annotate src/ffmpeg/libavcodec/i386/idct_mmx_xvid.c @ 830:68562d99230f trunk

[svn] - more symbol cleanups.
author nenolod
date Mon, 12 Mar 2007 15:23:10 -0700
parents e8776388b02a
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
808
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
1 ///****************************************************************************
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
2 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
3 // * XVID MPEG-4 VIDEO CODEC
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
4 // * - MMX and XMM forward discrete cosine transform -
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
5 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
6 // * Copyright(C) 2001 Peter Ross <pross@xvid.org>
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
7 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
8 // * This file is part of FFmpeg.
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
9 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
10 // * FFmpeg is free software; you can redistribute it and/or
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
11 // * modify it under the terms of the GNU Lesser General Public
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
12 // * License as published by the Free Software Foundation; either
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
13 // * version 2.1 of the License, or (at your option) any later version.
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
14 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
15 // * FFmpeg is distributed in the hope that it will be useful,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
16 // * but WITHOUT ANY WARRANTY; without even the implied warranty of
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
17 // * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
18 // * Lesser General Public License for more details.
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
19 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
20 // * You should have received a copy of the GNU Lesser General Public License
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
21 // * along with FFmpeg; if not, write to the Free Software Foundation,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
22 // * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
23 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
24 // * $Id: idct_mmx_xvid.c 6577 2006-10-07 15:30:46Z diego $
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
25 // *
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
26 // ***************************************************************************/
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
27
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
28 // ****************************************************************************
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
29 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
30 // Originally provided by Intel at AP-922
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
31 // http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
32 // (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
33 // but in a limited edition.
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
34 // New macro implements a column part for precise iDCT
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
35 // The routine precision now satisfies IEEE standard 1180-1990.
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
36 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
37 // Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru>
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
38 // Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org>
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
39 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
40 // http://www.elecard.com/peter/idct.html
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
41 // http://www.linuxvideo.org/mpeg2dec/
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
42 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
43 // ***************************************************************************/
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
44 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
45 // These examples contain code fragments for first stage iDCT 8x8
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
46 // (for rows) and first stage DCT 8x8 (for columns)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
47 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
48
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
49 // conversion to gcc syntax by michael niedermayer
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
50
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
51
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
52 #include <inttypes.h>
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
53 #include "../avcodec.h"
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
54
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
55 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
56 // Macros and other preprocessor constants
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
57 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
58
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
59 #define BITS_INV_ACC 5 // 4 or 5 for IEEE
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
60 #define SHIFT_INV_ROW (16 - BITS_INV_ACC) //11
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
61 #define SHIFT_INV_COL (1 + BITS_INV_ACC) //6
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
62 #define RND_INV_ROW (1024 * (6 - BITS_INV_ACC))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
63 #define RND_INV_COL (16 * (BITS_INV_ACC - 3))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
64 #define RND_INV_CORR (RND_INV_COL - 1)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
65
