diff --git a/programs/demos/3DS/BUMP3.INC b/programs/demos/3DS/BUMP3.INC
index 4db1fed748..c86f068c64 100644
--- a/programs/demos/3DS/BUMP3.INC
+++ b/programs/demos/3DS/BUMP3.INC
@@ -436,13 +436,13 @@ bump_line:
 ;--------------    edi - pointer to screen buffer
 ;stack - another parameters :
 .y    equ dword [ebp+4]
-.bx1  equ dword [ebp+8]   ;   ---
+.bx1  equ	[ebp+8]   ;   ---
 .by1  equ dword [ebp+12]  ;       |
-.bx2  equ dword [ebp+16]  ;       |
+.bx2  equ	[ebp+16]  ;       |
 .by2  equ dword [ebp+20]  ;       |>   bump and env coords
-.ex1  equ dword [ebp+24]  ;       |>   shifted shl ROUND
+.ex1  equ	[ebp+24]  ;       |>   shifted shl ROUND
 .ey1  equ dword [ebp+28]  ;       |
-.ex2  equ dword [ebp+32]  ;       |
+.ex2  equ	[ebp+32]  ;       |
 .ey2  equ dword [ebp+36]  ;   ---
 .bmap equ dword [ebp+40]
 .emap equ dword [ebp+44]
@@ -450,13 +450,13 @@ bump_line:
 .x1   equ dword [ebp-4]
 .x2   equ dword [ebp-8]
 .dbx  equ dword [ebp-12]
-.dby  equ dword [ebp-16]
+.dby  equ	[ebp-16]
 .dex  equ dword [ebp-20]
-.dey  equ dword [ebp-24]
+.dey  equ	[ebp-24]
 .cbx  equ dword [ebp-28]
-.cby  equ dword [ebp-32]
+.cby  equ	[ebp-32]
 .cex  equ dword [ebp-36]
-.cey  equ dword [ebp-40]
+.cey  equ	[ebp-40]
 	mov	ebp,esp
 
 	mov	ecx,.y
@@ -470,7 +470,7 @@ bump_line:
 	je	.bl_end
 
 	xchg	eax,ebx
-
+if Ext = NON
 	mov	edx,.bx1
 	xchg	edx,.bx2
 	mov	.bx1,edx
@@ -484,6 +484,16 @@ bump_line:
 	mov	edx,.ey1
 	xchg	edx,.ey2
 	mov	.ey1,edx
+else
+	movq	mm0,.bx1
+	movq	mm1,.bx2
+	movq	mm2,.ex1
+	movq	mm3,.ex2
+	movq	.bx2,mm0
+	movq	.bx1,mm1
+	movq	.ex1,mm3
+	movq	.ex2,mm2
+end if
   .bl_ok:
 	push	eax
 	push	ebx	      ;store x1, x2
@@ -523,9 +533,9 @@ bump_line:
 	idiv	ecx
 	push	eax
 
-	push	.bx1
+	push	dword .bx1
 	push	.by1
-	push	.ex1
+	push	dword .ex1
 	push	.ey1
      .draw:
     ; if TEX = SHIFTING   ;bump drawing only in shifting mode
@@ -591,7 +601,14 @@ bump_line:
      .put_pixel:
 	stosd
 	dec	edi
-
+;if Ext >= MMX
+;        movq    mm0,.cby
+;        movq    mm1,.cey
+;        paddd   mm0,.dby
+;        paddd   mm1,.dey
+;        movq    .cby,mm0
+;        movq    .cey,mm1
+;else
 	mov	eax,.dbx
 	add	.cbx,eax
 	mov	eax,.dby
@@ -600,7 +617,7 @@ bump_line:
 	add	.cex,eax
 	mov	eax,.dey
 	add	.cey,eax
-
+;end if
 	dec	ecx
 	jnz	.draw
    ;   end if
diff --git a/programs/demos/3DS/BUMP_CAT.INC b/programs/demos/3DS/BUMP_CAT.INC
index aa0eb34ce6..0ab609f52e 100644
--- a/programs/demos/3DS/BUMP_CAT.INC
+++ b/programs/demos/3DS/BUMP_CAT.INC
@@ -710,20 +710,14 @@ bump_line_z:
 ;--------------    edi - pointer to screen buffer
 ;stack - another parameters :
 .y	equ dword [ebp+4]
-.bx1q	equ	  [ebp+8]
-.bx2q	equ	  [ebp+16]
-.ex1q	equ	  [ebp+24]
-.ex2q	equ	  [ebp+32]
-.bx1	equ dword [ebp+8]   ;   ---
+.bx1	equ	  [ebp+8]   ;   ---
 .by1	equ dword [ebp+12]  ;       |
-.bx2	equ dword [ebp+16]  ;       |
+.bx2	equ	  [ebp+16]  ;       |
 .by2	equ dword [ebp+20]  ;       |>   bump and env coords
-.ex1	equ dword [ebp+24]  ;       |>   shifted shl ROUND
+.ex1	equ	  [ebp+24]  ;       |>   shifted shl ROUND
 .ey1	equ dword [ebp+28]  ;       |
-.ex2	equ dword [ebp+32]  ;       |
+.ex2	equ	  [ebp+32]  ;       |
 .ey2	equ dword [ebp+36]  ;   ---
-;.bx1q   equ qword [ebp+8]   ; - new
-;.ex1q   equ qword [ebp+24]  ; - new
 .bmap	equ dword [ebp+40]
 .emap	equ dword [ebp+44]
 .z_buff equ dword [ebp+48]
@@ -733,18 +727,14 @@ bump_line_z:
 .x1	equ dword [ebp-4]
 .x2	equ dword [ebp-8]
 .dbx	equ dword [ebp-12]
-.dby	equ dword [ebp-16]
-.dbyq	equ qword [ebp-16]  ; - new
+.dby	equ	  [ebp-16]
 .dex	equ dword [ebp-20]
-.dey	equ dword [ebp-24]
-.deyq	equ qword [ebp-24]  ; - new
+.dey	equ	  [ebp-24]
 .dz	equ dword [ebp-28]
 .cbx	equ dword [ebp-32]
-.cby	equ dword [ebp-36]
-.cbyq	equ qword [ebp-36]  ; - new
+.cby	equ	  [ebp-36]
 .cex	equ dword [ebp-40]
-.cey	equ dword [ebp-44]
-.ceyq	equ qword [ebp-44]  ; - new
+.cey	equ	  [ebp-44]
 .cz	equ dword [ebp-48]
 .czbuff equ dword [ebp-52]
 .temp1	equ	   ebp-60
@@ -782,14 +772,14 @@ if Ext=NON
 	mov	.ey1,edx
 else
 
