword[ebp-64]
76 
.dey13q equ	[ebp-64]
 76 
.dey13q equ	[ebp-64]
77 
.dz13  equ dword[ebp-68]
 77 
.dz13  equ dword[ebp-68]
78 
 
 78 
 
79 
.dx23  equ dword[ebp-72]
 79 
.dx23  equ dword[ebp-72]
80 
.dbx23 equ dword[ebp-76]
 80 
.dbx23 equ dword[ebp-76]
81 
.dby23 equ dword[ebp-80]
 81 
.dby23 equ dword[ebp-80]
82 
.dby23q equ	[ebp-80]
 82 
.dby23q equ	[ebp-80]
83 
.dex23 equ dword[ebp-84]
 83 
.dex23 equ dword[ebp-84]
84 
.dey23 equ dword[ebp-88]
 84 
.dey23 equ dword[ebp-88]
85 
.dey23q equ	[ebp-88]
 85 
.dey23q equ	[ebp-88]
86 
.dz23  equ dword[ebp-92]
 86 
.dz23  equ dword[ebp-92]
87 
 
 87 
 
88 
.cx1   equ dword[ebp-96]   ; current variables
 88 
.cx1   equ dword[ebp-96]   ; current variables
89 
.cx2   equ dword[ebp-100]
 89 
.cx2   equ dword[ebp-100]
90 
.cbx1  equ dword[ebp-104]
 90 
.cbx1  equ dword[ebp-104]
91 
.cby1  equ [ebp-108]
 91 
.cby1  equ [ebp-108]
92 
.cex1  equ dword[ebp-112]
 92 
.cex1  equ dword[ebp-112]
93 
.cey1  equ [ebp-116]
 93 
.cey1  equ [ebp-116]
94 
.cbx2  equ dword[ebp-120]
 94 
.cbx2  equ dword[ebp-120]
95 
.cby2  equ [ebp-124]
 95 
.cby2  equ [ebp-124]
96 
.cex2  equ dword[ebp-128]
 96 
.cex2  equ dword[ebp-128]
97 
.cey2  equ [ebp-132]
 97 
.cey2  equ [ebp-132]
98 
 
 98 
 
99 
.cz1   equ dword[ebp-136]
 99 
.cz1   equ dword[ebp-136]
100 
.cz2   equ dword[ebp-140]
 100 
.cz2   equ dword[ebp-140]
101 
 
 101 
 
102 
    if Ext >= MMX
 102 
    if Ext >= MMX
103 
       emms
 103 
       emms
104 
    else
 104 
    else
105 
       cld
 105 
       cld
106 
    end if
 106 
    end if
107 
       mov     ebp,esp
 107 
       mov     ebp,esp
108 
       push    edx esi	     ; store bump map
 108 
       push    edx esi	     ; store bump map
109 
;       push    esi        ; store e. map
 109 
;       push    esi        ; store e. map
110 
     ; sub     esp,120
 110 
     ; sub     esp,120
111 
 .sort3:		  ; sort triangle coordinates...
 111 
 .sort3:		  ; sort triangle coordinates...
112 
       cmp     ax,bx
 112 
       cmp     ax,bx
113 
       jle     .sort1
 113 
       jle     .sort1
114 
       xchg    eax,ebx
 114 
       xchg    eax,ebx
115 
       mov     edx,dword[.b_x1]
 115 
       mov     edx,dword[.b_x1]
116 
       xchg    edx,dword[.b_x2]
 116 
       xchg    edx,dword[.b_x2]
117 
       mov     dword[.b_x1],edx
 117 
       mov     dword[.b_x1],edx
118 
       mov     edx,dword[.e_x1]
 118 
       mov     edx,dword[.e_x1]
119 
       xchg    edx,dword[.e_x2]
 119 
       xchg    edx,dword[.e_x2]
120 
       mov     dword[.e_x1],edx
 120 
       mov     dword[.e_x1],edx
121 
       mov     dx,.z1
 121 
       mov     dx,.z1
122 
       xchg    dx,.z2
 122 
       xchg    dx,.z2
123 
       mov     .z1,dx
 123 
       mov     .z1,dx
124 
 .sort1:
 124 
 .sort1:
125 
       cmp	bx,cx
 125 
       cmp	bx,cx
126 
       jle	.sort2
 126 
       jle	.sort2
127 
       xchg	ebx,ecx
 127 
       xchg	ebx,ecx
128 
       mov	edx,dword[.b_x2]
 128 
       mov	edx,dword[.b_x2]
129 
       xchg	edx,dword[.b_x3]
 129 
       xchg	edx,dword[.b_x3]
130 
       mov	dword[.b_x2],edx
 130 
       mov	dword[.b_x2],edx
131 
       mov	edx,dword[.e_x2]
 131 
       mov	edx,dword[.e_x2]
132 
       xchg	edx,dword[.e_x3]
 132 
       xchg	edx,dword[.e_x3]
133 
       mov	dword[.e_x2],edx
 133 
       mov	dword[.e_x2],edx
134 
       mov     dx,.z2
 134 
       mov     dx,.z2
135 
       xchg    dx,.z3
 135 
       xchg    dx,.z3
136 
       mov     .z2,dx
 136 
       mov     .z2,dx
137 
       jmp	.sort3
 137 
       jmp	.sort3
138 
 .sort2:
 138 
 .sort2:
139 
       push	eax ebx ecx    ; store triangle coords in variables
 139 
       push	eax ebx ecx    ; store triangle coords in variables
140 
;       push     ebx
 140 
;       push     ebx
141 
;       push     ecx
 141 
;       push     ecx
142 
 
 142 
 
143 
	 mov	  edx,80008000h  ; eax,ebx,ecx are ANDd together into edx which means that
 143 
	 mov	  edx,80008000h  ; eax,ebx,ecx are ANDd together into edx which means that
144 
	 and	  edx,ebx	 ; if *all* of them are negative a sign flag is raised
 144 
	 and	  edx,ebx	 ; if *all* of them are negative a sign flag is raised
145 
	 and	  edx,ecx
 145 
	 and	  edx,ecx
146 
	 and	  edx,eax
 146 
	 and	  edx,eax
147 
	 test	  edx,80008000h  ; Check both X&Y at once
 147 
	 test	  edx,80008000h  ; Check both X&Y at once
148 
	 jne	  .loop23_done
 148 
	 jne	  .loop23_done
149 
    ;   mov     edx,eax         ; eax,ebx,ecx are ORd together into edx which means that
 149 
    ;   mov     edx,eax         ; eax,ebx,ecx are ORd together into edx which means that
150 
    ;   or      edx,ebx         ; if any *one* of them is negative a sign flag is raised
 150 
    ;   or      edx,ebx         ; if any *one* of them is negative a sign flag is raised
151 
    ;   or      edx,ecx
 151 
    ;   or      edx,ecx
152 
    ;   test    edx,80000000h   ; Check only X
 152 
    ;   test    edx,80000000h   ; Check only X
153 
    ;   jne     .loop23_done
 153 
    ;   jne     .loop23_done
154 
 
 154 
 
155 
    ;   cmp     .x1,SIZE_X    ; {
 155 
    ;   cmp     .x1,SIZE_X    ; {
156 
    ;   jg      .loop23_done
 156 
    ;   jg      .loop23_done
157 
    ;   cmp     .x2,SIZE_X     ; This can be optimized with effort
 157 
    ;   cmp     .x2,SIZE_X     ; This can be optimized with effort
158 
    ;   jg      .loop23_done
 158 
    ;   jg      .loop23_done
159 
    ;   cmp     .x3,SIZE_X
 159 
    ;   cmp     .x3,SIZE_X
160 
    ;   jg      .loop23_done    ; {
 160 
    ;   jg      .loop23_done    ; {
161 
 
 161 
 
162 
 
 162 
 
163 
       mov	bx,.y2	     ; calc delta 12
 163 
       mov	bx,.y2	     ; calc delta 12
164 
       sub	bx,.y1
 164 
       sub	bx,.y1
165 
       jnz	.bt_dx12_make
 165 
       jnz	.bt_dx12_make
166 
       mov	ecx,6
 166 
       mov	ecx,6
167 
       xor	edx,edx
 167 
       xor	edx,edx
168 
     @@:
 168 
     @@:
169 
       push	edx   ;dword 0
 169 
       push	edx   ;dword 0
170 
       loop	@b
 170 
       loop	@b
171 
       jmp	.bt_dx12_done
 171 
       jmp	.bt_dx12_done
172 
 .bt_dx12_make:
 172 
 .bt_dx12_make:
173 
       mov	ax,.x2
 173 
       mov	ax,.x2
174 
       sub	ax,.x1
 174 
       sub	ax,.x1
175 
       cwde
 175 
       cwde
176 
       movsx	ebx,bx
 176 
       movsx	ebx,bx
177 
       shl	eax,ROUND
 177 
       shl	eax,ROUND
178 
       cdq
 178 
       cdq
179 
       idiv	ebx
 179 
       idiv	ebx
180 
;      mov      .dx12,eax
 180 
;      mov      .dx12,eax
181 
       push	 eax
 181 
       push	 eax
182 
 
 182 
 
183 
if Ext=SSE
 183 
if Ext=SSE
184 
 
 184 
 
185 
       sub	 esp,16
 185 
       sub	 esp,16
186 
       cvtsi2ss  xmm3,ebx	     ;rcps
 186 
       cvtsi2ss  xmm3,ebx	     ;rcps
187 
    ;   mov       eax,255
 187 
    ;   mov       eax,255
188 
       cvtsi2ss  xmm4,[i255d] ;eax
 188 
       cvtsi2ss  xmm4,[i255d] ;eax
189 
       divss	 xmm3,xmm4
 189 
       divss	 xmm3,xmm4
190 
       rcpss	 xmm3,xmm3
 190 
       rcpss	 xmm3,xmm3
191 
    ;   mulss     xmm3,xmm4
 191 
    ;   mulss     xmm3,xmm4
192 
       shufps	 xmm3,xmm3,0
 192 
       shufps	 xmm3,xmm3,0
193 
 
