annotate x86/fdct_mmx.c @ 12495:fac8063ed1e7 libavcodec

Allow float values for libmp3lame quality. Patch by James Darnley, james D darnley A gmail
author cehoyos
date Wed, 15 Sep 2010 22:10:13 +0000
parents 7be32921237f
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
1 /*
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
2 * MMX optimized forward DCT
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
3 * The gcc porting is Copyright (c) 2001 Fabrice Bellard.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
4 * cleanup/optimizations are Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
5 * SSE2 optimization is Copyright (c) 2004 Denes Balatoni.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
6 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
7 * from fdctam32.c - AP922 MMX(3D-Now) forward-DCT
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
8 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
9 * Intel Application Note AP-922 - fast, precise implementation of DCT
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
10 * http://developer.intel.com/vtune/cbts/appnotes.htm
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
11 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
12 * Also of inspiration:
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
13 * a page about fdct at http://www.geocities.com/ssavekar/dct.htm
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
14 * Skal's fdct at http://skal.planet-d.net/coding/dct.html
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
15 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
16 * This file is part of FFmpeg.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
17 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
18 * FFmpeg is free software; you can redistribute it and/or
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
19 * modify it under the terms of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
20 * License as published by the Free Software Foundation; either
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
21 * version 2.1 of the License, or (at your option) any later version.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
22 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
23 * FFmpeg is distributed in the hope that it will be useful,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
24 * but WITHOUT ANY WARRANTY; without even the implied warranty of
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
25 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
26 * Lesser General Public License for more details.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
27 *
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
28 * You should have received a copy of the GNU Lesser General Public
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
29 * License along with FFmpeg; if not, write to the Free Software
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
30 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
31 */
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
32
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
33 #include "libavutil/common.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
34 #include "libavcodec/dsputil.h"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
35
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
36 //////////////////////////////////////////////////////////////////////
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
37 //
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
38 // constants for the forward DCT
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
39 // -----------------------------
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
40 //
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
41 // Be sure to check that your compiler is aligning all constants to QWORD
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
42 // (8-byte) memory boundaries! Otherwise the unaligned memory access will
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
43 // severely stall MMX execution.
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
44 //
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
45 //////////////////////////////////////////////////////////////////////
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
46
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
47 #define BITS_FRW_ACC 3 //; 2 or 3 for accuracy
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
48 #define SHIFT_FRW_COL BITS_FRW_ACC
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
49 #define SHIFT_FRW_ROW (BITS_FRW_ACC + 17 - 3)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
50 #define RND_FRW_ROW (1 << (SHIFT_FRW_ROW-1))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
51 //#define RND_FRW_COL (1 << (SHIFT_FRW_COL-1))
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
52
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
53 #define X8(x) x,x,x,x,x,x,x,x
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
54
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
55 //concatenated table, for forward DCT transformation
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
56 DECLARE_ALIGNED(16, static const int16_t, fdct_tg_all_16)[24] = {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
57 X8(13036), // tg * (2<<16) + 0.5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
58 X8(27146), // tg * (2<<16) + 0.5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
59 X8(-21746) // tg * (2<<16) + 0.5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
60 };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
61
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
62 DECLARE_ALIGNED(16, static const int16_t, ocos_4_16)[8] = {
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
63 X8(23170) //cos * (2<<15) + 0.5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
64 };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
65
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
66 DECLARE_ALIGNED(16, static const int16_t, fdct_one_corr)[8] = { X8(1) };
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
67
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
68 DECLARE_ALIGNED(8, static const int32_t, fdct_r_row)[2] = {RND_FRW_ROW, RND_FRW_ROW };
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
69
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
70 static struct
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
71 {
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
72 DECLARE_ALIGNED(16, const int32_t, fdct_r_row_sse2)[4];
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
73 } fdct_r_row_sse2 =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
74 {{
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
75 RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
76 }};
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
77 //DECLARE_ALIGNED(16, static const long, fdct_r_row_sse2)[4] = {RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW, RND_FRW_ROW};
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
78
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
79 DECLARE_ALIGNED(8, static const int16_t, tab_frw_01234567)[] = { // forward_dct coeff table
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
80 16384, 16384, 22725, 19266,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
81 16384, 16384, 12873, 4520,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
82 21407, 8867, 19266, -4520,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
83 -8867, -21407, -22725, -12873,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
84 16384, -16384, 12873, -22725,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
85 -16384, 16384, 4520, 19266,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
86 8867, -21407, 4520, -12873,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
87 21407, -8867, 19266, -22725,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
88
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
89 22725, 22725, 31521, 26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
90 22725, 22725, 17855, 6270,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
91 29692, 12299, 26722, -6270,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
92 -12299, -29692, -31521, -17855,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
93 22725, -22725, 17855, -31521,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
94 -22725, 22725, 6270, 26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
95 12299, -29692, 6270, -17855,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
96 29692, -12299, 26722, -31521,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
97
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
98 21407, 21407, 29692, 25172,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
