1 |
|
;/**************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * - MMX and XMM forward discrete cosine transform - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
|
; * 2002 Jaan Kalda |
8 |
|
; * |
9 |
|
; * This program is free software; you can redistribute it and/or modify it |
10 |
|
; * under the terms of the GNU General Public License as published by |
11 |
|
; * the Free Software Foundation; either version 2 of the License, or |
12 |
|
; * (at your option) any later version. |
13 |
|
; * |
14 |
|
; * This program is distributed in the hope that it will be useful, |
15 |
|
; * but WITHOUT ANY WARRANTY; without even the implied warranty of |
16 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
17 |
|
; * GNU General Public License for more details. |
18 |
|
; * |
19 |
|
; * You should have received a copy of the GNU General Public License |
20 |
|
; * along with this program; if not, write to the Free Software |
21 |
|
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
|
; * |
23 |
|
; * $Id$ |
24 |
|
; * |
25 |
|
; ***************************************************************************/ |
26 |
|
|
27 |
|
; **************************************************************************** |
28 |
|
; |
29 |
; Originally provided by Intel at AP-922 |
; Originally provided by Intel at AP-922 |
30 |
; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm |
31 |
; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) |
33 |
; New macro implements a column part for precise iDCT |
; New macro implements a column part for precise iDCT |
34 |
; The routine precision now satisfies IEEE standard 1180-1990. |
; The routine precision now satisfies IEEE standard 1180-1990. |
35 |
; |
; |
36 |
; Copyright (c) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
; Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
37 |
; Rounding trick Copyright (c) 2000 Michel Lespinasse <walken@zoy.org> |
; Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> |
38 |
; |
; |
39 |
; http://www.elecard.com/peter/idct.html |
; http://www.elecard.com/peter/idct.html |
40 |
; http://www.linuxvideo.org/mpeg2dec/ |
; http://www.linuxvideo.org/mpeg2dec/ |
41 |
; |
; |
42 |
;============================================================================= |
; ***************************************************************************/ |
43 |
; |
; |
44 |
; These examples contain code fragments for first stage iDCT 8x8 |
; These examples contain code fragments for first stage iDCT 8x8 |
45 |
; (for rows) and first stage DCT 8x8 (for columns) |
; (for rows) and first stage DCT 8x8 (for columns) |
46 |
; |
; |
|
;============================================================================= |
|
|
; |
|
|
; 04.11.2001 nasm conversion; peter ross <pross@cs.rmit.edu.au> |
|
|
; |
|
47 |
|
|
48 |
|
; ***************************************************************************/ |
49 |
; this 3dne function is compatible with iSSE, but is optimized specifically for |
; this 3dne function is compatible with iSSE, but is optimized specifically for |
50 |
; K7 pipelines (ca 5% gain) |
; K7 pipelines (ca 5% gain), for implementation details see the idct_mmx.asm |
51 |
|
; file |
52 |
; |
; |
53 |
;------------------------------------------------------------------------------ |
; ---------------------------------------------------------------------------- |
54 |
; 09.12.2002 Athlon optimizations contributed by Jaan Kalda |
; Athlon optimizations contributed by Jaan Kalda |
55 |
;------------------------------------------------------------------------------ |
;----------------------------------------------------------------------------- |
56 |
|
|
57 |
bits 32 |
;============================================================================= |
58 |
|
; Macros and other preprocessor constants |
59 |
%macro cglobal 1 |
;============================================================================= |
60 |
%ifdef PREFIX |
|
61 |
global _%1 |
%include "nasm.inc" |
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
62 |
|
|
63 |
%define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
%define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
64 |
%define SHIFT_INV_ROW 16 - BITS_INV_ACC |
%define SHIFT_INV_ROW 16 - BITS_INV_ACC |
72 |
%define SHIFT_FRW_ROW BITS_FRW_ACC + 17 |
%define SHIFT_FRW_ROW BITS_FRW_ACC + 17 |
73 |
%define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
%define RND_FRW_ROW 262144 * (BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
74 |
|
|
75 |
%ifdef FORMAT_COFF |
;============================================================================= |
76 |
section .data data |
; Local Data (Read Only) |
77 |
%else |
;============================================================================= |
|
section .data data align=16 |
|
|
%endif |
|
78 |
|
|
79 |
align 16 |
DATA |
80 |
|
|
81 |
one_corr dw 1, 1, 1, 1 |
;----------------------------------------------------------------------------- |
82 |
round_inv_row dd RND_INV_ROW, RND_INV_ROW |
; Various memory constants (trigonometric values or rounding values) |
83 |
round_inv_col dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL |
;----------------------------------------------------------------------------- |
|
round_inv_corr dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR |
|
|
round_frw_row dd RND_FRW_ROW, RND_FRW_ROW |
|
|
tg_1_16 dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
|
|
tg_2_16 dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
|
|
tg_3_16 dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
|
|
cos_4_16 dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 |
|
|
ocos_4_16 dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
|
84 |
|
|
85 |
otg_3_16 dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 |
ALIGN SECTION_ALIGN |
86 |
|
one_corr: |
87 |
|
dw 1, 1, 1, 1 |
88 |
|
round_inv_row: |
89 |
|
dd RND_INV_ROW, RND_INV_ROW |
90 |
|
round_inv_col: |
91 |
|
dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL |
92 |
|
round_inv_corr: |
93 |
|
dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR |
94 |
|
round_frw_row: |
95 |
|
dd RND_FRW_ROW, RND_FRW_ROW |
96 |
|
tg_1_16: |
97 |
|
dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 |
98 |
|
tg_2_16: |
99 |
|
dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 |
100 |
|
tg_3_16: |
101 |
|
dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 |
102 |
|
cos_4_16: |
103 |
|
dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 |
104 |
|
ocos_4_16: |
105 |
|
dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 |
106 |
|
otg_3_16: |
107 |
|
dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 |
108 |
|
|
109 |
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
%if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 |
110 |
rounder_0 dd 65536, 65536 |
rounder_0: |
111 |
rounder_4 dd 0, 0 |
dd 65536, 65536 |
112 |
rounder_1 dd 7195, 7195 |
rounder_4: |
113 |
rounder_7 dd 1024, 1024 |
dd 0, 0 |