 193 
 
194 
       movd	 mm0,[.b_x1]
 194 
       movd	 mm0,[.b_x1]
195 
       movd	 mm1,[.b_x2]
 195 
       movd	 mm1,[.b_x2]
196 
       movd	 mm2,[.e_x1]
 196 
       movd	 mm2,[.e_x1]
197 
       movd	 mm3,[.e_x2]
 197 
       movd	 mm3,[.e_x2]
198 
     ;  psubsw    mm3,mm2
 198 
     ;  psubsw    mm3,mm2
199 
     ;  psubsw    mm1,mm0
 199 
     ;  psubsw    mm1,mm0
200 
       pxor	 mm4,mm4
 200 
       pxor	 mm4,mm4
201 
       punpcklwd mm0,mm4
 201 
       punpcklwd mm0,mm4
202 
       punpcklwd mm1,mm4
 202 
       punpcklwd mm1,mm4
203 
       punpcklwd mm2,mm4
 203 
       punpcklwd mm2,mm4
204 
       punpcklwd mm3,mm4
 204 
       punpcklwd mm3,mm4
205 
   ;    pslld     mm0,ROUND
 205 
   ;    pslld     mm0,ROUND
206 
   ;    pslld     mm1,ROUND
 206 
   ;    pslld     mm1,ROUND
207 
   ;    pslld     mm2,ROUND
 207 
   ;    pslld     mm2,ROUND
208 
   ;    pslld     mm3,ROUND
 208 
   ;    pslld     mm3,ROUND
209 
       cvtpi2ps  xmm0,mm0
 209 
       cvtpi2ps  xmm0,mm0
210 
       movlhps	 xmm0,xmm0
 210 
       movlhps	 xmm0,xmm0
211 
       cvtpi2ps  xmm0,mm2
 211 
       cvtpi2ps  xmm0,mm2
212 
       cvtpi2ps  xmm1,mm1
 212 
       cvtpi2ps  xmm1,mm1
213 
       movlhps	 xmm1,xmm1
 213 
       movlhps	 xmm1,xmm1
214 
       cvtpi2ps  xmm1,mm3
 214 
       cvtpi2ps  xmm1,mm3
215 
       subps	 xmm1,xmm0
 215 
       subps	 xmm1,xmm0
216 
 
 216 
 
217 
     ;  pxor      mm4,mm4
 217 
     ;  pxor      mm4,mm4
218 
     ;  movq      mm5,mm1
 218 
     ;  movq      mm5,mm1
219 
     ;  movq      mm6,mm1
 219 
     ;  movq      mm6,mm1
220 
     ;  pcmpeqb   mm5,mm4
 220 
     ;  pcmpeqb   mm5,mm4
221 
;       psubd     mm1,mm0
 221 
;       psubd     mm1,mm0
222 
;       psubd     mm3,mm2
 222 
;       psubd     mm3,mm2
223 
 
 223 
 
224 
    ;   movq      mm0,[.b_x1]      ; bx1  by1   bx2    by2
 224 
    ;   movq      mm0,[.b_x1]      ; bx1  by1   bx2    by2
225 
    ;   movq      mm1,[.e_x1]      ; ex1  ey1   ex2    ey2
 225 
    ;   movq      mm1,[.e_x1]      ; ex1  ey1   ex2    ey2
226 
    ;   pxor
 226 
    ;   pxor
227 
    ;   punpcklhd mm0,mm1   ; lwd  ;
 227 
    ;   punpcklhd mm0,mm1   ; lwd  ;
228 
    ;   psubw     mm1,mm0   ; mm1, mm0
 228 
    ;   psubw     mm1,mm0   ; mm1, mm0
229 
    ;   pxor      mm2,mm2
 229 
    ;   pxor      mm2,mm2
230 
  ;     pmovmaskb eax,mm1
 230 
  ;     pmovmaskb eax,mm1
231 
  ;     and       eax,10101010b
 231 
  ;     and       eax,10101010b
232 
    ;   pcmpgtw   mm2,mm1
 232 
    ;   pcmpgtw   mm2,mm1
233 
    ;   punpcklwd mm1,mm2
 233 
    ;   punpcklwd mm1,mm2
234 
 ;      psllw     mm0,ROUND
 234 
 ;      psllw     mm0,ROUND
235 
 ;      psllw     mm1,ROUND
 235 
 ;      psllw     mm1,ROUND
236 
 ;      movq      mm2,mm0
 236 
 ;      movq      mm2,mm0
237 
 ;      psrlq     mm0,32
 237 
 ;      psrlq     mm0,32
238 
 
 238 
 
239 
;       cvtpi2ps  xmm0,mm1
 239 
;       cvtpi2ps  xmm0,mm1
240 
;       movlhps   xmm0,xmm0
 240 
;       movlhps   xmm0,xmm0
241 
;       cvtpi2ps  xmm0,mm3
 241 
;       cvtpi2ps  xmm0,mm3
242 
  ;     divps     xmm1,xmm3
 242 
  ;     divps     xmm1,xmm3
243 
       mulps	 xmm1,xmm3
 243 
       mulps	 xmm1,xmm3
244 
       shufps	 xmm1,xmm1,10110001b
 244 
       shufps	 xmm1,xmm1,10110001b
245 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
 245 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
246 
       movhlps	 xmm1,xmm1
 246 
       movhlps	 xmm1,xmm1
247 
       cvtps2pi  mm1,xmm1
 247 
       cvtps2pi  mm1,xmm1
248 
       movq	 .dey12q,mm0
 248 
       movq	 .dey12q,mm0
249 
       movq	 .dby12q,mm1
 249 
       movq	 .dby12q,mm1
250 
 
 250 
 
251 
;       movd      .dex12,mm0
 251 
;       movd      .dex12,mm0
252 
;       psrlq     mm0,32
 252 
;       psrlq     mm0,32
253 
;       movd      .dey12,mm0
 253 
;       movd      .dey12,mm0
254 
;       movhlps   xmm1,xmm1
 254 
;       movhlps   xmm1,xmm1
255 
;       cvtps2pi  mm0,xmm1
 255 
;       cvtps2pi  mm0,xmm1
256 
;       movd      .dbx12,mm0
 256 
;       movd      .dbx12,mm0
257 
;       psrlq     mm0,32
 257 
;       psrlq     mm0,32
258 
;       movd      .dby12,mm0
 258 
;       movd      .dby12,mm0
259 
 
 259 
 
260 
else
 260 
else
261 
       mov	ax,word[.b_x2]
 261 
       mov	ax,word[.b_x2]
262 
       sub	ax,word[.b_x1]
 262 
       sub	ax,word[.b_x1]
263 
       cwde
 263 
       cwde
264 
       shl	eax,ROUND
 264 
       shl	eax,ROUND
265 
       cdq
 265 
       cdq
266 
       idiv	ebx
 266 
       idiv	ebx
267 
 ;     mov      .dbx12,eax
 267 
 ;     mov      .dbx12,eax
268 
       push	 eax
 268 
       push	 eax
269 
 
 269 
 
270 
       mov	ax,word[.b_y2]
 270 
       mov	ax,word[.b_y2]
271 
       sub	ax,word[.b_y1]
 271 
       sub	ax,word[.b_y1]
272 
       cwde
 272 
       cwde
273 
       shl	eax,ROUND
 273 
       shl	eax,ROUND
274 
       cdq
 274 
       cdq
275 
       idiv	ebx
 275 
       idiv	ebx
276 
 ;     mov      .dby12,eax
 276 
 ;     mov      .dby12,eax
277 
       push	 eax
 277 
       push	 eax
278 
 
 278 
 
279 
 ;     mov       eax,.dbx12
 279 
 ;     mov       eax,.dbx12
280 
 ;     mov       ebx,.dby12
 280 
 ;     mov       ebx,.dby12
281 
 ;     int3
 281 
 ;     int3
282 
 
 282 
 
283 
       mov	ax,word[.e_x2]
 283 
       mov	ax,word[.e_x2]
284 
       sub	ax,word[.e_x1]
 284 
       sub	ax,word[.e_x1]
285 
       cwde
 285 
       cwde
286 
       shl	eax,ROUND
 286 
       shl	eax,ROUND
287 
       cdq
 287 
       cdq
288 
       idiv	ebx
 288 
       idiv	ebx
289 
 ;      mov      .dex12,eax
 289 
 ;      mov      .dex12,eax
290 
       push	 eax
 290 
       push	 eax
291 
 
 291 
 
292 
       mov	ax,word[.e_y2]
 292 
       mov	ax,word[.e_y2]
293 
       sub	ax,word[.e_y1]
 293 
       sub	ax,word[.e_y1]
294 
       cwde
 294 
       cwde
295 
       shl	eax,ROUND
 295 
       shl	eax,ROUND
296 
       cdq
 296 
       cdq
297 
       idiv	ebx
 297 
       idiv	ebx
298 
 ;      mov      .dey12,eax
 298 
 ;      mov      .dey12,eax
299 
       push	 eax
 299 
       push	 eax
300 
 
 300 
 
301 
end if
 301 
end if
302 
	mov	ax,.z2
 302 
	mov	ax,.z2
303 
	sub	ax,.z1
 303 
	sub	ax,.z1
304 
	cwde
 304 
	cwde
305 
	shl	eax,CATMULL_SHIFT
 305 
	shl	eax,CATMULL_SHIFT
306 
	cdq
 306 
	cdq
307 
	idiv	ebx
 307 
	idiv	ebx
308 
	push	eax
 308 
	push	eax
309 
   .bt_dx12_done:
 309 
   .bt_dx12_done:
310 
 
