idctmmx.asm

H.263的编码程序,加了CPU指令优化,VC版.

	paddd mm0, mm2 				; a3+b3 a2+b2
	psrad mm4, SHIFT_INV_ROW 		; y6=a1-b1 y7=a0-b0

	psubd mm7, mm2 			; 2 	; a3-b3 a2-b2
	psrad mm0, SHIFT_INV_ROW 		; y3=a3+b3 y2=a2+b2

	psrad mm7, SHIFT_INV_ROW 		; y4=a3-b3 y5=a2-b2

	packssdw mm3, mm0 		; 0 	; y3 y2 y1 y0

	packssdw mm7, mm4 		; 4 	; y6 y7 y4 y5

	movq mptr [OUT], mm3 		; 3 	; save y3 y2 y1 y0
	pshufw mm7, mm7, 10110001b 		; y7 y6 y5 y4

	movq mptr [OUT+8], mm7 		; 7 	; save y7 y6 y5 y4
ENDM

;=============================================================================
;
;=============================================================================

;=============================================================================
;
; The first stage DCT 8x8 - forward DCTs of columns
;
; The outputs are multiplied
; for rows 0,4 - on cos_4_16,
; for rows 1,7 - on cos_1_16,
; for rows 2,6 - on cos_2_16,
; for rows 3,5 - on cos_3_16
; and are shifted to the left for rise of accuracy
;
;-----------------------------------------------------------------------------
;
; The 8-point scaled forward DCT algorithm (26a8m)
;
;-----------------------------------------------------------------------------
;
; #define DCT_8_FRW_COL(x, y)
;{
; short t0, t1, t2, t3, t4, t5, t6, t7;
; short tp03, tm03, tp12, tm12, tp65, tm65;
; short tp465, tm465, tp765, tm765;
;
; t0 = LEFT_SHIFT ( x[0] + x[7] );
; t1 = LEFT_SHIFT ( x[1] + x[6] );
; t2 = LEFT_SHIFT ( x[2] + x[5] );
; t3 = LEFT_SHIFT ( x[3] + x[4] );
; t4 = LEFT_SHIFT ( x[3] - x[4] );
; t5 = LEFT_SHIFT ( x[2] - x[5] );
; t6 = LEFT_SHIFT ( x[1] - x[6] );
; t7 = LEFT_SHIFT ( x[0] - x[7] );
;
; tp03 = t0 + t3;
; tm03 = t0 - t3;
; tp12 = t1 + t2;
; tm12 = t1 - t2;
;
; y[0] = tp03 + tp12;
; y[4] = tp03 - tp12;
;
; y[2] = tm03 + tm12 * tg_2_16;
; y[6] = tm03 * tg_2_16 - tm12;
;
; tp65 =(t6 +t5 )*cos_4_16;
; tm65 =(t6 -t5 )*cos_4_16;
;
; tp765 = t7 + tp65;
; tm765 = t7 - tp65;
; tp465 = t4 + tm65;
; tm465 = t4 - tm65;
;
; y[1] = tp765 + tp465 * tg_1_16;
; y[7] = tp765 * tg_1_16 - tp465;
; y[5] = tm765 * tg_3_16 + tm465;
; y[3] = tm765 - tm465 * tg_3_16;
;}
;
;=============================================================================
DCT_8_FRW_COL_4 MACRO INP:REQ, OUT:REQ
LOCAL x0, x1, x2, x3, x4, x5, x6, x7
LOCAL y0, y1, y2, y3, y4, y5, y6, y7
x0 equ [INP + 0*16]
x1 equ [INP + 1*16]
x2 equ [INP + 2*16]
x3 equ [INP + 3*16]
x4 equ [INP + 4*16]
x5 equ [INP + 5*16]
x6 equ [INP + 6*16]
x7 equ [INP + 7*16]
y0 equ [OUT + 0*16]
y1 equ [OUT + 1*16]
y2 equ [OUT + 2*16]
y3 equ [OUT + 3*16]
y4 equ [OUT + 4*16]
y5 equ [OUT + 5*16]
y6 equ [OUT + 6*16]
y7 equ [OUT + 7*16]
movq mm0, x1 ; 0 ; x1