114 |
rounder_2 dd 4520, 4520 |
rounder_1: |
115 |
rounder_6 dd 1024, 1024 |
dd 7195, 7195 |
116 |
rounder_3 dd 2407, 2407 |
rounder_7 |
117 |
rounder_5 dd 240, 240 |
dd 1024, 1024 |
118 |
|
rounder_2: |
119 |
|
dd 4520, 4520 |
120 |
|
rounder_6: |
121 |
|
dd 1024, 1024 |
122 |
|
rounder_3: |
123 |
|
dd 2407, 2407 |
124 |
|
rounder_5: |
125 |
|
dd 240, 240 |
126 |
|
|
127 |
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
%elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 |
128 |
rounder_0 dd 65536, 65536 |
rounder_0: |
129 |
rounder_4 dd 0, 0 |
dd 65536, 65536 |
130 |
rounder_1 dd 3597, 3597 |
rounder_4: |
131 |
rounder_7 dd 512, 512 |
dd 0, 0 |
132 |
rounder_2 dd 2260, 2260 |
rounder_1: |
133 |
rounder_6 dd 512, 512 |
dd 3597, 3597 |
134 |
rounder_3 dd 1203, 1203 |
rounder_7: |
135 |
rounder_5 dd 120, 120 |
dd 512, 512 |
136 |
|
rounder_2: |
137 |
|
dd 2260, 2260 |
138 |
|
rounder_6: |
139 |
|
dd 512, 512 |
140 |
|
rounder_3: |
141 |
|
dd 1203, 1203 |
142 |
|
rounder_5: |
143 |
|
dd 120, 120 |
144 |
%else |
%else |
145 |
|
|
146 |
%error invalid _SHIFT_INV_ROW_ |
%error invalid SHIFT_INV_ROW |
147 |
|
|
148 |
%endif |
%endif |
149 |
|
|
|
;============================================================================= |
|
|
; |
|
|
; The first stage iDCT 8x8 - inverse DCTs of rows |
|
|
; |
|
|
;----------------------------------------------------------------------------- |
|
|
; The 8-point inverse DCT direct algorithm |
|
|
;----------------------------------------------------------------------------- |
|
|
; |
|
|
; static const short w[32] = { |
|
|
; FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16), |
|
|
; FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16), |
|
|
; FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16), |
|
|
; FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16), |
|
|
; FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16), |
|
|
; FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16), |
|
|
; FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16), |
|
|
; FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) }; |
|
|
; |
|
|
; #define DCT_8_INV_ROW(x, y) |
|
|
; { |
|
|
; int a0, a1, a2, a3, b0, b1, b2, b3; |
|
|
; |
|
|
; a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3]; |
|
|
; a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7]; |
|
|
; a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11]; |
|
|
; a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15]; |
|
|
; b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19]; |
|
|
; b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23]; |
|
|
; b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27]; |
|
|
; b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31]; |
|
|
; |
|
|
; y[0] = SHIFT_ROUND ( a0 + b0 ); |
|
|
; y[1] = SHIFT_ROUND ( a1 + b1 ); |
|
|
; y[2] = SHIFT_ROUND ( a2 + b2 ); |
|
|
; y[3] = SHIFT_ROUND ( a3 + b3 ); |
|
|
; y[4] = SHIFT_ROUND ( a3 - b3 ); |
|
|
; y[5] = SHIFT_ROUND ( a2 - b2 ); |
|
|
; y[6] = SHIFT_ROUND ( a1 - b1 ); |
|
|
; y[7] = SHIFT_ROUND ( a0 - b0 ); |
|
|
; } |
|
|
; |
|
150 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
151 |
; |
; Tables for xmm processors |
|
; In this implementation the outputs of the iDCT-1D are multiplied |
|
|
; for rows 0,4 - by cos_4_16, |
|
|
; for rows 1,7 - by cos_1_16, |
|
|
; for rows 2,6 - by cos_2_16, |
|
|
; for rows 3,5 - by cos_3_16 |
|
|
; and are shifted to the left for better accuracy |
|
|
; |
|
|
; For the constants used, |
|
|
; FIX(float_const) = (short) (float_const * (1<<15) + 0.5) |
|
|
; |
|
|
;============================================================================= |
|
|
|
|
|
;============================================================================= |
|
|
; MMX code |
|
|
;============================================================================= |
|
|
|
|
|
; Table for rows 0,4 - constants are multiplied by cos_4_16 |
|
|
|
|
|
tab_i_04 dw 16384, 16384, 16384, -16384 ; movq-> w06 w04 w02 w00 |
|
|
dw 21407, 8867, 8867, -21407 ; w07 w05 w03 w01 |
|
|
dw 16384, -16384, 16384, 16384 ; w14 w12 w10 w08 |
|
|
dw -8867, 21407, -21407, -8867 ; w15 w13 w11 w09 |
|
|
dw 22725, 12873, 19266, -22725 ; w22 w20 w18 w16 |
|
|
dw 19266, 4520, -4520, -12873 ; w23 w21 w19 w17 |
|
|
dw 12873, 4520, 4520, 19266 ; w30 w28 w26 w24 |
|
|
dw -22725, 19266, -12873, -22725 ; w31 w29 w27 w25 |
|
|
|
|
|
; Table for rows 1,7 - constants are multiplied by cos_1_16 |
|
|
|
|
|
tab_i_17 dw 22725, 22725, 22725, -22725 ; movq-> w06 w04 w02 w00 |
|
|
dw 29692, 12299, 12299, -29692 ; w07 w05 w03 w01 |
|
|
dw 22725, -22725, 22725, 22725 ; w14 w12 w10 w08 |
|
|
dw -12299, 29692, -29692, -12299 ; w15 w13 w11 w09 |
|
|
dw 31521, 17855, 26722, -31521 ; w22 w20 w18 w16 |
|
|
dw 26722, 6270, -6270, -17855 ; w23 w21 w19 w17 |
|
|
dw 17855, 6270, 6270, 26722 ; w30 w28 w26 w24 |
|
|
dw -31521, 26722, -17855, -31521 ; w31 w29 w27 w25 |
|
|
|
|
|
; Table for rows 2,6 - constants are multiplied by cos_2_16 |
|
|
|
|
|
tab_i_26 dw 21407, 21407, 21407, -21407 ; movq-> w06 w04 w02 w00 |
|
|
dw 27969, 11585, 11585, -27969 ; w07 w05 w03 w01 |
|
|
dw 21407, -21407, 21407, 21407 ; w14 w12 w10 w08 |
|
|
dw -11585, 27969, -27969, -11585 ; w15 w13 w11 w09 |
|
|
dw 29692, 16819, 25172, -29692 ; w22 w20 w18 w16 |
|
|
dw 25172, 5906, -5906, -16819 ; w23 w21 w19 w17 |
|
|
dw 16819, 5906, 5906, 25172 ; w30 w28 w26 w24 |
|
|
dw -29692, 25172, -16819, -29692 ; w31 w29 w27 w25 |
|
|
|
|
|
; Table for rows 3,5 - constants are multiplied by cos_3_16 |
|
|
|
|
|
tab_i_35 dw 19266, 19266, 19266, -19266 ; movq-> w06 w04 w02 w00 |
|
|
dw 25172, 10426, 10426, -25172 ; w07 w05 w03 w01 |
|
|
dw 19266, -19266, 19266, 19266 ; w14 w12 w10 w08 |
|
|
dw -10426, 25172, -25172, -10426 ; w15 w13 w11 w09 |
|
|
dw 26722, 15137, 22654, -26722 ; w22 w20 w18 w16 |
|
|
dw 22654, 5315, -5315, -15137 ; w23 w21 w19 w17 |
|
|
dw 15137, 5315, 5315, 22654 ; w30 w28 w26 w24 |
|
|
dw -26722, 22654, -15137, -26722 ; w31 w29 w27 w25 |
|
|
|
|
152 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
153 |
|
|
|
; |
|
|
; DCT_8_INV_ROW_1 INP, OUT, TABLE, ROUNDER |
|
|
; |
|
|
|
|
|
%macro DCT_8_INV_ROW_1 4 |
|
|
|
|
|
movq mm0, [%1] ; 0 ; x3 x2 x1 x0 |
|
|
|
|
|
movq mm1, [%1+8] ; 1 ; x7 x6 x5 x4 |
|
|
movq mm2, mm0 ; 2 ; x3 x2 x1 x0 |
|
|
|
|
|
movq mm3, [%3] ; 3 ; w06 w04 w02 w00 |
|
|
punpcklwd mm0, mm1 ; x5 x1 x4 x0 |
|
|
|
|
|
movq mm5, mm0 ; 5 ; x5 x1 x4 x0 |
|
|
punpckldq mm0, mm0 ; x4 x0 x4 x0 |
|
|
|
|
|
movq mm4, [%3+8] ; 4 ; w07 w05 w03 w01 |
|
|
punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 |
|
|
|
|
|
pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 |
|
|
movq mm6, mm2 ; 6 ; x7 x3 x6 x2 |
|
|
|
|
|
movq mm1, [%3+32] ; 1 ; w22 w20 w18 w16 |
|
|
punpckldq mm2, mm2 ; x6 x2 x6 x2 |
|
|
|
|
|
pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 |
|
|
punpckhdq mm5, mm5 ; x5 x1 x5 x1 |
|
|
|
|
|
pmaddwd mm0, [%3+16] ; x4*w14+x0*w12 x4*w10+x0*w08 |
|
|
