WebSVN – Kolibri OS – Path Comparison – / – /programs/demos Rev 2191 and /programs/demos Rev 2192

Regard whitespace Rev 2191 → Rev 2192

 /programs/demos/3DS/BUMP_CAT.INC
 ,10 → 81,10
 .cz2   equ      [ebp-108]
 .cbx1  equ dword[ebp-112]
 .cby1  equ      [ebp-116]
-.cbx2  equ dword[ebp-120]
-.cby2  equ      [ebp-124]
-.cex1  equ dword[ebp-128]
-.cey1  equ      [ebp-132]
+.cex1  equ dword[ebp-120]
+.cey1  equ      [ebp-124]
+.cbx2  equ dword[ebp-128]
+.cby2  equ      [ebp-132]
 .cex2  equ dword[ebp-136]
 .cey2  equ      [ebp-140]
 ,8 → 533,25
        jge      .loop12_done
   .loop12:
        call     .call_bump_line
+if Ext >= SSE2
+       movups   xmm0,.cey2
+       movups   xmm1,.cey1
+       movups   xmm2,.dey12
+       movups   xmm3,.dey13
+       paddd    xmm0,xmm2
+       paddd    xmm1,xmm3
+       movups   .cey2,xmm0
+       movups   .cey1,xmm1
+       movq     mm4,.cz1
+       movq     mm5,.cz2
+       paddd    mm4,.dz13
+       paddd    mm5,.dz12
+       movq     .cz1,mm4
+       movq     .cz2,mm5
+end if
-if Ext >= MMX
+if (Ext = MMX) | (Ext = SSE)
        movq     mm0,.cby2
        movq     mm1,.cby1
        movq     mm2,.cey2
 ,7 → 570,7
        movq     .cey2,mm2
        movq     .cz1,mm4
        movq     .cz2,mm5
-else
+else if Ext = NON
        mov      edx,.dbx13
        add      .cbx1,edx
        mov      eax,.dbx12
 ,8 → 634,23
      .loop23:
        call     .call_bump_line
-if Ext >= MMX
+if Ext >=  SSE2
+       movups   xmm0,.cey2
+       movups   xmm1,.cey1
+       movups   xmm2,.dey23
+       movups   xmm3,.dey13
+       paddd    xmm0,xmm2
+       paddd    xmm1,xmm3
+       movups   .cey2,xmm0
+       movups   .cey1,xmm1
+       movq     mm4,.cz1
+       movq     mm5,.cz2
+       paddd    mm4,.dz13
+       paddd    mm5,.dz23
+       movq     .cz1,mm4
+       movq     .cz2,mm5
+end if
+if (Ext = MMX) | (Ext = SSE)
        movq     mm0,.cby2
        movq     mm1,.cby1
        movq     mm2,.cey2
 ,7 → 669,7
        movq     .cey2,mm2
        movq     .cz1,mm4
        movq     .cz2,mm5
-else
+else if Ext = NON
        mov      eax,.dx13
        add      .cx1,eax
        mov      ebx,.dx23

