quantize_mpeg_xmm.asm

这是一个压缩解压包,用C语言进行编程的,里面有详细的源代码.

  mov rsi, -14
  mov rbx, rsp
  sub rsp, byte 24	; 16 would be enough, but it isn't important
  lea rbx, [rsp+8]
  and rbx, byte -8 ;ALIGN 8
  pxor mm0, mm0
  pxor mm3, mm3
  movq [byte rbx],mm0
  movq [rbx+8],mm0
  cmp rcx, byte 1
  je near .q1loop
  cmp rcx, byte 19
  jg near .lloop

ALIGN 16
.loop
  movq mm1, [rax + 8*rsi+112]       ; mm0 = [1st]
  psubw mm0, mm1 ;-mm1
  movq mm4, [rax + 8*rsi + 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, [rdi + 640 + 8*rsi+112]
  paddw mm3, [rdi + 640 + 8*rsi+120]
  movq mm5, [rdi + 896 + 8*rsi+112]
  movq mm7, [rdi + 896 + 8*rsi+120]
  pmulhuw mm5, mm0
  pmulhuw mm7, mm3
;  mov esp, esp
  movq mm2, [rdi + 512 + 8*rsi+112]
  movq mm6, [rdi + 512 + 8*rsi+120]
  pmullw mm2, mm5
  pmullw mm6, mm7
  psubw mm0, mm2
  psubw mm3, mm6
  movq mm2, [byte rbx]
  lea r11, [mmx_divs wrt rip]
  movq mm6, [r11 + rcx * 8 - 8]
  pmulhuw mm0, [rdi + 768 + 8*rsi+112]
  pmulhuw mm3, [rdi + 768 + 8*rsi+120]
  paddw mm2, [rbx+8]    ;sum
  paddw mm5, mm0
  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 rsi, byte 2
  paddw mm2, mm5        ;sum += x1
  movq [rbx], 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 [rbx+8], mm2 ;store sum
  movq [rdx + 8*rsi+112-16], mm5
  movq [rdx + 8*rsi +120-16], mm7
  jng near .loop

.done
; calculate  data[0] // (int32_t)dcscalar)
  paddw mm2, [rbx]
  add rsp, byte 24
  pop rbx
  pmaddwd mm2, [mmx_one wrt rip]
  punpckldq mm0, mm2 ;get low dw to mm0:high
  paddd mm0,mm2
  punpckhdq mm0, mm0 ;get result to low
  movd rax, mm0

  ret

ALIGN 16
.q1loop
  movq mm1, [rax + 8*rsi+112]       ; mm0 = [1st]
  psubw mm0, mm1                    ;-mm1
  movq mm4, [rax + 8*rsi+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, [rdi + 640 + 8*rsi+112]    ;mm0 is to be divided
  paddw mm3, [rdi + 640 + 8*rsi+120]    ; inter1 contains fix for division by 1
  movq mm5, [rdi + 896 + 8*rsi+112] ;with rounding down
  movq mm7, [rdi + 896 + 8*rsi+120]
  pmulhuw mm5, mm0
  pmulhuw mm7, mm3                          ;mm7: first approx of division
;  mov esp, esp
  movq mm2, [rdi + 512 + 8*rsi+112]
  movq mm6, [rdi + 512 + 8*rsi+120]      ; divs for q<=16
  pmullw mm2, mm5                           ;test value <= original
  pmullw mm6, mm7
  psubw mm0, mm2                            ;mismatch
  psubw mm3, mm6
  movq mm2, [byte rbx]
  pmulhuw mm0, [rdi + 768 + 8*rsi+112]  ;correction
  pmulhuw mm3, [rdi + 768 + 8*rsi+120]
  paddw mm2, [rbx+8]    ;sum
  paddw mm5, mm0        ;final result
  paddw mm7, mm3
  pxor mm0, mm0
  pxor mm3, mm3
  psrlw mm5, 1          ;  (level ) /2  (quant = 1)
  psrlw mm7, 1
  add rsi, byte 2
  paddw mm2, mm5        ;sum += x1
  movq [rbx], mm7       ;store x2
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  movq [rbx+8], mm2     ;store sum
  movq [rdx + 8*rsi+112-16], mm5
  movq [rdx + 8*rsi +120-16], mm7
  jng near .q1loop
  jmp near .done