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
66 #define BITS_FRW_ACC 3 // 2 or 3 for accuracy
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
67 #define SHIFT_FRW_COL BITS_FRW_ACC
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
68 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
69 #define RND_FRW_ROW (262144*(BITS_FRW_ACC - 1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
70
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
71
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
72 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
73 // Various memory constants (trigonometric values or rounding values)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
74 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
75
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
76
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
77 static const int16_t tg_1_16[4*4] attribute_used __attribute__ ((aligned(8))) = {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
78 13036,13036,13036,13036, // tg * (2<<16) + 0.5
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
79 27146,27146,27146,27146, // tg * (2<<16) + 0.5
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
80 -21746,-21746,-21746,-21746, // tg * (2<<16) + 0.5
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
81 23170,23170,23170,23170}; // cos * (2<<15) + 0.5
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
82
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
83 static const int32_t rounder_0[2*8] attribute_used __attribute__ ((aligned(8))) = {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
84 65536,65536,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
85 3597,3597,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
86 2260,2260,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
87 1203,1203,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
88 0,0,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
89 120,120,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
90 512,512,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
91 512,512};
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
92
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
93 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
94 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
95 // The first stage iDCT 8x8 - inverse DCTs of rows
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
96 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
97 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
98 // The 8-point inverse DCT direct algorithm
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
99 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
100 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
101 // static const short w[32] = {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
102 // FIX(cos_4_16), FIX(cos_2_16), FIX(cos_4_16), FIX(cos_6_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
103 // FIX(cos_4_16), FIX(cos_6_16), -FIX(cos_4_16), -FIX(cos_2_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
104 // FIX(cos_4_16), -FIX(cos_6_16), -FIX(cos_4_16), FIX(cos_2_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
105 // FIX(cos_4_16), -FIX(cos_2_16), FIX(cos_4_16), -FIX(cos_6_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
106 // FIX(cos_1_16), FIX(cos_3_16), FIX(cos_5_16), FIX(cos_7_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
107 // FIX(cos_3_16), -FIX(cos_7_16), -FIX(cos_1_16), -FIX(cos_5_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
108 // FIX(cos_5_16), -FIX(cos_1_16), FIX(cos_7_16), FIX(cos_3_16),
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
109 // FIX(cos_7_16), -FIX(cos_5_16), FIX(cos_3_16), -FIX(cos_1_16) };
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
110 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
111 // #define DCT_8_INV_ROW(x, y)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
112 // {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
113 // int a0, a1, a2, a3, b0, b1, b2, b3;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
114 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
115 // a0 =x[0]*w[0]+x[2]*w[1]+x[4]*w[2]+x[6]*w[3];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
116 // a1 =x[0]*w[4]+x[2]*w[5]+x[4]*w[6]+x[6]*w[7];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
117 // a2 = x[0] * w[ 8] + x[2] * w[ 9] + x[4] * w[10] + x[6] * w[11];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
118 // a3 = x[0] * w[12] + x[2] * w[13] + x[4] * w[14] + x[6] * w[15];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
119 // b0 = x[1] * w[16] + x[3] * w[17] + x[5] * w[18] + x[7] * w[19];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
120 // b1 = x[1] * w[20] + x[3] * w[21] + x[5] * w[22] + x[7] * w[23];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
121 // b2 = x[1] * w[24] + x[3] * w[25] + x[5] * w[26] + x[7] * w[27];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
122 // b3 = x[1] * w[28] + x[3] * w[29] + x[5] * w[30] + x[7] * w[31];
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
123 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
124 // y[0] = SHIFT_ROUND ( a0 + b0 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
125 // y[1] = SHIFT_ROUND ( a1 + b1 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
126 // y[2] = SHIFT_ROUND ( a2 + b2 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
127 // y[3] = SHIFT_ROUND ( a3 + b3 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
128 // y[4] = SHIFT_ROUND ( a3 - b3 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
129 // y[5] = SHIFT_ROUND ( a2 - b2 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
130 // y[6] = SHIFT_ROUND ( a1 - b1 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
131 // y[7] = SHIFT_ROUND ( a0 - b0 );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
132 // }
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
133 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
134 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
