x86p_32.asm

网络MPEG4IP流媒体开发源代码

.L5:	    
	mov [edi+ecx*4-4],eax
.L6:	
	mov eax,[esi+ecx*8]

        shr ah,2
        mov ebx,[esi+ecx*8+4]

        shr eax,3
        mov edx,[esi+ecx*8+4]

        shr bh,2
        mov dl,[esi+ecx*8+2]

        shl ebx,13
        and eax,000007FFh
        
        shl edx,8
        and ebx,07FF0000h

        and edx,0F800F800h
        add eax,ebx

        add eax,edx
        inc ecx

        jnz .L5                 

	mov [edi+ecx*4-4],eax

    ; tail
	pop ecx
	test cl,1
	jz .L7
	
	mov bl,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov ah,[esi+2]    ; red
	shr ah,3
	and al,11111100b
	shl eax,3
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2

.L7:	
	jmp _x86return



	
;; 32 bit RGB 888 to 16 BIT BGR 565 

_ConvertX86p32_16BGR565:
	
	; check short
	cmp ecx,BYTE 16
	ja .L3

.L1	; short loop
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111100b
	shl eax,3
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx
	jnz .L1
.L2
	jmp _x86return

.L3	; head
	mov ebx,edi
	and ebx,BYTE 11b
	jz .L4   
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111100b
	shl eax,3
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx

.L4	; save count
	push ecx

	; unroll twice
	shr ecx,1
    
	; point arrays to end
	lea esi,[esi+ecx*8]
	lea edi,[edi+ecx*4]

	; negative count
	neg ecx
	jmp SHORT .L6

.L5     
	mov [edi+ecx*4-4],eax            
.L6     
	mov edx,[esi+ecx*8+4]

        mov bh,[esi+ecx*8+4]                       
        mov ah,[esi+ecx*8]                       

        shr bh,3
        mov al,[esi+ecx*8+1]             

        shr ah,3
        mov bl,[esi+ecx*8+5]           

        shl eax,3
        mov dl,[esi+ecx*8+2]

        shl ebx,19
        and eax,0000FFE0h              
                
        shr edx,3
        and ebx,0FFE00000h             
        
        and edx,001F001Fh               
        add eax,ebx

        add eax,edx
        inc ecx

        jnz .L5                 

	mov [edi+ecx*4-4],eax            

	; tail
	pop ecx
	and ecx,BYTE 1
	jz .L7
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111100b
	shl eax,3
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2

.L7 
	jmp _x86return


	
	
;; 32 BIT RGB TO 16 BIT RGB 555

_ConvertX86p32_16RGB555:

	; check short
	cmp ecx,BYTE 16
	ja .L3

.L1	; short loop
	mov bl,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov ah,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx
	jnz .L1
.L2
	jmp _x86return

.L3	; head
	mov ebx,edi
        and ebx,BYTE 11b
	jz .L4   
	mov bl,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov ah,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx

.L4	; save count
	push ecx

	; unroll twice
	shr ecx,1
    
	; point arrays to end
	lea esi,[esi+ecx*8]
	lea edi,[edi+ecx*4]

	; negative counter 
	neg ecx
	jmp SHORT .L6

.L5     
	mov [edi+ecx*4-4],eax
.L6     
	mov eax,[esi+ecx*8]

        shr ah,3
        mov ebx,[esi+ecx*8+4]

        shr eax,3
        mov edx,[esi+ecx*8+4]

        shr bh,3
        mov dl,[esi+ecx*8+2]

        shl ebx,13
        and eax,000007FFh
        
        shl edx,7
        and ebx,07FF0000h

        and edx,07C007C00h
        add eax,ebx

        add eax,edx
        inc ecx

        jnz .L5                 

	mov [edi+ecx*4-4],eax

	; tail
	pop ecx
	and ecx,BYTE 1
	jz .L7
	mov bl,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov ah,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2

.L7
	jmp _x86return




;; 32 BIT RGB TO 16 BIT BGR 555
	
_ConvertX86p32_16BGR555:
	