ALIGN 8
.lloop
  movq mm1, [rax + 8*rsi+112]       ; mm0 = [1st]
  psubw mm0,mm1         ;-mm1
  movq mm4, [rax + 8*rsi+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, [rdi + 640 + 8*rsi+112] ;mm0 is to be divided inter1 contains fix for division by 1
  paddw mm3, [rdi + 640 + 8*rsi+120]
  movq mm5,[rdi + 896 + 8*rsi+112]
  movq mm7,[rdi + 896 + 8*rsi+120]
  pmulhuw mm5,mm0
  pmulhuw mm7,mm3       ;mm7: first approx of division
;  mov esp,esp
  movq mm2,[rdi + 512 + 8*rsi+112]
  movq mm6,[rdi + 512 + 8*rsi+120]
  pmullw mm2,mm5        ;test value <= original
  pmullw mm6,mm7
  psubw mm0,mm2         ;mismatch
  psubw mm3,mm6
  movq mm2,[byte rbx]
  lea r11, [mmx_div wrt rip]
  movq mm6,[r11 + rcx * 8 - 8]  ; divs for q<=16
  pmulhuw mm0,[rdi + 768 + 8*rsi+112] ;correction
  pmulhuw mm3,[rdi + 768 + 8*rsi+120]
  paddw mm2,[rbx+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 rsi,byte 2
  psrlw mm5, 1          ; (level ) / (2*quant)
  paddw mm2,mm5         ;sum += x1
  psrlw mm7, 1
  movq [rbx],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 [rbx+8], mm2 ;store sum
  movq [rdx + 8*rsi+112-16], mm5
  movq [rdx + 8*rsi +120-16], mm7
  jng near .lloop
  jmp near .done
.endfunc

;-----------------------------------------------------------------------------
;
; uint32_t dequant_mpeg_intra_x86_64(int16_t *data,
;                                  const int16_t const *coeff,
;                                  const uint32_t quant,
;                                  const uint32_t dcscalar,
;                                  const uint16_t *mpeg_matrices);
; Ported from the 32bit 3dne cousin
;-----------------------------------------------------------------------------

  ;   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 rcx+ 16 * %1] ; mm0 = c  = coeff[i]
  movq mm4, [rcx+ 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, [rsp-16]
  movq mm7, [rsp-16]
%endif
  pmullw mm2, [rdi + 16 * %1 ]     ; matrix[i]*quant
  pmullw mm7, [rdi + 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, [rsp-8]
  pcmpgtw mm3, [rsp-8]
  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, [rsp-8]
  movq mm3, [rsp-8]
  psubw mm2, mm1    ; finish negating back
  psubw mm7, mm4    ; finish negating back
  movq [byte rdx + 16 * %1], mm2   ; data[i]
  movq [rdx + 16 * %1  +8], mm7   ; data[i+1]
%endmacro

ALIGN 16
dequant_mpeg_intra_x86_64:
  mov rax, rdx		; quant
  mov rdx, rdi		; data
  mov r9, rcx		; dcscalar
  mov rcx, rsi		; coeff

  lea r11, [mmx_mul_quant wrt rip]
  movq mm7, [r11  + rax*8 - 8]
  psllw mm7, 2      ; << 2. See comment.
  push rbx

  movsx ebx, word [rcx]
  pxor mm0, mm0
  pxor mm3, mm3

  movq [rsp-8], mm0
  movq [rsp-16], mm7

  imul ebx, r9d
  movq mm2, mm7
  mov rdi, r8				; 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

  DEQUANT4INTRAMMX 3

  mov [byte rdx], bx

  DEQUANT4INTRAMMX 4
  DEQUANT4INTRAMMX 5
  DEQUANT4INTRAMMX 6
  DEQUANT4INTRAMMX 7

  pop rbx

  xor rax, rax
  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);
; Ported from 32bit 3dne cousin
;-----------------------------------------------------------------------------

    ; 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_x86_64:
  mov rax, rdx			; quant
  mov rdx, rdi			; data
  mov rdi, rcx			; mpeg_matrices
  mov rcx, rsi			; coeff

  lea r11, [mmx_mul_quant wrt rip]
  movq mm7, [r11  + rax*8 - 8]
  mov rax, -14
  paddw mm7, mm7    ; << 1
  pxor mm6, mm6     ; mismatch sum
  pxor mm1, mm1
  pxor mm3, mm3

ALIGN 16
.loop
  movq mm0, [rcx+8*rax + 7*16   ]   ; mm0 = coeff[i]
  pcmpgtw mm1, mm0  ; mm1 = sgn(c)    (preserved)
  movq mm2, [rcx+8*rax + 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, [mmzero wrt rip]
  movq mm5, [mmzero wrt rip]
  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, [rdi + 512 + 8*rax + 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, [rdi + 512 + 8*rax + 7*16 + 8]
  add rax, byte 2

  pcmpgtw mm5, [mmzero wrt rip]
  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, [mmzero wrt rip]
  paddusw mm2, mm5
  psrlw mm2, 5
  pxor mm2, mm3     ; start restoring sign
  psubusw mm3, mm5
  psubw mm2, mm3    ; finish restoring sign
  movq mm1, [mmzero wrt rip]
  movq mm3, [byte mmzero wrt rip]
  pxor mm6, mm0                             ; mismatch control
  movq [rdx + 8*rax + 7*16 -2*8   ], mm0    ; data[i]
  pxor mm6, mm2                             ; mismatch control
  movq [rdx + 8*rax + 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 rax, mm6
  and rax, byte 1
  xor rax, byte 1
  xor word [rdx + 2*63], ax

  xor rax, rax
  ret
.endfunc