dehuff.asm

TMS320VC55XX 图像 处理程序

	AC1 = AC1 + (@(p_arg4) * T3)

	@(d_pass1) = AC1                            ; PASS1 = UvldTabDC[htype]
	XAR0 = #_lowmask                
	call #_uvld                                 ; s = Uvld()
	
	if (AC0 == #0) goto #storedc  
	
	@(d_pass0) = AC0                            ; PASS0 = s (PASS1 since SP-- is already done)

	call #_getdctcoeff
	
storedc:
	XAR2 = dbl(@(p_arg2))                       ; AR2 = lastdc
	XAR4 = dbl(@(p_uvldlenmaxac))
	T1 = @(p_arg4)
	T3 = #UVLDACTABSIZE
	AC1 = @(p_uvldtabac+1)
	AC1 = AC1 + (@(p_arg4) * T3)
	@(d_pass1) = AC1                            ; PASS1 = UvldTabAC[htype]

	AC0 = AC0 + *AR2
	*AR2 = AC0                                  ; *lastdc += GetDctCoeff()
	|| *AR7+ = AC0                              ; *out_p+ = *lastdc
	
;; AC decoding
      AR2=@InputAD
     ;XAR2 = #0x6000                                           ; set up function arguments to pass
    ;XAR2 = dbl(*AR5(#(4)))
	AR6 = #63                                   ; AR6: coefficient counter
	BRC0 = AR6
	AC1 = *AR4(T1)
	@(d_tmp) = AC1
	
	AC1 = @(p_uvldctltabac+1)
	AC1 = AC1 + (@(p_arg4) * T0)
	@(d_pass2) = AC1                            ; PASS2 = UvldCtlTabAC[htype]
    bit(ST2, #ST2_AR2LC) = #1
	AC1 = @(d_tmp)
	|| blockrepeat {
	@(d_pass0) = AC1                            ; CAREFUL! Use PASS1 since SP-- is already done
	
	XAR4 = dbl(@(p_uvldctltabac))                
    AR2 = *AR5(#(1))                            ; *AR5(1) = buf_ptr, A = stream->databuf[buf_ptr] 	
	AC1 = @(d_pass0) 
    AC1 = AC1 - *AR5                            ; B = nrbits-bit_ptr    
	AR3 = AR0 + @(d_pass0)                      ; A = nrbits;    
    AC0 = *AR2                                  ; A = stream->databuf[buf_ptr] 
 	
	if (AC1 <= #0) goto #($7)                   ; if (-bit_ptr>=0) goto$2
                                                ; if (bit_ptr <=0) 
    AC1 = AC1 + #(-16)                          ; -bit_ptr+=-16
	AC0 = *AR2+ << #16                          ; buffer=stream->databuf[buf_ptr]<<16
	AC0 = AC0 | *AR2                            ; buffer+=stream->databuf[++buf_ptr]

$7:	
	T3 = AC1                                    ; T = B
	|| AR4 = @(d_pass2)                         ; B = dcdctl;
	AC0 = AC0 << T3
	AC0 = AC0 ^ #(-1)                           ; reversed
	AC0 = AC0 & *AR3                            ; A = (buffer>>bit_ptr)&mask[nrbits]
    T3 = exp(AC0)                               ; index = 31 - T
    AR4 = AR4 - T3                              ; B = dcdctl-(31-index);
                                                ; get index
	@(d_dat) = AC0                              ; code stored in *AR5(-1)
	AC0 = uns(*AR4(#(31)))                      ; pattern = dcdctl[index]
                                                ; get coding pattern
	AR1 = AC0 & #7ffh                           ; offset = pattern&0x07ff;
	AC0 = AC0 <<< #-11                          ; shift  = (pattern>>11)&0xf
	AC0 = -AC0
	T3 = AC0                                    ; T = -shift	
	|| AC2 = uns(@(d_dat))
	AC2 = AC2 << T3
	AR1 = AR1 + AC2                 
	AR1 = AR1 + @(d_pass1)                      ; AR4 = dcdtab+code
	AC0 = uns(*AR1)                             ; pattern = dcdtab[code];

    AC1 = AC0 <<< #-11                          ; nrbits = (pattern>>11)&0x1f
                                                ; flush_bits(stream, nrbits)
	AC1 = AC1 - *AR5
	AC1 = -AC1                                  ; stream->bit_ptr -=nrbits
 