99 21407, 21407, 16819, 5906,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
100 27969, 11585, 25172, -5906,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
101 -11585, -27969, -29692, -16819,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
102 21407, -21407, 16819, -29692,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
103 -21407, 21407, 5906, 25172,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
104 11585, -27969, 5906, -16819,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
105 27969, -11585, 25172, -29692,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
106
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
107 19266, 19266, 26722, 22654,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
108 19266, 19266, 15137, 5315,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
109 25172, 10426, 22654, -5315,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
110 -10426, -25172, -26722, -15137,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
111 19266, -19266, 15137, -26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
112 -19266, 19266, 5315, 22654,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
113 10426, -25172, 5315, -15137,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
114 25172, -10426, 22654, -26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
115
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
116 16384, 16384, 22725, 19266,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
117 16384, 16384, 12873, 4520,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
118 21407, 8867, 19266, -4520,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
119 -8867, -21407, -22725, -12873,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
120 16384, -16384, 12873, -22725,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
121 -16384, 16384, 4520, 19266,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
122 8867, -21407, 4520, -12873,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
123 21407, -8867, 19266, -22725,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
124
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
125 19266, 19266, 26722, 22654,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
126 19266, 19266, 15137, 5315,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
127 25172, 10426, 22654, -5315,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
128 -10426, -25172, -26722, -15137,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
129 19266, -19266, 15137, -26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
130 -19266, 19266, 5315, 22654,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
131 10426, -25172, 5315, -15137,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
132 25172, -10426, 22654, -26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
133
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
134 21407, 21407, 29692, 25172,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
135 21407, 21407, 16819, 5906,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
136 27969, 11585, 25172, -5906,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
137 -11585, -27969, -29692, -16819,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
138 21407, -21407, 16819, -29692,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
139 -21407, 21407, 5906, 25172,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
140 11585, -27969, 5906, -16819,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
141 27969, -11585, 25172, -29692,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
142
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
143 22725, 22725, 31521, 26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
144 22725, 22725, 17855, 6270,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
145 29692, 12299, 26722, -6270,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
146 -12299, -29692, -31521, -17855,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
147 22725, -22725, 17855, -31521,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
148 -22725, 22725, 6270, 26722,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
149 12299, -29692, 6270, -17855,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
150 29692, -12299, 26722, -31521,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
151 };
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
152
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
153 static struct
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
154 {
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
155 DECLARE_ALIGNED(16, const int16_t, tab_frw_01234567_sse2)[256];
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
156 } tab_frw_01234567_sse2 =
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
157 {{
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
158 //DECLARE_ALIGNED(16, static const int16_t, tab_frw_01234567_sse2)[] = { // forward_dct coeff table
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
159 #define TABLE_SSE2 C4, C4, C1, C3, -C6, -C2, -C1, -C5, \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
160 C4, C4, C5, C7, C2, C6, C3, -C7, \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
161 -C4, C4, C7, C3, C6, -C2, C7, -C5, \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
162 C4, -C4, C5, -C1, C2, -C6, C3, -C1,
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
163 // c1..c7 * cos(pi/4) * 2^15
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
164 #define C1 22725
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
165 #define C2 21407
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
166 #define C3 19266
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
167 #define C4 16384
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
168 #define C5 12873
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
169 #define C6 8867
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
170 #define C7 4520
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
171 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
173 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
174 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
175 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
176 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
177 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
178 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
179 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
180 #define C1 31521
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
181 #define C2 29692
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
182 #define C3 26722
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
183 #define C4 22725
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
184 #define C5 17855
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
185 #define C6 12299
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
186 #define C7 6270
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
187 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
188
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
189 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
190 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
191 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
192 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
193 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
194 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
195 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
196 #define C1 29692
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
197 #define C2 27969
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
198 #define C3 25172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
199 #define C4 21407
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
200 #define C5 16819
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
201 #define C6 11585
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
202 #define C7 5906
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
203 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
204
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
205 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
206 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
207 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
208 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
209 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
210 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
211 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
212 #define C1 26722
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
213 #define C2 25172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
214 #define C3 22654
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
215 #define C4 19266
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
216 #define C5 15137
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