135 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
136 // In this implementation the outputs of the iDCT-1D are multiplied
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
137 // for rows 0,4 - by cos_4_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
138 // for rows 1,7 - by cos_1_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
139 // for rows 2,6 - by cos_2_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
140 // for rows 3,5 - by cos_3_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
141 // and are shifted to the left for better accuracy
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
142 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
143 // For the constants used,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
144 // FIX(float_const) = (short) (float_const * (1<<15) + 0.5)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
145 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
146 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
147
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
148 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
149 // Tables for mmx processors
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
150 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
151
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
152 // Table for rows 0,4 - constants are multiplied by cos_4_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
153 static const int16_t tab_i_04_mmx[32*4] attribute_used __attribute__ ((aligned(8))) = {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
154 16384,16384,16384,-16384, // movq-> w06 w04 w02 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
155 21407,8867,8867,-21407, // w07 w05 w03 w01
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
156 16384,-16384,16384,16384, // w14 w12 w10 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
157 -8867,21407,-21407,-8867, // w15 w13 w11 w09
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
158 22725,12873,19266,-22725, // w22 w20 w18 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
159 19266,4520,-4520,-12873, // w23 w21 w19 w17
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
160 12873,4520,4520,19266, // w30 w28 w26 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
161 -22725,19266,-12873,-22725, // w31 w29 w27 w25
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
162 // Table for rows 1,7 - constants are multiplied by cos_1_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
163 22725,22725,22725,-22725, // movq-> w06 w04 w02 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
164 29692,12299,12299,-29692, // w07 w05 w03 w01
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
165 22725,-22725,22725,22725, // w14 w12 w10 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
166 -12299,29692,-29692,-12299, // w15 w13 w11 w09
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
167 31521,17855,26722,-31521, // w22 w20 w18 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
168 26722,6270,-6270,-17855, // w23 w21 w19 w17
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
169 17855,6270,6270,26722, // w30 w28 w26 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
170 -31521,26722,-17855,-31521, // w31 w29 w27 w25
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
171 // Table for rows 2,6 - constants are multiplied by cos_2_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
172 21407,21407,21407,-21407, // movq-> w06 w04 w02 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
173 27969,11585,11585,-27969, // w07 w05 w03 w01
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
174 21407,-21407,21407,21407, // w14 w12 w10 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
175 -11585,27969,-27969,-11585, // w15 w13 w11 w09
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
176 29692,16819,25172,-29692, // w22 w20 w18 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
177 25172,5906,-5906,-16819, // w23 w21 w19 w17
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
178 16819,5906,5906,25172, // w30 w28 w26 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
179 -29692,25172,-16819,-29692, // w31 w29 w27 w25
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
180 // Table for rows 3,5 - constants are multiplied by cos_3_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
181 19266,19266,19266,-19266, // movq-> w06 w04 w02 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
182 25172,10426,10426,-25172, // w07 w05 w03 w01
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
183 19266,-19266,19266,19266, // w14 w12 w10 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
184 -10426,25172,-25172,-10426, // w15 w13 w11 w09
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
185 26722,15137,22654,-26722, // w22 w20 w18 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
186 22654,5315,-5315,-15137, // w23 w21 w19 w17
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
187 15137,5315,5315,22654, // w30 w28 w26 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
188 -26722,22654,-15137,-26722, // w31 w29 w27 w25
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
189 };
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
190 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
191 // Tables for xmm processors
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
192 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
193
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
194 // %3 for rows 0,4 - constants are multiplied by cos_4_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
195 static const int16_t tab_i_04_xmm[32*4] attribute_used __attribute__ ((aligned(8))) = {
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
196 16384,21407,16384,8867, // movq-> w05 w04 w01 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
197 16384,8867,-16384,-21407, // w07 w06 w03 w02
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