 /programs/demos/3DS/BUMP_TEX.INC
 ,24 → 292,6
        cvtps2pi  mm1,xmm1
        movq      .dty12,mm0
        movq      .dz12,mm1
-;temporaly conversion to keep upside down
-;     fninit
-;     fld      .dx12
-;     fistp    .dx12
-;     fld      dword .dz12
-;     fistp    dword .dz12
-;     fld      .dbx12
-;     fistp    .dbx12
-;     fld      dword .dby12
-;     fistp    dword .dby12
-;     fld      .dex12
-;     fistp    .dex12
-;     fld      dword .dey12
-;     fistp    dword .dey12
-;     fld      .dtx12
-;     fistp    .dtx12
-;     fld      dword .dty12
-;     fistp    dword .dty12
 ;----
 ;       mov       ax,.z2
 ;       sub       ax,.z1
 ,25 → 500,6
        movq      .dty13,mm0
        movq      .dz13,mm1
-;temporaly conversion to keep upside down
-;     fninit
-;     fld      .dx13
-;     fistp    .dx13
-;     fld      dword .dz13
-;     fistp    dword .dz13
-;     fld      .dbx13
-;     fistp    .dbx13
-;     fld      dword .dby13
-;     fistp    dword .dby13
-;     fld      .dex13
-;     fistp    .dex13
-;     fld      dword .dey13
-;     fistp    dword .dey13
-;     fld      .dtx13
-;     fistp    .dtx13
-;     fld      dword .dty13
-;     fistp    dword .dty13
 else
        mov      ax,.x3
 ,24 → 662,6
        movq      .dz23,mm1
-;temporaly conversion to keep upside down
-;     fninit
-;     fld      .dx23
-;     fistp    .dx23
-;     fld      dword .dz23
-;     fistp    dword .dz23
-;     fld      .dbx23
-;     fistp    .dbx23
-;     fld      dword .dby23
-;     fistp    dword .dby23
-;     fld      .dex23
-;     fistp    .dex23
-;     fld      dword .dey23
-;     fistp    dword .dey23
-;     fld      .dtx23
-;     fistp    .dtx23
-;     fld      dword .dty23
-;     fistp    dword .dty23
 else
        mov      ax,.x3
        sub      ax,.x2
 ,58 → 739,7
       ;  sub     esp,40
    .bt_dx23_done:
        sub       esp,64
-;if Ext>=SSE
-;       movsx    eax,.x1
-;       shl      eax,ROUND
-;       cvtsi2ss xmm0,eax
-;       movss    .cx1,xmm0
-;       movss    .cx2,xmm0
-;
-;       movsx    ebx,word[.b_x1]
-;       shl      ebx,ROUND
-;       cvtsi2ss xmm0,ebx
-;       movss    .cbx1,xmm0
-;       movss    .cbx2,xmm0
-;
-;       movsx    ecx,word[.b_y1]
-;       shl      ecx,ROUND
-;       cvtsi2ss xmm0,ecx
-;       movss    .cby1,xmm0
-;       movss    .cby2,xmm0
-;       movsx    edx,word[.e_x1]
-;       shl      edx,ROUND
-;       cvtsi2ss xmm0,edx
-;       movss    .cex1,xmm0
-;       movss    .cex2,xmm0
-;
-;       movsx    eax,word[.e_y1]
-;       shl      eax,ROUND
-;       cvtsi2ss xmm0,eax
-;       movss    .cey1,xmm0
-;       movss    .cey2,xmm0
-;
-;       movsx    ebx,.z1
-;       shl      ebx,CATMULL_SHIFT
-;       cvtsi2ss xmm0,ebx
-;       movss    .cz1,xmm0
-;       movss    .cz2,xmm0
-;
-;       movsx    ecx,word[.t_x1]
-;       shl      ecx,ROUND
-;       cvtsi2ss xmm0,ecx
-;       movss    .ctx1,xmm0
-;       movss    .ctx2,xmm0
-;       movsx    edx,word[.t_y1]
-;       shl      edx,ROUND
-;       cvtsi2ss xmm0,edx
-;       movss    .cty1,xmm0
-;       movss    .cty2,xmm0
-;else
        movsx    eax,.x1
        shl      eax,ROUND
        mov      .cx1,eax
 ,14 → 796,8
        mov      .cty2,edx
       ; push     edx
       ; push     edx
-;end if
-       movsx    ecx,.y1
-       cmp      cx,.y2
-       jge      .loop12_done
-  .loop12:
-       call     .call_line
-;if Ext >= SSE
+;if Ext >= SSE2
 ;       movups  xmm0,.cby1
 ;       movups  xmm1,.cty1
 ;       movups  xmm2,.cby2
 ,22 → 806,36
 ;       movups  xmm5,.dty13
 ;       movups  xmm6,.dby12
 ;       movups  xmm7,.dty12
-;       addps   xmm0,xmm4
-;       addps   xmm1,xmm5
-;       addps   xmm2,xmm6
-;       addps   xmm3,xmm7
-;
-;;       addps   xmm0,.dby12
-;;       addps   xmm1,.dty12
-;;       addps   xmm2,.dby13
-;;       addps   xmm3,.dty13
-;       movups  .cby1,xmm0
-;       movups  .cty1,xmm1
-;       movups  .cby2,xmm2
-;       movups  .cty2,xmm3
 ;end if
+       movsx    ecx,.y1
+       cmp      cx,.y2
+       jge      .loop12_done
+  .loop12:
+;if Ext >= SSE2
+;       fxsave  [sse_repository]
+;end if
+       call     .call_line
+if Ext >= SSE2
+;       fxrstor [sse_repository]
+       movups  xmm0,.cby1
+       movups  xmm1,.cty1
+       movups  xmm2,.cby2
+       movups  xmm3,.cty2
+       movups  xmm4,.dby13
+       movups  xmm5,.dty13
+       movups  xmm6,.dby12
+       movups  xmm7,.dty12
+       paddd   xmm0,xmm4
+       paddd   xmm1,xmm5
+       paddd   xmm2,xmm6
+       paddd   xmm3,xmm7
+       movups  .cby1,xmm0
+       movups  .cty1,xmm1
+       movups  .cby2,xmm2
+       movups  .cty2,xmm3
+end if
-if Ext >= MMX
+if (Ext = MMX) | (Ext = SSE)
        movq     mm0,.cby2
        movq     mm1,.cby1
        movq     mm2,.cey2
 ,7 → 907,6
        cmp      cx,.y3
        jge      .loop23_done
-;if Ext < SSE
        movsx    eax,.z2
        shl      eax,CATMULL_SHIFT
