quantize_h263_mmx.asm

wince下的xvidcore开发库,可用于MP4等视频播放开发

  movlhps xmm6, xmm6                        ; duplicate into high 8 bytes

  cmp al, 1
  jz near .qes2_q1loop

.qes2_not1
  movq mm0, [mmx_div + eax*8 - 8]           ; divider
  movq2dq xmm7, mm0
  movlhps xmm7, xmm7

ALIGN 16
.qes2_loop
  movdqa xmm0, [esi + ecx*8]                ; xmm0 = [1st]
  movdqa xmm3, [esi + ecx*8 + 16]           ; xmm3 = [2nd]
  pxor xmm1, xmm1
  pxor xmm4, xmm4
  pcmpgtw xmm1, xmm0
  pcmpgtw xmm4, xmm3
  pxor xmm0, xmm1
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  psubusw xmm0, xmm6
  psubusw xmm3, xmm6
  pmulhw xmm0, xmm7
  pmulhw xmm3, xmm7
  paddw xmm5, xmm0
  pxor xmm0, xmm1
  paddw xmm5, xmm3
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  movdqa [edi + ecx*8], xmm0
  movdqa [edi + ecx*8 + 16], xmm3

  add ecx, 4
  cmp ecx, 16
  jnz .qes2_loop

.qes2_done
  movdqu xmm6, [plus_one]
  pmaddwd xmm5, xmm6
  movhlps xmm6, xmm5
  paddd xmm5, xmm6
  movdq2q mm0, xmm5

  movq mm5, mm0
  psrlq mm5, 32
  paddd mm0, mm5

  movd eax, mm0         ; return sum

  pop edi
  pop esi

  ret

ALIGN 16
.qes2_q1loop
  movdqa xmm0, [esi + ecx*8]        ; xmm0 = [1st]
  movdqa xmm3, [esi + ecx*8 + 16]   ; xmm3 = [2nd]
  pxor xmm1, xmm1
  pxor xmm4, xmm4
  pcmpgtw xmm1, xmm0
  pcmpgtw xmm4, xmm3
  pxor xmm0, xmm1
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  psubusw xmm0, xmm6
  psubusw xmm3, xmm6
  psrlw xmm0, 1
  psrlw xmm3, 1
  paddw xmm5, xmm0
  pxor xmm0, xmm1
  paddw xmm5, xmm3
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  movdqa [edi + ecx*8], xmm0
  movdqa [edi + ecx*8 + 16], xmm3

  add ecx, 4
  cmp ecx, 16
  jnz .qes2_q1loop
  jmp .qes2_done


;-----------------------------------------------------------------------------
;
; uint32_t dequant_h263_intra_mmx(int16_t *data,
;                                 const int16_t const *coeff,
;                                 const uint32_t quant,
;                                 const uint32_t dcscalar,
;                                 const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

  ; note: we only saturate to +2047 *before* restoring the sign.
  ; Hence, final clamp really is [-2048,2047]

ALIGN 16
dequant_h263_intra_mmx:

  mov edx, [esp+ 4]                 ; data
  mov ecx, [esp+ 8]                 ; coeff
  mov eax, [esp+12]                 ; quant
  movq mm6, [mmx_add + eax*8 - 8]   ; quant or quant-1
  movq mm7, [mmx_mul + eax*8 - 8]   ; 2*quant
  mov eax, -16

ALIGN 16
.loop
  movq mm0, [ecx+8*eax+8*16]        ; c  = coeff[i]
  movq mm3, [ecx+8*eax+8*16 + 8]    ; c' = coeff[i+1]
  pxor mm1, mm1
  pxor mm4, mm4
  pcmpgtw mm1, mm0                  ; sign(c)
  pcmpgtw mm4, mm3                  ; sign(c')
  pxor mm2, mm2
  pxor mm5, mm5
  pcmpeqw mm2, mm0                  ; c is zero
  pcmpeqw mm5, mm3                  ; c' is zero
  pandn mm2, mm6                    ; offset = isZero ? 0 : quant_add
  pandn mm5, mm6
  pxor mm0, mm1                     ; negate if negative
  pxor mm3, mm4                     ; negate if negative
  psubw mm0, mm1
  psubw mm3, mm4
  pmullw mm0, mm7                   ; *= 2Q
  pmullw mm3, mm7                   ; *= 2Q
  paddw mm0, mm2                    ; + offset
  paddw mm3, mm5                    ; + offset
  paddw mm0, mm1                    ; negate back
  paddw mm3, mm4                    ; negate back