198 16384,-8867,16384,-21407, // w13 w12 w09 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
199 -16384,21407,16384,-8867, // w15 w14 w11 w10
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
200 22725,19266,19266,-4520, // w21 w20 w17 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
201 12873,4520,-22725,-12873, // w23 w22 w19 w18
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
202 12873,-22725,4520,-12873, // w29 w28 w25 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
203 4520,19266,19266,-22725, // w31 w30 w27 w26
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
204 // %3 for rows 1,7 - constants are multiplied by cos_1_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
205 22725,29692,22725,12299, // movq-> w05 w04 w01 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
206 22725,12299,-22725,-29692, // w07 w06 w03 w02
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
207 22725,-12299,22725,-29692, // w13 w12 w09 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
208 -22725,29692,22725,-12299, // w15 w14 w11 w10
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
209 31521,26722,26722,-6270, // w21 w20 w17 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
210 17855,6270,-31521,-17855, // w23 w22 w19 w18
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
211 17855,-31521,6270,-17855, // w29 w28 w25 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
212 6270,26722,26722,-31521, // w31 w30 w27 w26
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
213 // %3 for rows 2,6 - constants are multiplied by cos_2_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
214 21407,27969,21407,11585, // movq-> w05 w04 w01 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
215 21407,11585,-21407,-27969, // w07 w06 w03 w02
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
216 21407,-11585,21407,-27969, // w13 w12 w09 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
217 -21407,27969,21407,-11585, // w15 w14 w11 w10
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
218 29692,25172,25172,-5906, // w21 w20 w17 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
219 16819,5906,-29692,-16819, // w23 w22 w19 w18
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
220 16819,-29692,5906,-16819, // w29 w28 w25 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
221 5906,25172,25172,-29692, // w31 w30 w27 w26
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
222 // %3 for rows 3,5 - constants are multiplied by cos_3_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
223 19266,25172,19266,10426, // movq-> w05 w04 w01 w00
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
224 19266,10426,-19266,-25172, // w07 w06 w03 w02
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
225 19266,-10426,19266,-25172, // w13 w12 w09 w08
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
226 -19266,25172,19266,-10426, // w15 w14 w11 w10
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
227 26722,22654,22654,-5315, // w21 w20 w17 w16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
228 15137,5315,-26722,-15137, // w23 w22 w19 w18
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
229 15137,-26722,5315,-15137, // w29 w28 w25 w24
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
230 5315,22654,22654,-26722, // w31 w30 w27 w26
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
231 };
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
232 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
233 // Helper macros for the code
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
234 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
235
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
236 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
237 // DCT_8_INV_ROW_MMX( INP, OUT, TABLE, ROUNDER
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
238 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
239
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
240 #define DCT_8_INV_ROW_MMX(A1,A2,A3,A4)\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
241 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
242 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
243 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
244 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w06 w04 w02 w00*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
245 "punpcklwd %%mm1,%%mm0 \n\t"/* x5 x1 x4 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
246 "movq %%mm0,%%mm5 \n\t"/* 5 ; x5 x1 x4 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
247 "punpckldq %%mm0,%%mm0 \n\t"/* x4 x0 x4 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
248 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w05 w03 w01*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
249 "punpckhwd %%mm1,%%mm2 \n\t"/* 1 ; x7 x3 x6 x2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
250 "pmaddwd %%mm0,%%mm3 \n\t"/* x4*w06+x0*w04 x4*w02+x0*w00*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
251 "movq %%mm2,%%mm6 \n\t"/* 6 ; x7 x3 x6 x2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
252 "movq 32+" #A3 ",%%mm1 \n\t"/* 1 ; w22 w20 w18 w16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
253 "punpckldq %%mm2,%%mm2 \n\t"/* x6 x2 x6 x2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
254 "pmaddwd %%mm2,%%mm4 \n\t"/* x6*w07+x2*w05 x6*w03+x2*w01*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
255 "punpckhdq %%mm5,%%mm5 \n\t"/* x5 x1 x5 x1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
256 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x4*w14+x0*w12 x4*w10+x0*w08*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
257 "punpckhdq %%mm6,%%mm6 \n\t"/* x7 x3 x7 x3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
258 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w21 w19 w17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
259 "pmaddwd %%mm5,%%mm1 \n\t"/* x5*w22+x1*w20 x5*w18+x1*w16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
260 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
