quantize_mmx.asm

视频压缩标准MPEG4的视频参考代码xvid9.1

;/*****************************************************************************
; *
; *  XVID MPEG-4 VIDEO CODEC
; *  mmx optimized quantization/dequantization             
; *
; *  Copyright(C) 2002 Peter Ross <[e-mail]>
; *  Copyright(C) 2002 Michael Militzer <[e-mail]>
; *  Copyright(C) 2002 Pascal Massimino <[e-mail]>
; *
; *  This file is part of XviD, a free MPEG-4 video encoder/decoder
; *
; *  XviD 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
; *
; *  Under section 8 of the GNU General Public License, the copyright
; *  holders of XVID explicitly forbid distribution in the following
; *  countries:
; *
; *    - Japan
; *    - United States of America
; *
; *  Linking XviD statically or dynamically with other modules is making a
; *  combined work based on XviD.  Thus, the terms and conditions of the
; *  GNU General Public License cover the whole combination.
; *
; *  As a special exception, the copyright holders of XviD give you
; *  permission to link XviD with independent modules that communicate with
; *  XviD solely through the VFW1.1 and DShow interfaces, regardless of the
; *  license terms of these independent modules, and to copy and distribute
; *  the resulting combined work under terms of your choice, provided that
; *  every copy of the combined work is accompanied by a complete copy of
; *  the source code of XviD (the version of XviD used to produce the
; *  combined work), being distributed under the terms of the GNU General
; *  Public License plus this exception.  An independent module is a module
; *  which is not derived from or based on XviD.
; *
; *  Note that people who make modified versions of XviD are not obligated
; *  to grant this special exception for their modified versions; it is
; *  their choice whether to do so.  The GNU General Public License gives
; *  permission to release a modified version without this exception; this
; *  exception also makes it possible to release a modified version which
; *  carries forward this exception.
; *
; * $Id: quantize_mmx.asm,v 1.7 2002/11/17 00:41:20 edgomez Exp $
; *
; *************************************************************************/

; enable dequant saturate [-2048,2047], test purposes only.
%define SATURATE

; data/text alignment
%define ALIGN 8

bits 32

section .data


%macro cglobal 1 
	%ifdef PREFIX
		global _%1 
		%define %1 _%1
	%else
		global %1
	%endif
%endmacro

align 16

plus_one times 8	dw	 1

;===========================================================================
;
; subtract by Q/2 table
;
;===========================================================================

%macro MMX_SUB  1
times 4 dw %1 / 2
%endmacro

align 16
mmx_sub
		MMX_SUB 1
		MMX_SUB 2
		MMX_SUB 3
		MMX_SUB 4
		MMX_SUB 5
		MMX_SUB 6
		MMX_SUB 7
		MMX_SUB 8
		MMX_SUB 9
		MMX_SUB 10
		MMX_SUB 11
		MMX_SUB 12
		MMX_SUB 13
		MMX_SUB 14
		MMX_SUB 15
		MMX_SUB 16
		MMX_SUB 17
		MMX_SUB 18
		MMX_SUB 19
		MMX_SUB 20
		MMX_SUB 21
		MMX_SUB 22
		MMX_SUB 23
		MMX_SUB 24
		MMX_SUB 25
		MMX_SUB 26
		MMX_SUB 27
		MMX_SUB 28
		MMX_SUB 29
		MMX_SUB 30
		MMX_SUB 31



;===========================================================================
;
; divide by 2Q table 
;
; use a shift of 16 to take full advantage of _pmulhw_
; for q=1, _pmulhw_ will overflow so it is treated seperately
; (3dnow2 provides _pmulhuw_ which wont cause overflow)
;
;===========================================================================

%macro MMX_DIV  1
times 4 dw  (1 << 16) / (%1 * 2) + 1
%endmacro

align 16
mmx_div
		MMX_DIV 1
		MMX_DIV 2
		MMX_DIV 3
		MMX_DIV 4
		MMX_DIV 5
		MMX_DIV 6
		MMX_DIV 7
		MMX_DIV 8
		MMX_DIV 9
		MMX_DIV 10
		MMX_DIV 11
		MMX_DIV 12
		MMX_DIV 13
		MMX_DIV 14
		MMX_DIV 15
		MMX_DIV 16
		MMX_DIV 17
		MMX_DIV 18
		MMX_DIV 19
		MMX_DIV 20
		MMX_DIV 21
		MMX_DIV 22
		MMX_DIV 23
		MMX_DIV 24
		MMX_DIV 25
		MMX_DIV 26
		MMX_DIV 27
		MMX_DIV 28
		MMX_DIV 29
		MMX_DIV 30
		MMX_DIV 31



;===========================================================================
;
; add by (odd(Q) ? Q : Q - 1) table
;
;===========================================================================