 	AC0 = AC0 & #07ffh                          ; pattern&0x07ff
	
	if (AC1 <= #0) execute (AD_unit)
	*AR5(#(1)) = *AR5(#(1)) + #1                ; stream->buf_ptr++;
	if (AC1 <= #0) execute (AD_unit)
	*AR5(#(1)) = *AR5(#(1)) & #(MCUBUFFINLEN-1)
	if (AC1 <= #0) execute (D_unit)
	|| AC1 = AC1 + #16                          ; stream->bit_ptr+= c16;
   	*AR5 = AC1                                  ; store stream->bit_ptr
                                                ; if s = 0x00 (EOB)    
    if (AC0 == #0) goto #endjpegdechuff
	AC2 = AC0 - #0xf0   
	AC1 = AC0 <<C #-4
	AR6 = AR6 - AC1                             ; AR6 -= s>>4
    XAR7 = #(unzigzag_r)
    T1 = AR6                                    ; AR7 = i (loop index)
	AR6 = AR6 - #1
	BRC0 = AR6
	if (AC2 == #0) goto $0xF0
                                                ; if s is regular r/s
	AC0 = AC0 & #0fh
	@(d_pass0) = AC0                            ; PASS0 = s

	AC1 = *AR7(T1)                              ; B = unzigzag_r[i]
	AC1 = AC1 + dbl(@(p_arg3))                  ; out_p += unzigzag_r[i]
 
 	XAR7 = AC1                                  ; AR7 = out_p

	XAR4 = XAR0
	AC0 = *AR5                                  ; *AR5(0) = bit_ptr
    AC2 = AR0 + @(d_pass0)                      ; A = lowmask + ssss
	AR2 = *AR5(#(1))   
	AC0 = AC0 - @(d_pass0)                      ; A = bit_ptr - ssss
    *AR5 = AC0
	AC0 = -AC0    
	|| AR4 = AC2         						; AR4 = &lowmask[ssss]
    T3 = AC0                                    ; set up T for NORM B
	AC1 = *AR2+ << #16                          ; BH = current word 
	AC1 = AC1 | *AR2                            ; BL = next word	
	if (AC0 < #0) goto #getvalue2 		
	
	*AR5(#(1)) = *AR5(#(1)) + #1                ; buf_ptr++
	*AR5(#(1)) = *AR5(#(1)) & #(MCUBUFFINLEN-1)
	AC0 = AC0 + #(-16)
	AC0 = -AC0
	*AR5 = AC0                                  ; store bit_ptr

getvalue2:
	XAR3 = #(_bitmask)
	T1 = @(d_pass0)
	AC1 = AC1 << T3
	AC1 = AC1
    AC0 = *AR4                                  ; *AR4 -> lowmask[ssss]
	AC0 = AC0 & (AC1 <<< #-16)

	AC1 = *AR3(T1)
	
	AC1 = AC1 & AC0

	if (AC1 == #0) execute (AD_unit)
	|| AC0 = AC0 ^ #(-1)
	if (AC1 == #0) execute (AD_unit)
	|| 	AC0 = AC0 & *AR4
	if (AC1 == #0) execute (AD_unit)
	|| 	AC0 = -AC0

 	*AR7+ = AC0                                 ; *out_p+ = A
$0xF0:	                                        ; if s = 0xF0 (ZRL)
	AC1 = @(d_tmp)
	}	

endjpegdechuff:
	BSA23 = #0
	bit(ST2, #ST2_AR2LC) = #0
	bit(ST2, #ST2_ARMS) = #1
	bit(ST1, #ST1_CPL) = #1 
    SP = SP + #2
	AR6, T3 = pop()
	XAR5 = popboth()
	XAR7 = popboth()
    return

	.align	4
_getdctcoeff:
	AC0 = *AR5                                  ; *AR5(0) = bit_ptr
	XAR4 = XAR0
    AC2 = AR0 + @(d_pass0)                      ; A = lowmask + ssss
	;XAR2 = dbl(*AR5(#(4)))
	 AR2=@InputAD
	;XAR2= #0x6000
	AR2 = *AR5(#(1))                            ; B = &(stream_p->databuf[buf_ptr])			
	AC0 = AC0 - @(d_pass0)                      ; A = bit_ptr - ssss
    *AR5 = AC0
	AC0 = -AC0    
	|| AR4 = AC2         						; AR4 = &lowmask[ssss]
    T3 = AC0                                    ; set up T for NORM B
	AC1 = *AR2+ << #16                          ; BH = current word 
	|| circular()
	AC1 = AC1 | *AR2                            ; BL = next word	
	|| circular()
	if (AC0 < #0) goto #getvalue 		
	