  ; saturates to +2047
  movq mm2, [mmx_32767_minus_2047]
  add eax, 2
  paddsw mm0, mm2
  paddsw mm3, mm2
  psubsw mm0, mm2
  psubsw mm3, mm2

  pxor mm0, mm1
  pxor mm3, mm4
  movq [edx + 8*eax + 8*16   - 2*8], mm0
  movq [edx + 8*eax + 8*16+8 - 2*8], mm3
  jnz near .loop

 ; deal with DC
  movd mm0, [ecx]
  pmullw mm0, [esp+16]          ; dcscalar
  movq mm2, [mmx_32767_minus_2047]
  paddsw mm0, mm2
  psubsw mm0, mm2
  movq mm3, [mmx_32768_minus_2048]
  psubsw mm0, mm3
  paddsw mm0, mm3
  movd eax, mm0
  mov [edx], ax

  xor eax, eax              ; return(0);
  ret

;-----------------------------------------------------------------------------
;
; uint32_t dequant_h263_intra_xmm(int16_t *data,
;                                 const int16_t const *coeff,
;                                 const uint32_t quant,
;                                 const uint32_t dcscalar,
;                                 const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

  ; this is the same as dequant_inter_mmx, except that we're
  ; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster)

ALIGN 16
dequant_h263_intra_xmm:

  mov edx, [esp+ 4]                 ; data
  mov ecx, [esp+ 8]                 ; coeff
  mov eax, [esp+12]                 ; quant
  movq mm6, [mmx_add + eax*8 - 8]   ; quant or quant-1
  movq mm7, [mmx_mul + eax*8 - 8]   ; 2*quant
  mov eax, -16

ALIGN 16
.loop
  movq mm0, [ecx+8*eax+8*16]        ; c  = coeff[i]
  movq mm3, [ecx+8*eax+8*16 + 8]    ; c' = coeff[i+1]
  pxor mm1, mm1
  pxor mm4, mm4
  pcmpgtw mm1, mm0                  ; sign(c)
  pcmpgtw mm4, mm3                  ; sign(c')
  pxor mm2, mm2
  pxor mm5, mm5
  pcmpeqw mm2, mm0                  ; c is zero
  pcmpeqw mm5, mm3                  ; c' is zero
  pandn mm2, mm6                    ; offset = isZero ? 0 : quant_add
  pandn mm5, mm6
  pxor mm0, mm1                     ; negate if negative
  pxor mm3, mm4                     ; negate if negative
  psubw mm0, mm1
  psubw mm3, mm4
  pmullw mm0, mm7                   ; *= 2Q
  pmullw mm3, mm7                   ; *= 2Q
  paddw mm0, mm2                    ; + offset
  paddw mm3, mm5                    ; + offset
  paddw mm0, mm1                    ; negate back
  paddw mm3, mm4                    ; negate back

   ; saturates to +2047
  movq mm2, [mmx_2047]
  pminsw mm0, mm2
  add eax, 2
  pminsw mm3, mm2

  pxor mm0, mm1
  pxor mm3, mm4
  movq [edx + 8*eax + 8*16   - 2*8], mm0
  movq [edx + 8*eax + 8*16+8 - 2*8], mm3
  jnz near .loop

    ; deal with DC
  movd mm0, [ecx]
  pmullw mm0, [esp+16]    ; dcscalar
  movq mm2, [mmx_32767_minus_2047]
  paddsw mm0, mm2
  psubsw mm0, mm2
  movq mm2, [mmx_32768_minus_2048]
  psubsw mm0, mm2
  paddsw mm0, mm2
  movd eax, mm0
  mov [edx], ax

  xor eax, eax
  ret


;-----------------------------------------------------------------------------
;
; uint32_t dequant_h263_intra_sse2(int16_t *data,
;                                  const int16_t const *coeff,
;                                  const uint32_t quant,
;                                  const uint32_t dcscalar,
;                                  const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

ALIGN 16
dequant_h263_intra_sse2:
  mov edx, [esp+ 4]        ; data
  mov ecx, [esp+ 8]        ; coeff
  mov eax, [esp+12]        ; quant
  movq mm6, [mmx_add + eax * 8 - 8]
  movq mm7, [mmx_mul + eax * 8 - 8]
  movq2dq xmm6, mm6
  movq2dq xmm7, mm7
  movlhps xmm6, xmm6
  movlhps xmm7, xmm7
  mov eax, -16