punpckhdq mm6, mm6 ; x7 x3 x7 x3 |
|
|
|
|
|
movq mm7, [%3+40] ; 7 ; w23 w21 w19 w17 |
|
|
pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 |
|
|
|
|
|
paddd mm3, [%4] ; +%4 |
|
|
pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 |
|
|
|
|
|
pmaddwd mm2, [%3+24] ; x6*w15+x2*w13 x6*w11+x2*w09 |
|
|
paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) |
|
|
|
|
|
pmaddwd mm5, [%3+48] ; x5*w30+x1*w28 x5*w26+x1*w24 |
|
|
movq mm4, mm3 ; 4 ; a1 a0 |
|
|
|
|
|
pmaddwd mm6, [%3+56] ; x7*w31+x3*w29 x7*w27+x3*w25 |
|
|
paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) |
|
|
|
|
|
paddd mm0, [%4] ; +%4 |
|
|
psubd mm3, mm1 ; a1-b1 a0-b0 |
|
|
|
|
|
psrad mm3, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
|
|
paddd mm1, mm4 ; 4 ; a1+b1 a0+b0 |
|
|
|
|
|
paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) |
|
|
psrad mm1, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
|
|
|
|
|
paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) |
|
|
movq mm4, mm0 ; 4 ; a3 a2 |
|
|
|
|
|
paddd mm0, mm5 ; a3+b3 a2+b2 |
|
|
psubd mm4, mm5 ; 5 ; a3-b3 a2-b2 |
|
|
|
|
|
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
|
|
psrad mm4, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
|
|
|
|
|
packssdw mm1, mm0 ; 0 ; y3 y2 y1 y0 |
|
|
packssdw mm4, mm3 ; 3 ; y6 y7 y4 y5 |
|
|
|
|
|
movq mm7, mm4 ; 7 ; y6 y7 y4 y5 |
|
|
psrld mm4, 16 ; 0 y6 0 y4 |
|
|
|
|
|
pslld mm7, 16 ; y7 0 y5 0 |
|
|
movq [%2], mm1 ; 1 ; save y3 y2 y1 y0 |
|
|
|
|
|
por mm7, mm4 ; 4 ; y7 y6 y5 y4 |
|
|
movq [%2+8], mm7 ; 7 ; save y7 y6 y5 y4 |
|
|
%endmacro |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
;============================================================================= |
|
|
; code for Pentium III |
|
|
;============================================================================= |
|
|
|
|
154 |
; %3 for rows 0,4 - constants are multiplied by cos_4_16 |
; %3 for rows 0,4 - constants are multiplied by cos_4_16 |
155 |
|
tab_i_04_xmm: |
156 |
tab_i_04_sse dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 |
157 |
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 |
158 |
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 |
159 |
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 |
163 |
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 |
164 |
|
|
165 |
; %3 for rows 1,7 - constants are multiplied by cos_1_16 |
; %3 for rows 1,7 - constants are multiplied by cos_1_16 |
166 |
|
tab_i_17_xmm: |
167 |
tab_i_17_sse dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 |
168 |
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 |
169 |
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 |
170 |
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 |
174 |
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 |
175 |
|
|
176 |
; %3 for rows 2,6 - constants are multiplied by cos_2_16 |
; %3 for rows 2,6 - constants are multiplied by cos_2_16 |
177 |
|
tab_i_26_xmm: |
178 |
tab_i_26_sse dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 |
179 |
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 |
180 |
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 |
181 |
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 |
185 |
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 |
186 |
|
|
187 |
; %3 for rows 3,5 - constants are multiplied by cos_3_16 |
; %3 for rows 3,5 - constants are multiplied by cos_3_16 |
188 |
|
tab_i_35_xmm: |
189 |
tab_i_35_sse dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 |
190 |
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 |
191 |
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 |
192 |
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 |
196 |
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 |
197 |
|
|
198 |
;============================================================================= |
;============================================================================= |
199 |
; |
; Code |
|
;============================================================================= |
|
|
|
|
|
;============================================================================= |
|
|
; |
|
|
; The first stage DCT 8x8 - forward DCTs of columns |
|
|
; |
|
|
; The %2puts 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; |
|
|
;} |
|
|
; |
|
200 |
;============================================================================= |
;============================================================================= |
201 |
|
|
202 |
|
TEXT |
203 |
|
|
204 |
; |
cglobal idct_3dne |
|
; DCT_8_FRW_COL_4 INP, OUT |
|
|
; |
|
|
|
|
|
%macro DCT_8_FRW_COL_4 2 |
|
|
|
|
|
LOCAL x0, x1, x2, x3, x4, x5, x6, x7 |
|
|
LOCAL y0, y1, y2, y3, y4, y5, y6, y7 |
|
|
x0 equ [%1 + 0*16] |
|
|
x1 equ [%1 + 1*16] |
|
|
x2 equ [%1 + 2*16] |
|
|
x3 equ [%1 + 3*16] |
|
|
x4 equ [%1 + 4*16] |
|
|
x5 equ [%1 + 5*16] |
|
|
x6 equ [%1 + 6*16] |
|
|
x7 equ [%1 + 7*16] |
|
|
y0 equ [%2 + 0*16] |
|
|
y1 equ [%2 + 1*16] |
|
|
y2 equ [%2 + 2*16] |
|
|
y3 equ [%2 + 3*16] |
|
|
y4 equ [%2 + 4*16] |
|
|
y5 equ [%2 + 5*16] |
|
|
y6 equ [%2 + 6*16] |
|
|
y7 equ [%2 + 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, [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, [one_corr] ; correction y2 +0.5 |
|
|
psllw mm2, SHIFT_FRW_COL+1 ; t6 |
|
|
pmulhw mm5, [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, [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16 |
|
|
psubsw mm6, mm3 ; 3 ; t6 - t5 |
|
|
pmulhw mm6, [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16 |
|
|
psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12 |
|
|
por mm5, [one_corr] ; correction y6 +0.5 |
|
|
psllw mm1, SHIFT_FRW_COL ; t4 |
|
|
por mm2, [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, [tg_1_16] ; 0 ; tg_1_16 |
|
|
psllw mm3, SHIFT_FRW_COL ; t7 |
|
|
movq mm6, [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, [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, [tg_1_16] ; tp765*tg_1_16 |
|
|
por mm0, [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 |
|
|
%endmacro |
|
|
|
|
|
|
|
|
; |
|
|
; DCT_8_INV_COL_4 INP,OUT |
|
|
; |
|
|
|
|
|
%macro DCT_8_INV_COL_4 2 |
|
|
movq mm0, [tg_3_16] |
|
|
|
|
|
movq mm3, [%1+16*3] |
|
|
movq mm1, mm0 ; tg_3_16 |
|
|
|
|
|
movq mm5, [%1+16*5] |
|
|
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
|
|
|
|
|
movq mm4, [tg_1_16] |
|
|
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
|
|
|
|
|
movq mm7, [%1+16*7] |
|
|
movq mm2, mm4 ; tg_1_16 |
|
|
|
|
|
movq mm6, [%1+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, [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, [tg_2_16] |
|
|
movq mm1, mm4 ; t1 |
|
|
|
|
|
; movq [SCRATCH+0], mm5 ; save b0 |
|
|
movq [%2+3*16], mm5 ; save b0 |
|
|
paddsw mm1, mm2 ; t1+t2 |
|
|
|
|
|
; movq [SCRATCH+8], mm6 ; save b3 |
|
|
movq [%2+5*16], mm6 ; save b3 |
|
|
psubsw mm4, mm2 ; t1-t2 |
|
|
|
|
|
movq mm5, [%1+2*16] |
|
|
movq mm0, mm7 ; tg_2_16 |
|
|
|
|
|
