VESA 32bpp optimisation

git-svn-id: svn://kolibrios.org@1707 a494cfbc-eb01-0410-851d-a64ba20cac60

kernel/branches/Kolibri-A/trunk/hid/mousedrv.inc

@@ -154,7 +154,7 @@ drm:
 	add    eax,ecx	; save picture under mouse
 	add    ebx,edx
 	push   ecx
-	call   getpixel
+	call   get_pixel
 	mov    [COLOR_TEMP],ecx
 	pop    ecx
 	mov    eax,edx

kernel/branches/Kolibri-A/trunk/kernel.asm

@@ -379,8 +379,8 @@ high_code:
 ;        ===  EGA, VGA & Vesa 1.2 modes not supported ===
 setvesa20:
 v20ga32:
-	mov	[PUTPIXEL],dword Vesa20_putpixel32
+	mov	[PUTPIXEL],dword put_pixel
-	mov	[GETPIXEL],dword Vesa20_getpixel32
+	mov	[GETPIXEL],dword get_pixel
 
 ; -------- Fast System Call init ----------
 .SEnP:

kernel/branches/Kolibri-A/trunk/memmap.inc

@@ -143,14 +143,14 @@
 ;
 ; 0x80010000 ->  6CBFF   kernel, 32-bit run-time code (up to 371 Kb)
 ;       in the current version:
-;            ->  01726   16-bit code end
+;            ->  00B37   16-bit code end
-;            ->  01828   16-bit data end
+;            ->  00C40   16-bit data end
-;            ->  11828   32-bit code start
+;            ->  10C40   32-bit code start
-;            ->  2E19E   32-bit code end 
+;            ->  2D582   32-bit code end 
-; 	     ->  314F8..end_of_kernel	 zero-filled zone after preinit_mem  
+; 	     ->  30918..end_of_kernel	 zero-filled zone after preinit_mem  
-;            ->  34DFB   uninitialized globals start 
+;            ->  3421B   uninitialized globals start 
-;            ->  3CFEA   end_of_kernel 
+;            ->  3C40A   end_of_kernel 
-;	     ->  3D000   not used (190k)
+;	     ->  3D000   not used (194k)
 ; 0x80050000 -> 090000	 zero-filled zone after preinit_mem  
 ; 0x8006CC00 ->  6DBFF   stack at boot time (4Kb)
 ;

kernel/branches/Kolibri-A/trunk/video/vesa20.inc

@@ -20,17 +20,6 @@
 $Revision$
 
 
-; If you're planning to write your own video driver I suggest
-; you replace the VESA12.INC file and see those instructions.
-
-;Screen_Max_X             equ     0xfe00
-;Screen_Max_Y            equ     0xfe04
-;BytesPerScanLine        equ     0xfe08
-;LFBAddress              equ     0xfe80
-;ScreenBPP               equ     0xfbf1
-
-
-
 ;*************************************************
 ; getpixel
 ;
@@ -41,26 +30,12 @@ $Revision$
 ; ret:
 ; ecx = 00 RR GG BB
 
-getpixel:
-     push    eax ebx edx edi
-     call    dword [GETPIXEL]
-     pop     edi edx ebx eax
-     ret
 
-Vesa20_getpixel24:
+get_pixel:
-; eax = x
+     mov     ecx, [BytesPerScanLine]
-; ebx = y
+     imul    ecx, ebx
-     imul    ebx, [BytesPerScanLine]    ; ebx = y * y multiplier
+     lea     ecx, [ecx+eax*4]		; ecx = x*4+(y*y multiplier)
-     lea     edi, [eax+eax*2] ; edi = x*3
+     mov     ecx, [ecx+LFB_BASE]
-     add     edi, ebx         ; edi = x*3+(y*y multiplier)
-     mov     ecx, [LFB_BASE+edi]
-     and     ecx, 0xffffff
-     ret
-
-Vesa20_getpixel32:
-     imul    ebx, [BytesPerScanLine]    ; ebx = y * y multiplier
-     lea     edi, [ebx+eax*4] ; edi = x*4+(y*y multiplier)
-     mov     ecx, [LFB_BASE+edi]
      and     ecx, 0xffffff
      ret
 