261 "pmaddwd %%mm6,%%mm7 \n\t"/* x7*w23+x3*w21 x7*w19+x3*w17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
262 "pmaddwd 24+" #A3 ",%%mm2 \n\t"/* x6*w15+x2*w13 x6*w11+x2*w09*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
263 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
264 "pmaddwd 48+" #A3 ",%%mm5 \n\t"/* x5*w30+x1*w28 x5*w26+x1*w24*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
265 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
266 "pmaddwd 56+" #A3 ",%%mm6 \n\t"/* x7*w31+x3*w29 x7*w27+x3*w25*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
267 "paddd %%mm7,%%mm1 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
268 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
269 "psubd %%mm1,%%mm3 \n\t"/* a1-b1 a0-b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
270 "psrad $11,%%mm3 \n\t"/* y6=a1-b1 y7=a0-b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
271 "paddd %%mm4,%%mm1 \n\t"/* 4 ; a1+b1 a0+b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
272 "paddd %%mm2,%%mm0 \n\t"/* 2 ; a3=sum(even3) a2=sum(even2)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
273 "psrad $11,%%mm1 \n\t"/* y1=a1+b1 y0=a0+b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
274 "paddd %%mm6,%%mm5 \n\t"/* 6 ; b3=sum(odd3) b2=sum(odd2)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
275 "movq %%mm0,%%mm4 \n\t"/* 4 ; a3 a2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
276 "paddd %%mm5,%%mm0 \n\t"/* a3+b3 a2+b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
277 "psubd %%mm5,%%mm4 \n\t"/* 5 ; a3-b3 a2-b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
278 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
279 "psrad $11,%%mm4 \n\t"/* y4=a3-b3 y5=a2-b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
280 "packssdw %%mm0,%%mm1 \n\t"/* 0 ; y3 y2 y1 y0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
281 "packssdw %%mm3,%%mm4 \n\t"/* 3 ; y6 y7 y4 y5*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
282 "movq %%mm4,%%mm7 \n\t"/* 7 ; y6 y7 y4 y5*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
283 "psrld $16,%%mm4 \n\t"/* 0 y6 0 y4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
284 "pslld $16,%%mm7 \n\t"/* y7 0 y5 0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
285 "movq %%mm1," #A2 " \n\t"/* 1 ; save y3 y2 y1 y0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
286 "por %%mm4,%%mm7 \n\t"/* 4 ; y7 y6 y5 y4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
287 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
288
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
289
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
290 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
291 // DCT_8_INV_ROW_XMM( INP, OUT, TABLE, ROUNDER
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
292 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
293
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
294 #define DCT_8_INV_ROW_XMM(A1,A2,A3,A4)\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
295 "movq " #A1 ",%%mm0 \n\t"/* 0 ; x3 x2 x1 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
296 "movq 8+" #A1 ",%%mm1 \n\t"/* 1 ; x7 x6 x5 x4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
297 "movq %%mm0,%%mm2 \n\t"/* 2 ; x3 x2 x1 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
298 "movq " #A3 ",%%mm3 \n\t"/* 3 ; w05 w04 w01 w00*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
299 "pshufw $0x88,%%mm0,%%mm0 \n\t"/* x2 x0 x2 x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
300 "movq 8+" #A3 ",%%mm4 \n\t"/* 4 ; w07 w06 w03 w02*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
301 "movq %%mm1,%%mm5 \n\t"/* 5 ; x7 x6 x5 x4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
302 "pmaddwd %%mm0,%%mm3 \n\t"/* x2*w05+x0*w04 x2*w01+x0*w00*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
303 "movq 32+" #A3 ",%%mm6 \n\t"/* 6 ; w21 w20 w17 w16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
304 "pshufw $0x88,%%mm1,%%mm1 \n\t"/* x6 x4 x6 x4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
305 "pmaddwd %%mm1,%%mm4 \n\t"/* x6*w07+x4*w06 x6*w03+x4*w02*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
306 "movq 40+" #A3 ",%%mm7 \n\t"/* 7 ; w23 w22 w19 w18*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
307 "pshufw $0xdd,%%mm2,%%mm2 \n\t"/* x3 x1 x3 x1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
308 "pmaddwd %%mm2,%%mm6 \n\t"/* x3*w21+x1*w20 x3*w17+x1*w16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
309 "pshufw $0xdd,%%mm5,%%mm5 \n\t"/* x7 x5 x7 x5*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
310 "pmaddwd %%mm5,%%mm7 \n\t"/* x7*w23+x5*w22 x7*w19+x5*w18*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
311 "paddd " #A4 ",%%mm3 \n\t"/* +%4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
312 "pmaddwd 16+" #A3 ",%%mm0 \n\t"/* x2*w13+x0*w12 x2*w09+x0*w08*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
313 "paddd %%mm4,%%mm3 \n\t"/* 4 ; a1=sum(even1) a0=sum(even0)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
314 "pmaddwd 24+" #A3 ",%%mm1 \n\t"/* x6*w15+x4*w14 x6*w11+x4*w10*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
315 "movq %%mm3,%%mm4 \n\t"/* 4 ; a1 a0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
316 "pmaddwd 48+" #A3 ",%%mm2 \n\t"/* x3*w29+x1*w28 x3*w25+x1*w24*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
317 "paddd %%mm7,%%mm6 \n\t"/* 7 ; b1=sum(odd1) b0=sum(odd0)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
318 "pmaddwd 56+" #A3 ",%%mm5 \n\t"/* x7*w31+x5*w30 x7*w27+x5*w26*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
319 "paddd %%mm6,%%mm3 \n\t"/* a1+b1 a0+b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
320 "paddd " #A4 ",%%mm0 \n\t"/* +%4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
321 "psrad $11,%%mm3 \n\t"/* y1=a1+b1 y0=a0+b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
322 "paddd %%mm1,%%mm0 \n\t"/* 1 ; a3=sum(even3) a2=sum(even2)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