 310 
 
311 
       mov	bx,.y3	     ; calc delta13
 311 
       mov	bx,.y3	     ; calc delta13
312 
       sub	bx,.y1
 312 
       sub	bx,.y1
313 
       jnz	.bt_dx13_make
 313 
       jnz	.bt_dx13_make
314 
       mov	ecx,6
 314 
       mov	ecx,6
315 
       xor	edx,edx
 315 
       xor	edx,edx
316 
     @@:
 316 
     @@:
317 
       push	edx   ;dword 0
 317 
       push	edx   ;dword 0
318 
       loop	@b
 318 
       loop	@b
319 
       jmp	.bt_dx13_done
 319 
       jmp	.bt_dx13_done
320 
 .bt_dx13_make:
 320 
 .bt_dx13_make:
321 
       mov	ax,.x3
 321 
       mov	ax,.x3
322 
       sub	ax,.x1
 322 
       sub	ax,.x1
323 
       cwde
 323 
       cwde
324 
       movsx	ebx,bx
 324 
       movsx	ebx,bx
325 
       shl	eax,ROUND
 325 
       shl	eax,ROUND
326 
       cdq
 326 
       cdq
327 
       idiv	ebx
 327 
       idiv	ebx
328 
 ;      mov      .dx13,eax
 328 
 ;      mov      .dx13,eax
329 
       push	 eax
 329 
       push	 eax
330 
 
 330 
 
331 
if Ext=SSE
 331 
if Ext=SSE
332 
 
 332 
 
333 
       cvtsi2ss  xmm3,ebx
 333 
       cvtsi2ss  xmm3,ebx
334 
  ;     mov       eax,255
 334 
  ;     mov       eax,255
335 
       cvtsi2ss  xmm4,[i255d]
 335 
       cvtsi2ss  xmm4,[i255d]
336 
       divss	 xmm3,xmm4
 336 
       divss	 xmm3,xmm4
337 
       rcpss	 xmm3,xmm3
 337 
       rcpss	 xmm3,xmm3
338 
;       mulss     xmm3,xmm4
 338 
;       mulss     xmm3,xmm4
339 
       shufps	 xmm3,xmm3,0
 339 
       shufps	 xmm3,xmm3,0
340 
       sub	 esp,16
 340 
       sub	 esp,16
341 
 
 341 
 
342 
       movd	 mm0,[.b_x1]
 342 
       movd	 mm0,[.b_x1]
343 
       movd	 mm1,[.b_x3]
 343 
       movd	 mm1,[.b_x3]
344 
       movd	 mm2,[.e_x1]
 344 
       movd	 mm2,[.e_x1]
345 
       movd	 mm3,[.e_x3]
 345 
       movd	 mm3,[.e_x3]
346 
 
 346 
 
347 
       pxor	 mm4,mm4
 347 
       pxor	 mm4,mm4
348 
       punpcklwd mm0,mm4
 348 
       punpcklwd mm0,mm4
349 
       punpcklwd mm1,mm4
 349 
       punpcklwd mm1,mm4
350 
       punpcklwd mm2,mm4
 350 
       punpcklwd mm2,mm4
351 
       punpcklwd mm3,mm4
 351 
       punpcklwd mm3,mm4
352 
 
 352 
 
353 
       cvtpi2ps  xmm0,mm0
 353 
       cvtpi2ps  xmm0,mm0
354 
       movlhps	 xmm0,xmm0
 354 
       movlhps	 xmm0,xmm0
355 
       cvtpi2ps  xmm0,mm2
 355 
       cvtpi2ps  xmm0,mm2
356 
       cvtpi2ps  xmm1,mm1
 356 
       cvtpi2ps  xmm1,mm1
357 
       movlhps	 xmm1,xmm1
 357 
       movlhps	 xmm1,xmm1
358 
       cvtpi2ps  xmm1,mm3
 358 
       cvtpi2ps  xmm1,mm3
359 
       subps	 xmm1,xmm0
 359 
       subps	 xmm1,xmm0
360 
 
 360 
 
361 
   ;    divps     xmm1,xmm3
 361 
   ;    divps     xmm1,xmm3
362 
       mulps	 xmm1,xmm3
 362 
       mulps	 xmm1,xmm3
363 
       shufps	 xmm1,xmm1,10110001b
 363 
       shufps	 xmm1,xmm1,10110001b
364 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
 364 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
365 
       movhlps	 xmm1,xmm1
 365 
       movhlps	 xmm1,xmm1
366 
       cvtps2pi  mm1,xmm1
 366 
       cvtps2pi  mm1,xmm1
367 
       movq	 .dey13q,mm0
 367 
       movq	 .dey13q,mm0
368 
       movq	 .dby13q,mm1
 368 
       movq	 .dby13q,mm1
369 
 
 369 
 
370 
else
 370 
else
371 
 
 371 
 
372 
       mov	ax,word[.b_x3]
 372 
       mov	ax,word[.b_x3]
373 
       sub	ax,word[.b_x1]
 373 
       sub	ax,word[.b_x1]
374 
       cwde
 374 
       cwde
375 
       shl	eax,ROUND
 375 
       shl	eax,ROUND
376 
       cdq
 376 
       cdq
377 
       idiv	ebx
 377 
       idiv	ebx
378 
 ;      mov      .dbx13,eax
 378 
 ;      mov      .dbx13,eax
379 
       push	 eax
 379 
       push	 eax
380 
 
 380 
 
381 
       mov	ax,word[.b_y3]
 381 
       mov	ax,word[.b_y3]
382 
       sub	ax,word[.b_y1]
 382 
       sub	ax,word[.b_y1]
383 
       cwde
 383 
       cwde
384 
       shl	eax,ROUND
 384 
       shl	eax,ROUND
385 
       cdq
 385 
       cdq
386 
       idiv	ebx
 386 
       idiv	ebx
387 
 ;      mov      .dby13,eax
 387 
 ;      mov      .dby13,eax
388 
       push	 eax
 388 
       push	 eax
389 
 
 389 
 
390 
       mov	ax,word[.e_x3]
 390 
       mov	ax,word[.e_x3]
391 
       sub	ax,word[.e_x1]
 391 
       sub	ax,word[.e_x1]
392 
       cwde
 392 
       cwde
393 
       shl	eax,ROUND
 393 
       shl	eax,ROUND
394 
       cdq
 394 
       cdq
395 
       idiv	ebx
 395 
       idiv	ebx
396 
 ;      mov      .dex13,eax
 396 
 ;      mov      .dex13,eax
397 
       push	 eax
 397 
       push	 eax
398 
 
 398 
 
399 
       mov	ax,word[.e_y3]
 399 
       mov	ax,word[.e_y3]
400 
       sub	ax,word[.e_y1]
 400 
       sub	ax,word[.e_y1]
401 
       cwde
 401 
       cwde
402 
       shl	eax,ROUND
 402 
       shl	eax,ROUND
403 
       cdq
 403 
       cdq
404 
       idiv	ebx
 404 
       idiv	ebx
405 
 ;      mov      .dey13,eax
 405 
 ;      mov      .dey13,eax
406 
       push	 eax
 406 
       push	 eax
407 
 
 407 
 
408 
end if
 408 
end if
409 
 
 409 
 
410 
       mov     ax,.z3
 410 
       mov     ax,.z3
411 
       sub     ax,.z1
 411 
       sub     ax,.z1
412 
       cwde
 412 
       cwde
413 
       shl     eax,CATMULL_SHIFT
 413 
       shl     eax,CATMULL_SHIFT
414 
       cdq
 414 
       cdq
415 
       idiv    ebx
 415 
       idiv    ebx
416 
  ;    mov    .dz13,eax
 416 
  ;    mov    .dz13,eax
417 
       push    eax
 417 
       push    eax
418 
   .bt_dx13_done:
 418 
   .bt_dx13_done:
419 
 
 419 
 
420 
       mov	bx,.y3	     ; calc delta23
 420 
       mov	bx,.y3	     ; calc delta23
421 
       sub	bx,.y2
 421 
       sub	bx,.y2
422 
       jnz	.bt_dx23_make
 422 
       jnz	.bt_dx23_make
423 
       mov	ecx,6
 423 
       mov	ecx,6
424 
       xor	edx,edx
 424 
       xor	edx,edx
425 
     @@:
 425 
     @@:
426 
       push	edx   ;dword 0
 426 
       push	edx   ;dword 0
427 
       loop	@b
 427 
       loop	@b
428 
       jmp	.bt_dx23_done
 428 
       jmp	.bt_dx23_done
429 
 .bt_dx23_make:
 429 
 .bt_dx23_make:
430 
       mov	ax,.x3
 430 
       mov	ax,.x3
431 
       sub	ax,.x2
 431 
       sub	ax,.x2
432 
       cwde
 432 
       cwde
433 
       movsx	ebx,bx
 433 
       movsx	ebx,bx
434 
       shl	eax,ROUND
 434 
       shl	eax,ROUND
435 
       cdq
 435 
       cdq
436 
       idiv	ebx
 436 
       idiv	ebx
437 
 ;      mov      .dx23,eax
 437 
 ;      mov      .dx23,eax
438 
       push	 eax
 438 
       push	 eax
439 
 
 439 
 
440 
if Ext=SSE
 440 
if Ext=SSE
441 
 
 441 
 
442 
       cvtsi2ss  xmm3,ebx
 442 
       cvtsi2ss  xmm3,ebx
443 
     ;  mov       eax,255
 443 
     ;  mov       eax,255
444 
       cvtsi2ss  xmm4,[i255d] ;eax
 444 
       cvtsi2ss  xmm4,[i255d] ;eax
445 
       divss	 xmm3,xmm4
 445 
       divss	 xmm3,xmm4
446 
       shufps	 xmm3,xmm3,0
 446 
       shufps	 xmm3,xmm3,0
447 
       sub	 esp,16
 447 
       sub	 esp,16
448 
 