@@ -129,10 +104,7 @@ vesa20_putimage:
      mov     [putimg.abs_cy], ebx
 ; real_sx = MIN(wnd_sx-image_cx, image_sx);
      mov     ebx, [eax-twdw + WDATA.box.width] ; ebx = wnd_sx
-; \begin{diamond}[20.08.2006]
+     inc     ebx	; WDATA.box.width is one pixel less than real window x-size
-; note that WDATA.box.width is one pixel less than real window x-size
-     inc     ebx
-; \end{diamond}[20.08.2006]
      sub     ebx, [putimg.image_cx]
      ja      @f
      add     esp, putimg.stack_data
@@ -146,9 +118,7 @@ vesa20_putimage:
      mov     [putimg.real_sx], ebx
 ; init real_sy
      mov     ebx, [eax-twdw + WDATA.box.height] ; ebx = wnd_sy
-; \begin{diamond}[20.08.2006]
      inc     ebx
-; \end{diamond}[20.08.2006]
      sub     ebx, [putimg.image_cy]
      ja      @f
      add     esp, putimg.stack_data
@@ -164,8 +134,6 @@ vesa20_putimage:
      mov     eax, [putimg.image_sx]
      mov     ecx, [putimg.real_sx]
      sub     eax, ecx
-;;     imul    eax, [putimg.source_bpp]
-;     lea     eax, [eax + eax * 2]
      call    esi
      add     eax, [putimg.arg_0]
      mov     [putimg.line_increment], eax
@@ -176,9 +144,8 @@ vesa20_putimage:
      mov     [putimg.winmap_newline], eax
 ; screen new line increment
      mov     eax, [BytesPerScanLine]
-     movzx   ebx, byte [ScreenBPP]
+     shl     ecx, 1
-     shr     ebx, 3
+     shl     ecx, 1
-     imul    ecx, ebx
      sub     eax, ecx
      mov     [putimg.screen_newline], eax
 ; pointer to image
@@ -187,9 +154,8 @@ vesa20_putimage:
      mov     edx, [putimg.abs_cy]
      imul    edx, [BytesPerScanLine]
      mov     eax, [putimg.abs_cx]
-     movzx   ebx, byte [ScreenBPP]
+     shl     eax, 1
-     shr     ebx, 3
+     shl     eax, 1
-     imul    eax, ebx
      add     edx, eax
 ; pointer to pixel map
      mov     eax, [putimg.abs_cy]
@@ -200,59 +166,12 @@ vesa20_putimage:
      xchg    eax, ebp
 ; get process number
      mov     ebx, [CURRENT_TASK]
-     cmp     byte [ScreenBPP], 32
-     je      put_image_end_32
-;put_image_end_24:
-     mov     edi, [putimg.real_sy]
-align   4
-.new_line:
-     mov     ecx, [putimg.real_sx]
-;     push    ebp edx
-align   4
-.new_x:
-     push    [putimg.edi]
-     mov     eax, [putimg.ebp+4]
-     call    eax
-     cmp     [ebp], bl
-     jne     .skip
-;     mov     eax, [esi]        ; eax = RRBBGGRR
-     mov     [LFB_BASE+edx], ax
-     shr     eax, 16
-     mov     [LFB_BASE+edx+2], al
-.skip:
-;     add     esi, 3 ;[putimg.source_bpp]
-     add     edx, 3
-     inc     ebp
-     dec     ecx
-     jnz     .new_x
-;     pop     edx ebp
-     add     esi, [putimg.line_increment]
-     add     edx, [putimg.screen_newline] ;[BytesPerScanLine]
-     add     ebp, [putimg.winmap_newline] ;[Screen_Max_X]
-;     inc     ebp
-     cmp     [putimg.ebp], putimage_get1bpp
-     jz      .correct
-     cmp     [putimg.ebp], putimage_get2bpp
-     jz      .correct
-     cmp     [putimg.ebp], putimage_get4bpp
-     jnz     @f
-.correct:
-     mov     eax, [putimg.edi]
-     mov     byte [eax], 80h
-@@:
-     dec     edi
-     jnz     .new_line
-.finish:
-     add     esp, putimg.stack_data
-     popad
-     ret
 
