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.

;/****************************************************************************
; *
; *  XVID MPEG-4 VIDEO CODEC
; *  - 3dne Quantization/Dequantization -
; *
; *  Copyright (C) 2002-2003 Peter Ross <[e-mail]>
; *                2002      Jaan Kalda
; *
; *  This program is free software ; you can redistribute it and/or modify
; *  it under the terms of the GNU General Public License as published by
; *  the Free Software Foundation ; either version 2 of the License, or
; *  (at your option) any later version.
; *
; *  This program is distributed in the hope that it will be useful,
; *  but WITHOUT ANY WARRANTY ; without even the implied warranty of
; *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
; *  GNU General Public License for more details.
; *
; *  You should have received a copy of the GNU General Public License
; *  along with this program ; if not, write to the Free Software
; *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
; *
; * $Id: quantize_mpeg_xmm.asm,v 1.5 2004/08/29 10:02:38 edgomez Exp $
; *
; ***************************************************************************/

; _3dne functions are compatible with iSSE, but are optimized specifically
; for K7 pipelines

%define SATURATE

BITS 32

%macro cglobal 1
	%ifdef PREFIX
		%ifdef MARK_FUNCS
			global _%1:function %1.endfunc-%1
			%define %1 _%1:function %1.endfunc-%1
		%else
			global _%1
			%define %1 _%1
		%endif
	%else
		%ifdef MARK_FUNCS
			global %1:function %1.endfunc-%1
		%else
			global %1
		%endif
	%endif
%endmacro

%macro cextern 1
	%ifdef PREFIX
		extern _%1
		%define %1 _%1
	%else
		extern %1
	%endif
%endmacro

;=============================================================================
; Local data
;=============================================================================

%ifdef FORMAT_COFF
SECTION .rodata
%else
SECTION .rodata align=16
%endif

ALIGN 8
mmzero:
	dd 0,0
mmx_one:
	times 4 dw 1

;-----------------------------------------------------------------------------
; divide by 2Q table
;-----------------------------------------------------------------------------

ALIGN 16
mmx_divs:		;i>2
%assign i 1
%rep 31
	times 4 dw  ((1 << 15) / i + 1)
	%assign i i+1
%endrep

ALIGN 16
mmx_div:		;quant>2
	times 4 dw 65535 ; the div by 2 formula will overflow for the case
	                 ; quant=1 but we don't care much because quant=1
	                 ; is handled by a different piece of code that
	                 ; doesn't use this table.
%assign quant 2
%rep 31
	times 4 dw  ((1 << 16) / quant + 1)
	%assign quant quant+1
%endrep

%macro FIXX 1
dw (1 << 16) / (%1) + 1
%endmacro

%define nop4	db	08Dh, 074h, 026h,0
%define nop3	add	esp, byte 0
%define nop2	mov	esp, esp
%define nop7	db	08dh, 02ch, 02dh,0,0,0,0
%define nop6	add	ebp, dword 0

;-----------------------------------------------------------------------------
; quantd table
;-----------------------------------------------------------------------------

%define VM18P	3
%define VM18Q	4

ALIGN 16
quantd:
%assign i 1
%rep 31
	times 4 dw  (((VM18P*i) + (VM18Q/2)) / VM18Q)
	%assign i i+1
%endrep

;-----------------------------------------------------------------------------
; multiple by 2Q table
;-----------------------------------------------------------------------------

ALIGN 16
mmx_mul_quant:
%assign i 1
%rep 31
	times 4 dw  i
	%assign i i+1
%endrep

;-----------------------------------------------------------------------------
; saturation limits
;-----------------------------------------------------------------------------

ALIGN 16
mmx_32767_minus_2047:
	times 4 dw (32767-2047)
mmx_32768_minus_2048:
	times 4 dw (32768-2048)
mmx_2047:
	times 4 dw 2047
mmx_minus_2048:
	times 4 dw (-2048)
zero:
	times 4 dw 0

int_div:
dd 0
%assign i 1
%rep 255
	dd  (1 << 17) / ( i) + 1
	%assign i i+1
%endrep

;=============================================================================
; Code
;=============================================================================

SECTION .text

cglobal quant_mpeg_intra_xmm
cglobal quant_mpeg_inter_xmm
cglobal dequant_mpeg_intra_3dne
cglobal dequant_mpeg_inter_3dne

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

ALIGN 16
quant_mpeg_intra_xmm:
  mov eax, [esp  + 8]       ; data
  mov ecx, [esp  + 12]      ; quant
  mov edx, [esp  + 4]       ; coeff
  push esi
  push edi
  push ebx
  nop
  mov edi, [esp + 12 + 20]		; mpeg_quant_matrices
  mov esi, -14
  pxor mm0, mm0
  pxor mm3, mm3
  cmp ecx, byte 1
  je near .q1loop
  cmp ecx, byte 19
  jg near .lloop
  nop6

ALIGN 16
.loop
  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 + 128 + 8*esi+112]
  paddw mm3, [edi + 128 + 8*esi+120]
  movq mm5, [edi + 384 + 8*esi+112]
  movq mm7, [edi + 384 + 8*esi+120]
  pmulhuw mm5, mm0
  pmulhuw mm7, mm3
  mov esp, esp
  movq mm2, [edi + 8*esi+112]
  movq mm6, [edi + 8*esi+120]
  pmullw mm2, mm5
  pmullw mm6, mm7
  psubw mm0, mm2
  psubw mm3, mm6
  nop4
  movq mm2, [quantd + ecx * 8 - 8]
  movq mm6, [mmx_divs + ecx * 8 - 8]
  paddw mm5, mm2
  paddw mm7, mm2
  mov esp, esp
  pmulhuw mm0, [edi + 256 + 8*esi+112]
  pmulhuw mm3, [edi + 256 + 8*esi+120]
  paddw mm5, mm0
  paddw mm7, mm3
  pxor mm0, mm0
  pxor mm3, mm3
  pmulhuw mm5, mm6      ; mm0 = (mm0 / 2Q) >> 16
  pmulhuw mm7, mm6      ;  (level + quantd) / quant (0<quant<32)
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  movq [edx + 8*esi+112], mm5
  movq [edx + 8*esi +120], mm7
  add esi, byte 2
  jng near .loop

