WebSVN – Kolibri OS – Diff – /programs/demos/3DS/BUMP_TEX.INC

        push	edx   ;dword 0
        loop	@b
        jmp	.bt_dx12_done
  .bt_dx12_make:
        movsx	ebx,bx
        mov	ax,.x2
        sub	ax,.x1
        cwde
-       movsx	ebx,bx
        shl	eax,ROUND
        cdq
        idiv	ebx
  ;     mov      .dx12,eax
        push	 eax
-       mov     ax,.z2
-       sub     ax,.z1
-       cwde
-       shl     eax,CATMULL_SHIFT
-       cdq
-       idiv    ebx
-       push    eax
 if Ext>=SSE
-       sub	 esp,16
+       sub	 esp,28
    ;    mov       eax,256
        cvtsi2ss  xmm4,[i255d]
        cvtsi2ss  xmm3,ebx	     ;rcps
        divps	 xmm1,xmm3   ;xmm1--> | dby | dbx | dey | dex |
        shufps	 xmm1,xmm1,10110001b
 			     ;xmm1--> | dbx | dby | dex | dey |
-       cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
+       cvtps2pi  mm0,xmm1 ;mm0,xmm1          ; mm0 -> 2 delta dwords
        movhlps	 xmm1,xmm1
-       cvtps2pi  mm1,xmm1
+       cvtps2pi  mm1,xmm1 ;mm1,xmm1
        movq	 .dey12,mm0
        movq	 .dby12,mm1
+;-------------
+  ;    pxor      mm0,mm0
+  ;    pxor      mm1,mm1
+   ;/   pinsrw    mm0,.z1,1
+   ;/   pinsrw    mm0,.x1,0
+   ;/   pinsrw    mm1,.z2,1
+   ;/   pinsrw    mm1,.x2,0
+       mov	 ax,.z2
+       sub	 ax,.z1
+       cwde
+   ;/    movd      mm1,eax
+   ;/    punpcklwd  mm0,mm4
+   ;/    punpcklwd  mm1,mm4
+  ;     cvtpi2ps   xmm1,mm1
+  ;     cvtpi2ps   xmm2,mm0
+  ;     subps      xmm1,xmm2
+   ;/   psubd      mm1,mm0
+       movd	  mm2,[.t_x1]
+       movd	  mm3,[.t_x2]
+       punpcklwd  mm2,mm4
+       punpcklwd  mm3,mm4
+       psubd	  mm3,mm2
+   ;/  cvtpi2ps  xmm1,mm1
+       cvtsi2ss  xmm1,eax
+       movlhps	 xmm1,xmm1
+       cvtpi2ps  xmm1,mm3
+       divps	 xmm1,xmm3   ; xmm1--> | xxx | dz | dty | dtx |
+       shufps	 xmm1,xmm1,11100001b
+			     ; xmm1--> | xxx | dz | dtx | dty |
+       cvtps2pi  mm0,xmm1    ; mm0 -> 2 delta dwords  | dtx | dty |
+       movhlps	 xmm1,xmm1
+       cvtss2si  eax,xmm1    ; eax --> 1 delta dword | dz |
+       movq	 .dty12,mm0
+       mov	 .dz12,eax
-;-------------
+;----
 ;       mov       ax,.z2
 ;       sub       ax,.z1
 ;       cwde
 ;       mov       bx,.x2
 ;       sub       bx,.x1
 ;       movsx     ebx,bx
 ;       movd      mm1,eax
 ;       psllq     mm1,32
+;       movd      mm1,ebx
 ;       movd      mm1,ebx
 ;;       push      ebx
 ;;       push      eax
 ;;       movq      mm1,[esp]
 ;;       add       esp,8
 ;       cvtps2pi  mm1,xmm1    ; mm1 --> 2 delta dwords | dx | dz |
 ;       movq      .dty12,mm0
 ;       movq      .dz12,mm1
 else
        mov     ax,.z2
+       sub     ax,.z1
+       cwde
+       shl     eax,CATMULL_SHIFT
+       cdq
+       idiv    ebx
+       push    eax
        mov	ax,word[.b_x2]
        sub	ax,word[.b_x1]
        cwde
        shl	eax,ROUND
        idiv	ebx
  ;     mov      .dey12,eax
        push	 eax
-end if
+;end if
        mov	ax,word[.t_x2]
        sub	ax,word[.t_x1]
        cwde
        cdq
        idiv	ebx
  ;     mov      .dty12,eax
        push	 eax
 end if
    .bt_dx12_done:
        mov	bx,.y3	     ; calc delta13
        sub	bx,.y1
        jnz	.bt_dx13_make
        idiv	ebx
  ;     mov      .dx13,eax
        push	 eax
-       mov     ax,.z3
-       sub     ax,.z1
-       cwde
-       shl     eax,CATMULL_SHIFT
-       cdq
-       idiv    ebx
-  ;    mov    .dz13,eax
-       push    eax
 if Ext>=SSE
-       sub	 esp,16
+       sub	 esp,28
    ;    mov       eax,255