movq mm6, [%1+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, [%1+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, [%1+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, [SCRATCH+0] ; load b0 |
|
|
movq mm1, [%2+3*16] ; load b0 |
|
|
psraw mm7, SHIFT_INV_COL ; dst6 |
|
|
|
|
|
movq mm4, mm5 ; a0 |
|
|
psraw mm0, SHIFT_INV_COL ; dst5 |
|
|
|
|
|
movq [%2+1*16], mm3 |
|
|
paddsw mm5, mm1 ; a0+b0 |
|
|
|
|
|
movq [%2+2*16], mm6 |
|
|
psubsw mm4, mm1 ; a0-b0 |
|
|
|
|
|
; movq mm3, [SCRATCH+8] ; load b3 |
|
|
movq mm3, [%2+5*16] ; load b3 |
|
|
psraw mm5, SHIFT_INV_COL ; dst0 |
|
|
|
|
|
movq mm6, mm2 ; a3 |
|
|
psraw mm4, SHIFT_INV_COL ; dst7 |
|
|
|
|
|
movq [%2+5*16], mm0 |
|
|
paddsw mm2, mm3 ; a3+b3 |
|
|
|
|
|
movq [%2+6*16], mm7 |
|
|
psubsw mm6, mm3 ; a3-b3 |
|
|
|
|
|
movq [%2+0*16], mm5 |
|
|
psraw mm2, SHIFT_INV_COL ; dst3 |
|
|
|
|
|
movq [%2+7*16], mm4 |
|
|
psraw mm6, SHIFT_INV_COL ; dst4 |
|
|
|
|
|
movq [%2+3*16], mm2 |
|
|
|
|
|
movq [%2+4*16], mm6 |
|
|
%endmacro |
|
|
|
|
205 |
|
|
206 |
;============================================================================= |
;----------------------------------------------------------------------------- |
207 |
; |
; void idct_3dne(uint16_t block[64]); |
208 |
; void idct_sse (short * const src_result); |
;----------------------------------------------------------------------------- |
|
; |
|
|
;============================================================================= |
|
209 |
|
|
210 |
align 16 |
ALIGN SECTION_ALIGN |
211 |
cglobal idct_3dne |
idct_3dne: |
212 |
idct_3dne |
mov _ECX, prm1 |
213 |
mov eax,[esp+4] |
|
214 |
; DCT_8_INV_ROW_1_s [eax+64], [eax+64], tab_i_04_sse, rounder_4 ;rounder_4=0 |
; DCT_8_INV_ROW_1_s [_ECX+64], [_ECX+64], tab_i_04_sse, rounder_4 ;rounder_4=0 |
215 |
pshufw mm0, [eax+64],10001000b ; x2 x0 x2 x0 |
pshufw mm0, [_ECX+64],10001000b ; x2 x0 x2 x0 |
216 |
movq mm3, [tab_i_04_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 |
217 |
pshufw mm1, [eax+64+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+64+8],10001000b ; x6 x4 x6 x4 |
218 |
movq mm4, [tab_i_04_sse+8] ; 4 ; w07 w06 w03 w02 |
movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 |
219 |
pshufw mm2, [eax+64],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+64],11011101b ; x3 x1 x3 x1 |
220 |
pshufw mm5, [eax+64+8],11011101b ; x7 x5 x7 x5 |
pshufw mm5, [_ECX+64+8],11011101b ; x7 x5 x7 x5 |
221 |
movq mm6, [tab_i_04_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 |
222 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
223 |
movq mm7, [tab_i_04_sse+40] ; 7 ; w23 w22 w19 w18 ; |
movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 ; |
224 |
pmaddwd mm0, [tab_i_04_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm0, [tab_i_04_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
225 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
226 |
pmaddwd mm1, [tab_i_04_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_04_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
227 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
228 |
pmaddwd mm2, [tab_i_04_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_04_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
229 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
230 |
pmaddwd mm5, [tab_i_04_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm5, [tab_i_04_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
231 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
232 |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
233 |
pshufw mm1, [eax+80+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+80+8],10001000b ; x6 x4 x6 x4 |
234 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
235 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
236 |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
237 |
pshufw mm5, [eax+80],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
pshufw mm5, [_ECX+80],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
238 |
movq mm7, mm0 ; 7 ; a3 a2 |
movq mm7, mm0 ; 7 ; a3 a2 |
239 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
240 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
241 |
movq mm3, [tab_i_35_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 |
242 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
243 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
244 |
pshufw mm2, [eax+80],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+80],11011101b; x3 x1 x3 x1 |
245 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
246 |
pmaddwd mm5, [tab_i_35_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm5, [tab_i_35_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
247 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
248 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
249 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
250 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
251 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
252 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
253 |
pshufw mm0, [eax+80+8],11011101b ; x7 x5 x7 x5 |
pshufw mm0, [_ECX+80+8],11011101b ; x7 x5 x7 x5 |
254 |
movq [eax+64], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
movq [_ECX+64], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
255 |
; DCT_8_INV_ROW_1_s [eax+80], [eax+80], tab_i_35_sse, rounder_5 |
|
256 |
movq mm4, [tab_i_35_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+80], [_ECX+80], tab_i_35_xmm, rounder_5 |
257 |
movq mm6, [tab_i_35_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 |
258 |
|
movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 |
259 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
260 |
paddd mm3, [rounder_5] ; +rounder stall 6 |
paddd mm3, [rounder_5] ; +rounder stall 6 |
261 |
paddd mm5, [rounder_5] ; +rounder |
paddd mm5, [rounder_5] ; +rounder |
262 |
movq [eax+64+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [_ECX+64+8], mm7 ; 7 ; save y7 y6 y5 y4 |
263 |
movq mm7, [tab_i_35_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 |
264 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
265 |
pmaddwd mm1, [tab_i_35_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_35_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
266 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
267 |
pmaddwd mm2, [tab_i_35_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_35_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
268 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
269 |
pmaddwd mm0, [tab_i_35_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm0, [tab_i_35_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
270 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
271 |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
272 |
pshufw mm1, [eax+96+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+96+8],10001000b ; x6 x4 x6 x4 |
273 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
274 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
275 |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