%macro MMX_ADD  1
%if %1 % 2 != 0
times 4 dw %1
%else
times 4 dw %1 - 1
%endif
%endmacro

align 16
mmx_add
		MMX_ADD 1
		MMX_ADD 2
		MMX_ADD 3
		MMX_ADD 4
		MMX_ADD 5
		MMX_ADD 6
		MMX_ADD 7
		MMX_ADD 8
		MMX_ADD 9
		MMX_ADD 10
		MMX_ADD 11
		MMX_ADD 12
		MMX_ADD 13
		MMX_ADD 14
		MMX_ADD 15
		MMX_ADD 16
		MMX_ADD 17
		MMX_ADD 18
		MMX_ADD 19
		MMX_ADD 20
		MMX_ADD 21
		MMX_ADD 22
		MMX_ADD 23
		MMX_ADD 24
		MMX_ADD 25
		MMX_ADD 26
		MMX_ADD 27
		MMX_ADD 28
		MMX_ADD 29
		MMX_ADD 30
		MMX_ADD 31


;===========================================================================
;
; multiple by 2Q table
;
;===========================================================================

%macro MMX_MUL  1
times 4 dw %1 * 2
%endmacro

align 16
mmx_mul
		MMX_MUL 1
		MMX_MUL 2
		MMX_MUL 3
		MMX_MUL 4
		MMX_MUL 5
		MMX_MUL 6
		MMX_MUL 7
		MMX_MUL 8
		MMX_MUL 9
		MMX_MUL 10
		MMX_MUL 11
		MMX_MUL 12
		MMX_MUL 13
		MMX_MUL 14
		MMX_MUL 15
		MMX_MUL 16
		MMX_MUL 17
		MMX_MUL 18
		MMX_MUL 19
		MMX_MUL 20
		MMX_MUL 21
		MMX_MUL 22
		MMX_MUL 23
		MMX_MUL 24
		MMX_MUL 25
		MMX_MUL 26
		MMX_MUL 27
		MMX_MUL 28
		MMX_MUL 29
		MMX_MUL 30
		MMX_MUL 31


;===========================================================================
;
; saturation limits 
;
;===========================================================================

align 16
sse2_2047	times 8 dw 2047

align 16
mmx_2047	times 4 dw 2047

align 8
mmx_32768_minus_2048				times 4 dw (32768-2048)
mmx_32767_minus_2047				times 4 dw (32767-2047)


section .text


;===========================================================================
;
; void quant_intra_mmx(int16_t * coeff, 
;					const int16_t const * data,
;					const uint32_t quant,
;					const uint32_t dcscalar);
;
;===========================================================================

align ALIGN
cglobal quant_intra_mmx
quant_intra_mmx

		push	ecx
		push	esi
		push	edi

		mov	edi, [esp + 12 + 4]		; coeff
		mov	esi, [esp + 12 + 8]		; data
		mov	eax, [esp + 12 + 12]		; quant

		xor ecx, ecx
		cmp	al, 1
		jz	.q1loop

		movq	mm7, [mmx_div + eax * 8 - 8]
align ALIGN
.loop
		movq	mm0, [esi + 8*ecx]		; mm0 = [1st]
		movq	mm3, [esi + 8*ecx + 8]	; 
		pxor	mm1, mm1		; mm1 = 0
		pxor	mm4, mm4		;
		pcmpgtw	mm1, mm0		; mm1 = (0 > mm0)
		pcmpgtw	mm4, mm3		; 
		pxor	mm0, mm1		; mm0 = |mm0|
		pxor	mm3, mm4		;
		psubw	mm0, mm1		; displace
		psubw	mm3, mm4		;
		pmulhw	mm0, mm7		; mm0 = (mm0 / 2Q) >> 16
		pmulhw	mm3, mm7		;
		pxor	mm0, mm1		; mm0 *= sign(mm0)
		pxor	mm3, mm4		;
		psubw	mm0, mm1		; undisplace
		psubw	mm3, mm4		;
		movq	[edi + 8*ecx], mm0
		movq	[edi + 8*ecx + 8], mm3
		
		add ecx,2
		cmp ecx,16
		jnz 	.loop 

.done	
	; caclulate  data[0] // (int32_t)dcscalar)

		mov 	ecx, [esp + 12 + 16]	; dcscalar
		mov 	edx, ecx
		movsx 	eax, word [esi]	; data[0]
		shr 	edx, 1			; edx = dcscalar /2
		cmp		eax, 0
		jg		.gtzero

		sub		eax, edx
		jmp		short .mul
.gtzero
		add		eax, edx
.mul
		cdq 				; expand eax -> edx:eax
		idiv	ecx			; eax = edx:eax / dcscalar
		
