kolibrios-gitea/kernel/trunk/video/blitter.inc
CleverMouse 4d704be765 speedup software blitter for 32bpp modes
git-svn-id: svn://kolibrios.org@4444 a494cfbc-eb01-0410-851d-a64ba20cac60
2014-01-13 11:16:28 +00:00

465 lines
11 KiB
PHP

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;; ;;
;; Copyright (C) KolibriOS team 2011-2012. All rights reserved. ;;
;; Distributed under terms of the GNU General Public License ;;
;; ;;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
struct BLITTER_BLOCK
xmin dd ?
ymin dd ?
xmax dd ?
ymax dd ?
ends
struct BLITTER
dc RECT
sc RECT
dst_x dd ? ; 32
dst_y dd ? ; 36
src_x dd ? ; 40
src_y dd ? ; 44
w dd ? ; 48
h dd ? ; 52
bitmap dd ? ; 56
stride dd ? ; 60
ends
align 4
block_clip:
;esi= clip RECT ptr
;edi= RECT ptr
;return code:
;eax= 0 - draw, 1 - don't draw
push ebx
mov eax, [edi+RECT.left]
mov ebx, [edi+RECT.right]
mov ecx, [esi+RECT.left] ;clip.left
mov edx, [esi+RECT.right] ;clip.right
cmp eax, edx ;left >= clip.right
jge .fail
cmp ebx, ecx ;right < clip.left
jl .fail
cmp eax, ecx ;left >= clip.left
jae @F
mov eax, ecx
@@:
mov [edi+RECT.left], eax
cmp ebx, edx ;right <= clip.right
jle @f
mov ebx, edx
@@:
mov [edi+RECT.right], ebx
mov eax, [edi+RECT.top]
mov ebx, [edi+RECT.bottom]
mov ecx, [esi+RECT.top] ;clip.top
mov edx, [esi+RECT.bottom] ;clip.bottom
cmp eax, edx ;top >= clip.bottom
jge .fail
cmp ebx, ecx ;bottom < clip.top
jl .fail
cmp eax, ecx ;top >= clip.top
jae @F
mov eax, ecx
@@:
mov [edi+RECT.top], eax
cmp ebx, edx ;bottom <= clip.bottom
jle @f
mov ebx, edx
@@:
mov [edi+RECT.bottom], ebx
pop ebx
xor eax, eax
ret
.fail:
pop ebx
mov eax, 1
ret
align 4
blit_clip:
.sx0 equ 8
.sy0 equ 12
.sx1 equ 16
.sy1 equ 20
.dx0 equ 24
.dy0 equ 28
.dx1 equ 32
.dy1 equ 36
push edi
push esi
push ebx
sub esp, 40
mov ebx, ecx
mov edx, [ecx+BLITTER.src_x]
mov [esp+.sx0], edx
mov eax, [ecx+BLITTER.src_y]
mov [esp+.sy0], eax
add edx, [ecx+BLITTER.w]
add eax, [ecx+BLITTER.h]
mov [esp+.sx1], edx
mov [esp+.sy1], eax
lea edi, [esp+.sx0]
lea esi, [ebx+BLITTER.sc]
call block_clip
test eax, eax
mov esi, 1
jnz .done
mov edi, [esp+.sx0]
mov edx, [ebx+BLITTER.dst_x]
add edx, edi
sub edx, [ebx+BLITTER.src_x]
mov [esp+.dx0], edx
mov ecx, [esp+.sy0]
mov eax, [ebx+BLITTER.dst_y]
add eax, ecx
sub eax, [ebx+BLITTER.src_y]
mov [esp+.dy0], eax
sub edx, edi
add edx, [esp+.sx1]
mov [esp+.dx1], edx
sub eax, ecx
add eax, [esp+.sy1]
mov [esp+.dy1], eax
lea edi, [esp+.dx0]
lea esi, [ebx+BLITTER.dc]
call block_clip
test eax, eax
mov esi, 1
jnz .done
mov edx, [esp+.dx0]
mov eax, [esp+.dx1]
sub eax, edx
mov [ebx+BLITTER.w], eax
mov eax, [esp+.dy0]
mov ecx, [esp+.dy1]
sub ecx, eax
mov [ebx+BLITTER.h], ecx
mov ecx, [ebx+BLITTER.src_x]
add ecx, edx
sub ecx, [ebx+BLITTER.dst_x]
mov [ebx+BLITTER.src_x], ecx
mov ecx, [ebx+BLITTER.src_y]
add ecx, eax
sub ecx, [ebx+BLITTER.dst_y]
mov [ebx+BLITTER.src_y], ecx
mov [ebx+BLITTER.dst_x], edx
mov [ebx+BLITTER.dst_y], eax
xor esi, esi
.done:
mov eax, esi
add esp, 40
pop ebx
pop esi
pop edi
purge .sx0
purge .sy0
purge .sx1
purge .sy1
purge .dx0
purge .dy0
purge .dx1
purge .dy1
ret
align 4
blit_32:
push ebp
push edi
push esi
push ebx
virtual at sizeof.BLITTER
.position dd ? ; (x shl 16) + y
; ???
.extra_var1 dd ?
