quantize_mpeg_xmm.asm

来自「从FFMPEG转换而来的H264解码程序,VC下编译..」· 汇编 代码 · 共 607 行 · 第 1/2 页

  psrlw mm7, 1
  add esi, byte 2
  paddw mm2, mm5        ;sum += x1
  movq [ebx], mm7       ;store x2
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  movq [ebx+8], mm2     ;store sum
  movq [edx + 8*esi+112-16], mm5
  movq [edx + 8*esi +120-16], mm7
  jng near .q1loop
  jmp near .done

ALIGN 8
.lloop
  movq mm1, [eax + 8*esi+112]       ; mm0 = [1st]
  psubw mm0,mm1         ;-mm1
  movq mm4, [eax + 8*esi+120]
  psubw mm3,mm4         ;-mm4
  pmaxsw mm0,mm1        ;|src|
  pmaxsw mm3,mm4
  nop2
  psraw mm1,15          ;sign src
  psraw mm4,15
  psllw mm0, 4          ; level << 4
  psllw mm3, 4          ;
  paddw mm0, [edi + 640 + 8*esi+112] ;mm0 is to be divided inter1 contains fix for division by 1
  paddw mm3, [edi + 640 + 8*esi+120]
  movq mm5,[edi + 896 + 8*esi+112]
  movq mm7,[edi + 896 + 8*esi+120]
  pmulhuw mm5,mm0
  pmulhuw mm7,mm3       ;mm7: first approx of division
  mov esp,esp
  movq mm2,[edi + 512 + 8*esi+112]
  movq mm6,[edi + 512 + 8*esi+120]
  pmullw mm2,mm5        ;test value <= original
  pmullw mm6,mm7
  psubw mm0,mm2         ;mismatch
  psubw mm3,mm6
  movq mm2,[byte ebx]
  movq mm6,[mmx_div + ecx * 8 - 8]  ; divs for q<=16
  pmulhuw mm0,[edi + 768 + 8*esi+112] ;correction
  pmulhuw mm3,[edi + 768 + 8*esi+120]
  paddw mm2,[ebx+8]     ;sum
  paddw mm5,mm0         ;final result
  paddw mm7,mm3
  pxor mm0,mm0
  pxor mm3,mm3
  pmulhuw mm5, mm6      ; mm0 = (mm0 / 2Q) >> 16
  pmulhuw mm7, mm6      ;  (level ) / quant (0<quant<32)
  add esi,byte 2
  psrlw mm5, 1          ; (level ) / (2*quant)
  paddw mm2,mm5         ;sum += x1
  psrlw mm7, 1
  movq [ebx],mm7        ;store x2
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  db 0Fh, 7Fh, 54h, 23h, 08 ;movq   [ebx+8],mm2 ;store sum
  movq [edx + 8*esi+112-16], mm5
  movq [edx + 8*esi +120-16], mm7
  jng near .lloop
  jmp near .done
.endfunc


;-----------------------------------------------------------------------------
;
; uint32_t dequant_mpeg_intra_3dne(int16_t *data,
;                                  const int16_t const *coeff,
;                                  const uint32_t quant,
;                                  const uint32_t dcscalar,
;                                  const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

  ;   Note: in order to saturate 'easily', we pre-shift the quantifier
  ; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to
  ; build a saturating mask. It is non-zero only when an overflow occured.
  ; We thus avoid packing/unpacking toward double-word.
  ; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g.,
  ; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not
  ; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a
  ; and quant in [1..31].
  ;

%macro DEQUANT4INTRAMMX 1
  movq mm1, [byte ecx+ 16 * %1] ; mm0 = c  = coeff[i]
  movq mm4, [ecx+ 16 * %1 +8]   ; mm3 = c' = coeff[i+1]
  psubw mm0, mm1
  psubw mm3, mm4
  pmaxsw mm0, mm1
  pmaxsw mm3, mm4
  psraw mm1, 15
  psraw mm4, 15
%if %1
  movq mm2, [eax+8] ;preshifted quant
  movq mm7, [eax+8]
%endif
  pmullw mm2, [edi + 16 * %1 ]     ; matrix[i]*quant
  pmullw mm7, [edi + 16 * %1 +8]   ; matrix[i+1]*quant
  movq mm5, mm0
  movq mm6, mm3
  pmulhw mm0, mm2   ; high of coeff*(matrix*quant)
  pmulhw mm3, mm7   ; high of coeff*(matrix*quant)
  pmullw mm2, mm5   ; low  of coeff*(matrix*quant)
  pmullw mm7, mm6   ; low  of coeff*(matrix*quant)
  pcmpgtw mm0, [eax]
  pcmpgtw mm3, [eax]
  paddusw mm2, mm0
  paddusw mm7, mm3
  psrlw mm2, 5
  psrlw mm7, 5
  pxor mm2, mm1     ; start negating back
  pxor mm7, mm4     ; start negating back
  psubusw mm1, mm0
  psubusw mm4, mm3
  movq mm0, [eax]   ;zero
  movq mm3, [eax]   ;zero
  psubw mm2, mm1    ; finish negating back
  psubw mm7, mm4    ; finish negating back
  movq [byte edx + 16 * %1], mm2   ; data[i]
  movq [edx + 16 * %1  +8], mm7   ; data[i+1]
%endmacro