ALIGN 16
.loop
  movdqa xmm0, [ecx + 8*16 + 8*eax]      ; c  = coeff[i]
  movdqa xmm3, [ecx + 8*16 + 8*eax+ 16]
  pxor xmm1, xmm1
  pxor xmm4, xmm4
  pcmpgtw xmm1, xmm0    ; sign(c)
  pcmpgtw xmm4, xmm3
  pxor xmm2, xmm2
  pxor xmm5, xmm5
  pcmpeqw xmm2, xmm0    ; c is zero
  pcmpeqw xmm5, xmm3
  pandn xmm2, xmm6      ; offset = isZero ? 0 : quant_add
  pandn xmm5, xmm6
  pxor xmm0, xmm1       ; negate if negative
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  pmullw xmm0, xmm7     ; *= 2Q
  pmullw xmm3, xmm7
  paddw xmm0, xmm2      ; + offset
  paddw xmm3, xmm5
  paddw xmm0, xmm1      ; negate back
  paddw xmm3, xmm4

    ; saturates to +2047
  movdqa xmm2, [sse2_2047]
  pminsw xmm0, xmm2
  add eax, 4
  pminsw xmm3, xmm2

  pxor xmm0, xmm1
  pxor xmm3, xmm4
  movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0
  movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3
  jnz near .loop

 ; deal with DC
  movd mm0, [ecx]
  pmullw mm0, [esp+16]    ; dcscalar
  movq mm2, [mmx_32767_minus_2047]
  paddsw mm0, mm2
  psubsw mm0, mm2
  movq mm2, [mmx_32768_minus_2048]
  psubsw mm0, mm2
  paddsw mm0, mm2
  movd eax, mm0
  mov [edx], ax

  xor eax, eax
  ret

;-----------------------------------------------------------------------------
;
; uint32t dequant_h263_inter_mmx(int16_t * data,
;                                const int16_t * const coeff,
;                                const uint32_t quant,
;                                const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

ALIGN 16
dequant_h263_inter_mmx:

  mov edx, [esp+ 4]        ; data
  mov ecx, [esp+ 8]        ; coeff
  mov eax, [esp+12]        ; quant
  movq mm6, [mmx_add + eax*8 - 8]  ; quant or quant-1
  movq mm7, [mmx_mul + eax*8 - 8]  ; 2*quant
  mov eax, -16

ALIGN 16
.loop
  movq mm0, [ecx+8*eax+8*16]      ; c  = coeff[i]
  movq mm3, [ecx+8*eax+8*16 + 8]  ; c' = coeff[i+1]
  pxor mm1, mm1
  pxor mm4, mm4
  pcmpgtw mm1, mm0  ; sign(c)
  pcmpgtw mm4, mm3  ; sign(c')
  pxor mm2, mm2
  pxor mm5, mm5
  pcmpeqw mm2, mm0  ; c is zero
  pcmpeqw mm5, mm3  ; c' is zero
  pandn mm2, mm6    ; offset = isZero ? 0 : quant_add
  pandn mm5, mm6
  pxor mm0, mm1     ; negate if negative
  pxor mm3, mm4     ; negate if negative
  psubw mm0, mm1
  psubw mm3, mm4
  pmullw mm0, mm7   ; *= 2Q
  pmullw mm3, mm7   ; *= 2Q
  paddw mm0, mm2    ; + offset
  paddw mm3, mm5    ; + offset
  paddw mm0, mm1    ; negate back
  paddw mm3, mm4    ; negate back
  ; saturates to +2047
  movq mm2, [mmx_32767_minus_2047]
  add eax, 2
  paddsw mm0, mm2
  paddsw mm3, mm2
  psubsw mm0, mm2
  psubsw mm3, mm2

  pxor mm0, mm1
  pxor mm3, mm4
  movq [edx + 8*eax + 8*16   - 2*8], mm0
  movq [edx + 8*eax + 8*16+8 - 2*8], mm3
  jnz near .loop

  xor eax, eax
  ret

;-----------------------------------------------------------------------------
;
; uint32_t dequant_h263_inter_xmm(int16_t * data,
;                                 const int16_t * const coeff,
;                                 const uint32_t quant,
;                                 const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