323 "psubd %%mm6,%%mm4 \n\t"/* 6 ; a1-b1 a0-b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
324 "movq %%mm0,%%mm7 \n\t"/* 7 ; a3 a2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
325 "paddd %%mm5,%%mm2 \n\t"/* 5 ; b3=sum(odd3) b2=sum(odd2)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
326 "paddd %%mm2,%%mm0 \n\t"/* a3+b3 a2+b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
327 "psrad $11,%%mm4 \n\t"/* y6=a1-b1 y7=a0-b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
328 "psubd %%mm2,%%mm7 \n\t"/* 2 ; a3-b3 a2-b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
329 "psrad $11,%%mm0 \n\t"/* y3=a3+b3 y2=a2+b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
330 "psrad $11,%%mm7 \n\t"/* y4=a3-b3 y5=a2-b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
331 "packssdw %%mm0,%%mm3 \n\t"/* 0 ; y3 y2 y1 y0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
332 "packssdw %%mm4,%%mm7 \n\t"/* 4 ; y6 y7 y4 y5*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
333 "movq %%mm3, " #A2 " \n\t"/* 3 ; save y3 y2 y1 y0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
334 "pshufw $0xb1,%%mm7,%%mm7 \n\t"/* y7 y6 y5 y4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
335 "movq %%mm7,8 +" #A2 "\n\t"/* 7 ; save y7 y6 y5 y4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
336
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
337
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
338 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
339 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
340 // The first stage DCT 8x8 - forward DCTs of columns
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
341 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
342 // The %2puts are multiplied
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
343 // for rows 0,4 - on cos_4_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
344 // for rows 1,7 - on cos_1_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
345 // for rows 2,6 - on cos_2_16,
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
346 // for rows 3,5 - on cos_3_16
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
347 // and are shifted to the left for rise of accuracy
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
348 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
349 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
350 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
351 // The 8-point scaled forward DCT algorithm (26a8m)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
352 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
353 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
354 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
355 // #define DCT_8_FRW_COL(x, y)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
356 //{
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
357 // short t0, t1, t2, t3, t4, t5, t6, t7;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
358 // short tp03, tm03, tp12, tm12, tp65, tm65;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
359 // short tp465, tm465, tp765, tm765;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
360 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
361 // t0 = LEFT_SHIFT ( x[0] + x[7] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
362 // t1 = LEFT_SHIFT ( x[1] + x[6] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
363 // t2 = LEFT_SHIFT ( x[2] + x[5] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
364 // t3 = LEFT_SHIFT ( x[3] + x[4] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
365 // t4 = LEFT_SHIFT ( x[3] - x[4] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
366 // t5 = LEFT_SHIFT ( x[2] - x[5] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
367 // t6 = LEFT_SHIFT ( x[1] - x[6] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
368 // t7 = LEFT_SHIFT ( x[0] - x[7] );
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
369 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
370 // tp03 = t0 + t3;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
371 // tm03 = t0 - t3;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
372 // tp12 = t1 + t2;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
373 // tm12 = t1 - t2;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
374 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
375 // y[0] = tp03 + tp12;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
376 // y[4] = tp03 - tp12;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
377 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
378 // y[2] = tm03 + tm12 * tg_2_16;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
379 // y[6] = tm03 * tg_2_16 - tm12;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
380 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
381 // tp65 =(t6 +t5 )*cos_4_16;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
382 // tm65 =(t6 -t5 )*cos_4_16;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
383 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
384 // tp765 = t7 + tp65;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
385 // tm765 = t7 - tp65;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
386 // tp465 = t4 + tm65;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
387 // tm465 = t4 - tm65;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
388 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
389 // y[1] = tp765 + tp465 * tg_1_16;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
390 // y[7] = tp765 * tg_1_16 - tp465;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
391 // y[5] = tm765 * tg_3_16 + tm465;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
392 // y[3] = tm765 - tm465 * tg_3_16;
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
393 //}
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
394 //
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
395 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
396
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