 put_image_end_32:
      mov     edi, [putimg.real_sy]
 align	4
 .new_line:
      mov     ecx, [putimg.real_sx]
-;     push    ebp edx
 align	4
 .new_x:
      push    [putimg.edi]
@@ -260,19 +179,15 @@ align   4
      call    eax
      cmp     [ebp], bl
      jne     .skip
-;     mov     eax, [esi]        ; ecx = RRBBGGRR
      mov     [LFB_BASE+edx], eax
 .skip:
-;     add     esi, [putimg.source_bpp]
      add     edx, 4
      inc     ebp
      dec     ecx
      jnz     .new_x
-;     pop     edx ebp
      add     esi, [putimg.line_increment]
      add     edx, [putimg.screen_newline] ;[BytesPerScanLine]
      add     ebp, [putimg.winmap_newline] ;[Screen_Max_X]
-;     inc     ebp
      cmp     [putimg.ebp], putimage_get1bpp
      jz      .correct
      cmp     [putimg.ebp], putimage_get2bpp
@@ -288,11 +203,8 @@ align   4
 .finish:
      add     esp, putimg.stack_data
      popad
-;     call   VGA__putimage
-;     mov     [EGA_counter],1
      ret
 
-
 ;*************************************************
 align 4
 __sys_putpixel:
@@ -302,55 +214,44 @@ __sys_putpixel:
 ; ecx = ?? RR GG BB    ; 0x01000000 negation
 ; edi = 0x00000001 force
 
-;;;        mov  [novesachecksum], dword 0
-     pushad
      cmp   [Screen_Max_X], eax
      jb    .exit
      cmp   [Screen_Max_Y], ebx
      jb    .exit
+.check_forced:
      test    edi,1		 ; force ?
-     jnz   .forced
+     jnz     .checked
 
-; not forced:
+.not_forced:
-
+     push     edx
-     push eax
      mov      edx,[_display.width]	   ; screen x size
      imul     edx, ebx
-     add  eax, [_WinMapAddress]
+     add      edx, [_WinMapAddress]
      movzx    edx, byte [eax+edx]
      cmp      edx, [CURRENT_TASK]
-     pop eax
+     pop      edx
      jne      .exit
 
-.forced:
+; OK to set pixel
-; check if negation
+.checked:
+     push  ebx
+     imul  ebx, [BytesPerScanLine]
+     lea   ebx, [ebx+eax*4]
      test  ecx,0x01000000
      jz    .noneg
-     call  getpixel
+     mov   ecx, [LFB_BASE+ebx]
      not   ecx
-     mov   [esp+32-8],ecx
+     and   ecx, 0x01FFFFFF
 .noneg:
-; OK to set pixel
+     mov   [LFB_BASE+ebx], ecx
-     call  dword [PUTPIXEL]    ; call the real put_pixel function
+     pop   ebx
 .exit:
-     popad
-     ret
-
-align 4
-Vesa20_putpixel24:
-; eax = x
-; ebx = y
-     imul    ebx, [BytesPerScanLine]     ; ebx = y * y multiplier
-     lea     edi, [eax+eax*2]  ; edi = x*3
-     mov     eax, [esp+32-8+4]
-     mov     [LFB_BASE+ebx+edi], ax
-     shr     eax, 16
-     mov     [LFB_BASE+ebx+edi+2], al
      ret
 
 
+
 align 4
-Vesa20_putpixel32:
+put_pixel:	; left for compatibility with Vesa20_putpixel32
 ; eax = x
 ; ebx = y
      imul    ebx, [BytesPerScanLine]	 ; ebx = y * y multiplier
@@ -359,6 +260,7 @@ Vesa20_putpixel32:
      mov     [LFB_BASE+edi], eax
      ret
 
+
 ;*************************************************
 
 ;align 4
@@ -492,50 +394,117 @@ align 4
 .exit:
      add     esp, 6*4
      popa
-;     dec   [mouse_pause]
      call   [draw_pointer]
      ret
 