	*AR5(#(1)) = *AR5(#(1)) + #1                ; buf_ptr++
	*AR5(#(1)) = *AR5(#(1)) & #(MCUBUFFINLEN-1)
	AC0 = AC0 + #(-16)
	AC0 = -AC0
	*AR5 = AC0                                  ; store bit_ptr

getvalue:
	XAR3 = #(_bitmask)
	T1 = @(d_pass0)
	AC1 = AC1 << T3
	AC1 = AC1
    AC0 = *AR4                                  ; *AR4 -> lowmask[ssss]
	AC0 = AC0 & (AC1 <<< #-16)

	AC1 = *AR3(T1)
	
	AC1 = AC1 & AC0

	if (AC1 == #0) execute (AD_unit)
	|| AC0 = AC0 ^ #(-1)
	if (AC1 == #0) execute (AD_unit)
	|| 	AC0 = AC0 & *AR4
	if (AC1 == #0) execute (AD_unit)
	|| 	AC0 = -AC0

	return


	
	.align	4
_uvld:
    XAR3 = XAR0
   ; XAR2 = dbl(*AR5(#(4)))
    ; XAR2 = dbl(*AR5(#(2)))
     AR2=@InputAD
    ;XAR2=#0x6000
   	BSA23 = @AR2_L
	|| mmap()
    XAR2=#0
    AR2 = *AR5(#(1))                            ; *AR5(1) = buf_ptr,
	AC1 = @(d_pass0) 
    AC1 = AC1 - *AR5                            ; B = nrbits-bit_ptr    
    ; AC0=#0
     AC0 = *AR2                                  ; A = stream->databuf[buf_ptr] 
    || circular()
	AR3 = AR0 + @(d_pass0)                      ; A = nrbits;    
    || nop
                                                ; core part of the kernel
	if (AC1 <= #0) goto #($9)                   ; if (-bit_ptr>=0) goto$2
                                                ; if (bit_ptr <=0) 
    AC1 = AC1 + #(-16)                          ; -bit_ptr+=-16
	AC0 = *AR2+ << #16                          ; buffer=stream->databuf[buf_ptr]<<16
	|| circular()
	AC0 = AC0 | *AR2                            ; buffer+=stream->databuf[++buf_ptr]
	|| circular()

$9:	
	T3 = AC1                                    ; T = B
	|| AR4 = @(d_pass2)                         ; B = dcdctl;
	AC0 = AC0 << T3
	AC0 = AC0 ^ #(-1)                           ; reversed
	AC0 = AC0 & *AR3                            ; A = (buffer>>bit_ptr)&mask[nrbits]
    T3 = exp(AC0)                               ; index = 31 - T
    AR4 = AR4 - T3                              ; B = dcdctl-(31-index);
                                                ; get index
	@(d_dat) = AC0                                  ; code stored in *AR5(-1)
	AC0 = uns(*AR4(#(31)))                      ; pattern = dcdctl[index]
                                                ; get coding pattern
	AR1 = AC0 & #7ffh                           ; offset = pattern&0x07ff;
	AC0 = AC0 <<< #-11                          ; shift  = (pattern>>11)&0xf
	AC0 = -AC0
	T3 = AC0                                    ; T = -shift	
	|| AC2 = uns(@(d_dat))
	AC2 = AC2 << T3
	AR1 = AR1 + AC2
	AR1 = AR1 + @(d_pass1)                      ; AR4 = dcdtab+code
	AC0 = uns(*AR1)                             ; pattern = dcdtab[code];

    AC1 = AC0 <<< #-11                          ; nrbits = (pattern>>11)&0x1f
                                                ; flush_bits(stream, nrbits)
	AC1 = AC1 - *AR5
	AC1 = -AC1                                  ; stream->bit_ptr -=nrbits
 
 	AC0 = AC0 & #07ffh                          ; pattern&0x07ff
	
	if (AC1 <= #0) execute (AD_unit)
	*AR5(#(1)) = *AR5(#(1)) + #1                ; stream->buf_ptr++;
	if (AC1 <= #0) execute (AD_unit)
	*AR5(#(1)) = *AR5(#(1)) & #(MCUBUFFINLEN-1)
	if (AC1 <= #0) execute (D_unit)
	|| AC1 = AC1 + #16                          ; stream->bit_ptr+= c16;
   	*AR5 = AC1                                  ; store stream->bit_ptr

    return
	

    .end