217 #define C6 10426
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
218 #define C7 5315
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
219 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
220
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
221 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
222 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
223 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
224 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
225 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
226 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
227 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
228 #define C1 22725
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
229 #define C2 21407
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
230 #define C3 19266
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
231 #define C4 16384
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
232 #define C5 12873
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
233 #define C6 8867
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
234 #define C7 4520
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
235 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
236
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
237 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
238 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
239 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
240 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
241 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
242 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
243 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
244 #define C1 26722
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
245 #define C2 25172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
246 #define C3 22654
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
247 #define C4 19266
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
248 #define C5 15137
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
249 #define C6 10426
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
250 #define C7 5315
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
251 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
252
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
253 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
254 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
255 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
256 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
257 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
258 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
259 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
260 #define C1 29692
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
261 #define C2 27969
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
262 #define C3 25172
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
263 #define C4 21407
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
264 #define C5 16819
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
265 #define C6 11585
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
266 #define C7 5906
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
267 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
268
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
269 #undef C1
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
270 #undef C2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
271 #undef C3
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
272 #undef C4
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
273 #undef C5
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
274 #undef C6
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
275 #undef C7
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
276 #define C1 31521
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
277 #define C2 29692
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
278 #define C3 26722
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
279 #define C4 22725
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
280 #define C5 17855
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
281 #define C6 12299
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
282 #define C7 6270
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
283 TABLE_SSE2
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
284 }};
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
285
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
286 #define S(s) AV_TOSTRING(s) //AV_STRINGIFY is too long
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
287
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
288 #define FDCT_COL(cpu, mm, mov)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
289 static av_always_inline void fdct_col_##cpu(const int16_t *in, int16_t *out, int offset)\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
290 {\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
291 __asm__ volatile (\
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
292 #mov" 16(%0), %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
293 #mov" 96(%0), %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
294 #mov" %%"#mm"0, %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
295 #mov" 32(%0), %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
296 "paddsw %%"#mm"1, %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
297 #mov" 80(%0), %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
298 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
299 #mov" (%0), %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
300 "paddsw %%"#mm"3, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
301 "paddsw 112(%0), %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
302 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
303 #mov" %%"#mm"0, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
304 "psubsw %%"#mm"1, %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
305 #mov" 16(%1), %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
306 "psubsw %%"#mm"4, %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
307 #mov" 48(%0), %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
308 "pmulhw %%"#mm"0, %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
309 "paddsw 64(%0), %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
310 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
311 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
312 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
313 #mov" %%"#mm"5, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
314 "psubsw %%"#mm"7, %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
315 "paddsw %%"#mm"5, %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
316 "paddsw %%"#mm"7, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
317 "por (%2), %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
318 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
319 "pmulhw 16(%1), %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
320 #mov" %%"#mm"4, %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
321 "psubsw 80(%0), %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
322 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
323 #mov" %%"#mm"1, 32(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
324 "paddsw %%"#mm"6, %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
325 #mov" 48(%0), %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
326 "psllw $"S(SHIFT_FRW_COL)"+1, %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
327 "psubsw 64(%0), %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
328 #mov" %%"#mm"2, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
329 #mov" %%"#mm"4, 64(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
330 "paddsw %%"#mm"3, %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
331 "pmulhw (%4), %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
332 "psubsw %%"#mm"3, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
333 "pmulhw (%4), %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
334 "psubsw %%"#mm"0, %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
335 "por (%2), %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
336 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
337 "por (%2), %%"#mm"2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
338 #mov" %%"#mm"1, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
339 #mov" (%0), %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
340 "paddsw %%"#mm"6, %%"#mm"1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
341 "psubsw 112(%0), %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
342 "psubsw %%"#mm"6, %%"#mm"4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
343 #mov" (%1), %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