  ; this is the same as dequant_inter_mmx,
  ; except that we're saturating using 'pminsw' (saves 2 cycles/loop)

ALIGN 16
dequant_h263_inter_xmm:

  mov edx, [esp+ 4]        ; data
  mov ecx, [esp+ 8]        ; coeff
  mov eax, [esp+12]        ; quant
  movq mm6, [mmx_add + eax*8 - 8]  ; quant or quant-1
  movq mm7, [mmx_mul + eax*8 - 8]  ; 2*quant
  mov eax, -16

ALIGN 16
.loop
  movq mm0, [ecx+8*eax+8*16]      ; c  = coeff[i]
  movq mm3, [ecx+8*eax+8*16 + 8]  ; c' = coeff[i+1]
  pxor mm1, mm1
  pxor mm4, mm4
  pcmpgtw mm1, mm0  ; sign(c)
  pcmpgtw mm4, mm3  ; sign(c')
  pxor mm2, mm2
  pxor mm5, mm5
  pcmpeqw mm2, mm0  ; c is zero
  pcmpeqw mm5, mm3  ; c' is zero
  pandn mm2, mm6    ; offset = isZero ? 0 : quant_add
  pandn mm5, mm6
  pxor mm0, mm1     ; negate if negative
  pxor mm3, mm4     ; negate if negative
  psubw mm0, mm1
  psubw mm3, mm4
  pmullw mm0, mm7   ; *= 2Q
  pmullw mm3, mm7   ; *= 2Q
  paddw mm0, mm2    ; + offset
  paddw mm3, mm5    ; + offset
  paddw mm0, mm1    ; start restoring sign
  paddw mm3, mm4    ; start restoring sign
                            ; saturates to +2047
  movq mm2, [mmx_2047]
  pminsw mm0, mm2
  add eax, 2
  pminsw mm3, mm2

  pxor mm0, mm1 ; finish restoring sign
  pxor mm3, mm4 ; finish restoring sign
  movq [edx + 8*eax + 8*16   - 2*8], mm0
  movq [edx + 8*eax + 8*16+8 - 2*8], mm3
  jnz near .loop

  xor eax, eax
  ret

;-----------------------------------------------------------------------------
;
; uint32_t dequant_h263_inter_sse2(int16_t * data,
;                                  const int16_t * const coeff,
;                                  const uint32_t quant,
;                                  const uint16_t *mpeg_matrices);
;
;-----------------------------------------------------------------------------

ALIGN 16
dequant_h263_inter_sse2:
  mov edx, [esp + 4]    ; data
  mov ecx, [esp + 8]    ; coeff
  mov eax, [esp + 12]   ; quant
  movq mm6, [mmx_add + eax * 8 - 8]
  movq mm7, [mmx_mul + eax * 8 - 8]
  movq2dq xmm6, mm6
  movq2dq xmm7, mm7
  movlhps xmm6, xmm6
  movlhps xmm7, xmm7
  mov eax, -16

ALIGN 16
.loop
  movdqa xmm0, [ecx + 8*16 + 8*eax]  ; c  = coeff[i]
  movdqa xmm3, [ecx + 8*16 + 8*eax + 16]

  pxor xmm1, xmm1
  pxor xmm4, xmm4
  pcmpgtw xmm1, xmm0    ; sign(c)
  pcmpgtw xmm4, xmm3
  pxor xmm2, xmm2
  pxor xmm5, xmm5
  pcmpeqw xmm2, xmm0    ; c is zero
  pcmpeqw xmm5, xmm3
  pandn xmm2, xmm6
  pandn xmm5, xmm6
  pxor xmm0, xmm1       ; negate if negative
  pxor xmm3, xmm4
  psubw xmm0, xmm1
  psubw xmm3, xmm4
  pmullw xmm0, xmm7     ; *= 2Q
  pmullw xmm3, xmm7
  paddw xmm0, xmm2      ; + offset
  paddw xmm3, xmm5

  paddw xmm0, xmm1      ; start restoring sign
  paddw xmm3, xmm4

 ; saturates to +2047
  movdqa xmm2, [sse2_2047]
  pminsw xmm0, xmm2
  add eax, 4
  pminsw xmm3, xmm2

  pxor xmm0, xmm1 ; finish restoring sign
  pxor xmm3, xmm4
  movdqa [edx + 8*16 - 8*4 + 8*eax], xmm0
  movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3
  jnz near .loop
 
   xor eax, eax
   ret