movq mm1, x6 ; 1 ; x6
movq mm2, mm0 ; 2 ; x1
movq mm3, x2 ; 3 ; x2
paddsw mm0, mm1 ; t1 = x[1] + x[6]
movq mm4, x5 ; 4 ; x5
psllw mm0, SHIFT_FRW_COL ; t1
movq mm5, x0 ; 5 ; x0
paddsw mm4, mm3 ; t2 = x[2] + x[5]
paddsw mm5, x7 ; t0 = x[0] + x[7]
psllw mm4, SHIFT_FRW_COL ; t2
movq mm6, mm0 ; 6 ; t1
psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6]
movq mm1, mptr tg_2_16 ; 1 ; tg_2_16
psubsw mm0, mm4 ; tm12 = t1 - t2
movq mm7, x3 ; 7 ; x3
pmulhw mm1, mm0 ; tm12*tg_2_16
paddsw mm7, x4 ; t3 = x[3] + x[4]
psllw mm5, SHIFT_FRW_COL ; t0
paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2
psllw mm7, SHIFT_FRW_COL ; t3
movq mm4, mm5 ; 4 ; t0
psubsw mm5, mm7 ; tm03 = t0 - t3
paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16
paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3
por mm1, mptr one_corr ; correction y2 +0.5
psllw mm2, SHIFT_FRW_COL+1 ; t6
pmulhw mm5, mptr tg_2_16 ; tm03*tg_2_16
movq mm7, mm4 ; 7 ; tp03
psubsw mm3, x5 ; t5 = x[2] - x[5]
psubsw mm4, mm6 ; y4 = tp03 - tp12
movq y2, mm1 ; 1 ; save y2
paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12
movq mm1, x3 ; 1 ; x3
psllw mm3, SHIFT_FRW_COL+1 ; t5
psubsw mm1, x4 ; t4 = x[3] - x[4]
movq mm6, mm2 ; 6 ; t6
movq y4, mm4 ; 4 ; save y4
paddsw mm2, mm3 ; t6 + t5
pmulhw mm2, mptr ocos_4_16 ; tp65 = (t6 + t5)*cos_4_16
psubsw mm6, mm3 ; 3 ; t6 - t5
pmulhw mm6, mptr ocos_4_16 ; tm65 = (t6 - t5)*cos_4_16
psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12
por mm5, mptr one_corr ; correction y6 +0.5
psllw mm1, SHIFT_FRW_COL ; t4
por mm2, mptr one_corr ; correction tp65 +0.5
movq mm4, mm1 ; 4 ; t4
movq mm3, x0 ; 3 ; x0
paddsw mm1, mm6 ; tp465 = t4 + tm65
psubsw mm3, x7 ; t7 = x[0] - x[7]
psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65
movq mm0, mptr tg_1_16 ; 0 ; tg_1_16
psllw mm3, SHIFT_FRW_COL ; t7
movq mm6, mptr tg_3_16 ; 6 ; tg_3_16
pmulhw mm0, mm1 ; tp465*tg_1_16
movq y0, mm7 ; 7 ; save y0
pmulhw mm6, mm4 ; tm465*tg_3_16
movq y6, mm5 ; 5 ; save y6
movq mm7, mm3 ; 7 ; t7
movq mm5, mptr tg_3_16 ; 5 ; tg_3_16
psubsw mm7, mm2 ; tm765 = t7 - tp65
paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65
pmulhw mm5, mm7 ; tm765*tg_3_16
paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16
paddsw mm6, mm4 ; tm465*tg_3_16
pmulhw mm3, mptr tg_1_16 ; tp765*tg_1_16
por mm0, mptr one_corr ; correction y1 +0.5
paddsw mm5, mm7 ; tm765*tg_3_16
psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16
movq y1, mm0 ; 0 ; save y1
paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465
movq y3, mm7 ; 7 ; save y3
psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465
movq y5, mm5 ; 5 ; save y5
movq y7, mm3 ; 3 ; save y7
ENDM