344 "psllw $"S(SHIFT_FRW_COL)", %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
345 #mov" 32(%1), %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
346 "pmulhw %%"#mm"1, %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
347 #mov" %%"#mm"7, (%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
348 "pmulhw %%"#mm"4, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
349 #mov" %%"#mm"5, 96(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
350 #mov" %%"#mm"3, %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
351 #mov" 32(%1), %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
352 "psubsw %%"#mm"2, %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
353 "paddsw %%"#mm"2, %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
354 "pmulhw %%"#mm"7, %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
355 "paddsw %%"#mm"3, %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
356 "paddsw %%"#mm"4, %%"#mm"6 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
357 "pmulhw (%1), %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
358 "por (%2), %%"#mm"0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
359 "paddsw %%"#mm"7, %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
360 "psubsw %%"#mm"6, %%"#mm"7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
361 #mov" %%"#mm"0, 16(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
362 "paddsw %%"#mm"4, %%"#mm"5 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
363 #mov" %%"#mm"7, 48(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
364 "psubsw %%"#mm"1, %%"#mm"3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
365 #mov" %%"#mm"5, 80(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
366 #mov" %%"#mm"3, 112(%3) \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
367 : \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
368 : "r" (in + offset), "r" (fdct_tg_all_16), "r" (fdct_one_corr), \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
369 "r" (out + offset), "r" (ocos_4_16)); \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
370 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
371
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
372 FDCT_COL(mmx, mm, movq)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
373 FDCT_COL(sse2, xmm, movdqa)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
374
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
375 static av_always_inline void fdct_row_sse2(const int16_t *in, int16_t *out)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
376 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
377 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
378 #define FDCT_ROW_SSE2_H1(i,t) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
379 "movq " #i "(%0), %%xmm2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
380 "movq " #i "+8(%0), %%xmm0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
381 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
382 "movdqa " #t "+48(%1), %%xmm7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
383 "movdqa " #t "(%1), %%xmm4 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
384 "movdqa " #t "+16(%1), %%xmm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
385
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
386 #define FDCT_ROW_SSE2_H2(i,t) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
387 "movq " #i "(%0), %%xmm2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
388 "movq " #i "+8(%0), %%xmm0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
389 "movdqa " #t "+32(%1), %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
390 "movdqa " #t "+48(%1), %%xmm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
391
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
392 #define FDCT_ROW_SSE2(i) \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
393 "movq %%xmm2, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
394 "pshuflw $27, %%xmm0, %%xmm0 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
395 "paddsw %%xmm0, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
396 "psubsw %%xmm0, %%xmm2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
397 "punpckldq %%xmm2, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
398 "pshufd $78, %%xmm1, %%xmm2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
399 "pmaddwd %%xmm2, %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
400 "pmaddwd %%xmm1, %%xmm7 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
401 "pmaddwd %%xmm5, %%xmm2 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
402 "pmaddwd %%xmm4, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
403 "paddd %%xmm7, %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
404 "paddd %%xmm2, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
405 "paddd %%xmm6, %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
406 "paddd %%xmm6, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
407 "psrad %3, %%xmm3 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
408 "psrad %3, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
409 "packssdw %%xmm3, %%xmm1 \n\t" \
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
410 "movdqa %%xmm1, " #i "(%4) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
411
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
412 "movdqa (%2), %%xmm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
413 FDCT_ROW_SSE2_H1(0,0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
414 FDCT_ROW_SSE2(0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
415 FDCT_ROW_SSE2_H2(64,0)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
416 FDCT_ROW_SSE2(64)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
417
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
418 FDCT_ROW_SSE2_H1(16,64)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
419 FDCT_ROW_SSE2(16)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
420 FDCT_ROW_SSE2_H2(112,64)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
421 FDCT_ROW_SSE2(112)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
422
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
423 FDCT_ROW_SSE2_H1(32,128)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
424 FDCT_ROW_SSE2(32)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
425 FDCT_ROW_SSE2_H2(96,128)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
426 FDCT_ROW_SSE2(96)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
427
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
428 FDCT_ROW_SSE2_H1(48,192)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
429 FDCT_ROW_SSE2(48)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
430 FDCT_ROW_SSE2_H2(80,192)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
431 FDCT_ROW_SSE2(80)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
432 :
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
433 : "r" (in), "r" (tab_frw_01234567_sse2.tab_frw_01234567_sse2), "r" (fdct_r_row_sse2.fdct_r_row_sse2), "i" (SHIFT_FRW_ROW), "r" (out)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
434 );
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
435 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
436
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
437 static av_always_inline void fdct_row_mmx2(const int16_t *in, int16_t *out, const int16_t *table)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
438 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
439 __asm__ volatile (
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
440 "pshufw $0x1B, 8(%0), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
441 "movq (%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
442 "movq %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
443 "paddsw %%mm5, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
444 "psubsw %%mm5, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
445 "movq %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
446 "punpckldq %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
447 "punpckhdq %%mm1, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
448 "movq (%1), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
449 "movq 8(%1), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
450 "movq 16(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
451 "movq 24(%1), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
452 "movq 32(%1), %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
453 "movq 40(%1), %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
454 "pmaddwd %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
455 "pmaddwd %%mm2, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
456 "pmaddwd %%mm0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
457 "pmaddwd %%mm2, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
458 "pmaddwd %%mm0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