397 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
398 // DCT_8_INV_COL_4 INP,OUT
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
399 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
400
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
401 #define DCT_8_INV_COL(A1,A2)\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
402 "movq 2*8(%3),%%mm0\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
403 "movq 16*3+" #A1 ",%%mm3\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
404 "movq %%mm0,%%mm1 \n\t"/* tg_3_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
405 "movq 16*5+" #A1 ",%%mm5\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
406 "pmulhw %%mm3,%%mm0 \n\t"/* x3*(tg_3_16-1)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
407 "movq (%3),%%mm4\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
408 "pmulhw %%mm5,%%mm1 \n\t"/* x5*(tg_3_16-1)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
409 "movq 16*7+" #A1 ",%%mm7\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
410 "movq %%mm4,%%mm2 \n\t"/* tg_1_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
411 "movq 16*1+" #A1 ",%%mm6\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
412 "pmulhw %%mm7,%%mm4 \n\t"/* x7*tg_1_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
413 "paddsw %%mm3,%%mm0 \n\t"/* x3*tg_3_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
414 "pmulhw %%mm6,%%mm2 \n\t"/* x1*tg_1_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
415 "paddsw %%mm3,%%mm1 \n\t"/* x3+x5*(tg_3_16-1)*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
416 "psubsw %%mm5,%%mm0 \n\t"/* x3*tg_3_16-x5 = tm35*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
417 "movq 3*8(%3),%%mm3\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
418 "paddsw %%mm5,%%mm1 \n\t"/* x3+x5*tg_3_16 = tp35*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
419 "paddsw %%mm6,%%mm4 \n\t"/* x1+tg_1_16*x7 = tp17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
420 "psubsw %%mm7,%%mm2 \n\t"/* x1*tg_1_16-x7 = tm17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
421 "movq %%mm4,%%mm5 \n\t"/* tp17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
422 "movq %%mm2,%%mm6 \n\t"/* tm17*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
423 "paddsw %%mm1,%%mm5 \n\t"/* tp17+tp35 = b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
424 "psubsw %%mm0,%%mm6 \n\t"/* tm17-tm35 = b3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
425 "psubsw %%mm1,%%mm4 \n\t"/* tp17-tp35 = t1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
426 "paddsw %%mm0,%%mm2 \n\t"/* tm17+tm35 = t2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
427 "movq 1*8(%3),%%mm7\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
428 "movq %%mm4,%%mm1 \n\t"/* t1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
429 "movq %%mm5,3*16 +" #A2 "\n\t"/* save b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
430 "paddsw %%mm2,%%mm1 \n\t"/* t1+t2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
431 "movq %%mm6,5*16 +" #A2 "\n\t"/* save b3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
432 "psubsw %%mm2,%%mm4 \n\t"/* t1-t2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
433 "movq 2*16+" #A1 ",%%mm5\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
434 "movq %%mm7,%%mm0 \n\t"/* tg_2_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
435 "movq 6*16+" #A1 ",%%mm6\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
436 "pmulhw %%mm5,%%mm0 \n\t"/* x2*tg_2_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
437 "pmulhw %%mm6,%%mm7 \n\t"/* x6*tg_2_16*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
438 "pmulhw %%mm3,%%mm1 \n\t"/* ocos_4_16*(t1+t2) = b1/2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
439 "movq 0*16+" #A1 ",%%mm2\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
440 "pmulhw %%mm3,%%mm4 \n\t"/* ocos_4_16*(t1-t2) = b2/2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
441 "psubsw %%mm6,%%mm0 \n\t"/* t2*tg_2_16-x6 = tm26*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
442 "movq %%mm2,%%mm3 \n\t"/* x0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
443 "movq 4*16+" #A1 ",%%mm6\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
444 "paddsw %%mm5,%%mm7 \n\t"/* x2+x6*tg_2_16 = tp26*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
445 "paddsw %%mm6,%%mm2 \n\t"/* x0+x4 = tp04*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
446 "psubsw %%mm6,%%mm3 \n\t"/* x0-x4 = tm04*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
447 "movq %%mm2,%%mm5 \n\t"/* tp04*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
448 "movq %%mm3,%%mm6 \n\t"/* tm04*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
449 "psubsw %%mm7,%%mm2 \n\t"/* tp04-tp26 = a3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
450 "paddsw %%mm0,%%mm3 \n\t"/* tm04+tm26 = a1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
451 "paddsw %%mm1,%%mm1 \n\t"/* b1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
452 "paddsw %%mm4,%%mm4 \n\t"/* b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
453 "paddsw %%mm7,%%mm5 \n\t"/* tp04+tp26 = a0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
454 "psubsw %%mm0,%%mm6 \n\t"/* tm04-tm26 = a2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
455 "movq %%mm3,%%mm7 \n\t"/* a1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
456 "movq %%mm6,%%mm0 \n\t"/* a2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
457 "paddsw %%mm1,%%mm3 \n\t"/* a1+b1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
458 "paddsw %%mm4,%%mm6 \n\t"/* a2+b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
459 "psraw $6,%%mm3 \n\t"/* dst1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
460 "psubsw %%mm1,%%mm7 \n\t"/* a1-b1*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
461 "psraw $6,%%mm6 \n\t"/* dst2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
462 "psubsw %%mm4,%%mm0 \n\t"/* a2-b2*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