-	movq	mm0,.bx1q
-	movq	mm1,.ex1q
-	movq	mm2,.bx2q
-	movq	mm3,.ex2q
-	movq	.bx2q,mm0
-	movq	.ex2q,mm1
-	movq	.bx1q,mm2
-	movq	.ex1q,mm3
+	movq	mm0,.bx1
+	movq	mm1,.ex1
+	movq	mm2,.bx2
+	movq	mm3,.ex2
+	movq	.bx2,mm0
+	movq	.ex2,mm1
+	movq	.bx1,mm2
+	movq	.ex1,mm3
 
 end if
 
@@ -816,12 +806,12 @@ if Ext >= SSE
        cvtsi2ss  xmm3,ebx	     ;rcps
        shufps	 xmm3,xmm3,0
 
-       cvtpi2ps  xmm0,.bx1q ;mm0
+       cvtpi2ps  xmm0,.bx1 ;mm0
        movlhps	 xmm0,xmm0
-       cvtpi2ps  xmm0,.ex1q ;mm2
-       cvtpi2ps  xmm1,.bx2q ;mm1
+       cvtpi2ps  xmm0,.ex1 ;mm2
+       cvtpi2ps  xmm1,.bx2 ;mm1
        movlhps	 xmm1,xmm1
-       cvtpi2ps  xmm1,.ex2q ;mm3
+       cvtpi2ps  xmm1,.ex2 ;mm3
        subps	 xmm1,xmm0
 
        divps	 xmm1,xmm3
@@ -830,8 +820,8 @@ if Ext >= SSE
        cvtps2pi  mm0,xmm1	   ; mm0 -> 2 delta dwords
        movhlps	 xmm1,xmm1
        cvtps2pi  mm1,xmm1
-       movq	 .deyq,mm0
-       movq	 .dbyq,mm1
+       movq	 .dey,mm0
+       movq	 .dby,mm1
 
 else
 
@@ -911,9 +901,9 @@ end if
 	mov	ecx,.x2
 	sub	ecx,.x1
 	; init current variables
-	push	.bx1
+	push	dword .bx1
 	push	.by1
-	push	.ex1
+	push	dword .ex1
 	push	.ey1
 
 	push	.z1		 ; current z shl CATMULL_SHIFT
@@ -1072,6 +1062,14 @@ end if
      .no_skip:
 	add	.czbuff,4
 ;;   if Ext = NON
+if Ext >= MMX
+	movq	mm0,.cby
+	movq	mm1,.cey
+	paddd	mm0,.dby
+	paddd	mm1,.dey
+	movq	.cby,mm0
+	movq	.cey,mm1
+else
 	mov	eax,.dbx
 	add	.cbx,eax
 	mov	eax,.dby
@@ -1080,6 +1078,7 @@ end if
 	add	.cex,eax
 	mov	eax,.dey
 	add	.cey,eax
+end if
 ;;    else
 ;;        paddd   mm0,mm2
 ;;        paddd   mm1,mm3
diff --git a/programs/demos/3DS/BUMP_TEX.INC b/programs/demos/3DS/BUMP_TEX.INC
index 3d80ff5574..7cfefeae72 100644
--- a/programs/demos/3DS/BUMP_TEX.INC
+++ b/programs/demos/3DS/BUMP_TEX.INC
@@ -178,28 +178,20 @@ bump_tex_triangle_z:
        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
@@ -228,12 +220,54 @@ if Ext>=SSE
 
        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
@@ -243,6 +277,7 @@ if Ext>=SSE
 ;       movd      mm1,eax
 ;       psllq     mm1,32
 ;       movd      mm1,ebx
+
 ;;       push      ebx
 ;;       push      eax
 ;;       movq      mm1,[esp]
@@ -282,6 +317,13 @@ if Ext>=SSE
 ;       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]
@@ -319,7 +361,7 @@ else
  ;     mov      .dey12,eax
        push	 eax
 
-end if
+;end if
 
        mov	ax,word[.t_x2]
        sub	ax,word[.t_x1]
@@ -338,7 +380,7 @@ end if
        idiv	ebx
  ;     mov      .dty12,eax
        push	 eax
-
+end if
    .bt_dx12_done:
 
        mov	bx,.y3	     ; calc delta13
@@ -361,18 +403,9 @@ end if
  ;     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
@@ -406,7 +439,42 @@ if Ext>=SSE
        cvtps2pi  mm1,xmm1
        movq	 .dey13,mm0
        movq	 .dby13,mm1
+;----------
+       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,.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
@@ -442,7 +510,6 @@ else
        idiv	ebx
  ;     mov      .dey13,eax
        push	 eax
-end if
 
        mov	ax,word[.t_x3]
        sub	ax,word[.t_x1]
@@ -461,7 +528,7 @@ end if
        idiv	ebx
  ;     mov      .dty13,eax
        push	 eax
-
+end if
    .bt_dx13_done:
 
        mov	bx,.y3	     ; calc delta23
@@ -484,18 +551,9 @@ end if
  ;     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
@@ -529,7 +587,42 @@ if Ext>=SSE
        cvtps2pi  mm1,xmm1
        movq	 .dey23,mm0
        movq	 .dby23,mm1
+;----------
+       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,.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
@@ -565,7 +658,7 @@ else
        idiv	ebx
  ;     mov      .dey23,eax
        push	 eax
-end if
+
 
        mov	ax,word[.t_x3]
        sub	ax,word[.t_x2]
@@ -584,7 +677,7 @@ end if
        idiv	ebx
  ;     mov      .dty23,eax
        push	 eax
-
+end if
       ;  sub     esp,40
    .bt_dx23_done:
        sub	 esp,64
diff --git a/programs/demos/3DS/GRD3.INC b/programs/demos/3DS/GRD3.INC
index ade6e1be47..1cbe2d00ac 100644
--- a/programs/demos/3DS/GRD3.INC
+++ b/programs/demos/3DS/GRD3.INC
@@ -22,27 +22,25 @@ gouraud_triangle:
 .x3    equ word[ebp-10]
 .y3    equ word[ebp-12]
 