 ,54 → 939,7
        movzx    ebx,word[.t_y2]
        shl      ebx,ROUND
        mov      .cty2,ebx
-;else
-;       movsx    eax,.z2
-;       shl      eax,CATMULL_SHIFT
-;       cvtsi2ss xmm0,eax
-;       movss    .cz2,xmm0
-;
-;       movsx    ebx,.x2
-;       shl      ebx,ROUND
-;       cvtsi2ss xmm1,ebx
-;       movss    .cx2,xmm1
-;
-;       movzx    edx,word[.b_x2]
-;       shl      edx,ROUND
-;       cvtsi2ss xmm2,edx
-;       movss    .cbx2,xmm2
-;
-;       movzx    eax,word[.b_y2]
-;       shl      eax,ROUND
-;       cvtsi2ss xmm0,eax
-;       movss    .cby2,xmm0
-;
-;       movzx    ebx,word[.e_x2]
-;       shl      ebx,ROUND
-;       cvtsi2ss xmm1,ebx
-;       movss    .cex2,xmm1
-;
-;       movzx    edx,word[.e_y2]
-;       shl      edx,ROUND
-;       cvtsi2ss xmm2,edx
-;       movss    .cey2,xmm2
-;
-;       movzx    eax,word[.t_x2]
-;       shl      eax,ROUND
-;       cvtsi2ss xmm0,eax
-;       movss    .ctx2,xmm0
-;
-;       movzx    ebx,word[.t_y2]
-;       shl      ebx,ROUND
-;       cvtsi2ss xmm1,ebx
-;       movss    .cty2,xmm1
-;end if
-     .loop23:
-       call     .call_line
-;if Ext >= SSE
+;if Ext >= SSE2
 ;       movups  xmm0,.cby1
 ;       movups  xmm1,.cty1
 ;       movups  xmm2,.cby2
 ,21 → 948,34
 ;       movups  xmm5,.dty13
 ;       movups  xmm6,.dby23
 ;       movups  xmm7,.dty23
-;       addps   xmm0,xmm4
-;       addps   xmm1,xmm5
-;       addps   xmm2,xmm6
-;       addps   xmm3,xmm7
-; ;      addps   xmm0,.dby13
-; ;      addps   xmm1,.dty13
-; ;      addps   xmm2,.dby23
-; ;      addps   xmm3,.dty23
-;       movups  .cby1,xmm0
-;       movups  .cty1,xmm1
-;       movups  .cby2,xmm2
-;       movups  .cty2,xmm3
+;end if
+     .loop23:
+;if Ext >= SSE2
+;       fxsave  [sse_repository]
+;end if
+       call     .call_line
+if Ext >= SSE2
+;       fxrstor [sse_repository]
+       movups  xmm0,.cby1
+       movups  xmm1,.cty1
+       movups  xmm2,.cby2
+       movups  xmm3,.cty2
+       movups  xmm4,.dby13
+       movups  xmm5,.dty13
+       movups  xmm6,.dby23
+       movups  xmm7,.dty23
+       paddd   xmm0,xmm4
+       paddd   xmm1,xmm5
+       paddd   xmm2,xmm6
+       paddd   xmm3,xmm7
+       movups  .cby1,xmm0
+       movups  .cty1,xmm1
+       movups  .cby2,xmm2
+       movups  .cty2,xmm3
 ;
-;end if
-if Ext >= MMX
+end if
+if (Ext = MMX) | (Ext = SSE)
        movq     mm0,.cby2
        movq     mm1,.cby1
        movq     mm2,.cey2
 ,90 → 1047,7
 ret   50
 .call_line:
-;if Ext >= SSE
-;       pushad
-;       push     .tex_ptr
-;       fninit
-;       fld      dword .cty1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     dword .cty1
-;       fld      .ctx1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     .ctx1
-;       fld      dword .cz1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;    ;   push     dwod .cz1
-;       fld      dword .cty2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;    ;   push     .cty2
-;       fld      .ctx2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;  ;     push     dword .ctx2
-;       fld      dword .cz2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     dword .cz2
-;
-;       push     .z_buff
-;       push     .t_emap
-;       push     .t_bmap
-;--------------------------------------
-;       fld      dword .cey2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     dword .cey2
-;
-;       fld      .cex2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     .cex2
-;
-;       fld      dword .cby2
-;       fistp    dword [esp-4]
-;       sub      esp,4
-; ;      push     dword .cby2
-;       fld      .cbx2
-;       fistp    dword [esp-4]
-;       sub      esp,4
- ;      push     .cbx2
-;------------------------------------
-;       fld      dword .cey1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     dword .cey1
-;       fld      .cex1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;    ;   push     .cex1
-;       fld      dword .cby1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;       push     dword .cby1
-;       fld      .cbx1
-;       fistp    dword [esp-4]
-;       sub      esp,4
-;;      push     .cbx1
-;       push     ecx
-;       fld      .cx1
-;       fistp    dword [esp-4]
-;       mov      eax,[esp-4]
-;       sar      eax,ROUND
-;       fld      .cx2
-;       fistp    dword [esp-4]
-;       mov      ebx,[esp-4]
-;       sar      ebx,ROUND
-;       call     bump_tex_line_z
-;
-;       popad
-;else
        pushad
        push     .tex_ptr
        push     dword .cty1
 ,8 → 1065,16
        push     .cbx2
        push     dword .cey1
        push     .cex1
+;if Ext >= SSE2
+;       sub      esp,8
+;       shufps   xmm0,xmm0,10110100b
+;       movhps   [esp],xmm0 ;================================
+;else
        push     dword .cby1
        push     .cbx1
+;end if
        push     ecx
        mov      eax,.cx1