		mov	[edi], ax		; coeff[0] = ax

		pop	edi
		pop	esi
		pop	ecx

		ret				

align ALIGN
.q1loop
		movq	mm0, [esi + 8*ecx]		; mm0 = [1st]
		movq	mm3, [esi + 8*ecx + 8]	; 
		pxor	mm1, mm1		; mm1 = 0
		pxor	mm4, mm4		;
		pcmpgtw	mm1, mm0		; mm1 = (0 > mm0)
		pcmpgtw	mm4, mm3		; 
		pxor	mm0, mm1		; mm0 = |mm0|
		pxor	mm3, mm4		; 
		psubw	mm0, mm1		; displace
		psubw	mm3, mm4		; 
		psrlw	mm0, 1			; mm0 >>= 1   (/2)
		psrlw	mm3, 1			;
		pxor	mm0, mm1		; mm0 *= sign(mm0)
		pxor	mm3, mm4        ;
		psubw	mm0, mm1		; undisplace
		psubw	mm3, mm4		;
		movq	[edi + 8*ecx], mm0
		movq	[edi + 8*ecx + 8], mm3

		add ecx,2
		cmp ecx,16
		jnz	.q1loop
		jmp	short .done



;===========================================================================
;
; void quant_intra_sse2(int16_t * coeff, 
;					const int16_t const * data,
;					const uint32_t quant,
;					const uint32_t dcscalar);
;
;===========================================================================

align ALIGN
cglobal quant_intra_sse2
quant_intra_sse2

		push	esi
		push	edi

		mov		edi, [esp + 8 + 4]			; coeff
		mov		esi, [esp + 8 + 8]			; data
		mov		eax, [esp + 8 + 12]			; quant

		xor		ecx, ecx
		cmp		al, 1
		jz		near .qas2_q1loop

.qas2_not1
		movq	mm7, [mmx_div + eax*8 - 8]
		movq2dq	xmm7, mm7
		movlhps	xmm7, xmm7

align 16
.qas2_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
		pmulhw	xmm0, xmm7
		pmulhw	xmm3, xmm7
		pxor	xmm0, xmm1
		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 	.qas2_loop 

.qas2_done	
		mov 	ecx, [esp + 8 + 16]	; dcscalar
		mov 	edx, ecx
		movsx 	eax, word [esi]
		shr 	edx, 1
		cmp		eax, 0
		jg		.qas2_gtzero

		sub		eax, edx
		jmp		short .qas2_mul
.qas2_gtzero
		add		eax, edx
.qas2_mul
		cdq
		idiv	ecx
		
		mov		[edi], ax

		pop		edi
		pop		esi

		ret		

align 16
.qas2_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
		psrlw	xmm0, 1
		psrlw	xmm3, 1
		pxor	xmm0, xmm1
		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		.qas2_q1loop
		jmp		near .qas2_done



;===========================================================================
;
; uint32_t quant_inter_mmx(int16_t * coeff,
;					const int16_t const * data,
;					const uint32_t quant);
;
;===========================================================================

align ALIGN
cglobal quant_inter_mmx
		quant_inter_mmx

		push	ecx
		push	esi
		push	edi

		mov	edi, [esp + 12 + 4]		; coeff
		mov	esi, [esp + 12 + 8]		; data
		mov	eax, [esp + 12 + 12]	; quant

		xor ecx, ecx

		pxor mm5, mm5					; sum
		movq mm6, [mmx_sub + eax * 8 - 8]	; sub

		cmp	al, 1
		jz  .q1loop

		movq	mm7, [mmx_div + eax * 8 - 8]	; divider

align ALIGN
.loop
		movq	mm0, [esi + 8*ecx]		; mm0 = [1st]
		movq	mm3, [esi + 8*ecx + 8]	; 
		pxor	mm1, mm1		; mm1 = 0
		pxor	mm4, mm4		;
		pcmpgtw	mm1, mm0		; mm1 = (0 > mm0)
		pcmpgtw	mm4, mm3		; 
		pxor	mm0, mm1		; mm0 = |mm0|
		pxor	mm3, mm4		; 
		psubw	mm0, mm1		; displace
		psubw	mm3, mm4		; 
		psubusw	mm0, mm6		; mm0 -= sub (unsigned, dont go < 0)
		psubusw	mm3, mm6		;
		pmulhw	mm0, mm7		; mm0 = (mm0 / 2Q) >> 16
		pmulhw	mm3, mm7		; 
		paddw	mm5, mm0		; sum += mm0
		pxor	mm0, mm1		; mm0 *= sign(mm0)
		paddw	mm5, mm3		;
		pxor	mm3, mm4		;
		psubw	mm0, mm1		; undisplace
		psubw	mm3, mm4
		movq	[edi + 8*ecx], mm0
		movq	[edi + 8*ecx + 8], mm3

		add ecx, 2	
		cmp ecx, 16
		jnz .loop

.done
		pmaddwd mm5, [plus_one]
		movq    mm0, mm5
		psrlq   mm5, 32
		paddd   mm0, mm5