-.dc12r equ dword[ebp-16]
-.dc12g equ dword[ebp-20]
-.dc12b equ dword[ebp-24]
-.dc13r equ dword[ebp-28]
-.dc13g equ dword[ebp-32]
-.dc13b equ dword[ebp-36]
-.dc23r equ dword[ebp-40]
-.dc23g equ dword[ebp-44]
-.dc23b equ dword[ebp-48]
-
-.c1r   equ dword[ebp-52]
-.c1g   equ dword[ebp-56]
-.c1b   equ dword[ebp-60]
-.c2r   equ dword[ebp-64]
-.c2g   equ dword[ebp-68]
-.c2b   equ dword[ebp-72]
-
-.dx12  equ dword[ebp-76]
-.dx13  equ dword[ebp-80]
-.dx23  equ dword[ebp-84]
+.dx12  equ dword[ebp-16]
+.dc12r equ dword[ebp-20]
+.dc12g equ [ebp-24]
+.dc12b equ dword[ebp-28]
+.dx13  equ dword[ebp-32]
+.dc13r equ dword[ebp-36]
+.dc13g equ [ebp-40]
+.dc13b equ dword[ebp-44]
+.dx23  equ dword[ebp-48]
+.dc23r equ dword[ebp-52]
+.dc23g equ [ebp-56]
+.dc23b equ dword[ebp-60]
 
+.c1r   equ dword[ebp-64]
+.c1g   equ [ebp-68]
+.c1b   equ dword[ebp-72]
+.c2r   equ dword[ebp-76]
+.c2g   equ [ebp-80]
+.c2b   equ dword[ebp-84]
 
 
        mov ebp,esp
@@ -73,7 +71,7 @@ gouraud_triangle:
        push eax   ;store triangle coordinates in user friendly variables
        push ebx
        push ecx
-       sub esp,72 ; set correctly value of esp
+    ;   sub esp,72 ; set correctly value of esp
 
        mov edx,eax    ; check only X triangle coordinate
        or edx,ebx
@@ -94,10 +92,15 @@ gouraud_triangle:
        mov bx,.y2	; calc deltas
        sub bx,.y1
        jnz .gt_dx12_make
-       mov .dx12,0
-       mov .dc12r,0
-       mov .dc12g,0
-       mov .dc12b,0
+       xor edx,edx
+       mov ecx,4
+     @@:
+       push edx
+       loop @b
+     ;  mov .dx12,0
+     ;  mov .dc12r,0
+     ;  mov .dc12g,0
+     ;  mov .dc12b,0
        jmp .gt_dx12_done
   .gt_dx12_make:
 
@@ -108,7 +111,8 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dx12,eax
+     ;  mov .dx12,eax
+       push eax
 
        mov ax,word[.col2r]
        sub ax,word[.col1r]
@@ -116,30 +120,39 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc12r,eax
+      ; mov .dc12r,eax
+       push eax
+
        mov ax,word[.col2g]
        sub ax,word[.col1g]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc12g,eax
+     ;  mov .dc12g,eax
+       push eax
+
        mov ax,word[.col2b]
        sub ax,word[.col1b]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc12b,eax
+     ;  mov .dc12b,eax
+       push eax
 .gt_dx12_done:
-
        mov bx,.y3
        sub bx,.y1
        jnz .gt_dx13_make
-       mov .dx13,0
-       mov .dc13r,0
-       mov .dc13g,0
-       mov .dc13b,0
+       xor edx,edx
+       mov ecx,4
+     @@:
+       push edx
+       loop @b
+   ;    mov .dx13,0
+   ;    mov .dc13r,0
+   ;    mov .dc13g,0
+   ;    mov .dc13b,0
        jmp .gt_dx13_done
 .gt_dx13_make:
        mov ax,.x3
@@ -149,7 +162,8 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dx13,eax
+   ;    mov .dx13,eax
+       push eax
 
        mov ax,word[.col3r]
        sub ax,word[.col1r]
@@ -157,30 +171,38 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc13r,eax
+    ;   mov .dc13r,eax
+       push eax
+
        mov ax,word[.col3g]
        sub ax,word[.col1g]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc13g,eax
+     ;  mov .dc13g,eax
+       push eax
        mov ax,word[.col3b]
        sub ax,word[.col1b]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc13b,eax
+    ;   mov .dc13b,eax
+       push eax
 .gt_dx13_done:
-
        mov bx,.y3
        sub bx,.y2
        jnz .gt_dx23_make
-       mov .dx23,0
-       mov .dc23r,0
-       mov .dc23g,0
-       mov .dc23b,0
+       xor edx,edx
+       mov ecx,4
+     @@:
+       push edx
+       loop @b
+   ;    mov .dx23,0
+   ;    mov .dc23r,0
+   ;    mov .dc23g,0
+   ;    mov .dc23b,0
        jmp .gt_dx23_done
 .gt_dx23_make:
        mov ax,.x3
@@ -190,7 +212,8 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dx23,eax
+     ; mov .dx23,eax
+       push eax
 
        mov ax,word[.col3r]
        sub ax,word[.col2r]
@@ -198,23 +221,30 @@ gouraud_triangle:
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc23r,eax
+      ; mov .dc23r,eax
+       push eax
+
+
        mov ax,word[.col3g]
        sub ax,word[.col2g]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc23g,eax
+     ;  mov .dc23g,eax
+       push eax
+
        mov ax,word[.col3b]
        sub ax,word[.col2b]
        cwde
        shl eax,ROUND
        cdq
        idiv ebx
-       mov .dc23b,eax
-.gt_dx23_done:
+    ;   mov .dc23b,eax
+       push eax
 