-
+align 4
 hline:
-; draw an horizontal line
+; ------------  draw a horizontal line -------------
 ; eax = x1
 ; edx = x2
 ; ebx = y
 ; ecx = color
 ; edi = force ?
-     push    eax edx
+     cmp     ebx, [Screen_Max_Y]
-     cmp     edx, eax      ; make sure x2 is above x1
+     jge     .out
+     push    eax ebp esi ebx edx
+     mov     ebp, [_display.width]	; ebp = screen co-ords base
+     imul    ebp, ebx
+     add     ebp, [_WinMapAddress]
+
+     cmp     edx, eax		; to make sure x2 > x1
      jge     @f
      xchg    eax, edx
-align   4
 @@:
-     call    [putpixel]
+     cmp     eax, [Screen_Max_X]
+     jge     .exit
+     imul    ebx, [BytesPerScanLine]
+     add     ebx, LFB_BASE
+     cmp     edx, [Screen_Max_X]	; last check
+     jb      .draw
+     mov     edx, [Screen_Max_X]
+
+.draw:	; -- the line ---
+     test    edi,1			; forced  ?
+     jnz    .checked
+; check whether the line covered by other windows
+     movzx   esi, byte [ebp+eax]
+     cmp     esi, [CURRENT_TASK]
+     jne     .nextpixel
+.checked:
+     test    ecx,0x01000000
+     jz     .noneg
+     mov     ecx, [ebx+eax*4]
+     not     ecx
+     and     ecx, 0x01FFFFFF	; keep bit[24] high !
+.noneg:
+     mov     [ebx+eax*4], ecx
+.nextpixel:
      inc     eax
      cmp     eax, edx
-     jle     @b
+     jle     .draw
-     pop     edx eax
+.exit:
+     pop     edx ebx esi ebp eax
+.out:
      ret
 
-
+align 4
 vline:
-; draw a vertical line
+; ---------  draw a vertical line  ------------
 ; eax = x
 ; ebx = y1
 ; edx = y2
 ; ecx = color
 ; edi = force ?
-     push    ebx edx
+     cmp     eax, [Screen_Max_X]
-     cmp     edx, ebx      ; make sure y2 is above y1
+     jge     .out
+     push    eax ebp esi ebx edx
+     mov     ebp, [_display.width]	; ebp = screen co-ords base
+     imul    ebp, ebx
+     add     ebp, [_WinMapAddress]
+     add     ebp, eax
+
+     cmp     edx, ebx			; to make sure y2 > y1
      jge     @f
      xchg    ebx, edx
-align   4
 @@:
-     call    [putpixel]
+     cmp     ebx, [Screen_Max_Y]
+     jge     .exit
+     push    ebx
+     imul    ebx, [BytesPerScanLine]
+     shl     eax, 1
+     shl     eax, 1
+     add     eax, ebx
+     add     eax, LFB_BASE
+     pop     ebx        		; restore ebx = y1
+     cmp     edx, [Screen_Max_Y]	; the last check
+     jb     .draw
+     mov     edx, [Screen_Max_Y]	; to prevent off-screen drawing
+
+.draw:	  ; (vertical line itself)
+     test    edi,1			; forced  ?
+     jnz    .checked
+; check whether the line covered by other windows
+     movzx   esi, byte [ebp]
+     cmp     esi, [CURRENT_TASK]
+     jne     .nextpixel
+.checked:
+     test    ecx,0x01000000
+     jz     .noneg
+     mov     ecx, [eax]
+     not     ecx
+     and     ecx, 0x01FFFFFF		; keep bit[24] high !
+.noneg:
+     mov     [eax], ecx
+.nextpixel:
+     add     eax, [BytesPerScanLine]
+     add     ebp, [_display.width]
      inc     ebx
      cmp     ebx, edx
-     jle     @b
+     jle    .draw
-     pop     edx ebx
+.exit:
+     pop     edx ebx esi ebp eax
+.out:
      ret
 