 448 
 
449 
       movd	 mm0,[.b_x2]
 449 
       movd	 mm0,[.b_x2]
450 
       movd	 mm1,[.b_x3]
 450 
       movd	 mm1,[.b_x3]
451 
       movd	 mm2,[.e_x2]
 451 
       movd	 mm2,[.e_x2]
452 
       movd	 mm3,[.e_x3]
 452 
       movd	 mm3,[.e_x3]
453 
 
 453 
 
454 
       pxor	 mm4,mm4
 454 
       pxor	 mm4,mm4
455 
       punpcklwd mm0,mm4
 455 
       punpcklwd mm0,mm4
456 
       punpcklwd mm1,mm4
 456 
       punpcklwd mm1,mm4
457 
       punpcklwd mm2,mm4
 457 
       punpcklwd mm2,mm4
458 
       punpcklwd mm3,mm4
 458 
       punpcklwd mm3,mm4
459 
 
 459 
 
460 
       cvtpi2ps  xmm0,mm0
 460 
       cvtpi2ps  xmm0,mm0
461 
       movlhps	 xmm0,xmm0
 461 
       movlhps	 xmm0,xmm0
462 
       cvtpi2ps  xmm0,mm2
 462 
       cvtpi2ps  xmm0,mm2
463 
       cvtpi2ps  xmm1,mm1
 463 
       cvtpi2ps  xmm1,mm1
464 
       movlhps	 xmm1,xmm1
 464 
       movlhps	 xmm1,xmm1
465 
       cvtpi2ps  xmm1,mm3
 465 
       cvtpi2ps  xmm1,mm3
466 
       subps	 xmm1,xmm0
 466 
       subps	 xmm1,xmm0
467 
 
 467 
 
468 
       divps	 xmm1,xmm3
 468 
       divps	 xmm1,xmm3
469 
       shufps	 xmm1,xmm1,10110001b
 469 
       shufps	 xmm1,xmm1,10110001b
470 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
 470 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
471 
       movhlps	 xmm1,xmm1
 471 
       movhlps	 xmm1,xmm1
472 
       cvtps2pi  mm1,xmm1
 472 
       cvtps2pi  mm1,xmm1
473 
       movq	 .dey23q,mm0
 473 
       movq	 .dey23q,mm0
474 
       movq	 .dby23q,mm1
 474 
       movq	 .dby23q,mm1
475 
 
 475 
 
476 
else
 476 
else
477 
 
 477 
 
478 
       mov	ax,word[.b_x3]
 478 
       mov	ax,word[.b_x3]
479 
       sub	ax,word[.b_x2]
 479 
       sub	ax,word[.b_x2]
480 
       cwde
 480 
       cwde
481 
       shl	eax,ROUND
 481 
       shl	eax,ROUND
482 
       cdq
 482 
       cdq
483 
       idiv	ebx
 483 
       idiv	ebx
484 
 ;      mov      .dbx23,eax
 484 
 ;      mov      .dbx23,eax
485 
       push	 eax
 485 
       push	 eax
486 
 
 486 
 
487 
       mov	ax,word[.b_y3]
 487 
       mov	ax,word[.b_y3]
488 
       sub	ax,word[.b_y2]
 488 
       sub	ax,word[.b_y2]
489 
       cwde
 489 
       cwde
490 
       shl	eax,ROUND
 490 
       shl	eax,ROUND
491 
       cdq
 491 
       cdq
492 
       idiv	ebx
 492 
       idiv	ebx
493 
 ;      mov      .dby23,eax
 493 
 ;      mov      .dby23,eax
494 
       push	 eax
 494 
       push	 eax
495 
 
 495 
 
496 
       mov	ax,word[.e_x3]
 496 
       mov	ax,word[.e_x3]
497 
       sub	ax,word[.e_x2]
 497 
       sub	ax,word[.e_x2]
498 
       cwde
 498 
       cwde
499 
       shl	eax,ROUND
 499 
       shl	eax,ROUND
500 
       cdq
 500 
       cdq
501 
       idiv	ebx
 501 
       idiv	ebx
502 
 ;      mov      .dex23,eax
 502 
 ;      mov      .dex23,eax
503 
       push	 eax
 503 
       push	 eax
504 
 
 504 
 
505 
       mov	ax,word[.e_y3]
 505 
       mov	ax,word[.e_y3]
506 
       sub	ax,word[.e_y2]
 506 
       sub	ax,word[.e_y2]
507 
       cwde
 507 
       cwde
508 
       shl	eax,ROUND
 508 
       shl	eax,ROUND
509 
       cdq
 509 
       cdq
510 
       idiv	ebx
 510 
       idiv	ebx
511 
 ;      mov      .dey23,eax
 511 
 ;      mov      .dey23,eax
512 
       push	 eax
 512 
       push	 eax
513 
end if
 513 
end if
514 
       mov     ax,.z3
 514 
       mov     ax,.z3
515 
       sub     ax,.z2
 515 
       sub     ax,.z2
516 
       cwde
 516 
       cwde
517 
       shl     eax,CATMULL_SHIFT
 517 
       shl     eax,CATMULL_SHIFT
518 
       cdq
 518 
       cdq
519 
       idiv    ebx
 519 
       idiv    ebx
520 
      ; mov     .dz23,eax
 520 
      ; mov     .dz23,eax
521 
       push    eax
 521 
       push    eax
522 
      ;  sub     esp,40
 522 
      ;  sub     esp,40
523 
   .bt_dx23_done:
 523 
   .bt_dx23_done:
524 
       movsx	eax,.x1
 524 
       movsx	eax,.x1
525 
       shl	eax,ROUND
 525 
       shl	eax,ROUND
526 
     ;  mov      .cx1,eax
 526 
     ;  mov      .cx1,eax
527 
     ;  mov      .cx2,eax
 527 
     ;  mov      .cx2,eax
528 
       push	eax eax
 528 
       push	eax eax
529 
     ; push     eax
 529 
     ; push     eax
530 
 
 530 
 
531 
       movsx	eax,word[.b_x1]
 531 
       movsx	eax,word[.b_x1]
532 
       shl	eax,ROUND
 532 
       shl	eax,ROUND
533 
       mov	.cbx1,eax
 533 
       mov	.cbx1,eax
534 
       mov	.cbx2,eax
 534 
       mov	.cbx2,eax
535 
     ;  push     eax eax
 535 
     ;  push     eax eax
536 
     ; push     eax
 536 
     ; push     eax
537 
 
 537 
 
538 
       movsx	eax,word[.b_y1]
 538 
       movsx	eax,word[.b_y1]
539 
       shl	eax,ROUND
 539 
       shl	eax,ROUND
540 
       mov	.cby1,eax
 540 
       mov	.cby1,eax
541 
       mov	.cby2,eax
 541 
       mov	.cby2,eax
542 
     ;  push     eax eax
 542 
     ;  push     eax eax
543 
     ; push     eax
 543 
     ; push     eax
544 
 
 544 
 
545 
       movsx	eax,word[.e_x1]
 545 
       movsx	eax,word[.e_x1]
546 
       shl	eax,ROUND
 546 
       shl	eax,ROUND
547 
       mov	.cex1,eax
 547 
       mov	.cex1,eax
548 
       mov	.cex2,eax
 548 
       mov	.cex2,eax
549 
      ; push     eax eax
 549 
      ; push     eax eax
550 
      ;push     eax
 550 
      ;push     eax
551 
 
 551 
 
552 
       movsx	eax,word[.e_y1]
 552 
       movsx	eax,word[.e_y1]
553 
       shl	eax,ROUND
 553 
       shl	eax,ROUND
554 
       mov	.cey1,eax
 554 
       mov	.cey1,eax
555 
       mov	.cey2,eax
 555 
       mov	.cey2,eax
556 
       sub	esp,32
 556 
       sub	esp,32
557 
      ; push     eax eax
 557 
      ; push     eax eax
558 
      ;push     eax
 558 
      ;push     eax
559 
 
 559 
 
560 
       movsx	eax,.z1
 560 
       movsx	eax,.z1
561 
       shl	eax,CATMULL_SHIFT
 561 
       shl	eax,CATMULL_SHIFT
562 
      ; mov      .cz1,eax
 562 
      ; mov      .cz1,eax
563 
      ; mov      .cz2,eax
 563 
      ; mov      .cz2,eax
564 
      push     eax eax
 564 
      push     eax eax
565 
      ;push     eax
 565 
      ;push     eax
566 
 
 566 
 
567 
       movsx	ecx,.y1
 567 
       movsx	ecx,.y1
568 
       cmp	cx,.y2
 568 
       cmp	cx,.y2
569 
       jge	.loop12_done
 569 
       jge	.loop12_done
570 
  .loop12:
 570 
  .loop12:
571 
       call	.call_line
 571 
       call	.call_line
572 
 