+.gt_dx23_done:
+       sub esp,24
        movsx eax,.x1
        shl eax,ROUND
        mov ebx,eax
@@ -240,6 +270,8 @@ gouraud_triangle:
        push edi
        push ebp
 
+       sar ebx,ROUND
+       push bx
        mov edx,.c2r		 ; c2r,c2g,c2b,c1r,c1g,c1b - current colors
        sar edx,ROUND
        push dx
@@ -249,6 +281,9 @@ gouraud_triangle:
        mov edx,.c2b
        sar edx,ROUND
        push dx
+
+       sar eax,ROUND
+       push ax
        mov edx,.c1r
        sar edx,ROUND
        push dx
@@ -259,10 +294,6 @@ gouraud_triangle:
        sar edx,ROUND
        push dx
        push cx
-       sar ebx,ROUND
-       push bx
-       sar eax,ROUND
-       push ax
        call gouraud_line
 
        pop ebp
@@ -271,16 +302,28 @@ gouraud_triangle:
        pop ebx
        pop eax
 
+if Ext >= MMX
+       movq mm0,.c1g
+       paddd mm0,.dc13g
+       movq .c1g,mm0
+else
        mov edx,.dc13r
        add .c1r,edx
        mov edx,.dc13g
        add .c1g,edx
+end if
        mov edx,.dc13b
        add .c1b,edx
+if Ext >= MMX
+       movq mm0,.c2g
+       paddd mm0,.dc12g
+       movq .c2g,mm0
+else
        mov edx,.dc12r
        add .c2r,edx
        mov edx,.dc12g
        add .c2g,edx
+end if
        mov edx,.dc12b
        add .c2b,edx
 
@@ -313,6 +356,8 @@ gouraud_triangle:
        push edi
        push ebp
 
+       sar ebx,ROUND
+       push bx
        mov edx,.c2r
        sar edx,ROUND
        push dx
@@ -322,6 +367,9 @@ gouraud_triangle:
        mov edx,.c2b
        sar edx,ROUND
        push dx
+
+       sar eax,ROUND
+       push ax
        mov edx,.c1r
        sar edx,ROUND
        push dx
@@ -332,10 +380,6 @@ gouraud_triangle:
        sar edx,ROUND
        push dx
        push cx
-       sar ebx,ROUND
-       push bx
-       sar eax,ROUND
-       push ax
        call gouraud_line
 
        pop ebp
@@ -344,16 +388,28 @@ gouraud_triangle:
        pop ebx
        pop eax
 
+if Ext >= MMX
+       movq mm0,.c1g
+       paddd mm0,.dc13g
+       movq .c1g,mm0
+else
        mov edx,.dc13r
        add .c1r,edx
        mov edx,.dc13g
        add .c1g,edx
+end if
        mov edx,.dc13b
        add .c1b,edx
+if Ext >= MMX
+       movq mm0,.c2g
+       paddd mm0,.dc23g
+       movq .c2g,mm0
+else
        mov edx,.dc23r
        add .c2r,edx
        mov edx,.dc23g
        add .c2g,edx
+end if
        mov edx,.dc23b
        add .c2b,edx
 
@@ -370,15 +426,15 @@ ret 18
 gouraud_line:
 ;-------------in - edi - pointer to screen buffer
 ;----------------- stack - another parameters
-.x1 equ word [ebp+4]
-.x2 equ word [ebp+6]
-.y equ word [ebp+8]
-.col1b equ ebp+10
-.col1g equ ebp+12
-.col1r equ ebp+14
-.col2b equ ebp+16
-.col2g equ ebp+18
-.col2r equ ebp+20
+.y equ word [ebp+4]
+.col1b equ ebp+6
+.col1g equ ebp+8
+.col1r equ ebp+10
+.x1    equ [ebp+12]
+.col2b equ ebp+14
+.col2g equ ebp+16
+.col2r equ ebp+18
+.x2    equ [ebp+20]
 .dc_r equ dword[ebp-4]
 .dc_g equ dword[ebp-8]
 .dc_b equ dword[ebp-12]
@@ -395,14 +451,19 @@ gouraud_line:
        je .gl_quit
        jl .gl_ok
 