.local_vars_size = $
end virtual
sub esp, .local_vars_size
mov eax, [TASK_BASE]
mov ebx, [eax-twdw + WDATA.box.width]
mov edx, [eax-twdw + WDATA.box.height]
inc ebx
inc edx
xor eax, eax
mov [esp+BLITTER.dc.left], eax
mov [esp+BLITTER.dc.top], eax
mov [esp+BLITTER.dc.right], ebx
mov [esp+BLITTER.dc.bottom], edx
mov [esp+BLITTER.sc.left], eax
mov [esp+BLITTER.sc.top], eax
mov eax, [ecx+24]
mov [esp+BLITTER.sc.right], eax
mov eax, [ecx+28]
mov [esp+BLITTER.sc.bottom], eax
mov eax, [ecx]
mov [esp+BLITTER.dst_x], eax
mov eax, [ecx+4]
mov [esp+BLITTER.dst_y], eax
mov eax, [ecx+16]
mov [esp+BLITTER.src_x], eax
mov eax, [ecx+20]
mov [esp+BLITTER.src_y], eax
mov eax, [ecx+8]
mov [esp+BLITTER.w], eax
mov eax, [ecx+12]
mov [esp+BLITTER.h], eax
mov eax, [ecx+32]
mov [esp+BLITTER.bitmap], eax
mov eax, [ecx+36]
mov [esp+BLITTER.stride], eax
mov ecx, esp
call blit_clip
test eax, eax
jne .L57
mov eax, [TASK_BASE]
mov ebx, [esp+BLITTER.dst_x]
mov ebp, [esp+BLITTER.dst_y]
add ebx, [eax-twdw + WDATA.box.left]
add ebp, [eax-twdw + WDATA.box.top]
mov ecx, ebx
add ecx, [esp+BLITTER.w]
shl ecx, 16
mov cx, bp
add ecx, [esp+BLITTER.h]
mov eax, ebx
shl eax, 16
mov ax, bp
mov [esp+.position], eax
mov edi, ebp
; imul edi, [_display.pitch]
mov edi, [BPSLine_calc_area+edi*4]
; imul ebp, [_display.width]
mov ebp, [d_width_calc_area+ebp*4]
add ebp, ebx
add ebp, [_WinMapAddress]
mov eax, [esp+BLITTER.src_y]
imul eax, [esp+BLITTER.stride]
mov esi, [esp+BLITTER.src_x]
lea esi, [eax+esi*4]
add esi, [esp+BLITTER.bitmap]
mov eax, ecx
mov ecx, [esp+BLITTER.h]
mov edx, [esp+BLITTER.w]
test ecx, ecx ;FIXME check clipping
jz .L57
test edx, edx
jz .L57
cmp [_display.bpp], 32
jne .core_24
lea edi, [edi+ebx*4]
mov ebx, [CURRENT_TASK]
; check for hardware cursor
cmp [_display.select_cursor], select_cursor
je .core_32.software_cursor
cmp [_display.select_cursor], 0
jne .core_32.hardware_cursor
;--------------------------------------
.core_32.software_cursor:
align 4
.outer32:
align 4
.inner32:
cmp [ebp], bl
jne .skip
;--------------------------------------
mov eax, [esi]
mov ecx, [esp+.position]
; check mouse area for putpixel
call [_display.check_mouse]
;--------------------------------------
; store to real LFB
mov [LFB_BASE+edi], eax
;--------------------------------------
align 4
.skip:
add esi, 4
add edi, 4
inc ebp
add [esp+.position], 1 shl 16
dec edx
jnz .inner32
add esi, [esp+BLITTER.stride]
add edi, [_display.pitch]
add ebp, [_display.width]
mov edx, [esp+BLITTER.w]
mov eax, edx
inc [esp+.position]
sub ebp, edx
shl eax, 2
sub esi, eax
sub edi, eax
shl eax, 16-2
sub [esp+.position], eax
dec [esp+BLITTER.h]
jnz .outer32
jmp .done
.core_32.hardware_cursor:
align 4
.hw.outer32:
xor ecx, ecx
align 4
.hw.inner32:
cmp [ebp+ecx], bl
jne .hw.skip
mov eax, [esi+ecx*4]
mov [LFB_BASE+edi+ecx*4], eax
align 4
.hw.skip:
inc ecx
dec edx
jnz .hw.inner32
add esi, [esp+BLITTER.stride]
add edi, [_display.pitch]
add ebp, [_display.width]
mov edx, [esp+BLITTER.w]
dec [esp+BLITTER.h]
jnz .hw.outer32
.done:
; call [draw_pointer]
; call __sys_draw_pointer
.L57:
add esp, .local_vars_size
pop ebx
pop esi
pop edi
pop ebp
ret
.core_24:
lea ebx, [ebx+ebx*2]
lea edi, [LFB_BASE+edi+ebx]
mov ebx, [CURRENT_TASK]
align 4
.outer24:
mov [esp+.extra_var1], edi
xor ecx, ecx
align 4
.inner24:
cmp [ebp+ecx], bl
jne .skip_1
;--------------------------------------
push eax
mov eax, [esi+ecx*4]
lea edi, [edi+ecx*2]
; check for hardware cursor
cmp [_display.select_cursor], select_cursor
je @f
cmp [_display.select_cursor], 0
jne .no_mouseunder_1
;--------------------------------------
align 4
@@:
push ecx
mov ecx, [esp+4]
ror ecx, 16
sub ecx, edx
rol ecx, 16
sub ecx, [esp+BLITTER.h + 8]
; check mouse area for putpixel
call [_display.check_mouse]
pop ecx
;--------------------------------------
align 4
.no_mouseunder_1:
mov [edi+ecx], ax
shr eax, 16
mov [edi+ecx+2], al
pop eax
;--------------------------------------
align 4
.skip_1:
mov edi, [esp+.extra_var1]
inc ecx
dec edx
jnz .inner24
add esi, [esp+BLITTER.stride]
add edi, [_display.pitch]
add ebp, [_display.width]
mov edx, [esp+BLITTER.w]
dec [esp+BLITTER.h]
jnz .outer24
jmp .done