 
@@ -570,9 +539,9 @@ vesa20_drawbar:
      sub     esp, drbar.stack_data
      mov     [drbar.color], edi
      sub     edx, ebx
-     jle     .exit          ;// mike.dld, 2005-01-29
+     jle     .exit	    
      sub     ecx, eax
-     jle     .exit          ;// mike.dld, 2005-01-29
+     jle     .exit	    
      mov     [drbar.bar_sy], edx
      mov     [drbar.bar_sx], ecx
      mov     [drbar.bar_cx], eax
@@ -584,13 +553,11 @@ vesa20_drawbar:
      mov     [drbar.abs_cy], ebx
 ; real_sx = MIN(wnd_sx-bar_cx, bar_sx);
      mov     ebx, [edi-twdw + WDATA.box.width] ; ebx = wnd_sx
-; \begin{diamond}[20.08.2006]
 ; note that WDATA.box.width is one pixel less than real window x-size
      inc     ebx
-; \end{diamond}[20.08.2006]
      sub     ebx, [drbar.bar_cx]
      ja      @f
-.exit:                       ;// mike.dld, 2005-01-29
+.exit:			     
      add     esp, drbar.stack_data
      popad
      xor     eax, eax
@@ -604,9 +571,7 @@ vesa20_drawbar:
      mov     [drbar.real_sx], ebx
 ; real_sy = MIN(wnd_sy-bar_cy, bar_sy);
      mov     ebx, [edi-twdw + WDATA.box.height] ; ebx = wnd_sy
-; \begin{diamond}[20.08.2006]
      inc     ebx
-; \end{diamond}
      sub     ebx, [drbar.bar_cy]
      ja      @f
      add     esp, drbar.stack_data
@@ -627,9 +592,8 @@ vesa20_drawbar:
      mov     [drbar.line_inc_map], eax
 ; line_inc_scr
      mov     eax, [drbar.real_sx]
-     movzx   ebx, byte [ScreenBPP]
+     shl     eax, 1
-     shr     ebx, 3
+     shl     eax, 1
-     imul    eax, ebx
      neg     eax
      add     eax, [BytesPerScanLine]
      mov     [drbar.line_inc_scr], eax
@@ -637,9 +601,8 @@ vesa20_drawbar:
      mov     edx, [drbar.abs_cy]
      imul    edx, [BytesPerScanLine]
      mov     eax, [drbar.abs_cx]
-;     movzx   ebx, byte [ScreenBPP]
+     shl     eax, 1
-;     shr     ebx, 3
+     shl     eax, 1
-     imul    eax, ebx
      add     edx, eax
 ; pointer to pixel map
      mov     eax, [drbar.abs_cy]
@@ -650,12 +613,8 @@ vesa20_drawbar:
      xchg    eax, ebp
 ; get process number
      mov     ebx, [CURRENT_TASK]
-     cmp     byte [ScreenBPP], 24
+
-     jne     draw_bar_end_32
+draw_bar_end_32:
-draw_bar_end_24:
-     mov     eax, [drbar.color]    ;; BBGGRR00
-     mov     bh, al                ;; bh  = BB
-     shr     eax, 8                ;; eax = RRGG
 ; eax - color high   RRGG
 ; bl - process num
 ; bh - color low    BB
@@ -663,42 +622,6 @@ draw_bar_end_24:
 ; edx - pointer to screen
 ; esi - counter
 ; edi - counter
-     mov     esi, [drbar.real_sy]
-align   4
-.new_y:
-     mov     edi, [drbar.real_sx]
-align   4
-.new_x:
-     cmp     byte [ebp], bl
-     jne     .skip
-
-     mov     [LFB_BASE+edx], bh
-     mov     [LFB_BASE+edx + 1], ax
-.skip:
-; add pixel
-     add     edx, 3
-     inc     ebp
-     dec     edi
-     jnz     .new_x
-; add line
-     add     edx, [drbar.line_inc_scr]
-     add     ebp, [drbar.line_inc_map]
-; <Ivan 15.10.04> drawing gradient bars
-     test    eax, 0x00800000
-     jz      @f
-     test    bh, bh
-     jz      @f
-     dec     bh
-@@:
-; </Ivan 15.10.04>
-     dec     esi
-     jnz     .new_y
-     add     esp, drbar.stack_data
-     popad
-     xor     eax, eax
-     ret
-
-draw_bar_end_32:
      mov     eax, [drbar.color]    ;; BBGGRR00
      mov     esi, [drbar.real_sy]
 align	4
@@ -731,8 +654,10 @@ align   4
      jnz     .new_y
      add     esp, drbar.stack_data
      popad
+     xor     eax, eax
      ret
 