463 "movq 3*16+" #A2 ",%%mm1 \n\t"/* load b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
464 "psraw $6,%%mm7 \n\t"/* dst6*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
465 "movq %%mm5,%%mm4 \n\t"/* a0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
466 "psraw $6,%%mm0 \n\t"/* dst5*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
467 "movq %%mm3,1*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
468 "paddsw %%mm1,%%mm5 \n\t"/* a0+b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
469 "movq %%mm6,2*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
470 "psubsw %%mm1,%%mm4 \n\t"/* a0-b0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
471 "movq 5*16+" #A2 ",%%mm3 \n\t"/* load b3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
472 "psraw $6,%%mm5 \n\t"/* dst0*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
473 "movq %%mm2,%%mm6 \n\t"/* a3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
474 "psraw $6,%%mm4 \n\t"/* dst7*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
475 "movq %%mm0,5*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
476 "paddsw %%mm3,%%mm2 \n\t"/* a3+b3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
477 "movq %%mm7,6*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
478 "psubsw %%mm3,%%mm6 \n\t"/* a3-b3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
479 "movq %%mm5,0*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
480 "psraw $6,%%mm2 \n\t"/* dst3*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
481 "movq %%mm4,7*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
482 "psraw $6,%%mm6 \n\t"/* dst4*/\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
483 "movq %%mm2,3*16+" #A2 "\n\t"\
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
484 "movq %%mm6,4*16+" #A2 "\n\t"
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
485
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
486 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
487 // Code
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
488 //=============================================================================
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
489
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
490 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
491 // void idct_mmx(uint16_t block[64]);
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
492 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
493
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
494
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
495 void ff_idct_xvid_mmx(short *block){
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
496 asm volatile(
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
497 //# Process each row
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
498 DCT_8_INV_ROW_MMX(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
499 DCT_8_INV_ROW_MMX(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
500 DCT_8_INV_ROW_MMX(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
501 DCT_8_INV_ROW_MMX(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
502 DCT_8_INV_ROW_MMX(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
503 DCT_8_INV_ROW_MMX(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
504 DCT_8_INV_ROW_MMX(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
505 DCT_8_INV_ROW_MMX(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
506
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
507 //# Process the columns (4 at a time)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
508 DCT_8_INV_COL(0(%0), 0(%0))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
509 DCT_8_INV_COL(8(%0), 8(%0))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
510 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_mmx), "r"(tg_1_16));
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
511 }
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
512
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
513 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
514 // void idct_xmm(uint16_t block[64]);
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
515 //-----------------------------------------------------------------------------
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
516
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
517
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
518 void ff_idct_xvid_mmx2(short *block){
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
519 asm volatile(
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
520 //# Process each row
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
521 DCT_8_INV_ROW_XMM(0*16(%0), 0*16(%0), 64*0(%2), 8*0(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
522 DCT_8_INV_ROW_XMM(1*16(%0), 1*16(%0), 64*1(%2), 8*1(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
523 DCT_8_INV_ROW_XMM(2*16(%0), 2*16(%0), 64*2(%2), 8*2(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
524 DCT_8_INV_ROW_XMM(3*16(%0), 3*16(%0), 64*3(%2), 8*3(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
525 DCT_8_INV_ROW_XMM(4*16(%0), 4*16(%0), 64*0(%2), 8*4(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
526 DCT_8_INV_ROW_XMM(5*16(%0), 5*16(%0), 64*3(%2), 8*5(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
527 DCT_8_INV_ROW_XMM(6*16(%0), 6*16(%0), 64*2(%2), 8*6(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
528 DCT_8_INV_ROW_XMM(7*16(%0), 7*16(%0), 64*1(%2), 8*7(%1))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
529
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
530 //# Process the columns (4 at a time)
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
531 DCT_8_INV_COL(0(%0), 0(%0))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
532 DCT_8_INV_COL(8(%0), 8(%0))
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
533 :: "r"(block), "r"(rounder_0), "r"(tab_i_04_xmm), "r"(tg_1_16));
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
534 }
e8776388b02a [svn] - add ffmpeg
nenolod
parents:
diff changeset
535