-       xchg ax,.x2
-       mov .x1,ax
-       mov eax,dword[.col1b]
-       xchg eax,dword[.col2b]
-       mov  dword[.col1b],eax
-       mov ax,word[.col1r]
-       xchg ax,word[.col2r]
-       mov word[.col1r],ax
+if Ext >= MMX
+       movq mm0,[.col1b]
+       movq mm1,[.col2b]
+       movq [.col1b],mm1
+       movq [.col2b],mm0
+else
+       mov eax,[.col1b]
+       xchg eax,[.col2b]
+       mov  [.col1b],eax
+       mov eax,[.col1r]
+       xchg eax,[.col2r]
+       mov [.col1r],eax
+end if
 .gl_ok:
   ;     cmp .x1,SIZE_X-1  ;check
   ;     jg .gl_quit
@@ -420,7 +481,7 @@ gouraud_line:
        movsx ecx,.y
        mov eax,SIZE_X*3
        mul ecx
-       movsx ebx,.x1
+       movsx ebx,word .x1
        lea ecx,[ebx*2+eax]
        add edi,ecx
        add edi,ebx
@@ -478,4 +539,152 @@ gouraud_line:
 .gl_quit:
       ; add esp,12
        mov esp,ebp
-ret 18
\ No newline at end of file
+ret 18
+if 0
+gouraud_line_SSE:	 ; new
+;-------------in - edi - pointer to screen buffer
+;----------------- stack - another parameters
+.y equ word [ebp+4]
+.col1b equ ebp+6
+.col1g equ ebp+8
+.col1r equ ebp+10
+.x1    equ [ebp+12]
+.col2b equ ebp+14
+.col2g equ ebp+16
+.col2r equ ebp+18
+.x2    equ [ebp+20]
+.dc_r equ dword[ebp-4]
+.dc_g equ dword[ebp-8]
+.dc_b equ dword[ebp-12]
+.lenght equ [ebp-16]
+.factor equ [ebp-24]   ;new
+       mov ebp,esp
+
+       mov ax,.y
+       or ax,ax
+       jl .gl_quit
+       cmp ax,SIZE_Y-1
+       jg .gl_quit
+
+       mov ax,.x1
+       cmp ax,.x2
+       je .gl_quit
+       jl .gl_ok
+
+if Ext >= MMX
+       movq mm0,[.col1b]
+       movq mm1,[.col2b]
+       movq [.col1b],mm1
+       movq [.col2b],mm0
+else
+       mov eax,[.col1b]
+       xchg eax,[.col2b]
+       mov  [.col1b],eax
+       mov eax,[.col1r]
+       xchg eax,[.col2r]
+       mov [.col1r],eax
+end if
+.gl_ok:
+  ;     cmp .x1,SIZE_X-1  ;check
+  ;     jg .gl_quit
+  ;     cmp .x2,SIZE_X-1
+  ;     jl @f
+  ;     mov .x2,SIZE_X-1
+  ;  @@:
+  ;     cmp .x1,0
+  ;     jg @f
+  ;     mov .x1,0
+  ;  @@:
+  ;     cmp .x2,0
+  ;     jl .gl_quit
+
+       movsx ecx,.y
+       mov eax,SIZE_X*3
+       mul ecx
+       movsx ebx,word .x1
+       lea ecx,[ebx*2+eax]
+       add edi,ecx
+       add edi,ebx
+
+       mov ax,word[.col2r]
+       sub ax,word[.col1r]
+       cwde
+       shl eax,ROUND
+       cdq
+       mov cx,.x2
+       sub cx,.x1
+       movsx ecx,cx
+       idiv ecx
+       ;mov .dc_r,eax           ;first delta
+       push eax
+
+       mov ax,word[.col2g]
+       sub ax,word[.col1g]
+       cwde
+       shl eax,ROUND
+       cdq
+       idiv ecx
+       ;mov .dc_g,eax
+       push eax
+
+       mov ax,word[.col2b]
+       sub ax,word[.col1b]
+       cwde
+       shl eax,ROUND
+       cdq
+       idiv ecx
+      ; mov .dc_b,eax
+       push eax
+
+       movsx ebx,word[.col1r]
+       shl ebx,ROUND
+       movsx edx,word[.col1g]
+       shl edx,ROUND
+       movsx esi,word[.col1b]
+       shl esi,ROUND
+
+       push	 ecx	     ; store line lenght
+       movd	 mm3,.dc_r
+       psrlq	 mm3,16      ; load dr to lowest word of mm3
+       pxor	 mm2,mm2     ; clear mm2
+       movd	 mm4,.dc_g
+       punpcklwd mm3,mm3     ; unpack dr to lower 2 words in in mm3
+       psrlq	 mm4,16      ; load dg to lowest word of mm4
+       movd	 mm5,.dc_b
+       psrlq	 mm5,16      ; load db to lowest word of mm5
+       punpcklwd mm4,mm4     ; unpack dg to lower 2 words in in mm3
+       lea	 ecx,[factor]
+       punpckldq mm3,mm3
+       punpcklwd mm5,mm5     ; unpack db to lower 2 words in in mm5
+       movq	 mm6,[.col1b]
+       xor	 eax,eax
+       pinsrw	 mm6,eax,3     ; clear the highest word in mm6
+       mov	 eax,010000h
+       punpckldq mm4,mm4     ; unpack dg to 4 words in mm4
+       mov	 [ecx],eax
+       mov	 eax,030002h
+       punpckldq mm5,mm5    ; unpack db to 4 words in mm5
+       movq	 mm7,mm6    ; load r1r1,g1g1,b1b1 to the first three
+			    ; words of mm7
+       pxor	 mm1,mm1    ; clear mm1
+
+
+.gl_draw:
+       mov eax,ebx
+       sar eax,ROUND
+       stosb
+       mov eax,edx
+       sar eax,ROUND
+       stosb
+       mov eax,esi
+       sar eax,ROUND
+       stosb
+       add ebx,.dc_r
+       add edx,.dc_g
+       add esi,.dc_b
+       loop .gl_draw
+.gl_quit:
+      ; add esp,12
+       mov esp,ebp
+ret 18
+end if
diff --git a/programs/demos/3DS/GRD_TEX.INC b/programs/demos/3DS/GRD_TEX.INC
index 191c2016b0..4577110c32 100644
--- a/programs/demos/3DS/GRD_TEX.INC
+++ b/programs/demos/3DS/GRD_TEX.INC
@@ -65,46 +65,44 @@ tex_plus_grd_triangle:
 
    .dx12      equ  dword[ebp-28] ;dd ?
    .tex_dx12  equ  dword[ebp-32] ;dd ?
-   .tex_dy12  equ  dword[ebp-36] ;dd ?
+   .tex_dy12  equ	[ebp-36] ;dd ?
    .dz12      equ  dword[ebp-40] ;dd ?
-   .dc12r     equ  dword[ebp-44] ;dd ?
+   .dc12r     equ	[ebp-44] ;dd ?
    .dc12g     equ  dword[ebp-48] ;dd ?
-   .dc12b     equ  dword[ebp-52] ;dd ?
+   .dc12b     equ	[ebp-52] ;dd ?
 
    .dx23      equ  dword[ebp-56] ;dd ?
    .tex_dx23  equ  dword[ebp-60] ;dd ?
-   .tex_dy23  equ  dword[ebp-64] ;dd ?
+   .tex_dy23  equ	[ebp-64] ;dd ?
    .dz23      equ  dword[ebp-68] ;dd ?
-   .dc23r     equ  dword[ebp-72] ;dd ?
+   .dc23r     equ	[ebp-72] ;dd ?
    .dc23g     equ  dword[ebp-76] ;dd ?
-   .dc23b     equ  dword[ebp-80] ;dword[ebp-8]dd ?
+   .dc23b     equ	[ebp-80] ;dword[ebp-8]dd ?
 