 572 
 
573 
       mov	eax,.dx13
 573 
       mov	eax,.dx13
574 
       add	.cx1,eax
 574 
       add	.cx1,eax
575 
       mov	ebx,.dx12
 575 
       mov	ebx,.dx12
576 
       add	.cx2,ebx
 576 
       add	.cx2,ebx
577 
if Ext>= SSE2
 577 
if Ext>= SSE2
578 
       movups	xmm0,.cey1
 578 
       movups	xmm0,.cey1
579 
       movups	xmm1,.cey2
 579 
       movups	xmm1,.cey2
580 
       movups	xmm2,.dey12q
 580 
       movups	xmm2,.dey12q
581 
       movups	xmm3,.dey13q
 581 
       movups	xmm3,.dey13q
582 
       paddd	xmm0,xmm3
 582 
       paddd	xmm0,xmm3
583 
       paddd	xmm1,xmm2
 583 
       paddd	xmm1,xmm2
584 
       movups	.cey1,xmm0
 584 
       movups	.cey1,xmm0
585 
       movups	.cey2,xmm1
 585 
       movups	.cey2,xmm1
586 
else if (Ext = MMX) | (Ext=SSE)
 586 
else if (Ext = MMX) | (Ext=SSE)
587 
       movq	mm0,.cby2	; with this optimization object
 587 
       movq	mm0,.cby2	; with this optimization object
588 
       movq	mm1,.cby1	; looks bit annoying
 588 
       movq	mm1,.cby1	; looks bit annoying
589 
       movq	mm2,.cey2
 589 
       movq	mm2,.cey2
590 
       movq	mm3,.cey1
 590 
       movq	mm3,.cey1
591 
       paddd	mm0,.dby12q
 591 
       paddd	mm0,.dby12q
592 
       paddd	mm1,.dby13q
 592 
       paddd	mm1,.dby13q
593 
       paddd	mm2,.dey12q
 593 
       paddd	mm2,.dey12q
594 
       paddd	mm3,.dey13q
 594 
       paddd	mm3,.dey13q
595 
       movq	.cby2,mm0
 595 
       movq	.cby2,mm0
596 
       movq	.cby1,mm1
 596 
       movq	.cby1,mm1
597 
       movq	.cey1,mm3
 597 
       movq	.cey1,mm3
598 
       movq	.cey2,mm2
 598 
       movq	.cey2,mm2
599 
else
 599 
else
600 
       mov	edx,.dbx13
 600 
       mov	edx,.dbx13
601 
       add	.cbx1,edx
 601 
       add	.cbx1,edx
602 
       mov	eax,.dbx12
 602 
       mov	eax,.dbx12
603 
       add	.cbx2,eax
 603 
       add	.cbx2,eax
604 
       mov	ebx,.dby13
 604 
       mov	ebx,.dby13
605 
       add	.cby1,ebx
 605 
       add	.cby1,ebx
606 
       mov	edx,.dby12
 606 
       mov	edx,.dby12
607 
       add	.cby2,edx
 607 
       add	.cby2,edx
608 
 
 608 
 
609 
       mov	eax,.dex13
 609 
       mov	eax,.dex13
610 
       add	.cex1,eax
 610 
       add	.cex1,eax
611 
       mov	ebx,.dex12
 611 
       mov	ebx,.dex12
612 
       add	.cex2,ebx
 612 
       add	.cex2,ebx
613 
       mov	edx,.dey13
 613 
       mov	edx,.dey13
614 
       add	.cey1,edx
 614 
       add	.cey1,edx
615 
       mov	eax,.dey12
 615 
       mov	eax,.dey12
616 
       add	.cey2,eax
 616 
       add	.cey2,eax
617 
 
 617 
 
618 
end if
 618 
end if
619 
       mov	ebx,.dz13
 619 
       mov	ebx,.dz13
620 
       add	.cz1,ebx
 620 
       add	.cz1,ebx
621 
       mov	edx,.dz12
 621 
       mov	edx,.dz12
622 
       add	.cz2,edx
 622 
       add	.cz2,edx
623 
 
 623 
 
624 
       inc	ecx
 624 
       inc	ecx
625 
       cmp	cx,.y2
 625 
       cmp	cx,.y2
626 
       jl	.loop12
 626 
       jl	.loop12
627 
    .loop12_done:
 627 
    .loop12_done:
628 
 
 628 
 
629 
       movsx	ecx,.y2
 629 
       movsx	ecx,.y2
630 
       cmp	cx,.y3
 630 
       cmp	cx,.y3
631 
       jge	.loop23_done
 631 
       jge	.loop23_done
632 
 
 632 
 
633 
       movsx	eax,.z2
 633 
       movsx	eax,.z2
634 
       shl	eax,CATMULL_SHIFT
 634 
       shl	eax,CATMULL_SHIFT
635 
       mov	.cz2,eax
 635 
       mov	.cz2,eax
636 
 
 636 
 
637 
       movsx	eax,.x2
 637 
       movsx	eax,.x2
638 
       shl	eax,ROUND
 638 
       shl	eax,ROUND
639 
       mov	.cx2,eax
 639 
       mov	.cx2,eax
640 
 
 640 
 
641 
       movzx	eax,word[.b_x2]
 641 
       movzx	eax,word[.b_x2]
642 
       shl	eax,ROUND
 642 
       shl	eax,ROUND
643 
       mov	.cbx2,eax
 643 
       mov	.cbx2,eax
644 
 
 644 
 
645 
       movzx	eax,word[.b_y2]
 645 
       movzx	eax,word[.b_y2]
646 
       shl	eax,ROUND
 646 
       shl	eax,ROUND
647 
       mov	.cby2,eax
 647 
       mov	.cby2,eax
648 
 
 648 
 
649 
       movzx	eax,word[.e_x2]
 649 
       movzx	eax,word[.e_x2]
650 
       shl	eax,ROUND
 650 
       shl	eax,ROUND
651 
       mov	.cex2,eax
 651 
       mov	.cex2,eax
652 
 
 652 
 
653 
       movzx	eax,word[.e_y2]
 653 
       movzx	eax,word[.e_y2]
654 
       shl	eax,ROUND
 654 
       shl	eax,ROUND
655 
       mov	.cey2,eax
 655 
       mov	.cey2,eax
656 
 
 656 
 
657 
     .loop23:
 657 
     .loop23:
658 
       call	.call_line
 658 
       call	.call_line
659 
;if Ext = NON
 659 
;if Ext = NON
660 
       mov	eax,.dx13
 660 
       mov	eax,.dx13
661 
       add	.cx1,eax
 661 
       add	.cx1,eax
662 
       mov	ebx,.dx23
 662 
       mov	ebx,.dx23
663 
       add	.cx2,ebx
 663 
       add	.cx2,ebx
664 
if Ext>= SSE2
 664 
if Ext>= SSE2
665 
       movups	xmm0,.cey1
 665 
       movups	xmm0,.cey1
666 
       movups	xmm1,.cey2
 666 
       movups	xmm1,.cey2
667 
       movups	xmm2,.dey23q
 667 
       movups	xmm2,.dey23q
668 
       movups	xmm3,.dey13q
 668 
       movups	xmm3,.dey13q
669 
       paddd	xmm0,xmm3
 669 
       paddd	xmm0,xmm3
670 
       paddd	xmm1,xmm2
 670 
       paddd	xmm1,xmm2
671 
       movups	.cey1,xmm0
 671 
       movups	.cey1,xmm0
672 
       movups	.cey2,xmm1
 672 
       movups	.cey2,xmm1
673 
else if (Ext = MMX) | ( Ext = SSE)
 673 
else if (Ext = MMX) | ( Ext = SSE)
674 
       movq	mm0,.cby2	 ;    with this mmx optimization object looks bit
 674 
       movq	mm0,.cby2	 ;    with this mmx optimization object looks bit
675 
       movq	mm1,.cby1	 ;    annoying
 675 
       movq	mm1,.cby1	 ;    annoying
676 
       movq	mm2,.cey2
 676 
       movq	mm2,.cey2
677 
       movq	mm3,.cey1
 677 
       movq	mm3,.cey1
678 
       paddd	mm0,.dby23q
 678 
       paddd	mm0,.dby23q
679 
       paddd	mm1,.dby13q
 679 
       paddd	mm1,.dby13q
680 
       paddd	mm2,.dey23q
 680 
       paddd	mm2,.dey23q
681 
       paddd	mm3,.dey13q
 681 
       paddd	mm3,.dey13q
682 
       movq	.cby2,mm0
 682 
       movq	.cby2,mm0
683 
       movq	.cby1,mm1
 683 
       movq	.cby1,mm1
684 
       movq	.cey2,mm2
 684 
       movq	.cey2,mm2
685 
       movq	.cey1,mm3
 685 
       movq	.cey1,mm3
686 
 
 686 
 
687 
else
 687 
else
688 
       mov	edx,.dbx13
 688 
       mov	edx,.dbx13
689 
       add	.cbx1,edx
 689 
       add	.cbx1,edx
690 
       mov	eax,.dbx23
 690 
       mov	eax,.dbx23
691 
       add	.cbx2,eax
 691 
       add	.cbx2,eax
692 
       mov	ebx,.dby13
 692 
       mov	ebx,.dby13
693 
       add	.cby1,ebx
 693 
       add	.cby1,ebx
694 
       mov	edx,.dby23
 694 
       mov	edx,.dby23
695 
       add	.cby2,edx
 695 
       add	.cby2,edx
696 
 
 696 
 
697 
       mov	eax,.dex13
 697 
       mov	eax,.dex13
698 
       add	.cex1,eax
 698 
       add	.cex1,eax
699 
       mov	ebx,.dex23
 699 
       mov	ebx,.dex23
700 
       add	.cex2,ebx
 700 
       add	.cex2,ebx
701 
       mov	edx,.dey13
 701 
       mov	edx,.dey13
702 
       add	.cey1,edx
 702 
       add	.cey1,edx
703 
       mov	eax,.dey23
 703 
       mov	eax,.dey23
704 
       add	.cey2,eax
 704 
       add	.cey2,eax
705 
end if
 705 
end if
706 
 
 706 
 
707 
       mov	ebx,.dz13
 707 
       mov	ebx,.dz13
708 
       add	.cz1,ebx
 708 
       add	.cz1,ebx
709 
       mov	edx,.dz23
 709 
       mov	edx,.dz23
710 
       add	.cz2,edx
 710 
       add	.cz2,edx
711 
;else
 711 
;else
712 
;       movq     mm0,.db13q
 712 
;       movq     mm0,.db13q
713 
;       movq     mm1,.cbx1q
 713 
;       movq     mm1,.cbx1q
714 
 
 714 
 
715 
       inc	ecx
 715 
       inc	ecx
716 
       cmp	cx,.y3
 716 
       cmp	cx,.y3
717 
       jl	.loop23
 717 
       jl	.loop23
718 
    .loop23_done:
 718 
    .loop23_done:
719 
 
 719 
 
720 
       mov	esp,ebp
 720 
       mov	esp,ebp
721 
ret   34
 721 
ret   34
722 
 