276 |
pshufw mm0, [eax+96],10001000b ; x2 x0 x2 x0 |
pshufw mm0, [_ECX+96],10001000b ; x2 x0 x2 x0 |
277 |
movq mm7, mm5 ; 7 ; a3 a2 |
movq mm7, mm5 ; 7 ; a3 a2 |
278 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
279 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
280 |
movq mm3, [tab_i_26_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 |
281 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
282 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
283 |
pshufw mm2, [eax+96],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+96],11011101b; x3 x1 x3 x1 |
284 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
285 |
pmaddwd mm0, [tab_i_26_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm0, [tab_i_26_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
286 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
287 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
288 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
289 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
290 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
291 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
292 |
pshufw mm5, [eax+96+8],11011101b ; x7 x5 x7 x5 |
pshufw mm5, [_ECX+96+8],11011101b ; x7 x5 x7 x5 |
293 |
movq [eax+80], mm6 ; 3 ; save y3 y2 y1 y0 |
movq [_ECX+80], mm6 ; 3 ; save y3 y2 y1 y0 |
294 |
; DCT_8_INV_ROW_1_s [eax+96], [eax+96], tab_i_26_sse, rounder_6 |
|
295 |
movq mm4, [tab_i_26_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+96], [_ECX+96], tab_i_26_xmm, rounder_6 |
296 |
movq mm6, [tab_i_26_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 |
297 |
|
movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 |
298 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
299 |
paddd mm3, [rounder_6] ; +rounder |
paddd mm3, [rounder_6] ; +rounder |
300 |
paddd mm0, [rounder_6] ; +rounder |
paddd mm0, [rounder_6] ; +rounder |
301 |
movq [eax+80+8], mm7 ; 7 ; save y7 y6 |
movq [_ECX+80+8], mm7 ; 7 ; save y7 y6 |
302 |
movq mm7, [tab_i_26_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 |
303 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
304 |
pmaddwd mm1, [tab_i_26_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_26_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
305 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
306 |
pmaddwd mm2, [tab_i_26_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_26_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
307 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
308 |
pmaddwd mm5, [tab_i_26_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm5, [tab_i_26_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
309 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
310 |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
311 |
pshufw mm1, [eax+112+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+112+8],10001000b ; x6 x4 x6 x4 |
312 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
313 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
314 |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
315 |
pshufw mm5, [eax+112],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
pshufw mm5, [_ECX+112],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
316 |
movq mm7, mm0 ; 7 ; a3 a2 |
movq mm7, mm0 ; 7 ; a3 a2 |
317 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
318 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
319 |
movq mm3, [tab_i_17_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 |
320 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
321 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
322 |
pshufw mm2, [eax+112],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+112],11011101b; x3 x1 x3 x1 |
323 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
324 |
pmaddwd mm5, [tab_i_17_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm5, [tab_i_17_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
325 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
326 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
327 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
328 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
329 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
330 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
331 |
pshufw mm0, [eax+112+8],11011101b ; x7 x5 x7 x5 |
pshufw mm0, [_ECX+112+8],11011101b ; x7 x5 x7 x5 |
332 |
movq [eax+96], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
movq [_ECX+96], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
333 |
; DCT_8_INV_ROW_1_s [eax+112], [eax+112], tab_i_17_sse, rounder_7 |
|
334 |
movq mm4, [tab_i_17_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+112], [_ECX+112], tab_i_17_xmm, rounder_7 |
335 |
movq mm6, [tab_i_17_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 |
336 |
|
movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 |
337 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
338 |
paddd mm3, [rounder_7] ; +rounder stall 6 |
paddd mm3, [rounder_7] ; +rounder stall 6 |
339 |
paddd mm5, [rounder_7] ; +rounder |
paddd mm5, [rounder_7] ; +rounder |
340 |
movq [eax+96+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [_ECX+96+8], mm7 ; 7 ; save y7 y6 y5 y4 |
341 |
movq mm7, [tab_i_17_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 |
342 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
343 |
pmaddwd mm1, [tab_i_17_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_17_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
344 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
345 |
pmaddwd mm2, [tab_i_17_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_17_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
346 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
347 |
pmaddwd mm0, [tab_i_17_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm0, [tab_i_17_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
348 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
349 |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
350 |
pshufw mm1, [eax+0+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+0+8],10001000b; x6 x4 x6 x4 |
351 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
352 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
353 |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
354 |
pshufw mm0, [eax+0],10001000b ; x2 x0 x2 x0 |
pshufw mm0, [_ECX+0],10001000b ; x2 x0 x2 x0 |
355 |
movq mm7, mm5 ; 7 ; a3 a2 |
movq mm7, mm5 ; 7 ; a3 a2 |
356 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
357 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
358 |
movq mm3, [tab_i_04_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 |
359 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
360 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
361 |
pshufw mm2, [eax+0],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+0],11011101b ; x3 x1 x3 x1 |
362 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