        cvtsi2ss  xmm4,[i255d]
        cvtsi2ss  xmm3,ebx	     ;rcps
        divss	 xmm3,xmm4
        movhlps	 xmm1,xmm1
        cvtps2pi  mm1,xmm1
        movq	 .dey13,mm0
        movq	 .dby13,mm1
-else
+;----------
+       mov	 ax,.z3
+       sub	 ax,.z1
+       cwde
+       movd	  mm2,[.t_x1]
+       movd	  mm3,[.t_x3]
+       punpcklwd  mm2,mm4
+       punpcklwd  mm3,mm4
+       psubd	  mm3,mm2
+       cvtsi2ss  xmm1,eax
+       movlhps	 xmm1,xmm1
+       cvtpi2ps  xmm1,mm3
+       divps	 xmm1,xmm3   ; xmm1--> | xxx | dz | dty | dtx |
+       shufps	 xmm1,xmm1,11100001b
+			     ; xmm1--> | xxx | dz | dtx | dty |
+       cvtps2pi  mm0,xmm1    ; mm0 -> 2 delta dwords  | dtx | dty |
+       movhlps	 xmm1,xmm1
+       cvtss2si  eax,xmm1    ; eax --> 1 delta dword | dz |
+       movq	 .dty13,mm0
+       mov	 .dz13,eax
+else
-       mov	ax,word[.b_x3]
+       mov     ax,.z3
+       sub     ax,.z1
+       cwde
+       shl     eax,CATMULL_SHIFT
+       cdq
+       idiv    ebx
+  ;    mov    .dz13,eax
+       push    eax
+       mov	ax,word[.b_x3]
        sub	ax,word[.b_x1]
        cwde
        shl	eax,ROUND
        cdq
        idiv	ebx
        cdq
        idiv	ebx
  ;     mov      .dey13,eax
        push	 eax
 end if
        mov	ax,word[.t_x3]
        sub	ax,word[.t_x1]
        cwde
        shl	eax,ROUND
        cdq
        idiv	ebx
  ;     mov      .dty13,eax
        push	 eax
 end if
    .bt_dx13_done:
        mov	bx,.y3	     ; calc delta23
        sub	bx,.y2
        jnz	.bt_dx23_make
        idiv	ebx
  ;     mov      .dx23,eax
        push	 eax
-       mov     ax,.z3
-       sub     ax,.z2
-       cwde
-       shl     eax,CATMULL_SHIFT
-       cdq
-       idiv    ebx
-     ; mov     .dz23,eax
-       push    eax
 if Ext>=SSE
-       sub	 esp,16
+       sub	 esp,28
     ;   mov       eax,255
        cvtsi2ss  xmm4,[i255d]
        cvtsi2ss  xmm3,ebx	     ;rcps
        divss	 xmm3,xmm4
        movhlps	 xmm1,xmm1
        cvtps2pi  mm1,xmm1
        movq	 .dey23,mm0
        movq	 .dby23,mm1
-else
+;----------
+       mov	 ax,.z3
+       sub	 ax,.z2
+       cwde
+       movd	  mm2,[.t_x2]
+       movd	  mm3,[.t_x3]
+       punpcklwd  mm2,mm4
+       punpcklwd  mm3,mm4
+       psubd	  mm3,mm2
+       cvtsi2ss  xmm1,eax
+       movlhps	 xmm1,xmm1
+       cvtpi2ps  xmm1,mm3
+       divps	 xmm1,xmm3   ; xmm1--> | xxx | dz | dty | dtx |
+       shufps	 xmm1,xmm1,11100001b
+			     ; xmm1--> | xxx | dz | dtx | dty |
+       cvtps2pi  mm0,xmm1    ; mm0 -> 2 delta dwords  | dtx | dty |
+       movhlps	 xmm1,xmm1
+       cvtss2si  eax,xmm1    ; eax --> 1 delta dword | dz |
+       movq	 .dty23,mm0
+       mov	 .dz23,eax
+else
        mov	ax,word[.b_x3]
+       mov     ax,.z3
+       sub     ax,.z2
+       cwde
+       shl     eax,CATMULL_SHIFT
+       cdq
+       idiv    ebx
+     ; mov     .dz23,eax
+       push    eax
+       mov	ax,word[.b_x3]
        sub	ax,word[.b_x2]
        cwde
        shl	eax,ROUND
        cdq
        idiv	ebx
        cdq
        idiv	ebx
  ;     mov      .dey23,eax
        push	 eax
 end if
        mov	ax,word[.t_x3]
        sub	ax,word[.t_x2]
        cwde
        shl	eax,ROUND
        cdq
        idiv	ebx
  ;     mov      .dty23,eax
        push	 eax
 end if
       ;  sub     esp,40
    .bt_dx23_done:
        sub	 esp,64
        movsx	eax,.x1
        shl	eax,ROUND
        mov	.cx1,eax

Subversion Repositories Kolibri OS

(root)/programs/demos/3DS/BUMP_TEX.INC – Rev 1245 → 1776