 722 
 
723 
.call_line:
 723 
.call_line:
724 
 
 724 
 
725 
       pushad
 725 
       pushad
726 
 
 726 
 
727 
       push	.cz1
 727 
       push	.cz1
728 
       push	.cz2
 728 
       push	.cz2
729 
       push	.z_buff
 729 
       push	.z_buff
730 
       push	.t_bmap
 730 
       push	.t_bmap
731 
       push	.t_emap
 731 
       push	.t_emap
732 
       push	dword .cey2
 732 
       push	dword .cey2
733 
       push	.cex2
 733 
       push	.cex2
734 
       push	dword .cey1
 734 
       push	dword .cey1
735 
       push	.cex1
 735 
       push	.cex1
736 
       push	dword .cby2
 736 
       push	dword .cby2
737 
       push	.cbx2
 737 
       push	.cbx2
738 
       push	dword .cby1
 738 
       push	dword .cby1
739 
       push	.cbx1
 739 
       push	.cbx1
740 
       push	ecx
 740 
       push	ecx
741 
 
 741 
 
742 
       mov	eax,.cx1
 742 
       mov	eax,.cx1
743 
       sar	eax,ROUND
 743 
       sar	eax,ROUND
744 
       mov	ebx,.cx2
 744 
       mov	ebx,.cx2
745 
       sar	ebx,ROUND
 745 
       sar	ebx,ROUND
746 
 
 746 
 
747 
       call	two_tex_line_z
 747 
       call	two_tex_line_z
748 
 
 748 
 
749 
       popad
 749 
       popad
750 
ret
 750 
ret
751 
two_tex_line_z:
 751 
two_tex_line_z:
752 
;--------------in: eax - x1
 752 
;--------------in: eax - x1
753 
;--------------    ebx - x2
 753 
;--------------    ebx - x2
754 
;--------------    edi - pointer to screen buffer
 754 
;--------------    edi - pointer to screen buffer
755 
;stack - another parameters :
 755 
;stack - another parameters :
756 
.y	equ dword [ebp+4]
 756 
.y	equ dword [ebp+4]
757 
.bx1	equ  [ebp+8]   ;   ---
 757 
.bx1	equ  [ebp+8]   ;   ---
758 
.by1	equ  [ebp+12]  ;       |
 758 
.by1	equ  [ebp+12]  ;       |
759 
.bx2	equ  [ebp+16]  ;       |
 759 
.bx2	equ  [ebp+16]  ;       |
760 
.by2	equ  [ebp+20]  ;       |>   b. texture and e. texture coords
 760 
.by2	equ  [ebp+20]  ;       |>   b. texture and e. texture coords
761 
.ex1	equ  [ebp+24]  ;       |>   shifted shl ROUND
 761 
.ex1	equ  [ebp+24]  ;       |>   shifted shl ROUND
762 
.ey1	equ  [ebp+28]  ;       |
 762 
.ey1	equ  [ebp+28]  ;       |
763 
.ex2	equ  [ebp+32]  ;       |
 763 
.ex2	equ  [ebp+32]  ;       |
764 
.ey2	equ  [ebp+36]  ;   ---
 764 
.ey2	equ  [ebp+36]  ;   ---
765 
.emap	equ  [ebp+40]  ; b texture offset
 765 
.emap	equ  [ebp+40]  ; b texture offset
766 
.bmap	equ  [ebp+44]  ; e texture offset
 766 
.bmap	equ  [ebp+44]  ; e texture offset
767 
.z_buff equ dword [ebp+48]
 767 
.z_buff equ dword [ebp+48]
768 
.z2	equ dword [ebp+52]  ;   -- |>   z coords shifted
 768 
.z2	equ dword [ebp+52]  ;   -- |>   z coords shifted
769 
.z1	equ dword [ebp+56]  ;   --       shl  CATMULL_SHIFT
 769 
.z1	equ dword [ebp+56]  ;   --       shl  CATMULL_SHIFT
770 
 
 770 
 
771 
.x1	equ dword [ebp-4]
 771 
.x1	equ dword [ebp-4]
772 
.x2	equ dword [ebp-8]
 772 
.x2	equ dword [ebp-8]
773 
.dbx	equ [ebp-12]
 773 
.dbx	equ [ebp-12]
774 
.dex	equ [ebp-16]
 774 
.dex	equ [ebp-16]
775 
.dby	equ [ebp-20]
 775 
.dby	equ [ebp-20]
776 
.dey	equ [ebp-24]
 776 
.dey	equ [ebp-24]
777 
.dz	equ dword [ebp-28]
 777 
.dz	equ dword [ebp-28]
778 
.cbx	equ [ebp-32]
 778 
.cbx	equ [ebp-32]
779 
.cex	equ [ebp-36]
 779 
.cex	equ [ebp-36]
780 
.cby	equ [ebp-40]
 780 
.cby	equ [ebp-40]
781 
.cey	equ [ebp-44]
 781 
.cey	equ [ebp-44]
782 
.cz	equ dword [ebp-48]
 782 
.cz	equ dword [ebp-48]
783 
.czbuff equ dword [ebp-52]
 783 
.czbuff equ dword [ebp-52]
784 
 
 784 
 
785 
	mov	ebp,esp
 785 
	mov	ebp,esp
786 
 
 786 
 
787 
	mov	ecx,.y
 787 
	mov	ecx,.y
788 
	or	ecx,ecx
 788 
	or	ecx,ecx
789 
	jl	.bl_end
 789 
	jl	.bl_end
790 
	cmp	cx,word[size_y_var]  ;SIZE_Y
 790 
	mov	dx,word[size_y_var]
- 
 
 791 
	dec	dx
- 
 
 792 
	cmp	cx,dx	;word[size_y_var]  ;SIZE_Y
791 
	jge	.bl_end
 793 
	jge	.bl_end
792 
 
 794 
 
793 
	cmp	eax,ebx
 795 
	cmp	eax,ebx
794 
	jl	@f
 796 
	jl	@f
795 
	je	.bl_end
 797 
	je	.bl_end
796 
 
 798 
 
797 
	xchg	eax,ebx
 799 
	xchg	eax,ebx
798 
if Ext=NON
 800 
if Ext=NON
799 
	mov	edx,.bx1
 801 
	mov	edx,.bx1
800 
	xchg	edx,.bx2
 802 
	xchg	edx,.bx2
801 
	mov	.bx1,edx
 803 
	mov	.bx1,edx
802 
	mov	edx,.by1
 804 
	mov	edx,.by1
803 
	xchg	edx,.by2
 805 
	xchg	edx,.by2
804 
	mov	.by1,edx
 806 
	mov	.by1,edx
805 
 
 807 
 
806 
	mov	edx,.ex1
 808 
	mov	edx,.ex1
807 
	xchg	edx,.ex2
 809 
	xchg	edx,.ex2
808 
	mov	.ex1,edx
 810 
	mov	.ex1,edx
809 
	mov	edx,.ey1
 811 
	mov	edx,.ey1
810 
	xchg	edx,.ey2
 812 
	xchg	edx,.ey2
811 
	mov	.ey1,edx
 813 
	mov	.ey1,edx
812 
else
 814 
else
813 
	movq	mm0,.bx1
 815 
	movq	mm0,.bx1
814 
	movq	mm1,.ex1
 816 
	movq	mm1,.ex1
815 
	movq	mm2,.bx2
 817 
	movq	mm2,.bx2
816 
	movq	mm3,.ex2
 818 
	movq	mm3,.ex2
817 
	movq	.bx2,mm0
 819 
	movq	.bx2,mm0
818 
	movq	.ex2,mm1
 820 
	movq	.ex2,mm1
819 
	movq	.bx1,mm2
 821 
	movq	.bx1,mm2
820 
	movq	.ex1,mm3
 822 
	movq	.ex1,mm3
821 
end if
 823 
end if
822 
	mov	edx,.z1
 824 
	mov	edx,.z1
823 
	xchg	edx,.z2
 825 
	xchg	edx,.z2
824 
	mov	.z1,edx
 826 
	mov	.z1,edx
825 
    @@:
 827 
    @@:
826 
	push	eax ebx    ;store x1, x2
 828 
	push	eax ebx    ;store x1, x2
827 
	mov	ebx,.x1
 829 
	mov	ebx,.x1
828 
	movzx	edx,word[size_x_var]
 830 
	movzx	edx,word[size_x_var]
829 
	cmp	ebx,edx
 831 
	dec	edx
- 
 
 832 
	cmp	ebx,edx
830 
  ;      cmp     bx,word[size_x_var]  ;SIZE_X
 833 
  ;      cmp     bx,word[size_x_var]  ;SIZE_X
831 
	jg	.bl_end
 834 
	jg	.bl_end
832 
	cmp	.x2,0
 835 
	cmp	.x2,0
833 
	jle	.bl_end
 836 
	jle	.bl_end
834 
 
 837 
 
835 
	mov	ebx,.x2
 838 
	mov	ebx,.x2
836 
	sub	ebx,.x1
 839 
	sub	ebx,.x1
837 
 
 840 
 
838 
if Ext>=SSE
 841 
if  Ext >= SSE
839 
 
 842 
 
840 
       sub	 esp,16
 843 
       sub	 esp,16
841 
       cvtsi2ss  xmm3,ebx	     ;rcps
 844 
       cvtsi2ss  xmm3,ebx	     ;rcps
842 
       shufps	 xmm3,xmm3,0
 845 
       shufps	 xmm3,xmm3,0
843 
 
 846 
 
844 
  ;     movq      mm0,.bx1q
 847 
  ;     movq      mm0,.bx1q
845 
  ;     movq      mm1,.bx2q
 848 
  ;     movq      mm1,.bx2q
846 
  ;     movq      mm2,.ex1q
 849 
  ;     movq      mm2,.ex1q
847 
  ;     movq      mm3,.ex2q
 850 
  ;     movq      mm3,.ex2q
848 
  ;     psubd     mm1,mm0
 851 
  ;     psubd     mm1,mm0
849 
  ;     psubd     mm3,mm2
 852 
  ;     psubd     mm3,mm2
850 
  ;     cvtpi2ps  xmm1,mm1
 853 
  ;     cvtpi2ps  xmm1,mm1
851 
  ;     movlhps   xmm1,xmm1
 854 
  ;     movlhps   xmm1,xmm1
852 
  ;     cvtpi2ps  xmm1,mm3
 855 
  ;     cvtpi2ps  xmm1,mm3
853 
 