363 |
pmaddwd mm0, [tab_i_04_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm0, [tab_i_04_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
364 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
365 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
366 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
367 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
368 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
369 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
370 |
pshufw mm5, [eax+0+8],11011101b ; x7 x5 x7 x5 |
pshufw mm5, [_ECX+0+8],11011101b; x7 x5 x7 x5 |
371 |
movq [eax+112], mm6 ; 3 ; save y3 y2 y1 y0 |
movq [_ECX+112], mm6 ; 3 ; save y3 y2 y1 y0 |
372 |
; DCT_8_INV_ROW_1_s [eax+0], 0, tab_i_04_sse, rounder_0 |
|
373 |
movq mm4, [tab_i_04_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+0], 0, tab_i_04_xmm, rounder_0 |
374 |
movq mm6, [tab_i_04_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 |
375 |
|
movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 |
376 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
377 |
paddd mm3, [rounder_0] ; +rounder |
paddd mm3, [rounder_0] ; +rounder |
378 |
paddd mm0, [rounder_0] ; +rounder |
paddd mm0, [rounder_0] ; +rounder |
379 |
movq [eax+112+8], mm7 ; 7 ; save y7 y6 |
movq [_ECX+112+8], mm7 ; 7 ; save y7 y6 |
380 |
movq mm7, [tab_i_04_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 |
381 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
382 |
pmaddwd mm1, [tab_i_04_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_04_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
383 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
384 |
pmaddwd mm2, [tab_i_04_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_04_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
385 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
386 |
pmaddwd mm5, [tab_i_04_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm5, [tab_i_04_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
387 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
388 |
paddd mm0, mm1 ; 1 |
paddd mm0, mm1 ; 1 |
389 |
pshufw mm1, [eax+16+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+16+8],10001000b ; x6 x4 x6 x4 |
390 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
391 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
392 |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
393 |
pshufw mm5, [eax+16],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
pshufw mm5, [_ECX+16],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
394 |
movq mm7, mm0 ; 7 ; a3 a2 |
movq mm7, mm0 ; 7 ; a3 a2 |
395 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
396 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
397 |
movq mm3, [tab_i_17_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 |
398 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
399 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
400 |
pshufw mm2, [eax+16],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+16],11011101b; x3 x1 x3 x1 |
401 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
402 |
pmaddwd mm5, [tab_i_17_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm5, [tab_i_17_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
403 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
404 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
405 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
406 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
407 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
408 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
409 |
pshufw mm0, [eax+16+8],11011101b ; x7 x5 x7 x5 |
pshufw mm0, [_ECX+16+8],11011101b ; x7 x5 x7 x5 |
410 |
movq [eax+0], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
movq [_ECX+0], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
411 |
; DCT_8_INV_ROW_1_s [eax+16], 16, tab_i_17_sse, rounder_1 |
|
412 |
movq mm4, [tab_i_17_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+16], 16, tab_i_17_xmm, rounder_1 |
413 |
movq mm6, [tab_i_17_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 |
414 |
|
movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 |
415 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
416 |
paddd mm3, [rounder_1] ; +rounder stall 6 |
paddd mm3, [rounder_1] ; +rounder stall 6 |
417 |
paddd mm5, [rounder_1] ; +rounder |
paddd mm5, [rounder_1] ; +rounder |
418 |
movq [eax+0+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [_ECX+0+8], mm7 ; 7 ; save y7 y6 y5 y4 |
419 |
movq mm7, [tab_i_17_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 |
420 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
421 |
pmaddwd mm1, [tab_i_17_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_17_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
422 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
423 |
pmaddwd mm2, [tab_i_17_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_17_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
424 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
425 |
pmaddwd mm0, [tab_i_17_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm0, [tab_i_17_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
426 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
427 |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
428 |
pshufw mm1, [eax+32+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+32+8],10001000b ; x6 x4 x6 x4 |
429 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
430 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
431 |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
432 |
pshufw mm0, [eax+32],10001000b ; x2 x0 x2 x0 |
pshufw mm0, [_ECX+32],10001000b; x2 x0 x2 x0 |
433 |
movq mm7, mm5 ; 7 ; a3 a2 |
movq mm7, mm5 ; 7 ; a3 a2 |
434 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 |
435 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
436 |
movq mm3, [tab_i_26_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 |
437 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
438 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm5, mm2 ; 0 free a3+b3 a2+b2 |
439 |
pshufw mm2, [eax+32],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+32],11011101b; x3 x1 x3 x1 |
440 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 |
441 |
pmaddwd mm0, [tab_i_26_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm0, [tab_i_26_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
442 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
443 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
444 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
445 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
446 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
447 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 |
448 |
pshufw mm5, [eax+32+8],11011101b ; x7 x5 x7 x5 |
pshufw mm5, [_ECX+32+8],11011101b ; x7 x5 x7 x5 |
449 |
movq [eax+16], mm6 ; 3 ; save y3 y2 y1 y0 |
movq [_ECX+16], mm6 ; 3 ; save y3 y2 y1 y0 |