 /programs/demos/3DS/DATA.INC
 ,7 → 271,7
       if Ext=SSE2
         db   ' (SSE2)'
       end if
-        db   ' 0.059',0
+        db   ' 0.060',0
     labellen:
         STRdata db '-1        '

 /programs/demos/3DS/GRD_TEX.INC
 ,7 → 486,7
       ; pop      ebp ebx eax
        popad
-if Ext >= MMX
+if (Ext = MMX)|(Ext=SSE)
        movq     mm0,.cur1b
        movq     mm1,.cur1r
        movq     mm2,.scan_y1
 ,7 → 505,25
        movq     .cur2b,mm3
        movq     .cur2r,mm4
        movq     .scan_y2,mm5
-else
+end if
+if Ext >= SSE2
+       movups   xmm0,.cur1b
+       movups   xmm1,.dc13b
+       movups   xmm2,.cur2b
+       movups   xmm3,.dc12b
+       movq     mm2,.scan_y1
+       movq     mm5,.scan_y2
+       paddd    xmm0,xmm1
+       paddd    xmm2,xmm3
+       paddd    mm2,.tex_dy13
+       paddd    mm5,.tex_dy12
+       movq     .scan_y1,mm2
+       movq     .scan_y2,mm5
+       movups   .cur1b,xmm0
+       movups   .cur2b,xmm2
+end if
+if Ext = NON
        mov      edx,.dc13b
        add      .cur1b,edx
        mov      esi,.dc13g
 ,7 → 620,7
        popad
-if Ext >= MMX
+if (Ext = MMX)|(Ext=SSE)
        movq     mm0,.cur1b
        movq     mm1,.cur1r
        movq     mm2,.scan_y1
 ,7 → 639,24
        movq     .cur2b,mm3
        movq     .cur2r,mm4
        movq     .scan_y2,mm5
-else
+end if
+if Ext >= SSE2
+       movups   xmm0,.cur1b
+       movups   xmm1,.dc13b
+       movups   xmm2,.cur2b
+       movups   xmm3,.dc23b
+       movq     mm2,.scan_y1
+       movq     mm5,.scan_y2
+       paddd    xmm0,xmm1
+       paddd    xmm2,xmm3
+       paddd    mm2,.tex_dy13
+       paddd    mm5,.tex_dy23
+       movq     .scan_y1,mm2
+       movq     .scan_y2,mm5
+       movups   .cur1b,xmm0
+       movups   .cur2b,xmm2
+end if
+if Ext = NON
        mov      edx,.dc13b
        add      .cur1b,edx
        mov      esi,.dc13g
 ,8 → 775,8
         mov     ecx,dword .z1
         xchg    ecx, .z2
         mov     dword .z1, ecx
-else
+end if
+if (Ext=MMX)
         movq    mm0,.b1        ; b, g
         movq    mm1,.b2
         movq    .b1, mm1
 ,6 → 791,16
         movq    .tex_x2,mm4
 end if
+if Ext>=SSE
+       movups  xmm0,.b1
+       movups  xmm1,.b2
+       movups  .b1,xmm1
+       movups  .b2,xmm0
+       movq    mm4,.tex_x1    ; x, z
+       movq    mm5,.tex_x2
+       movq    .tex_x1,mm5
+       movq    .tex_x2,mm4
+end if
     @@:
         or      bx,bx