.done
; calculate  data[0] // (int32_t)dcscalar)
  mov esi, [esp + 12 + 16]  ; dcscalar
  movsx ecx, word [eax]
  mov edi, ecx
  mov edx, [esp + 12 + 16]
  shr edx, 1            ; ebx = dcscalar /2
  sar edi, 31           ; cdq is vectorpath
  xor edx, edi          ; ebx = eax V -eax -1
  sub ecx, edi
  add ecx, edx
  mov edx, [dword esp + 12 + 4]
  mov esi, [int_div+4*esi]
  imul ecx, esi
  sar ecx, 17
  lea ebx, [byte ecx + 1]
  cmovs ecx, ebx
  ; idiv    cx          ; ecx = edi:ecx / dcscalar

  mov ebx, [esp]
  mov edi, [esp+4]
  mov esi, [esp+8]
  add esp, byte 12
  mov [edx], cx     ; coeff[0] = ax

  xor eax, eax
  ret

ALIGN 16
.q1loop
  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 + 128 + 8*esi+112]    ;mm0 is to be divided
  paddw mm3, [edi + 128 + 8*esi+120]    ;intra1 contains fix for division by 1
  movq mm5, [edi + 384 + 8*esi+112] ;with rounding down
  movq mm7, [edi + 384 + 8*esi+120]
  pmulhuw mm5, mm0
  pmulhuw mm7, mm3      ;mm7: first approx of division
  mov esp, esp
  movq mm2, [edi + 8*esi+112]
  movq mm6, [edi + 8*esi+120]      ; divs for q<=16
  pmullw mm2, mm5       ;test value <= original
  pmullw mm6, mm7
  psubw mm0, mm2        ;mismatch
  psubw mm3, mm6
  nop4
  movq mm2, [quantd + ecx * 8 - 8]
  paddw mm5, mm2        ;first approx with quantd
  paddw mm7, mm2
  mov esp, esp
  pmulhuw mm0, [edi + 256 + 8*esi+112]   ;correction
  pmulhuw mm3, [edi + 256 + 8*esi+120]
  paddw mm5, mm0        ;final result with quantd
  paddw mm7, mm3
  pxor mm0, mm0
  pxor mm3, mm3
  mov esp, esp
  psrlw mm5, 1          ;  (level + quantd) /2  (quant = 1)
  psrlw mm7, 1
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  movq [edx + 8*esi+112], mm5
  movq [edx + 8*esi +120], mm7
  add esi, byte 2
  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 + 128 + 8*esi+112] ;mm0 is to be divided intra1 contains fix for division by 1
  paddw mm3, [edi + 128 + 8*esi+120]
  movq mm5, [edi + 384 + 8*esi+112]
  movq mm7, [edi + 384 + 8*esi+120]
  pmulhuw mm5, mm0
  pmulhuw mm7, mm3      ;mm7: first approx of division
  mov esp, esp
  movq mm2, [edi + 8*esi+112]
  movq mm6, [edi + 8*esi+120]
  pmullw mm2, mm5       ;test value <= original
  pmullw mm6, mm7
  psubw mm0, mm2        ;mismatch
  psubw mm3, mm6
  nop4
  movq mm2, [quantd + ecx * 8 - 8]
  movq mm6, [mmx_div + ecx * 8 - 8] ; divs for q<=16
  paddw mm5, mm2        ;first approx with quantd
  paddw mm7, mm2
  mov esp, esp
  pmulhuw mm0, [edi + 256 + 8*esi+112] ;correction
  pmulhuw mm3, [edi + 256 + 8*esi+120]
  paddw mm5, mm0        ;final result with quantd
  paddw mm7, mm3
  pxor mm0, mm0
  pxor mm3, mm3
  mov esp, esp
  pmulhuw mm5, mm6      ; mm0 = (mm0 / 2Q) >> 16
  pmulhuw mm7, mm6      ;  (level + quantd) / quant (0<quant<32)
  psrlw mm5, 1          ; (level + quantd) / (2*quant)
  psrlw mm7, 1
  pxor mm5, mm1         ; mm0 *= sign(mm0)
  pxor mm7, mm4         ;
  psubw mm5, mm1        ; undisplace
  psubw mm7, mm4        ;
  movq [edx + 8*esi+112], mm5
  movq [edx + 8*esi +120], mm7
  add esi,byte 2
  jng near .lloop
  jmp near .done
.endfunc

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

ALIGN 16
quant_mpeg_inter_xmm:
  mov eax, [esp  + 8]       ; data
  mov ecx, [esp  + 12]      ; quant
  mov edx, [esp  + 4]       ; coeff
  push esi
  push edi
  push ebx
  nop
  mov edi, [esp + 12 + 16]
  mov esi, -14
  mov ebx, esp
  sub esp, byte 24
  lea ebx, [esp+8]
  and ebx, byte -8 ;ALIGN 8
  pxor mm0, mm0
  pxor mm3, mm3
  movq [byte ebx],mm0