 856 
 
854 
       cvtpi2ps  xmm0,.bx1 ;mm0     ; bx1; by1
 857 
       cvtpi2ps  xmm0,.bx1 ;mm0     ; bx1; by1
855 
       movlhps	 xmm0,xmm0
 858 
       movlhps	 xmm0,xmm0
856 
       cvtpi2ps  xmm0,.ex1 ;mm2     ; ex1; ey1
 859 
       cvtpi2ps  xmm0,.ex1 ;mm2     ; ex1; ey1
857 
       cvtpi2ps  xmm1,.bx2 ;mm1     ; bx2; by2
 860 
       cvtpi2ps  xmm1,.bx2 ;mm1     ; bx2; by2
858 
       movlhps	 xmm1,xmm1
 861 
       movlhps	 xmm1,xmm1
859 
       cvtpi2ps  xmm1,.ex2 ;mm3     ; ex2; ey2
 862 
       cvtpi2ps  xmm1,.ex2 ;mm3     ; ex2; ey2
860 
       subps	 xmm1,xmm0
 863 
       subps	 xmm1,xmm0
861 
				    ; hi             lo
 864 
				    ; hi             lo
862 
       divps	 xmm1,xmm3 ; xmm1 -> dby; dbx; dey; dex
 865 
       divps	 xmm1,xmm3 ; xmm1 -> dby; dbx; dey; dex
863 
 
 866 
 
864 
       shufps	 xmm1,xmm1,11011000b
 867 
       shufps	 xmm1,xmm1,11011000b
865 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
 868 
       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
866 
       movhlps	 xmm1,xmm1
 869 
       movhlps	 xmm1,xmm1
867 
       cvtps2pi  mm1,xmm1
 870 
       cvtps2pi  mm1,xmm1
868 
       movq	 .dex,mm0 ; hi - lo  ->  dbx, dex
 871 
       movq	 .dex,mm0 ; hi - lo  ->  dbx, dex
869 
       movq	 .dey,mm1 ; hi - lo  ->  dby, dey
 872 
       movq	 .dey,mm1 ; hi - lo  ->  dby, dey
870 
 
 873 
 
871 
else
 874 
else
872 
 
 875 
 
873 
	mov	eax,.bx2       ; calc .dbx
 876 
	mov	eax,.bx2       ; calc .dbx
874 
	sub	eax,.bx1
 877 
	sub	eax,.bx1
875 
	cdq
 878 
	cdq
876 
	idiv	ebx
 879 
	idiv	ebx
877 
	push	eax
 880 
	push	eax
878 
 
 881 
 
879 
	mov	eax,.ex2       ; calc .dby
 882 
	mov	eax,.ex2       ; calc .dby
880 
	sub	eax,.ex1
 883 
	sub	eax,.ex1
881 
	cdq
 884 
	cdq
882 
	idiv	ebx
 885 
	idiv	ebx
883 
	push	eax
 886 
	push	eax
884 
 
 887 
 
885 
	mov	eax,.by2       ; calc .dex
 888 
	mov	eax,.by2       ; calc .dex
886 
	sub	eax,.by1
 889 
	sub	eax,.by1
887 
	cdq
 890 
	cdq
888 
	idiv	ebx
 891 
	idiv	ebx
889 
	push	eax
 892 
	push	eax
890 
 
 893 
 
891 
	mov	eax,.ey2       ; calc .dey
 894 
	mov	eax,.ey2       ; calc .dey
892 
	sub	eax,.ey1
 895 
	sub	eax,.ey1
893 
	cdq
 896 
	cdq
894 
	idiv	ebx
 897 
	idiv	ebx
895 
	push	eax
 898 
	push	eax
896 
 
 899 
 
897 
end if
 900 
end if
898 
 
 901 
 
899 
	mov	eax,.z2        ; calc .dz
 902 
	mov	eax,.z2        ; calc .dz
900 
	sub	eax,.z1
 903 
	sub	eax,.z1
901 
	cdq
 904 
	cdq
902 
	idiv	ebx
 905 
	idiv	ebx
903 
	push	eax
 906 
	push	eax
904 
 
 907 
 
905 
	cmp	.x1,0	      ; set correctly begin variable
 908 
	cmp	.x1,0	      ; set correctly begin variable
906 
	jge	@f	      ; CLIPPING ON FUNCTION
 909 
	jge	@f	      ; CLIPPING ON FUNCTION
907 
			      ; cutting triangle exceedes screen
 910 
			      ; cutting triangle exceedes screen
908 
	mov	ebx,.x1
 911 
	mov	ebx,.x1
909 
	neg	ebx
 912 
	neg	ebx
910 
	imul	ebx	      ; eax = .dz * abs(.x1)
 913 
	imul	ebx	      ; eax = .dz * abs(.x1)
911 
	add	.z1,eax
 914 
	add	.z1,eax
912 
	mov	.x1,0
 915 
	mov	.x1,0
913 
 
 916 
 
914 
	mov	eax,.dbx
 917 
	mov	eax,.dbx
915 
	imul	ebx
 918 
	imul	ebx
916 
	add    .bx1,eax
 919 
	add    .bx1,eax
917 
 
 920 
 
918 
	mov	eax,.dby
 921 
	mov	eax,.dby
919 
	imul	ebx
 922 
	imul	ebx
920 
	add	.by1,eax
 923 
	add	.by1,eax
921 
 
 924 
 
922 
	mov	eax,.dex
 925 
	mov	eax,.dex
923 
	imul	ebx
 926 
	imul	ebx
924 
	add	.ex1,eax
 927 
	add	.ex1,eax
925 
 
 928 
 
926 
	mov	eax,.dey
 929 
	mov	eax,.dey
927 
	imul	ebx
 930 
	imul	ebx
928 
	add	.ey1,eax
 931 
	add	.ey1,eax
929 
      @@:
 932 
      @@:
930 
	movzx	eax,word[size_x_var] ;SIZE_X  ;word[size_x_var]
 933 
	movzx	eax,word[size_x_var] ;SIZE_X  ;word[size_x_var]
931 
	mov	ebx,.x2
 934 
	mov	ebx,.x2
932 
	cmp	eax,ebx
 935 
	cmp	eax,ebx
933 
	jg	@f
 936 
	jg	@f
934 
	mov	.x2,eax
 937 
	mov	.x2,eax
935 
      @@:
 938 
      @@:
936 
      ;  movd    mm0,eax
 939 
      ;  movd    mm0,eax
937 
      ;  movd    mm1,.x2
 940 
      ;  movd    mm1,.x2
938 
      ;  pminsw  mm0,mm1
 941 
      ;  pminsw  mm0,mm1
939 
      ;  movd    .x2,mm0
 942 
      ;  movd    .x2,mm0
940 
 ;       cmp     .x2,SIZE_X  ;eax   |
 943 
 ;       cmp     .x2,SIZE_X  ;eax   |
941 
 ;       jl      @f                 |>       this dont work idk cause
 944 
 ;       jl      @f                 |>       this dont work idk cause
942 
 ;       mov     .x2,SIZE_X ;eax    |
 945 
 ;       mov     .x2,SIZE_X ;eax    |
943 
      @@:
 946 
      @@:
944 
      ;  movzx   eax,word[size_x_var]       ;calc memory begin in buffers
 947 
      ;  movzx   eax,word[size_x_var]       ;calc memory begin in buffers
945 
	mov	ebx,.y
 948 
	mov	ebx,.y
946 
	mul	ebx
 949 
	mul	ebx
947 
	mov	ebx,.x1
 950 
	mov	ebx,.x1
948 
	add	eax,ebx
 951 
	add	eax,ebx
949 
	mov	ebx,eax
 952 
	mov	ebx,eax
950 
	lea	eax,[eax*3]
 953 
	lea	eax,[eax*3]
951 
	add	edi,eax 	  ; edi - screen
 954 
	add	edi,eax 	  ; edi - screen
952 
	mov	esi,.z_buff	  ; z-buffer filled with dd variables
 955 
	mov	esi,.z_buff	  ; z-buffer filled with dd variables
953 
	shl	ebx,2
 956 
	shl	ebx,2
954 
	add	esi,ebx 	  ; esi - Z buffer
 957 
	add	esi,ebx 	  ; esi - Z buffer
955 
 
 958 
 
956 
	mov	ecx,.x2
 959 
	mov	ecx,.x2
957 
	sub	ecx,.x1
 960 
	sub	ecx,.x1
958 
	; init current variables
 961 
	; init current variables
959 
	push	dword .bx1 ;.by1 .ex1 .ey1 .z1 esi
 962 
	push	dword .bx1 ;.by1 .ex1 .ey1 .z1 esi
960 
	push	dword .ex1
 963 
	push	dword .ex1
961 
	push	dword .by1
 964 
	push	dword .by1
962 
	push	dword .ey1
 965 
	push	dword .ey1
963 
 
 966 
 
964 
	push	.z1		 ; current z shl CATMULL_SHIFT
 967 
	push	.z1		 ; current z shl CATMULL_SHIFT
965 
	push	esi
 968 
	push	esi
966 
 