ALIGN 16
dequant_mpeg_intra_3dne:
  mov eax, [esp+12] ; quant
  mov ecx, [esp+8]  ; coeff
  movq mm7, [mmx_mul_quant  + eax*8 - 8]
  psllw mm7, 2      ; << 2. See comment.
  mov edx, [esp+4]  ; data
  push ebx
  movsx ebx, word [ecx]
  pxor mm0, mm0
  pxor mm3, mm3
  push esi
  lea eax, [esp-28]
  sub esp, byte 32
  and eax, byte -8  ;points to qword ALIGNed space on stack
  movq [eax], mm0
  movq [eax+8], mm7
  imul ebx, [esp+16+8+32]    ; dcscalar
  movq mm2, mm7
  push edi
  mov edi, [esp + 32 + 12 + 20] ; mpeg_quant_matrices	
ALIGN 4

  DEQUANT4INTRAMMX 0

  mov esi, -2048
  nop
  cmp ebx, esi

  DEQUANT4INTRAMMX 1

  cmovl ebx, esi
  neg esi
  sub esi, byte 1 ;2047

  DEQUANT4INTRAMMX 2

  cmp ebx, esi
  cmovg ebx, esi
  lea ebp, [byte ebp]

  DEQUANT4INTRAMMX 3

  mov esi, [esp+36]
  mov [byte edx], bx
  mov ebx, [esp+36+4]

  DEQUANT4INTRAMMX 4
  DEQUANT4INTRAMMX 5
  DEQUANT4INTRAMMX 6
  DEQUANT4INTRAMMX 7

  pop edi

  add esp, byte 32+8

  xor eax, eax
  ret
.endfunc

;-----------------------------------------------------------------------------
;
; uint32_t dequant_mpeg_inter_3dne(int16_t * data,
;                                  const int16_t * const coeff,
;                                  const uint32_t quant,
;                                  const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

    ; Note:  We use (2*c + sgn(c) - sgn(-c)) as multiplier
    ; so we handle the 3 cases: c<0, c==0, and c>0 in one shot.
    ; sgn(x) is the result of 'pcmpgtw 0,x':  0 if x>=0, -1 if x<0.
    ; It's mixed with the extraction of the absolute value.

ALIGN 16
dequant_mpeg_inter_3dne:
  mov edx, [esp+ 4]        ; data
  mov ecx, [esp+ 8]        ; coeff
  mov eax, [esp+12]        ; quant
  movq mm7, [mmx_mul_quant  + eax*8 - 8]
  mov eax, -14
  paddw mm7, mm7    ; << 1
  pxor mm6, mm6     ; mismatch sum
  push esi
  push edi
  mov esi, mmzero
  pxor mm1, mm1
  pxor mm3, mm3
  mov edi, [esp + 8 + 16] ; mpeg_quant_matrices
  nop
  nop4

ALIGN 16
.loop
  movq mm0, [ecx+8*eax + 7*16   ]   ; mm0 = coeff[i]
  pcmpgtw mm1, mm0  ; mm1 = sgn(c)    (preserved)
  movq mm2, [ecx+8*eax + 7*16 +8]   ; mm2 = coeff[i+1]
  pcmpgtw mm3, mm2  ; mm3 = sgn(c')   (preserved)
  paddsw mm0, mm1   ; c += sgn(c)
  paddsw mm2, mm3   ; c += sgn(c')
  paddw mm0, mm0    ; c *= 2
  paddw mm2, mm2    ; c'*= 2

  movq mm4, [esi]
  movq mm5, [esi]
  psubw mm4, mm0    ; -c
  psubw mm5, mm2    ; -c'

  psraw mm4, 16     ; mm4 = sgn(-c)
  psraw mm5, 16     ; mm5 = sgn(-c')
  psubsw mm0, mm4   ; c  -= sgn(-c)
  psubsw mm2, mm5   ; c' -= sgn(-c')
  pxor mm0, mm1     ; finish changing sign if needed
  pxor mm2, mm3     ; finish changing sign if needed

 ; we're short on register, here. Poor pairing...

  movq mm4, mm7     ; (matrix*quant)
  nop
  pmullw mm4, [edi + 512 + 8*eax + 7*16]
  movq mm5, mm4
  pmulhw mm5, mm0   ; high of c*(matrix*quant)
  pmullw mm0, mm4   ; low  of c*(matrix*quant)

  movq mm4, mm7     ; (matrix*quant)
  pmullw mm4, [edi + 512 + 8*eax + 7*16 + 8]
  add eax, byte 2

  pcmpgtw mm5, [esi]
  paddusw mm0, mm5
  psrlw mm0, 5
  pxor mm0, mm1     ; start restoring sign
  psubusw mm1, mm5

  movq mm5, mm4
  pmulhw mm5, mm2   ; high of c*(matrix*quant)
  pmullw mm2, mm4   ; low  of c*(matrix*quant)
  psubw mm0, mm1    ; finish restoring sign

  pcmpgtw mm5, [esi]
  paddusw mm2, mm5
  psrlw mm2, 5
  pxor mm2, mm3     ; start restoring sign
  psubusw mm3, mm5
  psubw mm2, mm3    ; finish restoring sign
  movq mm1, [esi]
  movq mm3, [byte esi]
  pxor mm6, mm0                             ; mismatch control
  movq [edx + 8*eax + 7*16 -2*8   ], mm0    ; data[i]
  pxor mm6, mm2                             ; mismatch control
  movq [edx + 8*eax + 7*16 -2*8 +8], mm2    ; data[i+1]

  jng .loop
  nop

 ; mismatch control

  pshufw mm0, mm6, 01010101b
  pshufw mm1, mm6, 10101010b
  pshufw mm2, mm6, 11111111b
  pxor mm6, mm0
  pxor mm1, mm2
  pxor mm6, mm1
  movd eax, mm6
  pop edi
  and eax, byte 1
  xor eax, byte 1
  mov esi, [esp]
  add esp, byte 4
  xor word [edx + 2*63], ax

  xor eax, eax
  ret
.endfunc