	; check short
	cmp ecx,BYTE 16
	ja .L3


.L1	; short loop
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx
	jnz .L1
.L2 
	jmp _x86return

.L3	; head
	mov ebx,edi
        and ebx,BYTE 11b
	jz .L4   
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2
	dec ecx

.L4	; save count
	push ecx

	; unroll twice
	shr ecx,1
    
	; point arrays to end
	lea esi,[esi+ecx*8]
	lea edi,[edi+ecx*4]

	; negative counter 
	neg ecx
	jmp SHORT .L6

.L5     
	mov [edi+ecx*4-4],eax            
.L6     
	mov edx,[esi+ecx*8+4]

        mov bh,[esi+ecx*8+4]                       
        mov ah,[esi+ecx*8]                       

        shr bh,3
        mov al,[esi+ecx*8+1]             

        shr ah,3
        mov bl,[esi+ecx*8+5]           

        shl eax,2
        mov dl,[esi+ecx*8+2]

        shl ebx,18
        and eax,00007FE0h              
                
        shr edx,3
        and ebx,07FE00000h             
        
        and edx,001F001Fh               
        add eax,ebx

        add eax,edx
        inc ecx

        jnz .L5                 

	mov [edi+ecx*4-4],eax            

	; tail
	pop ecx
	and ecx,BYTE 1
	jz .L7
	mov ah,[esi+0]    ; blue
	mov al,[esi+1]    ; green
	mov bl,[esi+2]    ; red
	shr ah,3
	and al,11111000b
	shl eax,2
	shr bl,3
	add al,bl
	mov [edi+0],al
	mov [edi+1],ah
	add esi,BYTE 4
	add edi,BYTE 2

.L7
	jmp _x86return




	
;; FROM 32 BIT RGB to 8 BIT RGB (rrrgggbbb)
;; This routine writes FOUR pixels at once (dword) and then, if they exist
;; the trailing three pixels
_ConvertX86p32_8RGB332:

	
.L_ALIGNED
	push ecx

	shr ecx,2		; We will draw 4 pixels at once
	jnz .L1
	
	jmp .L2			; short jump out of range :(
	
.L1:
	mov eax,[esi]		; first pair of pixels
	mov edx,[esi+4]

	shr dl,6
	mov ebx,eax

	shr al,6
	and ah,0e0h

	shr ebx,16
	and dh,0e0h
	
	shr ah,3
	and bl,0e0h

	shr dh,3
	
	or al,bl
	
	mov ebx,edx	
	or al,ah
	
	shr ebx,16
	or dl,dh

	and bl,0e0h
	
	or dl,bl

	mov ah,dl

	
		
	mov ebx,[esi+8]		; second pair of pixels

	mov edx,ebx
	and bh,0e0h

	shr bl,6
	and edx,0e00000h

	shr edx,16

	shr bh,3

	ror eax,16
	or bl,dl

	mov edx,[esi+12]
	or bl,bh
	
	mov al,bl

	mov ebx,edx
	and dh,0e0h

	shr dl,6
	and ebx,0e00000h
	
	shr dh,3
	mov ah,dl

	shr ebx,16
	or ah,dh

	or ah,bl

	rol eax,16
	add esi,BYTE 16
			
	mov [edi],eax	
	add edi,BYTE 4
	
	dec ecx
	jz .L2			; L1 out of range for short jump :(
	
	jmp .L1
.L2:
	
	pop ecx
	and ecx,BYTE 3		; mask out number of pixels to draw
	
	jz .L4			; Nothing to do anymore

.L3:
	mov eax,[esi]		; single pixel conversion for trailing pixels

        mov ebx,eax

        shr al,6
        and ah,0e0h

        shr ebx,16

        shr ah,3
        and bl,0e0h

        or al,ah
        or al,bl

        mov [edi],al

        inc edi
        add esi,BYTE 4

	dec ecx
	jnz .L3
	
.L4:	
	jmp _x86return