    .dx13      equ  dword[ebp-84] ;dd ?
    .tex_dx13  equ  dword[ebp-88] ;dd ?
-   .tex_dy13  equ  dword[ebp-92] ;dd ?
+   .tex_dy13  equ	[ebp-92] ;dd ?
    .dz13      equ  dword[ebp-96] ;dd ?
-   .dc13r     equ  dword[ebp-100] ;dd ?
+   .dc13r     equ	[ebp-100] ;dd ?
    .dc13g     equ  dword[ebp-104] ;dd ?
-   .dc13b     equ  dword[ebp-108] ;dd ?
+   .dc13b     equ	[ebp-108] ;dd ?
 
-   .zz1       equ  dword[ebp-112] ;dw ?
-   .zz2       equ  dword[ebp-116] ;dw ?
-   .cur1r     equ  dword[ebp-120] ;dw ?
-   .cur1g     equ  dword[ebp-124] ;dw ?
-   .cur1b     equ  dword[ebp-128] ;dw ?
-   .cur2r     equ  dword[ebp-132] ;dw ?
-   .cur2g     equ  dword[ebp-136] ;dw ?
-   .cur2b     equ  dword[ebp-140] ;dw ?
-   .scan_x1   equ  dword[ebp-144] ;dd ?
-   .scan_x2   equ  dword[ebp-148] ;dd ?
-   .scan_y1   equ  dword[ebp-152] ;dd ?
-   .scan_y2   equ  dword[ebp-156] ;dd ?
+   .scan_x1   equ  dword[ebp-112] ;dd ?
+   .scan_y1   equ	[ebp-116] ;dd ?
+   .zz1       equ  dword[ebp-120] ;dw ?
+   .cur1r     equ	[ebp-124] ;dw ?
+   .cur1g     equ  dword[ebp-128] ;dw ?
+   .cur1b     equ	[ebp-132] ;dw ?
+
+   .scan_x2   equ  dword[ebp-136] ;dd ?
+   .scan_y2   equ	[ebp-140] ;dd ?
+   .zz2       equ  dword[ebp-144] ;dw ?
+   .cur2r     equ	[ebp-148] ;dw ?
+   .cur2g     equ  dword[ebp-152] ;dw ?
+   .cur2b     equ	[ebp-156] ;dw ?
 
 
        mov    ebp,esp
-if Ext>=MMX
-       emms
-end if
 
    ;    mov     .tex_ptr,edx
    ;    mov     .z_ptr,esi
@@ -468,18 +466,18 @@ end if
        push	.zz2
 
        push	.scan_x2
-       push	.scan_y2
-       push	.cur2r
+       push	dword .scan_y2
+       push	dword .cur2r
        push	.cur2g
-       push	.cur2b
+       push	dword .cur2b
 
        push	.zz1
 
        push	.scan_x1
-       push	.scan_y1
-       push	.cur1r
+       push	dword .scan_y1
+       push	dword .cur1r
        push	.cur1g
-       push	.cur1b
+       push	dword .cur1b
 
        sar	eax,ROUND
        sar	ebx,ROUND
@@ -487,18 +485,39 @@ end if
 
       ; pop      ebp ebx eax
        popad
+
+if Ext >= MMX
+       movq	mm0,.cur1b
+       movq	mm1,.cur1r
+       movq	mm2,.scan_y1
+       movq	mm3,.cur2b
+       movq	mm4,.cur2r
+       movq	mm5,.scan_y2
+       paddd	mm0,.dc13b
+       paddd	mm1,.dc13r
+       paddd	mm2,.tex_dy13
+       paddd	mm3,.dc12b
+       paddd	mm4,.dc12r
+       paddd	mm5,.tex_dy12
+       movq	.cur1b,mm0
+       movq	.cur1r,mm1
+       movq	.scan_y1,mm2
+       movq	.cur2b,mm3
+       movq	.cur2r,mm4
+       movq	.scan_y2,mm5
+else
        mov	edx,.dc13b
        add	.cur1b,edx
        mov	esi,.dc13g
        add	.cur1g,esi
        mov	edi,.dc13r
        add	.cur1r,edi
+       mov	edx,.dz13
+       add	.zz1,edx
        mov	edx,.tex_dx13
        add	.scan_x1,edx
        mov	esi,.tex_dy13
        add	.scan_y1,esi
-       mov	edx,.dz13
-       add	.zz1,edx
 
        mov	edi,.dc12b
        add	.cur2b,edi
@@ -512,6 +531,7 @@ end if
        add	.scan_y2,esi
        mov	edx,.dz12
        add	.zz2,edx
+end if
        add	eax,.dx13
        add	ebx,.dx12
        inc	cx
@@ -563,24 +583,45 @@ end if
        push	.zz2
 
        push	.scan_x2
-       push	.scan_y2
-       push	.cur2r
+       push	dword .scan_y2
+       push	dword .cur2r
        push	.cur2g
-       push	.cur2b
+       push	dword .cur2b
 
        push	.zz1
 
        push	.scan_x1
-       push	.scan_y1
-       push	.cur1r
+       push	dword .scan_y1
+       push	dword .cur1r
        push	.cur1g
-       push	.cur1b
+       push	dword .cur1b
 
        sar	eax,ROUND
        sar	ebx,ROUND
        call	horizontal_tex_grd_line
 
        popad
+
+if Ext >= MMX
+       movq	mm0,.cur1b
+       movq	mm1,.cur1r
+       movq	mm2,.scan_y1
+       movq	mm3,.cur2b
+       movq	mm4,.cur2r
+       movq	mm5,.scan_y2
+       paddd	mm0,.dc13b
+       paddd	mm1,.dc13r
+       paddd	mm2,.tex_dy13
+       paddd	mm3,.dc23b
+       paddd	mm4,.dc23r
+       paddd	mm5,.tex_dy23
+       movq	.cur1b,mm0
+       movq	.cur1r,mm1
+       movq	.scan_y1,mm2
+       movq	.cur2b,mm3
+       movq	.cur2r,mm4
+       movq	.scan_y2,mm5
+else
        mov	edx,.dc13b
        add	.cur1b,edx
        mov	esi,.dc13g
@@ -606,6 +647,7 @@ end if
        add	.scan_y2,esi
        mov	edx,.dz23
        add	.zz2,edx
+end if
        add	eax,.dx13
        add	ebx,.dx23
        inc	cx
diff --git a/programs/demos/3DS/TEX_CAT.INC b/programs/demos/3DS/TEX_CAT.INC
index 68a3ee87d1..8fb9db5df9 100644
--- a/programs/demos/3DS/TEX_CAT.INC
+++ b/programs/demos/3DS/TEX_CAT.INC
@@ -59,9 +59,6 @@ tex_triangle_z:
 .cz1	  equ dword[ebp-88]
 .cz2	  equ dword[ebp-92]
 