450 |
; DCT_8_INV_ROW_1_s [eax+32], 32, tab_i_26_sse, rounder_2 |
|
451 |
movq mm4, [tab_i_26_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+32], 32, tab_i_26_xmm, rounder_2 |
452 |
movq mm6, [tab_i_26_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 |
453 |
|
movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 |
454 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 |
455 |
paddd mm3, [rounder_2] ; +rounder |
paddd mm3, [rounder_2] ; +rounder |
456 |
paddd mm0, [rounder_2] ; +rounder |
paddd mm0, [rounder_2] ; +rounder |
457 |
movq [eax+16+8], mm7 ; 7 ; save y7 y6 |
movq [_ECX+16+8], mm7 ; 7 ; save y7 y6 |
458 |
movq mm7, [tab_i_26_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 |
459 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
460 |
pmaddwd mm1, [tab_i_26_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_26_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
461 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
462 |
pmaddwd mm2, [tab_i_26_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_26_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
463 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
464 |
pmaddwd mm5, [tab_i_26_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm5, [tab_i_26_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
465 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
466 |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) |
467 |
pshufw mm1, [eax+48+8],10001000b ; x6 x4 x6 x4 |
pshufw mm1, [_ECX+48+8],10001000b ; x6 x4 x6 x4 |
468 |
movq mm4, mm3 ; 4 ; a1 a0 |
movq mm4, mm3 ; 4 ; a1 a0 |
469 |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) |
470 |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) |
471 |
pshufw mm5, [eax+48],10001000b ; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
pshufw mm5, [_ECX+48],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle |
472 |
movq mm7, mm0 ; 7 ; a3 a2 |
movq mm7, mm0 ; 7 ; a3 a2 |
473 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 |
474 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
475 |
movq mm3, [tab_i_35_sse] ; 3 ; w05 w04 w01 w00 |
movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 |
476 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
477 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
paddd mm0, mm2 ; 0 free a3+b3 a2+b2 |
478 |
pshufw mm2, [eax+48],11011101b ; x3 x1 x3 x1 |
pshufw mm2, [_ECX+48],11011101b; x3 x1 x3 x1 |
479 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 |
480 |
pmaddwd mm5, [tab_i_35_sse+16] ; x2*w13+x0*w12 x2*w09+x0*w08 |
pmaddwd mm5, [tab_i_35_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 |
481 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
482 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
483 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
484 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
485 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
486 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 |
487 |
pshufw mm0, [eax+48+8],11011101b ; x7 x5 x7 x5 |
pshufw mm0, [_ECX+48+8],11011101b ; x7 x5 x7 x5 |
488 |
movq [eax+32], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
movq [_ECX+32], mm6 ; 3 ; save y3 y2 y1 y0 stall2 |
489 |
; DCT_8_INV_ROW_1_s [eax+48], [eax+48], tab_i_35_sse, rounder_3 |
|
490 |
movq mm4, [tab_i_35_sse+8] ; 4 ; w07 w06 w03 w02 |
; DCT_8_INV_ROW_1_s [_ECX+48], [_ECX+48], tab_i_35_xmm, rounder_3 |
491 |
movq mm6, [tab_i_35_sse+32] ; 6 ; w21 w20 w17 w16 |
movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 |
492 |
|
movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 |
493 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 |
494 |
paddd mm3, [rounder_3] ; +rounder stall 6 |
paddd mm3, [rounder_3] ; +rounder stall 6 |
495 |
paddd mm5, [rounder_3] ; +rounder |
paddd mm5, [rounder_3] ; +rounder |
496 |
movq [eax+32+8], mm7 ; 7 ; save y7 y6 y5 y4 |
movq [_ECX+32+8], mm7 ; 7 ; save y7 y6 y5 y4 |
497 |
movq mm7, [tab_i_35_sse+40] ; 7 ; w23 w22 w19 w18 |
movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 |
498 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 |
499 |
pmaddwd mm1, [tab_i_35_sse+24] ; x6*w15+x4*w14 x6*w11+x4*w10 |
pmaddwd mm1, [tab_i_35_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 |
500 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 |
501 |
pmaddwd mm2, [tab_i_35_sse+48] ; x3*w29+x1*w28 x3*w25+x1*w24 |
pmaddwd mm2, [tab_i_35_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 |
502 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 |
503 |
pmaddwd mm0, [tab_i_35_sse+56] ; x7*w31+x5*w30 x7*w27+x5*w26 |
pmaddwd mm0, [tab_i_35_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 |
504 |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) |
505 |
paddd mm5, mm1 ; mm1 free ; a3=sum(even3) a2=sum(even2) |
paddd mm5, mm1 ; mm1 free ; a3=sum(even3) a2=sum(even2) |
506 |
movq mm1, [tg_3_16] |
movq mm1, [tg_3_16] |
513 |
paddd mm3, mm6 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
paddd mm3, mm6 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 |
514 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
psubd mm7, mm2 ; ; a3-b3 a2-b2 |
515 |
paddd mm2, mm5 ; 0 free a3+b3 a2+b2 |
paddd mm2, mm5 ; 0 free a3+b3 a2+b2 |
516 |
movq mm5, [eax+16*5] |
movq mm5, [_ECX+16*5] |
517 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 |
518 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 |
519 |
psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 |
520 |
psrad mm2, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
psrad mm2, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 |
521 |
movq mm6, [eax+16*1] |
movq mm6, [_ECX+16*1] |
522 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 |
523 |
movq mm4, [tg_1_16] |
movq mm4, [tg_1_16] |
524 |
packssdw mm3, mm2 ; 0 free ; y3 y2 y1 y0 |
packssdw mm3, mm2 ; 0 free ; y3 y2 y1 y0 |
525 |
pshufw mm2, mm7, 10110001b ; y7 y6 y5 y4 |
pshufw mm2, mm7, 10110001b ; y7 y6 y5 y4 |
526 |
; DCT_8_INV_COL_4 [eax+0],[eax+0] |
|
527 |
; movq mm3,mmword ptr [eax+16*3] |
; DCT_8_INV_COL_4 [_ECX+0],[_ECX+0] |
528 |
movq mm7, [eax+16*7] |
; movq mm3,mmword ptr [_ECX+16*3] |
529 |
|
movq mm7, [_ECX+16*7] |
530 |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
531 |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
532 |
movq [eax+48+8], mm2 ; 7 ; save y7 y6 y5 y4 |
movq [_ECX+48+8], mm2 ; 7 ; save y7 y6 y5 y4 |
533 |
movq mm2, mm4 ; tg_1_16 |
movq mm2, mm4 ; tg_1_16 |
534 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
535 |
paddsw mm0, mm3 ; x3*tg_3_16 |
paddsw mm0, mm3 ; x3*tg_3_16 |
536 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
537 |
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) |
538 |
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 |
539 |
movq [eax+48], mm3 ; 3 ; save y3 y2 y1 y0 |
movq [_ECX+48], mm3 ; 3 ; save y3 y2 y1 y0 |
540 |
movq mm3, [ocos_4_16] |
movq mm3, [ocos_4_16] |
541 |