+
 align 4
 vesa20_drawbackground_tiled:
 	call	[_display.disable_mouse]
@@ -749,10 +674,7 @@ dp2:
 	add	ebp, eax
 	add	ebp, eax
 	add	ebp, eax
-        cmp     [ScreenBPP], byte 24    ; 24 or 32 bpp ? - x size
-        jz      @f
 	add	ebp, eax
-@@:
 	add	ebp, LFB_BASE
 ; ebp:=Y*BytesPerScanLine+X*BytesPerPixel+AddrLFB
 	call	calculate_edi
@@ -796,9 +718,7 @@ nbgp:
 	add	esi, 3
 	add	edi, 3
 @@:
-        cmp     [ScreenBPP], byte 25    ; 24 or 32 bpp?
+	inc	edi   	 	; +1 for 32 bpp
-        sbb     edi, -1         ; +1 for 32 bpp
-; I do not use 'inc eax' because this is slightly slower then 'add eax,1'
 	add	ebp, edx
 	add	eax, edx
 	cmp	eax, [draw_data+32+RECT.right]
@@ -855,10 +775,8 @@ vesa20_drawbackground_stretch:
 	add	ebp, eax
 	add	ebp, eax
 	add	ebp, eax
-        cmp     [ScreenBPP], byte 24    ; 24 or 32 bpp ? - x size
-        jz      @f
 	add	ebp, eax
-@@:
+
 ; ebp:=Y*BytesPerScanLine+X*BytesPerPixel+AddrLFB
 	call	calculate_edi
 	xchg	edi, ebp
@@ -941,11 +859,10 @@ sdp3a:
 
 	mov	[LFB_BASE+edi+2], al
 snbgp:
-        cmp     [ScreenBPP], byte 25
+	add	edi, 4
-        sbb     edi, -4
+	inc	ebp
-        add     ebp, 1
 	mov	eax, [esp+20]
-        add     eax, 1
+	inc	eax
 	mov	[esp+20], eax
 	add	esi, 4
 	cmp	eax, [draw_data+32+RECT.right]
@@ -953,7 +870,7 @@ snbgp:
 sdp4:
 ; next y
 	mov	ebx, [esp+24]
-        add     ebx, 1
+	inc	ebx
 	mov	[esp+24], ebx
 	cmp	ebx, [draw_data+32+RECT.bottom]
 	ja	sdpdone
@@ -961,14 +878,11 @@ sdp4:
 	sub	eax, [draw_data+32+RECT.left]
 	sub	ebp, eax
 	add	ebp, [Screen_Max_X]
-        add     ebp, 1
+	inc	ebp
 	sub	edi, eax
 	sub	edi, eax
 	sub	edi, eax
-        cmp     [ScreenBPP], byte 24
-        jz      @f
 	sub	edi, eax
-@@:
 	add	edi, [BytesPerScanLine]
 ; restore ecx,edx; advance esi to next background line
 	mov	eax, [esp+28]
@@ -1017,7 +931,7 @@ smooth_line:
 @@:
 	stosd
 	mov	eax, [esp+20+8]
-        add     eax, 1
+	inc	eax
 	mov	[esp+20+8], eax
 	cmp	eax, [draw_data+32+RECT.right]
 	ja	@f
@@ -1035,38 +949,6 @@ smooth_line:
 
 align 16
 overlapping_of_points:
-if 0
-; this version of procedure works, but is slower than next version
-        push    ecx edx
-        mov     edx, eax
-        push    esi
-        shr     ecx, 24
-        mov     esi, ecx
-        mov     ecx, ebx
-        movzx   ebx, dl
-        movzx   eax, cl
-        sub     eax, ebx
-        movzx   ebx, dh
-        imul    eax, esi
-        add     dl, ah
-        movzx   eax, ch
-        sub     eax, ebx
-        imul    eax, esi
-        add     dh, ah
-        ror     ecx, 16
-        ror     edx, 16
-        movzx   eax, cl
-        movzx   ebx, dl
-        sub     eax, ebx
-        imul    eax, esi
-        pop     esi
-        add     dl, ah
-        mov     eax, edx
-        pop     edx
-        ror     eax, 16
-        pop     ecx
-        ret
-else
 	push	ecx edx
 	mov	edx, eax
 	push	esi
@@ -1094,7 +976,6 @@ else
 	ror	eax, 16
 	pop	ecx
 	ret
-end if
 
 iglobal
 align 4