-if Ext >= MMX
-	emms
-end if
 	mov	ebp,esp
 	push	esi		  ; store memory pointers
 	push	edx
diff --git a/programs/demos/3DS/VIEW3DS.ASM b/programs/demos/3DS/VIEW3DS.ASM
index 7a4751c306..f0c331d022 100644
--- a/programs/demos/3DS/VIEW3DS.ASM
+++ b/programs/demos/3DS/VIEW3DS.ASM
@@ -1,6 +1,6 @@
 
-; application : View3ds ver. 0.053 - tiny .3ds files viewer.
-; compiler    : FASM  1.65.13
+; application : View3ds ver. 0.054 - tiny .3ds files viewer.
+; compiler    : FASM  1.67.11
 ; system      : KolibriOS/MenuetOS
 ; author      : Macgub aka Maciej Guba
 ; email       : macgub3@wp.pl
@@ -215,6 +215,13 @@ still:
      .next_m:
 	cmp	 ah,18
 	jne	 .next_m2
+     if USE_LFN
+	mov	 [re_alloc_flag],1	 ; reallocate memory
+	mov	 [triangles_count_var],1000
+	mov	 [points_count_var],1000
+	call	 alloc_mem_for_tp
+	mov	 [re_alloc_flag],0
+     end if
 	mov	 bl,[generator_flag]
 	or	 bl,bl
 	jz	 .next_m2
@@ -244,8 +251,11 @@ still:
      @@:
 	mov	 edi,bumpmap
 	call	 calc_bumpmap
-      .next_m3:
-
+     .next_m3:
+	cmp	ah,21		 ; re map bumps, texture coordinates
+	jne	@f
+	call	calc_bumpmap_coords
+      @@:
 	jmp	noclose
 
 
@@ -595,14 +605,16 @@ calc_bumpmap_coords:	  ; map texture, bump
 ;      mov   ebp,esp
 ;      sub   esp,4
 
-      fninit			 ; spherical mapping around y axle
+      fninit
       fldpi
       fadd	st,st
-;      fstp      .Pi2
       mov	esi,[points_ptr]
       mov	edi,tex_points
       movzx	ecx,[points_count_var]
       inc	ecx
+;      cmp       [map_tex_flag],1
+;      jne       .cylindric
+      ; spherical mapping around y axle
 
    @@:
       fld	dword[esi]     ; x coord
@@ -635,8 +647,27 @@ calc_bumpmap_coords:	  ; map texture, bump
       add      edi,4
       loop     @b
       ffree    st0
+;      jmp      .end_map
+;  .cylindric:
+;       fld     dword[esi]     ; around y axle
+;       fld     dword[esi+8]
+;       fpatan
+;       fdiv    st0,st1
+;       fimul   [tex_x_div2]
+;       fiadd   [tex_x_div2]
+;       fistp   word[edi]
 
-;      mov      esp,ebp
+;       fld     dword[esi+4]
+;       fimul   [tex_y_div2]
+;       fiadd   [tex_y_div2]
+;       fistp   word[edi+2]
+
+;       add     esi,12
+;       add     edi,4
+;       loop    .cylindric
+;       ffree    st0
+;;      mov      esp,ebp
+;   .end_map:
 ret
 
 
@@ -2399,11 +2430,18 @@ ret
 if USE_LFN
 alloc_mem_for_tp:
 	mov	eax, 68
+	cmp	[re_alloc_flag],1
+	jz	@f
 	mov	ebx, 12
+	jmp	.alloc
+    @@:
+	mov	ebx,20
+    .alloc:
 	movzx	ecx, [triangles_count_var]
 	inc	ecx
 	lea	ecx, [ecx*3]
 	add	ecx, ecx
+	mov	edx,[triangles_ptr]
 	int	0x40		       ;  -> allocate memory to triangles
 	mov	[triangles_ptr], eax   ;  -> eax = pointer to allocated mem
 
@@ -2419,6 +2457,7 @@ alloc_mem_for_tp:
 	movzx	ecx, [triangles_count_var]
 	lea	ecx, [3+ecx*3]
 	shl	ecx, 2
+	mov	edx,[triangles_normals_ptr]
 	int	0x40			       ;  -> allocate memory for triangles normals
 	mov	[triangles_normals_ptr], eax   ;  -> eax = pointer to allocated mem
 
@@ -2426,22 +2465,26 @@ alloc_mem_for_tp:
 	movzx	ecx, [points_count_var]
 	lea	ecx,[3+ecx*3]
 	shl	ecx, 2
+	mov	edx,[points_normals_ptr]
 	int	0x40
 	mov	[points_normals_ptr], eax
 
 	mov	eax, 68
-	mov	ebx, 12
+    ;    mov     ebx, 12
 	movzx	ecx, [points_count_var]
 	lea	ecx,[3+ecx*3]
 	shl	ecx, 2
+	mov	edx,[points_normals_rot_ptr]
 	int	0x40
 	mov	[points_normals_rot_ptr], eax
 
 	mov	eax, 68
+	mov	edx,[points_ptr]
 	int	0x40
 	mov	[points_ptr], eax
 
 	mov	eax, 68
+	mov	edx,[points_rotated_ptr]
 	int	0x40
 	mov	[points_rotated_ptr], eax
 