 /programs/demos/3DS/TWO_TEX.INC
 ,16 → 88,12
 .cx1   equ dword[ebp-96]   ; current variables
 .cx2   equ dword[ebp-100]
-;.cbx1q equ      [ebp-104]
 .cbx1  equ dword[ebp-104]
 .cby1  equ [ebp-108]
-;.cbx2q          [ebp-112]
-.cbx2  equ dword[ebp-112]
-.cby2  equ [ebp-116]
-;.cex1q equ      [ebp-120]
-.cex1  equ dword[ebp-120]
-.cey1  equ [ebp-124]
-;.cex2q equ      [ebp-128]
+.cex1  equ dword[ebp-112]
+.cey1  equ [ebp-116]
+.cbx2  equ dword[ebp-120]
+.cby2  equ [ebp-124]
 .cex2  equ dword[ebp-128]
 .cey2  equ [ebp-132]
 ,8 → 575,16
        add      .cx1,eax
        mov      ebx,.dx12
        add      .cx2,ebx
-if Ext >= MMX
+if Ext>= SSE2
+       movups   xmm0,.cey1
+       movups   xmm1,.cey2
+       movups   xmm2,.dey12q
+       movups   xmm3,.dey13q
+       paddd    xmm0,xmm3
+       paddd    xmm1,xmm2
+       movups   .cey1,xmm0
+       movups   .cey2,xmm1
+else if (Ext = MMX) | (Ext=SSE)
        movq     mm0,.cby2       ; with this optimization object
        movq     mm1,.cby1       ; looks bit annoying
        movq     mm2,.cey2
 ,8 → 662,16
        add      .cx1,eax
        mov      ebx,.dx23
        add      .cx2,ebx
-if Ext >= MMX
+if Ext>= SSE2
+       movups   xmm0,.cey1
+       movups   xmm1,.cey2
+       movups   xmm2,.dey23q
+       movups   xmm3,.dey13q
+       paddd    xmm0,xmm3
+       paddd    xmm1,xmm2
+       movups   .cey1,xmm0
+       movups   .cey2,xmm1
+else if (Ext = MMX) | ( Ext = SSE)
        movq     mm0,.cby2        ;    with this mmx optimization object looks bit
        movq     mm1,.cby1        ;    annoying
        movq     mm2,.cey2

/programs/demos/3DS/VIEW3DS.ASM
1,5 → 1,5

; application : View3ds ver. 0.059 - tiny .3ds files viewer.
; application : View3ds ver. 0.060 - tiny .3ds files viewer.
; compiler : FASM
; system : KolibriOS
; author : Macgub aka Maciej Guba

 /programs/demos/3DS/readme.txt
 ,12 → 1,9
-View3ds 0.059 - tiny viewer to .3ds files.
+View3ds 0.060 - tiny viewer to .3ds files.
 What's new?
-. Bump and pararell two texture  mapping functions optimizations.
-   (files bump_cat.inc & two_tex.inc)
-   On my P4 changes are rather non visable, but on dual core in KlbrInWin
-   optimizations runs preety nice.
+. Header fix by Leency.
+. SSE2 optimizations by me. (Most visable in BUMP_TEX mode.)
 Buttons description:
 . rotary: choosing rotary axle: x, y, x+y.
 . shd. model: choosing shading model: flat, grd (smooth), env (spherical
 ,4 → 30,4
 . 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                  Jun 2011
+                         Macgub                  Aug 2011

Subversion Repositories Kolibri OS

Compare Revisions

Regard whitespace Rev 2191 → Rev 2192