DCT_8_INV_COL_4 MACRO INP:REQ, OUT:REQ
	movq	mm0, mmword ptr tg_3_16

	movq	mm3, mmword ptr [INP+16*3]
	movq	mm1, mm0			; tg_3_16

	movq	mm5, mmword ptr [INP+16*5]
	pmulhw	mm0, mm3			; x3*(tg_3_16-1)

	movq	mm4, mmword ptr tg_1_16
	pmulhw	mm1, mm5			; x5*(tg_3_16-1)

	movq	mm7, mmword ptr [INP+16*7]
	movq	mm2, mm4			; tg_1_16

	movq	mm6, mmword ptr [INP+16*1]
	pmulhw	mm4, mm7			; x7*tg_1_16

	paddsw	mm0, mm3			; x3*tg_3_16
	pmulhw	mm2, mm6			; x1*tg_1_16

	paddsw	mm1, mm3			; x3+x5*(tg_3_16-1)
	psubsw	mm0, mm5			; x3*tg_3_16-x5 = tm35

	movq	mm3, mmword ptr ocos_4_16
	paddsw	mm1, mm5			; x3+x5*tg_3_16 = tp35

	paddsw	mm4, mm6			; x1+tg_1_16*x7 = tp17
	psubsw	mm2, mm7			; x1*tg_1_16-x7 = tm17

	movq	mm5, mm4			; tp17
	movq	mm6, mm2			; tm17

	paddsw	mm5, mm1			; tp17+tp35 = b0
	psubsw	mm6, mm0			; tm17-tm35 = b3

	psubsw	mm4, mm1			; tp17-tp35 = t1
	paddsw	mm2, mm0			; tm17+tm35 = t2

	movq	mm7, mmword ptr tg_2_16
	movq	mm1, mm4			; t1

;	movq	mmword ptr [SCRATCH+0], mm5	; save b0
	movq	mmword ptr [OUT+3*16], mm5	; save b0
	paddsw	mm1, mm2			; t1+t2

;	movq	mmword ptr [SCRATCH+8], mm6	; save b3
	movq	mmword ptr [OUT+5*16], mm6	; save b3
	psubsw	mm4, mm2			; t1-t2

	movq	mm5, mmword ptr [INP+2*16]
	movq	mm0, mm7			; tg_2_16

	movq	mm6, mmword ptr [INP+6*16]
	pmulhw	mm0, mm5			; x2*tg_2_16

	pmulhw	mm7, mm6			; x6*tg_2_16
; slot
	pmulhw	mm1, mm3			; ocos_4_16*(t1+t2) = b1/2
; slot
	movq	mm2, mmword ptr [INP+0*16]
	pmulhw	mm4, mm3			; ocos_4_16*(t1-t2) = b2/2

	psubsw	mm0, mm6			; t2*tg_2_16-x6 = tm26
	movq	mm3, mm2			; x0

	movq	mm6, mmword ptr [INP+4*16]
	paddsw	mm7, mm5			; x2+x6*tg_2_16 = tp26

	paddsw	mm2, mm6			; x0+x4 = tp04
	psubsw	mm3, mm6			; x0-x4 = tm04

	movq	mm5, mm2			; tp04
	movq	mm6, mm3			; tm04

	psubsw	mm2, mm7			; tp04-tp26 = a3
	paddsw	mm3, mm0			; tm04+tm26 = a1

	paddsw mm1, mm1				; b1
	paddsw mm4, mm4				; b2

	paddsw	mm5, mm7			; tp04+tp26 = a0
	psubsw	mm6, mm0			; tm04-tm26 = a2

	movq	mm7, mm3			; a1
	movq	mm0, mm6			; a2

	paddsw	mm3, mm1			; a1+b1
	paddsw	mm6, mm4			; a2+b2

	psraw	mm3, SHIFT_INV_COL		; dst1
	psubsw	mm7, mm1			; a1-b1

	psraw	mm6, SHIFT_INV_COL		; dst2
	psubsw	mm0, mm4			; a2-b2

;	movq	mm1, mmword ptr [SCRATCH+0]	; load b0
	movq	mm1, mmword ptr [OUT+3*16]	; load b0
	psraw	mm7, SHIFT_INV_COL		; dst6