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 |
542 |
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 |
549 |
paddsw mm2, mm0 ; tm17+tm35 = t2 |
paddsw mm2, mm0 ; tm17+tm35 = t2 |
550 |
movq mm7, [tg_2_16] |
movq mm7, [tg_2_16] |
551 |
movq mm1, mm4 ; t1 |
movq mm1, mm4 ; t1 |
552 |
movq [eax+3*16], mm5 ; save b0 |
movq [_ECX+3*16], mm5 ; save b0 |
553 |
paddsw mm1, mm2 ; t1+t2 |
paddsw mm1, mm2 ; t1+t2 |
554 |
movq [eax+5*16], mm6 ; save b3 |
movq [_ECX+5*16], mm6 ; save b3 |
555 |
psubsw mm4, mm2 ; t1-t2 |
psubsw mm4, mm2 ; t1-t2 |
556 |
movq mm5, [eax+2*16] |
movq mm5, [_ECX+2*16] |
557 |
movq mm0, mm7 ; tg_2_16 |
movq mm0, mm7 ; tg_2_16 |
558 |
movq mm6, [eax+6*16] |
movq mm6, [_ECX+6*16] |
559 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
560 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
561 |
; slot |
; slot |
562 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
563 |
; slot |
; slot |
564 |
movq mm2, [eax+0*16] |
movq mm2, [_ECX+0*16] |
565 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
566 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
567 |
movq mm3, [eax+0*16] ; x0 |
movq mm3, [_ECX+0*16] ; x0 |
568 |
movq mm6, [eax+4*16] |
movq mm6, [_ECX+4*16] |
569 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
570 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
571 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
585 |
psubsw mm7, mm1 ; a1-b1 |
psubsw mm7, mm1 ; a1-b1 |
586 |
psraw mm6, SHIFT_INV_COL ; dst2 |
psraw mm6, SHIFT_INV_COL ; dst2 |
587 |
psubsw mm0, mm4 ; a2-b2 |
psubsw mm0, mm4 ; a2-b2 |
588 |
movq mm1, [eax+3*16] ; load b0 |
movq mm1, [_ECX+3*16] ; load b0 |
589 |
psraw mm7, SHIFT_INV_COL ; dst6 |
psraw mm7, SHIFT_INV_COL ; dst6 |
590 |
movq mm4, mm5 ; a0 |
movq mm4, mm5 ; a0 |
591 |
psraw mm0, SHIFT_INV_COL ; dst5 |
psraw mm0, SHIFT_INV_COL ; dst5 |
592 |
movq [eax+1*16], mm3 |
movq [_ECX+1*16], mm3 |
593 |
paddsw mm5, mm1 ; a0+b0 |
paddsw mm5, mm1 ; a0+b0 |
594 |
movq [eax+2*16], mm6 |
movq [_ECX+2*16], mm6 |
595 |
psubsw mm4, mm1 ; a0-b0 |
psubsw mm4, mm1 ; a0-b0 |
596 |
movq mm3, [eax+5*16] ; load b3 |
movq mm3, [_ECX+5*16] ; load b3 |
597 |
psraw mm5, SHIFT_INV_COL ; dst0 |
psraw mm5, SHIFT_INV_COL ; dst0 |
598 |
movq mm6, mm2 ; a3 |
movq mm6, mm2 ; a3 |
599 |
psraw mm4, SHIFT_INV_COL ; dst7 |
psraw mm4, SHIFT_INV_COL ; dst7 |
600 |
movq [eax+5*16], mm0 |
movq [_ECX+5*16], mm0 |
601 |
movq mm0, [tg_3_16] |
movq mm0, [tg_3_16] |
602 |
paddsw mm2, mm3 ; a3+b3 |
paddsw mm2, mm3 ; a3+b3 |
603 |
movq [eax+6*16], mm7 |
movq [_ECX+6*16], mm7 |
604 |
psubsw mm6, mm3 ; a3-b3 |
psubsw mm6, mm3 ; a3-b3 |
605 |
movq mm3, [eax+8+16*3] |
movq mm3, [_ECX+8+16*3] |
606 |
movq [eax+0*16], mm5 |
movq [_ECX+0*16], mm5 |
607 |
psraw mm2, SHIFT_INV_COL ; dst3 |
psraw mm2, SHIFT_INV_COL ; dst3 |
608 |
movq [eax+7*16], mm4 |
movq [_ECX+7*16], mm4 |
|
; DCT_8_INV_COL_4 [eax+8],[eax+8] |
|
609 |
|
|
610 |
|
; DCT_8_INV_COL_4 [_ECX+8],[_ECX+8] |
611 |
movq mm1, mm0 ; tg_3_16 |
movq mm1, mm0 ; tg_3_16 |
612 |
movq mm5, [eax+8+16*5] |
movq mm5, [_ECX+8+16*5] |
613 |
psraw mm6, SHIFT_INV_COL ; dst4 |
psraw mm6, SHIFT_INV_COL ; dst4 |
614 |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
pmulhw mm0, mm3 ; x3*(tg_3_16-1) |
615 |
movq mm4, [tg_1_16] |
movq mm4, [tg_1_16] |
616 |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
pmulhw mm1, mm5 ; x5*(tg_3_16-1) |
617 |
movq mm7, [eax+8+16*7] |
movq mm7, [_ECX+8+16*7] |
618 |
movq [eax+3*16], mm2 |
movq [_ECX+3*16], mm2 |
619 |
movq mm2, mm4 ; tg_1_16 |
movq mm2, mm4 ; tg_1_16 |
620 |
movq [eax+4*16], mm6 |
movq [_ECX+4*16], mm6 |
621 |
movq mm6, [eax+8+16*1] |
movq mm6, [_ECX+8+16*1] |
622 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
pmulhw mm4, mm7 ; x7*tg_1_16 |
623 |
paddsw mm0, mm3 ; x3*tg_3_16 |
paddsw mm0, mm3 ; x3*tg_3_16 |
624 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
pmulhw mm2, mm6 ; x1*tg_1_16 |
636 |
movq mm7, [tg_2_16] |
movq mm7, [tg_2_16] |
637 |
movq mm1, mm4 ; t1 |
movq mm1, mm4 ; t1 |
638 |
psubsw mm6, mm0 ; tm17-tm35 = b3 |
psubsw mm6, mm0 ; tm17-tm35 = b3 |
639 |
movq [eax+8+3*16], mm5 ; save b0 |
movq [_ECX+8+3*16], mm5 ; save b0 |
640 |
movq [eax+8+5*16], mm6 ; save b3 |
movq [_ECX+8+5*16], mm6 ; save b3 |
641 |
psubsw mm4, mm2 ; t1-t2 |
psubsw mm4, mm2 ; t1-t2 |
642 |
movq mm5, [eax+8+2*16] |
movq mm5, [_ECX+8+2*16] |
643 |
movq mm0, mm7 ; tg_2_16 |
movq mm0, mm7 ; tg_2_16 |
644 |
movq mm6, [eax+8+6*16] |
movq mm6, [_ECX+8+6*16] |
645 |
paddsw mm1, mm2 ; t1+t2 |
paddsw mm1, mm2 ; t1+t2 |
646 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
pmulhw mm0, mm5 ; x2*tg_2_16 |
647 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
pmulhw mm7, mm6 ; x6*tg_2_16 |
648 |
movq mm2, [eax+8+0*16] |
movq mm2, [_ECX+8+0*16] |
649 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 |
650 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 |
651 |
; slot |
; slot |
652 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
653 |
; slot |
; slot |
654 |
movq mm3, [eax+8+0*16] ; x0 |
movq mm3, [_ECX+8+0*16] ; x0 |
655 |
movq mm6, [eax+8+4*16] |
movq mm6, [_ECX+8+4*16] |
656 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 |
657 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
paddsw mm2, mm6 ; x0+x4 = tp04 |
658 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
psubsw mm3, mm6 ; x0-x4 = tm04 |
672 |
psubsw mm7, mm1 ; a1-b1 |
psubsw mm7, mm1 ; a1-b1 |
673 |
psraw mm6, SHIFT_INV_COL ; dst2 |
psraw mm6, SHIFT_INV_COL ; dst2 |
674 |
psubsw mm0, mm4 ; a2-b2 |
psubsw mm0, mm4 ; a2-b2 |
675 |
movq mm1, [eax+8+3*16] ; load b0 |
movq mm1, [_ECX+8+3*16] ; load b0 |
676 |
psraw mm7, SHIFT_INV_COL ; dst6 |
psraw mm7, SHIFT_INV_COL ; dst6 |
677 |
movq mm4, mm5 ; a0 |
movq mm4, mm5 ; a0 |
678 |
psraw mm0, SHIFT_INV_COL ; dst5 |
psraw mm0, SHIFT_INV_COL ; dst5 |
679 |
movq [eax+8+1*16], mm3 |
movq [_ECX+8+1*16], mm3 |
680 |
paddsw mm5, mm1 ; a0+b0 |
paddsw mm5, mm1 ; a0+b0 |
681 |
movq [eax+8+2*16], mm6 |
movq [_ECX+8+2*16], mm6 |
682 |
psubsw mm4, mm1 ; a0-b0 |
psubsw mm4, mm1 ; a0-b0 |
683 |
movq mm3, [eax+8+5*16] ; load b3 |
movq mm3, [_ECX+8+5*16] ; load b3 |
684 |
psraw mm5, SHIFT_INV_COL ; dst0 |
psraw mm5, SHIFT_INV_COL ; dst0 |
685 |
movq mm6, mm2 ; a3 |
movq mm6, mm2 ; a3 |
686 |
psraw mm4, SHIFT_INV_COL ; dst7 |
psraw mm4, SHIFT_INV_COL ; dst7 |
687 |
movq [eax+8+5*16], mm0 |
movq [_ECX+8+5*16], mm0 |
688 |
paddsw mm2, mm3 ; a3+b3 |
paddsw mm2, mm3 ; a3+b3 |
689 |
movq [eax+8+6*16], mm7 |
movq [_ECX+8+6*16], mm7 |
690 |
psubsw mm6, mm3 ; a3-b3 |
psubsw mm6, mm3 ; a3-b3 |
691 |
movq [eax+8+0*16], mm5 |
movq [_ECX+8+0*16], mm5 |
692 |
psraw mm2, SHIFT_INV_COL ; dst3 |
psraw mm2, SHIFT_INV_COL ; dst3 |
693 |
movq [eax+8+7*16], mm4 |
movq [_ECX+8+7*16], mm4 |
694 |
psraw mm6, SHIFT_INV_COL ; dst4 |
psraw mm6, SHIFT_INV_COL ; dst4 |
695 |
movq [eax+8+3*16], mm2 |
movq [_ECX+8+3*16], mm2 |
696 |
movq [eax+8+4*16], mm6 |
movq [_ECX+8+4*16], mm6 |
697 |
|
|
698 |
ret |
ret |
699 |
|
ENDFUNC |
700 |
|
|
701 |
|
NON_EXEC_STACK |