459 "pmaddwd %%mm2, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
460 "pmaddwd 48(%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
461 "pmaddwd 56(%1), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
462 "paddd %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
463 "paddd %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
464 "paddd %%mm6, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
465 "paddd %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
466 "movq (%2), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
467 "paddd %%mm0, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
468 "paddd %%mm0, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
469 "paddd %%mm0, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
470 "paddd %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
471 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
472 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
473 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
474 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
475 "packssdw %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
476 "packssdw %%mm2, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
477 "movq %%mm3, (%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
478 "movq %%mm7, 8(%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
479 :
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
480 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
481 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
482
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
483 static av_always_inline void fdct_row_mmx(const int16_t *in, int16_t *out, const int16_t *table)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
484 {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
485 //FIXME reorder (I do not have an old MMX-only CPU here to benchmark ...)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
486 __asm__ volatile(
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
487 "movd 12(%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
488 "punpcklwd 8(%0), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
489 "movq %%mm1, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
490 "psrlq $0x20, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
491 "movq 0(%0), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
492 "punpcklwd %%mm2, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
493 "movq %%mm0, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
494 "paddsw %%mm1, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
495 "psubsw %%mm1, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
496 "movq %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
497 "punpckldq %%mm5, %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
498 "punpckhdq %%mm5, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
499 "movq 0(%1), %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
500 "movq 8(%1), %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
501 "movq 16(%1), %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
502 "movq 24(%1), %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
503 "movq 32(%1), %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
504 "movq 40(%1), %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
505 "pmaddwd %%mm0, %%mm1 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
506 "pmaddwd %%mm2, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
507 "pmaddwd %%mm0, %%mm4 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
508 "pmaddwd %%mm2, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
509 "pmaddwd %%mm0, %%mm6 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
510 "pmaddwd %%mm2, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
511 "pmaddwd 48(%1), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
512 "pmaddwd 56(%1), %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
513 "paddd %%mm1, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
514 "paddd %%mm4, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
515 "paddd %%mm6, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
516 "paddd %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
517 "movq (%2), %%mm0 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
518 "paddd %%mm0, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
519 "paddd %%mm0, %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
520 "paddd %%mm0, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
521 "paddd %%mm0, %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
522 "psrad $"S(SHIFT_FRW_ROW)", %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
523 "psrad $"S(SHIFT_FRW_ROW)", %%mm5 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
524 "psrad $"S(SHIFT_FRW_ROW)", %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
525 "psrad $"S(SHIFT_FRW_ROW)", %%mm2 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
526 "packssdw %%mm5, %%mm3 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
527 "packssdw %%mm2, %%mm7 \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
528 "movq %%mm3, 0(%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
529 "movq %%mm7, 8(%3) \n\t"
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
530 :
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
531 : "r" (in), "r" (table), "r" (fdct_r_row), "r" (out));
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
532 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
533
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
534 void ff_fdct_mmx(int16_t *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
535 {
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
536 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
537 int16_t * block1= (int16_t*)align_tmp;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
538 const int16_t *table= tab_frw_01234567;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
539 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
540
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
541 fdct_col_mmx(block, block1, 0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
542 fdct_col_mmx(block, block1, 4);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
543
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
544 for(i=8;i>0;i--) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
545 fdct_row_mmx(block1, block, table);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
546 block1 += 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
547 table += 32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
548 block += 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
549 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
550 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
551
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
552 void ff_fdct_mmx2(int16_t *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
553 {
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
554 DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
555 int16_t *block1= (int16_t*)align_tmp;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
556 const int16_t *table= tab_frw_01234567;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
557 int i;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
558
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
559 fdct_col_mmx(block, block1, 0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
560 fdct_col_mmx(block, block1, 4);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
561
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
562 for(i=8;i>0;i--) {
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
563 fdct_row_mmx2(block1, block, table);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
564 block1 += 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
565 table += 32;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
566 block += 8;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
567 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
568 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
569
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
570 void ff_fdct_sse2(int16_t *block)
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
571 {
11508
7be32921237f Replace remaining uses of ATTR_ALIGNED with DECLARE_ALIGNED
mru
parents: 8430
diff changeset
572 DECLARE_ALIGNED(16, int64_t, align_tmp)[16];
8430
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
573 int16_t * const block1= (int16_t*)align_tmp;
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
574
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
575 fdct_col_sse2(block, block1, 0);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
576 fdct_row_sse2(block1, block);
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
577 }
7768bdfd4f7b Rename libavcodec/i386/ --> libavcodec/x86/.
diego
parents:
diff changeset
578