 969 
 
967 
if Ext >= MMX
 970 
if Ext >= MMX
968 
     pxor   mm0,mm0
 971 
     pxor   mm0,mm0
969 
     movq   mm3,.cex   ; hi - lo -> cbx; cex
 972 
     movq   mm3,.cex   ; hi - lo -> cbx; cex
970 
     movq   mm4,.cey   ; hi - lo -> cby; cey
 973 
     movq   mm4,.cey   ; hi - lo -> cby; cey
971 
;     movq   mm5,mm3
 974 
;     movq   mm5,mm3
972 
;     movq   mm6,mm4
 975 
;     movq   mm6,mm4
973 
;     psrad  mm5,ROUND
 976 
;     psrad  mm5,ROUND
974 
;     psrad  mm6,ROUND
 977 
;     psrad  mm6,ROUND
975 
;     movq   .ceyq,mm5
 978 
;     movq   .ceyq,mm5
976 
;     movq   .cbyq,mm6
 979 
;     movq   .cbyq,mm6
977 
     mov    edx,.czbuff
 980 
     mov    edx,.czbuff
978 
else
 981 
else
979 
     cld
 982 
     cld
980 
end if
 983 
end if
981 
     .draw:
 984 
     .draw:
982 
    ; if TEX = SHIFTING   ;bump drawing only in shifting mode
 985 
    ; if TEX = SHIFTING   ;bump drawing only in shifting mode
983 
if Ext=NON
 986 
if Ext=NON
984 
	mov	esi,.czbuff	 ; .czbuff current address in buffer
 987 
	mov	esi,.czbuff	 ; .czbuff current address in buffer
985 
	mov	ebx,.cz 	 ; .cz - cur z position
 988 
	mov	ebx,.cz 	 ; .cz - cur z position
986 
	cmp	ebx,dword[esi]
 989 
	cmp	ebx,dword[esi]
987 
else
 990 
else
988 
	mov	ebx,.cz
 991 
	mov	ebx,.cz
989 
	cmp	ebx,dword[edx]
 992 
	cmp	ebx,dword[edx]
990 
end if
 993 
end if
991 
	jge	.skip
 994 
	jge	.skip
992 
 
 995 
 
993 
if Ext=NON
 996 
if Ext=NON
994 
	mov	eax,.cby
 997 
	mov	eax,.cby
995 
	mov	esi,.cbx
 998 
	mov	esi,.cbx
996 
	sar	eax,ROUND
 999 
	sar	eax,ROUND
997 
	sar	esi,ROUND
 1000 
	sar	esi,ROUND
998 
	shl	eax,TEX_SHIFT	;-
 1001 
	shl	eax,TEX_SHIFT	;-
999 
	add	esi,eax
 1002 
	add	esi,eax
1000 
	lea	esi,[esi*3]	    ;-  ; esi - current b. texture addres
 1003 
	lea	esi,[esi*3]	    ;-  ; esi - current b. texture addres
1001 
	add	esi,.bmap
 1004 
	add	esi,.bmap
1002 
 
 1005 
 
1003 
	mov	ebx,.cex       ;.cex - current env map X
 1006 
	mov	ebx,.cex       ;.cex - current env map X
1004 
	mov	eax,.cey       ;.cey - current  env map y
 1007 
	mov	eax,.cey       ;.cey - current  env map y
1005 
	sar	ebx,ROUND
 1008 
	sar	ebx,ROUND
1006 
	sar	eax,ROUND
 1009 
	sar	eax,ROUND
1007 
 
 1010 
 
1008 
	shl	eax,TEX_SHIFT
 1011 
	shl	eax,TEX_SHIFT
1009 
	add	ebx,eax
 1012 
	add	ebx,eax
1010 
	lea	ebx,[ebx*3]
 1013 
	lea	ebx,[ebx*3]
1011 
	add	ebx,.emap
 1014 
	add	ebx,.emap
1012 
 
 1015 
 
1013 
 
 1016 
 
1014 
else
 1017 
else
1015 
	movq	mm5,mm4 ;.cey
 1018 
	movq	mm5,mm4 ;.cey
1016 
	psrad	mm5,ROUND
 1019 
	psrad	mm5,ROUND
1017 
	pslld	mm5,TEX_SHIFT
 1020 
	pslld	mm5,TEX_SHIFT
1018 
	movq	mm6,mm3 ;.cex
 1021 
	movq	mm6,mm3 ;.cex
1019 
	psrad	mm6,ROUND
 1022 
	psrad	mm6,ROUND
1020 
	paddd	mm5,mm6
 1023 
	paddd	mm5,mm6
1021 
	movq	mm6,mm5
 1024 
	movq	mm6,mm5
1022 
	paddd	mm5,mm5
 1025 
	paddd	mm5,mm5
1023 
	paddd	mm5,mm6
 1026 
	paddd	mm5,mm6
1024 
	paddd	mm5,.emap
 1027 
	paddd	mm5,.emap
1025 
	movd	esi,mm5
 1028 
	movd	esi,mm5
1026 
	psrlq	mm5,32
 1029 
	psrlq	mm5,32
1027 
	movd	ebx,mm5
 1030 
	movd	ebx,mm5
1028 
end if
 1031 
end if
1029 
if Ext>=MMX
 1032 
if Ext>=MMX
1030 
	movd	  mm1,[esi]
 1033 
	movd	  mm1,[esi]
1031 
	movd	  mm2,[ebx]
 1034 
	movd	  mm2,[ebx]
1032 
	punpcklbw mm1,mm0
 1035 
	punpcklbw mm1,mm0
1033 
	punpcklbw mm2,mm0
 1036 
	punpcklbw mm2,mm0
1034 
	pmullw	  mm1,mm2
 1037 
	pmullw	  mm1,mm2
1035 
	psrlw	  mm1,8
 1038 
	psrlw	  mm1,8
1036 
	packuswb  mm1,mm0
 1039 
	packuswb  mm1,mm0
1037 
	movd	  [edi],mm1
 1040 
	movd	  [edi],mm1
1038 
	mov	  ebx,.cz
 1041 
	mov	  ebx,.cz
1039 
	mov	  dword[edx],ebx
 1042 
	mov	  dword[edx],ebx
1040 
else
 1043 
else
1041 
	cld			; esi - tex e.
 1044 
	cld			; esi - tex e.
1042 
	lodsb			; ebx - tex b.
 1045 
	lodsb			; ebx - tex b.
1043 
	mov	dl,[ebx]
 1046 
	mov	dl,[ebx]
1044 
	mul	dl
 1047 
	mul	dl
1045 
	shr	ax,8
 1048 
	shr	ax,8
1046 
	stosb
 1049 
	stosb
1047 
	inc	ebx
 1050 
	inc	ebx
1048 
	lodsb
 1051 
	lodsb
1049 
	mov	dl,[ebx]
 1052 
	mov	dl,[ebx]
1050 
	mul	dl
 1053 
	mul	dl
1051 
	shr	ax,8
 1054 
	shr	ax,8
1052 
	stosb
 1055 
	stosb
1053 
	inc	ebx
 1056 
	inc	ebx
1054 
	lodsb
 1057 
	lodsb
1055 
	mov	dl,[ebx]
 1058 
	mov	dl,[ebx]
1056 
	mul	dl
 1059 
	mul	dl
1057 
	shr	ax,8
 1060 
	shr	ax,8
1058 
	stosb
 1061 
	stosb
1059 
	mov	ebx,.cz
 1062 
	mov	ebx,.cz
1060 
	mov	esi,.czbuff
 1063 
	mov	esi,.czbuff
1061 
	mov	dword[esi],ebx
 1064 
	mov	dword[esi],ebx
1062 
	jmp	.no_skip
 1065 
	jmp	.no_skip
1063 
end if
 1066 
end if
1064 
     .skip:
 1067 
     .skip:
1065 
	add	edi,3
 1068 
	add	edi,3
1066 
 
 1069 
 
1067 
   if Ext = NON
 1070 
   if Ext = NON
1068 
     .no_skip:
 1071 
     .no_skip:
1069 
	add	.czbuff,4
 1072 
	add	.czbuff,4
1070 
	mov	eax,.dbx
 1073 
	mov	eax,.dbx
1071 
	add	.cbx,eax
 1074 
	add	.cbx,eax
1072 
	mov	eax,.dby
 1075 
	mov	eax,.dby
1073 
	add	.cby,eax
 1076 
	add	.cby,eax
1074 
	mov	eax,.dex
 1077 
	mov	eax,.dex
1075 
	add	.cex,eax
 1078 
	add	.cex,eax
1076 
	mov	eax,.dey
 1079 
	mov	eax,.dey
1077 
	add	.cey,eax
 1080 
	add	.cey,eax
1078 
    else
 1081 
    else
1079 
	add	edx,4
 1082 
	add	edx,4
1080 
	paddd	mm3,.dex
 1083 
	paddd	mm3,.dex
1081 
	paddd	mm4,.dey
 1084 
	paddd	mm4,.dey
1082 
  ;      movq    mm5,mm3
 1085 
  ;      movq    mm5,mm3
1083 
  ;      movq    mm6,mm4
 1086 
  ;      movq    mm6,mm4
1084 
  ;      psrad   mm5,ROUND
 1087 
  ;      psrad   mm5,ROUND
1085 
  ;      psrad   mm6,ROUND
 1088 
  ;      psrad   mm6,ROUND
1086 
     ;   movq    .cex,mm3
 1089 
     ;   movq    .cex,mm3
1087 
     ;   movq    .cey,mm4
 1090 
     ;   movq    .cey,mm4
1088 
    end if
 1091 
    end if
1089 
	mov	eax,.dz
 1092 
	mov	eax,.dz
1090 
	add	.cz,eax
 1093 
	add	.cz,eax
1091 
    if Ext = NON
 1094 
    if Ext = NON
1092 
	dec	ecx
 1095 
	dec	ecx
1093 
	jnz	.draw
 1096 
	jnz	.draw
1094 
    else
 1097 
    else
1095 
	loop	.draw
 1098 
	loop	.draw
1096 
    end if
 1099 
    end if
1097 
 
 1100 
 
1098 
  .bl_end:
 1101 
  .bl_end:
1099 
	mov	esp,ebp
 1102 
	mov	esp,ebp
1100 
ret 56
 1103 
ret 56