	movq	mm4, mm5			; a0
	psraw	mm0, SHIFT_INV_COL		; dst5

	movq	mmword ptr [OUT+1*16], mm3
	paddsw	mm5, mm1			; a0+b0

	movq	mmword ptr [OUT+2*16], mm6
	psubsw	mm4, mm1			; a0-b0

;	movq	mm3, mmword ptr [SCRATCH+8]	; load b3
	movq	mm3, mmword ptr [OUT+5*16]	; load b3
	psraw	mm5, SHIFT_INV_COL		; dst0

	movq	mm6, mm2			; a3
	psraw	mm4, SHIFT_INV_COL		; dst7

	movq	mmword ptr [OUT+5*16], mm0
	paddsw	mm2, mm3			; a3+b3

	movq	mmword ptr [OUT+6*16], mm7
	psubsw	mm6, mm3			; a3-b3

	movq	mmword ptr [OUT+0*16], mm5
	psraw	mm2, SHIFT_INV_COL		; dst3

	movq	mmword ptr [OUT+7*16], mm4
	psraw	mm6, SHIFT_INV_COL		; dst4

	movq	mmword ptr [OUT+3*16], mm2

	movq	mmword ptr [OUT+4*16], mm6
ENDM

_TEXT SEGMENT PARA PUBLIC USE32 'CODE'

;
; extern "C" __cdecl void idct8x8_mmx (short *src_result);
;
public  ?idct_mmx@@YAXPA_F0@Z

?idct_mmx@@YAXPA_F0@Z proc near
	mov     eax, dword ptr [esp + 4]          ; source
	mov     ecx, dword ptr [esp + 8]

	DCT_8_INV_ROW_1	[eax+0], [ecx+0], tab_i_04, rounder_0
	DCT_8_INV_ROW_1	[eax+16], [ecx+16], tab_i_17, rounder_1
	DCT_8_INV_ROW_1	[eax+32], [ecx+32], tab_i_26, rounder_2
	DCT_8_INV_ROW_1	[eax+48], [ecx+48], tab_i_35, rounder_3
	DCT_8_INV_ROW_1	[eax+64], [ecx+64], tab_i_04, rounder_4
	DCT_8_INV_ROW_1	[eax+80], [ecx+80], tab_i_35, rounder_5
	DCT_8_INV_ROW_1	[eax+96], [ecx+96], tab_i_26, rounder_6
	DCT_8_INV_ROW_1	[eax+112], [ecx+112], tab_i_17, rounder_7

	DCT_8_INV_COL_4 [eax+0],[ecx+0]
	DCT_8_INV_COL_4 [eax+8],[ecx+8]
	emms

	ret    

?idct_mmx@@YAXPA_F0@Z ENDP

_TEXT ENDS

_TEXT SEGMENT PARA PUBLIC USE32 'CODE'

;
; extern "C" __fastcall void idct8x8_sse (short *src_result);
;
public  ?IDCTSSE@@YIXPA_F@Z

?IDCTSSE@@YIXPA_F@Z proc near
	mov     eax, ecx          ; source

	DCT_8_INV_ROW_1_s [eax+0], [eax+0], tab_i_04_s, rounder_0
	DCT_8_INV_ROW_1_s [eax+16], [eax+16], tab_i_17_s, rounder_1
	DCT_8_INV_ROW_1_s [eax+32], [eax+32], tab_i_26_s, rounder_2
	DCT_8_INV_ROW_1_s [eax+48], [eax+48], tab_i_35_s, rounder_3
	DCT_8_INV_ROW_1_s [eax+64], [eax+64], tab_i_04_s, rounder_4
	DCT_8_INV_ROW_1_s [eax+80], [eax+80], tab_i_35_s, rounder_5
	DCT_8_INV_ROW_1_s [eax+96], [eax+96], tab_i_26_s, rounder_6
	DCT_8_INV_ROW_1_s [eax+112], [eax+112], tab_i_17_s, rounder_7

	DCT_8_INV_COL_4 [eax+0],[eax+0]
	DCT_8_INV_COL_4 [eax+8],[eax+8]
	emms

	ret

?IDCTSSE@@YIXPA_F@Z ENDP

_TEXT ENDS

END