@@ -2449,6 +2492,7 @@ alloc_mem_for_tp:
 	movzx	ecx, [points_count_var]
 	inc	ecx
 	shl	ecx, 3
+	mov	edx,[points_translated_ptr]
 	int	0x40
 	mov	[points_translated_ptr], eax
 ret
@@ -2592,7 +2636,7 @@ ret
 	add	bx,110
 	mov	cx,[size_y]
 	add	cx,30
-	mov	edx,0x02000000	; color of work area RRGGBB,8->color gl
+	mov	edx,0x14000000	; color of work area RRGGBB,8->color gl
 	mov	esi,0x805080d0	; color of grab bar  RRGGBB,8->color gl
 	mov	edi,0x005080d0	; color of frames    RRGGBB
 	int	0x40
@@ -2606,15 +2650,15 @@ ret
 	int	0x40
 
 	; CLOSE BUTTON
-	mov	eax,8		; function 8 : define and draw button
-	movzx	ebx,[size_x]
-	shl	ebx,16
-	add	ebx, 91 shl 16 + 12
-;        mov     ebx,(SIZE_X+80+30-19)*65536+12     ; [x start] *65536 + [x size]
-	mov	ecx,5*65536+12	; [y start] *65536 + [y size]
-	mov	edx,1		; button id
-	mov	esi,0x6688dd	; button color RRGGBB
-	int	0x40
+;       mov     eax,8           ; function 8 : define and draw button
+;       movzx   ebx,[size_x]
+;       shl     ebx,16
+;       add     ebx, 91 shl 16 + 12
+;;        mov     ebx,(SIZE_X+80+30-19)*65536+12     ; [x start] *65536 + [x size]
+;       mov     ecx,5*65536+12  ; [y start] *65536 + [y size]
+;       mov     edx,1           ; button id
+;       mov     esi,0x6688dd    ; button color RRGGBB
+;       int     0x40
 
 	call	buttons 	; more buttons
 
diff --git a/programs/demos/3DS/data.inc b/programs/demos/3DS/data.inc
index e6352b799e..0c753d684b 100644
--- a/programs/demos/3DS/data.inc
+++ b/programs/demos/3DS/data.inc
@@ -15,6 +15,7 @@
 	size_x		dw	SIZE_X
 	size_y		dw	SIZE_Y
 
+	re_alloc_flag db 0
 	angle_counter dw 0
 	piD180	      dd 0.017453292519943295769236907684886
 	piD128	      dd 0.024544
@@ -146,6 +147,11 @@ bumps_flag	  db 0
 bumps_deep_flag   db 3
 		  dd bumps_d_f
 
+		  db 21
+		  db 're-map tex'
+		  db 255
+map_tex_flag	  db ? ;1
+		  dd ?	;bumps_d_f
 ;                  db 21
 ;                  db 'light No. '
 ;                  db 3
@@ -262,7 +268,7 @@ base_vector:
       if Ext=SSE
 	db   ' (SSE)'
       end if
-	db   ' 0.053'
+	db   ' 0.054'
     labellen:
 	STRdata db '-1        '
 
@@ -277,7 +283,7 @@ if USE_LFN
 	 fsize	dd	0 ;180000 ; sizeof(workarea)
 	 fptr	dd	0 ;workarea
 	file_name:
-		db	'/rd/1/3d/teapot.3ds',0
+		db	'/rd/1/3d/house.3ds',0
 
 else
 
@@ -288,7 +294,7 @@ else
 		dd workarea
 		dd hash_table
 	file_name:
-		db   '/rd/1/teapot.3ds',0
+		db   '/rd/1/house.3ds',0
 end if
 
 I_END:
@@ -390,7 +396,7 @@ align 8
 	rand_seed	    dw ?
 align 8
 	buffer	dq	?
-	err_	dd	?
+	err	dd	?
 	drr	dd	?
 	xx	dd	?
 	yy	dd	?
diff --git a/programs/demos/3DS/readme.txt b/programs/demos/3DS/readme.txt
index 09818815f8..6c93385b82 100644
--- a/programs/demos/3DS/readme.txt
+++ b/programs/demos/3DS/readme.txt
@@ -1,8 +1,11 @@
-View3ds 0.053 - tiny viewer to .3ds files.
+View3ds 0.054 - tiny viewer to .3ds files.
 
 What's new?
-1. Optimizations.
-
+1. Skinned window by Leency.
+2. Optimizations.
+3. Re map texture, bumps option - allow spherical mapping around each axle (X,Y,Z).
+4. Problem with too small memory to generate object fixed. (Problem ocurred with
+   house.3ds object and others objects contains less than 1000 faces and points).
 
 Buttons description:
 1. rotary: choosing rotary axle: x, y, x+y.
@@ -14,8 +17,8 @@ Buttons description:
    mapping).
 3. speed: idle, full.
 4,5. zoom in, out: no comment.
-6. catmull: on( use z buffer ( z coordinate interpolation), off( depth sorting, painters alghoritm).
-   txgrd and 2tex models only with catmull  = on.
+6. catmull: on( use z buffer ( z coordinate interpolation), off( depth sorting, painters
+   alghoritm).Txgrd and 2tex models only with catmull  = on.
 7. culling: backface culling on/ off.
 8. rand. light: Randomize 3 unlinear lights( so called Phong's illumination).
 9. Blur: blur N times; N=0,1,2,3,4,5
@@ -26,6 +29,9 @@ Buttons description:
 14. move: changes meaning x,y,z +/- buttons  ->  obj: moving object, camr: moving camera.
 15. generate: Generates some objects: node, Thorn Crown, heart...
 16. bumps: random, according to texture.
-15. bumps deep -> create bumps deeper or lighter.
+17. bumps deep -> create bumps deeper or lighter.
+18. re-map tex -> re-map texture and bump map coordinates, to change spherical mapping 
+    around axle use 'xchg' and 'mirror' buttons, then press 're-map tex' button.
 
-			 Macgub 		X 2009
\ No newline at end of file
+			 Macgub 		december 2009
+                                                                                  Last edited Jan 2010
\ No newline at end of file